xref: /freebsd/sys/vm/uma_core.c (revision b820822823cc5f76af5cc8be6d96483a31b36151)
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 /*
1419  * Frees pages from a keg back to the system.  This is done on demand from
1420  * the pageout daemon.
1421  *
1422  * Returns nothing.
1423  */
1424 static void
1425 keg_drain(uma_keg_t keg)
1426 {
1427 	struct slabhead freeslabs;
1428 	uma_domain_t dom;
1429 	uma_slab_t slab, tmp;
1430 	int i, n;
1431 
1432 	if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL)
1433 		return;
1434 
1435 	for (i = 0; i < vm_ndomains; i++) {
1436 		CTR4(KTR_UMA, "keg_drain %s(%p) domain %d free items: %u",
1437 		    keg->uk_name, keg, i, dom->ud_free_items);
1438 		dom = &keg->uk_domain[i];
1439 		LIST_INIT(&freeslabs);
1440 
1441 		KEG_LOCK(keg, i);
1442 		if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0) {
1443 			LIST_FOREACH(slab, &dom->ud_free_slab, us_link)
1444 				UMA_HASH_REMOVE(&keg->uk_hash, slab);
1445 		}
1446 		n = dom->ud_free_slabs;
1447 		LIST_SWAP(&freeslabs, &dom->ud_free_slab, uma_slab, us_link);
1448 		dom->ud_free_slabs = 0;
1449 		dom->ud_free_items -= n * keg->uk_ipers;
1450 		dom->ud_pages -= n * keg->uk_ppera;
1451 		KEG_UNLOCK(keg, i);
1452 
1453 		LIST_FOREACH_SAFE(slab, &freeslabs, us_link, tmp)
1454 			keg_free_slab(keg, slab, keg->uk_ipers);
1455 	}
1456 }
1457 
1458 static void
1459 zone_reclaim(uma_zone_t zone, int waitok, bool drain)
1460 {
1461 
1462 	/*
1463 	 * Set draining to interlock with zone_dtor() so we can release our
1464 	 * locks as we go.  Only dtor() should do a WAITOK call since it
1465 	 * is the only call that knows the structure will still be available
1466 	 * when it wakes up.
1467 	 */
1468 	ZONE_LOCK(zone);
1469 	while (zone->uz_flags & UMA_ZFLAG_RECLAIMING) {
1470 		if (waitok == M_NOWAIT)
1471 			goto out;
1472 		msleep(zone, &ZDOM_GET(zone, 0)->uzd_lock, PVM, "zonedrain",
1473 		    1);
1474 	}
1475 	zone->uz_flags |= UMA_ZFLAG_RECLAIMING;
1476 	ZONE_UNLOCK(zone);
1477 	bucket_cache_reclaim(zone, drain);
1478 
1479 	/*
1480 	 * The DRAINING flag protects us from being freed while
1481 	 * we're running.  Normally the uma_rwlock would protect us but we
1482 	 * must be able to release and acquire the right lock for each keg.
1483 	 */
1484 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0)
1485 		keg_drain(zone->uz_keg);
1486 	ZONE_LOCK(zone);
1487 	zone->uz_flags &= ~UMA_ZFLAG_RECLAIMING;
1488 	wakeup(zone);
1489 out:
1490 	ZONE_UNLOCK(zone);
1491 }
1492 
1493 static void
1494 zone_drain(uma_zone_t zone, void *unused)
1495 {
1496 
1497 	zone_reclaim(zone, M_NOWAIT, true);
1498 }
1499 
1500 static void
1501 zone_trim(uma_zone_t zone, void *unused)
1502 {
1503 
1504 	zone_reclaim(zone, M_NOWAIT, false);
1505 }
1506 
1507 /*
1508  * Allocate a new slab for a keg and inserts it into the partial slab list.
1509  * The keg should be unlocked on entry.  If the allocation succeeds it will
1510  * be locked on return.
1511  *
1512  * Arguments:
1513  *	flags   Wait flags for the item initialization routine
1514  *	aflags  Wait flags for the slab allocation
1515  *
1516  * Returns:
1517  *	The slab that was allocated or NULL if there is no memory and the
1518  *	caller specified M_NOWAIT.
1519  */
1520 static uma_slab_t
1521 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int domain, int flags,
1522     int aflags)
1523 {
1524 	uma_domain_t dom;
1525 	uma_alloc allocf;
1526 	uma_slab_t slab;
1527 	unsigned long size;
1528 	uint8_t *mem;
1529 	uint8_t sflags;
1530 	int i;
1531 
1532 	KASSERT(domain >= 0 && domain < vm_ndomains,
1533 	    ("keg_alloc_slab: domain %d out of range", domain));
1534 
1535 	allocf = keg->uk_allocf;
1536 	slab = NULL;
1537 	mem = NULL;
1538 	if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) {
1539 		uma_hash_slab_t hslab;
1540 		hslab = zone_alloc_item(slabzone(keg->uk_ipers), NULL,
1541 		    domain, aflags);
1542 		if (hslab == NULL)
1543 			goto fail;
1544 		slab = &hslab->uhs_slab;
1545 	}
1546 
1547 	/*
1548 	 * This reproduces the old vm_zone behavior of zero filling pages the
1549 	 * first time they are added to a zone.
1550 	 *
1551 	 * Malloced items are zeroed in uma_zalloc.
1552 	 */
1553 
1554 	if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
1555 		aflags |= M_ZERO;
1556 	else
1557 		aflags &= ~M_ZERO;
1558 
1559 	if (keg->uk_flags & UMA_ZONE_NODUMP)
1560 		aflags |= M_NODUMP;
1561 
1562 	/* zone is passed for legacy reasons. */
1563 	size = keg->uk_ppera * PAGE_SIZE;
1564 	mem = allocf(zone, size, domain, &sflags, aflags);
1565 	if (mem == NULL) {
1566 		if (keg->uk_flags & UMA_ZFLAG_OFFPAGE)
1567 			zone_free_item(slabzone(keg->uk_ipers),
1568 			    slab_tohashslab(slab), NULL, SKIP_NONE);
1569 		goto fail;
1570 	}
1571 	uma_total_inc(size);
1572 
1573 	/* For HASH zones all pages go to the same uma_domain. */
1574 	if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
1575 		domain = 0;
1576 
1577 	/* Point the slab into the allocated memory */
1578 	if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE))
1579 		slab = (uma_slab_t )(mem + keg->uk_pgoff);
1580 	else
1581 		slab_tohashslab(slab)->uhs_data = mem;
1582 
1583 	if (keg->uk_flags & UMA_ZFLAG_VTOSLAB)
1584 		for (i = 0; i < keg->uk_ppera; i++)
1585 			vsetzoneslab((vm_offset_t)mem + (i * PAGE_SIZE),
1586 			    zone, slab);
1587 
1588 	slab->us_freecount = keg->uk_ipers;
1589 	slab->us_flags = sflags;
1590 	slab->us_domain = domain;
1591 
1592 	BIT_FILL(keg->uk_ipers, &slab->us_free);
1593 #ifdef INVARIANTS
1594 	BIT_ZERO(keg->uk_ipers, slab_dbg_bits(slab, keg));
1595 #endif
1596 
1597 	if (keg->uk_init != NULL) {
1598 		for (i = 0; i < keg->uk_ipers; i++)
1599 			if (keg->uk_init(slab_item(slab, keg, i),
1600 			    keg->uk_size, flags) != 0)
1601 				break;
1602 		if (i != keg->uk_ipers) {
1603 			keg_free_slab(keg, slab, i);
1604 			goto fail;
1605 		}
1606 	}
1607 	KEG_LOCK(keg, domain);
1608 
1609 	CTR3(KTR_UMA, "keg_alloc_slab: allocated slab %p for %s(%p)",
1610 	    slab, keg->uk_name, keg);
1611 
1612 	if (keg->uk_flags & UMA_ZFLAG_HASH)
1613 		UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
1614 
1615 	/*
1616 	 * If we got a slab here it's safe to mark it partially used
1617 	 * and return.  We assume that the caller is going to remove
1618 	 * at least one item.
1619 	 */
1620 	dom = &keg->uk_domain[domain];
1621 	LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
1622 	dom->ud_pages += keg->uk_ppera;
1623 	dom->ud_free_items += keg->uk_ipers;
1624 
1625 	return (slab);
1626 
1627 fail:
1628 	return (NULL);
1629 }
1630 
1631 /*
1632  * This function is intended to be used early on in place of page_alloc() so
1633  * that we may use the boot time page cache to satisfy allocations before
1634  * the VM is ready.
1635  */
1636 static void *
1637 startup_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1638     int wait)
1639 {
1640 	vm_paddr_t pa;
1641 	vm_page_t m;
1642 	void *mem;
1643 	int pages;
1644 	int i;
1645 
1646 	pages = howmany(bytes, PAGE_SIZE);
1647 	KASSERT(pages > 0, ("%s can't reserve 0 pages", __func__));
1648 
1649 	*pflag = UMA_SLAB_BOOT;
1650 	m = vm_page_alloc_contig_domain(NULL, 0, domain,
1651 	    malloc2vm_flags(wait) | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED, pages,
1652 	    (vm_paddr_t)0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT);
1653 	if (m == NULL)
1654 		return (NULL);
1655 
1656 	pa = VM_PAGE_TO_PHYS(m);
1657 	for (i = 0; i < pages; i++, pa += PAGE_SIZE) {
1658 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
1659     defined(__riscv) || defined(__powerpc64__)
1660 		if ((wait & M_NODUMP) == 0)
1661 			dump_add_page(pa);
1662 #endif
1663 	}
1664 	/* Allocate KVA and indirectly advance bootmem. */
1665 	mem = (void *)pmap_map(&bootmem, m->phys_addr,
1666 	    m->phys_addr + (pages * PAGE_SIZE), VM_PROT_READ | VM_PROT_WRITE);
1667         if ((wait & M_ZERO) != 0)
1668                 bzero(mem, pages * PAGE_SIZE);
1669 
1670         return (mem);
1671 }
1672 
1673 static void
1674 startup_free(void *mem, vm_size_t bytes)
1675 {
1676 	vm_offset_t va;
1677 	vm_page_t m;
1678 
1679 	va = (vm_offset_t)mem;
1680 	m = PHYS_TO_VM_PAGE(pmap_kextract(va));
1681 	pmap_remove(kernel_pmap, va, va + bytes);
1682 	for (; bytes != 0; bytes -= PAGE_SIZE, m++) {
1683 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
1684     defined(__riscv) || defined(__powerpc64__)
1685 		dump_drop_page(VM_PAGE_TO_PHYS(m));
1686 #endif
1687 		vm_page_unwire_noq(m);
1688 		vm_page_free(m);
1689 	}
1690 }
1691 
1692 /*
1693  * Allocates a number of pages from the system
1694  *
1695  * Arguments:
1696  *	bytes  The number of bytes requested
1697  *	wait  Shall we wait?
1698  *
1699  * Returns:
1700  *	A pointer to the alloced memory or possibly
1701  *	NULL if M_NOWAIT is set.
1702  */
1703 static void *
1704 page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1705     int wait)
1706 {
1707 	void *p;	/* Returned page */
1708 
1709 	*pflag = UMA_SLAB_KERNEL;
1710 	p = (void *)kmem_malloc_domainset(DOMAINSET_FIXED(domain), bytes, wait);
1711 
1712 	return (p);
1713 }
1714 
1715 static void *
1716 pcpu_page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1717     int wait)
1718 {
1719 	struct pglist alloctail;
1720 	vm_offset_t addr, zkva;
1721 	int cpu, flags;
1722 	vm_page_t p, p_next;
1723 #ifdef NUMA
1724 	struct pcpu *pc;
1725 #endif
1726 
1727 	MPASS(bytes == (mp_maxid + 1) * PAGE_SIZE);
1728 
1729 	TAILQ_INIT(&alloctail);
1730 	flags = VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | VM_ALLOC_NOOBJ |
1731 	    malloc2vm_flags(wait);
1732 	*pflag = UMA_SLAB_KERNEL;
1733 	for (cpu = 0; cpu <= mp_maxid; cpu++) {
1734 		if (CPU_ABSENT(cpu)) {
1735 			p = vm_page_alloc(NULL, 0, flags);
1736 		} else {
1737 #ifndef NUMA
1738 			p = vm_page_alloc(NULL, 0, flags);
1739 #else
1740 			pc = pcpu_find(cpu);
1741 			if (__predict_false(VM_DOMAIN_EMPTY(pc->pc_domain)))
1742 				p = NULL;
1743 			else
1744 				p = vm_page_alloc_domain(NULL, 0,
1745 				    pc->pc_domain, flags);
1746 			if (__predict_false(p == NULL))
1747 				p = vm_page_alloc(NULL, 0, flags);
1748 #endif
1749 		}
1750 		if (__predict_false(p == NULL))
1751 			goto fail;
1752 		TAILQ_INSERT_TAIL(&alloctail, p, listq);
1753 	}
1754 	if ((addr = kva_alloc(bytes)) == 0)
1755 		goto fail;
1756 	zkva = addr;
1757 	TAILQ_FOREACH(p, &alloctail, listq) {
1758 		pmap_qenter(zkva, &p, 1);
1759 		zkva += PAGE_SIZE;
1760 	}
1761 	return ((void*)addr);
1762 fail:
1763 	TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1764 		vm_page_unwire_noq(p);
1765 		vm_page_free(p);
1766 	}
1767 	return (NULL);
1768 }
1769 
1770 /*
1771  * Allocates a number of pages from within an object
1772  *
1773  * Arguments:
1774  *	bytes  The number of bytes requested
1775  *	wait   Shall we wait?
1776  *
1777  * Returns:
1778  *	A pointer to the alloced memory or possibly
1779  *	NULL if M_NOWAIT is set.
1780  */
1781 static void *
1782 noobj_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags,
1783     int wait)
1784 {
1785 	TAILQ_HEAD(, vm_page) alloctail;
1786 	u_long npages;
1787 	vm_offset_t retkva, zkva;
1788 	vm_page_t p, p_next;
1789 	uma_keg_t keg;
1790 
1791 	TAILQ_INIT(&alloctail);
1792 	keg = zone->uz_keg;
1793 
1794 	npages = howmany(bytes, PAGE_SIZE);
1795 	while (npages > 0) {
1796 		p = vm_page_alloc_domain(NULL, 0, domain, VM_ALLOC_INTERRUPT |
1797 		    VM_ALLOC_WIRED | VM_ALLOC_NOOBJ |
1798 		    ((wait & M_WAITOK) != 0 ? VM_ALLOC_WAITOK :
1799 		    VM_ALLOC_NOWAIT));
1800 		if (p != NULL) {
1801 			/*
1802 			 * Since the page does not belong to an object, its
1803 			 * listq is unused.
1804 			 */
1805 			TAILQ_INSERT_TAIL(&alloctail, p, listq);
1806 			npages--;
1807 			continue;
1808 		}
1809 		/*
1810 		 * Page allocation failed, free intermediate pages and
1811 		 * exit.
1812 		 */
1813 		TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1814 			vm_page_unwire_noq(p);
1815 			vm_page_free(p);
1816 		}
1817 		return (NULL);
1818 	}
1819 	*flags = UMA_SLAB_PRIV;
1820 	zkva = keg->uk_kva +
1821 	    atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
1822 	retkva = zkva;
1823 	TAILQ_FOREACH(p, &alloctail, listq) {
1824 		pmap_qenter(zkva, &p, 1);
1825 		zkva += PAGE_SIZE;
1826 	}
1827 
1828 	return ((void *)retkva);
1829 }
1830 
1831 /*
1832  * Allocate physically contiguous pages.
1833  */
1834 static void *
1835 contig_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1836     int wait)
1837 {
1838 
1839 	*pflag = UMA_SLAB_KERNEL;
1840 	return ((void *)kmem_alloc_contig_domainset(DOMAINSET_FIXED(domain),
1841 	    bytes, wait, 0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT));
1842 }
1843 
1844 /*
1845  * Frees a number of pages to the system
1846  *
1847  * Arguments:
1848  *	mem   A pointer to the memory to be freed
1849  *	size  The size of the memory being freed
1850  *	flags The original p->us_flags field
1851  *
1852  * Returns:
1853  *	Nothing
1854  */
1855 static void
1856 page_free(void *mem, vm_size_t size, uint8_t flags)
1857 {
1858 
1859 	if ((flags & UMA_SLAB_BOOT) != 0) {
1860 		startup_free(mem, size);
1861 		return;
1862 	}
1863 
1864 	KASSERT((flags & UMA_SLAB_KERNEL) != 0,
1865 	    ("UMA: page_free used with invalid flags %x", flags));
1866 
1867 	kmem_free((vm_offset_t)mem, size);
1868 }
1869 
1870 /*
1871  * Frees pcpu zone allocations
1872  *
1873  * Arguments:
1874  *	mem   A pointer to the memory to be freed
1875  *	size  The size of the memory being freed
1876  *	flags The original p->us_flags field
1877  *
1878  * Returns:
1879  *	Nothing
1880  */
1881 static void
1882 pcpu_page_free(void *mem, vm_size_t size, uint8_t flags)
1883 {
1884 	vm_offset_t sva, curva;
1885 	vm_paddr_t paddr;
1886 	vm_page_t m;
1887 
1888 	MPASS(size == (mp_maxid+1)*PAGE_SIZE);
1889 
1890 	if ((flags & UMA_SLAB_BOOT) != 0) {
1891 		startup_free(mem, size);
1892 		return;
1893 	}
1894 
1895 	sva = (vm_offset_t)mem;
1896 	for (curva = sva; curva < sva + size; curva += PAGE_SIZE) {
1897 		paddr = pmap_kextract(curva);
1898 		m = PHYS_TO_VM_PAGE(paddr);
1899 		vm_page_unwire_noq(m);
1900 		vm_page_free(m);
1901 	}
1902 	pmap_qremove(sva, size >> PAGE_SHIFT);
1903 	kva_free(sva, size);
1904 }
1905 
1906 /*
1907  * Zero fill initializer
1908  *
1909  * Arguments/Returns follow uma_init specifications
1910  */
1911 static int
1912 zero_init(void *mem, int size, int flags)
1913 {
1914 	bzero(mem, size);
1915 	return (0);
1916 }
1917 
1918 #ifdef INVARIANTS
1919 static struct noslabbits *
1920 slab_dbg_bits(uma_slab_t slab, uma_keg_t keg)
1921 {
1922 
1923 	return ((void *)((char *)&slab->us_free + BITSET_SIZE(keg->uk_ipers)));
1924 }
1925 #endif
1926 
1927 /*
1928  * Actual size of embedded struct slab (!OFFPAGE).
1929  */
1930 static size_t
1931 slab_sizeof(int nitems)
1932 {
1933 	size_t s;
1934 
1935 	s = sizeof(struct uma_slab) + BITSET_SIZE(nitems) * SLAB_BITSETS;
1936 	return (roundup(s, UMA_ALIGN_PTR + 1));
1937 }
1938 
1939 #define	UMA_FIXPT_SHIFT	31
1940 #define	UMA_FRAC_FIXPT(n, d)						\
1941 	((uint32_t)(((uint64_t)(n) << UMA_FIXPT_SHIFT) / (d)))
1942 #define	UMA_FIXPT_PCT(f)						\
1943 	((u_int)(((uint64_t)100 * (f)) >> UMA_FIXPT_SHIFT))
1944 #define	UMA_PCT_FIXPT(pct)	UMA_FRAC_FIXPT((pct), 100)
1945 #define	UMA_MIN_EFF	UMA_PCT_FIXPT(100 - UMA_MAX_WASTE)
1946 
1947 /*
1948  * Compute the number of items that will fit in a slab.  If hdr is true, the
1949  * item count may be limited to provide space in the slab for an inline slab
1950  * header.  Otherwise, all slab space will be provided for item storage.
