xref: /freebsd/sys/vm/uma_core.c (revision 3cbb4cc200f8a0ad7ed08233425ea54524a21f1c)
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 #define	BUCKET_MIN	2
252 
253 struct uma_bucket_zone bucket_zones[] = {
254 	/* Literal bucket sizes. */
255 	{ NULL, "2 Bucket", 2, 4096 },
256 	{ NULL, "4 Bucket", 4, 3072 },
257 	{ NULL, "8 Bucket", 8, 2048 },
258 	{ NULL, "16 Bucket", 16, 1024 },
259 	/* Rounded down power of 2 sizes for efficiency. */
260 	{ NULL, "32 Bucket", BUCKET_SIZE(32), 512 },
261 	{ NULL, "64 Bucket", BUCKET_SIZE(64), 256 },
262 	{ NULL, "128 Bucket", BUCKET_SIZE(128), 128 },
263 	{ NULL, "256 Bucket", BUCKET_SIZE(256), 64 },
264 	{ NULL, NULL, 0}
265 };
266 
267 /*
268  * Flags and enumerations to be passed to internal functions.
269  */
270 enum zfreeskip {
271 	SKIP_NONE =	0,
272 	SKIP_CNT =	0x00000001,
273 	SKIP_DTOR =	0x00010000,
274 	SKIP_FINI =	0x00020000,
275 };
276 
277 /* Prototypes.. */
278 
279 void	uma_startup1(vm_offset_t);
280 void	uma_startup2(void);
281 
282 static void *noobj_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
283 static void *page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
284 static void *pcpu_page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
285 static void *startup_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
286 static void *contig_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
287 static void page_free(void *, vm_size_t, uint8_t);
288 static void pcpu_page_free(void *, vm_size_t, uint8_t);
289 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int, int, int);
290 static void cache_drain(uma_zone_t);
291 static void bucket_drain(uma_zone_t, uma_bucket_t);
292 static void bucket_cache_reclaim(uma_zone_t zone, bool);
293 static int keg_ctor(void *, int, void *, int);
294 static void keg_dtor(void *, int, void *);
295 static int zone_ctor(void *, int, void *, int);
296 static void zone_dtor(void *, int, void *);
297 static inline void item_dtor(uma_zone_t zone, void *item, int size,
298     void *udata, enum zfreeskip skip);
299 static int zero_init(void *, int, int);
300 static void zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata,
301     int itemdomain, bool ws);
302 static void zone_foreach(void (*zfunc)(uma_zone_t, void *), void *);
303 static void zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *), void *);
304 static void zone_timeout(uma_zone_t zone, void *);
305 static int hash_alloc(struct uma_hash *, u_int);
306 static int hash_expand(struct uma_hash *, struct uma_hash *);
307 static void hash_free(struct uma_hash *hash);
308 static void uma_timeout(void *);
309 static void uma_shutdown(void);
310 static void *zone_alloc_item(uma_zone_t, void *, int, int);
311 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip);
312 static int zone_alloc_limit(uma_zone_t zone, int count, int flags);
313 static void zone_free_limit(uma_zone_t zone, int count);
314 static void bucket_enable(void);
315 static void bucket_init(void);
316 static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int);
317 static void bucket_free(uma_zone_t zone, uma_bucket_t, void *);
318 static void bucket_zone_drain(void);
319 static uma_bucket_t zone_alloc_bucket(uma_zone_t, void *, int, int);
320 static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab);
321 static void slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item);
322 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit,
323     uma_fini fini, int align, uint32_t flags);
324 static int zone_import(void *, void **, int, int, int);
325 static void zone_release(void *, void **, int);
326 static bool cache_alloc(uma_zone_t, uma_cache_t, void *, int);
327 static bool cache_free(uma_zone_t, uma_cache_t, void *, void *, int);
328 
329 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS);
330 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS);
331 static int sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS);
332 static int sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS);
333 static int sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS);
334 static int sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS);
335 static int sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS);
336 
337 static uint64_t uma_zone_get_allocs(uma_zone_t zone);
338 
339 static SYSCTL_NODE(_vm, OID_AUTO, debug, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
340     "Memory allocation debugging");
341 
342 #ifdef INVARIANTS
343 static uint64_t uma_keg_get_allocs(uma_keg_t zone);
344 static inline struct noslabbits *slab_dbg_bits(uma_slab_t slab, uma_keg_t keg);
345 
346 static bool uma_dbg_kskip(uma_keg_t keg, void *mem);
347 static bool uma_dbg_zskip(uma_zone_t zone, void *mem);
348 static void uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item);
349 static void uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item);
350 
351 static u_int dbg_divisor = 1;
352 SYSCTL_UINT(_vm_debug, OID_AUTO, divisor,
353     CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &dbg_divisor, 0,
354     "Debug & thrash every this item in memory allocator");
355 
356 static counter_u64_t uma_dbg_cnt = EARLY_COUNTER;
357 static counter_u64_t uma_skip_cnt = EARLY_COUNTER;
358 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, trashed, CTLFLAG_RD,
359     &uma_dbg_cnt, "memory items debugged");
360 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, skipped, CTLFLAG_RD,
361     &uma_skip_cnt, "memory items skipped, not debugged");
362 #endif
363 
364 SYSCTL_NODE(_vm, OID_AUTO, uma, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
365     "Universal Memory Allocator");
366 
367 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLFLAG_MPSAFE|CTLTYPE_INT,
368     0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones");
369 
370 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLFLAG_MPSAFE|CTLTYPE_STRUCT,
371     0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats");
372 
373 static int zone_warnings = 1;
374 SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RWTUN, &zone_warnings, 0,
375     "Warn when UMA zones becomes full");
376 
377 static int multipage_slabs = 1;
378 TUNABLE_INT("vm.debug.uma_multipage_slabs", &multipage_slabs);
379 SYSCTL_INT(_vm_debug, OID_AUTO, uma_multipage_slabs,
380     CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &multipage_slabs, 0,
381     "UMA may choose larger slab sizes for better efficiency");
382 
383 /*
384  * Select the slab zone for an offpage slab with the given maximum item count.
385  */
386 static inline uma_zone_t
387 slabzone(int ipers)
388 {
389 
390 	return (slabzones[ipers > SLABZONE0_SETSIZE]);
391 }
392 
393 /*
394  * This routine checks to see whether or not it's safe to enable buckets.
395  */
396 static void
397 bucket_enable(void)
398 {
399 
400 	KASSERT(booted >= BOOT_KVA, ("Bucket enable before init"));
401 	bucketdisable = vm_page_count_min();
402 }
403 
404 /*
405  * Initialize bucket_zones, the array of zones of buckets of various sizes.
406  *
407  * For each zone, calculate the memory required for each bucket, consisting
408  * of the header and an array of pointers.
409  */
410 static void
411 bucket_init(void)
412 {
413 	struct uma_bucket_zone *ubz;
414 	int size;
415 
416 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) {
417 		size = roundup(sizeof(struct uma_bucket), sizeof(void *));
418 		size += sizeof(void *) * ubz->ubz_entries;
419 		ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
420 		    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
421 		    UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET |
422 		    UMA_ZONE_FIRSTTOUCH);
423 	}
424 }
425 
426 /*
427  * Given a desired number of entries for a bucket, return the zone from which
428  * to allocate the bucket.
429  */
430 static struct uma_bucket_zone *
431 bucket_zone_lookup(int entries)
432 {
433 	struct uma_bucket_zone *ubz;
434 
435 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
436 		if (ubz->ubz_entries >= entries)
437 			return (ubz);
438 	ubz--;
439 	return (ubz);
440 }
441 
442 static struct uma_bucket_zone *
443 bucket_zone_max(uma_zone_t zone, int nitems)
444 {
445 	struct uma_bucket_zone *ubz;
446 	int bpcpu;
447 
448 	bpcpu = 2;
449 	if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
450 		/* Count the cross-domain bucket. */
451 		bpcpu++;
452 
453 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
454 		if (ubz->ubz_entries * bpcpu * mp_ncpus > nitems)
455 			break;
456 	if (ubz == &bucket_zones[0])
457 		ubz = NULL;
458 	else
459 		ubz--;
460 	return (ubz);
461 }
462 
463 static int
464 bucket_select(int size)
465 {
466 	struct uma_bucket_zone *ubz;
467 
468 	ubz = &bucket_zones[0];
469 	if (size > ubz->ubz_maxsize)
470 		return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1);
471 
472 	for (; ubz->ubz_entries != 0; ubz++)
473 		if (ubz->ubz_maxsize < size)
474 			break;
475 	ubz--;
476 	return (ubz->ubz_entries);
477 }
478 
479 static uma_bucket_t
480 bucket_alloc(uma_zone_t zone, void *udata, int flags)
481 {
482 	struct uma_bucket_zone *ubz;
483 	uma_bucket_t bucket;
484 
485 	/*
486 	 * Don't allocate buckets early in boot.
487 	 */
488 	if (__predict_false(booted < BOOT_KVA))
489 		return (NULL);
490 
491 	/*
492 	 * To limit bucket recursion we store the original zone flags
493 	 * in a cookie passed via zalloc_arg/zfree_arg.  This allows the
494 	 * NOVM flag to persist even through deep recursions.  We also
495 	 * store ZFLAG_BUCKET once we have recursed attempting to allocate
496 	 * a bucket for a bucket zone so we do not allow infinite bucket
497 	 * recursion.  This cookie will even persist to frees of unused
498 	 * buckets via the allocation path or bucket allocations in the
499 	 * free path.
500 	 */
501 	if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
502 		udata = (void *)(uintptr_t)zone->uz_flags;
503 	else {
504 		if ((uintptr_t)udata & UMA_ZFLAG_BUCKET)
505 			return (NULL);
506 		udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET);
507 	}
508 	if (((uintptr_t)udata & UMA_ZONE_VM) != 0)
509 		flags |= M_NOVM;
510 	ubz = bucket_zone_lookup(zone->uz_bucket_size);
511 	if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0)
512 		ubz++;
513 	bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags);
514 	if (bucket) {
515 #ifdef INVARIANTS
516 		bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
517 #endif
518 		bucket->ub_cnt = 0;
519 		bucket->ub_entries = ubz->ubz_entries;
520 		bucket->ub_seq = SMR_SEQ_INVALID;
521 		CTR3(KTR_UMA, "bucket_alloc: zone %s(%p) allocated bucket %p",
522 		    zone->uz_name, zone, bucket);
523 	}
524 
525 	return (bucket);
526 }
527 
528 static void
529 bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata)
530 {
531 	struct uma_bucket_zone *ubz;
532 
533 	if (bucket->ub_cnt != 0)
534 		bucket_drain(zone, bucket);
535 
536 	KASSERT(bucket->ub_cnt == 0,
537 	    ("bucket_free: Freeing a non free bucket."));
538 	KASSERT(bucket->ub_seq == SMR_SEQ_INVALID,
539 	    ("bucket_free: Freeing an SMR bucket."));
540 	if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
541 		udata = (void *)(uintptr_t)zone->uz_flags;
542 	ubz = bucket_zone_lookup(bucket->ub_entries);
543 	uma_zfree_arg(ubz->ubz_zone, bucket, udata);
544 }
545 
546 static void
547 bucket_zone_drain(void)
548 {
549 	struct uma_bucket_zone *ubz;
550 
551 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
552 		uma_zone_reclaim(ubz->ubz_zone, UMA_RECLAIM_DRAIN);
553 }
554 
555 /*
556  * Acquire the domain lock and record contention.
557  */
558 static uma_zone_domain_t
559 zone_domain_lock(uma_zone_t zone, int domain)
560 {
561 	uma_zone_domain_t zdom;
562 	bool lockfail;
563 
564 	zdom = ZDOM_GET(zone, domain);
565 	lockfail = false;
566 	if (ZDOM_OWNED(zdom))
567 		lockfail = true;
568 	ZDOM_LOCK(zdom);
569 	/* This is unsynchronized.  The counter does not need to be precise. */
570 	if (lockfail && zone->uz_bucket_size < zone->uz_bucket_size_max)
571 		zone->uz_bucket_size++;
572 	return (zdom);
573 }
574 
575 /*
576  * Search for the domain with the least cached items and return it if it
577  * is out of balance with the preferred domain.
578  */
579 static __noinline int
580 zone_domain_lowest(uma_zone_t zone, int pref)
581 {
582 	long least, nitems, prefitems;
583 	int domain;
584 	int i;
585 
586 	prefitems = least = LONG_MAX;
587 	domain = 0;
588 	for (i = 0; i < vm_ndomains; i++) {
589 		nitems = ZDOM_GET(zone, i)->uzd_nitems;
590 		if (nitems < least) {
591 			domain = i;
592 			least = nitems;
593 		}
594 		if (domain == pref)
595 			prefitems = nitems;
596 	}
597 	if (prefitems < least * 2)
598 		return (pref);
599 
600 	return (domain);
601 }
602 
603 /*
604  * Search for the domain with the most cached items and return it or the
605  * preferred domain if it has enough to proceed.
606  */
607 static __noinline int
608 zone_domain_highest(uma_zone_t zone, int pref)
609 {
610 	long most, nitems;
611 	int domain;
612 	int i;
613 
614 	if (ZDOM_GET(zone, pref)->uzd_nitems > BUCKET_MAX)
615 		return (pref);
616 
617 	most = 0;
618 	domain = 0;
619 	for (i = 0; i < vm_ndomains; i++) {
620 		nitems = ZDOM_GET(zone, i)->uzd_nitems;
621 		if (nitems > most) {
622 			domain = i;
623 			most = nitems;
624 		}
625 	}
626 
627 	return (domain);
628 }
629 
630 /*
631  * Safely subtract cnt from imax.
632  */
633 static void
634 zone_domain_imax_sub(uma_zone_domain_t zdom, int cnt)
635 {
636 	long new;
637 	long old;
638 
639 	old = zdom->uzd_imax;
640 	do {
641 		if (old <= cnt)
642 			new = 0;
643 		else
644 			new = old - cnt;
645 	} while (atomic_fcmpset_long(&zdom->uzd_imax, &old, new) == 0);
646 }
647 
648 /*
649  * Set the maximum imax value.
650  */
651 static void
652 zone_domain_imax_set(uma_zone_domain_t zdom, int nitems)
653 {
654 	long old;
655 
656 	old = zdom->uzd_imax;
657 	do {
658 		if (old >= nitems)
659 			break;
660 	} while (atomic_fcmpset_long(&zdom->uzd_imax, &old, nitems) == 0);
661 }
662 
663 /*
664  * Attempt to satisfy an allocation by retrieving a full bucket from one of the
665  * zone's caches.  If a bucket is found the zone is not locked on return.
666  */
667 static uma_bucket_t
668 zone_fetch_bucket(uma_zone_t zone, uma_zone_domain_t zdom, bool reclaim)
669 {
670 	uma_bucket_t bucket;
671 	int i;
672 	bool dtor = false;
673 
674 	ZDOM_LOCK_ASSERT(zdom);
675 
676 	if ((bucket = STAILQ_FIRST(&zdom->uzd_buckets)) == NULL)
677 		return (NULL);
678 
679 	/* SMR Buckets can not be re-used until readers expire. */
680 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
681 	    bucket->ub_seq != SMR_SEQ_INVALID) {
682 		if (!smr_poll(zone->uz_smr, bucket->ub_seq, false))
683 			return (NULL);
684 		bucket->ub_seq = SMR_SEQ_INVALID;
685 		dtor = (zone->uz_dtor != NULL) || UMA_ALWAYS_CTORDTOR;
686 		if (STAILQ_NEXT(bucket, ub_link) != NULL)
687 			zdom->uzd_seq = STAILQ_NEXT(bucket, ub_link)->ub_seq;
688 	}
689 	STAILQ_REMOVE_HEAD(&zdom->uzd_buckets, ub_link);
690 
691 	KASSERT(zdom->uzd_nitems >= bucket->ub_cnt,
692 	    ("%s: item count underflow (%ld, %d)",
693 	    __func__, zdom->uzd_nitems, bucket->ub_cnt));
694 	KASSERT(bucket->ub_cnt > 0,
695 	    ("%s: empty bucket in bucket cache", __func__));
696 	zdom->uzd_nitems -= bucket->ub_cnt;
697 
698 	/*
699 	 * Shift the bounds of the current WSS interval to avoid
700 	 * perturbing the estimate.
701 	 */
702 	if (reclaim) {
703 		zdom->uzd_imin -= lmin(zdom->uzd_imin, bucket->ub_cnt);
704 		zone_domain_imax_sub(zdom, bucket->ub_cnt);
705 	} else if (zdom->uzd_imin > zdom->uzd_nitems)
706 		zdom->uzd_imin = zdom->uzd_nitems;
707 
708 	ZDOM_UNLOCK(zdom);
709 	if (dtor)
710 		for (i = 0; i < bucket->ub_cnt; i++)
711 			item_dtor(zone, bucket->ub_bucket[i], zone->uz_size,
712 			    NULL, SKIP_NONE);
713 
714 	return (bucket);
715 }
716 
717 /*
718  * Insert a full bucket into the specified cache.  The "ws" parameter indicates
719  * whether the bucket's contents should be counted as part of the zone's working
720  * set.  The bucket may be freed if it exceeds the bucket limit.
721  */
722 static void
723 zone_put_bucket(uma_zone_t zone, int domain, uma_bucket_t bucket, void *udata,
724     const bool ws)
725 {
726 	uma_zone_domain_t zdom;
727 
728 	/* We don't cache empty buckets.  This can happen after a reclaim. */
729 	if (bucket->ub_cnt == 0)
730 		goto out;
731 	zdom = zone_domain_lock(zone, domain);
732 
733 	/*
734 	 * Conditionally set the maximum number of items.
735 	 */
736 	zdom->uzd_nitems += bucket->ub_cnt;
737 	if (__predict_true(zdom->uzd_nitems < zone->uz_bucket_max)) {
738 		if (ws)
739 			zone_domain_imax_set(zdom, zdom->uzd_nitems);
740 		if (STAILQ_EMPTY(&zdom->uzd_buckets))
741 			zdom->uzd_seq = bucket->ub_seq;
742 
743 		/*
744 		 * Try to promote reuse of recently used items.  For items
745 		 * protected by SMR, try to defer reuse to minimize polling.
746 		 */
747 		if (bucket->ub_seq == SMR_SEQ_INVALID)
748 			STAILQ_INSERT_HEAD(&zdom->uzd_buckets, bucket, ub_link);
749 		else
750 			STAILQ_INSERT_TAIL(&zdom->uzd_buckets, bucket, ub_link);
751 		ZDOM_UNLOCK(zdom);
752 		return;
753 	}
754 	zdom->uzd_nitems -= bucket->ub_cnt;
755 	ZDOM_UNLOCK(zdom);
756 out:
757 	bucket_free(zone, bucket, udata);
758 }
759 
760 /* Pops an item out of a per-cpu cache bucket. */
761 static inline void *
762 cache_bucket_pop(uma_cache_t cache, uma_cache_bucket_t bucket)
763 {
764 	void *item;
765 
766 	CRITICAL_ASSERT(curthread);
767 
768 	bucket->ucb_cnt--;
769 	item = bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt];
770 #ifdef INVARIANTS
771 	bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = NULL;
772 	KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled."));
773 #endif
774 	cache->uc_allocs++;
775 
776 	return (item);
777 }
778 
779 /* Pushes an item into a per-cpu cache bucket. */
780 static inline void
781 cache_bucket_push(uma_cache_t cache, uma_cache_bucket_t bucket, void *item)
782 {
783 
784 	CRITICAL_ASSERT(curthread);
785 	KASSERT(bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] == NULL,
786 	    ("uma_zfree: Freeing to non free bucket index."));
787 
788 	bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = item;
789 	bucket->ucb_cnt++;
790 	cache->uc_frees++;
791 }
792 
793 /*
794  * Unload a UMA bucket from a per-cpu cache.
795  */
796 static inline uma_bucket_t
797 cache_bucket_unload(uma_cache_bucket_t bucket)
798 {
799 	uma_bucket_t b;
800 
801 	b = bucket->ucb_bucket;
802 	if (b != NULL) {
803 		MPASS(b->ub_entries == bucket->ucb_entries);
804 		b->ub_cnt = bucket->ucb_cnt;
805 		bucket->ucb_bucket = NULL;
806 		bucket->ucb_entries = bucket->ucb_cnt = 0;
807 	}
808 
809 	return (b);
810 }
811 
812 static inline uma_bucket_t
813 cache_bucket_unload_alloc(uma_cache_t cache)
814 {
815 
816 	return (cache_bucket_unload(&cache->uc_allocbucket));
817 }
818 
819 static inline uma_bucket_t
820 cache_bucket_unload_free(uma_cache_t cache)
821 {
822 
823 	return (cache_bucket_unload(&cache->uc_freebucket));
824 }
825 
826 static inline uma_bucket_t
827 cache_bucket_unload_cross(uma_cache_t cache)
828 {
829 
830 	return (cache_bucket_unload(&cache->uc_crossbucket));
831 }
832 
833 /*
834  * Load a bucket into a per-cpu cache bucket.
835  */
836 static inline void
837 cache_bucket_load(uma_cache_bucket_t bucket, uma_bucket_t b)
838 {
839 
840 	CRITICAL_ASSERT(curthread);
841 	MPASS(bucket->ucb_bucket == NULL);
842 	MPASS(b->ub_seq == SMR_SEQ_INVALID);
843 
844 	bucket->ucb_bucket = b;
845 	bucket->ucb_cnt = b->ub_cnt;
846 	bucket->ucb_entries = b->ub_entries;
847 }
848 
849 static inline void
850 cache_bucket_load_alloc(uma_cache_t cache, uma_bucket_t b)
851 {
852 
853 	cache_bucket_load(&cache->uc_allocbucket, b);
854 }
855 
856 static inline void
857 cache_bucket_load_free(uma_cache_t cache, uma_bucket_t b)
858 {
859 
860 	cache_bucket_load(&cache->uc_freebucket, b);
861 }
862 
863 #ifdef NUMA
864 static inline void
865 cache_bucket_load_cross(uma_cache_t cache, uma_bucket_t b)
866 {
867 
868 	cache_bucket_load(&cache->uc_crossbucket, b);
869 }
870 #endif
871 
872 /*
873  * Copy and preserve ucb_spare.
874  */
875 static inline void
876 cache_bucket_copy(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
877 {
878 
879 	b1->ucb_bucket = b2->ucb_bucket;
880 	b1->ucb_entries = b2->ucb_entries;
881 	b1->ucb_cnt = b2->ucb_cnt;
882 }
883 
884 /*
885  * Swap two cache buckets.
886  */
887 static inline void
888 cache_bucket_swap(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
889 {
890 	struct uma_cache_bucket b3;
891 
892 	CRITICAL_ASSERT(curthread);
893 
894 	cache_bucket_copy(&b3, b1);
895 	cache_bucket_copy(b1, b2);
896 	cache_bucket_copy(b2, &b3);
897 }
898 
899 /*
900  * Attempt to fetch a bucket from a zone on behalf of the current cpu cache.
901  */
902 static uma_bucket_t
903 cache_fetch_bucket(uma_zone_t zone, uma_cache_t cache, int domain)
904 {
905 	uma_zone_domain_t zdom;
906 	uma_bucket_t bucket;
907 
908 	/*
909 	 * Avoid the lock if possible.
910 	 */
911 	zdom = ZDOM_GET(zone, domain);
912 	if (zdom->uzd_nitems == 0)
913 		return (NULL);
914 
915 	if ((cache_uz_flags(cache) & UMA_ZONE_SMR) != 0 &&
916 	    !smr_poll(zone->uz_smr, zdom->uzd_seq, false))
917 		return (NULL);
918 
919 	/*
920 	 * Check the zone's cache of buckets.
921 	 */
922 	zdom = zone_domain_lock(zone, domain);
923 	if ((bucket = zone_fetch_bucket(zone, zdom, false)) != NULL)
924 		return (bucket);
925 	ZDOM_UNLOCK(zdom);
926 
927 	return (NULL);
928 }
929 
930 static void
931 zone_log_warning(uma_zone_t zone)
932 {
933 	static const struct timeval warninterval = { 300, 0 };
934 
935 	if (!zone_warnings || zone->uz_warning == NULL)
936 		return;
937 
938 	if (ratecheck(&zone->uz_ratecheck, &warninterval))
939 		printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning);
940 }
941 
942 static inline void
943 zone_maxaction(uma_zone_t zone)
944 {
945 
946 	if (zone->uz_maxaction.ta_func != NULL)
947 		taskqueue_enqueue(taskqueue_thread, &zone->uz_maxaction);
948 }
949 
950 /*
951  * Routine called by timeout which is used to fire off some time interval
952  * based calculations.  (stats, hash size, etc.)
