xref: /freebsd/sys/vm/uma_core.c (revision 2d4e511ca269f1908d27f4e5779c53475527391d)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2002-2019 Jeffrey Roberson <jeff@FreeBSD.org>
5  * Copyright (c) 2004, 2005 Bosko Milekic <bmilekic@FreeBSD.org>
6  * Copyright (c) 2004-2006 Robert N. M. Watson
7  * All rights reserved.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice unmodified, this list of conditions, and the following
14  *    disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
20  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
21  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
22  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
23  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
24  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
25  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
26  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
27  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
28  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
29  */
30 
31 /*
32  * uma_core.c  Implementation of the Universal Memory allocator
33  *
34  * This allocator is intended to replace the multitude of similar object caches
35  * in the standard FreeBSD kernel.  The intent is to be flexible as well as
36  * efficient.  A primary design goal is to return unused memory to the rest of
37  * the system.  This will make the system as a whole more flexible due to the
38  * ability to move memory to subsystems which most need it instead of leaving
39  * pools of reserved memory unused.
40  *
41  * The basic ideas stem from similar slab/zone based allocators whose algorithms
42  * are well known.
43  *
44  */
45 
46 /*
47  * TODO:
48  *	- Improve memory usage for large allocations
49  *	- Investigate cache size adjustments
50  */
51 
52 #include <sys/cdefs.h>
53 __FBSDID("$FreeBSD$");
54 
55 #include "opt_ddb.h"
56 #include "opt_param.h"
57 #include "opt_vm.h"
58 
59 #include <sys/param.h>
60 #include <sys/systm.h>
61 #include <sys/bitset.h>
62 #include <sys/domainset.h>
63 #include <sys/eventhandler.h>
64 #include <sys/kernel.h>
65 #include <sys/types.h>
66 #include <sys/limits.h>
67 #include <sys/queue.h>
68 #include <sys/malloc.h>
69 #include <sys/ktr.h>
70 #include <sys/lock.h>
71 #include <sys/sysctl.h>
72 #include <sys/mutex.h>
73 #include <sys/proc.h>
74 #include <sys/random.h>
75 #include <sys/rwlock.h>
76 #include <sys/sbuf.h>
77 #include <sys/sched.h>
78 #include <sys/sleepqueue.h>
79 #include <sys/smp.h>
80 #include <sys/smr.h>
81 #include <sys/taskqueue.h>
82 #include <sys/vmmeter.h>
83 
84 #include <vm/vm.h>
85 #include <vm/vm_domainset.h>
86 #include <vm/vm_object.h>
87 #include <vm/vm_page.h>
88 #include <vm/vm_pageout.h>
89 #include <vm/vm_param.h>
90 #include <vm/vm_phys.h>
91 #include <vm/vm_pagequeue.h>
92 #include <vm/vm_map.h>
93 #include <vm/vm_kern.h>
94 #include <vm/vm_extern.h>
95 #include <vm/uma.h>
96 #include <vm/uma_int.h>
97 #include <vm/uma_dbg.h>
98 
99 #include <ddb/ddb.h>
100 
101 #ifdef DEBUG_MEMGUARD
102 #include <vm/memguard.h>
103 #endif
104 
105 #include <machine/md_var.h>
106 
107 #ifdef INVARIANTS
108 #define	UMA_ALWAYS_CTORDTOR	1
109 #else
110 #define	UMA_ALWAYS_CTORDTOR	0
111 #endif
112 
113 /*
114  * This is the zone and keg from which all zones are spawned.
115  */
116 static uma_zone_t kegs;
117 static uma_zone_t zones;
118 
119 /*
120  * These are the two zones from which all offpage uma_slab_ts are allocated.
121  *
122  * One zone is for slab headers that can represent a larger number of items,
123  * making the slabs themselves more efficient, and the other zone is for
124  * headers that are smaller and represent fewer items, making the headers more
125  * efficient.
126  */
127 #define	SLABZONE_SIZE(setsize)					\
128     (sizeof(struct uma_hash_slab) + BITSET_SIZE(setsize) * SLAB_BITSETS)
129 #define	SLABZONE0_SETSIZE	(PAGE_SIZE / 16)
130 #define	SLABZONE1_SETSIZE	SLAB_MAX_SETSIZE
131 #define	SLABZONE0_SIZE	SLABZONE_SIZE(SLABZONE0_SETSIZE)
132 #define	SLABZONE1_SIZE	SLABZONE_SIZE(SLABZONE1_SETSIZE)
133 static uma_zone_t slabzones[2];
134 
135 /*
136  * The initial hash tables come out of this zone so they can be allocated
137  * prior to malloc coming up.
138  */
139 static uma_zone_t hashzone;
140 
141 /* The boot-time adjusted value for cache line alignment. */
142 int uma_align_cache = 64 - 1;
143 
144 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets");
145 static MALLOC_DEFINE(M_UMA, "UMA", "UMA Misc");
146 
147 /*
148  * Are we allowed to allocate buckets?
149  */
150 static int bucketdisable = 1;
151 
152 /* Linked list of all kegs in the system */
153 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs);
154 
155 /* Linked list of all cache-only zones in the system */
156 static LIST_HEAD(,uma_zone) uma_cachezones =
157     LIST_HEAD_INITIALIZER(uma_cachezones);
158 
159 /* This RW lock protects the keg list */
160 static struct rwlock_padalign __exclusive_cache_line uma_rwlock;
161 
162 /*
163  * First available virual address for boot time allocations.
164  */
165 static vm_offset_t bootstart;
166 static vm_offset_t bootmem;
167 
168 static struct sx uma_reclaim_lock;
169 
170 /*
171  * kmem soft limit, initialized by uma_set_limit().  Ensure that early
172  * allocations don't trigger a wakeup of the reclaim thread.
173  */
174 unsigned long uma_kmem_limit = LONG_MAX;
175 SYSCTL_ULONG(_vm, OID_AUTO, uma_kmem_limit, CTLFLAG_RD, &uma_kmem_limit, 0,
176     "UMA kernel memory soft limit");
177 unsigned long uma_kmem_total;
178 SYSCTL_ULONG(_vm, OID_AUTO, uma_kmem_total, CTLFLAG_RD, &uma_kmem_total, 0,
179     "UMA kernel memory usage");
180 
181 /* Is the VM done starting up? */
182 static enum {
183 	BOOT_COLD,
184 	BOOT_KVA,
185 	BOOT_RUNNING,
186 	BOOT_SHUTDOWN,
187 } booted = BOOT_COLD;
188 
189 /*
190  * This is the handle used to schedule events that need to happen
191  * outside of the allocation fast path.
192  */
193 static struct callout uma_callout;
194 #define	UMA_TIMEOUT	20		/* Seconds for callout interval. */
195 
196 /*
197  * This structure is passed as the zone ctor arg so that I don't have to create
198  * a special allocation function just for zones.
199  */
200 struct uma_zctor_args {
201 	const char *name;
202 	size_t size;
203 	uma_ctor ctor;
204 	uma_dtor dtor;
205 	uma_init uminit;
206 	uma_fini fini;
207 	uma_import import;
208 	uma_release release;
209 	void *arg;
210 	uma_keg_t keg;
211 	int align;
212 	uint32_t flags;
213 };
214 
215 struct uma_kctor_args {
216 	uma_zone_t zone;
217 	size_t size;
218 	uma_init uminit;
219 	uma_fini fini;
220 	int align;
221 	uint32_t flags;
222 };
223 
224 struct uma_bucket_zone {
225 	uma_zone_t	ubz_zone;
226 	const char	*ubz_name;
227 	int		ubz_entries;	/* Number of items it can hold. */
228 	int		ubz_maxsize;	/* Maximum allocation size per-item. */
229 };
230 
231 /*
232  * Compute the actual number of bucket entries to pack them in power
233  * of two sizes for more efficient space utilization.
234  */
235 #define	BUCKET_SIZE(n)						\
236     (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *))
237 
238 #define	BUCKET_MAX	BUCKET_SIZE(256)
239 #define	BUCKET_MIN	2
240 
241 struct uma_bucket_zone bucket_zones[] = {
242 	/* Literal bucket sizes. */
243 	{ NULL, "2 Bucket", 2, 4096 },
244 	{ NULL, "4 Bucket", 4, 3072 },
245 	{ NULL, "8 Bucket", 8, 2048 },
246 	{ NULL, "16 Bucket", 16, 1024 },
247 	/* Rounded down power of 2 sizes for efficiency. */
248 	{ NULL, "32 Bucket", BUCKET_SIZE(32), 512 },
249 	{ NULL, "64 Bucket", BUCKET_SIZE(64), 256 },
250 	{ NULL, "128 Bucket", BUCKET_SIZE(128), 128 },
251 	{ NULL, "256 Bucket", BUCKET_SIZE(256), 64 },
252 	{ NULL, NULL, 0}
253 };
254 
255 /*
256  * Flags and enumerations to be passed to internal functions.