1951  */
1952 static u_int
1953 slab_ipers_hdr(u_int size, u_int rsize, u_int slabsize, bool hdr)
1954 {
1955 	u_int ipers;
1956 	u_int padpi;
1957 
1958 	/* The padding between items is not needed after the last item. */
1959 	padpi = rsize - size;
1960 
1961 	if (hdr) {
1962 		/*
1963 		 * Start with the maximum item count and remove items until
1964 		 * the slab header first alongside the allocatable memory.
1965 		 */
1966 		for (ipers = MIN(SLAB_MAX_SETSIZE,
1967 		    (slabsize + padpi - slab_sizeof(1)) / rsize);
1968 		    ipers > 0 &&
1969 		    ipers * rsize - padpi + slab_sizeof(ipers) > slabsize;
1970 		    ipers--)
1971 			continue;
1972 	} else {
1973 		ipers = MIN((slabsize + padpi) / rsize, SLAB_MAX_SETSIZE);
1974 	}
1975 
1976 	return (ipers);
1977 }
1978 
1979 struct keg_layout_result {
1980 	u_int format;
1981 	u_int slabsize;
1982 	u_int ipers;
1983 	u_int eff;
1984 };
1985 
1986 static void
1987 keg_layout_one(uma_keg_t keg, u_int rsize, u_int slabsize, u_int fmt,
1988     struct keg_layout_result *kl)
1989 {
1990 	u_int total;
1991 
1992 	kl->format = fmt;
1993 	kl->slabsize = slabsize;
1994 
1995 	/* Handle INTERNAL as inline with an extra page. */
1996 	if ((fmt & UMA_ZFLAG_INTERNAL) != 0) {
1997 		kl->format &= ~UMA_ZFLAG_INTERNAL;
1998 		kl->slabsize += PAGE_SIZE;
1999 	}
2000 
2001 	kl->ipers = slab_ipers_hdr(keg->uk_size, rsize, kl->slabsize,
2002 	    (fmt & UMA_ZFLAG_OFFPAGE) == 0);
2003 
2004 	/* Account for memory used by an offpage slab header. */
2005 	total = kl->slabsize;
2006 	if ((fmt & UMA_ZFLAG_OFFPAGE) != 0)
2007 		total += slabzone(kl->ipers)->uz_keg->uk_rsize;
2008 
2009 	kl->eff = UMA_FRAC_FIXPT(kl->ipers * rsize, total);
2010 }
2011 
2012 /*
2013  * Determine the format of a uma keg.  This determines where the slab header
2014  * will be placed (inline or offpage) and calculates ipers, rsize, and ppera.
2015  *
2016  * Arguments
2017  *	keg  The zone we should initialize
2018  *
2019  * Returns
2020  *	Nothing
2021  */
2022 static void
2023 keg_layout(uma_keg_t keg)
2024 {
2025 	struct keg_layout_result kl = {}, kl_tmp;
2026 	u_int fmts[2];
2027 	u_int alignsize;
2028 	u_int nfmt;
2029 	u_int pages;
2030 	u_int rsize;
2031 	u_int slabsize;
2032 	u_int i, j;
2033 
2034 	KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
2035 	    (keg->uk_size <= UMA_PCPU_ALLOC_SIZE &&
2036 	     (keg->uk_flags & UMA_ZONE_CACHESPREAD) == 0),
2037 	    ("%s: cannot configure for PCPU: keg=%s, size=%u, flags=0x%b",
2038 	     __func__, keg->uk_name, keg->uk_size, keg->uk_flags,
2039 	     PRINT_UMA_ZFLAGS));
2040 	KASSERT((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) == 0 ||
2041 	    (keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0,
2042 	    ("%s: incompatible flags 0x%b", __func__, keg->uk_flags,
2043 	     PRINT_UMA_ZFLAGS));
2044 
2045 	alignsize = keg->uk_align + 1;
2046 
2047 	/*
2048 	 * Calculate the size of each allocation (rsize) according to
2049 	 * alignment.  If the requested size is smaller than we have
2050 	 * allocation bits for we round it up.
2051 	 */
2052 	rsize = MAX(keg->uk_size, UMA_SMALLEST_UNIT);
2053 	rsize = roundup2(rsize, alignsize);
2054 
2055 	if ((keg->uk_flags & UMA_ZONE_CACHESPREAD) != 0) {
2056 		/*
2057 		 * We want one item to start on every align boundary in a page.
2058 		 * To do this we will span pages.  We will also extend the item
2059 		 * by the size of align if it is an even multiple of align.
2060 		 * Otherwise, it would fall on the same boundary every time.
2061 		 */
2062 		if ((rsize & alignsize) == 0)
2063 			rsize += alignsize;
2064 		slabsize = rsize * (PAGE_SIZE / alignsize);
2065 		slabsize = MIN(slabsize, rsize * SLAB_MAX_SETSIZE);
2066 		slabsize = MIN(slabsize, UMA_CACHESPREAD_MAX_SIZE);
2067 		slabsize = round_page(slabsize);
2068 	} else {
2069 		/*
2070 		 * Start with a slab size of as many pages as it takes to
2071 		 * represent a single item.  We will try to fit as many
2072 		 * additional items into the slab as possible.
2073 		 */
2074 		slabsize = round_page(keg->uk_size);
2075 	}
2076 
2077 	/* Build a list of all of the available formats for this keg. */
2078 	nfmt = 0;
2079 
2080 	/* Evaluate an inline slab layout. */
2081 	if ((keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0)
2082 		fmts[nfmt++] = 0;
2083 
2084 	/* TODO: vm_page-embedded slab. */
2085 
2086 	/*
2087 	 * We can't do OFFPAGE if we're internal or if we've been
2088 	 * asked to not go to the VM for buckets.  If we do this we
2089 	 * may end up going to the VM for slabs which we do not want
2090 	 * to do if we're UMA_ZONE_VM, which clearly forbids it.
2091 	 * In those cases, evaluate a pseudo-format called INTERNAL
2092 	 * which has an inline slab header and one extra page to
2093 	 * guarantee that it fits.
2094 	 *
2095 	 * Otherwise, see if using an OFFPAGE slab will improve our
2096 	 * efficiency.
2097 	 */
2098 	if ((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) != 0)
2099 		fmts[nfmt++] = UMA_ZFLAG_INTERNAL;
2100 	else
2101 		fmts[nfmt++] = UMA_ZFLAG_OFFPAGE;
2102 
2103 	/*
2104 	 * Choose a slab size and format which satisfy the minimum efficiency.
2105 	 * Prefer the smallest slab size that meets the constraints.
2106 	 *
2107 	 * Start with a minimum slab size, to accommodate CACHESPREAD.  Then,
2108 	 * for small items (up to PAGE_SIZE), the iteration increment is one
2109 	 * page; and for large items, the increment is one item.
2110 	 */
2111 	i = (slabsize + rsize - keg->uk_size) / MAX(PAGE_SIZE, rsize);
2112 	KASSERT(i >= 1, ("keg %s(%p) flags=0x%b slabsize=%u, rsize=%u, i=%u",
2113 	    keg->uk_name, keg, keg->uk_flags, PRINT_UMA_ZFLAGS, slabsize,
2114 	    rsize, i));
2115 	for ( ; ; i++) {
2116 		slabsize = (rsize <= PAGE_SIZE) ? ptoa(i) :
2117 		    round_page(rsize * (i - 1) + keg->uk_size);
2118 
2119 		for (j = 0; j < nfmt; j++) {
2120 			/* Only if we have no viable format yet. */
2121 			if ((fmts[j] & UMA_ZFLAG_INTERNAL) != 0 &&
2122 			    kl.ipers > 0)
2123 				continue;
2124 
2125 			keg_layout_one(keg, rsize, slabsize, fmts[j], &kl_tmp);
2126 			if (kl_tmp.eff <= kl.eff)
2127 				continue;
2128 
2129 			kl = kl_tmp;
2130 
2131 			CTR6(KTR_UMA, "keg %s layout: format %#x "
2132 			    "(ipers %u * rsize %u) / slabsize %#x = %u%% eff",
2133 			    keg->uk_name, kl.format, kl.ipers, rsize,
2134 			    kl.slabsize, UMA_FIXPT_PCT(kl.eff));
2135 
2136 			/* Stop when we reach the minimum efficiency. */
2137 			if (kl.eff >= UMA_MIN_EFF)
2138 				break;
2139 		}
2140 
2141 		if (kl.eff >= UMA_MIN_EFF || !multipage_slabs ||
2142 		    slabsize >= SLAB_MAX_SETSIZE * rsize ||
2143 		    (keg->uk_flags & (UMA_ZONE_PCPU | UMA_ZONE_CONTIG)) != 0)
2144 			break;
2145 	}
2146 
2147 	pages = atop(kl.slabsize);
2148 	if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
2149 		pages *= mp_maxid + 1;
2150 
2151 	keg->uk_rsize = rsize;
2152 	keg->uk_ipers = kl.ipers;
2153 	keg->uk_ppera = pages;
2154 	keg->uk_flags |= kl.format;
2155 
2156 	/*
2157 	 * How do we find the slab header if it is offpage or if not all item
2158 	 * start addresses are in the same page?  We could solve the latter
2159 	 * case with vaddr alignment, but we don't.
2160 	 */
2161 	if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0 ||
2162 	    (keg->uk_ipers - 1) * rsize >= PAGE_SIZE) {
2163 		if ((keg->uk_flags & UMA_ZONE_NOTPAGE) != 0)
2164 			keg->uk_flags |= UMA_ZFLAG_HASH;
2165 		else
2166 			keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2167 	}
2168 
2169 	CTR6(KTR_UMA, "%s: keg=%s, flags=%#x, rsize=%u, ipers=%u, ppera=%u",
2170 	    __func__, keg->uk_name, keg->uk_flags, rsize, keg->uk_ipers,
2171 	    pages);
2172 	KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_MAX_SETSIZE,
2173 	    ("%s: keg=%s, flags=0x%b, rsize=%u, ipers=%u, ppera=%u", __func__,
2174 	     keg->uk_name, keg->uk_flags, PRINT_UMA_ZFLAGS, rsize,
2175 	     keg->uk_ipers, pages));
2176 }
2177 
2178 /*
2179  * Keg header ctor.  This initializes all fields, locks, etc.  And inserts
2180  * the keg onto the global keg list.
2181  *
2182  * Arguments/Returns follow uma_ctor specifications
2183  *	udata  Actually uma_kctor_args
2184  */
2185 static int
2186 keg_ctor(void *mem, int size, void *udata, int flags)
2187 {
2188 	struct uma_kctor_args *arg = udata;
2189 	uma_keg_t keg = mem;
2190 	uma_zone_t zone;
2191 	int i;
2192 
2193 	bzero(keg, size);
2194 	keg->uk_size = arg->size;
2195 	keg->uk_init = arg->uminit;
2196 	keg->uk_fini = arg->fini;
2197 	keg->uk_align = arg->align;
2198 	keg->uk_reserve = 0;
2199 	keg->uk_flags = arg->flags;
2200 
2201 	/*
2202 	 * We use a global round-robin policy by default.  Zones with
2203 	 * UMA_ZONE_FIRSTTOUCH set will use first-touch instead, in which
2204 	 * case the iterator is never run.
2205 	 */
2206 	keg->uk_dr.dr_policy = DOMAINSET_RR();
2207 	keg->uk_dr.dr_iter = 0;
2208 
2209 	/*
2210 	 * The primary zone is passed to us at keg-creation time.
2211 	 */
2212 	zone = arg->zone;
2213 	keg->uk_name = zone->uz_name;
2214 
2215 	if (arg->flags & UMA_ZONE_ZINIT)
2216 		keg->uk_init = zero_init;
2217 
2218 	if (arg->flags & UMA_ZONE_MALLOC)
2219 		keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2220 
2221 #ifndef SMP
2222 	keg->uk_flags &= ~UMA_ZONE_PCPU;
2223 #endif
2224 
2225 	keg_layout(keg);
2226 
2227 	/*
2228 	 * Use a first-touch NUMA policy for kegs that pmap_extract() will
2229 	 * work on.  Use round-robin for everything else.
2230 	 *
2231 	 * Zones may override the default by specifying either.
2232 	 */
2233 #ifdef NUMA
2234 	if ((keg->uk_flags &
2235 	    (UMA_ZONE_ROUNDROBIN | UMA_ZFLAG_CACHE | UMA_ZONE_NOTPAGE)) == 0)
2236 		keg->uk_flags |= UMA_ZONE_FIRSTTOUCH;
2237 	else if ((keg->uk_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2238 		keg->uk_flags |= UMA_ZONE_ROUNDROBIN;
2239 #endif
2240 
2241 	/*
2242 	 * If we haven't booted yet we need allocations to go through the
2243 	 * startup cache until the vm is ready.
2244 	 */
2245 #ifdef UMA_MD_SMALL_ALLOC
2246 	if (keg->uk_ppera == 1)
2247 		keg->uk_allocf = uma_small_alloc;
2248 	else
2249 #endif
2250 	if (booted < BOOT_KVA)
2251 		keg->uk_allocf = startup_alloc;
2252 	else if (keg->uk_flags & UMA_ZONE_PCPU)
2253 		keg->uk_allocf = pcpu_page_alloc;
2254 	else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 && keg->uk_ppera > 1)
2255 		keg->uk_allocf = contig_alloc;
2256 	else
2257 		keg->uk_allocf = page_alloc;
2258 #ifdef UMA_MD_SMALL_ALLOC
2259 	if (keg->uk_ppera == 1)
2260 		keg->uk_freef = uma_small_free;
2261 	else
2262 #endif
2263 	if (keg->uk_flags & UMA_ZONE_PCPU)
2264 		keg->uk_freef = pcpu_page_free;
2265 	else
2266 		keg->uk_freef = page_free;
2267 
2268 	/*
2269 	 * Initialize keg's locks.
2270 	 */
2271 	for (i = 0; i < vm_ndomains; i++)
2272 		KEG_LOCK_INIT(keg, i, (arg->flags & UMA_ZONE_MTXCLASS));
2273 
2274 	/*
2275 	 * If we're putting the slab header in the actual page we need to
2276 	 * figure out where in each page it goes.  See slab_sizeof
2277 	 * definition.
2278 	 */
2279 	if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE)) {
2280 		size_t shsize;
2281 
2282 		shsize = slab_sizeof(keg->uk_ipers);
2283 		keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - shsize;
2284 		/*
2285 		 * The only way the following is possible is if with our
2286 		 * UMA_ALIGN_PTR adjustments we are now bigger than
2287 		 * UMA_SLAB_SIZE.  I haven't checked whether this is
2288 		 * mathematically possible for all cases, so we make
2289 		 * sure here anyway.
2290 		 */
2291 		KASSERT(keg->uk_pgoff + shsize <= PAGE_SIZE * keg->uk_ppera,
2292 		    ("zone %s ipers %d rsize %d size %d slab won't fit",
2293 		    zone->uz_name, keg->uk_ipers, keg->uk_rsize, keg->uk_size));
2294 	}
2295 
2296 	if (keg->uk_flags & UMA_ZFLAG_HASH)
2297 		hash_alloc(&keg->uk_hash, 0);
2298 
2299 	CTR3(KTR_UMA, "keg_ctor %p zone %s(%p)", keg, zone->uz_name, zone);
2300 
2301 	LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
2302 
2303 	rw_wlock(&uma_rwlock);
2304 	LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
2305 	rw_wunlock(&uma_rwlock);
2306 	return (0);
2307 }
2308 
2309 static void
2310 zone_kva_available(uma_zone_t zone, void *unused)
2311 {
2312 	uma_keg_t keg;
2313 
2314 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
2315 		return;
2316 	KEG_GET(zone, keg);
2317 
2318 	if (keg->uk_allocf == startup_alloc) {
2319 		/* Switch to the real allocator. */
2320 		if (keg->uk_flags & UMA_ZONE_PCPU)
2321 			keg->uk_allocf = pcpu_page_alloc;
2322 		else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 &&
2323 		    keg->uk_ppera > 1)
2324 			keg->uk_allocf = contig_alloc;
2325 		else
2326 			keg->uk_allocf = page_alloc;
2327 	}
2328 }
2329 
2330 static void
2331 zone_alloc_counters(uma_zone_t zone, void *unused)
2332 {
2333 
2334 	zone->uz_allocs = counter_u64_alloc(M_WAITOK);
2335 	zone->uz_frees = counter_u64_alloc(M_WAITOK);
2336 	zone->uz_fails = counter_u64_alloc(M_WAITOK);
2337 	zone->uz_xdomain = counter_u64_alloc(M_WAITOK);
2338 }
2339 
2340 static void
2341 zone_alloc_sysctl(uma_zone_t zone, void *unused)
2342 {
2343 	uma_zone_domain_t zdom;
2344 	uma_domain_t dom;
2345 	uma_keg_t keg;
2346 	struct sysctl_oid *oid, *domainoid;
2347 	int domains, i, cnt;
2348 	static const char *nokeg = "cache zone";
2349 	char *c;
2350 
2351 	/*
2352 	 * Make a sysctl safe copy of the zone name by removing
2353 	 * any special characters and handling dups by appending
2354 	 * an index.
2355 	 */
2356 	if (zone->uz_namecnt != 0) {
2357 		/* Count the number of decimal digits and '_' separator. */
2358 		for (i = 1, cnt = zone->uz_namecnt; cnt != 0; i++)
2359 			cnt /= 10;
2360 		zone->uz_ctlname = malloc(strlen(zone->uz_name) + i + 1,
2361 		    M_UMA, M_WAITOK);
2362 		sprintf(zone->uz_ctlname, "%s_%d", zone->uz_name,
2363 		    zone->uz_namecnt);
2364 	} else
2365 		zone->uz_ctlname = strdup(zone->uz_name, M_UMA);
2366 	for (c = zone->uz_ctlname; *c != '\0'; c++)
2367 		if (strchr("./\\ -", *c) != NULL)
2368 			*c = '_';
2369 
2370 	/*
2371 	 * Basic parameters at the root.
2372 	 */
2373 	zone->uz_oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_vm_uma),
2374 	    OID_AUTO, zone->uz_ctlname, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2375 	oid = zone->uz_oid;
2376 	SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2377 	    "size", CTLFLAG_RD, &zone->uz_size, 0, "Allocation size");
2378 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2379 	    "flags", CTLFLAG_RD | CTLTYPE_STRING | CTLFLAG_MPSAFE,
2380 	    zone, 0, sysctl_handle_uma_zone_flags, "A",
2381 	    "Allocator configuration flags");
2382 	SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2383 	    "bucket_size", CTLFLAG_RD, &zone->uz_bucket_size, 0,
2384 	    "Desired per-cpu cache size");
2385 	SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2386 	    "bucket_size_max", CTLFLAG_RD, &zone->uz_bucket_size_max, 0,
2387 	    "Maximum allowed per-cpu cache size");
2388 
2389 	/*
2390 	 * keg if present.