953  *
954  * Arguments:
955  *	arg   Unused
956  *
957  * Returns:
958  *	Nothing
959  */
960 static void
961 uma_timeout(void *unused)
962 {
963 	bucket_enable();
964 	zone_foreach(zone_timeout, NULL);
965 
966 	/* Reschedule this event */
967 	callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
968 }
969 
970 /*
971  * Update the working set size estimate for the zone's bucket cache.
972  * The constants chosen here are somewhat arbitrary.  With an update period of
973  * 20s (UMA_TIMEOUT), this estimate is dominated by zone activity over the
974  * last 100s.
975  */
976 static void
977 zone_domain_update_wss(uma_zone_domain_t zdom)
978 {
979 	long wss;
980 
981 	ZDOM_LOCK(zdom);
982 	MPASS(zdom->uzd_imax >= zdom->uzd_imin);
983 	wss = zdom->uzd_imax - zdom->uzd_imin;
984 	zdom->uzd_imax = zdom->uzd_imin = zdom->uzd_nitems;
985 	zdom->uzd_wss = (4 * wss + zdom->uzd_wss) / 5;
986 	ZDOM_UNLOCK(zdom);
987 }
988 
989 /*
990  * Routine to perform timeout driven calculations.  This expands the
991  * hashes and does per cpu statistics aggregation.
992  *
993  *  Returns nothing.
994  */
995 static void
996 zone_timeout(uma_zone_t zone, void *unused)
997 {
998 	uma_keg_t keg;
999 	u_int slabs, pages;
1000 
1001 	if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
1002 		goto update_wss;
1003 
1004 	keg = zone->uz_keg;
1005 
1006 	/*
1007 	 * Hash zones are non-numa by definition so the first domain
1008 	 * is the only one present.
1009 	 */
1010 	KEG_LOCK(keg, 0);
1011 	pages = keg->uk_domain[0].ud_pages;
1012 
1013 	/*
1014 	 * Expand the keg hash table.
1015 	 *
1016 	 * This is done if the number of slabs is larger than the hash size.
1017 	 * What I'm trying to do here is completely reduce collisions.  This
1018 	 * may be a little aggressive.  Should I allow for two collisions max?
1019 	 */
1020 	if ((slabs = pages / keg->uk_ppera) > keg->uk_hash.uh_hashsize) {
1021 		struct uma_hash newhash;
1022 		struct uma_hash oldhash;
1023 		int ret;
1024 
1025 		/*
1026 		 * This is so involved because allocating and freeing
1027 		 * while the keg lock is held will lead to deadlock.
1028 		 * I have to do everything in stages and check for
1029 		 * races.
1030 		 */
1031 		KEG_UNLOCK(keg, 0);
1032 		ret = hash_alloc(&newhash, 1 << fls(slabs));
1033 		KEG_LOCK(keg, 0);
1034 		if (ret) {
1035 			if (hash_expand(&keg->uk_hash, &newhash)) {
1036 				oldhash = keg->uk_hash;
1037 				keg->uk_hash = newhash;
1038 			} else
1039 				oldhash = newhash;
1040 
1041 			KEG_UNLOCK(keg, 0);
1042 			hash_free(&oldhash);
1043 			goto update_wss;
1044 		}
1045 	}
1046 	KEG_UNLOCK(keg, 0);
1047 
1048 update_wss:
1049 	for (int i = 0; i < vm_ndomains; i++)
1050 		zone_domain_update_wss(ZDOM_GET(zone, i));
1051 }
1052 
1053 /*
1054  * Allocate and zero fill the next sized hash table from the appropriate
1055  * backing store.
1056  *
1057  * Arguments:
1058  *	hash  A new hash structure with the old hash size in uh_hashsize
1059  *
1060  * Returns:
1061  *	1 on success and 0 on failure.
1062  */
1063 static int
1064 hash_alloc(struct uma_hash *hash, u_int size)
1065 {
1066 	size_t alloc;
1067 
1068 	KASSERT(powerof2(size), ("hash size must be power of 2"));
1069 	if (size > UMA_HASH_SIZE_INIT)  {
1070 		hash->uh_hashsize = size;
1071 		alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
1072 		hash->uh_slab_hash = malloc(alloc, M_UMAHASH, M_NOWAIT);
1073 	} else {
1074 		alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
1075 		hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
1076 		    UMA_ANYDOMAIN, M_WAITOK);
1077 		hash->uh_hashsize = UMA_HASH_SIZE_INIT;
1078 	}
1079 	if (hash->uh_slab_hash) {
1080 		bzero(hash->uh_slab_hash, alloc);
1081 		hash->uh_hashmask = hash->uh_hashsize - 1;
1082 		return (1);
1083 	}
1084 
1085 	return (0);
1086 }
1087 
1088 /*
1089  * Expands the hash table for HASH zones.  This is done from zone_timeout
1090  * to reduce collisions.  This must not be done in the regular allocation
1091  * path, otherwise, we can recurse on the vm while allocating pages.
1092  *
1093  * Arguments:
1094  *	oldhash  The hash you want to expand
1095  *	newhash  The hash structure for the new table
1096  *
1097  * Returns:
1098  *	Nothing
1099  *
1100  * Discussion:
1101  */
1102 static int
1103 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
1104 {
1105 	uma_hash_slab_t slab;
1106 	u_int hval;
1107 	u_int idx;
1108 
1109 	if (!newhash->uh_slab_hash)
1110 		return (0);
1111 
1112 	if (oldhash->uh_hashsize >= newhash->uh_hashsize)
1113 		return (0);
1114 
1115 	/*
1116 	 * I need to investigate hash algorithms for resizing without a
1117 	 * full rehash.
1118 	 */
1119 
1120 	for (idx = 0; idx < oldhash->uh_hashsize; idx++)
1121 		while (!LIST_EMPTY(&oldhash->uh_slab_hash[idx])) {
1122 			slab = LIST_FIRST(&oldhash->uh_slab_hash[idx]);
1123 			LIST_REMOVE(slab, uhs_hlink);
1124 			hval = UMA_HASH(newhash, slab->uhs_data);
1125 			LIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
1126 			    slab, uhs_hlink);
1127 		}
1128 
1129 	return (1);
1130 }
1131 
1132 /*
1133  * Free the hash bucket to the appropriate backing store.
1134  *
1135  * Arguments:
1136  *	slab_hash  The hash bucket we're freeing
1137  *	hashsize   The number of entries in that hash bucket
1138  *
1139  * Returns:
1140  *	Nothing
1141  */
1142 static void
1143 hash_free(struct uma_hash *hash)
1144 {
1145 	if (hash->uh_slab_hash == NULL)
1146 		return;
1147 	if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
1148 		zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE);
1149 	else
1150 		free(hash->uh_slab_hash, M_UMAHASH);
1151 }
1152 
1153 /*
1154  * Frees all outstanding items in a bucket
1155  *
1156  * Arguments:
1157  *	zone   The zone to free to, must be unlocked.
1158  *	bucket The free/alloc bucket with items.
1159  *
1160  * Returns:
1161  *	Nothing
1162  */
1163 static void
1164 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
1165 {
1166 	int i;
1167 
1168 	if (bucket->ub_cnt == 0)
1169 		return;
1170 
1171 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
1172 	    bucket->ub_seq != SMR_SEQ_INVALID) {
1173 		smr_wait(zone->uz_smr, bucket->ub_seq);
1174 		bucket->ub_seq = SMR_SEQ_INVALID;
1175 		for (i = 0; i < bucket->ub_cnt; i++)
1176 			item_dtor(zone, bucket->ub_bucket[i],
1177 			    zone->uz_size, NULL, SKIP_NONE);
1178 	}
1179 	if (zone->uz_fini)
1180 		for (i = 0; i < bucket->ub_cnt; i++)
1181 			zone->uz_fini(bucket->ub_bucket[i], zone->uz_size);
1182 	zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt);
1183 	if (zone->uz_max_items > 0)
1184 		zone_free_limit(zone, bucket->ub_cnt);
1185 #ifdef INVARIANTS
1186 	bzero(bucket->ub_bucket, sizeof(void *) * bucket->ub_cnt);
1187 #endif
1188 	bucket->ub_cnt = 0;
1189 }
1190 
1191 /*
1192  * Drains the per cpu caches for a zone.
1193  *
1194  * NOTE: This may only be called while the zone is being torn down, and not
1195  * during normal operation.  This is necessary in order that we do not have
1196  * to migrate CPUs to drain the per-CPU caches.
1197  *
1198  * Arguments:
1199  *	zone     The zone to drain, must be unlocked.
1200  *
1201  * Returns:
1202  *	Nothing
1203  */
1204 static void
1205 cache_drain(uma_zone_t zone)
1206 {
1207 	uma_cache_t cache;
1208 	uma_bucket_t bucket;
1209 	smr_seq_t seq;
1210 	int cpu;
1211 
1212 	/*
1213 	 * XXX: It is safe to not lock the per-CPU caches, because we're
1214 	 * tearing down the zone anyway.  I.e., there will be no further use
1215 	 * of the caches at this point.
1216 	 *
1217 	 * XXX: It would good to be able to assert that the zone is being
1218 	 * torn down to prevent improper use of cache_drain().
1219 	 */
1220 	seq = SMR_SEQ_INVALID;
1221 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
1222 		seq = smr_advance(zone->uz_smr);
1223 	CPU_FOREACH(cpu) {
1224 		cache = &zone->uz_cpu[cpu];
1225 		bucket = cache_bucket_unload_alloc(cache);
1226 		if (bucket != NULL)
1227 			bucket_free(zone, bucket, NULL);
1228 		bucket = cache_bucket_unload_free(cache);
1229 		if (bucket != NULL) {
1230 			bucket->ub_seq = seq;
1231 			bucket_free(zone, bucket, NULL);
1232 		}
1233 		bucket = cache_bucket_unload_cross(cache);
1234 		if (bucket != NULL) {
1235 			bucket->ub_seq = seq;
1236 			bucket_free(zone, bucket, NULL);
1237 		}
1238 	}
1239 	bucket_cache_reclaim(zone, true);
1240 }
1241 
1242 static void
1243 cache_shrink(uma_zone_t zone, void *unused)
1244 {
1245 
1246 	if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1247 		return;
1248 
1249 	zone->uz_bucket_size =
1250 	    (zone->uz_bucket_size_min + zone->uz_bucket_size) / 2;
1251 }
1252 
1253 static void
1254 cache_drain_safe_cpu(uma_zone_t zone, void *unused)
1255 {
1256 	uma_cache_t cache;
1257 	uma_bucket_t b1, b2, b3;
1258 	int domain;
1259 
1260 	if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1261 		return;
1262 
1263 	b1 = b2 = b3 = NULL;
1264 	critical_enter();
1265 	cache = &zone->uz_cpu[curcpu];
1266 	domain = PCPU_GET(domain);
1267 	b1 = cache_bucket_unload_alloc(cache);
1268 
1269 	/*
1270 	 * Don't flush SMR zone buckets.  This leaves the zone without a
1271 	 * bucket and forces every free to synchronize().
1272 	 */
1273 	if ((zone->uz_flags & UMA_ZONE_SMR) == 0) {
1274 		b2 = cache_bucket_unload_free(cache);
1275 		b3 = cache_bucket_unload_cross(cache);
1276 	}
1277 	critical_exit();
1278 
1279 	if (b1 != NULL)
1280 		zone_free_bucket(zone, b1, NULL, domain, false);
1281 	if (b2 != NULL)
1282 		zone_free_bucket(zone, b2, NULL, domain, false);
1283 	if (b3 != NULL) {
1284 		/* Adjust the domain so it goes to zone_free_cross. */
1285 		domain = (domain + 1) % vm_ndomains;
1286 		zone_free_bucket(zone, b3, NULL, domain, false);
1287 	}
1288 }
1289 
1290 /*
1291  * Safely drain per-CPU caches of a zone(s) to alloc bucket.
1292  * This is an expensive call because it needs to bind to all CPUs
1293  * one by one and enter a critical section on each of them in order
1294  * to safely access their cache buckets.
1295  * Zone lock must not be held on call this function.
1296  */
1297 static void
1298 pcpu_cache_drain_safe(uma_zone_t zone)
1299 {
1300 	int cpu;
1301 
1302 	/*
1303 	 * Polite bucket sizes shrinking was not enough, shrink aggressively.
1304 	 */
1305 	if (zone)
1306 		cache_shrink(zone, NULL);
1307 	else
1308 		zone_foreach(cache_shrink, NULL);
1309 
1310 	CPU_FOREACH(cpu) {
1311 		thread_lock(curthread);
1312 		sched_bind(curthread, cpu);
1313 		thread_unlock(curthread);
1314 
1315 		if (zone)
1316 			cache_drain_safe_cpu(zone, NULL);
1317 		else
1318 			zone_foreach(cache_drain_safe_cpu, NULL);
1319 	}
1320 	thread_lock(curthread);
1321 	sched_unbind(curthread);
1322 	thread_unlock(curthread);
1323 }
1324 
1325 /*
1326  * Reclaim cached buckets from a zone.  All buckets are reclaimed if the caller
1327  * requested a drain, otherwise the per-domain caches are trimmed to either
1328  * estimated working set size.
1329  */
1330 static void
1331 bucket_cache_reclaim(uma_zone_t zone, bool drain)
1332 {
1333 	uma_zone_domain_t zdom;
1334 	uma_bucket_t bucket;
1335 	long target;
1336 	int i;
1337 
1338 	/*
1339 	 * Shrink the zone bucket size to ensure that the per-CPU caches
1340 	 * don't grow too large.
1341 	 */
1342 	if (zone->uz_bucket_size > zone->uz_bucket_size_min)
1343 		zone->uz_bucket_size--;
1344 
1345 	for (i = 0; i < vm_ndomains; i++) {
1346 		/*
1347 		 * The cross bucket is partially filled and not part of
1348 		 * the item count.  Reclaim it individually here.
1349 		 */
1350 		zdom = ZDOM_GET(zone, i);
1351 		if ((zone->uz_flags & UMA_ZONE_SMR) == 0 || drain) {
1352 			ZONE_CROSS_LOCK(zone);
1353 			bucket = zdom->uzd_cross;
1354 			zdom->uzd_cross = NULL;
1355 			ZONE_CROSS_UNLOCK(zone);
1356 			if (bucket != NULL)
1357 				bucket_free(zone, bucket, NULL);
1358 		}
1359 
1360 		/*
1361 		 * If we were asked to drain the zone, we are done only once
1362 		 * this bucket cache is empty.  Otherwise, we reclaim items in
1363 		 * excess of the zone's estimated working set size.  If the
1364 		 * difference nitems - imin is larger than the WSS estimate,
1365 		 * then the estimate will grow at the end of this interval and
1366 		 * we ignore the historical average.
1367 		 */
1368 		ZDOM_LOCK(zdom);
1369 		target = drain ? 0 : lmax(zdom->uzd_wss, zdom->uzd_nitems -
1370 		    zdom->uzd_imin);
1371 		while (zdom->uzd_nitems > target) {
1372 			bucket = zone_fetch_bucket(zone, zdom, true);
1373 			if (bucket == NULL)
1374 				break;
1375 			bucket_free(zone, bucket, NULL);
1376 			ZDOM_LOCK(zdom);
1377 		}
1378 		ZDOM_UNLOCK(zdom);
1379 	}
1380 }
1381 
1382 static void
1383 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start)
1384 {
1385 	uint8_t *mem;
1386 	int i;
1387 	uint8_t flags;
1388 
1389 	CTR4(KTR_UMA, "keg_free_slab keg %s(%p) slab %p, returning %d bytes",
1390 	    keg->uk_name, keg, slab, PAGE_SIZE * keg->uk_ppera);
1391 
1392 	mem = slab_data(slab, keg);
1393 	flags = slab->us_flags;
1394 	i = start;
1395 	if (keg->uk_fini != NULL) {
1396 		for (i--; i > -1; i--)
1397 #ifdef INVARIANTS
1398 		/*
1399 		 * trash_fini implies that dtor was trash_dtor. trash_fini
1400 		 * would check that memory hasn't been modified since free,
1401 		 * which executed trash_dtor.
1402 		 * That's why we need to run uma_dbg_kskip() check here,
1403 		 * albeit we don't make skip check for other init/fini
1404 		 * invocations.
1405 		 */
1406 		if (!uma_dbg_kskip(keg, slab_item(slab, keg, i)) ||
1407 		    keg->uk_fini != trash_fini)
1408 #endif
1409 			keg->uk_fini(slab_item(slab, keg, i), keg->uk_size);
1410 	}
1411 	if (keg->uk_flags & UMA_ZFLAG_OFFPAGE)
1412 		zone_free_item(slabzone(keg->uk_ipers), slab_tohashslab(slab),
1413 		    NULL, SKIP_NONE);
1414 	keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags);
1415 	uma_total_dec(PAGE_SIZE * keg->uk_ppera);
1416 }
1417 
1418 static void
1419 keg_drain_domain(uma_keg_t keg, int domain)
1420 {
1421 	struct slabhead freeslabs;
1422 	uma_domain_t dom;
1423 	uma_slab_t slab, tmp;
1424 	uint32_t i, stofree, stokeep, partial;
1425 
1426 	dom = &keg->uk_domain[domain];
1427 	LIST_INIT(&freeslabs);
1428 
1429 	CTR4(KTR_UMA, "keg_drain %s(%p) domain %d free items: %u",
1430 	    keg->uk_name, keg, domain, dom->ud_free_items);
1431 
1432 	KEG_LOCK(keg, domain);
1433 
1434 	/*
1435 	 * Are the free items in partially allocated slabs sufficient to meet
1436 	 * the reserve? If not, compute the number of fully free slabs that must
1437 	 * be kept.
1438 	 */
1439 	partial = dom->ud_free_items - dom->ud_free_slabs * keg->uk_ipers;
1440 	if (partial < keg->uk_reserve) {
1441 		stokeep = min(dom->ud_free_slabs,
1442 		    howmany(keg->uk_reserve - partial, keg->uk_ipers));
1443 	} else {
1444 		stokeep = 0;
1445 	}
1446 	stofree = dom->ud_free_slabs - stokeep;
1447 
1448 	/*
1449 	 * Partition the free slabs into two sets: those that must be kept in
1450 	 * order to maintain the reserve, and those that may be released back to
1451 	 * the system.  Since one set may be much larger than the other,
1452 	 * populate the smaller of the two sets and swap them if necessary.
1453 	 */
1454 	for (i = min(stofree, stokeep); i > 0; i--) {
1455 		slab = LIST_FIRST(&dom->ud_free_slab);
1456 		LIST_REMOVE(slab, us_link);
1457 		LIST_INSERT_HEAD(&freeslabs, slab, us_link);
1458 	}
1459 	if (stofree > stokeep)
1460 		LIST_SWAP(&freeslabs, &dom->ud_free_slab, uma_slab, us_link);
1461 
1462 	if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0) {
1463 		LIST_FOREACH(slab, &freeslabs, us_link)
1464 			UMA_HASH_REMOVE(&keg->uk_hash, slab);
1465 	}
1466 	dom->ud_free_items -= stofree * keg->uk_ipers;
1467 	dom->ud_free_slabs -= stofree;
1468 	dom->ud_pages -= stofree * keg->uk_ppera;
1469 	KEG_UNLOCK(keg, domain);
1470 
1471 	LIST_FOREACH_SAFE(slab, &freeslabs, us_link, tmp)
1472 		keg_free_slab(keg, slab, keg->uk_ipers);
1473 }
1474 
1475 /*
1476  * Frees pages from a keg back to the system.  This is done on demand from
1477  * the pageout daemon.
1478  *
1479  * Returns nothing.
1480  */
1481 static void
1482 keg_drain(uma_keg_t keg)
1483 {
1484 	int i;
1485 
1486 	if ((keg->uk_flags & UMA_ZONE_NOFREE) != 0)
1487 		return;
1488 	for (i = 0; i < vm_ndomains; i++)
1489 		keg_drain_domain(keg, i);
1490 }
1491 
1492 static void
1493 zone_reclaim(uma_zone_t zone, int waitok, bool drain)
1494 {
1495 
1496 	/*
1497 	 * Set draining to interlock with zone_dtor() so we can release our
1498 	 * locks as we go.  Only dtor() should do a WAITOK call since it
1499 	 * is the only call that knows the structure will still be available
1500 	 * when it wakes up.
1501 	 */
1502 	ZONE_LOCK(zone);
1503 	while (zone->uz_flags & UMA_ZFLAG_RECLAIMING) {
1504 		if (waitok == M_NOWAIT)
1505 			goto out;
1506 		msleep(zone, &ZDOM_GET(zone, 0)->uzd_lock, PVM, "zonedrain",
1507 		    1);
1508 	}
1509 	zone->uz_flags |= UMA_ZFLAG_RECLAIMING;
1510 	ZONE_UNLOCK(zone);
1511 	bucket_cache_reclaim(zone, drain);
1512 
1513 	/*
1514 	 * The DRAINING flag protects us from being freed while
1515 	 * we're running.  Normally the uma_rwlock would protect us but we
1516 	 * must be able to release and acquire the right lock for each keg.
1517 	 */
1518 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0)
1519 		keg_drain(zone->uz_keg);
1520 	ZONE_LOCK(zone);
1521 	zone->uz_flags &= ~UMA_ZFLAG_RECLAIMING;
1522 	wakeup(zone);
1523 out:
1524 	ZONE_UNLOCK(zone);
1525 }
1526 
1527 static void
1528 zone_drain(uma_zone_t zone, void *unused)
1529 {
1530 
1531 	zone_reclaim(zone, M_NOWAIT, true);
1532 }
1533 
1534 static void
1535 zone_trim(uma_zone_t zone, void *unused)
1536 {
1537 
1538 	zone_reclaim(zone, M_NOWAIT, false);
1539 }
1540 
1541 /*
1542  * Allocate a new slab for a keg and inserts it into the partial slab list.
1543  * The keg should be unlocked on entry.  If the allocation succeeds it will
1544  * be locked on return.
1545  *
1546  * Arguments:
1547  *	flags   Wait flags for the item initialization routine
1548  *	aflags  Wait flags for the slab allocation
1549  *
1550  * Returns:
1551  *	The slab that was allocated or NULL if there is no memory and the
1552  *	caller specified M_NOWAIT.
1553  */
1554 static uma_slab_t
1555 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int domain, int flags,
1556     int aflags)
1557 {
1558 	uma_domain_t dom;
1559 	uma_alloc allocf;
1560 	uma_slab_t slab;
1561 	unsigned long size;
1562 	uint8_t *mem;
1563 	uint8_t sflags;
1564 	int i;
1565 
1566 	KASSERT(domain >= 0 && domain < vm_ndomains,
1567 	    ("keg_alloc_slab: domain %d out of range", domain));
1568 
1569 	allocf = keg->uk_allocf;
1570 	slab = NULL;
1571 	mem = NULL;
1572 	if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) {
1573 		uma_hash_slab_t hslab;
1574 		hslab = zone_alloc_item(slabzone(keg->uk_ipers), NULL,
1575 		    domain, aflags);
1576 		if (hslab == NULL)
1577 			goto fail;
1578 		slab = &hslab->uhs_slab;
1579 	}
1580 
1581 	/*
1582 	 * This reproduces the old vm_zone behavior of zero filling pages the
1583 	 * first time they are added to a zone.
1584 	 *
1585 	 * Malloced items are zeroed in uma_zalloc.