257  */
258 enum zfreeskip {
259 	SKIP_NONE =	0,
260 	SKIP_CNT =	0x00000001,
261 	SKIP_DTOR =	0x00010000,
262 	SKIP_FINI =	0x00020000,
263 };
264 
265 /* Prototypes.. */
266 
267 void	uma_startup1(vm_offset_t);
268 void	uma_startup2(void);
269 
270 static void *noobj_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
271 static void *page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
272 static void *pcpu_page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
273 static void *startup_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
274 static void *contig_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
275 static void page_free(void *, vm_size_t, uint8_t);
276 static void pcpu_page_free(void *, vm_size_t, uint8_t);
277 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int, int, int);
278 static void cache_drain(uma_zone_t);
279 static void bucket_drain(uma_zone_t, uma_bucket_t);
280 static void bucket_cache_reclaim(uma_zone_t zone, bool);
281 static int keg_ctor(void *, int, void *, int);
282 static void keg_dtor(void *, int, void *);
283 static int zone_ctor(void *, int, void *, int);
284 static void zone_dtor(void *, int, void *);
285 static inline void item_dtor(uma_zone_t zone, void *item, int size,
286     void *udata, enum zfreeskip skip);
287 static int zero_init(void *, int, int);
288 static void zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata,
289     int itemdomain, bool ws);
290 static void zone_foreach(void (*zfunc)(uma_zone_t, void *), void *);
291 static void zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *), void *);
292 static void zone_timeout(uma_zone_t zone, void *);
293 static int hash_alloc(struct uma_hash *, u_int);
294 static int hash_expand(struct uma_hash *, struct uma_hash *);
295 static void hash_free(struct uma_hash *hash);
296 static void uma_timeout(void *);
297 static void uma_startup3(void);
298 static void uma_shutdown(void);
299 static void *zone_alloc_item(uma_zone_t, void *, int, int);
300 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip);
301 static int zone_alloc_limit(uma_zone_t zone, int count, int flags);
302 static void zone_free_limit(uma_zone_t zone, int count);
303 static void bucket_enable(void);
304 static void bucket_init(void);
305 static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int);
306 static void bucket_free(uma_zone_t zone, uma_bucket_t, void *);
307 static void bucket_zone_drain(void);
308 static uma_bucket_t zone_alloc_bucket(uma_zone_t, void *, int, int);
309 static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab);
310 static void slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item);
311 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit,
312     uma_fini fini, int align, uint32_t flags);
313 static int zone_import(void *, void **, int, int, int);
314 static void zone_release(void *, void **, int);
315 static bool cache_alloc(uma_zone_t, uma_cache_t, void *, int);
316 static bool cache_free(uma_zone_t, uma_cache_t, void *, void *, int);
317 
318 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS);
319 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS);
320 static int sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS);
321 static int sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS);
322 static int sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS);
323 static int sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS);
324 static int sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS);
325 
326 static uint64_t uma_zone_get_allocs(uma_zone_t zone);
327 
328 static SYSCTL_NODE(_vm, OID_AUTO, debug, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
329     "Memory allocation debugging");
330 
331 #ifdef INVARIANTS
332 static uint64_t uma_keg_get_allocs(uma_keg_t zone);
333 static inline struct noslabbits *slab_dbg_bits(uma_slab_t slab, uma_keg_t keg);
334 
335 static bool uma_dbg_kskip(uma_keg_t keg, void *mem);
336 static bool uma_dbg_zskip(uma_zone_t zone, void *mem);
337 static void uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item);
338 static void uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item);
339 
340 static u_int dbg_divisor = 1;
341 SYSCTL_UINT(_vm_debug, OID_AUTO, divisor,
342     CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &dbg_divisor, 0,
343     "Debug & thrash every this item in memory allocator");
344 
345 static counter_u64_t uma_dbg_cnt = EARLY_COUNTER;
346 static counter_u64_t uma_skip_cnt = EARLY_COUNTER;
347 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, trashed, CTLFLAG_RD,
348     &uma_dbg_cnt, "memory items debugged");
349 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, skipped, CTLFLAG_RD,
350     &uma_skip_cnt, "memory items skipped, not debugged");
351 #endif
352 
353 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
354 
355 SYSCTL_NODE(_vm, OID_AUTO, uma, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
356     "Universal Memory Allocator");
357 
358 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLFLAG_MPSAFE|CTLTYPE_INT,
359     0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones");
360 
361 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLFLAG_MPSAFE|CTLTYPE_STRUCT,
362     0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats");
363 
364 static int zone_warnings = 1;
365 SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RWTUN, &zone_warnings, 0,
366     "Warn when UMA zones becomes full");
367 
368 static int multipage_slabs = 1;
369 TUNABLE_INT("vm.debug.uma_multipage_slabs", &multipage_slabs);
370 SYSCTL_INT(_vm_debug, OID_AUTO, uma_multipage_slabs,
371     CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &multipage_slabs, 0,
372     "UMA may choose larger slab sizes for better efficiency");
373 
374 /*
375  * Select the slab zone for an offpage slab with the given maximum item count.
376  */
377 static inline uma_zone_t
378 slabzone(int ipers)
379 {
380 
381 	return (slabzones[ipers > SLABZONE0_SETSIZE]);
382 }
383 
384 /*
385  * This routine checks to see whether or not it's safe to enable buckets.
386  */
387 static void
388 bucket_enable(void)
389 {
390 
391 	KASSERT(booted >= BOOT_KVA, ("Bucket enable before init"));
392 	bucketdisable = vm_page_count_min();
393 }
394 
395 /*
396  * Initialize bucket_zones, the array of zones of buckets of various sizes.
397  *
398  * For each zone, calculate the memory required for each bucket, consisting
399  * of the header and an array of pointers.
400  */
401 static void
402 bucket_init(void)
403 {
404 	struct uma_bucket_zone *ubz;
405 	int size;
406 
407 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) {
408 		size = roundup(sizeof(struct uma_bucket), sizeof(void *));
409 		size += sizeof(void *) * ubz->ubz_entries;
410 		ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
411 		    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
412 		    UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET |
413 		    UMA_ZONE_FIRSTTOUCH);
414 	}
415 }
416 
417 /*
418  * Given a desired number of entries for a bucket, return the zone from which
419  * to allocate the bucket.
420  */
421 static struct uma_bucket_zone *
422 bucket_zone_lookup(int entries)
423 {
424 	struct uma_bucket_zone *ubz;
425 
426 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
427 		if (ubz->ubz_entries >= entries)
428 			return (ubz);
429 	ubz--;
430 	return (ubz);
431 }
432 
433 static struct uma_bucket_zone *
434 bucket_zone_max(uma_zone_t zone, int nitems)
435 {
436 	struct uma_bucket_zone *ubz;
437 	int bpcpu;
438 
439 	bpcpu = 2;
440 	if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
441 		/* Count the cross-domain bucket. */
442 		bpcpu++;
443 
444 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
445 		if (ubz->ubz_entries * bpcpu * mp_ncpus > nitems)
446 			break;
447 	if (ubz == &bucket_zones[0])
448 		ubz = NULL;
449 	else
450 		ubz--;
451 	return (ubz);
452 }
453 
454 static int
455 bucket_select(int size)
456 {
457 	struct uma_bucket_zone *ubz;
458 
459 	ubz = &bucket_zones[0];
460 	if (size > ubz->ubz_maxsize)
461 		return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1);
462 
463 	for (; ubz->ubz_entries != 0; ubz++)
464 		if (ubz->ubz_maxsize < size)
465 			break;
466 	ubz--;
467 	return (ubz->ubz_entries);
468 }
469 
470 static uma_bucket_t
471 bucket_alloc(uma_zone_t zone, void *udata, int flags)
472 {
473 	struct uma_bucket_zone *ubz;
474 	uma_bucket_t bucket;
475 
476 	/*
477 	 * Don't allocate buckets early in boot.