2391 	 */
2392 	if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
2393 		domains = vm_ndomains;
2394 	else
2395 		domains = 1;
2396 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2397 	    "keg", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2398 	keg = zone->uz_keg;
2399 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0) {
2400 		SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2401 		    "name", CTLFLAG_RD, keg->uk_name, "Keg name");
2402 		SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2403 		    "rsize", CTLFLAG_RD, &keg->uk_rsize, 0,
2404 		    "Real object size with alignment");
2405 		SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2406 		    "ppera", CTLFLAG_RD, &keg->uk_ppera, 0,
2407 		    "pages per-slab allocation");
2408 		SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2409 		    "ipers", CTLFLAG_RD, &keg->uk_ipers, 0,
2410 		    "items available per-slab");
2411 		SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2412 		    "align", CTLFLAG_RD, &keg->uk_align, 0,
2413 		    "item alignment mask");
2414 		SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2415 		    "efficiency", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2416 		    keg, 0, sysctl_handle_uma_slab_efficiency, "I",
2417 		    "Slab utilization (100 - internal fragmentation %)");
2418 		domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(oid),
2419 		    OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2420 		for (i = 0; i < domains; i++) {
2421 			dom = &keg->uk_domain[i];
2422 			oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2423 			    OID_AUTO, VM_DOMAIN(i)->vmd_name,
2424 			    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2425 			SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2426 			    "pages", CTLFLAG_RD, &dom->ud_pages, 0,
2427 			    "Total pages currently allocated from VM");
2428 			SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2429 			    "free_items", CTLFLAG_RD, &dom->ud_free_items, 0,
2430 			    "items free in the slab layer");
2431 		}
2432 	} else
2433 		SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2434 		    "name", CTLFLAG_RD, nokeg, "Keg name");
2435 
2436 	/*
2437 	 * Information about zone limits.
2438 	 */
2439 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2440 	    "limit", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2441 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2442 	    "items", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2443 	    zone, 0, sysctl_handle_uma_zone_items, "QU",
2444 	    "current number of allocated items if limit is set");
2445 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2446 	    "max_items", CTLFLAG_RD, &zone->uz_max_items, 0,
2447 	    "Maximum number of cached items");
2448 	SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2449 	    "sleepers", CTLFLAG_RD, &zone->uz_sleepers, 0,
2450 	    "Number of threads sleeping at limit");
2451 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2452 	    "sleeps", CTLFLAG_RD, &zone->uz_sleeps, 0,
2453 	    "Total zone limit sleeps");
2454 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2455 	    "bucket_max", CTLFLAG_RD, &zone->uz_bucket_max, 0,
2456 	    "Maximum number of items in each domain's bucket cache");
2457 
2458 	/*
2459 	 * Per-domain zone information.
2460 	 */
2461 	domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid),
2462 	    OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2463 	for (i = 0; i < domains; i++) {
2464 		zdom = ZDOM_GET(zone, i);
2465 		oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2466 		    OID_AUTO, VM_DOMAIN(i)->vmd_name,
2467 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2468 		SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2469 		    "nitems", CTLFLAG_RD, &zdom->uzd_nitems,
2470 		    "number of items in this domain");
2471 		SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2472 		    "imax", CTLFLAG_RD, &zdom->uzd_imax,
2473 		    "maximum item count in this period");
2474 		SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2475 		    "imin", CTLFLAG_RD, &zdom->uzd_imin,
2476 		    "minimum item count in this period");
2477 		SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2478 		    "wss", CTLFLAG_RD, &zdom->uzd_wss,
2479 		    "Working set size");
2480 	}
2481 
2482 	/*
2483 	 * General statistics.
2484 	 */
2485 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2486 	    "stats", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2487 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2488 	    "current", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2489 	    zone, 1, sysctl_handle_uma_zone_cur, "I",
2490 	    "Current number of allocated items");
2491 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2492 	    "allocs", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2493 	    zone, 0, sysctl_handle_uma_zone_allocs, "QU",
2494 	    "Total allocation calls");
2495 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2496 	    "frees", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2497 	    zone, 0, sysctl_handle_uma_zone_frees, "QU",
2498 	    "Total free calls");
2499 	SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2500 	    "fails", CTLFLAG_RD, &zone->uz_fails,
2501 	    "Number of allocation failures");
2502 	SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2503 	    "xdomain", CTLFLAG_RD, &zone->uz_xdomain,
2504 	    "Free calls from the wrong domain");
2505 }
2506 
2507 struct uma_zone_count {
2508 	const char	*name;
2509 	int		count;
2510 };
2511 
2512 static void
2513 zone_count(uma_zone_t zone, void *arg)
2514 {
2515 	struct uma_zone_count *cnt;
2516 
2517 	cnt = arg;
2518 	/*
2519 	 * Some zones are rapidly created with identical names and
2520 	 * destroyed out of order.  This can lead to gaps in the count.
2521 	 * Use one greater than the maximum observed for this name.
2522 	 */
2523 	if (strcmp(zone->uz_name, cnt->name) == 0)
2524 		cnt->count = MAX(cnt->count,
2525 		    zone->uz_namecnt + 1);
2526 }
2527 
2528 static void
2529 zone_update_caches(uma_zone_t zone)
2530 {
2531 	int i;
2532 
2533 	for (i = 0; i <= mp_maxid; i++) {
2534 		cache_set_uz_size(&zone->uz_cpu[i], zone->uz_size);
2535 		cache_set_uz_flags(&zone->uz_cpu[i], zone->uz_flags);
2536 	}
2537 }
2538 
2539 /*
2540  * Zone header ctor.  This initializes all fields, locks, etc.
2541  *
2542  * Arguments/Returns follow uma_ctor specifications
2543  *	udata  Actually uma_zctor_args
2544  */
2545 static int
2546 zone_ctor(void *mem, int size, void *udata, int flags)
2547 {
2548 	struct uma_zone_count cnt;
2549 	struct uma_zctor_args *arg = udata;
2550 	uma_zone_domain_t zdom;
2551 	uma_zone_t zone = mem;
2552 	uma_zone_t z;
2553 	uma_keg_t keg;
2554 	int i;
2555 
2556 	bzero(zone, size);
2557 	zone->uz_name = arg->name;
2558 	zone->uz_ctor = arg->ctor;
2559 	zone->uz_dtor = arg->dtor;
2560 	zone->uz_init = NULL;
2561 	zone->uz_fini = NULL;
2562 	zone->uz_sleeps = 0;
2563 	zone->uz_bucket_size = 0;
2564 	zone->uz_bucket_size_min = 0;
2565 	zone->uz_bucket_size_max = BUCKET_MAX;
2566 	zone->uz_flags = (arg->flags & UMA_ZONE_SMR);
2567 	zone->uz_warning = NULL;
2568 	/* The domain structures follow the cpu structures. */
2569 	zone->uz_bucket_max = ULONG_MAX;
2570 	timevalclear(&zone->uz_ratecheck);
2571 
2572 	/* Count the number of duplicate names. */
2573 	cnt.name = arg->name;
2574 	cnt.count = 0;
2575 	zone_foreach(zone_count, &cnt);
2576 	zone->uz_namecnt = cnt.count;
2577 	ZONE_CROSS_LOCK_INIT(zone);
2578 
2579 	for (i = 0; i < vm_ndomains; i++) {
2580 		zdom = ZDOM_GET(zone, i);
2581 		ZDOM_LOCK_INIT(zone, zdom, (arg->flags & UMA_ZONE_MTXCLASS));
2582 		STAILQ_INIT(&zdom->uzd_buckets);
2583 	}
2584 
2585 #ifdef INVARIANTS
2586 	if (arg->uminit == trash_init && arg->fini == trash_fini)
2587 		zone->uz_flags |= UMA_ZFLAG_TRASH | UMA_ZFLAG_CTORDTOR;
2588 #endif
2589 
2590 	/*
2591 	 * This is a pure cache zone, no kegs.
2592 	 */
2593 	if (arg->import) {
2594 		KASSERT((arg->flags & UMA_ZFLAG_CACHE) != 0,
2595 		    ("zone_ctor: Import specified for non-cache zone."));
2596 		zone->uz_flags = arg->flags;
2597 		zone->uz_size = arg->size;
2598 		zone->uz_import = arg->import;
2599 		zone->uz_release = arg->release;
2600 		zone->uz_arg = arg->arg;
2601 #ifdef NUMA
2602 		/*
2603 		 * Cache zones are round-robin unless a policy is
2604 		 * specified because they may have incompatible
2605 		 * constraints.
2606 		 */
2607 		if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2608 			zone->uz_flags |= UMA_ZONE_ROUNDROBIN;
2609 #endif
2610 		rw_wlock(&uma_rwlock);
2611 		LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link);
2612 		rw_wunlock(&uma_rwlock);
2613 		goto out;
2614 	}
2615 
2616 	/*
2617 	 * Use the regular zone/keg/slab allocator.
2618 	 */
2619 	zone->uz_import = zone_import;
2620 	zone->uz_release = zone_release;
2621 	zone->uz_arg = zone;
2622 	keg = arg->keg;
2623 
2624 	if (arg->flags & UMA_ZONE_SECONDARY) {
2625 		KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
2626 		    ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
2627 		KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
2628 		zone->uz_init = arg->uminit;
2629 		zone->uz_fini = arg->fini;
2630 		zone->uz_flags |= UMA_ZONE_SECONDARY;
2631 		rw_wlock(&uma_rwlock);
2632 		ZONE_LOCK(zone);
2633 		LIST_FOREACH(z, &keg->uk_zones, uz_link) {
2634 			if (LIST_NEXT(z, uz_link) == NULL) {
2635 				LIST_INSERT_AFTER(z, zone, uz_link);
2636 				break;
2637 			}
2638 		}
2639 		ZONE_UNLOCK(zone);
2640 		rw_wunlock(&uma_rwlock);
2641 	} else if (keg == NULL) {
2642 		if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
2643 		    arg->align, arg->flags)) == NULL)
2644 			return (ENOMEM);
2645 	} else {
2646 		struct uma_kctor_args karg;
2647 		int error;
2648 
2649 		/* We should only be here from uma_startup() */
2650 		karg.size = arg->size;
2651 		karg.uminit = arg->uminit;
2652 		karg.fini = arg->fini;
2653 		karg.align = arg->align;
2654 		karg.flags = (arg->flags & ~UMA_ZONE_SMR);
2655 		karg.zone = zone;
2656 		error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
2657 		    flags);
2658 		if (error)
2659 			return (error);
2660 	}
2661 
2662 	/* Inherit properties from the keg. */
2663 	zone->uz_keg = keg;
2664 	zone->uz_size = keg->uk_size;
2665 	zone->uz_flags |= (keg->uk_flags &
2666 	    (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
2667 
2668 out:
2669 	if (booted >= BOOT_PCPU) {
2670 		zone_alloc_counters(zone, NULL);
2671 		if (booted >= BOOT_RUNNING)
2672 			zone_alloc_sysctl(zone, NULL);
2673 	} else {
2674 		zone->uz_allocs = EARLY_COUNTER;
2675 		zone->uz_frees = EARLY_COUNTER;
2676 		zone->uz_fails = EARLY_COUNTER;
2677 	}
2678 
2679 	/* Caller requests a private SMR context. */
2680 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
2681 		zone->uz_smr = smr_create(zone->uz_name, 0, 0);
2682 
2683 	KASSERT((arg->flags & (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET)) !=
2684 	    (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET),
2685 	    ("Invalid zone flag combination"));
2686 	if (arg->flags & UMA_ZFLAG_INTERNAL)
2687 		zone->uz_bucket_size_max = zone->uz_bucket_size = 0;
2688 	if ((arg->flags & UMA_ZONE_MAXBUCKET) != 0)
2689 		zone->uz_bucket_size = BUCKET_MAX;
2690 	else if ((arg->flags & UMA_ZONE_MINBUCKET) != 0)
2691 		zone->uz_bucket_size_max = zone->uz_bucket_size = BUCKET_MIN;
2692 	else if ((arg->flags & UMA_ZONE_NOBUCKET) != 0)
2693 		zone->uz_bucket_size = 0;
2694 	else
2695 		zone->uz_bucket_size = bucket_select(zone->uz_size);
2696 	zone->uz_bucket_size_min = zone->uz_bucket_size;
2697 	if (zone->uz_dtor != NULL || zone->uz_ctor != NULL)
2698 		zone->uz_flags |= UMA_ZFLAG_CTORDTOR;
2699 	zone_update_caches(zone);
2700 
2701 	return (0);
2702 }
2703 
2704 /*
2705  * Keg header dtor.  This frees all data, destroys locks, frees the hash
2706  * table and removes the keg from the global list.
2707  *
2708  * Arguments/Returns follow uma_dtor specifications
2709  *	udata  unused
2710  */
2711 static void
2712 keg_dtor(void *arg, int size, void *udata)
2713 {
2714 	uma_keg_t keg;
2715 	uint32_t free, pages;
2716 	int i;
2717 
2718 	keg = (uma_keg_t)arg;
2719 	free = pages = 0;
2720 	for (i = 0; i < vm_ndomains; i++) {
2721 		free += keg->uk_domain[i].ud_free_items;
2722 		pages += keg->uk_domain[i].ud_pages;
2723 		KEG_LOCK_FINI(keg, i);
2724 	}
2725 	if (pages != 0)
2726 		printf("Freed UMA keg (%s) was not empty (%u items). "
2727 		    " Lost %u pages of memory.\n",
2728 		    keg->uk_name ? keg->uk_name : "",
2729 		    pages / keg->uk_ppera * keg->uk_ipers - free, pages);
2730 
2731 	hash_free(&keg->uk_hash);
2732 }
2733 
2734 /*
2735  * Zone header dtor.
2736  *
2737  * Arguments/Returns follow uma_dtor specifications
2738  *	udata  unused
2739  */
2740 static void
2741 zone_dtor(void *arg, int size, void *udata)
2742 {
2743 	uma_zone_t zone;
2744 	uma_keg_t keg;
2745 	int i;
2746 
2747 	zone = (uma_zone_t)arg;
2748 
2749 	sysctl_remove_oid(zone->uz_oid, 1, 1);
2750 
2751 	if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
2752 		cache_drain(zone);
2753 
2754 	rw_wlock(&uma_rwlock);
2755 	LIST_REMOVE(zone, uz_link);
2756 	rw_wunlock(&uma_rwlock);
2757 	zone_reclaim(zone, M_WAITOK, true);
2758 
2759 	/*
2760 	 * We only destroy kegs from non secondary/non cache zones.
2761 	 */
2762 	if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) {
2763 		keg = zone->uz_keg;
2764 		rw_wlock(&uma_rwlock);
2765 		LIST_REMOVE(keg, uk_link);
2766 		rw_wunlock(&uma_rwlock);
2767 		zone_free_item(kegs, keg, NULL, SKIP_NONE);
2768 	}
2769 	counter_u64_free(zone->uz_allocs);
2770 	counter_u64_free(zone->uz_frees);
2771 	counter_u64_free(zone->uz_fails);
2772 	counter_u64_free(zone->uz_xdomain);
2773 	free(zone->uz_ctlname, M_UMA);
2774 	for (i = 0; i < vm_ndomains; i++)
2775 		ZDOM_LOCK_FINI(ZDOM_GET(zone, i));
2776 	ZONE_CROSS_LOCK_FINI(zone);
2777 }
2778 
2779 static void
2780 zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *arg), void *arg)
2781 {
2782 	uma_keg_t keg;
2783 	uma_zone_t zone;
2784 
2785 	LIST_FOREACH(keg, &uma_kegs, uk_link) {
2786 		LIST_FOREACH(zone, &keg->uk_zones, uz_link)
2787 			zfunc(zone, arg);
2788 	}
2789 	LIST_FOREACH(zone, &uma_cachezones, uz_link)
2790 		zfunc(zone, arg);
2791 }
2792 
2793 /*
2794  * Traverses every zone in the system and calls a callback
2795  *
2796  * Arguments:
2797  *	zfunc  A pointer to a function which accepts a zone
2798  *		as an argument.
2799  *
2800  * Returns:
2801  *	Nothing
2802  */
2803 static void
2804 zone_foreach(void (*zfunc)(uma_zone_t, void *arg), void *arg)
2805 {
2806 
2807 	rw_rlock(&uma_rwlock);
2808 	zone_foreach_unlocked(zfunc, arg);
2809 	rw_runlock(&uma_rwlock);
2810 }
2811 
2812 /*
2813  * Initialize the kernel memory allocator.  This is done after pages can be
2814  * allocated but before general KVA is available.
2815  */
2816 void
2817 uma_startup1(vm_offset_t virtual_avail)
2818 {
2819 	struct uma_zctor_args args;
2820 	size_t ksize, zsize, size;
2821 	uma_keg_t primarykeg;
2822 	uintptr_t m;
2823 	int domain;
2824 	uint8_t pflag;
2825 
2826 	bootstart = bootmem = virtual_avail;
2827 
2828 	rw_init(&uma_rwlock, "UMA lock");
2829 	sx_init(&uma_reclaim_lock, "umareclaim");
2830 
2831 	ksize = sizeof(struct uma_keg) +
2832 	    (sizeof(struct uma_domain) * vm_ndomains);
2833 	ksize = roundup(ksize, UMA_SUPER_ALIGN);
2834 	zsize = sizeof(struct uma_zone) +
2835 	    (sizeof(struct uma_cache) * (mp_maxid + 1)) +
2836 	    (sizeof(struct uma_zone_domain) * vm_ndomains);
2837 	zsize = roundup(zsize, UMA_SUPER_ALIGN);
2838 
2839 	/* Allocate the zone of zones, zone of kegs, and zone of zones keg. */
2840 	size = (zsize * 2) + ksize;
2841 	for (domain = 0; domain < vm_ndomains; domain++) {
2842 		m = (uintptr_t)startup_alloc(NULL, size, domain, &pflag,
2843 		    M_NOWAIT | M_ZERO);
2844 		if (m != 0)
2845 			break;
2846 	}
2847 	zones = (uma_zone_t)m;
2848 	m += zsize;
2849 	kegs = (uma_zone_t)m;
2850 	m += zsize;
2851 	primarykeg = (uma_keg_t)m;
2852 
2853 	/* "manually" create the initial zone */
2854 	memset(&args, 0, sizeof(args));
2855 	args.name = "UMA Kegs";
2856 	args.size = ksize;
2857 	args.ctor = keg_ctor;
2858 	args.dtor = keg_dtor;
2859 	args.uminit = zero_init;
2860 	args.fini = NULL;
2861 	args.keg = primarykeg;
2862 	args.align = UMA_SUPER_ALIGN - 1;
2863 	args.flags = UMA_ZFLAG_INTERNAL;
2864 	zone_ctor(kegs, zsize, &args, M_WAITOK);
2865 
2866 	args.name = "UMA Zones";
2867 	args.size = zsize;
2868 	args.ctor = zone_ctor;
2869 	args.dtor = zone_dtor;
2870 	args.uminit = zero_init;
2871 	args.fini = NULL;
2872 	args.keg = NULL;
2873 	args.align = UMA_SUPER_ALIGN - 1;
2874 	args.flags = UMA_ZFLAG_INTERNAL;
2875 	zone_ctor(zones, zsize, &args, M_WAITOK);
2876 
2877 	/* Now make zones for slab headers */
2878 	slabzones[0] = uma_zcreate("UMA Slabs 0", SLABZONE0_SIZE,
2879 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2880 	slabzones[1] = uma_zcreate("UMA Slabs 1", SLABZONE1_SIZE,
2881 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2882 
2883 	hashzone = uma_zcreate("UMA Hash",
2884 	    sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
2885 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2886 
2887 	bucket_init();
2888 	smr_init();
2889 }
2890 
2891 #ifndef UMA_MD_SMALL_ALLOC
2892 extern void vm_radix_reserve_kva(void);
2893 #endif
2894 
2895 /*
2896  * Advertise the availability of normal kva allocations and switch to
2897  * the default back-end allocator.  Marks the KVA we consumed on startup
2898  * as used in the map.