1586 	 */
1587 
1588 	if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
1589 		aflags |= M_ZERO;
1590 	else
1591 		aflags &= ~M_ZERO;
1592 
1593 	if (keg->uk_flags & UMA_ZONE_NODUMP)
1594 		aflags |= M_NODUMP;
1595 
1596 	/* zone is passed for legacy reasons. */
1597 	size = keg->uk_ppera * PAGE_SIZE;
1598 	mem = allocf(zone, size, domain, &sflags, aflags);
1599 	if (mem == NULL) {
1600 		if (keg->uk_flags & UMA_ZFLAG_OFFPAGE)
1601 			zone_free_item(slabzone(keg->uk_ipers),
1602 			    slab_tohashslab(slab), NULL, SKIP_NONE);
1603 		goto fail;
1604 	}
1605 	uma_total_inc(size);
1606 
1607 	/* For HASH zones all pages go to the same uma_domain. */
1608 	if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
1609 		domain = 0;
1610 
1611 	/* Point the slab into the allocated memory */
1612 	if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE))
1613 		slab = (uma_slab_t )(mem + keg->uk_pgoff);
1614 	else
1615 		slab_tohashslab(slab)->uhs_data = mem;
1616 
1617 	if (keg->uk_flags & UMA_ZFLAG_VTOSLAB)
1618 		for (i = 0; i < keg->uk_ppera; i++)
1619 			vsetzoneslab((vm_offset_t)mem + (i * PAGE_SIZE),
1620 			    zone, slab);
1621 
1622 	slab->us_freecount = keg->uk_ipers;
1623 	slab->us_flags = sflags;
1624 	slab->us_domain = domain;
1625 
1626 	BIT_FILL(keg->uk_ipers, &slab->us_free);
1627 #ifdef INVARIANTS
1628 	BIT_ZERO(keg->uk_ipers, slab_dbg_bits(slab, keg));
1629 #endif
1630 
1631 	if (keg->uk_init != NULL) {
1632 		for (i = 0; i < keg->uk_ipers; i++)
1633 			if (keg->uk_init(slab_item(slab, keg, i),
1634 			    keg->uk_size, flags) != 0)
1635 				break;
1636 		if (i != keg->uk_ipers) {
1637 			keg_free_slab(keg, slab, i);
1638 			goto fail;
1639 		}
1640 	}
1641 	KEG_LOCK(keg, domain);
1642 
1643 	CTR3(KTR_UMA, "keg_alloc_slab: allocated slab %p for %s(%p)",
1644 	    slab, keg->uk_name, keg);
1645 
1646 	if (keg->uk_flags & UMA_ZFLAG_HASH)
1647 		UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
1648 
1649 	/*
1650 	 * If we got a slab here it's safe to mark it partially used
1651 	 * and return.  We assume that the caller is going to remove
1652 	 * at least one item.
1653 	 */
1654 	dom = &keg->uk_domain[domain];
1655 	LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
1656 	dom->ud_pages += keg->uk_ppera;
1657 	dom->ud_free_items += keg->uk_ipers;
1658 
1659 	return (slab);
1660 
1661 fail:
1662 	return (NULL);
1663 }
1664 
1665 /*
1666  * This function is intended to be used early on in place of page_alloc() so
1667  * that we may use the boot time page cache to satisfy allocations before
1668  * the VM is ready.
1669  */
1670 static void *
1671 startup_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1672     int wait)
1673 {
1674 	vm_paddr_t pa;
1675 	vm_page_t m;
1676 	void *mem;
1677 	int pages;
1678 	int i;
1679 
1680 	pages = howmany(bytes, PAGE_SIZE);
1681 	KASSERT(pages > 0, ("%s can't reserve 0 pages", __func__));
1682 
1683 	*pflag = UMA_SLAB_BOOT;
1684 	m = vm_page_alloc_contig_domain(NULL, 0, domain,
1685 	    malloc2vm_flags(wait) | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED, pages,
1686 	    (vm_paddr_t)0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT);
1687 	if (m == NULL)
1688 		return (NULL);
1689 
1690 	pa = VM_PAGE_TO_PHYS(m);
1691 	for (i = 0; i < pages; i++, pa += PAGE_SIZE) {
1692 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
1693     defined(__riscv) || defined(__powerpc64__)
1694 		if ((wait & M_NODUMP) == 0)
1695 			dump_add_page(pa);
1696 #endif
1697 	}
1698 	/* Allocate KVA and indirectly advance bootmem. */
1699 	mem = (void *)pmap_map(&bootmem, m->phys_addr,
1700 	    m->phys_addr + (pages * PAGE_SIZE), VM_PROT_READ | VM_PROT_WRITE);
1701         if ((wait & M_ZERO) != 0)
1702                 bzero(mem, pages * PAGE_SIZE);
1703 
1704         return (mem);
1705 }
1706 
1707 static void
1708 startup_free(void *mem, vm_size_t bytes)
1709 {
1710 	vm_offset_t va;
1711 	vm_page_t m;
1712 
1713 	va = (vm_offset_t)mem;
1714 	m = PHYS_TO_VM_PAGE(pmap_kextract(va));
1715 	pmap_remove(kernel_pmap, va, va + bytes);
1716 	for (; bytes != 0; bytes -= PAGE_SIZE, m++) {
1717 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
1718     defined(__riscv) || defined(__powerpc64__)
1719 		dump_drop_page(VM_PAGE_TO_PHYS(m));
1720 #endif
1721 		vm_page_unwire_noq(m);
1722 		vm_page_free(m);
1723 	}
1724 }
1725 
1726 /*
1727  * Allocates a number of pages from the system
1728  *
1729  * Arguments:
1730  *	bytes  The number of bytes requested
1731  *	wait  Shall we wait?
1732  *
1733  * Returns:
1734  *	A pointer to the alloced memory or possibly
1735  *	NULL if M_NOWAIT is set.
1736  */
1737 static void *
1738 page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1739     int wait)
1740 {
1741 	void *p;	/* Returned page */
1742 
1743 	*pflag = UMA_SLAB_KERNEL;
1744 	p = (void *)kmem_malloc_domainset(DOMAINSET_FIXED(domain), bytes, wait);
1745 
1746 	return (p);
1747 }
1748 
1749 static void *
1750 pcpu_page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1751     int wait)
1752 {
1753 	struct pglist alloctail;
1754 	vm_offset_t addr, zkva;
1755 	int cpu, flags;
1756 	vm_page_t p, p_next;
1757 #ifdef NUMA
1758 	struct pcpu *pc;
1759 #endif
1760 
1761 	MPASS(bytes == (mp_maxid + 1) * PAGE_SIZE);
1762 
1763 	TAILQ_INIT(&alloctail);
1764 	flags = VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | VM_ALLOC_NOOBJ |
1765 	    malloc2vm_flags(wait);
1766 	*pflag = UMA_SLAB_KERNEL;
1767 	for (cpu = 0; cpu <= mp_maxid; cpu++) {
1768 		if (CPU_ABSENT(cpu)) {
1769 			p = vm_page_alloc(NULL, 0, flags);
1770 		} else {
1771 #ifndef NUMA
1772 			p = vm_page_alloc(NULL, 0, flags);
1773 #else
1774 			pc = pcpu_find(cpu);
1775 			if (__predict_false(VM_DOMAIN_EMPTY(pc->pc_domain)))
1776 				p = NULL;
1777 			else
1778 				p = vm_page_alloc_domain(NULL, 0,
1779 				    pc->pc_domain, flags);
1780 			if (__predict_false(p == NULL))
1781 				p = vm_page_alloc(NULL, 0, flags);
1782 #endif
1783 		}
1784 		if (__predict_false(p == NULL))
1785 			goto fail;
1786 		TAILQ_INSERT_TAIL(&alloctail, p, listq);
1787 	}
1788 	if ((addr = kva_alloc(bytes)) == 0)
1789 		goto fail;
1790 	zkva = addr;
1791 	TAILQ_FOREACH(p, &alloctail, listq) {
1792 		pmap_qenter(zkva, &p, 1);
1793 		zkva += PAGE_SIZE;
1794 	}
1795 	return ((void*)addr);
1796 fail:
1797 	TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1798 		vm_page_unwire_noq(p);
1799 		vm_page_free(p);
1800 	}
1801 	return (NULL);
1802 }
1803 
1804 /*
1805  * Allocates a number of pages from within an object
1806  *
1807  * Arguments:
1808  *	bytes  The number of bytes requested
1809  *	wait   Shall we wait?
1810  *
1811  * Returns:
1812  *	A pointer to the alloced memory or possibly
1813  *	NULL if M_NOWAIT is set.
1814  */
1815 static void *
1816 noobj_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags,
1817     int wait)
1818 {
1819 	TAILQ_HEAD(, vm_page) alloctail;
1820 	u_long npages;
1821 	vm_offset_t retkva, zkva;
1822 	vm_page_t p, p_next;
1823 	uma_keg_t keg;
1824 
1825 	TAILQ_INIT(&alloctail);
1826 	keg = zone->uz_keg;
1827 
1828 	npages = howmany(bytes, PAGE_SIZE);
1829 	while (npages > 0) {
1830 		p = vm_page_alloc_domain(NULL, 0, domain, VM_ALLOC_INTERRUPT |
1831 		    VM_ALLOC_WIRED | VM_ALLOC_NOOBJ |
1832 		    ((wait & M_WAITOK) != 0 ? VM_ALLOC_WAITOK :
1833 		    VM_ALLOC_NOWAIT));
1834 		if (p != NULL) {
1835 			/*
1836 			 * Since the page does not belong to an object, its
1837 			 * listq is unused.
1838 			 */
1839 			TAILQ_INSERT_TAIL(&alloctail, p, listq);
1840 			npages--;
1841 			continue;
1842 		}
1843 		/*
1844 		 * Page allocation failed, free intermediate pages and
1845 		 * exit.
1846 		 */
1847 		TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1848 			vm_page_unwire_noq(p);
1849 			vm_page_free(p);
1850 		}
1851 		return (NULL);
1852 	}
1853 	*flags = UMA_SLAB_PRIV;
1854 	zkva = keg->uk_kva +
1855 	    atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
1856 	retkva = zkva;
1857 	TAILQ_FOREACH(p, &alloctail, listq) {
1858 		pmap_qenter(zkva, &p, 1);
1859 		zkva += PAGE_SIZE;
1860 	}
1861 
1862 	return ((void *)retkva);
1863 }
1864 
1865 /*
1866  * Allocate physically contiguous pages.
1867  */
1868 static void *
1869 contig_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1870     int wait)
1871 {
1872 
1873 	*pflag = UMA_SLAB_KERNEL;
1874 	return ((void *)kmem_alloc_contig_domainset(DOMAINSET_FIXED(domain),
1875 	    bytes, wait, 0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT));
1876 }
1877 
1878 /*
1879  * Frees a number of pages to the system
1880  *
1881  * Arguments:
1882  *	mem   A pointer to the memory to be freed
1883  *	size  The size of the memory being freed
1884  *	flags The original p->us_flags field
1885  *
1886  * Returns:
1887  *	Nothing
1888  */
1889 static void
1890 page_free(void *mem, vm_size_t size, uint8_t flags)
1891 {
1892 
1893 	if ((flags & UMA_SLAB_BOOT) != 0) {
1894 		startup_free(mem, size);
1895 		return;
1896 	}
1897 
1898 	KASSERT((flags & UMA_SLAB_KERNEL) != 0,
1899 	    ("UMA: page_free used with invalid flags %x", flags));
1900 
1901 	kmem_free((vm_offset_t)mem, size);
1902 }
1903 
1904 /*
1905  * Frees pcpu zone allocations
1906  *
1907  * Arguments:
1908  *	mem   A pointer to the memory to be freed
1909  *	size  The size of the memory being freed
1910  *	flags The original p->us_flags field
1911  *
1912  * Returns:
1913  *	Nothing
1914  */
1915 static void
1916 pcpu_page_free(void *mem, vm_size_t size, uint8_t flags)
1917 {
1918 	vm_offset_t sva, curva;
1919 	vm_paddr_t paddr;
1920 	vm_page_t m;
1921 
1922 	MPASS(size == (mp_maxid+1)*PAGE_SIZE);
1923 
1924 	if ((flags & UMA_SLAB_BOOT) != 0) {
1925 		startup_free(mem, size);
1926 		return;
1927 	}
1928 
1929 	sva = (vm_offset_t)mem;
1930 	for (curva = sva; curva < sva + size; curva += PAGE_SIZE) {
1931 		paddr = pmap_kextract(curva);
1932 		m = PHYS_TO_VM_PAGE(paddr);
1933 		vm_page_unwire_noq(m);
1934 		vm_page_free(m);
1935 	}
1936 	pmap_qremove(sva, size >> PAGE_SHIFT);
1937 	kva_free(sva, size);
1938 }
1939 
1940 /*
1941  * Zero fill initializer
1942  *
1943  * Arguments/Returns follow uma_init specifications
1944  */
1945 static int
1946 zero_init(void *mem, int size, int flags)
1947 {
1948 	bzero(mem, size);
1949 	return (0);
1950 }
1951 
1952 #ifdef INVARIANTS
1953 static struct noslabbits *
1954 slab_dbg_bits(uma_slab_t slab, uma_keg_t keg)
1955 {
1956 
1957 	return ((void *)((char *)&slab->us_free + BITSET_SIZE(keg->uk_ipers)));
1958 }
1959 #endif
1960 
1961 /*
1962  * Actual size of embedded struct slab (!OFFPAGE).
1963  */
1964 static size_t
1965 slab_sizeof(int nitems)
1966 {
1967 	size_t s;
1968 
1969 	s = sizeof(struct uma_slab) + BITSET_SIZE(nitems) * SLAB_BITSETS;
1970 	return (roundup(s, UMA_ALIGN_PTR + 1));
1971 }
1972 
1973 #define	UMA_FIXPT_SHIFT	31
1974 #define	UMA_FRAC_FIXPT(n, d)						\
1975 	((uint32_t)(((uint64_t)(n) << UMA_FIXPT_SHIFT) / (d)))
1976 #define	UMA_FIXPT_PCT(f)						\
1977 	((u_int)(((uint64_t)100 * (f)) >> UMA_FIXPT_SHIFT))
1978 #define	UMA_PCT_FIXPT(pct)	UMA_FRAC_FIXPT((pct), 100)
1979 #define	UMA_MIN_EFF	UMA_PCT_FIXPT(100 - UMA_MAX_WASTE)
1980 
1981 /*
1982  * Compute the number of items that will fit in a slab.  If hdr is true, the
1983  * item count may be limited to provide space in the slab for an inline slab
1984  * header.  Otherwise, all slab space will be provided for item storage.
1985  */
1986 static u_int
1987 slab_ipers_hdr(u_int size, u_int rsize, u_int slabsize, bool hdr)
1988 {
1989 	u_int ipers;
1990 	u_int padpi;
1991 
1992 	/* The padding between items is not needed after the last item. */
1993 	padpi = rsize - size;
1994 
1995 	if (hdr) {
1996 		/*
1997 		 * Start with the maximum item count and remove items until
1998 		 * the slab header first alongside the allocatable memory.
1999 		 */
2000 		for (ipers = MIN(SLAB_MAX_SETSIZE,
2001 		    (slabsize + padpi - slab_sizeof(1)) / rsize);
2002 		    ipers > 0 &&
2003 		    ipers * rsize - padpi + slab_sizeof(ipers) > slabsize;
2004 		    ipers--)
2005 			continue;
2006 	} else {
2007 		ipers = MIN((slabsize + padpi) / rsize, SLAB_MAX_SETSIZE);
2008 	}
2009 
2010 	return (ipers);
2011 }
2012 
2013 struct keg_layout_result {
2014 	u_int format;
2015 	u_int slabsize;
2016 	u_int ipers;
2017 	u_int eff;
2018 };
2019 
2020 static void
2021 keg_layout_one(uma_keg_t keg, u_int rsize, u_int slabsize, u_int fmt,
2022     struct keg_layout_result *kl)
2023 {
2024 	u_int total;
2025 
2026 	kl->format = fmt;
2027 	kl->slabsize = slabsize;
2028 
2029 	/* Handle INTERNAL as inline with an extra page. */
2030 	if ((fmt & UMA_ZFLAG_INTERNAL) != 0) {
2031 		kl->format &= ~UMA_ZFLAG_INTERNAL;
2032 		kl->slabsize += PAGE_SIZE;
2033 	}
2034 
2035 	kl->ipers = slab_ipers_hdr(keg->uk_size, rsize, kl->slabsize,
2036 	    (fmt & UMA_ZFLAG_OFFPAGE) == 0);
2037 
2038 	/* Account for memory used by an offpage slab header. */
2039 	total = kl->slabsize;
2040 	if ((fmt & UMA_ZFLAG_OFFPAGE) != 0)
2041 		total += slabzone(kl->ipers)->uz_keg->uk_rsize;
2042 
2043 	kl->eff = UMA_FRAC_FIXPT(kl->ipers * rsize, total);
2044 }
2045 
2046 /*
2047  * Determine the format of a uma keg.  This determines where the slab header
2048  * will be placed (inline or offpage) and calculates ipers, rsize, and ppera.
2049  *
2050  * Arguments
2051  *	keg  The zone we should initialize
2052  *
2053  * Returns
2054  *	Nothing
2055  */
2056 static void
2057 keg_layout(uma_keg_t keg)
2058 {
2059 	struct keg_layout_result kl = {}, kl_tmp;
2060 	u_int fmts[2];
2061 	u_int alignsize;
2062 	u_int nfmt;
2063 	u_int pages;
2064 	u_int rsize;
2065 	u_int slabsize;
2066 	u_int i, j;
2067 
2068 	KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
2069 	    (keg->uk_size <= UMA_PCPU_ALLOC_SIZE &&
2070 	     (keg->uk_flags & UMA_ZONE_CACHESPREAD) == 0),
2071 	    ("%s: cannot configure for PCPU: keg=%s, size=%u, flags=0x%b",
2072 	     __func__, keg->uk_name, keg->uk_size, keg->uk_flags,
2073 	     PRINT_UMA_ZFLAGS));
2074 	KASSERT((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) == 0 ||
2075 	    (keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0,
2076 	    ("%s: incompatible flags 0x%b", __func__, keg->uk_flags,
2077 	     PRINT_UMA_ZFLAGS));
2078 
2079 	alignsize = keg->uk_align + 1;
2080 
2081 	/*
2082 	 * Calculate the size of each allocation (rsize) according to
2083 	 * alignment.  If the requested size is smaller than we have
2084 	 * allocation bits for we round it up.
2085 	 */
2086 	rsize = MAX(keg->uk_size, UMA_SMALLEST_UNIT);
2087 	rsize = roundup2(rsize, alignsize);
2088 
2089 	if ((keg->uk_flags & UMA_ZONE_CACHESPREAD) != 0) {
2090 		/*
2091 		 * We want one item to start on every align boundary in a page.
2092 		 * To do this we will span pages.  We will also extend the item
2093 		 * by the size of align if it is an even multiple of align.
2094 		 * Otherwise, it would fall on the same boundary every time.
2095 		 */
2096 		if ((rsize & alignsize) == 0)
2097 			rsize += alignsize;
2098 		slabsize = rsize * (PAGE_SIZE / alignsize);
2099 		slabsize = MIN(slabsize, rsize * SLAB_MAX_SETSIZE);
2100 		slabsize = MIN(slabsize, UMA_CACHESPREAD_MAX_SIZE);
2101 		slabsize = round_page(slabsize);
2102 	} else {
2103 		/*
2104 		 * Start with a slab size of as many pages as it takes to
2105 		 * represent a single item.  We will try to fit as many
2106 		 * additional items into the slab as possible.
2107 		 */
2108 		slabsize = round_page(keg->uk_size);
2109 	}
2110 
2111 	/* Build a list of all of the available formats for this keg. */
2112 	nfmt = 0;
2113 
2114 	/* Evaluate an inline slab layout. */
2115 	if ((keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0)
2116 		fmts[nfmt++] = 0;
2117 
2118 	/* TODO: vm_page-embedded slab. */
2119 
2120 	/*
2121 	 * We can't do OFFPAGE if we're internal or if we've been
2122 	 * asked to not go to the VM for buckets.  If we do this we
2123 	 * may end up going to the VM for slabs which we do not want
2124 	 * to do if we're UMA_ZONE_VM, which clearly forbids it.
2125 	 * In those cases, evaluate a pseudo-format called INTERNAL
2126 	 * which has an inline slab header and one extra page to
2127 	 * guarantee that it fits.
2128 	 *
2129 	 * Otherwise, see if using an OFFPAGE slab will improve our
2130 	 * efficiency.
2131 	 */
2132 	if ((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) != 0)
2133 		fmts[nfmt++] = UMA_ZFLAG_INTERNAL;
2134 	else
2135 		fmts[nfmt++] = UMA_ZFLAG_OFFPAGE;
2136 
2137 	/*
2138 	 * Choose a slab size and format which satisfy the minimum efficiency.
2139 	 * Prefer the smallest slab size that meets the constraints.
2140 	 *
2141 	 * Start with a minimum slab size, to accommodate CACHESPREAD.  Then,
2142 	 * for small items (up to PAGE_SIZE), the iteration increment is one
2143 	 * page; and for large items, the increment is one item.
2144 	 */
2145 	i = (slabsize + rsize - keg->uk_size) / MAX(PAGE_SIZE, rsize);
2146 	KASSERT(i >= 1, ("keg %s(%p) flags=0x%b slabsize=%u, rsize=%u, i=%u",
2147 	    keg->uk_name, keg, keg->uk_flags, PRINT_UMA_ZFLAGS, slabsize,
2148 	    rsize, i));
2149 	for ( ; ; i++) {
2150 		slabsize = (rsize <= PAGE_SIZE) ? ptoa(i) :
2151 		    round_page(rsize * (i - 1) + keg->uk_size);
2152 
2153 		for (j = 0; j < nfmt; j++) {
2154 			/* Only if we have no viable format yet. */
2155 			if ((fmts[j] & UMA_ZFLAG_INTERNAL) != 0 &&
2156 			    kl.ipers > 0)
2157 				continue;
2158 
2159 			keg_layout_one(keg, rsize, slabsize, fmts[j], &kl_tmp);
2160 			if (kl_tmp.eff <= kl.eff)
2161 				continue;
2162 
2163 			kl = kl_tmp;
2164 
2165 			CTR6(KTR_UMA, "keg %s layout: format %#x "
2166 			    "(ipers %u * rsize %u) / slabsize %#x = %u%% eff",
2167 			    keg->uk_name, kl.format, kl.ipers, rsize,
2168 			    kl.slabsize, UMA_FIXPT_PCT(kl.eff));
2169 
2170 			/* Stop when we reach the minimum efficiency. */
2171 			if (kl.eff >= UMA_MIN_EFF)
2172 				break;
2173 		}
2174 
2175 		if (kl.eff >= UMA_MIN_EFF || !multipage_slabs ||
2176 		    slabsize >= SLAB_MAX_SETSIZE * rsize ||
2177 		    (keg->uk_flags & (UMA_ZONE_PCPU | UMA_ZONE_CONTIG)) != 0)
2178 			break;
2179 	}
2180 
2181 	pages = atop(kl.slabsize);
2182 	if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
2183 		pages *= mp_maxid + 1;
2184 
2185 	keg->uk_rsize = rsize;
2186 	keg->uk_ipers = kl.ipers;
2187 	keg->uk_ppera = pages;
2188 	keg->uk_flags |= kl.format;
2189 
2190 	/*
2191 	 * How do we find the slab header if it is offpage or if not all item
2192 	 * start addresses are in the same page?  We could solve the latter
2193 	 * case with vaddr alignment, but we don't.
2194 	 */
2195 	if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0 ||
2196 	    (keg->uk_ipers - 1) * rsize >= PAGE_SIZE) {
2197 		if ((keg->uk_flags & UMA_ZONE_NOTPAGE) != 0)
2198 			keg->uk_flags |= UMA_ZFLAG_HASH;
2199 		else
2200 			keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2201 	}
2202 
2203 	CTR6(KTR_UMA, "%s: keg=%s, flags=%#x, rsize=%u, ipers=%u, ppera=%u",
2204 	    __func__, keg->uk_name, keg->uk_flags, rsize, keg->uk_ipers,
2205 	    pages);
2206 	KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_MAX_SETSIZE,
2207 	    ("%s: keg=%s, flags=0x%b, rsize=%u, ipers=%u, ppera=%u", __func__,
2208 	     keg->uk_name, keg->uk_flags, PRINT_UMA_ZFLAGS, rsize,
2209 	     keg->uk_ipers, pages));
2210 }
2211 
2212 /*
2213  * Keg header ctor.  This initializes all fields, locks, etc.  And inserts
2214  * the keg onto the global keg list.
2215  *
2216  * Arguments/Returns follow uma_ctor specifications
2217  *	udata  Actually uma_kctor_args
2218  */
2219 static int
2220 keg_ctor(void *mem, int size, void *udata, int flags)
2221 {
2222 	struct uma_kctor_args *arg = udata;
2223 	uma_keg_t keg = mem;
2224 	uma_zone_t zone;
2225 	int i;
2226 
2227 	bzero(keg, size);
2228 	keg->uk_size = arg->size;
2229 	keg->uk_init = arg->uminit;
2230 	keg->uk_fini = arg->fini;
2231 	keg->uk_align = arg->align;
2232 	keg->uk_reserve = 0;
2233 	keg->uk_flags = arg->flags;
2234 
2235 	/*
2236 	 * We use a global round-robin policy by default.  Zones with
2237 	 * UMA_ZONE_FIRSTTOUCH set will use first-touch instead, in which
2238 	 * case the iterator is never run.
2239 	 */
2240 	keg->uk_dr.dr_policy = DOMAINSET_RR();
2241 	keg->uk_dr.dr_iter = 0;
2242 
2243 	/*
2244 	 * The primary zone is passed to us at keg-creation time.