478 	 */
479 	if (__predict_false(booted < BOOT_KVA))
480 		return (NULL);
481 
482 	/*
483 	 * To limit bucket recursion we store the original zone flags
484 	 * in a cookie passed via zalloc_arg/zfree_arg.  This allows the
485 	 * NOVM flag to persist even through deep recursions.  We also
486 	 * store ZFLAG_BUCKET once we have recursed attempting to allocate
487 	 * a bucket for a bucket zone so we do not allow infinite bucket
488 	 * recursion.  This cookie will even persist to frees of unused
489 	 * buckets via the allocation path or bucket allocations in the
490 	 * free path.
491 	 */
492 	if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
493 		udata = (void *)(uintptr_t)zone->uz_flags;
494 	else {
495 		if ((uintptr_t)udata & UMA_ZFLAG_BUCKET)
496 			return (NULL);
497 		udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET);
498 	}
499 	if (((uintptr_t)udata & UMA_ZONE_VM) != 0)
500 		flags |= M_NOVM;
501 	ubz = bucket_zone_lookup(zone->uz_bucket_size);
502 	if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0)
503 		ubz++;
504 	bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags);
505 	if (bucket) {
506 #ifdef INVARIANTS
507 		bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
508 #endif
509 		bucket->ub_cnt = 0;
510 		bucket->ub_entries = ubz->ubz_entries;
511 		bucket->ub_seq = SMR_SEQ_INVALID;
512 		CTR3(KTR_UMA, "bucket_alloc: zone %s(%p) allocated bucket %p",
513 		    zone->uz_name, zone, bucket);
514 	}
515 
516 	return (bucket);
517 }
518 
519 static void
520 bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata)
521 {
522 	struct uma_bucket_zone *ubz;
523 
524 	if (bucket->ub_cnt != 0)
525 		bucket_drain(zone, bucket);
526 
527 	KASSERT(bucket->ub_cnt == 0,
528 	    ("bucket_free: Freeing a non free bucket."));
529 	KASSERT(bucket->ub_seq == SMR_SEQ_INVALID,
530 	    ("bucket_free: Freeing an SMR bucket."));
531 	if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
532 		udata = (void *)(uintptr_t)zone->uz_flags;
533 	ubz = bucket_zone_lookup(bucket->ub_entries);
534 	uma_zfree_arg(ubz->ubz_zone, bucket, udata);
535 }
536 
537 static void
538 bucket_zone_drain(void)
539 {
540 	struct uma_bucket_zone *ubz;
541 
542 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
543 		uma_zone_reclaim(ubz->ubz_zone, UMA_RECLAIM_DRAIN);
544 }
545 
546 /*
547  * Acquire the domain lock and record contention.
548  */
549 static uma_zone_domain_t
550 zone_domain_lock(uma_zone_t zone, int domain)
551 {
552 	uma_zone_domain_t zdom;
553 	bool lockfail;
554 
555 	zdom = ZDOM_GET(zone, domain);
556 	lockfail = false;
557 	if (ZDOM_OWNED(zdom))
558 		lockfail = true;
559 	ZDOM_LOCK(zdom);
560 	/* This is unsynchronized.  The counter does not need to be precise. */
561 	if (lockfail && zone->uz_bucket_size < zone->uz_bucket_size_max)
562 		zone->uz_bucket_size++;
563 	return (zdom);
564 }
565 
566 /*
567  * Search for the domain with the least cached items and return it if it
568  * is out of balance with the preferred domain.
569  */
570 static __noinline int
571 zone_domain_lowest(uma_zone_t zone, int pref)
572 {
573 	long least, nitems, prefitems;
574 	int domain;
575 	int i;
576 
577 	prefitems = least = LONG_MAX;
578 	domain = 0;
579 	for (i = 0; i < vm_ndomains; i++) {
580 		nitems = ZDOM_GET(zone, i)->uzd_nitems;
581 		if (nitems < least) {
582 			domain = i;
583 			least = nitems;
584 		}
585 		if (domain == pref)
586 			prefitems = nitems;
587 	}
588 	if (prefitems < least * 2)
589 		return (pref);
590 
591 	return (domain);
592 }
593 
594 /*
595  * Search for the domain with the most cached items and return it or the
596  * preferred domain if it has enough to proceed.
597  */
598 static __noinline int
599 zone_domain_highest(uma_zone_t zone, int pref)
600 {
601 	long most, nitems;
602 	int domain;
603 	int i;
604 
605 	if (ZDOM_GET(zone, pref)->uzd_nitems > BUCKET_MAX)
606 		return (pref);
607 
608 	most = 0;
609 	domain = 0;
610 	for (i = 0; i < vm_ndomains; i++) {
611 		nitems = ZDOM_GET(zone, i)->uzd_nitems;
612 		if (nitems > most) {
613 			domain = i;
614 			most = nitems;
615 		}
616 	}
617 
618 	return (domain);
619 }
620 
621 /*
622  * Safely subtract cnt from imax.
623  */
624 static void
625 zone_domain_imax_sub(uma_zone_domain_t zdom, int cnt)
626 {
627 	long new;
628 	long old;
629 
630 	old = zdom->uzd_imax;
631 	do {
632 		if (old <= cnt)
633 			new = 0;
634 		else
635 			new = old - cnt;
636 	} while (atomic_fcmpset_long(&zdom->uzd_imax, &old, new) == 0);
637 }
638 
639 /*
640  * Set the maximum imax value.
641  */
642 static void
643 zone_domain_imax_set(uma_zone_domain_t zdom, int nitems)
644 {
645 	long old;
646 
647 	old = zdom->uzd_imax;
648 	do {
649 		if (old >= nitems)
650 			break;
651 	} while (atomic_fcmpset_long(&zdom->uzd_imax, &old, nitems) == 0);
652 }
653 
654 /*
655  * Attempt to satisfy an allocation by retrieving a full bucket from one of the
656  * zone's caches.  If a bucket is found the zone is not locked on return.
657  */
658 static uma_bucket_t
659 zone_fetch_bucket(uma_zone_t zone, uma_zone_domain_t zdom, bool reclaim)
660 {
661 	uma_bucket_t bucket;
662 	int i;
663 	bool dtor = false;
664 
665 	ZDOM_LOCK_ASSERT(zdom);
666 
667 	if ((bucket = STAILQ_FIRST(&zdom->uzd_buckets)) == NULL)
668 		return (NULL);
669 
670 	/* SMR Buckets can not be re-used until readers expire. */
671 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
672 	    bucket->ub_seq != SMR_SEQ_INVALID) {
673 		if (!smr_poll(zone->uz_smr, bucket->ub_seq, false))
674 			return (NULL);
675 		bucket->ub_seq = SMR_SEQ_INVALID;
676 		dtor = (zone->uz_dtor != NULL) || UMA_ALWAYS_CTORDTOR;
677 		if (STAILQ_NEXT(bucket, ub_link) != NULL)
678 			zdom->uzd_seq = STAILQ_NEXT(bucket, ub_link)->ub_seq;
679 	}
680 	MPASS(zdom->uzd_nitems >= bucket->ub_cnt);
681 	STAILQ_REMOVE_HEAD(&zdom->uzd_buckets, ub_link);
682 	zdom->uzd_nitems -= bucket->ub_cnt;
683 
684 	/*
685 	 * Shift the bounds of the current WSS interval to avoid
686 	 * perturbing the estimate.
687 	 */
688 	if (reclaim) {
689 		zdom->uzd_imin -= lmin(zdom->uzd_imin, bucket->ub_cnt);
690 		zone_domain_imax_sub(zdom, bucket->ub_cnt);
691 	} else if (zdom->uzd_imin > zdom->uzd_nitems)
692 		zdom->uzd_imin = zdom->uzd_nitems;
693 
694 	ZDOM_UNLOCK(zdom);
695 	if (dtor)
696 		for (i = 0; i < bucket->ub_cnt; i++)
697 			item_dtor(zone, bucket->ub_bucket[i], zone->uz_size,
698 			    NULL, SKIP_NONE);
699 
700 	return (bucket);
701 }
702 
703 /*
704  * Insert a full bucket into the specified cache.  The "ws" parameter indicates
705  * whether the bucket's contents should be counted as part of the zone's working
706  * set.  The bucket may be freed if it exceeds the bucket limit.
707  */
708 static void
709 zone_put_bucket(uma_zone_t zone, int domain, uma_bucket_t bucket, void *udata,
710     const bool ws)
711 {
712 	uma_zone_domain_t zdom;
713 
714 	/* We don't cache empty buckets.  This can happen after a reclaim. */
715 	if (bucket->ub_cnt == 0)
716 		goto out;
717 	zdom = zone_domain_lock(zone, domain);
718 
719 	KASSERT(!ws || zdom->uzd_nitems < zone->uz_bucket_max,
720 	    ("%s: zone %p overflow", __func__, zone));
721 
722 	/*
723 	 * Conditionally set the maximum number of items.