2899  */
2900 void
2901 uma_startup2(void)
2902 {
2903 
2904 	if (bootstart != bootmem) {
2905 		vm_map_lock(kernel_map);
2906 		(void)vm_map_insert(kernel_map, NULL, 0, bootstart, bootmem,
2907 		    VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT);
2908 		vm_map_unlock(kernel_map);
2909 	}
2910 
2911 #ifndef UMA_MD_SMALL_ALLOC
2912 	/* Set up radix zone to use noobj_alloc. */
2913 	vm_radix_reserve_kva();
2914 #endif
2915 
2916 	booted = BOOT_KVA;
2917 	zone_foreach_unlocked(zone_kva_available, NULL);
2918 	bucket_enable();
2919 }
2920 
2921 /*
2922  * Allocate counters as early as possible so that boot-time allocations are
2923  * accounted more precisely.
2924  */
2925 static void
2926 uma_startup_pcpu(void *arg __unused)
2927 {
2928 
2929 	zone_foreach_unlocked(zone_alloc_counters, NULL);
2930 	booted = BOOT_PCPU;
2931 }
2932 SYSINIT(uma_startup_pcpu, SI_SUB_COUNTER, SI_ORDER_ANY, uma_startup_pcpu, NULL);
2933 
2934 /*
2935  * Finish our initialization steps.
2936  */
2937 static void
2938 uma_startup3(void *arg __unused)
2939 {
2940 
2941 #ifdef INVARIANTS
2942 	TUNABLE_INT_FETCH("vm.debug.divisor", &dbg_divisor);
2943 	uma_dbg_cnt = counter_u64_alloc(M_WAITOK);
2944 	uma_skip_cnt = counter_u64_alloc(M_WAITOK);
2945 #endif
2946 	zone_foreach_unlocked(zone_alloc_sysctl, NULL);
2947 	callout_init(&uma_callout, 1);
2948 	callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
2949 	booted = BOOT_RUNNING;
2950 
2951 	EVENTHANDLER_REGISTER(shutdown_post_sync, uma_shutdown, NULL,
2952 	    EVENTHANDLER_PRI_FIRST);
2953 }
2954 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
2955 
2956 static void
2957 uma_shutdown(void)
2958 {
2959 
2960 	booted = BOOT_SHUTDOWN;
2961 }
2962 
2963 static uma_keg_t
2964 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
2965 		int align, uint32_t flags)
2966 {
2967 	struct uma_kctor_args args;
2968 
2969 	args.size = size;
2970 	args.uminit = uminit;
2971 	args.fini = fini;
2972 	args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
2973 	args.flags = flags;
2974 	args.zone = zone;
2975 	return (zone_alloc_item(kegs, &args, UMA_ANYDOMAIN, M_WAITOK));
2976 }
2977 
2978 /* Public functions */
2979 /* See uma.h */
2980 void
2981 uma_set_align(int align)
2982 {
2983 
2984 	if (align != UMA_ALIGN_CACHE)
2985 		uma_align_cache = align;
2986 }
2987 
2988 /* See uma.h */
2989 uma_zone_t
2990 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
2991 		uma_init uminit, uma_fini fini, int align, uint32_t flags)
2992 
2993 {
2994 	struct uma_zctor_args args;
2995 	uma_zone_t res;
2996 
2997 	KASSERT(powerof2(align + 1), ("invalid zone alignment %d for \"%s\"",
2998 	    align, name));
2999 
3000 	/* This stuff is essential for the zone ctor */
3001 	memset(&args, 0, sizeof(args));
3002 	args.name = name;
3003 	args.size = size;
3004 	args.ctor = ctor;
3005 	args.dtor = dtor;
3006 	args.uminit = uminit;
3007 	args.fini = fini;
3008 #ifdef  INVARIANTS
3009 	/*
3010 	 * Inject procedures which check for memory use after free if we are
3011 	 * allowed to scramble the memory while it is not allocated.  This
3012 	 * requires that: UMA is actually able to access the memory, no init
3013 	 * or fini procedures, no dependency on the initial value of the
3014 	 * memory, and no (legitimate) use of the memory after free.  Note,
3015 	 * the ctor and dtor do not need to be empty.
3016 	 */
3017 	if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOTOUCH |
3018 	    UMA_ZONE_NOFREE))) && uminit == NULL && fini == NULL) {
3019 		args.uminit = trash_init;
3020 		args.fini = trash_fini;
3021 	}
3022 #endif
3023 	args.align = align;
3024 	args.flags = flags;
3025 	args.keg = NULL;
3026 
3027 	sx_slock(&uma_reclaim_lock);
3028 	res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3029 	sx_sunlock(&uma_reclaim_lock);
3030 
3031 	return (res);
3032 }
3033 
3034 /* See uma.h */
3035 uma_zone_t
3036 uma_zsecond_create(const char *name, uma_ctor ctor, uma_dtor dtor,
3037     uma_init zinit, uma_fini zfini, uma_zone_t primary)
3038 {
3039 	struct uma_zctor_args args;
3040 	uma_keg_t keg;
3041 	uma_zone_t res;
3042 
3043 	keg = primary->uz_keg;
3044 	memset(&args, 0, sizeof(args));
3045 	args.name = name;
3046 	args.size = keg->uk_size;
3047 	args.ctor = ctor;
3048 	args.dtor = dtor;
3049 	args.uminit = zinit;
3050 	args.fini = zfini;
3051 	args.align = keg->uk_align;
3052 	args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
3053 	args.keg = keg;
3054 
3055 	sx_slock(&uma_reclaim_lock);
3056 	res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3057 	sx_sunlock(&uma_reclaim_lock);
3058 
3059 	return (res);
3060 }
3061 
3062 /* See uma.h */
3063 uma_zone_t
3064 uma_zcache_create(const char *name, int size, uma_ctor ctor, uma_dtor dtor,
3065     uma_init zinit, uma_fini zfini, uma_import zimport, uma_release zrelease,
3066     void *arg, int flags)
3067 {
3068 	struct uma_zctor_args args;
3069 
3070 	memset(&args, 0, sizeof(args));
3071 	args.name = name;
3072 	args.size = size;
3073 	args.ctor = ctor;
3074 	args.dtor = dtor;
3075 	args.uminit = zinit;
3076 	args.fini = zfini;
3077 	args.import = zimport;
3078 	args.release = zrelease;
3079 	args.arg = arg;
3080 	args.align = 0;
3081 	args.flags = flags | UMA_ZFLAG_CACHE;
3082 
3083 	return (zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK));
3084 }
3085 
3086 /* See uma.h */
3087 void
3088 uma_zdestroy(uma_zone_t zone)
3089 {
3090 
3091 	/*
3092 	 * Large slabs are expensive to reclaim, so don't bother doing
3093 	 * unnecessary work if we're shutting down.
3094 	 */
3095 	if (booted == BOOT_SHUTDOWN &&
3096 	    zone->uz_fini == NULL && zone->uz_release == zone_release)
3097 		return;
3098 	sx_slock(&uma_reclaim_lock);
3099 	zone_free_item(zones, zone, NULL, SKIP_NONE);
3100 	sx_sunlock(&uma_reclaim_lock);
3101 }
3102 
3103 void
3104 uma_zwait(uma_zone_t zone)
3105 {
3106 
3107 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
3108 		uma_zfree_smr(zone, uma_zalloc_smr(zone, M_WAITOK));
3109 	else if ((zone->uz_flags & UMA_ZONE_PCPU) != 0)
3110 		uma_zfree_pcpu(zone, uma_zalloc_pcpu(zone, M_WAITOK));
3111 	else
3112 		uma_zfree(zone, uma_zalloc(zone, M_WAITOK));
3113 }
3114 
3115 void *
3116 uma_zalloc_pcpu_arg(uma_zone_t zone, void *udata, int flags)
3117 {
3118 	void *item, *pcpu_item;
3119 #ifdef SMP
3120 	int i;
3121 
3122 	MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3123 #endif
3124 	item = uma_zalloc_arg(zone, udata, flags & ~M_ZERO);
3125 	if (item == NULL)
3126 		return (NULL);
3127 	pcpu_item = zpcpu_base_to_offset(item);
3128 	if (flags & M_ZERO) {
3129 #ifdef SMP
3130 		for (i = 0; i <= mp_maxid; i++)
3131 			bzero(zpcpu_get_cpu(pcpu_item, i), zone->uz_size);
3132 #else
3133 		bzero(item, zone->uz_size);
3134 #endif
3135 	}
3136 	return (pcpu_item);
3137 }
3138 
3139 /*
3140  * A stub while both regular and pcpu cases are identical.
3141  */
3142 void
3143 uma_zfree_pcpu_arg(uma_zone_t zone, void *pcpu_item, void *udata)
3144 {
3145 	void *item;
3146 
3147 #ifdef SMP
3148 	MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3149 #endif
3150 	item = zpcpu_offset_to_base(pcpu_item);
3151 	uma_zfree_arg(zone, item, udata);
3152 }
3153 
3154 static inline void *
3155 item_ctor(uma_zone_t zone, int uz_flags, int size, void *udata, int flags,
3156     void *item)
3157 {
3158 #ifdef INVARIANTS
3159 	bool skipdbg;
3160 
3161 	skipdbg = uma_dbg_zskip(zone, item);
3162 	if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3163 	    zone->uz_ctor != trash_ctor)
3164 		trash_ctor(item, size, udata, flags);
3165 #endif
3166 	/* Check flags before loading ctor pointer. */
3167 	if (__predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0) &&
3168 	    __predict_false(zone->uz_ctor != NULL) &&
3169 	    zone->uz_ctor(item, size, udata, flags) != 0) {
3170 		counter_u64_add(zone->uz_fails, 1);
3171 		zone_free_item(zone, item, udata, SKIP_DTOR | SKIP_CNT);
3172 		return (NULL);
3173 	}
3174 #ifdef INVARIANTS
3175 	if (!skipdbg)
3176 		uma_dbg_alloc(zone, NULL, item);
3177 #endif
3178 	if (__predict_false(flags & M_ZERO))
3179 		return (memset(item, 0, size));
3180 
3181 	return (item);
3182 }
3183 
3184 static inline void
3185 item_dtor(uma_zone_t zone, void *item, int size, void *udata,
3186     enum zfreeskip skip)
3187 {
3188 #ifdef INVARIANTS
3189 	bool skipdbg;
3190 
3191 	skipdbg = uma_dbg_zskip(zone, item);
3192 	if (skip == SKIP_NONE && !skipdbg) {
3193 		if ((zone->uz_flags & UMA_ZONE_MALLOC) != 0)
3194 			uma_dbg_free(zone, udata, item);
3195 		else
3196 			uma_dbg_free(zone, NULL, item);
3197 	}
3198 #endif
3199 	if (__predict_true(skip < SKIP_DTOR)) {
3200 		if (zone->uz_dtor != NULL)
3201 			zone->uz_dtor(item, size, udata);
3202 #ifdef INVARIANTS
3203 		if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3204 		    zone->uz_dtor != trash_dtor)
3205 			trash_dtor(item, size, udata);
3206 #endif
3207 	}
3208 }
3209 
3210 #ifdef NUMA
3211 static int
3212 item_domain(void *item)
3213 {
3214 	int domain;
3215 
3216 	domain = _vm_phys_domain(vtophys(item));
3217 	KASSERT(domain >= 0 && domain < vm_ndomains,
3218 	    ("%s: unknown domain for item %p", __func__, item));
3219 	return (domain);
3220 }
3221 #endif
3222 
3223 #if defined(INVARIANTS) || defined(DEBUG_MEMGUARD) || defined(WITNESS)
3224 #define	UMA_ZALLOC_DEBUG
3225 static int
3226 uma_zalloc_debug(uma_zone_t zone, void **itemp, void *udata, int flags)
3227 {
3228 	int error;
3229 
3230 	error = 0;
3231 #ifdef WITNESS
3232 	if (flags & M_WAITOK) {
3233 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
3234 		    "uma_zalloc_debug: zone \"%s\"", zone->uz_name);
3235 	}
3236 #endif
3237 
3238 #ifdef INVARIANTS
3239 	KASSERT((flags & M_EXEC) == 0,
3240 	    ("uma_zalloc_debug: called with M_EXEC"));
3241 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3242 	    ("uma_zalloc_debug: called within spinlock or critical section"));
3243 	KASSERT((zone->uz_flags & UMA_ZONE_PCPU) == 0 || (flags & M_ZERO) == 0,
3244 	    ("uma_zalloc_debug: allocating from a pcpu zone with M_ZERO"));
3245 #endif
3246 
3247 #ifdef DEBUG_MEMGUARD
3248 	if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && memguard_cmp_zone(zone)) {
3249 		void *item;
3250 		item = memguard_alloc(zone->uz_size, flags);
3251 		if (item != NULL) {
3252 			error = EJUSTRETURN;
3253 			if (zone->uz_init != NULL &&
3254 			    zone->uz_init(item, zone->uz_size, flags) != 0) {
3255 				*itemp = NULL;
3256 				return (error);
3257 			}
3258 			if (zone->uz_ctor != NULL &&
3259 			    zone->uz_ctor(item, zone->uz_size, udata,
3260 			    flags) != 0) {
3261 				counter_u64_add(zone->uz_fails, 1);
3262 			    	zone->uz_fini(item, zone->uz_size);
3263 				*itemp = NULL;
3264 				return (error);
3265 			}
3266 			*itemp = item;
3267 			return (error);
3268 		}
3269 		/* This is unfortunate but should not be fatal. */
3270 	}
3271 #endif
3272 	return (error);
3273 }
3274 
3275 static int
3276 uma_zfree_debug(uma_zone_t zone, void *item, void *udata)
3277 {
3278 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3279 	    ("uma_zfree_debug: called with spinlock or critical section held"));
3280 
3281 #ifdef DEBUG_MEMGUARD
3282 	if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && is_memguard_addr(item)) {
3283 		if (zone->uz_dtor != NULL)
3284 			zone->uz_dtor(item, zone->uz_size, udata);
3285 		if (zone->uz_fini != NULL)
3286 			zone->uz_fini(item, zone->uz_size);
3287 		memguard_free(item);
3288 		return (EJUSTRETURN);
3289 	}
3290 #endif
3291 	return (0);
3292 }
3293 #endif
3294 
3295 static inline void *
3296 cache_alloc_item(uma_zone_t zone, uma_cache_t cache, uma_cache_bucket_t bucket,
3297     void *udata, int flags)
3298 {
3299 	void *item;
3300 	int size, uz_flags;
3301 
3302 	item = cache_bucket_pop(cache, bucket);
3303 	size = cache_uz_size(cache);
3304 	uz_flags = cache_uz_flags(cache);
3305 	critical_exit();
3306 	return (item_ctor(zone, uz_flags, size, udata, flags, item));
3307 }
3308 
3309 static __noinline void *
3310 cache_alloc_retry(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3311 {
3312 	uma_cache_bucket_t bucket;
3313 	int domain;
3314 
3315 	while (cache_alloc(zone, cache, udata, flags)) {
3316 		cache = &zone->uz_cpu[curcpu];
3317 		bucket = &cache->uc_allocbucket;
3318 		if (__predict_false(bucket->ucb_cnt == 0))
3319 			continue;
3320 		return (cache_alloc_item(zone, cache, bucket, udata, flags));
3321 	}
3322 	critical_exit();
3323 
3324 	/*
3325 	 * We can not get a bucket so try to return a single item.
3326 	 */
3327 	if (zone->uz_flags & UMA_ZONE_FIRSTTOUCH)
3328 		domain = PCPU_GET(domain);
3329 	else
3330 		domain = UMA_ANYDOMAIN;
3331 	return (zone_alloc_item(zone, udata, domain, flags));
3332 }
3333 
3334 /* See uma.h */
3335 void *
3336 uma_zalloc_smr(uma_zone_t zone, int flags)
3337 {
3338 	uma_cache_bucket_t bucket;
3339 	uma_cache_t cache;
3340 
3341 #ifdef UMA_ZALLOC_DEBUG
3342 	void *item;
3343 
3344 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
3345 	    ("uma_zalloc_arg: called with non-SMR zone."));
3346 	if (uma_zalloc_debug(zone, &item, NULL, flags) == EJUSTRETURN)
3347 		return (item);
3348 #endif
3349 
3350 	critical_enter();
3351 	cache = &zone->uz_cpu[curcpu];
3352 	bucket = &cache->uc_allocbucket;
3353 	if (__predict_false(bucket->ucb_cnt == 0))
3354 		return (cache_alloc_retry(zone, cache, NULL, flags));
3355 	return (cache_alloc_item(zone, cache, bucket, NULL, flags));
3356 }
3357 
3358 /* See uma.h */
3359 void *
3360 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
3361 {
3362 	uma_cache_bucket_t bucket;
3363 	uma_cache_t cache;
3364 
3365 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3366 	random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3367 
3368 	/* This is the fast path allocation */
3369 	CTR3(KTR_UMA, "uma_zalloc_arg zone %s(%p) flags %d", zone->uz_name,
3370 	    zone, flags);
3371 
3372 #ifdef UMA_ZALLOC_DEBUG
3373 	void *item;
3374 
3375 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3376 	    ("uma_zalloc_arg: called with SMR zone."));
3377 	if (uma_zalloc_debug(zone, &item, udata, flags) == EJUSTRETURN)
3378 		return (item);
3379 #endif
3380 
3381 	/*
3382 	 * If possible, allocate from the per-CPU cache.  There are two
3383 	 * requirements for safe access to the per-CPU cache: (1) the thread
3384 	 * accessing the cache must not be preempted or yield during access,
3385 	 * and (2) the thread must not migrate CPUs without switching which
3386 	 * cache it accesses.  We rely on a critical section to prevent
3387 	 * preemption and migration.  We release the critical section in
3388 	 * order to acquire the zone mutex if we are unable to allocate from
3389 	 * the current cache; when we re-acquire the critical section, we
3390 	 * must detect and handle migration if it has occurred.
3391 	 */
3392 	critical_enter();
3393 	cache = &zone->uz_cpu[curcpu];
3394 	bucket = &cache->uc_allocbucket;
3395 	if (__predict_false(bucket->ucb_cnt == 0))
3396 		return (cache_alloc_retry(zone, cache, udata, flags));
3397 	return (cache_alloc_item(zone, cache, bucket, udata, flags));
3398 }
3399 
3400 /*
3401  * Replenish an alloc bucket and possibly restore an old one.  Called in
3402  * a critical section.  Returns in a critical section.
3403  *
3404  * A false return value indicates an allocation failure.
3405  * A true return value indicates success and the caller should retry.