2245 	 */
2246 	zone = arg->zone;
2247 	keg->uk_name = zone->uz_name;
2248 
2249 	if (arg->flags & UMA_ZONE_ZINIT)
2250 		keg->uk_init = zero_init;
2251 
2252 	if (arg->flags & UMA_ZONE_MALLOC)
2253 		keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2254 
2255 #ifndef SMP
2256 	keg->uk_flags &= ~UMA_ZONE_PCPU;
2257 #endif
2258 
2259 	keg_layout(keg);
2260 
2261 	/*
2262 	 * Use a first-touch NUMA policy for kegs that pmap_extract() will
2263 	 * work on.  Use round-robin for everything else.
2264 	 *
2265 	 * Zones may override the default by specifying either.
2266 	 */
2267 #ifdef NUMA
2268 	if ((keg->uk_flags &
2269 	    (UMA_ZONE_ROUNDROBIN | UMA_ZFLAG_CACHE | UMA_ZONE_NOTPAGE)) == 0)
2270 		keg->uk_flags |= UMA_ZONE_FIRSTTOUCH;
2271 	else if ((keg->uk_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2272 		keg->uk_flags |= UMA_ZONE_ROUNDROBIN;
2273 #endif
2274 
2275 	/*
2276 	 * If we haven't booted yet we need allocations to go through the
2277 	 * startup cache until the vm is ready.
2278 	 */
2279 #ifdef UMA_MD_SMALL_ALLOC
2280 	if (keg->uk_ppera == 1)
2281 		keg->uk_allocf = uma_small_alloc;
2282 	else
2283 #endif
2284 	if (booted < BOOT_KVA)
2285 		keg->uk_allocf = startup_alloc;
2286 	else if (keg->uk_flags & UMA_ZONE_PCPU)
2287 		keg->uk_allocf = pcpu_page_alloc;
2288 	else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 && keg->uk_ppera > 1)
2289 		keg->uk_allocf = contig_alloc;
2290 	else
2291 		keg->uk_allocf = page_alloc;
2292 #ifdef UMA_MD_SMALL_ALLOC
2293 	if (keg->uk_ppera == 1)
2294 		keg->uk_freef = uma_small_free;
2295 	else
2296 #endif
2297 	if (keg->uk_flags & UMA_ZONE_PCPU)
2298 		keg->uk_freef = pcpu_page_free;
2299 	else
2300 		keg->uk_freef = page_free;
2301 
2302 	/*
2303 	 * Initialize keg's locks.
2304 	 */
2305 	for (i = 0; i < vm_ndomains; i++)
2306 		KEG_LOCK_INIT(keg, i, (arg->flags & UMA_ZONE_MTXCLASS));
2307 
2308 	/*
2309 	 * If we're putting the slab header in the actual page we need to
2310 	 * figure out where in each page it goes.  See slab_sizeof
2311 	 * definition.
2312 	 */
2313 	if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE)) {
2314 		size_t shsize;
2315 
2316 		shsize = slab_sizeof(keg->uk_ipers);
2317 		keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - shsize;
2318 		/*
2319 		 * The only way the following is possible is if with our
2320 		 * UMA_ALIGN_PTR adjustments we are now bigger than
2321 		 * UMA_SLAB_SIZE.  I haven't checked whether this is
2322 		 * mathematically possible for all cases, so we make
2323 		 * sure here anyway.
2324 		 */
2325 		KASSERT(keg->uk_pgoff + shsize <= PAGE_SIZE * keg->uk_ppera,
2326 		    ("zone %s ipers %d rsize %d size %d slab won't fit",
2327 		    zone->uz_name, keg->uk_ipers, keg->uk_rsize, keg->uk_size));
2328 	}
2329 
2330 	if (keg->uk_flags & UMA_ZFLAG_HASH)
2331 		hash_alloc(&keg->uk_hash, 0);
2332 
2333 	CTR3(KTR_UMA, "keg_ctor %p zone %s(%p)", keg, zone->uz_name, zone);
2334 
2335 	LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
2336 
2337 	rw_wlock(&uma_rwlock);
2338 	LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
2339 	rw_wunlock(&uma_rwlock);
2340 	return (0);
2341 }
2342 
2343 static void
2344 zone_kva_available(uma_zone_t zone, void *unused)
2345 {
2346 	uma_keg_t keg;
2347 
2348 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
2349 		return;
2350 	KEG_GET(zone, keg);
2351 
2352 	if (keg->uk_allocf == startup_alloc) {
2353 		/* Switch to the real allocator. */
2354 		if (keg->uk_flags & UMA_ZONE_PCPU)
2355 			keg->uk_allocf = pcpu_page_alloc;
2356 		else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 &&
2357 		    keg->uk_ppera > 1)
2358 			keg->uk_allocf = contig_alloc;
2359 		else
2360 			keg->uk_allocf = page_alloc;
2361 	}
2362 }
2363 
2364 static void
2365 zone_alloc_counters(uma_zone_t zone, void *unused)
2366 {
2367 
2368 	zone->uz_allocs = counter_u64_alloc(M_WAITOK);
2369 	zone->uz_frees = counter_u64_alloc(M_WAITOK);
2370 	zone->uz_fails = counter_u64_alloc(M_WAITOK);
2371 	zone->uz_xdomain = counter_u64_alloc(M_WAITOK);
2372 }
2373 
2374 static void
2375 zone_alloc_sysctl(uma_zone_t zone, void *unused)
2376 {
2377 	uma_zone_domain_t zdom;
2378 	uma_domain_t dom;
2379 	uma_keg_t keg;
2380 	struct sysctl_oid *oid, *domainoid;
2381 	int domains, i, cnt;
2382 	static const char *nokeg = "cache zone";
2383 	char *c;
2384 
2385 	/*
2386 	 * Make a sysctl safe copy of the zone name by removing
2387 	 * any special characters and handling dups by appending
2388 	 * an index.
2389 	 */
2390 	if (zone->uz_namecnt != 0) {
2391 		/* Count the number of decimal digits and '_' separator. */
2392 		for (i = 1, cnt = zone->uz_namecnt; cnt != 0; i++)
2393 			cnt /= 10;
2394 		zone->uz_ctlname = malloc(strlen(zone->uz_name) + i + 1,
2395 		    M_UMA, M_WAITOK);
2396 		sprintf(zone->uz_ctlname, "%s_%d", zone->uz_name,
2397 		    zone->uz_namecnt);
2398 	} else
2399 		zone->uz_ctlname = strdup(zone->uz_name, M_UMA);
2400 	for (c = zone->uz_ctlname; *c != '\0'; c++)
2401 		if (strchr("./\\ -", *c) != NULL)
2402 			*c = '_';
2403 
2404 	/*
2405 	 * Basic parameters at the root.
2406 	 */
2407 	zone->uz_oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_vm_uma),
2408 	    OID_AUTO, zone->uz_ctlname, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2409 	oid = zone->uz_oid;
2410 	SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2411 	    "size", CTLFLAG_RD, &zone->uz_size, 0, "Allocation size");
2412 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2413 	    "flags", CTLFLAG_RD | CTLTYPE_STRING | CTLFLAG_MPSAFE,
2414 	    zone, 0, sysctl_handle_uma_zone_flags, "A",
2415 	    "Allocator configuration flags");
2416 	SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2417 	    "bucket_size", CTLFLAG_RD, &zone->uz_bucket_size, 0,
2418 	    "Desired per-cpu cache size");
2419 	SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2420 	    "bucket_size_max", CTLFLAG_RD, &zone->uz_bucket_size_max, 0,
2421 	    "Maximum allowed per-cpu cache size");
2422 
2423 	/*
2424 	 * keg if present.
2425 	 */
2426 	if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
2427 		domains = vm_ndomains;
2428 	else
2429 		domains = 1;
2430 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2431 	    "keg", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2432 	keg = zone->uz_keg;
2433 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0) {
2434 		SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2435 		    "name", CTLFLAG_RD, keg->uk_name, "Keg name");
2436 		SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2437 		    "rsize", CTLFLAG_RD, &keg->uk_rsize, 0,
2438 		    "Real object size with alignment");
2439 		SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2440 		    "ppera", CTLFLAG_RD, &keg->uk_ppera, 0,
2441 		    "pages per-slab allocation");
2442 		SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2443 		    "ipers", CTLFLAG_RD, &keg->uk_ipers, 0,
2444 		    "items available per-slab");
2445 		SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2446 		    "align", CTLFLAG_RD, &keg->uk_align, 0,
2447 		    "item alignment mask");
2448 		SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2449 		    "reserve", CTLFLAG_RD, &keg->uk_reserve, 0,
2450 		    "number of reserved items");
2451 		SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2452 		    "efficiency", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2453 		    keg, 0, sysctl_handle_uma_slab_efficiency, "I",
2454 		    "Slab utilization (100 - internal fragmentation %)");
2455 		domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(oid),
2456 		    OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2457 		for (i = 0; i < domains; i++) {
2458 			dom = &keg->uk_domain[i];
2459 			oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2460 			    OID_AUTO, VM_DOMAIN(i)->vmd_name,
2461 			    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2462 			SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2463 			    "pages", CTLFLAG_RD, &dom->ud_pages, 0,
2464 			    "Total pages currently allocated from VM");
2465 			SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2466 			    "free_items", CTLFLAG_RD, &dom->ud_free_items, 0,
2467 			    "items free in the slab layer");
2468 		}
2469 	} else
2470 		SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2471 		    "name", CTLFLAG_RD, nokeg, "Keg name");
2472 
2473 	/*
2474 	 * Information about zone limits.
2475 	 */
2476 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2477 	    "limit", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2478 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2479 	    "items", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2480 	    zone, 0, sysctl_handle_uma_zone_items, "QU",
2481 	    "current number of allocated items if limit is set");
2482 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2483 	    "max_items", CTLFLAG_RD, &zone->uz_max_items, 0,
2484 	    "Maximum number of cached items");
2485 	SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2486 	    "sleepers", CTLFLAG_RD, &zone->uz_sleepers, 0,
2487 	    "Number of threads sleeping at limit");
2488 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2489 	    "sleeps", CTLFLAG_RD, &zone->uz_sleeps, 0,
2490 	    "Total zone limit sleeps");
2491 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2492 	    "bucket_max", CTLFLAG_RD, &zone->uz_bucket_max, 0,
2493 	    "Maximum number of items in each domain's bucket cache");
2494 
2495 	/*
2496 	 * Per-domain zone information.
2497 	 */
2498 	domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid),
2499 	    OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2500 	for (i = 0; i < domains; i++) {
2501 		zdom = ZDOM_GET(zone, i);
2502 		oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2503 		    OID_AUTO, VM_DOMAIN(i)->vmd_name,
2504 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2505 		SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2506 		    "nitems", CTLFLAG_RD, &zdom->uzd_nitems,
2507 		    "number of items in this domain");
2508 		SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2509 		    "imax", CTLFLAG_RD, &zdom->uzd_imax,
2510 		    "maximum item count in this period");
2511 		SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2512 		    "imin", CTLFLAG_RD, &zdom->uzd_imin,
2513 		    "minimum item count in this period");
2514 		SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2515 		    "wss", CTLFLAG_RD, &zdom->uzd_wss,
2516 		    "Working set size");
2517 	}
2518 
2519 	/*
2520 	 * General statistics.
2521 	 */
2522 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2523 	    "stats", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2524 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2525 	    "current", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2526 	    zone, 1, sysctl_handle_uma_zone_cur, "I",
2527 	    "Current number of allocated items");
2528 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2529 	    "allocs", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2530 	    zone, 0, sysctl_handle_uma_zone_allocs, "QU",
2531 	    "Total allocation calls");
2532 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2533 	    "frees", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2534 	    zone, 0, sysctl_handle_uma_zone_frees, "QU",
2535 	    "Total free calls");
2536 	SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2537 	    "fails", CTLFLAG_RD, &zone->uz_fails,
2538 	    "Number of allocation failures");
2539 	SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2540 	    "xdomain", CTLFLAG_RD, &zone->uz_xdomain,
2541 	    "Free calls from the wrong domain");
2542 }
2543 
2544 struct uma_zone_count {
2545 	const char	*name;
2546 	int		count;
2547 };
2548 
2549 static void
2550 zone_count(uma_zone_t zone, void *arg)
2551 {
2552 	struct uma_zone_count *cnt;
2553 
2554 	cnt = arg;
2555 	/*
2556 	 * Some zones are rapidly created with identical names and
2557 	 * destroyed out of order.  This can lead to gaps in the count.
2558 	 * Use one greater than the maximum observed for this name.
2559 	 */
2560 	if (strcmp(zone->uz_name, cnt->name) == 0)
2561 		cnt->count = MAX(cnt->count,
2562 		    zone->uz_namecnt + 1);
2563 }
2564 
2565 static void
2566 zone_update_caches(uma_zone_t zone)
2567 {
2568 	int i;
2569 
2570 	for (i = 0; i <= mp_maxid; i++) {
2571 		cache_set_uz_size(&zone->uz_cpu[i], zone->uz_size);
2572 		cache_set_uz_flags(&zone->uz_cpu[i], zone->uz_flags);
2573 	}
2574 }
2575 
2576 /*
2577  * Zone header ctor.  This initializes all fields, locks, etc.
2578  *
2579  * Arguments/Returns follow uma_ctor specifications
2580  *	udata  Actually uma_zctor_args
2581  */
2582 static int
2583 zone_ctor(void *mem, int size, void *udata, int flags)
2584 {
2585 	struct uma_zone_count cnt;
2586 	struct uma_zctor_args *arg = udata;
2587 	uma_zone_domain_t zdom;
2588 	uma_zone_t zone = mem;
2589 	uma_zone_t z;
2590 	uma_keg_t keg;
2591 	int i;
2592 
2593 	bzero(zone, size);
2594 	zone->uz_name = arg->name;
2595 	zone->uz_ctor = arg->ctor;
2596 	zone->uz_dtor = arg->dtor;
2597 	zone->uz_init = NULL;
2598 	zone->uz_fini = NULL;
2599 	zone->uz_sleeps = 0;
2600 	zone->uz_bucket_size = 0;
2601 	zone->uz_bucket_size_min = 0;
2602 	zone->uz_bucket_size_max = BUCKET_MAX;
2603 	zone->uz_flags = (arg->flags & UMA_ZONE_SMR);
2604 	zone->uz_warning = NULL;
2605 	/* The domain structures follow the cpu structures. */
2606 	zone->uz_bucket_max = ULONG_MAX;
2607 	timevalclear(&zone->uz_ratecheck);
2608 
2609 	/* Count the number of duplicate names. */
2610 	cnt.name = arg->name;
2611 	cnt.count = 0;
2612 	zone_foreach(zone_count, &cnt);
2613 	zone->uz_namecnt = cnt.count;
2614 	ZONE_CROSS_LOCK_INIT(zone);
2615 
2616 	for (i = 0; i < vm_ndomains; i++) {
2617 		zdom = ZDOM_GET(zone, i);
2618 		ZDOM_LOCK_INIT(zone, zdom, (arg->flags & UMA_ZONE_MTXCLASS));
2619 		STAILQ_INIT(&zdom->uzd_buckets);
2620 	}
2621 
2622 #ifdef INVARIANTS
2623 	if (arg->uminit == trash_init && arg->fini == trash_fini)
2624 		zone->uz_flags |= UMA_ZFLAG_TRASH | UMA_ZFLAG_CTORDTOR;
2625 #endif
2626 
2627 	/*
2628 	 * This is a pure cache zone, no kegs.
2629 	 */
2630 	if (arg->import) {
2631 		KASSERT((arg->flags & UMA_ZFLAG_CACHE) != 0,
2632 		    ("zone_ctor: Import specified for non-cache zone."));
2633 		zone->uz_flags = arg->flags;
2634 		zone->uz_size = arg->size;
2635 		zone->uz_import = arg->import;
2636 		zone->uz_release = arg->release;
2637 		zone->uz_arg = arg->arg;
2638 #ifdef NUMA
2639 		/*
2640 		 * Cache zones are round-robin unless a policy is
2641 		 * specified because they may have incompatible
2642 		 * constraints.
2643 		 */
2644 		if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2645 			zone->uz_flags |= UMA_ZONE_ROUNDROBIN;
2646 #endif
2647 		rw_wlock(&uma_rwlock);
2648 		LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link);
2649 		rw_wunlock(&uma_rwlock);
2650 		goto out;
2651 	}
2652 
2653 	/*
2654 	 * Use the regular zone/keg/slab allocator.
2655 	 */
2656 	zone->uz_import = zone_import;
2657 	zone->uz_release = zone_release;
2658 	zone->uz_arg = zone;
2659 	keg = arg->keg;
2660 
2661 	if (arg->flags & UMA_ZONE_SECONDARY) {
2662 		KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
2663 		    ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
2664 		KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
2665 		zone->uz_init = arg->uminit;
2666 		zone->uz_fini = arg->fini;
2667 		zone->uz_flags |= UMA_ZONE_SECONDARY;
2668 		rw_wlock(&uma_rwlock);
2669 		ZONE_LOCK(zone);
2670 		LIST_FOREACH(z, &keg->uk_zones, uz_link) {
2671 			if (LIST_NEXT(z, uz_link) == NULL) {
2672 				LIST_INSERT_AFTER(z, zone, uz_link);
2673 				break;
2674 			}
2675 		}
2676 		ZONE_UNLOCK(zone);
2677 		rw_wunlock(&uma_rwlock);
2678 	} else if (keg == NULL) {
2679 		if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
2680 		    arg->align, arg->flags)) == NULL)
2681 			return (ENOMEM);
2682 	} else {
2683 		struct uma_kctor_args karg;
2684 		int error;
2685 
2686 		/* We should only be here from uma_startup() */
2687 		karg.size = arg->size;
2688 		karg.uminit = arg->uminit;
2689 		karg.fini = arg->fini;
2690 		karg.align = arg->align;
2691 		karg.flags = (arg->flags & ~UMA_ZONE_SMR);
2692 		karg.zone = zone;
2693 		error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
2694 		    flags);
2695 		if (error)
2696 			return (error);
2697 	}
2698 
2699 	/* Inherit properties from the keg. */
2700 	zone->uz_keg = keg;
2701 	zone->uz_size = keg->uk_size;
2702 	zone->uz_flags |= (keg->uk_flags &
2703 	    (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
2704 
2705 out:
2706 	if (booted >= BOOT_PCPU) {
2707 		zone_alloc_counters(zone, NULL);
2708 		if (booted >= BOOT_RUNNING)
2709 			zone_alloc_sysctl(zone, NULL);
2710 	} else {
2711 		zone->uz_allocs = EARLY_COUNTER;
2712 		zone->uz_frees = EARLY_COUNTER;
2713 		zone->uz_fails = EARLY_COUNTER;
2714 	}
2715 
2716 	/* Caller requests a private SMR context. */
2717 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
2718 		zone->uz_smr = smr_create(zone->uz_name, 0, 0);
2719 
2720 	KASSERT((arg->flags & (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET)) !=
2721 	    (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET),
2722 	    ("Invalid zone flag combination"));
2723 	if (arg->flags & UMA_ZFLAG_INTERNAL)
2724 		zone->uz_bucket_size_max = zone->uz_bucket_size = 0;
2725 	if ((arg->flags & UMA_ZONE_MAXBUCKET) != 0)
2726 		zone->uz_bucket_size = BUCKET_MAX;
2727 	else if ((arg->flags & UMA_ZONE_MINBUCKET) != 0)
2728 		zone->uz_bucket_size_max = zone->uz_bucket_size = BUCKET_MIN;
2729 	else if ((arg->flags & UMA_ZONE_NOBUCKET) != 0)
2730 		zone->uz_bucket_size = 0;
2731 	else
2732 		zone->uz_bucket_size = bucket_select(zone->uz_size);
2733 	zone->uz_bucket_size_min = zone->uz_bucket_size;
2734 	if (zone->uz_dtor != NULL || zone->uz_ctor != NULL)
2735 		zone->uz_flags |= UMA_ZFLAG_CTORDTOR;
2736 	zone_update_caches(zone);
2737 
2738 	return (0);
2739 }
2740 
2741 /*
2742  * Keg header dtor.  This frees all data, destroys locks, frees the hash
2743  * table and removes the keg from the global list.
2744  *
2745  * Arguments/Returns follow uma_dtor specifications
2746  *	udata  unused
2747  */
2748 static void
2749 keg_dtor(void *arg, int size, void *udata)
2750 {
2751 	uma_keg_t keg;
2752 	uint32_t free, pages;
2753 	int i;
2754 
2755 	keg = (uma_keg_t)arg;
2756 	free = pages = 0;
2757 	for (i = 0; i < vm_ndomains; i++) {
2758 		free += keg->uk_domain[i].ud_free_items;
2759 		pages += keg->uk_domain[i].ud_pages;
2760 		KEG_LOCK_FINI(keg, i);
2761 	}
2762 	if (pages != 0)
2763 		printf("Freed UMA keg (%s) was not empty (%u items). "
2764 		    " Lost %u pages of memory.\n",
2765 		    keg->uk_name ? keg->uk_name : "",
2766 		    pages / keg->uk_ppera * keg->uk_ipers - free, pages);
2767 
2768 	hash_free(&keg->uk_hash);
2769 }
2770 
2771 /*
2772  * Zone header dtor.
2773  *
2774  * Arguments/Returns follow uma_dtor specifications
2775  *	udata  unused
2776  */
2777 static void
2778 zone_dtor(void *arg, int size, void *udata)
2779 {
2780 	uma_zone_t zone;
2781 	uma_keg_t keg;
2782 	int i;
2783 
2784 	zone = (uma_zone_t)arg;
2785 
2786 	sysctl_remove_oid(zone->uz_oid, 1, 1);
2787 
2788 	if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
2789 		cache_drain(zone);
2790 
2791 	rw_wlock(&uma_rwlock);
2792 	LIST_REMOVE(zone, uz_link);
2793 	rw_wunlock(&uma_rwlock);
2794 	if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) {
2795 		keg = zone->uz_keg;
2796 		keg->uk_reserve = 0;
2797 	}
2798 	zone_reclaim(zone, M_WAITOK, true);
2799 
2800 	/*
2801 	 * We only destroy kegs from non secondary/non cache zones.
2802 	 */
2803 	if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) {
2804 		keg = zone->uz_keg;
2805 		rw_wlock(&uma_rwlock);
2806 		LIST_REMOVE(keg, uk_link);
2807 		rw_wunlock(&uma_rwlock);
2808 		zone_free_item(kegs, keg, NULL, SKIP_NONE);
2809 	}
2810 	counter_u64_free(zone->uz_allocs);
2811 	counter_u64_free(zone->uz_frees);
2812 	counter_u64_free(zone->uz_fails);
2813 	counter_u64_free(zone->uz_xdomain);
2814 	free(zone->uz_ctlname, M_UMA);
2815 	for (i = 0; i < vm_ndomains; i++)
2816 		ZDOM_LOCK_FINI(ZDOM_GET(zone, i));
2817 	ZONE_CROSS_LOCK_FINI(zone);
2818 }
2819 
2820 static void
2821 zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *arg), void *arg)
2822 {
2823 	uma_keg_t keg;
2824 	uma_zone_t zone;
2825 
2826 	LIST_FOREACH(keg, &uma_kegs, uk_link) {
2827 		LIST_FOREACH(zone, &keg->uk_zones, uz_link)
2828 			zfunc(zone, arg);
2829 	}
2830 	LIST_FOREACH(zone, &uma_cachezones, uz_link)
2831 		zfunc(zone, arg);
2832 }
2833 
2834 /*
2835  * Traverses every zone in the system and calls a callback
2836  *
2837  * Arguments:
2838  *	zfunc  A pointer to a function which accepts a zone
2839  *		as an argument.
2840  *
2841  * Returns:
2842  *	Nothing
2843  */
2844 static void
2845 zone_foreach(void (*zfunc)(uma_zone_t, void *arg), void *arg)
2846 {
2847 
2848 	rw_rlock(&uma_rwlock);
2849 	zone_foreach_unlocked(zfunc, arg);
2850 	rw_runlock(&uma_rwlock);
2851 }
2852 
2853 /*
2854  * Initialize the kernel memory allocator.  This is done after pages can be
2855  * allocated but before general KVA is available.