724 	 */
725 	zdom->uzd_nitems += bucket->ub_cnt;
726 	if (__predict_true(zdom->uzd_nitems < zone->uz_bucket_max)) {
727 		if (ws)
728 			zone_domain_imax_set(zdom, zdom->uzd_nitems);
729 		if (STAILQ_EMPTY(&zdom->uzd_buckets))
730 			zdom->uzd_seq = bucket->ub_seq;
731 		STAILQ_INSERT_TAIL(&zdom->uzd_buckets, bucket, ub_link);
732 		ZDOM_UNLOCK(zdom);
733 		return;
734 	}
735 	zdom->uzd_nitems -= bucket->ub_cnt;
736 	ZDOM_UNLOCK(zdom);
737 out:
738 	bucket_free(zone, bucket, udata);
739 }
740 
741 /* Pops an item out of a per-cpu cache bucket. */
742 static inline void *
743 cache_bucket_pop(uma_cache_t cache, uma_cache_bucket_t bucket)
744 {
745 	void *item;
746 
747 	CRITICAL_ASSERT(curthread);
748 
749 	bucket->ucb_cnt--;
750 	item = bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt];
751 #ifdef INVARIANTS
752 	bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = NULL;
753 	KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled."));
754 #endif
755 	cache->uc_allocs++;
756 
757 	return (item);
758 }
759 
760 /* Pushes an item into a per-cpu cache bucket. */
761 static inline void
762 cache_bucket_push(uma_cache_t cache, uma_cache_bucket_t bucket, void *item)
763 {
764 
765 	CRITICAL_ASSERT(curthread);
766 	KASSERT(bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] == NULL,
767 	    ("uma_zfree: Freeing to non free bucket index."));
768 
769 	bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = item;
770 	bucket->ucb_cnt++;
771 	cache->uc_frees++;
772 }
773 
774 /*
775  * Unload a UMA bucket from a per-cpu cache.
776  */
777 static inline uma_bucket_t
778 cache_bucket_unload(uma_cache_bucket_t bucket)
779 {
780 	uma_bucket_t b;
781 
782 	b = bucket->ucb_bucket;
783 	if (b != NULL) {
784 		MPASS(b->ub_entries == bucket->ucb_entries);
785 		b->ub_cnt = bucket->ucb_cnt;
786 		bucket->ucb_bucket = NULL;
787 		bucket->ucb_entries = bucket->ucb_cnt = 0;
788 	}
789 
790 	return (b);
791 }
792 
793 static inline uma_bucket_t
794 cache_bucket_unload_alloc(uma_cache_t cache)
795 {
796 
797 	return (cache_bucket_unload(&cache->uc_allocbucket));
798 }
799 
800 static inline uma_bucket_t
801 cache_bucket_unload_free(uma_cache_t cache)
802 {
803 
804 	return (cache_bucket_unload(&cache->uc_freebucket));
805 }
806 
807 static inline uma_bucket_t
808 cache_bucket_unload_cross(uma_cache_t cache)
809 {
810 
811 	return (cache_bucket_unload(&cache->uc_crossbucket));
812 }
813 
814 /*
815  * Load a bucket into a per-cpu cache bucket.
816  */
817 static inline void
818 cache_bucket_load(uma_cache_bucket_t bucket, uma_bucket_t b)
819 {
820 
821 	CRITICAL_ASSERT(curthread);
822 	MPASS(bucket->ucb_bucket == NULL);
823 	MPASS(b->ub_seq == SMR_SEQ_INVALID);
824 
825 	bucket->ucb_bucket = b;
826 	bucket->ucb_cnt = b->ub_cnt;
827 	bucket->ucb_entries = b->ub_entries;
828 }
829 
830 static inline void
831 cache_bucket_load_alloc(uma_cache_t cache, uma_bucket_t b)
832 {
833 
834 	cache_bucket_load(&cache->uc_allocbucket, b);
835 }
836 
837 static inline void
838 cache_bucket_load_free(uma_cache_t cache, uma_bucket_t b)
839 {
840 
841 	cache_bucket_load(&cache->uc_freebucket, b);
842 }
843 
844 #ifdef NUMA
845 static inline void
846 cache_bucket_load_cross(uma_cache_t cache, uma_bucket_t b)
847 {
848 
849 	cache_bucket_load(&cache->uc_crossbucket, b);
850 }
851 #endif
852 
853 /*
854  * Copy and preserve ucb_spare.
855  */
856 static inline void
857 cache_bucket_copy(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
858 {
859 
860 	b1->ucb_bucket = b2->ucb_bucket;
861 	b1->ucb_entries = b2->ucb_entries;
862 	b1->ucb_cnt = b2->ucb_cnt;
863 }
864 
865 /*
866  * Swap two cache buckets.
867  */
868 static inline void
869 cache_bucket_swap(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
870 {
871 	struct uma_cache_bucket b3;
872 
873 	CRITICAL_ASSERT(curthread);
874 
875 	cache_bucket_copy(&b3, b1);
876 	cache_bucket_copy(b1, b2);
877 	cache_bucket_copy(b2, &b3);
878 }
879 
880 /*
881  * Attempt to fetch a bucket from a zone on behalf of the current cpu cache.
882  */
883 static uma_bucket_t
884 cache_fetch_bucket(uma_zone_t zone, uma_cache_t cache, int domain)
885 {
886 	uma_zone_domain_t zdom;
887 	uma_bucket_t bucket;
888 
889 	/*
890 	 * Avoid the lock if possible.
891 	 */
892 	zdom = ZDOM_GET(zone, domain);
893 	if (zdom->uzd_nitems == 0)
894 		return (NULL);
895 
896 	if ((cache_uz_flags(cache) & UMA_ZONE_SMR) != 0 &&
897 	    !smr_poll(zone->uz_smr, zdom->uzd_seq, false))
898 		return (NULL);
899 
900 	/*
901 	 * Check the zone's cache of buckets.
902 	 */
903 	zdom = zone_domain_lock(zone, domain);
904 	if ((bucket = zone_fetch_bucket(zone, zdom, false)) != NULL) {
905 		KASSERT(bucket->ub_cnt != 0,
906 		    ("cache_fetch_bucket: Returning an empty bucket."));
907 		return (bucket);
908 	}
909 	ZDOM_UNLOCK(zdom);
910 
911 	return (NULL);
912 }
913 
914 static void
915 zone_log_warning(uma_zone_t zone)
916 {
917 	static const struct timeval warninterval = { 300, 0 };
918 
919 	if (!zone_warnings || zone->uz_warning == NULL)
920 		return;
921 
922 	if (ratecheck(&zone->uz_ratecheck, &warninterval))
923 		printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning);
924 }
925 
926 static inline void
927 zone_maxaction(uma_zone_t zone)
928 {
929 
930 	if (zone->uz_maxaction.ta_func != NULL)
931 		taskqueue_enqueue(taskqueue_thread, &zone->uz_maxaction);
932 }
933 
934 /*
935  * Routine called by timeout which is used to fire off some time interval
936  * based calculations.  (stats, hash size, etc.)
937  *
938  * Arguments:
939  *	arg   Unused
940  *
941  * Returns:
942  *	Nothing
943  */
944 static void
945 uma_timeout(void *unused)
946 {
947 	bucket_enable();
948 	zone_foreach(zone_timeout, NULL);
949 
950 	/* Reschedule this event */
951 	callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
952 }
953 
954 /*
955  * Update the working set size estimate for the zone's bucket cache.
956  * The constants chosen here are somewhat arbitrary.  With an update period of
957  * 20s (UMA_TIMEOUT), this estimate is dominated by zone activity over the
958  * last 100s.
959  */
960 static void
961 zone_domain_update_wss(uma_zone_domain_t zdom)
962 {
963 	long wss;
964 
965 	ZDOM_LOCK(zdom);
966 	MPASS(zdom->uzd_imax >= zdom->uzd_imin);
967 	wss = zdom->uzd_imax - zdom->uzd_imin;
968 	zdom->uzd_imax = zdom->uzd_imin = zdom->uzd_nitems;
969 	zdom->uzd_wss = (4 * wss + zdom->uzd_wss) / 5;
970 	ZDOM_UNLOCK(zdom);
971 }
972 
973 /*
974  * Routine to perform timeout driven calculations.  This expands the
975  * hashes and does per cpu statistics aggregation.