3406  */
3407 static __noinline bool
3408 cache_alloc(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3409 {
3410 	uma_bucket_t bucket;
3411 	int curdomain, domain;
3412 	bool new;
3413 
3414 	CRITICAL_ASSERT(curthread);
3415 
3416 	/*
3417 	 * If we have run out of items in our alloc bucket see
3418 	 * if we can switch with the free bucket.
3419 	 *
3420 	 * SMR Zones can't re-use the free bucket until the sequence has
3421 	 * expired.
3422 	 */
3423 	if ((cache_uz_flags(cache) & UMA_ZONE_SMR) == 0 &&
3424 	    cache->uc_freebucket.ucb_cnt != 0) {
3425 		cache_bucket_swap(&cache->uc_freebucket,
3426 		    &cache->uc_allocbucket);
3427 		return (true);
3428 	}
3429 
3430 	/*
3431 	 * Discard any empty allocation bucket while we hold no locks.
3432 	 */
3433 	bucket = cache_bucket_unload_alloc(cache);
3434 	critical_exit();
3435 
3436 	if (bucket != NULL) {
3437 		KASSERT(bucket->ub_cnt == 0,
3438 		    ("cache_alloc: Entered with non-empty alloc bucket."));
3439 		bucket_free(zone, bucket, udata);
3440 	}
3441 
3442 	/*
3443 	 * Attempt to retrieve the item from the per-CPU cache has failed, so
3444 	 * we must go back to the zone.  This requires the zdom lock, so we
3445 	 * must drop the critical section, then re-acquire it when we go back
3446 	 * to the cache.  Since the critical section is released, we may be
3447 	 * preempted or migrate.  As such, make sure not to maintain any
3448 	 * thread-local state specific to the cache from prior to releasing
3449 	 * the critical section.
3450 	 */
3451 	domain = PCPU_GET(domain);
3452 	if ((cache_uz_flags(cache) & UMA_ZONE_ROUNDROBIN) != 0 ||
3453 	    VM_DOMAIN_EMPTY(domain))
3454 		domain = zone_domain_highest(zone, domain);
3455 	bucket = cache_fetch_bucket(zone, cache, domain);
3456 	if (bucket == NULL && zone->uz_bucket_size != 0 && !bucketdisable) {
3457 		bucket = zone_alloc_bucket(zone, udata, domain, flags);
3458 		new = true;
3459 	} else {
3460 		new = false;
3461 	}
3462 
3463 	CTR3(KTR_UMA, "uma_zalloc: zone %s(%p) bucket zone returned %p",
3464 	    zone->uz_name, zone, bucket);
3465 	if (bucket == NULL) {
3466 		critical_enter();
3467 		return (false);
3468 	}
3469 
3470 	/*
3471 	 * See if we lost the race or were migrated.  Cache the
3472 	 * initialized bucket to make this less likely or claim
3473 	 * the memory directly.
3474 	 */
3475 	critical_enter();
3476 	cache = &zone->uz_cpu[curcpu];
3477 	if (cache->uc_allocbucket.ucb_bucket == NULL &&
3478 	    ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) == 0 ||
3479 	    (curdomain = PCPU_GET(domain)) == domain ||
3480 	    VM_DOMAIN_EMPTY(curdomain))) {
3481 		if (new)
3482 			atomic_add_long(&ZDOM_GET(zone, domain)->uzd_imax,
3483 			    bucket->ub_cnt);
3484 		cache_bucket_load_alloc(cache, bucket);
3485 		return (true);
3486 	}
3487 
3488 	/*
3489 	 * We lost the race, release this bucket and start over.
3490 	 */
3491 	critical_exit();
3492 	zone_put_bucket(zone, domain, bucket, udata, false);
3493 	critical_enter();
3494 
3495 	return (true);
3496 }
3497 
3498 void *
3499 uma_zalloc_domain(uma_zone_t zone, void *udata, int domain, int flags)
3500 {
3501 #ifdef NUMA
3502 	uma_bucket_t bucket;
3503 	uma_zone_domain_t zdom;
3504 	void *item;
3505 #endif
3506 
3507 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3508 	random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3509 
3510 	/* This is the fast path allocation */
3511 	CTR4(KTR_UMA, "uma_zalloc_domain zone %s(%p) domain %d flags %d",
3512 	    zone->uz_name, zone, domain, flags);
3513 
3514 	if (flags & M_WAITOK) {
3515 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
3516 		    "uma_zalloc_domain: zone \"%s\"", zone->uz_name);
3517 	}
3518 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3519 	    ("uma_zalloc_domain: called with spinlock or critical section held"));
3520 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3521 	    ("uma_zalloc_domain: called with SMR zone."));
3522 #ifdef NUMA
3523 	KASSERT((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0,
3524 	    ("uma_zalloc_domain: called with non-FIRSTTOUCH zone."));
3525 
3526 	if (vm_ndomains == 1)
3527 		return (uma_zalloc_arg(zone, udata, flags));
3528 
3529 	/*
3530 	 * Try to allocate from the bucket cache before falling back to the keg.
3531 	 * We could try harder and attempt to allocate from per-CPU caches or
3532 	 * the per-domain cross-domain buckets, but the complexity is probably
3533 	 * not worth it.  It is more important that frees of previous
3534 	 * cross-domain allocations do not blow up the cache.
3535 	 */
3536 	zdom = zone_domain_lock(zone, domain);
3537 	if ((bucket = zone_fetch_bucket(zone, zdom, false)) != NULL) {
3538 		item = bucket->ub_bucket[bucket->ub_cnt - 1];
3539 #ifdef INVARIANTS
3540 		bucket->ub_bucket[bucket->ub_cnt - 1] = NULL;
3541 #endif
3542 		bucket->ub_cnt--;
3543 		zone_put_bucket(zone, domain, bucket, udata, true);
3544 		item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata,
3545 		    flags, item);
3546 		if (item != NULL) {
3547 			KASSERT(item_domain(item) == domain,
3548 			    ("%s: bucket cache item %p from wrong domain",
3549 			    __func__, item));
3550 			counter_u64_add(zone->uz_allocs, 1);
3551 		}
3552 		return (item);
3553 	}
3554 	ZDOM_UNLOCK(zdom);
3555 	return (zone_alloc_item(zone, udata, domain, flags));
3556 #else
3557 	return (uma_zalloc_arg(zone, udata, flags));
3558 #endif
3559 }
3560 
3561 /*
3562  * Find a slab with some space.  Prefer slabs that are partially used over those
3563  * that are totally full.  This helps to reduce fragmentation.
3564  *
3565  * If 'rr' is 1, search all domains starting from 'domain'.  Otherwise check
3566  * only 'domain'.
3567  */
3568 static uma_slab_t
3569 keg_first_slab(uma_keg_t keg, int domain, bool rr)
3570 {
3571 	uma_domain_t dom;
3572 	uma_slab_t slab;
3573 	int start;
3574 
3575 	KASSERT(domain >= 0 && domain < vm_ndomains,
3576 	    ("keg_first_slab: domain %d out of range", domain));
3577 	KEG_LOCK_ASSERT(keg, domain);
3578 
3579 	slab = NULL;
3580 	start = domain;
3581 	do {
3582 		dom = &keg->uk_domain[domain];
3583 		if ((slab = LIST_FIRST(&dom->ud_part_slab)) != NULL)
3584 			return (slab);
3585 		if ((slab = LIST_FIRST(&dom->ud_free_slab)) != NULL) {
3586 			LIST_REMOVE(slab, us_link);
3587 			dom->ud_free_slabs--;
3588 			LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
3589 			return (slab);
3590 		}
3591 		if (rr)
3592 			domain = (domain + 1) % vm_ndomains;
3593 	} while (domain != start);
3594 
3595 	return (NULL);
3596 }
3597 
3598 /*
3599  * Fetch an existing slab from a free or partial list.  Returns with the
3600  * keg domain lock held if a slab was found or unlocked if not.
3601  */
3602 static uma_slab_t
3603 keg_fetch_free_slab(uma_keg_t keg, int domain, bool rr, int flags)
3604 {
3605 	uma_slab_t slab;
3606 	uint32_t reserve;
3607 
3608 	/* HASH has a single free list. */
3609 	if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
3610 		domain = 0;
3611 
3612 	KEG_LOCK(keg, domain);
3613 	reserve = (flags & M_USE_RESERVE) != 0 ? 0 : keg->uk_reserve;
3614 	if (keg->uk_domain[domain].ud_free_items <= reserve ||
3615 	    (slab = keg_first_slab(keg, domain, rr)) == NULL) {
3616 		KEG_UNLOCK(keg, domain);
3617 		return (NULL);
3618 	}
3619 	return (slab);
3620 }
3621 
3622 static uma_slab_t
3623 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int rdomain, const int flags)
3624 {
3625 	struct vm_domainset_iter di;
3626 	uma_slab_t slab;
3627 	int aflags, domain;
3628 	bool rr;
3629 
3630 restart:
3631 	/*
3632 	 * Use the keg's policy if upper layers haven't already specified a
3633 	 * domain (as happens with first-touch zones).
3634 	 *
3635 	 * To avoid races we run the iterator with the keg lock held, but that
3636 	 * means that we cannot allow the vm_domainset layer to sleep.  Thus,
3637 	 * clear M_WAITOK and handle low memory conditions locally.
3638 	 */
3639 	rr = rdomain == UMA_ANYDOMAIN;
3640 	if (rr) {
3641 		aflags = (flags & ~M_WAITOK) | M_NOWAIT;
3642 		vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
3643 		    &aflags);
3644 	} else {
3645 		aflags = flags;
3646 		domain = rdomain;
3647 	}
3648 
3649 	for (;;) {
3650 		slab = keg_fetch_free_slab(keg, domain, rr, flags);
3651 		if (slab != NULL)
3652 			return (slab);
3653 
3654 		/*
3655 		 * M_NOVM means don't ask at all!
3656 		 */
3657 		if (flags & M_NOVM)
3658 			break;
3659 
3660 		slab = keg_alloc_slab(keg, zone, domain, flags, aflags);
3661 		if (slab != NULL)
3662 			return (slab);
3663 		if (!rr && (flags & M_WAITOK) == 0)
3664 			break;
3665 		if (rr && vm_domainset_iter_policy(&di, &domain) != 0) {
3666 			if ((flags & M_WAITOK) != 0) {
3667 				vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask, 0);
3668 				goto restart;
3669 			}
3670 			break;
3671 		}
3672 	}
3673 
3674 	/*
3675 	 * We might not have been able to get a slab but another cpu
3676 	 * could have while we were unlocked.  Check again before we
3677 	 * fail.
3678 	 */
3679 	if ((slab = keg_fetch_free_slab(keg, domain, rr, flags)) != NULL)
3680 		return (slab);
3681 
3682 	return (NULL);
3683 }
3684 
3685 static void *
3686 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
3687 {
3688 	uma_domain_t dom;
3689 	void *item;
3690 	int freei;
3691 
3692 	KEG_LOCK_ASSERT(keg, slab->us_domain);
3693 
3694 	dom = &keg->uk_domain[slab->us_domain];
3695 	freei = BIT_FFS(keg->uk_ipers, &slab->us_free) - 1;
3696 	BIT_CLR(keg->uk_ipers, freei, &slab->us_free);
3697 	item = slab_item(slab, keg, freei);
3698 	slab->us_freecount--;
3699 	dom->ud_free_items--;
3700 
3701 	/*
3702 	 * Move this slab to the full list.  It must be on the partial list, so
3703 	 * we do not need to update the free slab count.  In particular,
3704 	 * keg_fetch_slab() always returns slabs on the partial list.
3705 	 */
3706 	if (slab->us_freecount == 0) {
3707 		LIST_REMOVE(slab, us_link);
3708 		LIST_INSERT_HEAD(&dom->ud_full_slab, slab, us_link);
3709 	}
3710 
3711 	return (item);
3712 }
3713 
3714 static int
3715 zone_import(void *arg, void **bucket, int max, int domain, int flags)
3716 {
3717 	uma_domain_t dom;
3718 	uma_zone_t zone;
3719 	uma_slab_t slab;
3720 	uma_keg_t keg;
3721 #ifdef NUMA
3722 	int stripe;
3723 #endif
3724 	int i;
3725 
3726 	zone = arg;
3727 	slab = NULL;
3728 	keg = zone->uz_keg;
3729 	/* Try to keep the buckets totally full */
3730 	for (i = 0; i < max; ) {
3731 		if ((slab = keg_fetch_slab(keg, zone, domain, flags)) == NULL)
3732 			break;
3733 #ifdef NUMA
3734 		stripe = howmany(max, vm_ndomains);
3735 #endif
3736 		dom = &keg->uk_domain[slab->us_domain];
3737 		while (slab->us_freecount && i < max) {
3738 			bucket[i++] = slab_alloc_item(keg, slab);
3739 			if (dom->ud_free_items <= keg->uk_reserve)
3740 				break;
3741 #ifdef NUMA
3742 			/*
3743 			 * If the zone is striped we pick a new slab for every
3744 			 * N allocations.  Eliminating this conditional will
3745 			 * instead pick a new domain for each bucket rather
3746 			 * than stripe within each bucket.  The current option
3747 			 * produces more fragmentation and requires more cpu
3748 			 * time but yields better distribution.
3749 			 */
3750 			if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0 &&
3751 			    vm_ndomains > 1 && --stripe == 0)
3752 				break;
3753 #endif
3754 		}
3755 		KEG_UNLOCK(keg, slab->us_domain);
3756 		/* Don't block if we allocated any successfully. */
3757 		flags &= ~M_WAITOK;
3758 		flags |= M_NOWAIT;
3759 	}
3760 
3761 	return i;
3762 }
3763 
3764 static int
3765 zone_alloc_limit_hard(uma_zone_t zone, int count, int flags)
3766 {
3767 	uint64_t old, new, total, max;
3768 
3769 	/*
3770 	 * The hard case.  We're going to sleep because there were existing
3771 	 * sleepers or because we ran out of items.  This routine enforces
3772 	 * fairness by keeping fifo order.
3773 	 *
3774 	 * First release our ill gotten gains and make some noise.
3775 	 */
3776 	for (;;) {
3777 		zone_free_limit(zone, count);
3778 		zone_log_warning(zone);
3779 		zone_maxaction(zone);
3780 		if (flags & M_NOWAIT)
3781 			return (0);
3782 
3783 		/*
3784 		 * We need to allocate an item or set ourself as a sleeper
3785 		 * while the sleepq lock is held to avoid wakeup races.  This
3786 		 * is essentially a home rolled semaphore.
3787 		 */
3788 		sleepq_lock(&zone->uz_max_items);
3789 		old = zone->uz_items;
3790 		do {
3791 			MPASS(UZ_ITEMS_SLEEPERS(old) < UZ_ITEMS_SLEEPERS_MAX);
3792 			/* Cache the max since we will evaluate twice. */
3793 			max = zone->uz_max_items;
3794 			if (UZ_ITEMS_SLEEPERS(old) != 0 ||
3795 			    UZ_ITEMS_COUNT(old) >= max)
3796 				new = old + UZ_ITEMS_SLEEPER;
3797 			else
3798 				new = old + MIN(count, max - old);
3799 		} while (atomic_fcmpset_64(&zone->uz_items, &old, new) == 0);
3800 
3801 		/* We may have successfully allocated under the sleepq lock. */
3802 		if (UZ_ITEMS_SLEEPERS(new) == 0) {
3803 			sleepq_release(&zone->uz_max_items);
3804 			return (new - old);
3805 		}
3806 
3807 		/*
3808 		 * This is in a different cacheline from uz_items so that we
3809 		 * don't constantly invalidate the fastpath cacheline when we
3810 		 * adjust item counts.  This could be limited to toggling on
3811 		 * transitions.
3812 		 */
3813 		atomic_add_32(&zone->uz_sleepers, 1);
3814 		atomic_add_64(&zone->uz_sleeps, 1);
3815 
3816 		/*
3817 		 * We have added ourselves as a sleeper.  The sleepq lock
3818 		 * protects us from wakeup races.  Sleep now and then retry.
3819 		 */
3820 		sleepq_add(&zone->uz_max_items, NULL, "zonelimit", 0, 0);
3821 		sleepq_wait(&zone->uz_max_items, PVM);
3822 
3823 		/*
3824 		 * After wakeup, remove ourselves as a sleeper and try
3825 		 * again.  We no longer have the sleepq lock for protection.
3826 		 *
3827 		 * Subract ourselves as a sleeper while attempting to add
3828 		 * our count.
3829 		 */
3830 		atomic_subtract_32(&zone->uz_sleepers, 1);
3831 		old = atomic_fetchadd_64(&zone->uz_items,
3832 		    -(UZ_ITEMS_SLEEPER - count));
3833 		/* We're no longer a sleeper. */
3834 		old -= UZ_ITEMS_SLEEPER;
3835 
3836 		/*
3837 		 * If we're still at the limit, restart.  Notably do not
3838 		 * block on other sleepers.  Cache the max value to protect
3839 		 * against changes via sysctl.
3840 		 */
3841 		total = UZ_ITEMS_COUNT(old);
3842 		max = zone->uz_max_items;
3843 		if (total >= max)
3844 			continue;
3845 		/* Truncate if necessary, otherwise wake other sleepers. */
3846 		if (total + count > max) {
3847 			zone_free_limit(zone, total + count - max);
3848 			count = max - total;
3849 		} else if (total + count < max && UZ_ITEMS_SLEEPERS(old) != 0)
3850 			wakeup_one(&zone->uz_max_items);
3851 
3852 		return (count);
3853 	}
3854 }
3855 
3856 /*
3857  * Allocate 'count' items from our max_items limit.  Returns the number
3858  * available.  If M_NOWAIT is not specified it will sleep until at least
3859  * one item can be allocated.
3860  */
3861 static int
3862 zone_alloc_limit(uma_zone_t zone, int count, int flags)
3863 {
3864 	uint64_t old;
3865 	uint64_t max;
3866 
3867 	max = zone->uz_max_items;
3868 	MPASS(max > 0);
3869 
3870 	/*
3871 	 * We expect normal allocations to succeed with a simple
3872 	 * fetchadd.
3873 	 */
3874 	old = atomic_fetchadd_64(&zone->uz_items, count);
3875 	if (__predict_true(old + count <= max))
3876 		return (count);
3877 
3878 	/*
3879 	 * If we had some items and no sleepers just return the
3880 	 * truncated value.  We have to release the excess space
3881 	 * though because that may wake sleepers who weren't woken
3882 	 * because we were temporarily over the limit.
3883 	 */
3884 	if (old < max) {
3885 		zone_free_limit(zone, (old + count) - max);
3886 		return (max - old);
3887 	}
3888 	return (zone_alloc_limit_hard(zone, count, flags));
3889 }
3890 
3891 /*
3892  * Free a number of items back to the limit.
3893  */
3894 static void
3895 zone_free_limit(uma_zone_t zone, int count)
3896 {
3897 	uint64_t old;
3898 
3899 	MPASS(count > 0);
3900 
3901 	/*
3902 	 * In the common case we either have no sleepers or
3903 	 * are still over the limit and can just return.
3904 	 */
3905 	old = atomic_fetchadd_64(&zone->uz_items, -count);
3906 	if (__predict_true(UZ_ITEMS_SLEEPERS(old) == 0 ||
3907 	   UZ_ITEMS_COUNT(old) - count >= zone->uz_max_items))
3908 		return;
3909 
3910 	/*
3911 	 * Moderate the rate of wakeups.  Sleepers will continue
3912 	 * to generate wakeups if necessary.