2856  */
2857 void
2858 uma_startup1(vm_offset_t virtual_avail)
2859 {
2860 	struct uma_zctor_args args;
2861 	size_t ksize, zsize, size;
2862 	uma_keg_t primarykeg;
2863 	uintptr_t m;
2864 	int domain;
2865 	uint8_t pflag;
2866 
2867 	bootstart = bootmem = virtual_avail;
2868 
2869 	rw_init(&uma_rwlock, "UMA lock");
2870 	sx_init(&uma_reclaim_lock, "umareclaim");
2871 
2872 	ksize = sizeof(struct uma_keg) +
2873 	    (sizeof(struct uma_domain) * vm_ndomains);
2874 	ksize = roundup(ksize, UMA_SUPER_ALIGN);
2875 	zsize = sizeof(struct uma_zone) +
2876 	    (sizeof(struct uma_cache) * (mp_maxid + 1)) +
2877 	    (sizeof(struct uma_zone_domain) * vm_ndomains);
2878 	zsize = roundup(zsize, UMA_SUPER_ALIGN);
2879 
2880 	/* Allocate the zone of zones, zone of kegs, and zone of zones keg. */
2881 	size = (zsize * 2) + ksize;
2882 	for (domain = 0; domain < vm_ndomains; domain++) {
2883 		m = (uintptr_t)startup_alloc(NULL, size, domain, &pflag,
2884 		    M_NOWAIT | M_ZERO);
2885 		if (m != 0)
2886 			break;
2887 	}
2888 	zones = (uma_zone_t)m;
2889 	m += zsize;
2890 	kegs = (uma_zone_t)m;
2891 	m += zsize;
2892 	primarykeg = (uma_keg_t)m;
2893 
2894 	/* "manually" create the initial zone */
2895 	memset(&args, 0, sizeof(args));
2896 	args.name = "UMA Kegs";
2897 	args.size = ksize;
2898 	args.ctor = keg_ctor;
2899 	args.dtor = keg_dtor;
2900 	args.uminit = zero_init;
2901 	args.fini = NULL;
2902 	args.keg = primarykeg;
2903 	args.align = UMA_SUPER_ALIGN - 1;
2904 	args.flags = UMA_ZFLAG_INTERNAL;
2905 	zone_ctor(kegs, zsize, &args, M_WAITOK);
2906 
2907 	args.name = "UMA Zones";
2908 	args.size = zsize;
2909 	args.ctor = zone_ctor;
2910 	args.dtor = zone_dtor;
2911 	args.uminit = zero_init;
2912 	args.fini = NULL;
2913 	args.keg = NULL;
2914 	args.align = UMA_SUPER_ALIGN - 1;
2915 	args.flags = UMA_ZFLAG_INTERNAL;
2916 	zone_ctor(zones, zsize, &args, M_WAITOK);
2917 
2918 	/* Now make zones for slab headers */
2919 	slabzones[0] = uma_zcreate("UMA Slabs 0", SLABZONE0_SIZE,
2920 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2921 	slabzones[1] = uma_zcreate("UMA Slabs 1", SLABZONE1_SIZE,
2922 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2923 
2924 	hashzone = uma_zcreate("UMA Hash",
2925 	    sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
2926 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2927 
2928 	bucket_init();
2929 	smr_init();
2930 }
2931 
2932 #ifndef UMA_MD_SMALL_ALLOC
2933 extern void vm_radix_reserve_kva(void);
2934 #endif
2935 
2936 /*
2937  * Advertise the availability of normal kva allocations and switch to
2938  * the default back-end allocator.  Marks the KVA we consumed on startup
2939  * as used in the map.
2940  */
2941 void
2942 uma_startup2(void)
2943 {
2944 
2945 	if (bootstart != bootmem) {
2946 		vm_map_lock(kernel_map);
2947 		(void)vm_map_insert(kernel_map, NULL, 0, bootstart, bootmem,
2948 		    VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT);
2949 		vm_map_unlock(kernel_map);
2950 	}
2951 
2952 #ifndef UMA_MD_SMALL_ALLOC
2953 	/* Set up radix zone to use noobj_alloc. */
2954 	vm_radix_reserve_kva();
2955 #endif
2956 
2957 	booted = BOOT_KVA;
2958 	zone_foreach_unlocked(zone_kva_available, NULL);
2959 	bucket_enable();
2960 }
2961 
2962 /*
2963  * Allocate counters as early as possible so that boot-time allocations are
2964  * accounted more precisely.
2965  */
2966 static void
2967 uma_startup_pcpu(void *arg __unused)
2968 {
2969 
2970 	zone_foreach_unlocked(zone_alloc_counters, NULL);
2971 	booted = BOOT_PCPU;
2972 }
2973 SYSINIT(uma_startup_pcpu, SI_SUB_COUNTER, SI_ORDER_ANY, uma_startup_pcpu, NULL);
2974 
2975 /*
2976  * Finish our initialization steps.
2977  */
2978 static void
2979 uma_startup3(void *arg __unused)
2980 {
2981 
2982 #ifdef INVARIANTS
2983 	TUNABLE_INT_FETCH("vm.debug.divisor", &dbg_divisor);
2984 	uma_dbg_cnt = counter_u64_alloc(M_WAITOK);
2985 	uma_skip_cnt = counter_u64_alloc(M_WAITOK);
2986 #endif
2987 	zone_foreach_unlocked(zone_alloc_sysctl, NULL);
2988 	callout_init(&uma_callout, 1);
2989 	callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
2990 	booted = BOOT_RUNNING;
2991 
2992 	EVENTHANDLER_REGISTER(shutdown_post_sync, uma_shutdown, NULL,
2993 	    EVENTHANDLER_PRI_FIRST);
2994 }
2995 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
2996 
2997 static void
2998 uma_shutdown(void)
2999 {
3000 
3001 	booted = BOOT_SHUTDOWN;
3002 }
3003 
3004 static uma_keg_t
3005 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
3006 		int align, uint32_t flags)
3007 {
3008 	struct uma_kctor_args args;
3009 
3010 	args.size = size;
3011 	args.uminit = uminit;
3012 	args.fini = fini;
3013 	args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
3014 	args.flags = flags;
3015 	args.zone = zone;
3016 	return (zone_alloc_item(kegs, &args, UMA_ANYDOMAIN, M_WAITOK));
3017 }
3018 
3019 /* Public functions */
3020 /* See uma.h */
3021 void
3022 uma_set_align(int align)
3023 {
3024 
3025 	if (align != UMA_ALIGN_CACHE)
3026 		uma_align_cache = align;
3027 }
3028 
3029 /* See uma.h */
3030 uma_zone_t
3031 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
3032 		uma_init uminit, uma_fini fini, int align, uint32_t flags)
3033 
3034 {
3035 	struct uma_zctor_args args;
3036 	uma_zone_t res;
3037 
3038 	KASSERT(powerof2(align + 1), ("invalid zone alignment %d for \"%s\"",
3039 	    align, name));
3040 
3041 	/* This stuff is essential for the zone ctor */
3042 	memset(&args, 0, sizeof(args));
3043 	args.name = name;
3044 	args.size = size;
3045 	args.ctor = ctor;
3046 	args.dtor = dtor;
3047 	args.uminit = uminit;
3048 	args.fini = fini;
3049 #ifdef  INVARIANTS
3050 	/*
3051 	 * Inject procedures which check for memory use after free if we are
3052 	 * allowed to scramble the memory while it is not allocated.  This
3053 	 * requires that: UMA is actually able to access the memory, no init
3054 	 * or fini procedures, no dependency on the initial value of the
3055 	 * memory, and no (legitimate) use of the memory after free.  Note,
3056 	 * the ctor and dtor do not need to be empty.
3057 	 */
3058 	if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOTOUCH |
3059 	    UMA_ZONE_NOFREE))) && uminit == NULL && fini == NULL) {
3060 		args.uminit = trash_init;
3061 		args.fini = trash_fini;
3062 	}
3063 #endif
3064 	args.align = align;
3065 	args.flags = flags;
3066 	args.keg = NULL;
3067 
3068 	sx_slock(&uma_reclaim_lock);
3069 	res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3070 	sx_sunlock(&uma_reclaim_lock);
3071 
3072 	return (res);
3073 }
3074 
3075 /* See uma.h */
3076 uma_zone_t
3077 uma_zsecond_create(const char *name, uma_ctor ctor, uma_dtor dtor,
3078     uma_init zinit, uma_fini zfini, uma_zone_t primary)
3079 {
3080 	struct uma_zctor_args args;
3081 	uma_keg_t keg;
3082 	uma_zone_t res;
3083 
3084 	keg = primary->uz_keg;
3085 	memset(&args, 0, sizeof(args));
3086 	args.name = name;
3087 	args.size = keg->uk_size;
3088 	args.ctor = ctor;
3089 	args.dtor = dtor;
3090 	args.uminit = zinit;
3091 	args.fini = zfini;
3092 	args.align = keg->uk_align;
3093 	args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
3094 	args.keg = keg;
3095 
3096 	sx_slock(&uma_reclaim_lock);
3097 	res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3098 	sx_sunlock(&uma_reclaim_lock);
3099 
3100 	return (res);
3101 }
3102 
3103 /* See uma.h */
3104 uma_zone_t
3105 uma_zcache_create(const char *name, int size, uma_ctor ctor, uma_dtor dtor,
3106     uma_init zinit, uma_fini zfini, uma_import zimport, uma_release zrelease,
3107     void *arg, int flags)
3108 {
3109 	struct uma_zctor_args args;
3110 
3111 	memset(&args, 0, sizeof(args));
3112 	args.name = name;
3113 	args.size = size;
3114 	args.ctor = ctor;
3115 	args.dtor = dtor;
3116 	args.uminit = zinit;
3117 	args.fini = zfini;
3118 	args.import = zimport;
3119 	args.release = zrelease;
3120 	args.arg = arg;
3121 	args.align = 0;
3122 	args.flags = flags | UMA_ZFLAG_CACHE;
3123 
3124 	return (zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK));
3125 }
3126 
3127 /* See uma.h */
3128 void
3129 uma_zdestroy(uma_zone_t zone)
3130 {
3131 
3132 	/*
3133 	 * Large slabs are expensive to reclaim, so don't bother doing
3134 	 * unnecessary work if we're shutting down.
3135 	 */
3136 	if (booted == BOOT_SHUTDOWN &&
3137 	    zone->uz_fini == NULL && zone->uz_release == zone_release)
3138 		return;
3139 	sx_slock(&uma_reclaim_lock);
3140 	zone_free_item(zones, zone, NULL, SKIP_NONE);
3141 	sx_sunlock(&uma_reclaim_lock);
3142 }
3143 
3144 void
3145 uma_zwait(uma_zone_t zone)
3146 {
3147 
3148 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
3149 		uma_zfree_smr(zone, uma_zalloc_smr(zone, M_WAITOK));
3150 	else if ((zone->uz_flags & UMA_ZONE_PCPU) != 0)
3151 		uma_zfree_pcpu(zone, uma_zalloc_pcpu(zone, M_WAITOK));
3152 	else
3153 		uma_zfree(zone, uma_zalloc(zone, M_WAITOK));
3154 }
3155 
3156 void *
3157 uma_zalloc_pcpu_arg(uma_zone_t zone, void *udata, int flags)
3158 {
3159 	void *item, *pcpu_item;
3160 #ifdef SMP
3161 	int i;
3162 
3163 	MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3164 #endif
3165 	item = uma_zalloc_arg(zone, udata, flags & ~M_ZERO);
3166 	if (item == NULL)
3167 		return (NULL);
3168 	pcpu_item = zpcpu_base_to_offset(item);
3169 	if (flags & M_ZERO) {
3170 #ifdef SMP
3171 		for (i = 0; i <= mp_maxid; i++)
3172 			bzero(zpcpu_get_cpu(pcpu_item, i), zone->uz_size);
3173 #else
3174 		bzero(item, zone->uz_size);
3175 #endif
3176 	}
3177 	return (pcpu_item);
3178 }
3179 
3180 /*
3181  * A stub while both regular and pcpu cases are identical.
3182  */
3183 void
3184 uma_zfree_pcpu_arg(uma_zone_t zone, void *pcpu_item, void *udata)
3185 {
3186 	void *item;
3187 
3188 #ifdef SMP
3189 	MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3190 #endif
3191 	item = zpcpu_offset_to_base(pcpu_item);
3192 	uma_zfree_arg(zone, item, udata);
3193 }
3194 
3195 static inline void *
3196 item_ctor(uma_zone_t zone, int uz_flags, int size, void *udata, int flags,
3197     void *item)
3198 {
3199 #ifdef INVARIANTS
3200 	bool skipdbg;
3201 
3202 	skipdbg = uma_dbg_zskip(zone, item);
3203 	if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3204 	    zone->uz_ctor != trash_ctor)
3205 		trash_ctor(item, size, udata, flags);
3206 #endif
3207 	/* Check flags before loading ctor pointer. */
3208 	if (__predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0) &&
3209 	    __predict_false(zone->uz_ctor != NULL) &&
3210 	    zone->uz_ctor(item, size, udata, flags) != 0) {
3211 		counter_u64_add(zone->uz_fails, 1);
3212 		zone_free_item(zone, item, udata, SKIP_DTOR | SKIP_CNT);
3213 		return (NULL);
3214 	}
3215 #ifdef INVARIANTS
3216 	if (!skipdbg)
3217 		uma_dbg_alloc(zone, NULL, item);
3218 #endif
3219 	if (__predict_false(flags & M_ZERO))
3220 		return (memset(item, 0, size));
3221 
3222 	return (item);
3223 }
3224 
3225 static inline void
3226 item_dtor(uma_zone_t zone, void *item, int size, void *udata,
3227     enum zfreeskip skip)
3228 {
3229 #ifdef INVARIANTS
3230 	bool skipdbg;
3231 
3232 	skipdbg = uma_dbg_zskip(zone, item);
3233 	if (skip == SKIP_NONE && !skipdbg) {
3234 		if ((zone->uz_flags & UMA_ZONE_MALLOC) != 0)
3235 			uma_dbg_free(zone, udata, item);
3236 		else
3237 			uma_dbg_free(zone, NULL, item);
3238 	}
3239 #endif
3240 	if (__predict_true(skip < SKIP_DTOR)) {
3241 		if (zone->uz_dtor != NULL)
3242 			zone->uz_dtor(item, size, udata);
3243 #ifdef INVARIANTS
3244 		if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3245 		    zone->uz_dtor != trash_dtor)
3246 			trash_dtor(item, size, udata);
3247 #endif
3248 	}
3249 }
3250 
3251 #ifdef NUMA
3252 static int
3253 item_domain(void *item)
3254 {
3255 	int domain;
3256 
3257 	domain = _vm_phys_domain(vtophys(item));
3258 	KASSERT(domain >= 0 && domain < vm_ndomains,
3259 	    ("%s: unknown domain for item %p", __func__, item));
3260 	return (domain);
3261 }
3262 #endif
3263 
3264 #if defined(INVARIANTS) || defined(DEBUG_MEMGUARD) || defined(WITNESS)
3265 #define	UMA_ZALLOC_DEBUG
3266 static int
3267 uma_zalloc_debug(uma_zone_t zone, void **itemp, void *udata, int flags)
3268 {
3269 	int error;
3270 
3271 	error = 0;
3272 #ifdef WITNESS
3273 	if (flags & M_WAITOK) {
3274 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
3275 		    "uma_zalloc_debug: zone \"%s\"", zone->uz_name);
3276 	}
3277 #endif
3278 
3279 #ifdef INVARIANTS
3280 	KASSERT((flags & M_EXEC) == 0,
3281 	    ("uma_zalloc_debug: called with M_EXEC"));
3282 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3283 	    ("uma_zalloc_debug: called within spinlock or critical section"));
3284 	KASSERT((zone->uz_flags & UMA_ZONE_PCPU) == 0 || (flags & M_ZERO) == 0,
3285 	    ("uma_zalloc_debug: allocating from a pcpu zone with M_ZERO"));
3286 #endif
3287 
3288 #ifdef DEBUG_MEMGUARD
3289 	if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && memguard_cmp_zone(zone)) {
3290 		void *item;
3291 		item = memguard_alloc(zone->uz_size, flags);
3292 		if (item != NULL) {
3293 			error = EJUSTRETURN;
3294 			if (zone->uz_init != NULL &&
3295 			    zone->uz_init(item, zone->uz_size, flags) != 0) {
3296 				*itemp = NULL;
3297 				return (error);
3298 			}
3299 			if (zone->uz_ctor != NULL &&
3300 			    zone->uz_ctor(item, zone->uz_size, udata,
3301 			    flags) != 0) {
3302 				counter_u64_add(zone->uz_fails, 1);
3303 			    	zone->uz_fini(item, zone->uz_size);
3304 				*itemp = NULL;
3305 				return (error);
3306 			}
3307 			*itemp = item;
3308 			return (error);
3309 		}
3310 		/* This is unfortunate but should not be fatal. */
3311 	}
3312 #endif
3313 	return (error);
3314 }
3315 
3316 static int
3317 uma_zfree_debug(uma_zone_t zone, void *item, void *udata)
3318 {
3319 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3320 	    ("uma_zfree_debug: called with spinlock or critical section held"));
3321 
3322 #ifdef DEBUG_MEMGUARD
3323 	if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && is_memguard_addr(item)) {
3324 		if (zone->uz_dtor != NULL)
3325 			zone->uz_dtor(item, zone->uz_size, udata);
3326 		if (zone->uz_fini != NULL)
3327 			zone->uz_fini(item, zone->uz_size);
3328 		memguard_free(item);
3329 		return (EJUSTRETURN);
3330 	}
3331 #endif
3332 	return (0);
3333 }
3334 #endif
3335 
3336 static inline void *
3337 cache_alloc_item(uma_zone_t zone, uma_cache_t cache, uma_cache_bucket_t bucket,
3338     void *udata, int flags)
3339 {
3340 	void *item;
3341 	int size, uz_flags;
3342 
3343 	item = cache_bucket_pop(cache, bucket);
3344 	size = cache_uz_size(cache);
3345 	uz_flags = cache_uz_flags(cache);
3346 	critical_exit();
3347 	return (item_ctor(zone, uz_flags, size, udata, flags, item));
3348 }
3349 
3350 static __noinline void *
3351 cache_alloc_retry(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3352 {
3353 	uma_cache_bucket_t bucket;
3354 	int domain;
3355 
3356 	while (cache_alloc(zone, cache, udata, flags)) {
3357 		cache = &zone->uz_cpu[curcpu];
3358 		bucket = &cache->uc_allocbucket;
3359 		if (__predict_false(bucket->ucb_cnt == 0))
3360 			continue;
3361 		return (cache_alloc_item(zone, cache, bucket, udata, flags));
3362 	}
3363 	critical_exit();
3364 
3365 	/*
3366 	 * We can not get a bucket so try to return a single item.
3367 	 */
3368 	if (zone->uz_flags & UMA_ZONE_FIRSTTOUCH)
3369 		domain = PCPU_GET(domain);
3370 	else
3371 		domain = UMA_ANYDOMAIN;
3372 	return (zone_alloc_item(zone, udata, domain, flags));
3373 }
3374 
3375 /* See uma.h */
3376 void *
3377 uma_zalloc_smr(uma_zone_t zone, int flags)
3378 {
3379 	uma_cache_bucket_t bucket;
3380 	uma_cache_t cache;
3381 
3382 #ifdef UMA_ZALLOC_DEBUG
3383 	void *item;
3384 
3385 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
3386 	    ("uma_zalloc_arg: called with non-SMR zone."));
3387 	if (uma_zalloc_debug(zone, &item, NULL, flags) == EJUSTRETURN)
3388 		return (item);
3389 #endif
3390 
3391 	critical_enter();
3392 	cache = &zone->uz_cpu[curcpu];
3393 	bucket = &cache->uc_allocbucket;
3394 	if (__predict_false(bucket->ucb_cnt == 0))
3395 		return (cache_alloc_retry(zone, cache, NULL, flags));
3396 	return (cache_alloc_item(zone, cache, bucket, NULL, flags));
3397 }
3398 
3399 /* See uma.h */
3400 void *
3401 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
3402 {
3403 	uma_cache_bucket_t bucket;
3404 	uma_cache_t cache;
3405 
3406 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3407 	random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3408 
3409 	/* This is the fast path allocation */
3410 	CTR3(KTR_UMA, "uma_zalloc_arg zone %s(%p) flags %d", zone->uz_name,
3411 	    zone, flags);
3412 
3413 #ifdef UMA_ZALLOC_DEBUG
3414 	void *item;
3415 
3416 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3417 	    ("uma_zalloc_arg: called with SMR zone."));
3418 	if (uma_zalloc_debug(zone, &item, udata, flags) == EJUSTRETURN)
3419 		return (item);
3420 #endif
3421 
3422 	/*
3423 	 * If possible, allocate from the per-CPU cache.  There are two
3424 	 * requirements for safe access to the per-CPU cache: (1) the thread
3425 	 * accessing the cache must not be preempted or yield during access,
3426 	 * and (2) the thread must not migrate CPUs without switching which
3427 	 * cache it accesses.  We rely on a critical section to prevent
3428 	 * preemption and migration.  We release the critical section in
3429 	 * order to acquire the zone mutex if we are unable to allocate from
3430 	 * the current cache; when we re-acquire the critical section, we
3431 	 * must detect and handle migration if it has occurred.
3432 	 */
3433 	critical_enter();
3434 	cache = &zone->uz_cpu[curcpu];
3435 	bucket = &cache->uc_allocbucket;
3436 	if (__predict_false(bucket->ucb_cnt == 0))
3437 		return (cache_alloc_retry(zone, cache, udata, flags));
3438 	return (cache_alloc_item(zone, cache, bucket, udata, flags));
3439 }
3440 
3441 /*
3442  * Replenish an alloc bucket and possibly restore an old one.  Called in
3443  * a critical section.  Returns in a critical section.
3444  *
3445  * A false return value indicates an allocation failure.
3446  * A true return value indicates success and the caller should retry.
3447  */
3448 static __noinline bool
3449 cache_alloc(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3450 {
3451 	uma_bucket_t bucket;
3452 	int curdomain, domain;
3453 	bool new;
3454 
3455 	CRITICAL_ASSERT(curthread);
3456 
3457 	/*
3458 	 * If we have run out of items in our alloc bucket see
3459 	 * if we can switch with the free bucket.
3460 	 *
3461 	 * SMR Zones can't re-use the free bucket until the sequence has
3462 	 * expired.
3463 	 */
3464 	if ((cache_uz_flags(cache) & UMA_ZONE_SMR) == 0 &&
3465 	    cache->uc_freebucket.ucb_cnt != 0) {
3466 		cache_bucket_swap(&cache->uc_freebucket,
3467 		    &cache->uc_allocbucket);
3468 		return (true);
3469 	}
3470 
3471 	/*
3472 	 * Discard any empty allocation bucket while we hold no locks.
3473 	 */
3474 	bucket = cache_bucket_unload_alloc(cache);
3475 	critical_exit();
3476 
3477 	if (bucket != NULL) {
3478 		KASSERT(bucket->ub_cnt == 0,
3479 		    ("cache_alloc: Entered with non-empty alloc bucket."));
3480 		bucket_free(zone, bucket, udata);
3481 	}
3482 
3483 	/*
3484 	 * Attempt to retrieve the item from the per-CPU cache has failed, so
3485 	 * we must go back to the zone.  This requires the zdom lock, so we
3486 	 * must drop the critical section, then re-acquire it when we go back
3487 	 * to the cache.  Since the critical section is released, we may be
3488 	 * preempted or migrate.  As such, make sure not to maintain any
3489 	 * thread-local state specific to the cache from prior to releasing
3490 	 * the critical section.
3491 	 */
3492 	domain = PCPU_GET(domain);
3493 	if ((cache_uz_flags(cache) & UMA_ZONE_ROUNDROBIN) != 0 ||
3494 	    VM_DOMAIN_EMPTY(domain))
3495 		domain = zone_domain_highest(zone, domain);
3496 	bucket = cache_fetch_bucket(zone, cache, domain);
3497 	if (bucket == NULL && zone->uz_bucket_size != 0 && !bucketdisable) {
3498 		bucket = zone_alloc_bucket(zone, udata, domain, flags);
3499 		new = true;
3500 	} else {
3501 		new = false;
3502 	}
3503 
3504 	CTR3(KTR_UMA, "uma_zalloc: zone %s(%p) bucket zone returned %p",
3505 	    zone->uz_name, zone, bucket);
3506 	if (bucket == NULL) {
3507 		critical_enter();
3508 		return (false);
3509 	}
3510 
3511 	/*
3512 	 * See if we lost the race or were migrated.  Cache the
3513 	 * initialized bucket to make this less likely or claim
3514 	 * the memory directly.
3515 	 */
3516 	critical_enter();
3517 	cache = &zone->uz_cpu[curcpu];
3518 	if (cache->uc_allocbucket.ucb_bucket == NULL &&
3519 	    ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) == 0 ||
3520 	    (curdomain = PCPU_GET(domain)) == domain ||
3521 	    VM_DOMAIN_EMPTY(curdomain))) {
3522 		if (new)
3523 			atomic_add_long(&ZDOM_GET(zone, domain)->uzd_imax,
3524 			    bucket->ub_cnt);
3525 		cache_bucket_load_alloc(cache, bucket);
3526 		return (true);
3527 	}
3528 
3529 	/*
3530 	 * We lost the race, release this bucket and start over.