976  *
977  *  Returns nothing.
978  */
979 static void
980 zone_timeout(uma_zone_t zone, void *unused)
981 {
982 	uma_keg_t keg;
983 	u_int slabs, pages;
984 
985 	if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
986 		goto update_wss;
987 
988 	keg = zone->uz_keg;
989 
990 	/*
991 	 * Hash zones are non-numa by definition so the first domain
992 	 * is the only one present.
993 	 */
994 	KEG_LOCK(keg, 0);
995 	pages = keg->uk_domain[0].ud_pages;
996 
997 	/*
998 	 * Expand the keg hash table.
999 	 *
1000 	 * This is done if the number of slabs is larger than the hash size.
1001 	 * What I'm trying to do here is completely reduce collisions.  This
1002 	 * may be a little aggressive.  Should I allow for two collisions max?
1003 	 */
1004 	if ((slabs = pages / keg->uk_ppera) > keg->uk_hash.uh_hashsize) {
1005 		struct uma_hash newhash;
1006 		struct uma_hash oldhash;
1007 		int ret;
1008 
1009 		/*
1010 		 * This is so involved because allocating and freeing
1011 		 * while the keg lock is held will lead to deadlock.
1012 		 * I have to do everything in stages and check for
1013 		 * races.
1014 		 */
1015 		KEG_UNLOCK(keg, 0);
1016 		ret = hash_alloc(&newhash, 1 << fls(slabs));
1017 		KEG_LOCK(keg, 0);
1018 		if (ret) {
1019 			if (hash_expand(&keg->uk_hash, &newhash)) {
1020 				oldhash = keg->uk_hash;
1021 				keg->uk_hash = newhash;
1022 			} else
1023 				oldhash = newhash;
1024 
1025 			KEG_UNLOCK(keg, 0);
1026 			hash_free(&oldhash);
1027 			goto update_wss;
1028 		}
1029 	}
1030 	KEG_UNLOCK(keg, 0);
1031 
1032 update_wss:
1033 	for (int i = 0; i < vm_ndomains; i++)
1034 		zone_domain_update_wss(ZDOM_GET(zone, i));
1035 }
1036 
1037 /*
1038  * Allocate and zero fill the next sized hash table from the appropriate
1039  * backing store.
1040  *
1041  * Arguments:
1042  *	hash  A new hash structure with the old hash size in uh_hashsize
1043  *
1044  * Returns:
1045  *	1 on success and 0 on failure.
1046  */
1047 static int
1048 hash_alloc(struct uma_hash *hash, u_int size)
1049 {
1050 	size_t alloc;
1051 
1052 	KASSERT(powerof2(size), ("hash size must be power of 2"));
1053 	if (size > UMA_HASH_SIZE_INIT)  {
1054 		hash->uh_hashsize = size;
1055 		alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
1056 		hash->uh_slab_hash = malloc(alloc, M_UMAHASH, M_NOWAIT);
1057 	} else {
1058 		alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
1059 		hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
1060 		    UMA_ANYDOMAIN, M_WAITOK);
1061 		hash->uh_hashsize = UMA_HASH_SIZE_INIT;
1062 	}
1063 	if (hash->uh_slab_hash) {
1064 		bzero(hash->uh_slab_hash, alloc);
1065 		hash->uh_hashmask = hash->uh_hashsize - 1;
1066 		return (1);
1067 	}
1068 
1069 	return (0);
1070 }
1071 
1072 /*
1073  * Expands the hash table for HASH zones.  This is done from zone_timeout
1074  * to reduce collisions.  This must not be done in the regular allocation
1075  * path, otherwise, we can recurse on the vm while allocating pages.
1076  *
1077  * Arguments:
1078  *	oldhash  The hash you want to expand
1079  *	newhash  The hash structure for the new table
1080  *
1081  * Returns:
1082  *	Nothing
1083  *
1084  * Discussion:
1085  */
1086 static int
1087 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
1088 {
1089 	uma_hash_slab_t slab;
1090 	u_int hval;
1091 	u_int idx;
1092 
1093 	if (!newhash->uh_slab_hash)
1094 		return (0);
1095 
1096 	if (oldhash->uh_hashsize >= newhash->uh_hashsize)
1097 		return (0);
1098 
1099 	/*
1100 	 * I need to investigate hash algorithms for resizing without a
1101 	 * full rehash.
1102 	 */
1103 
1104 	for (idx = 0; idx < oldhash->uh_hashsize; idx++)
1105 		while (!LIST_EMPTY(&oldhash->uh_slab_hash[idx])) {
1106 			slab = LIST_FIRST(&oldhash->uh_slab_hash[idx]);
1107 			LIST_REMOVE(slab, uhs_hlink);
1108 			hval = UMA_HASH(newhash, slab->uhs_data);
1109 			LIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
1110 			    slab, uhs_hlink);
1111 		}
1112 
1113 	return (1);
1114 }
1115 
1116 /*
1117  * Free the hash bucket to the appropriate backing store.
1118  *
1119  * Arguments:
1120  *	slab_hash  The hash bucket we're freeing
1121  *	hashsize   The number of entries in that hash bucket
1122  *
1123  * Returns:
1124  *	Nothing
1125  */
1126 static void
1127 hash_free(struct uma_hash *hash)
1128 {
1129 	if (hash->uh_slab_hash == NULL)
1130 		return;
1131 	if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
1132 		zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE);
1133 	else
1134 		free(hash->uh_slab_hash, M_UMAHASH);
1135 }
1136 
1137 /*
1138  * Frees all outstanding items in a bucket
1139  *
1140  * Arguments:
1141  *	zone   The zone to free to, must be unlocked.
1142  *	bucket The free/alloc bucket with items.
1143  *
1144  * Returns:
1145  *	Nothing
1146  */
1147 static void
1148 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
1149 {
1150 	int i;
1151 
1152 	if (bucket->ub_cnt == 0)
1153 		return;
1154 
1155 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
1156 	    bucket->ub_seq != SMR_SEQ_INVALID) {
1157 		smr_wait(zone->uz_smr, bucket->ub_seq);
1158 		bucket->ub_seq = SMR_SEQ_INVALID;
1159 		for (i = 0; i < bucket->ub_cnt; i++)
1160 			item_dtor(zone, bucket->ub_bucket[i],
1161 			    zone->uz_size, NULL, SKIP_NONE);
1162 	}
1163 	if (zone->uz_fini)
1164 		for (i = 0; i < bucket->ub_cnt; i++)
1165 			zone->uz_fini(bucket->ub_bucket[i], zone->uz_size);
1166 	zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt);
1167 	if (zone->uz_max_items > 0)
1168 		zone_free_limit(zone, bucket->ub_cnt);
1169 #ifdef INVARIANTS
1170 	bzero(bucket->ub_bucket, sizeof(void *) * bucket->ub_cnt);
1171 #endif
1172 	bucket->ub_cnt = 0;
1173 }
1174 
1175 /*
1176  * Drains the per cpu caches for a zone.
1177  *
1178  * NOTE: This may only be called while the zone is being torn down, and not
1179  * during normal operation.  This is necessary in order that we do not have
1180  * to migrate CPUs to drain the per-CPU caches.
1181  *
1182  * Arguments:
1183  *	zone     The zone to drain, must be unlocked.
1184  *
1185  * Returns:
1186  *	Nothing
1187  */
1188 static void
1189 cache_drain(uma_zone_t zone)
1190 {
1191 	uma_cache_t cache;
1192 	uma_bucket_t bucket;
1193 	smr_seq_t seq;
1194 	int cpu;
1195 
1196 	/*
1197 	 * XXX: It is safe to not lock the per-CPU caches, because we're
1198 	 * tearing down the zone anyway.  I.e., there will be no further use
1199 	 * of the caches at this point.
1200 	 *
1201 	 * XXX: It would good to be able to assert that the zone is being
1202 	 * torn down to prevent improper use of cache_drain().