3913 	 */
3914 	wakeup_one(&zone->uz_max_items);
3915 }
3916 
3917 static uma_bucket_t
3918 zone_alloc_bucket(uma_zone_t zone, void *udata, int domain, int flags)
3919 {
3920 	uma_bucket_t bucket;
3921 	int maxbucket, cnt;
3922 
3923 	CTR3(KTR_UMA, "zone_alloc_bucket zone %s(%p) domain %d", zone->uz_name,
3924 	    zone, domain);
3925 
3926 	/* Avoid allocs targeting empty domains. */
3927 	if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
3928 		domain = UMA_ANYDOMAIN;
3929 	else if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
3930 		domain = UMA_ANYDOMAIN;
3931 
3932 	if (zone->uz_max_items > 0)
3933 		maxbucket = zone_alloc_limit(zone, zone->uz_bucket_size,
3934 		    M_NOWAIT);
3935 	else
3936 		maxbucket = zone->uz_bucket_size;
3937 	if (maxbucket == 0)
3938 		return (false);
3939 
3940 	/* Don't wait for buckets, preserve caller's NOVM setting. */
3941 	bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
3942 	if (bucket == NULL) {
3943 		cnt = 0;
3944 		goto out;
3945 	}
3946 
3947 	bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
3948 	    MIN(maxbucket, bucket->ub_entries), domain, flags);
3949 
3950 	/*
3951 	 * Initialize the memory if necessary.
3952 	 */
3953 	if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
3954 		int i;
3955 
3956 		for (i = 0; i < bucket->ub_cnt; i++)
3957 			if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size,
3958 			    flags) != 0)
3959 				break;
3960 		/*
3961 		 * If we couldn't initialize the whole bucket, put the
3962 		 * rest back onto the freelist.
3963 		 */
3964 		if (i != bucket->ub_cnt) {
3965 			zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
3966 			    bucket->ub_cnt - i);
3967 #ifdef INVARIANTS
3968 			bzero(&bucket->ub_bucket[i],
3969 			    sizeof(void *) * (bucket->ub_cnt - i));
3970 #endif
3971 			bucket->ub_cnt = i;
3972 		}
3973 	}
3974 
3975 	cnt = bucket->ub_cnt;
3976 	if (bucket->ub_cnt == 0) {
3977 		bucket_free(zone, bucket, udata);
3978 		counter_u64_add(zone->uz_fails, 1);
3979 		bucket = NULL;
3980 	}
3981 out:
3982 	if (zone->uz_max_items > 0 && cnt < maxbucket)
3983 		zone_free_limit(zone, maxbucket - cnt);
3984 
3985 	return (bucket);
3986 }
3987 
3988 /*
3989  * Allocates a single item from a zone.
3990  *
3991  * Arguments
3992  *	zone   The zone to alloc for.
3993  *	udata  The data to be passed to the constructor.
3994  *	domain The domain to allocate from or UMA_ANYDOMAIN.
3995  *	flags  M_WAITOK, M_NOWAIT, M_ZERO.
3996  *
3997  * Returns
3998  *	NULL if there is no memory and M_NOWAIT is set
3999  *	An item if successful
4000  */
4001 
4002 static void *
4003 zone_alloc_item(uma_zone_t zone, void *udata, int domain, int flags)
4004 {
4005 	void *item;
4006 
4007 	if (zone->uz_max_items > 0 && zone_alloc_limit(zone, 1, flags) == 0) {
4008 		counter_u64_add(zone->uz_fails, 1);
4009 		return (NULL);
4010 	}
4011 
4012 	/* Avoid allocs targeting empty domains. */
4013 	if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
4014 		domain = UMA_ANYDOMAIN;
4015 
4016 	if (zone->uz_import(zone->uz_arg, &item, 1, domain, flags) != 1)
4017 		goto fail_cnt;
4018 
4019 	/*
4020 	 * We have to call both the zone's init (not the keg's init)
4021 	 * and the zone's ctor.  This is because the item is going from
4022 	 * a keg slab directly to the user, and the user is expecting it
4023 	 * to be both zone-init'd as well as zone-ctor'd.
4024 	 */
4025 	if (zone->uz_init != NULL) {
4026 		if (zone->uz_init(item, zone->uz_size, flags) != 0) {
4027 			zone_free_item(zone, item, udata, SKIP_FINI | SKIP_CNT);
4028 			goto fail_cnt;
4029 		}
4030 	}
4031 	item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata, flags,
4032 	    item);
4033 	if (item == NULL)
4034 		goto fail;
4035 
4036 	counter_u64_add(zone->uz_allocs, 1);
4037 	CTR3(KTR_UMA, "zone_alloc_item item %p from %s(%p)", item,
4038 	    zone->uz_name, zone);
4039 
4040 	return (item);
4041 
4042 fail_cnt:
4043 	counter_u64_add(zone->uz_fails, 1);
4044 fail:
4045 	if (zone->uz_max_items > 0)
4046 		zone_free_limit(zone, 1);
4047 	CTR2(KTR_UMA, "zone_alloc_item failed from %s(%p)",
4048 	    zone->uz_name, zone);
4049 
4050 	return (NULL);
4051 }
4052 
4053 /* See uma.h */
4054 void
4055 uma_zfree_smr(uma_zone_t zone, void *item)
4056 {
4057 	uma_cache_t cache;
4058 	uma_cache_bucket_t bucket;
4059 	int itemdomain, uz_flags;
4060 
4061 #ifdef UMA_ZALLOC_DEBUG
4062 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
4063 	    ("uma_zfree_smr: called with non-SMR zone."));
4064 	KASSERT(item != NULL, ("uma_zfree_smr: Called with NULL pointer."));
4065 	SMR_ASSERT_NOT_ENTERED(zone->uz_smr);
4066 	if (uma_zfree_debug(zone, item, NULL) == EJUSTRETURN)
4067 		return;
4068 #endif
4069 	cache = &zone->uz_cpu[curcpu];
4070 	uz_flags = cache_uz_flags(cache);
4071 	itemdomain = 0;
4072 #ifdef NUMA
4073 	if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4074 		itemdomain = item_domain(item);
4075 #endif
4076 	critical_enter();
4077 	do {
4078 		cache = &zone->uz_cpu[curcpu];
4079 		/* SMR Zones must free to the free bucket. */
4080 		bucket = &cache->uc_freebucket;
4081 #ifdef NUMA
4082 		if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4083 		    PCPU_GET(domain) != itemdomain) {
4084 			bucket = &cache->uc_crossbucket;
4085 		}
4086 #endif
4087 		if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4088 			cache_bucket_push(cache, bucket, item);
4089 			critical_exit();
4090 			return;
4091 		}
4092 	} while (cache_free(zone, cache, NULL, item, itemdomain));
4093 	critical_exit();
4094 
4095 	/*
4096 	 * If nothing else caught this, we'll just do an internal free.
4097 	 */
4098 	zone_free_item(zone, item, NULL, SKIP_NONE);
4099 }
4100 
4101 /* See uma.h */
4102 void
4103 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
4104 {
4105 	uma_cache_t cache;
4106 	uma_cache_bucket_t bucket;
4107 	int itemdomain, uz_flags;
4108 
4109 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
4110 	random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
4111 
4112 	CTR2(KTR_UMA, "uma_zfree_arg zone %s(%p)", zone->uz_name, zone);
4113 
4114 #ifdef UMA_ZALLOC_DEBUG
4115 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
4116 	    ("uma_zfree_arg: called with SMR zone."));
4117 	if (uma_zfree_debug(zone, item, udata) == EJUSTRETURN)
4118 		return;
4119 #endif
4120         /* uma_zfree(..., NULL) does nothing, to match free(9). */
4121         if (item == NULL)
4122                 return;
4123 
4124 	/*
4125 	 * We are accessing the per-cpu cache without a critical section to
4126 	 * fetch size and flags.  This is acceptable, if we are preempted we
4127 	 * will simply read another cpu's line.
4128 	 */
4129 	cache = &zone->uz_cpu[curcpu];
4130 	uz_flags = cache_uz_flags(cache);
4131 	if (UMA_ALWAYS_CTORDTOR ||
4132 	    __predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0))
4133 		item_dtor(zone, item, cache_uz_size(cache), udata, SKIP_NONE);
4134 
4135 	/*
4136 	 * The race here is acceptable.  If we miss it we'll just have to wait
4137 	 * a little longer for the limits to be reset.
4138 	 */
4139 	if (__predict_false(uz_flags & UMA_ZFLAG_LIMIT)) {
4140 		if (zone->uz_sleepers > 0)
4141 			goto zfree_item;
4142 	}
4143 
4144 	/*
4145 	 * If possible, free to the per-CPU cache.  There are two
4146 	 * requirements for safe access to the per-CPU cache: (1) the thread
4147 	 * accessing the cache must not be preempted or yield during access,
4148 	 * and (2) the thread must not migrate CPUs without switching which
4149 	 * cache it accesses.  We rely on a critical section to prevent
4150 	 * preemption and migration.  We release the critical section in
4151 	 * order to acquire the zone mutex if we are unable to free to the
4152 	 * current cache; when we re-acquire the critical section, we must
4153 	 * detect and handle migration if it has occurred.
4154 	 */
4155 	itemdomain = 0;
4156 #ifdef NUMA
4157 	if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4158 		itemdomain = item_domain(item);
4159 #endif
4160 	critical_enter();
4161 	do {
4162 		cache = &zone->uz_cpu[curcpu];
4163 		/*
4164 		 * Try to free into the allocbucket first to give LIFO
4165 		 * ordering for cache-hot datastructures.  Spill over
4166 		 * into the freebucket if necessary.  Alloc will swap
4167 		 * them if one runs dry.
4168 		 */
4169 		bucket = &cache->uc_allocbucket;
4170 #ifdef NUMA
4171 		if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4172 		    PCPU_GET(domain) != itemdomain) {
4173 			bucket = &cache->uc_crossbucket;
4174 		} else
4175 #endif
4176 		if (bucket->ucb_cnt == bucket->ucb_entries &&
4177 		   cache->uc_freebucket.ucb_cnt <
4178 		   cache->uc_freebucket.ucb_entries)
4179 			cache_bucket_swap(&cache->uc_freebucket,
4180 			    &cache->uc_allocbucket);
4181 		if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4182 			cache_bucket_push(cache, bucket, item);
4183 			critical_exit();
4184 			return;
4185 		}
4186 	} while (cache_free(zone, cache, udata, item, itemdomain));
4187 	critical_exit();
4188 
4189 	/*
4190 	 * If nothing else caught this, we'll just do an internal free.
4191 	 */
4192 zfree_item:
4193 	zone_free_item(zone, item, udata, SKIP_DTOR);
4194 }
4195 
4196 #ifdef NUMA
4197 /*
4198  * sort crossdomain free buckets to domain correct buckets and cache
4199  * them.
4200  */
4201 static void
4202 zone_free_cross(uma_zone_t zone, uma_bucket_t bucket, void *udata)
4203 {
4204 	struct uma_bucketlist fullbuckets;
4205 	uma_zone_domain_t zdom;
4206 	uma_bucket_t b;
4207 	smr_seq_t seq;
4208 	void *item;
4209 	int domain;
4210 
4211 	CTR3(KTR_UMA,
4212 	    "uma_zfree: zone %s(%p) draining cross bucket %p",
4213 	    zone->uz_name, zone, bucket);
4214 
4215 	/*
4216 	 * It is possible for buckets to arrive here out of order so we fetch
4217 	 * the current smr seq rather than accepting the bucket's.
4218 	 */
4219 	seq = SMR_SEQ_INVALID;
4220 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
4221 		seq = smr_advance(zone->uz_smr);
4222 
4223 	/*
4224 	 * To avoid having ndomain * ndomain buckets for sorting we have a
4225 	 * lock on the current crossfree bucket.  A full matrix with
4226 	 * per-domain locking could be used if necessary.
4227 	 */
4228 	STAILQ_INIT(&fullbuckets);
4229 	ZONE_CROSS_LOCK(zone);
4230 	while (bucket->ub_cnt > 0) {
4231 		item = bucket->ub_bucket[bucket->ub_cnt - 1];
4232 		domain = item_domain(item);
4233 		zdom = ZDOM_GET(zone, domain);
4234 		if (zdom->uzd_cross == NULL) {
4235 			zdom->uzd_cross = bucket_alloc(zone, udata, M_NOWAIT);
4236 			if (zdom->uzd_cross == NULL)
4237 				break;
4238 		}
4239 		b = zdom->uzd_cross;
4240 		b->ub_bucket[b->ub_cnt++] = item;
4241 		b->ub_seq = seq;
4242 		if (b->ub_cnt == b->ub_entries) {
4243 			STAILQ_INSERT_HEAD(&fullbuckets, b, ub_link);
4244 			zdom->uzd_cross = NULL;
4245 		}
4246 		bucket->ub_cnt--;
4247 	}
4248 	ZONE_CROSS_UNLOCK(zone);
4249 	if (bucket->ub_cnt == 0)
4250 		bucket->ub_seq = SMR_SEQ_INVALID;
4251 	bucket_free(zone, bucket, udata);
4252 
4253 	while ((b = STAILQ_FIRST(&fullbuckets)) != NULL) {
4254 		STAILQ_REMOVE_HEAD(&fullbuckets, ub_link);
4255 		domain = item_domain(b->ub_bucket[0]);
4256 		zone_put_bucket(zone, domain, b, udata, true);
4257 	}
4258 }
4259 #endif
4260 
4261 static void
4262 zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata,
4263     int itemdomain, bool ws)
4264 {
4265 
4266 #ifdef NUMA
4267 	/*
4268 	 * Buckets coming from the wrong domain will be entirely for the
4269 	 * only other domain on two domain systems.  In this case we can
4270 	 * simply cache them.  Otherwise we need to sort them back to
4271 	 * correct domains.
4272 	 */
4273 	if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4274 	    vm_ndomains > 2 && PCPU_GET(domain) != itemdomain) {
4275 		zone_free_cross(zone, bucket, udata);
4276 		return;
4277 	}
4278 #endif
4279 
4280 	/*
4281 	 * Attempt to save the bucket in the zone's domain bucket cache.
4282 	 */
4283 	CTR3(KTR_UMA,
4284 	    "uma_zfree: zone %s(%p) putting bucket %p on free list",
4285 	    zone->uz_name, zone, bucket);
4286 	/* ub_cnt is pointing to the last free item */
4287 	if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
4288 		itemdomain = zone_domain_lowest(zone, itemdomain);
4289 	zone_put_bucket(zone, itemdomain, bucket, udata, ws);
4290 }
4291 
4292 /*
4293  * Populate a free or cross bucket for the current cpu cache.  Free any
4294  * existing full bucket either to the zone cache or back to the slab layer.
4295  *
4296  * Enters and returns in a critical section.  false return indicates that
4297  * we can not satisfy this free in the cache layer.  true indicates that
4298  * the caller should retry.
4299  */
4300 static __noinline bool
4301 cache_free(uma_zone_t zone, uma_cache_t cache, void *udata, void *item,
4302     int itemdomain)
4303 {
4304 	uma_cache_bucket_t cbucket;
4305 	uma_bucket_t newbucket, bucket;
4306 
4307 	CRITICAL_ASSERT(curthread);
4308 
4309 	if (zone->uz_bucket_size == 0)
4310 		return false;
4311 
4312 	cache = &zone->uz_cpu[curcpu];
4313 	newbucket = NULL;
4314 
4315 	/*
4316 	 * FIRSTTOUCH domains need to free to the correct zdom.  When
4317 	 * enabled this is the zdom of the item.   The bucket is the
4318 	 * cross bucket if the current domain and itemdomain do not match.
4319 	 */
4320 	cbucket = &cache->uc_freebucket;
4321 #ifdef NUMA
4322 	if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4323 		if (PCPU_GET(domain) != itemdomain) {
4324 			cbucket = &cache->uc_crossbucket;
4325 			if (cbucket->ucb_cnt != 0)
4326 				counter_u64_add(zone->uz_xdomain,
4327 				    cbucket->ucb_cnt);
4328 		}
4329 	}
4330 #endif
4331 	bucket = cache_bucket_unload(cbucket);
4332 	KASSERT(bucket == NULL || bucket->ub_cnt == bucket->ub_entries,
4333 	    ("cache_free: Entered with non-full free bucket."));
4334 
4335 	/* We are no longer associated with this CPU. */
4336 	critical_exit();
4337 
4338 	/*
4339 	 * Don't let SMR zones operate without a free bucket.  Force
4340 	 * a synchronize and re-use this one.  We will only degrade
4341 	 * to a synchronize every bucket_size items rather than every
4342 	 * item if we fail to allocate a bucket.
4343 	 */
4344 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0) {
4345 		if (bucket != NULL)
4346 			bucket->ub_seq = smr_advance(zone->uz_smr);
4347 		newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4348 		if (newbucket == NULL && bucket != NULL) {
4349 			bucket_drain(zone, bucket);
4350 			newbucket = bucket;
4351 			bucket = NULL;
4352 		}
4353 	} else if (!bucketdisable)
4354 		newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4355 
4356 	if (bucket != NULL)
4357 		zone_free_bucket(zone, bucket, udata, itemdomain, true);
4358 
4359 	critical_enter();
4360 	if ((bucket = newbucket) == NULL)
4361 		return (false);
4362 	cache = &zone->uz_cpu[curcpu];
4363 #ifdef NUMA
4364 	/*
4365 	 * Check to see if we should be populating the cross bucket.  If it
4366 	 * is already populated we will fall through and attempt to populate
4367 	 * the free bucket.
4368 	 */
4369 	if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4370 		if (PCPU_GET(domain) != itemdomain &&
4371 		    cache->uc_crossbucket.ucb_bucket == NULL) {
4372 			cache_bucket_load_cross(cache, bucket);
4373 			return (true);
4374 		}
4375 	}
4376 #endif
4377 	/*
4378 	 * We may have lost the race to fill the bucket or switched CPUs.