3531 	 */
3532 	critical_exit();
3533 	zone_put_bucket(zone, domain, bucket, udata, false);
3534 	critical_enter();
3535 
3536 	return (true);
3537 }
3538 
3539 void *
3540 uma_zalloc_domain(uma_zone_t zone, void *udata, int domain, int flags)
3541 {
3542 #ifdef NUMA
3543 	uma_bucket_t bucket;
3544 	uma_zone_domain_t zdom;
3545 	void *item;
3546 #endif
3547 
3548 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3549 	random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3550 
3551 	/* This is the fast path allocation */
3552 	CTR4(KTR_UMA, "uma_zalloc_domain zone %s(%p) domain %d flags %d",
3553 	    zone->uz_name, zone, domain, flags);
3554 
3555 	if (flags & M_WAITOK) {
3556 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
3557 		    "uma_zalloc_domain: zone \"%s\"", zone->uz_name);
3558 	}
3559 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3560 	    ("uma_zalloc_domain: called with spinlock or critical section held"));
3561 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3562 	    ("uma_zalloc_domain: called with SMR zone."));
3563 #ifdef NUMA
3564 	KASSERT((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0,
3565 	    ("uma_zalloc_domain: called with non-FIRSTTOUCH zone."));
3566 
3567 	if (vm_ndomains == 1)
3568 		return (uma_zalloc_arg(zone, udata, flags));
3569 
3570 	/*
3571 	 * Try to allocate from the bucket cache before falling back to the keg.
3572 	 * We could try harder and attempt to allocate from per-CPU caches or
3573 	 * the per-domain cross-domain buckets, but the complexity is probably
3574 	 * not worth it.  It is more important that frees of previous
3575 	 * cross-domain allocations do not blow up the cache.
3576 	 */
3577 	zdom = zone_domain_lock(zone, domain);
3578 	if ((bucket = zone_fetch_bucket(zone, zdom, false)) != NULL) {
3579 		item = bucket->ub_bucket[bucket->ub_cnt - 1];
3580 #ifdef INVARIANTS
3581 		bucket->ub_bucket[bucket->ub_cnt - 1] = NULL;
3582 #endif
3583 		bucket->ub_cnt--;
3584 		zone_put_bucket(zone, domain, bucket, udata, true);
3585 		item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata,
3586 		    flags, item);
3587 		if (item != NULL) {
3588 			KASSERT(item_domain(item) == domain,
3589 			    ("%s: bucket cache item %p from wrong domain",
3590 			    __func__, item));
3591 			counter_u64_add(zone->uz_allocs, 1);
3592 		}
3593 		return (item);
3594 	}
3595 	ZDOM_UNLOCK(zdom);
3596 	return (zone_alloc_item(zone, udata, domain, flags));
3597 #else
3598 	return (uma_zalloc_arg(zone, udata, flags));
3599 #endif
3600 }
3601 
3602 /*
3603  * Find a slab with some space.  Prefer slabs that are partially used over those
3604  * that are totally full.  This helps to reduce fragmentation.
3605  *
3606  * If 'rr' is 1, search all domains starting from 'domain'.  Otherwise check
3607  * only 'domain'.
3608  */
3609 static uma_slab_t
3610 keg_first_slab(uma_keg_t keg, int domain, bool rr)
3611 {
3612 	uma_domain_t dom;
3613 	uma_slab_t slab;
3614 	int start;
3615 
3616 	KASSERT(domain >= 0 && domain < vm_ndomains,
3617 	    ("keg_first_slab: domain %d out of range", domain));
3618 	KEG_LOCK_ASSERT(keg, domain);
3619 
3620 	slab = NULL;
3621 	start = domain;
3622 	do {
3623 		dom = &keg->uk_domain[domain];
3624 		if ((slab = LIST_FIRST(&dom->ud_part_slab)) != NULL)
3625 			return (slab);
3626 		if ((slab = LIST_FIRST(&dom->ud_free_slab)) != NULL) {
3627 			LIST_REMOVE(slab, us_link);
3628 			dom->ud_free_slabs--;
3629 			LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
3630 			return (slab);
3631 		}
3632 		if (rr)
3633 			domain = (domain + 1) % vm_ndomains;
3634 	} while (domain != start);
3635 
3636 	return (NULL);
3637 }
3638 
3639 /*
3640  * Fetch an existing slab from a free or partial list.  Returns with the
3641  * keg domain lock held if a slab was found or unlocked if not.
3642  */
3643 static uma_slab_t
3644 keg_fetch_free_slab(uma_keg_t keg, int domain, bool rr, int flags)
3645 {
3646 	uma_slab_t slab;
3647 	uint32_t reserve;
3648 
3649 	/* HASH has a single free list. */
3650 	if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
3651 		domain = 0;
3652 
3653 	KEG_LOCK(keg, domain);
3654 	reserve = (flags & M_USE_RESERVE) != 0 ? 0 : keg->uk_reserve;
3655 	if (keg->uk_domain[domain].ud_free_items <= reserve ||
3656 	    (slab = keg_first_slab(keg, domain, rr)) == NULL) {
3657 		KEG_UNLOCK(keg, domain);
3658 		return (NULL);
3659 	}
3660 	return (slab);
3661 }
3662 
3663 static uma_slab_t
3664 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int rdomain, const int flags)
3665 {
3666 	struct vm_domainset_iter di;
3667 	uma_slab_t slab;
3668 	int aflags, domain;
3669 	bool rr;
3670 
3671 restart:
3672 	/*
3673 	 * Use the keg's policy if upper layers haven't already specified a
3674 	 * domain (as happens with first-touch zones).
3675 	 *
3676 	 * To avoid races we run the iterator with the keg lock held, but that
3677 	 * means that we cannot allow the vm_domainset layer to sleep.  Thus,
3678 	 * clear M_WAITOK and handle low memory conditions locally.
3679 	 */
3680 	rr = rdomain == UMA_ANYDOMAIN;
3681 	if (rr) {
3682 		aflags = (flags & ~M_WAITOK) | M_NOWAIT;
3683 		vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
3684 		    &aflags);
3685 	} else {
3686 		aflags = flags;
3687 		domain = rdomain;
3688 	}
3689 
3690 	for (;;) {
3691 		slab = keg_fetch_free_slab(keg, domain, rr, flags);
3692 		if (slab != NULL)
3693 			return (slab);
3694 
3695 		/*
3696 		 * M_NOVM means don't ask at all!
3697 		 */
3698 		if (flags & M_NOVM)
3699 			break;
3700 
3701 		slab = keg_alloc_slab(keg, zone, domain, flags, aflags);
3702 		if (slab != NULL)
3703 			return (slab);
3704 		if (!rr && (flags & M_WAITOK) == 0)
3705 			break;
3706 		if (rr && vm_domainset_iter_policy(&di, &domain) != 0) {
3707 			if ((flags & M_WAITOK) != 0) {
3708 				vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask, 0);
3709 				goto restart;
3710 			}
3711 			break;
3712 		}
3713 	}
3714 
3715 	/*
3716 	 * We might not have been able to get a slab but another cpu
3717 	 * could have while we were unlocked.  Check again before we
3718 	 * fail.
3719 	 */
3720 	if ((slab = keg_fetch_free_slab(keg, domain, rr, flags)) != NULL)
3721 		return (slab);
3722 
3723 	return (NULL);
3724 }
3725 
3726 static void *
3727 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
3728 {
3729 	uma_domain_t dom;
3730 	void *item;
3731 	int freei;
3732 
3733 	KEG_LOCK_ASSERT(keg, slab->us_domain);
3734 
3735 	dom = &keg->uk_domain[slab->us_domain];
3736 	freei = BIT_FFS(keg->uk_ipers, &slab->us_free) - 1;
3737 	BIT_CLR(keg->uk_ipers, freei, &slab->us_free);
3738 	item = slab_item(slab, keg, freei);
3739 	slab->us_freecount--;
3740 	dom->ud_free_items--;
3741 
3742 	/*
3743 	 * Move this slab to the full list.  It must be on the partial list, so
3744 	 * we do not need to update the free slab count.  In particular,
3745 	 * keg_fetch_slab() always returns slabs on the partial list.
3746 	 */
3747 	if (slab->us_freecount == 0) {
3748 		LIST_REMOVE(slab, us_link);
3749 		LIST_INSERT_HEAD(&dom->ud_full_slab, slab, us_link);
3750 	}
3751 
3752 	return (item);
3753 }
3754 
3755 static int
3756 zone_import(void *arg, void **bucket, int max, int domain, int flags)
3757 {
3758 	uma_domain_t dom;
3759 	uma_zone_t zone;
3760 	uma_slab_t slab;
3761 	uma_keg_t keg;
3762 #ifdef NUMA
3763 	int stripe;
3764 #endif
3765 	int i;
3766 
3767 	zone = arg;
3768 	slab = NULL;
3769 	keg = zone->uz_keg;
3770 	/* Try to keep the buckets totally full */
3771 	for (i = 0; i < max; ) {
3772 		if ((slab = keg_fetch_slab(keg, zone, domain, flags)) == NULL)
3773 			break;
3774 #ifdef NUMA
3775 		stripe = howmany(max, vm_ndomains);
3776 #endif
3777 		dom = &keg->uk_domain[slab->us_domain];
3778 		do {
3779 			bucket[i++] = slab_alloc_item(keg, slab);
3780 			if (dom->ud_free_items <= keg->uk_reserve) {
3781 				/*
3782 				 * Avoid depleting the reserve after a
3783 				 * successful item allocation, even if
3784 				 * M_USE_RESERVE is specified.
3785 				 */
3786 				KEG_UNLOCK(keg, slab->us_domain);
3787 				goto out;
3788 			}
3789 #ifdef NUMA
3790 			/*
3791 			 * If the zone is striped we pick a new slab for every
3792 			 * N allocations.  Eliminating this conditional will
3793 			 * instead pick a new domain for each bucket rather
3794 			 * than stripe within each bucket.  The current option
3795 			 * produces more fragmentation and requires more cpu
3796 			 * time but yields better distribution.
3797 			 */
3798 			if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0 &&
3799 			    vm_ndomains > 1 && --stripe == 0)
3800 				break;
3801 #endif
3802 		} while (slab->us_freecount != 0 && i < max);
3803 		KEG_UNLOCK(keg, slab->us_domain);
3804 
3805 		/* Don't block if we allocated any successfully. */
3806 		flags &= ~M_WAITOK;
3807 		flags |= M_NOWAIT;
3808 	}
3809 out:
3810 	return i;
3811 }
3812 
3813 static int
3814 zone_alloc_limit_hard(uma_zone_t zone, int count, int flags)
3815 {
3816 	uint64_t old, new, total, max;
3817 
3818 	/*
3819 	 * The hard case.  We're going to sleep because there were existing
3820 	 * sleepers or because we ran out of items.  This routine enforces
3821 	 * fairness by keeping fifo order.
3822 	 *
3823 	 * First release our ill gotten gains and make some noise.
3824 	 */
3825 	for (;;) {
3826 		zone_free_limit(zone, count);
3827 		zone_log_warning(zone);
3828 		zone_maxaction(zone);
3829 		if (flags & M_NOWAIT)
3830 			return (0);
3831 
3832 		/*
3833 		 * We need to allocate an item or set ourself as a sleeper
3834 		 * while the sleepq lock is held to avoid wakeup races.  This
3835 		 * is essentially a home rolled semaphore.
3836 		 */
3837 		sleepq_lock(&zone->uz_max_items);
3838 		old = zone->uz_items;
3839 		do {
3840 			MPASS(UZ_ITEMS_SLEEPERS(old) < UZ_ITEMS_SLEEPERS_MAX);
3841 			/* Cache the max since we will evaluate twice. */
3842 			max = zone->uz_max_items;
3843 			if (UZ_ITEMS_SLEEPERS(old) != 0 ||
3844 			    UZ_ITEMS_COUNT(old) >= max)
3845 				new = old + UZ_ITEMS_SLEEPER;
3846 			else
3847 				new = old + MIN(count, max - old);
3848 		} while (atomic_fcmpset_64(&zone->uz_items, &old, new) == 0);
3849 
3850 		/* We may have successfully allocated under the sleepq lock. */
3851 		if (UZ_ITEMS_SLEEPERS(new) == 0) {
3852 			sleepq_release(&zone->uz_max_items);
3853 			return (new - old);
3854 		}
3855 
3856 		/*
3857 		 * This is in a different cacheline from uz_items so that we
3858 		 * don't constantly invalidate the fastpath cacheline when we
3859 		 * adjust item counts.  This could be limited to toggling on
3860 		 * transitions.
3861 		 */
3862 		atomic_add_32(&zone->uz_sleepers, 1);
3863 		atomic_add_64(&zone->uz_sleeps, 1);
3864 
3865 		/*
3866 		 * We have added ourselves as a sleeper.  The sleepq lock
3867 		 * protects us from wakeup races.  Sleep now and then retry.
3868 		 */
3869 		sleepq_add(&zone->uz_max_items, NULL, "zonelimit", 0, 0);
3870 		sleepq_wait(&zone->uz_max_items, PVM);
3871 
3872 		/*
3873 		 * After wakeup, remove ourselves as a sleeper and try
3874 		 * again.  We no longer have the sleepq lock for protection.
3875 		 *
3876 		 * Subract ourselves as a sleeper while attempting to add
3877 		 * our count.
3878 		 */
3879 		atomic_subtract_32(&zone->uz_sleepers, 1);
3880 		old = atomic_fetchadd_64(&zone->uz_items,
3881 		    -(UZ_ITEMS_SLEEPER - count));
3882 		/* We're no longer a sleeper. */
3883 		old -= UZ_ITEMS_SLEEPER;
3884 
3885 		/*
3886 		 * If we're still at the limit, restart.  Notably do not
3887 		 * block on other sleepers.  Cache the max value to protect
3888 		 * against changes via sysctl.
3889 		 */
3890 		total = UZ_ITEMS_COUNT(old);
3891 		max = zone->uz_max_items;
3892 		if (total >= max)
3893 			continue;
3894 		/* Truncate if necessary, otherwise wake other sleepers. */
3895 		if (total + count > max) {
3896 			zone_free_limit(zone, total + count - max);
3897 			count = max - total;
3898 		} else if (total + count < max && UZ_ITEMS_SLEEPERS(old) != 0)
3899 			wakeup_one(&zone->uz_max_items);
3900 
3901 		return (count);
3902 	}
3903 }
3904 
3905 /*
3906  * Allocate 'count' items from our max_items limit.  Returns the number
3907  * available.  If M_NOWAIT is not specified it will sleep until at least
3908  * one item can be allocated.
3909  */
3910 static int
3911 zone_alloc_limit(uma_zone_t zone, int count, int flags)
3912 {
3913 	uint64_t old;
3914 	uint64_t max;
3915 
3916 	max = zone->uz_max_items;
3917 	MPASS(max > 0);
3918 
3919 	/*
3920 	 * We expect normal allocations to succeed with a simple
3921 	 * fetchadd.
3922 	 */
3923 	old = atomic_fetchadd_64(&zone->uz_items, count);
3924 	if (__predict_true(old + count <= max))
3925 		return (count);
3926 
3927 	/*
3928 	 * If we had some items and no sleepers just return the
3929 	 * truncated value.  We have to release the excess space
3930 	 * though because that may wake sleepers who weren't woken
3931 	 * because we were temporarily over the limit.
3932 	 */
3933 	if (old < max) {
3934 		zone_free_limit(zone, (old + count) - max);
3935 		return (max - old);
3936 	}
3937 	return (zone_alloc_limit_hard(zone, count, flags));
3938 }
3939 
3940 /*
3941  * Free a number of items back to the limit.
3942  */
3943 static void
3944 zone_free_limit(uma_zone_t zone, int count)
3945 {
3946 	uint64_t old;
3947 
3948 	MPASS(count > 0);
3949 
3950 	/*
3951 	 * In the common case we either have no sleepers or
3952 	 * are still over the limit and can just return.
3953 	 */
3954 	old = atomic_fetchadd_64(&zone->uz_items, -count);
3955 	if (__predict_true(UZ_ITEMS_SLEEPERS(old) == 0 ||
3956 	   UZ_ITEMS_COUNT(old) - count >= zone->uz_max_items))
3957 		return;
3958 
3959 	/*
3960 	 * Moderate the rate of wakeups.  Sleepers will continue
3961 	 * to generate wakeups if necessary.
3962 	 */
3963 	wakeup_one(&zone->uz_max_items);
3964 }
3965 
3966 static uma_bucket_t
3967 zone_alloc_bucket(uma_zone_t zone, void *udata, int domain, int flags)
3968 {
3969 	uma_bucket_t bucket;
3970 	int maxbucket, cnt;
3971 
3972 	CTR3(KTR_UMA, "zone_alloc_bucket zone %s(%p) domain %d", zone->uz_name,
3973 	    zone, domain);
3974 
3975 	/* Avoid allocs targeting empty domains. */
3976 	if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
3977 		domain = UMA_ANYDOMAIN;
3978 	else if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
3979 		domain = UMA_ANYDOMAIN;
3980 
3981 	if (zone->uz_max_items > 0)
3982 		maxbucket = zone_alloc_limit(zone, zone->uz_bucket_size,
3983 		    M_NOWAIT);
3984 	else
3985 		maxbucket = zone->uz_bucket_size;
3986 	if (maxbucket == 0)
3987 		return (false);
3988 
3989 	/* Don't wait for buckets, preserve caller's NOVM setting. */
3990 	bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
3991 	if (bucket == NULL) {
3992 		cnt = 0;
3993 		goto out;
3994 	}
3995 
3996 	bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
3997 	    MIN(maxbucket, bucket->ub_entries), domain, flags);
3998 
3999 	/*
4000 	 * Initialize the memory if necessary.
4001 	 */
4002 	if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
4003 		int i;
4004 
4005 		for (i = 0; i < bucket->ub_cnt; i++)
4006 			if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size,
4007 			    flags) != 0)
4008 				break;
4009 		/*
4010 		 * If we couldn't initialize the whole bucket, put the
4011 		 * rest back onto the freelist.
4012 		 */
4013 		if (i != bucket->ub_cnt) {
4014 			zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
4015 			    bucket->ub_cnt - i);
4016 #ifdef INVARIANTS
4017 			bzero(&bucket->ub_bucket[i],
4018 			    sizeof(void *) * (bucket->ub_cnt - i));
4019 #endif
4020 			bucket->ub_cnt = i;
4021 		}
4022 	}
4023 
4024 	cnt = bucket->ub_cnt;
4025 	if (bucket->ub_cnt == 0) {
4026 		bucket_free(zone, bucket, udata);
4027 		counter_u64_add(zone->uz_fails, 1);
4028 		bucket = NULL;
4029 	}
4030 out:
4031 	if (zone->uz_max_items > 0 && cnt < maxbucket)
4032 		zone_free_limit(zone, maxbucket - cnt);
4033 
4034 	return (bucket);
4035 }
4036 
4037 /*
4038  * Allocates a single item from a zone.
4039  *
4040  * Arguments
4041  *	zone   The zone to alloc for.
4042  *	udata  The data to be passed to the constructor.
4043  *	domain The domain to allocate from or UMA_ANYDOMAIN.
4044  *	flags  M_WAITOK, M_NOWAIT, M_ZERO.
4045  *
4046  * Returns
4047  *	NULL if there is no memory and M_NOWAIT is set
4048  *	An item if successful
4049  */
4050 
4051 static void *
4052 zone_alloc_item(uma_zone_t zone, void *udata, int domain, int flags)
4053 {
4054 	void *item;
4055 
4056 	if (zone->uz_max_items > 0 && zone_alloc_limit(zone, 1, flags) == 0) {
4057 		counter_u64_add(zone->uz_fails, 1);
4058 		return (NULL);
4059 	}
4060 
4061 	/* Avoid allocs targeting empty domains. */
4062 	if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
4063 		domain = UMA_ANYDOMAIN;
4064 
4065 	if (zone->uz_import(zone->uz_arg, &item, 1, domain, flags) != 1)
4066 		goto fail_cnt;
4067 
4068 	/*
4069 	 * We have to call both the zone's init (not the keg's init)
4070 	 * and the zone's ctor.  This is because the item is going from
4071 	 * a keg slab directly to the user, and the user is expecting it
4072 	 * to be both zone-init'd as well as zone-ctor'd.
4073 	 */
4074 	if (zone->uz_init != NULL) {
4075 		if (zone->uz_init(item, zone->uz_size, flags) != 0) {
4076 			zone_free_item(zone, item, udata, SKIP_FINI | SKIP_CNT);
4077 			goto fail_cnt;
4078 		}
4079 	}
4080 	item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata, flags,
4081 	    item);
4082 	if (item == NULL)
4083 		goto fail;
4084 
4085 	counter_u64_add(zone->uz_allocs, 1);
4086 	CTR3(KTR_UMA, "zone_alloc_item item %p from %s(%p)", item,
4087 	    zone->uz_name, zone);
4088 
4089 	return (item);
4090 
4091 fail_cnt:
4092 	counter_u64_add(zone->uz_fails, 1);
4093 fail:
4094 	if (zone->uz_max_items > 0)
4095 		zone_free_limit(zone, 1);
4096 	CTR2(KTR_UMA, "zone_alloc_item failed from %s(%p)",
4097 	    zone->uz_name, zone);
4098 
4099 	return (NULL);
4100 }
4101 
4102 /* See uma.h */
4103 void
4104 uma_zfree_smr(uma_zone_t zone, void *item)
4105 {
4106 	uma_cache_t cache;
4107 	uma_cache_bucket_t bucket;
4108 	int itemdomain, uz_flags;
4109 
4110 #ifdef UMA_ZALLOC_DEBUG
4111 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
4112 	    ("uma_zfree_smr: called with non-SMR zone."));
4113 	KASSERT(item != NULL, ("uma_zfree_smr: Called with NULL pointer."));
4114 	SMR_ASSERT_NOT_ENTERED(zone->uz_smr);
4115 	if (uma_zfree_debug(zone, item, NULL) == EJUSTRETURN)
4116 		return;
4117 #endif
4118 	cache = &zone->uz_cpu[curcpu];
4119 	uz_flags = cache_uz_flags(cache);
4120 	itemdomain = 0;
4121 #ifdef NUMA
4122 	if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4123 		itemdomain = item_domain(item);
4124 #endif
4125 	critical_enter();
4126 	do {
4127 		cache = &zone->uz_cpu[curcpu];
4128 		/* SMR Zones must free to the free bucket. */
4129 		bucket = &cache->uc_freebucket;
4130 #ifdef NUMA
4131 		if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4132 		    PCPU_GET(domain) != itemdomain) {
4133 			bucket = &cache->uc_crossbucket;
4134 		}
4135 #endif
4136 		if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4137 			cache_bucket_push(cache, bucket, item);
4138 			critical_exit();
4139 			return;
4140 		}
4141 	} while (cache_free(zone, cache, NULL, item, itemdomain));
4142 	critical_exit();
4143 
4144 	/*
4145 	 * If nothing else caught this, we'll just do an internal free.
4146 	 */
4147 	zone_free_item(zone, item, NULL, SKIP_NONE);
4148 }
4149 
4150 /* See uma.h */
4151 void
4152 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
4153 {
4154 	uma_cache_t cache;
4155 	uma_cache_bucket_t bucket;
4156 	int itemdomain, uz_flags;
4157 
4158 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
4159 	random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
4160 
4161 	CTR2(KTR_UMA, "uma_zfree_arg zone %s(%p)", zone->uz_name, zone);
4162 
4163 #ifdef UMA_ZALLOC_DEBUG
4164 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
4165 	    ("uma_zfree_arg: called with SMR zone."));
4166 	if (uma_zfree_debug(zone, item, udata) == EJUSTRETURN)
4167 		return;
4168 #endif
4169         /* uma_zfree(..., NULL) does nothing, to match free(9). */
4170         if (item == NULL)
4171                 return;
4172 
4173 	/*
4174 	 * We are accessing the per-cpu cache without a critical section to
4175 	 * fetch size and flags.  This is acceptable, if we are preempted we
4176 	 * will simply read another cpu's line.
4177 	 */
4178 	cache = &zone->uz_cpu[curcpu];
4179 	uz_flags = cache_uz_flags(cache);
4180 	if (UMA_ALWAYS_CTORDTOR ||
4181 	    __predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0))
4182 		item_dtor(zone, item, cache_uz_size(cache), udata, SKIP_NONE);
4183 
4184 	/*
4185 	 * The race here is acceptable.  If we miss it we'll just have to wait
4186 	 * a little longer for the limits to be reset.