1203 	 */
1204 	seq = SMR_SEQ_INVALID;
1205 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
1206 		seq = smr_advance(zone->uz_smr);
1207 	CPU_FOREACH(cpu) {
1208 		cache = &zone->uz_cpu[cpu];
1209 		bucket = cache_bucket_unload_alloc(cache);
1210 		if (bucket != NULL)
1211 			bucket_free(zone, bucket, NULL);
1212 		bucket = cache_bucket_unload_free(cache);
1213 		if (bucket != NULL) {
1214 			bucket->ub_seq = seq;
1215 			bucket_free(zone, bucket, NULL);
1216 		}
1217 		bucket = cache_bucket_unload_cross(cache);
1218 		if (bucket != NULL) {
1219 			bucket->ub_seq = seq;
1220 			bucket_free(zone, bucket, NULL);
1221 		}
1222 	}
1223 	bucket_cache_reclaim(zone, true);
1224 }
1225 
1226 static void
1227 cache_shrink(uma_zone_t zone, void *unused)
1228 {
1229 
1230 	if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1231 		return;
1232 
1233 	zone->uz_bucket_size =
1234 	    (zone->uz_bucket_size_min + zone->uz_bucket_size) / 2;
1235 }
1236 
1237 static void
1238 cache_drain_safe_cpu(uma_zone_t zone, void *unused)
1239 {
1240 	uma_cache_t cache;
1241 	uma_bucket_t b1, b2, b3;
1242 	int domain;
1243 
1244 	if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1245 		return;
1246 
1247 	b1 = b2 = b3 = NULL;
1248 	critical_enter();
1249 	cache = &zone->uz_cpu[curcpu];
1250 	domain = PCPU_GET(domain);
1251 	b1 = cache_bucket_unload_alloc(cache);
1252 
1253 	/*
1254 	 * Don't flush SMR zone buckets.  This leaves the zone without a
1255 	 * bucket and forces every free to synchronize().
1256 	 */
1257 	if ((zone->uz_flags & UMA_ZONE_SMR) == 0) {
1258 		b2 = cache_bucket_unload_free(cache);
1259 		b3 = cache_bucket_unload_cross(cache);
1260 	}
1261 	critical_exit();
1262 
1263 	if (b1 != NULL)
1264 		zone_free_bucket(zone, b1, NULL, domain, false);
1265 	if (b2 != NULL)
1266 		zone_free_bucket(zone, b2, NULL, domain, false);
1267 	if (b3 != NULL) {
1268 		/* Adjust the domain so it goes to zone_free_cross. */
1269 		domain = (domain + 1) % vm_ndomains;
1270 		zone_free_bucket(zone, b3, NULL, domain, false);
1271 	}
1272 }
1273 
1274 /*
1275  * Safely drain per-CPU caches of a zone(s) to alloc bucket.
1276  * This is an expensive call because it needs to bind to all CPUs
1277  * one by one and enter a critical section on each of them in order
1278  * to safely access their cache buckets.
1279  * Zone lock must not be held on call this function.
1280  */
1281 static void
1282 pcpu_cache_drain_safe(uma_zone_t zone)
1283 {
1284 	int cpu;
1285 
1286 	/*
1287 	 * Polite bucket sizes shrinking was not enough, shrink aggressively.
1288 	 */
1289 	if (zone)
1290 		cache_shrink(zone, NULL);
1291 	else
1292 		zone_foreach(cache_shrink, NULL);
1293 
1294 	CPU_FOREACH(cpu) {
1295 		thread_lock(curthread);
1296 		sched_bind(curthread, cpu);
1297 		thread_unlock(curthread);
1298 
1299 		if (zone)
1300 			cache_drain_safe_cpu(zone, NULL);
1301 		else
1302 			zone_foreach(cache_drain_safe_cpu, NULL);
1303 	}
1304 	thread_lock(curthread);
1305 	sched_unbind(curthread);
1306 	thread_unlock(curthread);
1307 }
1308 
1309 /*
1310  * Reclaim cached buckets from a zone.  All buckets are reclaimed if the caller
1311  * requested a drain, otherwise the per-domain caches are trimmed to either
1312  * estimated working set size.
1313  */
1314 static void
1315 bucket_cache_reclaim(uma_zone_t zone, bool drain)
1316 {
1317 	uma_zone_domain_t zdom;
1318 	uma_bucket_t bucket;
1319 	long target;
1320 	int i;
1321 
1322 	/*
1323 	 * Shrink the zone bucket size to ensure that the per-CPU caches
1324 	 * don't grow too large.
1325 	 */
1326 	if (zone->uz_bucket_size > zone->uz_bucket_size_min)
1327 		zone->uz_bucket_size--;
1328 
1329 	for (i = 0; i < vm_ndomains; i++) {
1330 		/*
1331 		 * The cross bucket is partially filled and not part of
1332 		 * the item count.  Reclaim it individually here.
1333 		 */
1334 		zdom = ZDOM_GET(zone, i);
1335 		if ((zone->uz_flags & UMA_ZONE_SMR) == 0 || drain) {
1336 			ZONE_CROSS_LOCK(zone);
1337 			bucket = zdom->uzd_cross;
1338 			zdom->uzd_cross = NULL;
1339 			ZONE_CROSS_UNLOCK(zone);
1340 			if (bucket != NULL)
1341 				bucket_free(zone, bucket, NULL);
1342 		}
1343 
1344 		/*
1345 		 * If we were asked to drain the zone, we are done only once
1346 		 * this bucket cache is empty.  Otherwise, we reclaim items in
1347 		 * excess of the zone's estimated working set size.  If the
1348 		 * difference nitems - imin is larger than the WSS estimate,
1349 		 * then the estimate will grow at the end of this interval and
1350 		 * we ignore the historical average.
1351 		 */
1352 		ZDOM_LOCK(zdom);
1353 		target = drain ? 0 : lmax(zdom->uzd_wss, zdom->uzd_nitems -
1354 		    zdom->uzd_imin);
1355 		while (zdom->uzd_nitems > target) {
1356 			bucket = zone_fetch_bucket(zone, zdom, true);
1357 			if (bucket == NULL)
1358 				break;
1359 			bucket_free(zone, bucket, NULL);
1360 			ZDOM_LOCK(zdom);
1361 		}
1362 		ZDOM_UNLOCK(zdom);
1363 	}
1364 }
1365 
1366 static void
1367 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start)
1368 {
1369 	uint8_t *mem;
1370 	int i;
1371 	uint8_t flags;
1372 
1373 	CTR4(KTR_UMA, "keg_free_slab keg %s(%p) slab %p, returning %d bytes",
1374 	    keg->uk_name, keg, slab, PAGE_SIZE * keg->uk_ppera);
1375 
1376 	mem = slab_data(slab, keg);
1377 	flags = slab->us_flags;
1378 	i = start;
1379 	if (keg->uk_fini != NULL) {
1380 		for (i--; i > -1; i--)
1381 #ifdef INVARIANTS
1382 		/*
1383 		 * trash_fini implies that dtor was trash_dtor. trash_fini
1384 		 * would check that memory hasn't been modified since free,
1385 		 * which executed trash_dtor.
1386 		 * That's why we need to run uma_dbg_kskip() check here,
1387 		 * albeit we don't make skip check for other init/fini
1388 		 * invocations.
1389 		 */
1390 		if (!uma_dbg_kskip(keg, slab_item(slab, keg, i)) ||
1391 		    keg->uk_fini != trash_fini)
1392 #endif
1393 			keg->uk_fini(slab_item(slab, keg, i), keg->uk_size);
1394 	}
1395 	if (keg->uk_flags & UMA_ZFLAG_OFFPAGE)
1396 		zone_free_item(slabzone(keg->uk_ipers), slab_tohashslab(slab),
1397 		    NULL, SKIP_NONE);
1398 	keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags);
1399 	uma_total_dec(PAGE_SIZE * keg->uk_ppera);
1400 }
1401 
1402 /*
1403  * Frees pages from a keg back to the system.  This is done on demand from
1404  * the pageout daemon.
1405  *
1406  * Returns nothing.
1407  */
1408 static void
1409 keg_drain(uma_keg_t keg)
1410 {
1411 	struct slabhead freeslabs;
1412 	uma_domain_t dom;
1413 	uma_slab_t slab, tmp;
1414 	int i, n;
1415 
1416 	if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL)
1417 		return;
1418 
1419 	for (i = 0; i < vm_ndomains; i++) {
1420 		CTR4(KTR_UMA, "keg_drain %s(%p) domain %d free items: %u",
1421 		    keg->uk_name, keg, i, dom->ud_free_items);
1422 		dom = &keg->uk_domain[i];
1423 		LIST_INIT(&freeslabs);
1424 
1425 		KEG_LOCK(keg, i);
1426 		if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0) {
1427 			LIST_FOREACH(slab, &dom->ud_free_slab, us_link)
1428 				UMA_HASH_REMOVE(&keg->uk_hash, slab);
1429 		}
1430 		n = dom->ud_free_slabs;
1431 		LIST_SWAP(&freeslabs, &dom->ud_free_slab, uma_slab, us_link);
1432 		dom->ud_free_slabs = 0;
1433 		dom->ud_free_items -= n * keg->uk_ipers;
1434 		dom->ud_pages -= n * keg->uk_ppera;
1435 		KEG_UNLOCK(keg, i);
1436 
1437 		LIST_FOREACH_SAFE(slab, &freeslabs, us_link, tmp)
1438 			keg_free_slab(keg, slab, keg->uk_ipers);
1439 	}
1440 }
1441 
1442 static void
1443 zone_reclaim(uma_zone_t zone, int waitok, bool drain)
1444 {
1445 
1446 	/*
1447 	 * Set draining to interlock with zone_dtor() so we can release our
1448 	 * locks as we go.  Only dtor() should do a WAITOK call since it
1449 	 * is the only call that knows the structure will still be available
1450 	 * when it wakes up.