4379 	 */
4380 	if (cache->uc_freebucket.ucb_bucket != NULL) {
4381 		critical_exit();
4382 		bucket_free(zone, bucket, udata);
4383 		critical_enter();
4384 	} else
4385 		cache_bucket_load_free(cache, bucket);
4386 
4387 	return (true);
4388 }
4389 
4390 static void
4391 slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item)
4392 {
4393 	uma_keg_t keg;
4394 	uma_domain_t dom;
4395 	int freei;
4396 
4397 	keg = zone->uz_keg;
4398 	KEG_LOCK_ASSERT(keg, slab->us_domain);
4399 
4400 	/* Do we need to remove from any lists? */
4401 	dom = &keg->uk_domain[slab->us_domain];
4402 	if (slab->us_freecount + 1 == keg->uk_ipers) {
4403 		LIST_REMOVE(slab, us_link);
4404 		LIST_INSERT_HEAD(&dom->ud_free_slab, slab, us_link);
4405 		dom->ud_free_slabs++;
4406 	} else if (slab->us_freecount == 0) {
4407 		LIST_REMOVE(slab, us_link);
4408 		LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
4409 	}
4410 
4411 	/* Slab management. */
4412 	freei = slab_item_index(slab, keg, item);
4413 	BIT_SET(keg->uk_ipers, freei, &slab->us_free);
4414 	slab->us_freecount++;
4415 
4416 	/* Keg statistics. */
4417 	dom->ud_free_items++;
4418 }
4419 
4420 static void
4421 zone_release(void *arg, void **bucket, int cnt)
4422 {
4423 	struct mtx *lock;
4424 	uma_zone_t zone;
4425 	uma_slab_t slab;
4426 	uma_keg_t keg;
4427 	uint8_t *mem;
4428 	void *item;
4429 	int i;
4430 
4431 	zone = arg;
4432 	keg = zone->uz_keg;
4433 	lock = NULL;
4434 	if (__predict_false((zone->uz_flags & UMA_ZFLAG_HASH) != 0))
4435 		lock = KEG_LOCK(keg, 0);
4436 	for (i = 0; i < cnt; i++) {
4437 		item = bucket[i];
4438 		if (__predict_true((zone->uz_flags & UMA_ZFLAG_VTOSLAB) != 0)) {
4439 			slab = vtoslab((vm_offset_t)item);
4440 		} else {
4441 			mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
4442 			if ((zone->uz_flags & UMA_ZFLAG_HASH) != 0)
4443 				slab = hash_sfind(&keg->uk_hash, mem);
4444 			else
4445 				slab = (uma_slab_t)(mem + keg->uk_pgoff);
4446 		}
4447 		if (lock != KEG_LOCKPTR(keg, slab->us_domain)) {
4448 			if (lock != NULL)
4449 				mtx_unlock(lock);
4450 			lock = KEG_LOCK(keg, slab->us_domain);
4451 		}
4452 		slab_free_item(zone, slab, item);
4453 	}
4454 	if (lock != NULL)
4455 		mtx_unlock(lock);
4456 }
4457 
4458 /*
4459  * Frees a single item to any zone.
4460  *
4461  * Arguments:
4462  *	zone   The zone to free to
4463  *	item   The item we're freeing
4464  *	udata  User supplied data for the dtor
4465  *	skip   Skip dtors and finis
4466  */
4467 static __noinline void
4468 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
4469 {
4470 
4471 	/*
4472 	 * If a free is sent directly to an SMR zone we have to
4473 	 * synchronize immediately because the item can instantly
4474 	 * be reallocated. This should only happen in degenerate
4475 	 * cases when no memory is available for per-cpu caches.
4476 	 */
4477 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0 && skip == SKIP_NONE)
4478 		smr_synchronize(zone->uz_smr);
4479 
4480 	item_dtor(zone, item, zone->uz_size, udata, skip);
4481 
4482 	if (skip < SKIP_FINI && zone->uz_fini)
4483 		zone->uz_fini(item, zone->uz_size);
4484 
4485 	zone->uz_release(zone->uz_arg, &item, 1);
4486 
4487 	if (skip & SKIP_CNT)
4488 		return;
4489 
4490 	counter_u64_add(zone->uz_frees, 1);
4491 
4492 	if (zone->uz_max_items > 0)
4493 		zone_free_limit(zone, 1);
4494 }
4495 
4496 /* See uma.h */
4497 int
4498 uma_zone_set_max(uma_zone_t zone, int nitems)
4499 {
4500 	struct uma_bucket_zone *ubz;
4501 	int count;
4502 
4503 	/*
4504 	 * XXX This can misbehave if the zone has any allocations with
4505 	 * no limit and a limit is imposed.  There is currently no
4506 	 * way to clear a limit.
4507 	 */
4508 	ZONE_LOCK(zone);
4509 	ubz = bucket_zone_max(zone, nitems);
4510 	count = ubz != NULL ? ubz->ubz_entries : 0;
4511 	zone->uz_bucket_size_max = zone->uz_bucket_size = count;
4512 	if (zone->uz_bucket_size_min > zone->uz_bucket_size_max)
4513 		zone->uz_bucket_size_min = zone->uz_bucket_size_max;
4514 	zone->uz_max_items = nitems;
4515 	zone->uz_flags |= UMA_ZFLAG_LIMIT;
4516 	zone_update_caches(zone);
4517 	/* We may need to wake waiters. */
4518 	wakeup(&zone->uz_max_items);
4519 	ZONE_UNLOCK(zone);
4520 
4521 	return (nitems);
4522 }
4523 
4524 /* See uma.h */
4525 void
4526 uma_zone_set_maxcache(uma_zone_t zone, int nitems)
4527 {
4528 	struct uma_bucket_zone *ubz;
4529 	int bpcpu;
4530 
4531 	ZONE_LOCK(zone);
4532 	ubz = bucket_zone_max(zone, nitems);
4533 	if (ubz != NULL) {
4534 		bpcpu = 2;
4535 		if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4536 			/* Count the cross-domain bucket. */
4537 			bpcpu++;
4538 		nitems -= ubz->ubz_entries * bpcpu * mp_ncpus;
4539 		zone->uz_bucket_size_max = ubz->ubz_entries;
4540 	} else {
4541 		zone->uz_bucket_size_max = zone->uz_bucket_size = 0;
4542 	}
4543 	if (zone->uz_bucket_size_min > zone->uz_bucket_size_max)
4544 		zone->uz_bucket_size_min = zone->uz_bucket_size_max;
4545 	zone->uz_bucket_max = nitems / vm_ndomains;
4546 	ZONE_UNLOCK(zone);
4547 }
4548 
4549 /* See uma.h */
4550 int
4551 uma_zone_get_max(uma_zone_t zone)
4552 {
4553 	int nitems;
4554 
4555 	nitems = atomic_load_64(&zone->uz_max_items);
4556 
4557 	return (nitems);
4558 }
4559 
4560 /* See uma.h */
4561 void
4562 uma_zone_set_warning(uma_zone_t zone, const char *warning)
4563 {
4564 
4565 	ZONE_ASSERT_COLD(zone);
4566 	zone->uz_warning = warning;
4567 }
4568 
4569 /* See uma.h */
4570 void
4571 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction)
4572 {
4573 
4574 	ZONE_ASSERT_COLD(zone);
4575 	TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone);
4576 }
4577 
4578 /* See uma.h */
4579 int
4580 uma_zone_get_cur(uma_zone_t zone)
4581 {
4582 	int64_t nitems;
4583 	u_int i;
4584 
4585 	nitems = 0;
4586 	if (zone->uz_allocs != EARLY_COUNTER && zone->uz_frees != EARLY_COUNTER)
4587 		nitems = counter_u64_fetch(zone->uz_allocs) -
4588 		    counter_u64_fetch(zone->uz_frees);
4589 	CPU_FOREACH(i)
4590 		nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs) -
4591 		    atomic_load_64(&zone->uz_cpu[i].uc_frees);
4592 
4593 	return (nitems < 0 ? 0 : nitems);
4594 }
4595 
4596 static uint64_t
4597 uma_zone_get_allocs(uma_zone_t zone)
4598 {
4599 	uint64_t nitems;
4600 	u_int i;
4601 
4602 	nitems = 0;
4603 	if (zone->uz_allocs != EARLY_COUNTER)
4604 		nitems = counter_u64_fetch(zone->uz_allocs);
4605 	CPU_FOREACH(i)
4606 		nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs);
4607 
4608 	return (nitems);
4609 }
4610 
4611 static uint64_t
4612 uma_zone_get_frees(uma_zone_t zone)
4613 {
4614 	uint64_t nitems;
4615 	u_int i;
4616 
4617 	nitems = 0;
4618 	if (zone->uz_frees != EARLY_COUNTER)
4619 		nitems = counter_u64_fetch(zone->uz_frees);
4620 	CPU_FOREACH(i)
4621 		nitems += atomic_load_64(&zone->uz_cpu[i].uc_frees);
4622 
4623 	return (nitems);
4624 }
4625 
4626 #ifdef INVARIANTS
4627 /* Used only for KEG_ASSERT_COLD(). */
4628 static uint64_t
4629 uma_keg_get_allocs(uma_keg_t keg)
4630 {
4631 	uma_zone_t z;
4632 	uint64_t nitems;
4633 
4634 	nitems = 0;
4635 	LIST_FOREACH(z, &keg->uk_zones, uz_link)
4636 		nitems += uma_zone_get_allocs(z);
4637 
4638 	return (nitems);
4639 }
4640 #endif
4641 
4642 /* See uma.h */
4643 void
4644 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
4645 {
4646 	uma_keg_t keg;
4647 
4648 	KEG_GET(zone, keg);
4649 	KEG_ASSERT_COLD(keg);
4650 	keg->uk_init = uminit;
4651 }
4652 
4653 /* See uma.h */
4654 void
4655 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
4656 {
4657 	uma_keg_t keg;
4658 
4659 	KEG_GET(zone, keg);
4660 	KEG_ASSERT_COLD(keg);
4661 	keg->uk_fini = fini;
4662 }
4663 
4664 /* See uma.h */
4665 void
4666 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
4667 {
4668 
4669 	ZONE_ASSERT_COLD(zone);
4670 	zone->uz_init = zinit;
4671 }
4672 
4673 /* See uma.h */
4674 void
4675 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
4676 {
4677 
4678 	ZONE_ASSERT_COLD(zone);
4679 	zone->uz_fini = zfini;
4680 }
4681 
4682 /* See uma.h */
4683 void
4684 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
4685 {
4686 	uma_keg_t keg;
4687 
4688 	KEG_GET(zone, keg);
4689 	KEG_ASSERT_COLD(keg);
4690 	keg->uk_freef = freef;
4691 }
4692 
4693 /* See uma.h */
4694 void
4695 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
4696 {
4697 	uma_keg_t keg;
4698 
4699 	KEG_GET(zone, keg);
4700 	KEG_ASSERT_COLD(keg);
4701 	keg->uk_allocf = allocf;
4702 }
4703 
4704 /* See uma.h */
4705 void
4706 uma_zone_set_smr(uma_zone_t zone, smr_t smr)
4707 {
4708 
4709 	ZONE_ASSERT_COLD(zone);
4710 
4711 	KASSERT(smr != NULL, ("Got NULL smr"));
4712 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
4713 	    ("zone %p (%s) already uses SMR", zone, zone->uz_name));
4714 	zone->uz_flags |= UMA_ZONE_SMR;
4715 	zone->uz_smr = smr;
4716 	zone_update_caches(zone);
4717 }
4718 
4719 smr_t
4720 uma_zone_get_smr(uma_zone_t zone)
4721 {
4722 
4723 	return (zone->uz_smr);
4724 }
4725 
4726 /* See uma.h */
4727 void
4728 uma_zone_reserve(uma_zone_t zone, int items)
4729 {
4730 	uma_keg_t keg;
4731 
4732 	KEG_GET(zone, keg);
4733 	KEG_ASSERT_COLD(keg);
4734 	keg->uk_reserve = items;
4735 }
4736 
4737 /* See uma.h */
4738 int
4739 uma_zone_reserve_kva(uma_zone_t zone, int count)
4740 {
4741 	uma_keg_t keg;
4742 	vm_offset_t kva;
4743 	u_int pages;
4744 
4745 	KEG_GET(zone, keg);
4746 	KEG_ASSERT_COLD(keg);
4747 	ZONE_ASSERT_COLD(zone);
4748 
4749 	pages = howmany(count, keg->uk_ipers) * keg->uk_ppera;
4750 
4751 #ifdef UMA_MD_SMALL_ALLOC
4752 	if (keg->uk_ppera > 1) {
4753 #else
4754 	if (1) {
4755 #endif
4756 		kva = kva_alloc((vm_size_t)pages * PAGE_SIZE);
4757 		if (kva == 0)
4758 			return (0);
4759 	} else
4760 		kva = 0;
4761 
4762 	MPASS(keg->uk_kva == 0);
4763 	keg->uk_kva = kva;
4764 	keg->uk_offset = 0;
4765 	zone->uz_max_items = pages * keg->uk_ipers;
4766 #ifdef UMA_MD_SMALL_ALLOC
4767 	keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
4768 #else
4769 	keg->uk_allocf = noobj_alloc;
4770 #endif
4771 	keg->uk_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
4772 	zone->uz_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
4773 	zone_update_caches(zone);
4774 
4775 	return (1);
4776 }
4777 
4778 /* See uma.h */
4779 void
4780 uma_prealloc(uma_zone_t zone, int items)
4781 {
4782 	struct vm_domainset_iter di;
4783 	uma_domain_t dom;
4784 	uma_slab_t slab;
4785 	uma_keg_t keg;
4786 	int aflags, domain, slabs;
4787 
4788 	KEG_GET(zone, keg);
4789 	slabs = howmany(items, keg->uk_ipers);
4790 	while (slabs-- > 0) {
4791 		aflags = M_NOWAIT;
4792 		vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
4793 		    &aflags);
4794 		for (;;) {
4795 			slab = keg_alloc_slab(keg, zone, domain, M_WAITOK,
4796 			    aflags);
4797 			if (slab != NULL) {
4798 				dom = &keg->uk_domain[slab->us_domain];
4799 				/*
4800 				 * keg_alloc_slab() always returns a slab on the
4801 				 * partial list.
4802 				 */
4803 				LIST_REMOVE(slab, us_link);
4804 				LIST_INSERT_HEAD(&dom->ud_free_slab, slab,
4805 				    us_link);
4806 				dom->ud_free_slabs++;
4807 				KEG_UNLOCK(keg, slab->us_domain);
4808 				break;
4809 			}
4810 			if (vm_domainset_iter_policy(&di, &domain) != 0)
4811 				vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask, 0);
4812 		}
4813 	}
4814 }
4815 
4816 /*
4817  * Returns a snapshot of memory consumption in bytes.
4818  */
4819 size_t
4820 uma_zone_memory(uma_zone_t zone)
4821 {
4822 	size_t sz;
4823 	int i;
4824 
4825 	sz = 0;
4826 	if (zone->uz_flags & UMA_ZFLAG_CACHE) {
4827 		for (i = 0; i < vm_ndomains; i++)
4828 			sz += ZDOM_GET(zone, i)->uzd_nitems;
4829 		return (sz * zone->uz_size);
4830 	}
4831 	for (i = 0; i < vm_ndomains; i++)
4832 		sz += zone->uz_keg->uk_domain[i].ud_pages;
4833 
4834 	return (sz * PAGE_SIZE);
4835 }
4836 
4837 /* See uma.h */
4838 void
4839 uma_reclaim(int req)
4840 {
4841 
4842 	CTR0(KTR_UMA, "UMA: vm asked us to release pages!");
4843 	sx_xlock(&uma_reclaim_lock);
4844 	bucket_enable();
4845 
4846 	switch (req) {
4847 	case UMA_RECLAIM_TRIM:
4848 		zone_foreach(zone_trim, NULL);
4849 		break;
4850 	case UMA_RECLAIM_DRAIN:
4851 	case UMA_RECLAIM_DRAIN_CPU:
4852 		zone_foreach(zone_drain, NULL);
4853 		if (req == UMA_RECLAIM_DRAIN_CPU) {
4854 			pcpu_cache_drain_safe(NULL);
4855 			zone_foreach(zone_drain, NULL);
4856 		}
4857 		break;
4858 	default:
4859 		panic("unhandled reclamation request %d", req);
4860 	}
4861 
4862 	/*
4863 	 * Some slabs may have been freed but this zone will be visited early
4864 	 * we visit again so that we can free pages that are empty once other
4865 	 * zones are drained.  We have to do the same for buckets.
4866 	 */
4867 	zone_drain(slabzones[0], NULL);
4868 	zone_drain(slabzones[1], NULL);
4869 	bucket_zone_drain();
4870 	sx_xunlock(&uma_reclaim_lock);
4871 }
4872 
4873 static volatile int uma_reclaim_needed;
4874 
4875 void
4876 uma_reclaim_wakeup(void)
4877 {
4878 
4879 	if (atomic_fetchadd_int(&uma_reclaim_needed, 1) == 0)
4880 		wakeup(uma_reclaim);
4881 }
4882 
4883 void
4884 uma_reclaim_worker(void *arg __unused)
4885 {
4886 
4887 	for (;;) {
4888 		sx_xlock(&uma_reclaim_lock);
4889 		while (atomic_load_int(&uma_reclaim_needed) == 0)
4890 			sx_sleep(uma_reclaim, &uma_reclaim_lock, PVM, "umarcl",
4891 			    hz);
4892 		sx_xunlock(&uma_reclaim_lock);
4893 		EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM);
4894 		uma_reclaim(UMA_RECLAIM_DRAIN_CPU);
4895 		atomic_store_int(&uma_reclaim_needed, 0);
4896 		/* Don't fire more than once per-second. */
4897 		pause("umarclslp", hz);
4898 	}
4899 }
4900 
4901 /* See uma.h */
4902 void
4903 uma_zone_reclaim(uma_zone_t zone, int req)
4904 {
4905 
4906 	switch (req) {
4907 	case UMA_RECLAIM_TRIM:
4908 		zone_trim(zone, NULL);
4909 		break;
4910 	case UMA_RECLAIM_DRAIN:
4911 		zone_drain(zone, NULL);
4912 		break;
4913 	case UMA_RECLAIM_DRAIN_CPU:
4914 		pcpu_cache_drain_safe(zone);
4915 		zone_drain(zone, NULL);
4916 		break;
4917 	default:
4918 		panic("unhandled reclamation request %d", req);
4919 	}
4920 }
4921 
4922 /* See uma.h */
4923 int
4924 uma_zone_exhausted(uma_zone_t zone)
4925 {
4926 
4927 	return (atomic_load_32(&zone->uz_sleepers) > 0);
4928 }
4929 
4930 unsigned long
4931 uma_limit(void)
4932 {
4933 
4934 	return (uma_kmem_limit);
4935 }
4936 
4937 void
4938 uma_set_limit(unsigned long limit)
4939 {
4940 
4941 	uma_kmem_limit = limit;
4942 }
4943 
4944 unsigned long
4945 uma_size(void)
4946 {
4947 
4948 	return (atomic_load_long(&uma_kmem_total));
4949 }
4950 
4951 long
4952 uma_avail(void)
4953 {
4954 
4955 	return (uma_kmem_limit - uma_size());
4956 }
4957 
4958 #ifdef DDB
4959 /*
4960  * Generate statistics across both the zone and its per-cpu cache's.  Return
4961  * desired statistics if the pointer is non-NULL for that statistic.
4962  *
4963  * Note: does not update the zone statistics, as it can't safely clear the
4964  * per-CPU cache statistic.