4187 	 */
4188 	if (__predict_false(uz_flags & UMA_ZFLAG_LIMIT)) {
4189 		if (zone->uz_sleepers > 0)
4190 			goto zfree_item;
4191 	}
4192 
4193 	/*
4194 	 * If possible, free to the per-CPU cache.  There are two
4195 	 * requirements for safe access to the per-CPU cache: (1) the thread
4196 	 * accessing the cache must not be preempted or yield during access,
4197 	 * and (2) the thread must not migrate CPUs without switching which
4198 	 * cache it accesses.  We rely on a critical section to prevent
4199 	 * preemption and migration.  We release the critical section in
4200 	 * order to acquire the zone mutex if we are unable to free to the
4201 	 * current cache; when we re-acquire the critical section, we must
4202 	 * detect and handle migration if it has occurred.
4203 	 */
4204 	itemdomain = 0;
4205 #ifdef NUMA
4206 	if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4207 		itemdomain = item_domain(item);
4208 #endif
4209 	critical_enter();
4210 	do {
4211 		cache = &zone->uz_cpu[curcpu];
4212 		/*
4213 		 * Try to free into the allocbucket first to give LIFO
4214 		 * ordering for cache-hot datastructures.  Spill over
4215 		 * into the freebucket if necessary.  Alloc will swap
4216 		 * them if one runs dry.
4217 		 */
4218 		bucket = &cache->uc_allocbucket;
4219 #ifdef NUMA
4220 		if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4221 		    PCPU_GET(domain) != itemdomain) {
4222 			bucket = &cache->uc_crossbucket;
4223 		} else
4224 #endif
4225 		if (bucket->ucb_cnt == bucket->ucb_entries &&
4226 		   cache->uc_freebucket.ucb_cnt <
4227 		   cache->uc_freebucket.ucb_entries)
4228 			cache_bucket_swap(&cache->uc_freebucket,
4229 			    &cache->uc_allocbucket);
4230 		if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4231 			cache_bucket_push(cache, bucket, item);
4232 			critical_exit();
4233 			return;
4234 		}
4235 	} while (cache_free(zone, cache, udata, item, itemdomain));
4236 	critical_exit();
4237 
4238 	/*
4239 	 * If nothing else caught this, we'll just do an internal free.
4240 	 */
4241 zfree_item:
4242 	zone_free_item(zone, item, udata, SKIP_DTOR);
4243 }
4244 
4245 #ifdef NUMA
4246 /*
4247  * sort crossdomain free buckets to domain correct buckets and cache
4248  * them.
4249  */
4250 static void
4251 zone_free_cross(uma_zone_t zone, uma_bucket_t bucket, void *udata)
4252 {
4253 	struct uma_bucketlist fullbuckets;
4254 	uma_zone_domain_t zdom;
4255 	uma_bucket_t b;
4256 	smr_seq_t seq;
4257 	void *item;
4258 	int domain;
4259 
4260 	CTR3(KTR_UMA,
4261 	    "uma_zfree: zone %s(%p) draining cross bucket %p",
4262 	    zone->uz_name, zone, bucket);
4263 
4264 	/*
4265 	 * It is possible for buckets to arrive here out of order so we fetch
4266 	 * the current smr seq rather than accepting the bucket's.
4267 	 */
4268 	seq = SMR_SEQ_INVALID;
4269 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
4270 		seq = smr_advance(zone->uz_smr);
4271 
4272 	/*
4273 	 * To avoid having ndomain * ndomain buckets for sorting we have a
4274 	 * lock on the current crossfree bucket.  A full matrix with
4275 	 * per-domain locking could be used if necessary.
4276 	 */
4277 	STAILQ_INIT(&fullbuckets);
4278 	ZONE_CROSS_LOCK(zone);
4279 	while (bucket->ub_cnt > 0) {
4280 		item = bucket->ub_bucket[bucket->ub_cnt - 1];
4281 		domain = item_domain(item);
4282 		zdom = ZDOM_GET(zone, domain);
4283 		if (zdom->uzd_cross == NULL) {
4284 			zdom->uzd_cross = bucket_alloc(zone, udata, M_NOWAIT);
4285 			if (zdom->uzd_cross == NULL)
4286 				break;
4287 		}
4288 		b = zdom->uzd_cross;
4289 		b->ub_bucket[b->ub_cnt++] = item;
4290 		b->ub_seq = seq;
4291 		if (b->ub_cnt == b->ub_entries) {
4292 			STAILQ_INSERT_HEAD(&fullbuckets, b, ub_link);
4293 			zdom->uzd_cross = NULL;
4294 		}
4295 		bucket->ub_cnt--;
4296 	}
4297 	ZONE_CROSS_UNLOCK(zone);
4298 	if (bucket->ub_cnt == 0)
4299 		bucket->ub_seq = SMR_SEQ_INVALID;
4300 	bucket_free(zone, bucket, udata);
4301 
4302 	while ((b = STAILQ_FIRST(&fullbuckets)) != NULL) {
4303 		STAILQ_REMOVE_HEAD(&fullbuckets, ub_link);
4304 		domain = item_domain(b->ub_bucket[0]);
4305 		zone_put_bucket(zone, domain, b, udata, true);
4306 	}
4307 }
4308 #endif
4309 
4310 static void
4311 zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata,
4312     int itemdomain, bool ws)
4313 {
4314 
4315 #ifdef NUMA
4316 	/*
4317 	 * Buckets coming from the wrong domain will be entirely for the
4318 	 * only other domain on two domain systems.  In this case we can
4319 	 * simply cache them.  Otherwise we need to sort them back to
4320 	 * correct domains.
4321 	 */
4322 	if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4323 	    vm_ndomains > 2 && PCPU_GET(domain) != itemdomain) {
4324 		zone_free_cross(zone, bucket, udata);
4325 		return;
4326 	}
4327 #endif
4328 
4329 	/*
4330 	 * Attempt to save the bucket in the zone's domain bucket cache.
4331 	 */
4332 	CTR3(KTR_UMA,
4333 	    "uma_zfree: zone %s(%p) putting bucket %p on free list",
4334 	    zone->uz_name, zone, bucket);
4335 	/* ub_cnt is pointing to the last free item */
4336 	if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
4337 		itemdomain = zone_domain_lowest(zone, itemdomain);
4338 	zone_put_bucket(zone, itemdomain, bucket, udata, ws);
4339 }
4340 
4341 /*
4342  * Populate a free or cross bucket for the current cpu cache.  Free any
4343  * existing full bucket either to the zone cache or back to the slab layer.
4344  *
4345  * Enters and returns in a critical section.  false return indicates that
4346  * we can not satisfy this free in the cache layer.  true indicates that
4347  * the caller should retry.
4348  */
4349 static __noinline bool
4350 cache_free(uma_zone_t zone, uma_cache_t cache, void *udata, void *item,
4351     int itemdomain)
4352 {
4353 	uma_cache_bucket_t cbucket;
4354 	uma_bucket_t newbucket, bucket;
4355 
4356 	CRITICAL_ASSERT(curthread);
4357 
4358 	if (zone->uz_bucket_size == 0)
4359 		return false;
4360 
4361 	cache = &zone->uz_cpu[curcpu];
4362 	newbucket = NULL;
4363 
4364 	/*
4365 	 * FIRSTTOUCH domains need to free to the correct zdom.  When
4366 	 * enabled this is the zdom of the item.   The bucket is the
4367 	 * cross bucket if the current domain and itemdomain do not match.
4368 	 */
4369 	cbucket = &cache->uc_freebucket;
4370 #ifdef NUMA
4371 	if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4372 		if (PCPU_GET(domain) != itemdomain) {
4373 			cbucket = &cache->uc_crossbucket;
4374 			if (cbucket->ucb_cnt != 0)
4375 				counter_u64_add(zone->uz_xdomain,
4376 				    cbucket->ucb_cnt);
4377 		}
4378 	}
4379 #endif
4380 	bucket = cache_bucket_unload(cbucket);
4381 	KASSERT(bucket == NULL || bucket->ub_cnt == bucket->ub_entries,
4382 	    ("cache_free: Entered with non-full free bucket."));
4383 
4384 	/* We are no longer associated with this CPU. */
4385 	critical_exit();
4386 
4387 	/*
4388 	 * Don't let SMR zones operate without a free bucket.  Force
4389 	 * a synchronize and re-use this one.  We will only degrade
4390 	 * to a synchronize every bucket_size items rather than every
4391 	 * item if we fail to allocate a bucket.
4392 	 */
4393 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0) {
4394 		if (bucket != NULL)
4395 			bucket->ub_seq = smr_advance(zone->uz_smr);
4396 		newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4397 		if (newbucket == NULL && bucket != NULL) {
4398 			bucket_drain(zone, bucket);
4399 			newbucket = bucket;
4400 			bucket = NULL;
4401 		}
4402 	} else if (!bucketdisable)
4403 		newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4404 
4405 	if (bucket != NULL)
4406 		zone_free_bucket(zone, bucket, udata, itemdomain, true);
4407 
4408 	critical_enter();
4409 	if ((bucket = newbucket) == NULL)
4410 		return (false);
4411 	cache = &zone->uz_cpu[curcpu];
4412 #ifdef NUMA
4413 	/*
4414 	 * Check to see if we should be populating the cross bucket.  If it
4415 	 * is already populated we will fall through and attempt to populate
4416 	 * the free bucket.
4417 	 */
4418 	if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4419 		if (PCPU_GET(domain) != itemdomain &&
4420 		    cache->uc_crossbucket.ucb_bucket == NULL) {
4421 			cache_bucket_load_cross(cache, bucket);
4422 			return (true);
4423 		}
4424 	}
4425 #endif
4426 	/*
4427 	 * We may have lost the race to fill the bucket or switched CPUs.
4428 	 */
4429 	if (cache->uc_freebucket.ucb_bucket != NULL) {
4430 		critical_exit();
4431 		bucket_free(zone, bucket, udata);
4432 		critical_enter();
4433 	} else
4434 		cache_bucket_load_free(cache, bucket);
4435 
4436 	return (true);
4437 }
4438 
4439 static void
4440 slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item)
4441 {
4442 	uma_keg_t keg;
4443 	uma_domain_t dom;
4444 	int freei;
4445 
4446 	keg = zone->uz_keg;
4447 	KEG_LOCK_ASSERT(keg, slab->us_domain);
4448 
4449 	/* Do we need to remove from any lists? */
4450 	dom = &keg->uk_domain[slab->us_domain];
4451 	if (slab->us_freecount + 1 == keg->uk_ipers) {
4452 		LIST_REMOVE(slab, us_link);
4453 		LIST_INSERT_HEAD(&dom->ud_free_slab, slab, us_link);
4454 		dom->ud_free_slabs++;
4455 	} else if (slab->us_freecount == 0) {
4456 		LIST_REMOVE(slab, us_link);
4457 		LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
4458 	}
4459 
4460 	/* Slab management. */
4461 	freei = slab_item_index(slab, keg, item);
4462 	BIT_SET(keg->uk_ipers, freei, &slab->us_free);
4463 	slab->us_freecount++;
4464 
4465 	/* Keg statistics. */
4466 	dom->ud_free_items++;
4467 }
4468 
4469 static void
4470 zone_release(void *arg, void **bucket, int cnt)
4471 {
4472 	struct mtx *lock;
4473 	uma_zone_t zone;
4474 	uma_slab_t slab;
4475 	uma_keg_t keg;
4476 	uint8_t *mem;
4477 	void *item;
4478 	int i;
4479 
4480 	zone = arg;
4481 	keg = zone->uz_keg;
4482 	lock = NULL;
4483 	if (__predict_false((zone->uz_flags & UMA_ZFLAG_HASH) != 0))
4484 		lock = KEG_LOCK(keg, 0);
4485 	for (i = 0; i < cnt; i++) {
4486 		item = bucket[i];
4487 		if (__predict_true((zone->uz_flags & UMA_ZFLAG_VTOSLAB) != 0)) {
4488 			slab = vtoslab((vm_offset_t)item);
4489 		} else {
4490 			mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
4491 			if ((zone->uz_flags & UMA_ZFLAG_HASH) != 0)
4492 				slab = hash_sfind(&keg->uk_hash, mem);
4493 			else
4494 				slab = (uma_slab_t)(mem + keg->uk_pgoff);
4495 		}
4496 		if (lock != KEG_LOCKPTR(keg, slab->us_domain)) {
4497 			if (lock != NULL)
4498 				mtx_unlock(lock);
4499 			lock = KEG_LOCK(keg, slab->us_domain);
4500 		}
4501 		slab_free_item(zone, slab, item);
4502 	}
4503 	if (lock != NULL)
4504 		mtx_unlock(lock);
4505 }
4506 
4507 /*
4508  * Frees a single item to any zone.
4509  *
4510  * Arguments:
4511  *	zone   The zone to free to
4512  *	item   The item we're freeing
4513  *	udata  User supplied data for the dtor
4514  *	skip   Skip dtors and finis
4515  */
4516 static __noinline void
4517 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
4518 {
4519 
4520 	/*
4521 	 * If a free is sent directly to an SMR zone we have to
4522 	 * synchronize immediately because the item can instantly
4523 	 * be reallocated. This should only happen in degenerate
4524 	 * cases when no memory is available for per-cpu caches.
4525 	 */
4526 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0 && skip == SKIP_NONE)
4527 		smr_synchronize(zone->uz_smr);
4528 
4529 	item_dtor(zone, item, zone->uz_size, udata, skip);
4530 
4531 	if (skip < SKIP_FINI && zone->uz_fini)
4532 		zone->uz_fini(item, zone->uz_size);
4533 
4534 	zone->uz_release(zone->uz_arg, &item, 1);
4535 
4536 	if (skip & SKIP_CNT)
4537 		return;
4538 
4539 	counter_u64_add(zone->uz_frees, 1);
4540 
4541 	if (zone->uz_max_items > 0)
4542 		zone_free_limit(zone, 1);
4543 }
4544 
4545 /* See uma.h */
4546 int
4547 uma_zone_set_max(uma_zone_t zone, int nitems)
4548 {
4549 	struct uma_bucket_zone *ubz;
4550 	int count;
4551 
4552 	/*
4553 	 * XXX This can misbehave if the zone has any allocations with
4554 	 * no limit and a limit is imposed.  There is currently no
4555 	 * way to clear a limit.
4556 	 */
4557 	ZONE_LOCK(zone);
4558 	ubz = bucket_zone_max(zone, nitems);
4559 	count = ubz != NULL ? ubz->ubz_entries : 0;
4560 	zone->uz_bucket_size_max = zone->uz_bucket_size = count;
4561 	if (zone->uz_bucket_size_min > zone->uz_bucket_size_max)
4562 		zone->uz_bucket_size_min = zone->uz_bucket_size_max;
4563 	zone->uz_max_items = nitems;
4564 	zone->uz_flags |= UMA_ZFLAG_LIMIT;
4565 	zone_update_caches(zone);
4566 	/* We may need to wake waiters. */
4567 	wakeup(&zone->uz_max_items);
4568 	ZONE_UNLOCK(zone);
4569 
4570 	return (nitems);
4571 }
4572 
4573 /* See uma.h */
4574 void
4575 uma_zone_set_maxcache(uma_zone_t zone, int nitems)
4576 {
4577 	struct uma_bucket_zone *ubz;
4578 	int bpcpu;
4579 
4580 	ZONE_LOCK(zone);
4581 	ubz = bucket_zone_max(zone, nitems);
4582 	if (ubz != NULL) {
4583 		bpcpu = 2;
4584 		if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4585 			/* Count the cross-domain bucket. */
4586 			bpcpu++;
4587 		nitems -= ubz->ubz_entries * bpcpu * mp_ncpus;
4588 		zone->uz_bucket_size_max = ubz->ubz_entries;
4589 	} else {
4590 		zone->uz_bucket_size_max = zone->uz_bucket_size = 0;
4591 	}
4592 	if (zone->uz_bucket_size_min > zone->uz_bucket_size_max)
4593 		zone->uz_bucket_size_min = zone->uz_bucket_size_max;
4594 	zone->uz_bucket_max = nitems / vm_ndomains;
4595 	ZONE_UNLOCK(zone);
4596 }
4597 
4598 /* See uma.h */
4599 int
4600 uma_zone_get_max(uma_zone_t zone)
4601 {
4602 	int nitems;
4603 
4604 	nitems = atomic_load_64(&zone->uz_max_items);
4605 
4606 	return (nitems);
4607 }
4608 
4609 /* See uma.h */
4610 void
4611 uma_zone_set_warning(uma_zone_t zone, const char *warning)
4612 {
4613 
4614 	ZONE_ASSERT_COLD(zone);
4615 	zone->uz_warning = warning;
4616 }
4617 
4618 /* See uma.h */
4619 void
4620 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction)
4621 {
4622 
4623 	ZONE_ASSERT_COLD(zone);
4624 	TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone);
4625 }
4626 
4627 /* See uma.h */
4628 int
4629 uma_zone_get_cur(uma_zone_t zone)
4630 {
4631 	int64_t nitems;
4632 	u_int i;
4633 
4634 	nitems = 0;
4635 	if (zone->uz_allocs != EARLY_COUNTER && zone->uz_frees != EARLY_COUNTER)
4636 		nitems = counter_u64_fetch(zone->uz_allocs) -
4637 		    counter_u64_fetch(zone->uz_frees);
4638 	CPU_FOREACH(i)
4639 		nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs) -
4640 		    atomic_load_64(&zone->uz_cpu[i].uc_frees);
4641 
4642 	return (nitems < 0 ? 0 : nitems);
4643 }
4644 
4645 static uint64_t
4646 uma_zone_get_allocs(uma_zone_t zone)
4647 {
4648 	uint64_t nitems;
4649 	u_int i;
4650 
4651 	nitems = 0;
4652 	if (zone->uz_allocs != EARLY_COUNTER)
4653 		nitems = counter_u64_fetch(zone->uz_allocs);
4654 	CPU_FOREACH(i)
4655 		nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs);
4656 
4657 	return (nitems);
4658 }
4659 
4660 static uint64_t
4661 uma_zone_get_frees(uma_zone_t zone)
4662 {
4663 	uint64_t nitems;
4664 	u_int i;
4665 
4666 	nitems = 0;
4667 	if (zone->uz_frees != EARLY_COUNTER)
4668 		nitems = counter_u64_fetch(zone->uz_frees);
4669 	CPU_FOREACH(i)
4670 		nitems += atomic_load_64(&zone->uz_cpu[i].uc_frees);
4671 
4672 	return (nitems);
4673 }
4674 
4675 #ifdef INVARIANTS
4676 /* Used only for KEG_ASSERT_COLD(). */
4677 static uint64_t
4678 uma_keg_get_allocs(uma_keg_t keg)
4679 {
4680 	uma_zone_t z;
4681 	uint64_t nitems;
4682 
4683 	nitems = 0;
4684 	LIST_FOREACH(z, &keg->uk_zones, uz_link)
4685 		nitems += uma_zone_get_allocs(z);
4686 
4687 	return (nitems);
4688 }
4689 #endif
4690 
4691 /* See uma.h */
4692 void
4693 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
4694 {
4695 	uma_keg_t keg;
4696 
4697 	KEG_GET(zone, keg);
4698 	KEG_ASSERT_COLD(keg);
4699 	keg->uk_init = uminit;
4700 }
4701 
4702 /* See uma.h */
4703 void
4704 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
4705 {
4706 	uma_keg_t keg;
4707 
4708 	KEG_GET(zone, keg);
4709 	KEG_ASSERT_COLD(keg);
4710 	keg->uk_fini = fini;
4711 }
4712 
4713 /* See uma.h */
4714 void
4715 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
4716 {
4717 
4718 	ZONE_ASSERT_COLD(zone);
4719 	zone->uz_init = zinit;
4720 }
4721 
4722 /* See uma.h */
4723 void
4724 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
4725 {
4726 
4727 	ZONE_ASSERT_COLD(zone);
4728 	zone->uz_fini = zfini;
4729 }
4730 
4731 /* See uma.h */
4732 void
4733 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
4734 {
4735 	uma_keg_t keg;
4736 
4737 	KEG_GET(zone, keg);
4738 	KEG_ASSERT_COLD(keg);
4739 	keg->uk_freef = freef;
4740 }
4741 
4742 /* See uma.h */
4743 void
4744 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
4745 {
4746 	uma_keg_t keg;
4747 
4748 	KEG_GET(zone, keg);
4749 	KEG_ASSERT_COLD(keg);
4750 	keg->uk_allocf = allocf;
4751 }
4752 
4753 /* See uma.h */
4754 void
4755 uma_zone_set_smr(uma_zone_t zone, smr_t smr)
4756 {
4757 
4758 	ZONE_ASSERT_COLD(zone);
4759 
4760 	KASSERT(smr != NULL, ("Got NULL smr"));
4761 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
4762 	    ("zone %p (%s) already uses SMR", zone, zone->uz_name));
4763 	zone->uz_flags |= UMA_ZONE_SMR;
4764 	zone->uz_smr = smr;
4765 	zone_update_caches(zone);
4766 }
4767 
4768 smr_t
4769 uma_zone_get_smr(uma_zone_t zone)
4770 {
4771 
4772 	return (zone->uz_smr);
4773 }
4774 
4775 /* See uma.h */
4776 void
4777 uma_zone_reserve(uma_zone_t zone, int items)
4778 {
4779 	uma_keg_t keg;
4780 
4781 	KEG_GET(zone, keg);
4782 	KEG_ASSERT_COLD(keg);
4783 	keg->uk_reserve = items;
4784 }
4785 
4786 /* See uma.h */
4787 int
4788 uma_zone_reserve_kva(uma_zone_t zone, int count)
4789 {
4790 	uma_keg_t keg;
4791 	vm_offset_t kva;
4792 	u_int pages;
4793 
4794 	KEG_GET(zone, keg);
4795 	KEG_ASSERT_COLD(keg);
4796 	ZONE_ASSERT_COLD(zone);
4797 
4798 	pages = howmany(count, keg->uk_ipers) * keg->uk_ppera;
4799 
4800 #ifdef UMA_MD_SMALL_ALLOC
4801 	if (keg->uk_ppera > 1) {
4802 #else
4803 	if (1) {
4804 #endif
4805 		kva = kva_alloc((vm_size_t)pages * PAGE_SIZE);
4806 		if (kva == 0)
4807 			return (0);
4808 	} else
4809 		kva = 0;
4810 
4811 	MPASS(keg->uk_kva == 0);
4812 	keg->uk_kva = kva;
4813 	keg->uk_offset = 0;
4814 	zone->uz_max_items = pages * keg->uk_ipers;
4815 #ifdef UMA_MD_SMALL_ALLOC
4816 	keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
4817 #else
4818 	keg->uk_allocf = noobj_alloc;
4819 #endif
4820 	keg->uk_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
4821 	zone->uz_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
4822 	zone_update_caches(zone);
4823 
4824 	return (1);
4825 }
4826 
4827 /* See uma.h */
4828 void
4829 uma_prealloc(uma_zone_t zone, int items)
4830 {
4831 	struct vm_domainset_iter di;
4832 	uma_domain_t dom;
4833 	uma_slab_t slab;
4834 	uma_keg_t keg;
4835 	int aflags, domain, slabs;
4836 
4837 	KEG_GET(zone, keg);
4838 	slabs = howmany(items, keg->uk_ipers);
4839 	while (slabs-- > 0) {
4840 		aflags = M_NOWAIT;
4841 		vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
4842 		    &aflags);
4843 		for (;;) {
4844 			slab = keg_alloc_slab(keg, zone, domain, M_WAITOK,
4845 			    aflags);
4846 			if (slab != NULL) {
4847 				dom = &keg->uk_domain[slab->us_domain];
4848 				/*
4849 				 * keg_alloc_slab() always returns a slab on the
4850 				 * partial list.
4851 				 */
4852 				LIST_REMOVE(slab, us_link);
4853 				LIST_INSERT_HEAD(&dom->ud_free_slab, slab,
4854 				    us_link);
4855 				dom->ud_free_slabs++;
4856 				KEG_UNLOCK(keg, slab->us_domain);
4857 				break;
4858 			}
4859 			if (vm_domainset_iter_policy(&di, &domain) != 0)
4860 				vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask, 0);
4861 		}
4862 	}
4863 }
4864 
4865 /*
4866  * Returns a snapshot of memory consumption in bytes.