1451 	 */
1452 	ZONE_LOCK(zone);
1453 	while (zone->uz_flags & UMA_ZFLAG_RECLAIMING) {
1454 		if (waitok == M_NOWAIT)
1455 			goto out;
1456 		msleep(zone, &ZDOM_GET(zone, 0)->uzd_lock, PVM, "zonedrain",
1457 		    1);
1458 	}
1459 	zone->uz_flags |= UMA_ZFLAG_RECLAIMING;
1460 	ZONE_UNLOCK(zone);
1461 	bucket_cache_reclaim(zone, drain);
1462 
1463 	/*
1464 	 * The DRAINING flag protects us from being freed while
1465 	 * we're running.  Normally the uma_rwlock would protect us but we
1466 	 * must be able to release and acquire the right lock for each keg.
1467 	 */
1468 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0)
1469 		keg_drain(zone->uz_keg);
1470 	ZONE_LOCK(zone);
1471 	zone->uz_flags &= ~UMA_ZFLAG_RECLAIMING;
1472 	wakeup(zone);
1473 out:
1474 	ZONE_UNLOCK(zone);
1475 }
1476 
1477 static void
1478 zone_drain(uma_zone_t zone, void *unused)
1479 {
1480 
1481 	zone_reclaim(zone, M_NOWAIT, true);
1482 }
1483 
1484 static void
1485 zone_trim(uma_zone_t zone, void *unused)
1486 {
1487 
1488 	zone_reclaim(zone, M_NOWAIT, false);
1489 }
1490 
1491 /*
1492  * Allocate a new slab for a keg and inserts it into the partial slab list.
1493  * The keg should be unlocked on entry.  If the allocation succeeds it will
1494  * be locked on return.
1495  *
1496  * Arguments:
1497  *	flags   Wait flags for the item initialization routine
1498  *	aflags  Wait flags for the slab allocation
1499  *
1500  * Returns:
1501  *	The slab that was allocated or NULL if there is no memory and the
1502  *	caller specified M_NOWAIT.
1503  */
1504 static uma_slab_t
1505 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int domain, int flags,
1506     int aflags)
1507 {
1508 	uma_domain_t dom;
1509 	uma_alloc allocf;
1510 	uma_slab_t slab;
1511 	unsigned long size;
1512 	uint8_t *mem;
1513 	uint8_t sflags;
1514 	int i;
1515 
1516 	KASSERT(domain >= 0 && domain < vm_ndomains,
1517 	    ("keg_alloc_slab: domain %d out of range", domain));
1518 
1519 	allocf = keg->uk_allocf;
1520 	slab = NULL;
1521 	mem = NULL;
1522 	if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) {
1523 		uma_hash_slab_t hslab;
1524 		hslab = zone_alloc_item(slabzone(keg->uk_ipers), NULL,
1525 		    domain, aflags);
1526 		if (hslab == NULL)
1527 			goto fail;
1528 		slab = &hslab->uhs_slab;
1529 	}
1530 
1531 	/*
1532 	 * This reproduces the old vm_zone behavior of zero filling pages the
1533 	 * first time they are added to a zone.
1534 	 *
1535 	 * Malloced items are zeroed in uma_zalloc.
1536 	 */
1537 
1538 	if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
1539 		aflags |= M_ZERO;
1540 	else
1541 		aflags &= ~M_ZERO;
1542 
1543 	if (keg->uk_flags & UMA_ZONE_NODUMP)
1544 		aflags |= M_NODUMP;
1545 
1546 	/* zone is passed for legacy reasons. */
1547 	size = keg->uk_ppera * PAGE_SIZE;
1548 	mem = allocf(zone, size, domain, &sflags, aflags);
1549 	if (mem == NULL) {
1550 		if (keg->uk_flags & UMA_ZFLAG_OFFPAGE)
1551 			zone_free_item(slabzone(keg->uk_ipers),
1552 			    slab_tohashslab(slab), NULL, SKIP_NONE);
1553 		goto fail;
1554 	}
1555 	uma_total_inc(size);
1556 
1557 	/* For HASH zones all pages go to the same uma_domain. */
1558 	if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
1559 		domain = 0;
1560 
1561 	/* Point the slab into the allocated memory */
1562 	if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE))
1563 		slab = (uma_slab_t )(mem + keg->uk_pgoff);
1564 	else
1565 		slab_tohashslab(slab)->uhs_data = mem;
1566 
1567 	if (keg->uk_flags & UMA_ZFLAG_VTOSLAB)
1568 		for (i = 0; i < keg->uk_ppera; i++)
1569 			vsetzoneslab((vm_offset_t)mem + (i * PAGE_SIZE),
1570 			    zone, slab);
1571 
1572 	slab->us_freecount = keg->uk_ipers;
1573 	slab->us_flags = sflags;
1574 	slab->us_domain = domain;
1575 
1576 	BIT_FILL(keg->uk_ipers, &slab->us_free);
1577 #ifdef INVARIANTS
1578 	BIT_ZERO(keg->uk_ipers, slab_dbg_bits(slab, keg));
1579 #endif
1580 
1581 	if (keg->uk_init != NULL) {
1582 		for (i = 0; i < keg->uk_ipers; i++)
1583 			if (keg->uk_init(slab_item(slab, keg, i),
1584 			    keg->uk_size, flags) != 0)
1585 				break;
1586 		if (i != keg->uk_ipers) {
1587 			keg_free_slab(keg, slab, i);
1588 			goto fail;
1589 		}
1590 	}
1591 	KEG_LOCK(keg, domain);
1592 
1593 	CTR3(KTR_UMA, "keg_alloc_slab: allocated slab %p for %s(%p)",
1594 	    slab, keg->uk_name, keg);
1595 
1596 	if (keg->uk_flags & UMA_ZFLAG_HASH)
1597 		UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
1598 
1599 	/*
1600 	 * If we got a slab here it's safe to mark it partially used
1601 	 * and return.  We assume that the caller is going to remove
1602 	 * at least one item.
1603 	 */
1604 	dom = &keg->uk_domain[domain];
1605 	LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
1606 	dom->ud_pages += keg->uk_ppera;
1607 	dom->ud_free_items += keg->uk_ipers;
1608 
1609 	return (slab);
1610 
1611 fail:
1612 	return (NULL);
1613 }
1614 
1615 /*
1616  * This function is intended to be used early on in place of page_alloc() so
1617  * that we may use the boot time page cache to satisfy allocations before
1618  * the VM is ready.
1619  */
1620 static void *
1621 startup_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1622     int wait)
1623 {
1624 	vm_paddr_t pa;
1625 	vm_page_t m;
1626 	void *mem;
1627 	int pages;
1628 	int i;
1629 
1630 	pages = howmany(bytes, PAGE_SIZE);
1631 	KASSERT(pages > 0, ("%s can't reserve 0 pages", __func__));
1632 
1633 	*pflag = UMA_SLAB_BOOT;
1634 	m = vm_page_alloc_contig_domain(NULL, 0, domain,
1635 	    malloc2vm_flags(wait) | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED, pages,
1636 	    (vm_paddr_t)0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT);
1637 	if (m == NULL)
1638 		return (NULL);
1639 
1640 	pa = VM_PAGE_TO_PHYS(m);
1641 	for (i = 0; i < pages; i++, pa += PAGE_SIZE) {
1642 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
1643     defined(__riscv) || defined(__powerpc64__)
1644 		if ((wait & M_NODUMP) == 0)
1645 			dump_add_page(pa);
1646 #endif
1647 	}
1648 	/* Allocate KVA and indirectly advance bootmem. */
1649 	mem = (void *)pmap_map(&bootmem, m->phys_addr,
1650 	    m->phys_addr + (pages * PAGE_SIZE), VM_PROT_READ | VM_PROT_WRITE);
1651         if ((wait & M_ZERO) != 0)
1652                 bzero(mem, pages * PAGE_SIZE);
1653 
1654         return (mem);
1655 }
1656 
1657 static void
1658 startup_free(void *mem, vm_size_t bytes)
1659 {
1660 	vm_offset_t va;
1661 	vm_page_t m;
1662 
1663 	va = (vm_offset_t)mem;
1664 	m = PHYS_TO_VM_PAGE(pmap_kextract(va));
1665 	pmap_remove(kernel_pmap, va, va + bytes);
1666 	for (; bytes != 0; bytes -= PAGE_SIZE, m++) {
1667 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
1668     defined(__riscv) || defined(__powerpc64__)
1669 		dump_drop_page(VM_PAGE_TO_PHYS(m));
1670 #endif
1671 		vm_page_unwire_noq(m);
1672 		vm_page_free(m);
1673 	}
1674 }
1675 
1676 /*
1677  * Allocates a number of pages from the system
1678  *
1679  * Arguments:
1680  *	bytes  The number of bytes requested
1681  *	wait  Shall we wait?