4965  *
4966  */
4967 static void
4968 uma_zone_sumstat(uma_zone_t z, long *cachefreep, uint64_t *allocsp,
4969     uint64_t *freesp, uint64_t *sleepsp, uint64_t *xdomainp)
4970 {
4971 	uma_cache_t cache;
4972 	uint64_t allocs, frees, sleeps, xdomain;
4973 	int cachefree, cpu;
4974 
4975 	allocs = frees = sleeps = xdomain = 0;
4976 	cachefree = 0;
4977 	CPU_FOREACH(cpu) {
4978 		cache = &z->uz_cpu[cpu];
4979 		cachefree += cache->uc_allocbucket.ucb_cnt;
4980 		cachefree += cache->uc_freebucket.ucb_cnt;
4981 		xdomain += cache->uc_crossbucket.ucb_cnt;
4982 		cachefree += cache->uc_crossbucket.ucb_cnt;
4983 		allocs += cache->uc_allocs;
4984 		frees += cache->uc_frees;
4985 	}
4986 	allocs += counter_u64_fetch(z->uz_allocs);
4987 	frees += counter_u64_fetch(z->uz_frees);
4988 	xdomain += counter_u64_fetch(z->uz_xdomain);
4989 	sleeps += z->uz_sleeps;
4990 	if (cachefreep != NULL)
4991 		*cachefreep = cachefree;
4992 	if (allocsp != NULL)
4993 		*allocsp = allocs;
4994 	if (freesp != NULL)
4995 		*freesp = frees;
4996 	if (sleepsp != NULL)
4997 		*sleepsp = sleeps;
4998 	if (xdomainp != NULL)
4999 		*xdomainp = xdomain;
5000 }
5001 #endif /* DDB */
5002 
5003 static int
5004 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
5005 {
5006 	uma_keg_t kz;
5007 	uma_zone_t z;
5008 	int count;
5009 
5010 	count = 0;
5011 	rw_rlock(&uma_rwlock);
5012 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
5013 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
5014 			count++;
5015 	}
5016 	LIST_FOREACH(z, &uma_cachezones, uz_link)
5017 		count++;
5018 
5019 	rw_runlock(&uma_rwlock);
5020 	return (sysctl_handle_int(oidp, &count, 0, req));
5021 }
5022 
5023 static void
5024 uma_vm_zone_stats(struct uma_type_header *uth, uma_zone_t z, struct sbuf *sbuf,
5025     struct uma_percpu_stat *ups, bool internal)
5026 {
5027 	uma_zone_domain_t zdom;
5028 	uma_cache_t cache;
5029 	int i;
5030 
5031 	for (i = 0; i < vm_ndomains; i++) {
5032 		zdom = ZDOM_GET(z, i);
5033 		uth->uth_zone_free += zdom->uzd_nitems;
5034 	}
5035 	uth->uth_allocs = counter_u64_fetch(z->uz_allocs);
5036 	uth->uth_frees = counter_u64_fetch(z->uz_frees);
5037 	uth->uth_fails = counter_u64_fetch(z->uz_fails);
5038 	uth->uth_xdomain = counter_u64_fetch(z->uz_xdomain);
5039 	uth->uth_sleeps = z->uz_sleeps;
5040 
5041 	for (i = 0; i < mp_maxid + 1; i++) {
5042 		bzero(&ups[i], sizeof(*ups));
5043 		if (internal || CPU_ABSENT(i))
5044 			continue;
5045 		cache = &z->uz_cpu[i];
5046 		ups[i].ups_cache_free += cache->uc_allocbucket.ucb_cnt;
5047 		ups[i].ups_cache_free += cache->uc_freebucket.ucb_cnt;
5048 		ups[i].ups_cache_free += cache->uc_crossbucket.ucb_cnt;
5049 		ups[i].ups_allocs = cache->uc_allocs;
5050 		ups[i].ups_frees = cache->uc_frees;
5051 	}
5052 }
5053 
5054 static int
5055 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
5056 {
5057 	struct uma_stream_header ush;
5058 	struct uma_type_header uth;
5059 	struct uma_percpu_stat *ups;
5060 	struct sbuf sbuf;
5061 	uma_keg_t kz;
5062 	uma_zone_t z;
5063 	uint64_t items;
5064 	uint32_t kfree, pages;
5065 	int count, error, i;
5066 
5067 	error = sysctl_wire_old_buffer(req, 0);
5068 	if (error != 0)
5069 		return (error);
5070 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
5071 	sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL);
5072 	ups = malloc((mp_maxid + 1) * sizeof(*ups), M_TEMP, M_WAITOK);
5073 
5074 	count = 0;
5075 	rw_rlock(&uma_rwlock);
5076 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
5077 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
5078 			count++;
5079 	}
5080 
5081 	LIST_FOREACH(z, &uma_cachezones, uz_link)
5082 		count++;
5083 
5084 	/*
5085 	 * Insert stream header.
5086 	 */
5087 	bzero(&ush, sizeof(ush));
5088 	ush.ush_version = UMA_STREAM_VERSION;
5089 	ush.ush_maxcpus = (mp_maxid + 1);
5090 	ush.ush_count = count;
5091 	(void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
5092 
5093 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
5094 		kfree = pages = 0;
5095 		for (i = 0; i < vm_ndomains; i++) {
5096 			kfree += kz->uk_domain[i].ud_free_items;
5097 			pages += kz->uk_domain[i].ud_pages;
5098 		}
5099 		LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5100 			bzero(&uth, sizeof(uth));
5101 			strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5102 			uth.uth_align = kz->uk_align;
5103 			uth.uth_size = kz->uk_size;
5104 			uth.uth_rsize = kz->uk_rsize;
5105 			if (z->uz_max_items > 0) {
5106 				items = UZ_ITEMS_COUNT(z->uz_items);
5107 				uth.uth_pages = (items / kz->uk_ipers) *
5108 					kz->uk_ppera;
5109 			} else
5110 				uth.uth_pages = pages;
5111 			uth.uth_maxpages = (z->uz_max_items / kz->uk_ipers) *
5112 			    kz->uk_ppera;
5113 			uth.uth_limit = z->uz_max_items;
5114 			uth.uth_keg_free = kfree;
5115 
5116 			/*
5117 			 * A zone is secondary is it is not the first entry
5118 			 * on the keg's zone list.
5119 			 */
5120 			if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
5121 			    (LIST_FIRST(&kz->uk_zones) != z))
5122 				uth.uth_zone_flags = UTH_ZONE_SECONDARY;
5123 			uma_vm_zone_stats(&uth, z, &sbuf, ups,
5124 			    kz->uk_flags & UMA_ZFLAG_INTERNAL);
5125 			(void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5126 			for (i = 0; i < mp_maxid + 1; i++)
5127 				(void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5128 		}
5129 	}
5130 	LIST_FOREACH(z, &uma_cachezones, uz_link) {
5131 		bzero(&uth, sizeof(uth));
5132 		strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5133 		uth.uth_size = z->uz_size;
5134 		uma_vm_zone_stats(&uth, z, &sbuf, ups, false);
5135 		(void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5136 		for (i = 0; i < mp_maxid + 1; i++)
5137 			(void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5138 	}
5139 
5140 	rw_runlock(&uma_rwlock);
5141 	error = sbuf_finish(&sbuf);
5142 	sbuf_delete(&sbuf);
5143 	free(ups, M_TEMP);
5144 	return (error);
5145 }
5146 
5147 int
5148 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS)
5149 {
5150 	uma_zone_t zone = *(uma_zone_t *)arg1;
5151 	int error, max;
5152 
5153 	max = uma_zone_get_max(zone);
5154 	error = sysctl_handle_int(oidp, &max, 0, req);
5155 	if (error || !req->newptr)
5156 		return (error);
5157 
5158 	uma_zone_set_max(zone, max);
5159 
5160 	return (0);
5161 }
5162 
5163 int
5164 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS)
5165 {
5166 	uma_zone_t zone;
5167 	int cur;
5168 
5169 	/*
5170 	 * Some callers want to add sysctls for global zones that
5171 	 * may not yet exist so they pass a pointer to a pointer.
5172 	 */
5173 	if (arg2 == 0)
5174 		zone = *(uma_zone_t *)arg1;
5175 	else
5176 		zone = arg1;
5177 	cur = uma_zone_get_cur(zone);
5178 	return (sysctl_handle_int(oidp, &cur, 0, req));
5179 }
5180 
5181 static int
5182 sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS)
5183 {
5184 	uma_zone_t zone = arg1;
5185 	uint64_t cur;
5186 
5187 	cur = uma_zone_get_allocs(zone);
5188 	return (sysctl_handle_64(oidp, &cur, 0, req));
5189 }
5190 
5191 static int
5192 sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS)
5193 {
5194 	uma_zone_t zone = arg1;
5195 	uint64_t cur;
5196 
5197 	cur = uma_zone_get_frees(zone);
5198 	return (sysctl_handle_64(oidp, &cur, 0, req));
5199 }
5200 
5201 static int
5202 sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS)
5203 {
5204 	struct sbuf sbuf;
5205 	uma_zone_t zone = arg1;
5206 	int error;
5207 
5208 	sbuf_new_for_sysctl(&sbuf, NULL, 0, req);
5209 	if (zone->uz_flags != 0)
5210 		sbuf_printf(&sbuf, "0x%b", zone->uz_flags, PRINT_UMA_ZFLAGS);
5211 	else
5212 		sbuf_printf(&sbuf, "0");
5213 	error = sbuf_finish(&sbuf);
5214 	sbuf_delete(&sbuf);
5215 
5216 	return (error);
5217 }
5218 
5219 static int
5220 sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS)
5221 {
5222 	uma_keg_t keg = arg1;
5223 	int avail, effpct, total;
5224 
5225 	total = keg->uk_ppera * PAGE_SIZE;
5226 	if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0)
5227 		total += slabzone(keg->uk_ipers)->uz_keg->uk_rsize;
5228 	/*
5229 	 * We consider the client's requested size and alignment here, not the
5230 	 * real size determination uk_rsize, because we also adjust the real
5231 	 * size for internal implementation reasons (max bitset size).
5232 	 */
5233 	avail = keg->uk_ipers * roundup2(keg->uk_size, keg->uk_align + 1);
5234 	if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
5235 		avail *= mp_maxid + 1;
5236 	effpct = 100 * avail / total;
5237 	return (sysctl_handle_int(oidp, &effpct, 0, req));
5238 }
5239 
5240 static int
5241 sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS)
5242 {
5243 	uma_zone_t zone = arg1;
5244 	uint64_t cur;
5245 
5246 	cur = UZ_ITEMS_COUNT(atomic_load_64(&zone->uz_items));
5247 	return (sysctl_handle_64(oidp, &cur, 0, req));
5248 }
5249 
5250 #ifdef INVARIANTS
5251 static uma_slab_t
5252 uma_dbg_getslab(uma_zone_t zone, void *item)
5253 {
5254 	uma_slab_t slab;
5255 	uma_keg_t keg;
5256 	uint8_t *mem;
5257 
5258 	/*
5259 	 * It is safe to return the slab here even though the
5260 	 * zone is unlocked because the item's allocation state
5261 	 * essentially holds a reference.
5262 	 */
5263 	mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
5264 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5265 		return (NULL);
5266 	if (zone->uz_flags & UMA_ZFLAG_VTOSLAB)
5267 		return (vtoslab((vm_offset_t)mem));
5268 	keg = zone->uz_keg;
5269 	if ((keg->uk_flags & UMA_ZFLAG_HASH) == 0)
5270 		return ((uma_slab_t)(mem + keg->uk_pgoff));
5271 	KEG_LOCK(keg, 0);
5272 	slab = hash_sfind(&keg->uk_hash, mem);
5273 	KEG_UNLOCK(keg, 0);
5274 
5275 	return (slab);
5276 }
5277 
5278 static bool
5279 uma_dbg_zskip(uma_zone_t zone, void *mem)
5280 {
5281 
5282 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5283 		return (true);
5284 
5285 	return (uma_dbg_kskip(zone->uz_keg, mem));
5286 }
5287 
5288 static bool
5289 uma_dbg_kskip(uma_keg_t keg, void *mem)
5290 {
5291 	uintptr_t idx;
5292 
5293 	if (dbg_divisor == 0)
5294 		return (true);
5295 
5296 	if (dbg_divisor == 1)
5297 		return (false);
5298 
5299 	idx = (uintptr_t)mem >> PAGE_SHIFT;
5300 	if (keg->uk_ipers > 1) {
5301 		idx *= keg->uk_ipers;
5302 		idx += ((uintptr_t)mem & PAGE_MASK) / keg->uk_rsize;
5303 	}
5304 
5305 	if ((idx / dbg_divisor) * dbg_divisor != idx) {
5306 		counter_u64_add(uma_skip_cnt, 1);
5307 		return (true);
5308 	}
5309 	counter_u64_add(uma_dbg_cnt, 1);
5310 
5311 	return (false);
5312 }
5313 
5314 /*
5315  * Set up the slab's freei data such that uma_dbg_free can function.
5316  *
5317  */
5318 static void
5319 uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item)
5320 {
5321 	uma_keg_t keg;
5322 	int freei;
5323 
5324 	if (slab == NULL) {
5325 		slab = uma_dbg_getslab(zone, item);
5326 		if (slab == NULL)
5327 			panic("uma: item %p did not belong to zone %s",
5328 			    item, zone->uz_name);
5329 	}
5330 	keg = zone->uz_keg;
5331 	freei = slab_item_index(slab, keg, item);
5332 
5333 	if (BIT_ISSET(keg->uk_ipers, freei, slab_dbg_bits(slab, keg)))
5334 		panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)",
5335 		    item, zone, zone->uz_name, slab, freei);
5336 	BIT_SET_ATOMIC(keg->uk_ipers, freei, slab_dbg_bits(slab, keg));
5337 }
5338 
5339 /*
5340  * Verifies freed addresses.  Checks for alignment, valid slab membership
5341  * and duplicate frees.
5342  *
5343  */
5344 static void
5345 uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item)
5346 {
5347 	uma_keg_t keg;
5348 	int freei;
5349 
5350 	if (slab == NULL) {
5351 		slab = uma_dbg_getslab(zone, item);
5352 		if (slab == NULL)
5353 			panic("uma: Freed item %p did not belong to zone %s",
5354 			    item, zone->uz_name);
5355 	}
5356 	keg = zone->uz_keg;
5357 	freei = slab_item_index(slab, keg, item);
5358 
5359 	if (freei >= keg->uk_ipers)
5360 		panic("Invalid free of %p from zone %p(%s) slab %p(%d)",
5361 		    item, zone, zone->uz_name, slab, freei);
5362 
5363 	if (slab_item(slab, keg, freei) != item)
5364 		panic("Unaligned free of %p from zone %p(%s) slab %p(%d)",
5365 		    item, zone, zone->uz_name, slab, freei);
5366 
5367 	if (!BIT_ISSET(keg->uk_ipers, freei, slab_dbg_bits(slab, keg)))
5368 		panic("Duplicate free of %p from zone %p(%s) slab %p(%d)",
5369 		    item, zone, zone->uz_name, slab, freei);
5370 
5371 	BIT_CLR_ATOMIC(keg->uk_ipers, freei, slab_dbg_bits(slab, keg));
5372 }
5373 #endif /* INVARIANTS */
5374 
5375 #ifdef DDB
5376 static int64_t
5377 get_uma_stats(uma_keg_t kz, uma_zone_t z, uint64_t *allocs, uint64_t *used,
5378     uint64_t *sleeps, long *cachefree, uint64_t *xdomain)
5379 {
5380 	uint64_t frees;
5381 	int i;
5382 
5383 	if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
5384 		*allocs = counter_u64_fetch(z->uz_allocs);
5385 		frees = counter_u64_fetch(z->uz_frees);
5386 		*sleeps = z->uz_sleeps;
5387 		*cachefree = 0;
5388 		*xdomain = 0;
5389 	} else
5390 		uma_zone_sumstat(z, cachefree, allocs, &frees, sleeps,
5391 		    xdomain);
5392 	for (i = 0; i < vm_ndomains; i++) {
5393 		*cachefree += ZDOM_GET(z, i)->uzd_nitems;
5394 		if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
5395 		    (LIST_FIRST(&kz->uk_zones) != z)))
5396 			*cachefree += kz->uk_domain[i].ud_free_items;
5397 	}
5398 	*used = *allocs - frees;
5399 	return (((int64_t)*used + *cachefree) * kz->uk_size);
5400 }
5401 
5402 DB_SHOW_COMMAND(uma, db_show_uma)
5403 {
5404 	const char *fmt_hdr, *fmt_entry;
5405 	uma_keg_t kz;
5406 	uma_zone_t z;
5407 	uint64_t allocs, used, sleeps, xdomain;
5408 	long cachefree;
5409 	/* variables for sorting */
5410 	uma_keg_t cur_keg;
5411 	uma_zone_t cur_zone, last_zone;
5412 	int64_t cur_size, last_size, size;
5413 	int ties;
5414 
5415 	/* /i option produces machine-parseable CSV output */
5416 	if (modif[0] == 'i') {
5417 		fmt_hdr = "%s,%s,%s,%s,%s,%s,%s,%s,%s\n";
5418 		fmt_entry = "\"%s\",%ju,%jd,%ld,%ju,%ju,%u,%jd,%ju\n";
5419 	} else {
5420 		fmt_hdr = "%18s %6s %7s %7s %11s %7s %7s %10s %8s\n";
5421 		fmt_entry = "%18s %6ju %7jd %7ld %11ju %7ju %7u %10jd %8ju\n";
5422 	}
5423 
5424 	db_printf(fmt_hdr, "Zone", "Size", "Used", "Free", "Requests",
5425 	    "Sleeps", "Bucket", "Total Mem", "XFree");
5426 
5427 	/* Sort the zones with largest size first. */
5428 	last_zone = NULL;
5429 	last_size = INT64_MAX;
5430 	for (;;) {
5431 		cur_zone = NULL;
5432 		cur_size = -1;
5433 		ties = 0;
5434 		LIST_FOREACH(kz, &uma_kegs, uk_link) {
5435 			LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5436 				/*
5437 				 * In the case of size ties, print out zones
5438 				 * in the order they are encountered.  That is,
5439 				 * when we encounter the most recently output
5440 				 * zone, we have already printed all preceding
5441 				 * ties, and we must print all following ties.
5442 				 */
5443 				if (z == last_zone) {
5444 					ties = 1;
5445 					continue;
5446 				}
5447 				size = get_uma_stats(kz, z, &allocs, &used,
5448 				    &sleeps, &cachefree, &xdomain);
5449 				if (size > cur_size && size < last_size + ties)
5450 				{
5451 					cur_size = size;
5452 					cur_zone = z;
5453 					cur_keg = kz;
5454 				}
5455 			}
5456 		}
5457 		if (cur_zone == NULL)
5458 			break;
5459 
5460 		size = get_uma_stats(cur_keg, cur_zone, &allocs, &used,
5461 		    &sleeps, &cachefree, &xdomain);
5462 		db_printf(fmt_entry, cur_zone->uz_name,
5463 		    (uintmax_t)cur_keg->uk_size, (intmax_t)used, cachefree,
5464 		    (uintmax_t)allocs, (uintmax_t)sleeps,
5465 		    (unsigned)cur_zone->uz_bucket_size, (intmax_t)size,
5466 		    xdomain);
5467 
5468 		if (db_pager_quit)
5469 			return;
5470 		last_zone = cur_zone;
5471 		last_size = cur_size;
5472 	}
5473 }
5474 
5475 DB_SHOW_COMMAND(umacache, db_show_umacache)
5476 {
5477 	uma_zone_t z;
5478 	uint64_t allocs, frees;
5479 	long cachefree;
5480 	int i;
5481 
5482 	db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
5483 	    "Requests", "Bucket");
5484 	LIST_FOREACH(z, &uma_cachezones, uz_link) {
5485 		uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL, NULL);
5486 		for (i = 0; i < vm_ndomains; i++)
5487 			cachefree += ZDOM_GET(z, i)->uzd_nitems;
5488 		db_printf("%18s %8ju %8jd %8ld %12ju %8u\n",
5489 		    z->uz_name, (uintmax_t)z->uz_size,
5490 		    (intmax_t)(allocs - frees), cachefree,
5491 		    (uintmax_t)allocs, z->uz_bucket_size);
5492 		if (db_pager_quit)
5493 			return;
5494 	}
5495 }
5496 #endif	/* DDB */
5497