4867  */
4868 size_t
4869 uma_zone_memory(uma_zone_t zone)
4870 {
4871 	size_t sz;
4872 	int i;
4873 
4874 	sz = 0;
4875 	if (zone->uz_flags & UMA_ZFLAG_CACHE) {
4876 		for (i = 0; i < vm_ndomains; i++)
4877 			sz += ZDOM_GET(zone, i)->uzd_nitems;
4878 		return (sz * zone->uz_size);
4879 	}
4880 	for (i = 0; i < vm_ndomains; i++)
4881 		sz += zone->uz_keg->uk_domain[i].ud_pages;
4882 
4883 	return (sz * PAGE_SIZE);
4884 }
4885 
4886 /* See uma.h */
4887 void
4888 uma_reclaim(int req)
4889 {
4890 
4891 	CTR0(KTR_UMA, "UMA: vm asked us to release pages!");
4892 	sx_xlock(&uma_reclaim_lock);
4893 	bucket_enable();
4894 
4895 	switch (req) {
4896 	case UMA_RECLAIM_TRIM:
4897 		zone_foreach(zone_trim, NULL);
4898 		break;
4899 	case UMA_RECLAIM_DRAIN:
4900 	case UMA_RECLAIM_DRAIN_CPU:
4901 		zone_foreach(zone_drain, NULL);
4902 		if (req == UMA_RECLAIM_DRAIN_CPU) {
4903 			pcpu_cache_drain_safe(NULL);
4904 			zone_foreach(zone_drain, NULL);
4905 		}
4906 		break;
4907 	default:
4908 		panic("unhandled reclamation request %d", req);
4909 	}
4910 
4911 	/*
4912 	 * Some slabs may have been freed but this zone will be visited early
4913 	 * we visit again so that we can free pages that are empty once other
4914 	 * zones are drained.  We have to do the same for buckets.
4915 	 */
4916 	zone_drain(slabzones[0], NULL);
4917 	zone_drain(slabzones[1], NULL);
4918 	bucket_zone_drain();
4919 	sx_xunlock(&uma_reclaim_lock);
4920 }
4921 
4922 static volatile int uma_reclaim_needed;
4923 
4924 void
4925 uma_reclaim_wakeup(void)
4926 {
4927 
4928 	if (atomic_fetchadd_int(&uma_reclaim_needed, 1) == 0)
4929 		wakeup(uma_reclaim);
4930 }
4931 
4932 void
4933 uma_reclaim_worker(void *arg __unused)
4934 {
4935 
4936 	for (;;) {
4937 		sx_xlock(&uma_reclaim_lock);
4938 		while (atomic_load_int(&uma_reclaim_needed) == 0)
4939 			sx_sleep(uma_reclaim, &uma_reclaim_lock, PVM, "umarcl",
4940 			    hz);
4941 		sx_xunlock(&uma_reclaim_lock);
4942 		EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM);
4943 		uma_reclaim(UMA_RECLAIM_DRAIN_CPU);
4944 		atomic_store_int(&uma_reclaim_needed, 0);
4945 		/* Don't fire more than once per-second. */
4946 		pause("umarclslp", hz);
4947 	}
4948 }
4949 
4950 /* See uma.h */
4951 void
4952 uma_zone_reclaim(uma_zone_t zone, int req)
4953 {
4954 
4955 	switch (req) {
4956 	case UMA_RECLAIM_TRIM:
4957 		zone_trim(zone, NULL);
4958 		break;
4959 	case UMA_RECLAIM_DRAIN:
4960 		zone_drain(zone, NULL);
4961 		break;
4962 	case UMA_RECLAIM_DRAIN_CPU:
4963 		pcpu_cache_drain_safe(zone);
4964 		zone_drain(zone, NULL);
4965 		break;
4966 	default:
4967 		panic("unhandled reclamation request %d", req);
4968 	}
4969 }
4970 
4971 /* See uma.h */
4972 int
4973 uma_zone_exhausted(uma_zone_t zone)
4974 {
4975 
4976 	return (atomic_load_32(&zone->uz_sleepers) > 0);
4977 }
4978 
4979 unsigned long
4980 uma_limit(void)
4981 {
4982 
4983 	return (uma_kmem_limit);
4984 }
4985 
4986 void
4987 uma_set_limit(unsigned long limit)
4988 {
4989 
4990 	uma_kmem_limit = limit;
4991 }
4992 
4993 unsigned long
4994 uma_size(void)
4995 {
4996 
4997 	return (atomic_load_long(&uma_kmem_total));
4998 }
4999 
5000 long
5001 uma_avail(void)
5002 {
5003 
5004 	return (uma_kmem_limit - uma_size());
5005 }
5006 
5007 #ifdef DDB
5008 /*
5009  * Generate statistics across both the zone and its per-cpu cache's.  Return
5010  * desired statistics if the pointer is non-NULL for that statistic.
5011  *
5012  * Note: does not update the zone statistics, as it can't safely clear the
5013  * per-CPU cache statistic.
5014  *
5015  */
5016 static void
5017 uma_zone_sumstat(uma_zone_t z, long *cachefreep, uint64_t *allocsp,
5018     uint64_t *freesp, uint64_t *sleepsp, uint64_t *xdomainp)
5019 {
5020 	uma_cache_t cache;
5021 	uint64_t allocs, frees, sleeps, xdomain;
5022 	int cachefree, cpu;
5023 
5024 	allocs = frees = sleeps = xdomain = 0;
5025 	cachefree = 0;
5026 	CPU_FOREACH(cpu) {
5027 		cache = &z->uz_cpu[cpu];
5028 		cachefree += cache->uc_allocbucket.ucb_cnt;
5029 		cachefree += cache->uc_freebucket.ucb_cnt;
5030 		xdomain += cache->uc_crossbucket.ucb_cnt;
5031 		cachefree += cache->uc_crossbucket.ucb_cnt;
5032 		allocs += cache->uc_allocs;
5033 		frees += cache->uc_frees;
5034 	}
5035 	allocs += counter_u64_fetch(z->uz_allocs);
5036 	frees += counter_u64_fetch(z->uz_frees);
5037 	xdomain += counter_u64_fetch(z->uz_xdomain);
5038 	sleeps += z->uz_sleeps;
5039 	if (cachefreep != NULL)
5040 		*cachefreep = cachefree;
5041 	if (allocsp != NULL)
5042 		*allocsp = allocs;
5043 	if (freesp != NULL)
5044 		*freesp = frees;
5045 	if (sleepsp != NULL)
5046 		*sleepsp = sleeps;
5047 	if (xdomainp != NULL)
5048 		*xdomainp = xdomain;
5049 }
5050 #endif /* DDB */
5051 
5052 static int
5053 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
5054 {
5055 	uma_keg_t kz;
5056 	uma_zone_t z;
5057 	int count;
5058 
5059 	count = 0;
5060 	rw_rlock(&uma_rwlock);
5061 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
5062 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
5063 			count++;
5064 	}
5065 	LIST_FOREACH(z, &uma_cachezones, uz_link)
5066 		count++;
5067 
5068 	rw_runlock(&uma_rwlock);
5069 	return (sysctl_handle_int(oidp, &count, 0, req));
5070 }
5071 
5072 static void
5073 uma_vm_zone_stats(struct uma_type_header *uth, uma_zone_t z, struct sbuf *sbuf,
5074     struct uma_percpu_stat *ups, bool internal)
5075 {
5076 	uma_zone_domain_t zdom;
5077 	uma_cache_t cache;
5078 	int i;
5079 
5080 	for (i = 0; i < vm_ndomains; i++) {
5081 		zdom = ZDOM_GET(z, i);
5082 		uth->uth_zone_free += zdom->uzd_nitems;
5083 	}
5084 	uth->uth_allocs = counter_u64_fetch(z->uz_allocs);
5085 	uth->uth_frees = counter_u64_fetch(z->uz_frees);
5086 	uth->uth_fails = counter_u64_fetch(z->uz_fails);
5087 	uth->uth_xdomain = counter_u64_fetch(z->uz_xdomain);
5088 	uth->uth_sleeps = z->uz_sleeps;
5089 
5090 	for (i = 0; i < mp_maxid + 1; i++) {
5091 		bzero(&ups[i], sizeof(*ups));
5092 		if (internal || CPU_ABSENT(i))
5093 			continue;
5094 		cache = &z->uz_cpu[i];
5095 		ups[i].ups_cache_free += cache->uc_allocbucket.ucb_cnt;
5096 		ups[i].ups_cache_free += cache->uc_freebucket.ucb_cnt;
5097 		ups[i].ups_cache_free += cache->uc_crossbucket.ucb_cnt;
5098 		ups[i].ups_allocs = cache->uc_allocs;
5099 		ups[i].ups_frees = cache->uc_frees;
5100 	}
5101 }
5102 
5103 static int
5104 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
5105 {
5106 	struct uma_stream_header ush;
5107 	struct uma_type_header uth;
5108 	struct uma_percpu_stat *ups;
5109 	struct sbuf sbuf;
5110 	uma_keg_t kz;
5111 	uma_zone_t z;
5112 	uint64_t items;
5113 	uint32_t kfree, pages;
5114 	int count, error, i;
5115 
5116 	error = sysctl_wire_old_buffer(req, 0);
5117 	if (error != 0)
5118 		return (error);
5119 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
5120 	sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL);
5121 	ups = malloc((mp_maxid + 1) * sizeof(*ups), M_TEMP, M_WAITOK);
5122 
5123 	count = 0;
5124 	rw_rlock(&uma_rwlock);
5125 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
5126 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
5127 			count++;
5128 	}
5129 
5130 	LIST_FOREACH(z, &uma_cachezones, uz_link)
5131 		count++;
5132 
5133 	/*
5134 	 * Insert stream header.
5135 	 */
5136 	bzero(&ush, sizeof(ush));
5137 	ush.ush_version = UMA_STREAM_VERSION;
5138 	ush.ush_maxcpus = (mp_maxid + 1);
5139 	ush.ush_count = count;
5140 	(void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
5141 
5142 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
5143 		kfree = pages = 0;
5144 		for (i = 0; i < vm_ndomains; i++) {
5145 			kfree += kz->uk_domain[i].ud_free_items;
5146 			pages += kz->uk_domain[i].ud_pages;
5147 		}
5148 		LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5149 			bzero(&uth, sizeof(uth));
5150 			strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5151 			uth.uth_align = kz->uk_align;
5152 			uth.uth_size = kz->uk_size;
5153 			uth.uth_rsize = kz->uk_rsize;
5154 			if (z->uz_max_items > 0) {
5155 				items = UZ_ITEMS_COUNT(z->uz_items);
5156 				uth.uth_pages = (items / kz->uk_ipers) *
5157 					kz->uk_ppera;
5158 			} else
5159 				uth.uth_pages = pages;
5160 			uth.uth_maxpages = (z->uz_max_items / kz->uk_ipers) *
5161 			    kz->uk_ppera;
5162 			uth.uth_limit = z->uz_max_items;
5163 			uth.uth_keg_free = kfree;
5164 
5165 			/*
5166 			 * A zone is secondary is it is not the first entry
5167 			 * on the keg's zone list.
5168 			 */
5169 			if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
5170 			    (LIST_FIRST(&kz->uk_zones) != z))
5171 				uth.uth_zone_flags = UTH_ZONE_SECONDARY;
5172 			uma_vm_zone_stats(&uth, z, &sbuf, ups,
5173 			    kz->uk_flags & UMA_ZFLAG_INTERNAL);
5174 			(void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5175 			for (i = 0; i < mp_maxid + 1; i++)
5176 				(void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5177 		}
5178 	}
5179 	LIST_FOREACH(z, &uma_cachezones, uz_link) {
5180 		bzero(&uth, sizeof(uth));
5181 		strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5182 		uth.uth_size = z->uz_size;
5183 		uma_vm_zone_stats(&uth, z, &sbuf, ups, false);
5184 		(void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5185 		for (i = 0; i < mp_maxid + 1; i++)
5186 			(void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5187 	}
5188 
5189 	rw_runlock(&uma_rwlock);
5190 	error = sbuf_finish(&sbuf);
5191 	sbuf_delete(&sbuf);
5192 	free(ups, M_TEMP);
5193 	return (error);
5194 }
5195 
5196 int
5197 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS)
5198 {
5199 	uma_zone_t zone = *(uma_zone_t *)arg1;
5200 	int error, max;
5201 
5202 	max = uma_zone_get_max(zone);
5203 	error = sysctl_handle_int(oidp, &max, 0, req);
5204 	if (error || !req->newptr)
5205 		return (error);
5206 
5207 	uma_zone_set_max(zone, max);
5208 
5209 	return (0);
5210 }
5211 
5212 int
5213 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS)
5214 {
5215 	uma_zone_t zone;
5216 	int cur;
5217 
5218 	/*
5219 	 * Some callers want to add sysctls for global zones that
5220 	 * may not yet exist so they pass a pointer to a pointer.
5221 	 */
5222 	if (arg2 == 0)
5223 		zone = *(uma_zone_t *)arg1;
5224 	else
5225 		zone = arg1;
5226 	cur = uma_zone_get_cur(zone);
5227 	return (sysctl_handle_int(oidp, &cur, 0, req));
5228 }
5229 
5230 static int
5231 sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS)
5232 {
5233 	uma_zone_t zone = arg1;
5234 	uint64_t cur;
5235 
5236 	cur = uma_zone_get_allocs(zone);
5237 	return (sysctl_handle_64(oidp, &cur, 0, req));
5238 }
5239 
5240 static int
5241 sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS)
5242 {
5243 	uma_zone_t zone = arg1;
5244 	uint64_t cur;
5245 
5246 	cur = uma_zone_get_frees(zone);
5247 	return (sysctl_handle_64(oidp, &cur, 0, req));
5248 }
5249 
5250 static int
5251 sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS)
5252 {
5253 	struct sbuf sbuf;
5254 	uma_zone_t zone = arg1;
5255 	int error;
5256 
5257 	sbuf_new_for_sysctl(&sbuf, NULL, 0, req);
5258 	if (zone->uz_flags != 0)
5259 		sbuf_printf(&sbuf, "0x%b", zone->uz_flags, PRINT_UMA_ZFLAGS);
5260 	else
5261 		sbuf_printf(&sbuf, "0");
5262 	error = sbuf_finish(&sbuf);
5263 	sbuf_delete(&sbuf);
5264 
5265 	return (error);
5266 }
5267 
5268 static int
5269 sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS)
5270 {
5271 	uma_keg_t keg = arg1;
5272 	int avail, effpct, total;
5273 
5274 	total = keg->uk_ppera * PAGE_SIZE;
5275 	if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0)
5276 		total += slabzone(keg->uk_ipers)->uz_keg->uk_rsize;
5277 	/*
5278 	 * We consider the client's requested size and alignment here, not the
5279 	 * real size determination uk_rsize, because we also adjust the real
5280 	 * size for internal implementation reasons (max bitset size).
5281 	 */
5282 	avail = keg->uk_ipers * roundup2(keg->uk_size, keg->uk_align + 1);
5283 	if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
5284 		avail *= mp_maxid + 1;
5285 	effpct = 100 * avail / total;
5286 	return (sysctl_handle_int(oidp, &effpct, 0, req));
5287 }
5288 
5289 static int
5290 sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS)
5291 {
5292 	uma_zone_t zone = arg1;
5293 	uint64_t cur;
5294 
5295 	cur = UZ_ITEMS_COUNT(atomic_load_64(&zone->uz_items));
5296 	return (sysctl_handle_64(oidp, &cur, 0, req));
5297 }
5298 
5299 #ifdef INVARIANTS
5300 static uma_slab_t
5301 uma_dbg_getslab(uma_zone_t zone, void *item)
5302 {
5303 	uma_slab_t slab;
5304 	uma_keg_t keg;
5305 	uint8_t *mem;
5306 
5307 	/*
5308 	 * It is safe to return the slab here even though the
5309 	 * zone is unlocked because the item's allocation state
5310 	 * essentially holds a reference.
5311 	 */
5312 	mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
5313 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5314 		return (NULL);
5315 	if (zone->uz_flags & UMA_ZFLAG_VTOSLAB)
5316 		return (vtoslab((vm_offset_t)mem));
5317 	keg = zone->uz_keg;
5318 	if ((keg->uk_flags & UMA_ZFLAG_HASH) == 0)
5319 		return ((uma_slab_t)(mem + keg->uk_pgoff));
5320 	KEG_LOCK(keg, 0);
5321 	slab = hash_sfind(&keg->uk_hash, mem);
5322 	KEG_UNLOCK(keg, 0);
5323 
5324 	return (slab);
5325 }
5326 
5327 static bool
5328 uma_dbg_zskip(uma_zone_t zone, void *mem)
5329 {
5330 
5331 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5332 		return (true);
5333 
5334 	return (uma_dbg_kskip(zone->uz_keg, mem));
5335 }
5336 
5337 static bool
5338 uma_dbg_kskip(uma_keg_t keg, void *mem)
5339 {
5340 	uintptr_t idx;
5341 
5342 	if (dbg_divisor == 0)
5343 		return (true);
5344 
5345 	if (dbg_divisor == 1)
5346 		return (false);
5347 
5348 	idx = (uintptr_t)mem >> PAGE_SHIFT;
5349 	if (keg->uk_ipers > 1) {
5350 		idx *= keg->uk_ipers;
5351 		idx += ((uintptr_t)mem & PAGE_MASK) / keg->uk_rsize;
5352 	}
5353 
5354 	if ((idx / dbg_divisor) * dbg_divisor != idx) {
5355 		counter_u64_add(uma_skip_cnt, 1);
5356 		return (true);
5357 	}
5358 	counter_u64_add(uma_dbg_cnt, 1);
5359 
5360 	return (false);
5361 }
5362 
5363 /*
5364  * Set up the slab's freei data such that uma_dbg_free can function.
5365  *
5366  */
5367 static void
5368 uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item)
5369 {
5370 	uma_keg_t keg;
5371 	int freei;
5372 
5373 	if (slab == NULL) {
5374 		slab = uma_dbg_getslab(zone, item);
5375 		if (slab == NULL)
5376 			panic("uma: item %p did not belong to zone %s",
5377 			    item, zone->uz_name);
5378 	}
5379 	keg = zone->uz_keg;
5380 	freei = slab_item_index(slab, keg, item);
5381 
5382 	if (BIT_ISSET(keg->uk_ipers, freei, slab_dbg_bits(slab, keg)))
5383 		panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)",
5384 		    item, zone, zone->uz_name, slab, freei);
5385 	BIT_SET_ATOMIC(keg->uk_ipers, freei, slab_dbg_bits(slab, keg));
5386 }
5387 
5388 /*
5389  * Verifies freed addresses.  Checks for alignment, valid slab membership
5390  * and duplicate frees.
5391  *
5392  */
5393 static void
5394 uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item)
5395 {
5396 	uma_keg_t keg;
5397 	int freei;
5398 
5399 	if (slab == NULL) {
5400 		slab = uma_dbg_getslab(zone, item);
5401 		if (slab == NULL)
5402 			panic("uma: Freed item %p did not belong to zone %s",
5403 			    item, zone->uz_name);
5404 	}
5405 	keg = zone->uz_keg;
5406 	freei = slab_item_index(slab, keg, item);
5407 
5408 	if (freei >= keg->uk_ipers)
5409 		panic("Invalid free of %p from zone %p(%s) slab %p(%d)",
5410 		    item, zone, zone->uz_name, slab, freei);
5411 
5412 	if (slab_item(slab, keg, freei) != item)
5413 		panic("Unaligned free of %p from zone %p(%s) slab %p(%d)",
5414 		    item, zone, zone->uz_name, slab, freei);
5415 
5416 	if (!BIT_ISSET(keg->uk_ipers, freei, slab_dbg_bits(slab, keg)))
5417 		panic("Duplicate free of %p from zone %p(%s) slab %p(%d)",
5418 		    item, zone, zone->uz_name, slab, freei);
5419 
5420 	BIT_CLR_ATOMIC(keg->uk_ipers, freei, slab_dbg_bits(slab, keg));
5421 }
5422 #endif /* INVARIANTS */
5423 
5424 #ifdef DDB
5425 static int64_t
5426 get_uma_stats(uma_keg_t kz, uma_zone_t z, uint64_t *allocs, uint64_t *used,
5427     uint64_t *sleeps, long *cachefree, uint64_t *xdomain)
5428 {
5429 	uint64_t frees;
5430 	int i;
5431 
5432 	if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
5433 		*allocs = counter_u64_fetch(z->uz_allocs);
5434 		frees = counter_u64_fetch(z->uz_frees);
5435 		*sleeps = z->uz_sleeps;
5436 		*cachefree = 0;
5437 		*xdomain = 0;
5438 	} else
5439 		uma_zone_sumstat(z, cachefree, allocs, &frees, sleeps,
5440 		    xdomain);
5441 	for (i = 0; i < vm_ndomains; i++) {
5442 		*cachefree += ZDOM_GET(z, i)->uzd_nitems;
5443 		if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
5444 		    (LIST_FIRST(&kz->uk_zones) != z)))
5445 			*cachefree += kz->uk_domain[i].ud_free_items;
5446 	}
5447 	*used = *allocs - frees;
5448 	return (((int64_t)*used + *cachefree) * kz->uk_size);
5449 }
5450 
5451 DB_SHOW_COMMAND(uma, db_show_uma)
5452 {
5453 	const char *fmt_hdr, *fmt_entry;
5454 	uma_keg_t kz;
5455 	uma_zone_t z;
5456 	uint64_t allocs, used, sleeps, xdomain;
5457 	long cachefree;
5458 	/* variables for sorting */
5459 	uma_keg_t cur_keg;
5460 	uma_zone_t cur_zone, last_zone;
5461 	int64_t cur_size, last_size, size;
5462 	int ties;
5463 
5464 	/* /i option produces machine-parseable CSV output */
5465 	if (modif[0] == 'i') {
5466 		fmt_hdr = "%s,%s,%s,%s,%s,%s,%s,%s,%s\n";
5467 		fmt_entry = "\"%s\",%ju,%jd,%ld,%ju,%ju,%u,%jd,%ju\n";
5468 	} else {
5469 		fmt_hdr = "%18s %6s %7s %7s %11s %7s %7s %10s %8s\n";
5470 		fmt_entry = "%18s %6ju %7jd %7ld %11ju %7ju %7u %10jd %8ju\n";
5471 	}
5472 
5473 	db_printf(fmt_hdr, "Zone", "Size", "Used", "Free", "Requests",
5474 	    "Sleeps", "Bucket", "Total Mem", "XFree");
5475 
5476 	/* Sort the zones with largest size first. */
5477 	last_zone = NULL;
5478 	last_size = INT64_MAX;
5479 	for (;;) {
5480 		cur_zone = NULL;
5481 		cur_size = -1;
5482 		ties = 0;
5483 		LIST_FOREACH(kz, &uma_kegs, uk_link) {
5484 			LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5485 				/*
5486 				 * In the case of size ties, print out zones
5487 				 * in the order they are encountered.  That is,
5488 				 * when we encounter the most recently output
5489 				 * zone, we have already printed all preceding
5490 				 * ties, and we must print all following ties.
5491 				 */
5492 				if (z == last_zone) {
5493 					ties = 1;
5494 					continue;
5495 				}
5496 				size = get_uma_stats(kz, z, &allocs, &used,
5497 				    &sleeps, &cachefree, &xdomain);
5498 				if (size > cur_size && size < last_size + ties)
5499 				{
5500 					cur_size = size;
5501 					cur_zone = z;
5502 					cur_keg = kz;
5503 				}
5504 			}
5505 		}
5506 		if (cur_zone == NULL)
5507 			break;
5508 
5509 		size = get_uma_stats(cur_keg, cur_zone, &allocs, &used,
5510 		    &sleeps, &cachefree, &xdomain);
5511 		db_printf(fmt_entry, cur_zone->uz_name,
5512 		    (uintmax_t)cur_keg->uk_size, (intmax_t)used, cachefree,
5513 		    (uintmax_t)allocs, (uintmax_t)sleeps,
5514 		    (unsigned)cur_zone->uz_bucket_size, (intmax_t)size,
5515 		    xdomain);
5516 
5517 		if (db_pager_quit)
5518 			return;
5519 		last_zone = cur_zone;
5520 		last_size = cur_size;
5521 	}
5522 }
5523 
5524 DB_SHOW_COMMAND(umacache, db_show_umacache)
5525 {
5526 	uma_zone_t z;
5527 	uint64_t allocs, frees;
5528 	long cachefree;
5529 	int i;
5530 
5531 	db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
5532 	    "Requests", "Bucket");
5533 	LIST_FOREACH(z, &uma_cachezones, uz_link) {
5534 		uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL, NULL);
5535 		for (i = 0; i < vm_ndomains; i++)
5536 			cachefree += ZDOM_GET(z, i)->uzd_nitems;
5537 		db_printf("%18s %8ju %8jd %8ld %12ju %8u\n",
5538 		    z->uz_name, (uintmax_t)z->uz_size,
5539 		    (intmax_t)(allocs - frees), cachefree,
5540 		    (uintmax_t)allocs, z->uz_bucket_size);
5541 		if (db_pager_quit)
5542 			return;
5543 	}
5544 }
5545 #endif	/* DDB */
5546