1682  *
1683  * Returns:
1684  *	A pointer to the alloced memory or possibly
1685  *	NULL if M_NOWAIT is set.
1686  */
1687 static void *
1688 page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1689     int wait)
1690 {
1691 	void *p;	/* Returned page */
1692 
1693 	*pflag = UMA_SLAB_KERNEL;
1694 	p = (void *)kmem_malloc_domainset(DOMAINSET_FIXED(domain), bytes, wait);
1695 
1696 	return (p);
1697 }
1698 
1699 static void *
1700 pcpu_page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1701     int wait)
1702 {
1703 	struct pglist alloctail;
1704 	vm_offset_t addr, zkva;
1705 	int cpu, flags;
1706 	vm_page_t p, p_next;
1707 #ifdef NUMA
1708 	struct pcpu *pc;
1709 #endif
1710 
1711 	MPASS(bytes == (mp_maxid + 1) * PAGE_SIZE);
1712 
1713 	TAILQ_INIT(&alloctail);
1714 	flags = VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | VM_ALLOC_NOOBJ |
1715 	    malloc2vm_flags(wait);
1716 	*pflag = UMA_SLAB_KERNEL;
1717 	for (cpu = 0; cpu <= mp_maxid; cpu++) {
1718 		if (CPU_ABSENT(cpu)) {
1719 			p = vm_page_alloc(NULL, 0, flags);
1720 		} else {
1721 #ifndef NUMA
1722 			p = vm_page_alloc(NULL, 0, flags);
1723 #else
1724 			pc = pcpu_find(cpu);
1725 			if (__predict_false(VM_DOMAIN_EMPTY(pc->pc_domain)))
1726 				p = NULL;
1727 			else
1728 				p = vm_page_alloc_domain(NULL, 0,
1729 				    pc->pc_domain, flags);
1730 			if (__predict_false(p == NULL))
1731 				p = vm_page_alloc(NULL, 0, flags);
1732 #endif
1733 		}
1734 		if (__predict_false(p == NULL))
1735 			goto fail;
1736 		TAILQ_INSERT_TAIL(&alloctail, p, listq);
1737 	}
1738 	if ((addr = kva_alloc(bytes)) == 0)
1739 		goto fail;
1740 	zkva = addr;
1741 	TAILQ_FOREACH(p, &alloctail, listq) {
1742 		pmap_qenter(zkva, &p, 1);
1743 		zkva += PAGE_SIZE;
1744 	}
1745 	return ((void*)addr);
1746 fail:
1747 	TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1748 		vm_page_unwire_noq(p);
1749 		vm_page_free(p);
1750 	}
1751 	return (NULL);
1752 }
1753 
1754 /*
1755  * Allocates a number of pages from within an object
1756  *
1757  * Arguments:
1758  *	bytes  The number of bytes requested
1759  *	wait   Shall we wait?
1760  *
1761  * Returns:
1762  *	A pointer to the alloced memory or possibly
1763  *	NULL if M_NOWAIT is set.
1764  */
1765 static void *
1766 noobj_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags,
1767     int wait)
1768 {
1769 	TAILQ_HEAD(, vm_page) alloctail;
1770 	u_long npages;
1771 	vm_offset_t retkva, zkva;
1772 	vm_page_t p, p_next;
1773 	uma_keg_t keg;
1774 
1775 	TAILQ_INIT(&alloctail);
1776 	keg = zone->uz_keg;
1777 
1778 	npages = howmany(bytes, PAGE_SIZE);
1779 	while (npages > 0) {
1780 		p = vm_page_alloc_domain(NULL, 0, domain, VM_ALLOC_INTERRUPT |
1781 		    VM_ALLOC_WIRED | VM_ALLOC_NOOBJ |
1782 		    ((wait & M_WAITOK) != 0 ? VM_ALLOC_WAITOK :
1783 		    VM_ALLOC_NOWAIT));
1784 		if (p != NULL) {
1785 			/*
1786 			 * Since the page does not belong to an object, its
1787 			 * listq is unused.
1788 			 */
1789 			TAILQ_INSERT_TAIL(&alloctail, p, listq);
1790 			npages--;
1791 			continue;
1792 		}
1793 		/*
1794 		 * Page allocation failed, free intermediate pages and
1795 		 * exit.
1796 		 */
1797 		TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1798 			vm_page_unwire_noq(p);
1799 			vm_page_free(p);
1800 		}
1801 		return (NULL);
1802 	}
1803 	*flags = UMA_SLAB_PRIV;
1804 	zkva = keg->uk_kva +
1805 	    atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
1806 	retkva = zkva;
1807 	TAILQ_FOREACH(p, &alloctail, listq) {
1808 		pmap_qenter(zkva, &p, 1);
1809 		zkva += PAGE_SIZE;
1810 	}
1811 
1812 	return ((void *)retkva);
1813 }
1814 
1815 /*
1816  * Allocate physically contiguous pages.
1817  */
1818 static void *
1819 contig_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1820     int wait)
1821 {
1822 
1823 	*pflag = UMA_SLAB_KERNEL;
1824 	return ((void *)kmem_alloc_contig_domainset(DOMAINSET_FIXED(domain),
1825 	    bytes, wait, 0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT));
1826 }
1827 
1828 /*
1829  * Frees a number of pages to the system
1830  *
1831  * Arguments:
1832  *	mem   A pointer to the memory to be freed
1833  *	size  The size of the memory being freed
1834  *	flags The original p->us_flags field
1835  *
1836  * Returns:
1837  *	Nothing
1838  */
1839 static void
1840 page_free(void *mem, vm_size_t size, uint8_t flags)
1841 {
1842 
1843 	if ((flags & UMA_SLAB_BOOT) != 0) {
1844 		startup_free(mem, size);
1845 		return;
1846 	}
1847 
1848 	KASSERT((flags & UMA_SLAB_KERNEL) != 0,
1849 	    ("UMA: page_free used with invalid flags %x", flags));
1850 
1851 	kmem_free((vm_offset_t)mem, size);
1852 }
1853 
1854 /*
1855  * Frees pcpu zone allocations
1856  *
1857  * Arguments:
1858  *	mem   A pointer to the memory to be freed
1859  *	size  The size of the memory being freed
1860  *	flags The original p->us_flags field
1861  *
1862  * Returns:
1863  *	Nothing
1864  */
1865 static void
1866 pcpu_page_free(void *mem, vm_size_t size, uint8_t flags)
1867 {
1868 	vm_offset_t sva, curva;
1869 	vm_paddr_t paddr;
1870 	vm_page_t m;
1871 
1872 	MPASS(size == (mp_maxid+1)*PAGE_SIZE);
1873 
1874 	if ((flags & UMA_SLAB_BOOT) != 0) {
1875 		startup_free(mem, size);
1876 		return;
1877 	}
1878 
1879 	sva = (vm_offset_t)mem;
1880 	for (curva = sva; curva < sva + size; curva += PAGE_SIZE) {
1881 		paddr = pmap_kextract(curva);
1882 		m = PHYS_TO_VM_PAGE(paddr);
1883 		vm_page_unwire_noq(m);
1884 		vm_page_free(m);
1885 	}
1886 	pmap_qremove(sva, size >> PAGE_SHIFT);
1887 	kva_free(sva, size);
1888 }
1889 
1890 
1891 /*
1892  * Zero fill initializer
1893  *
1894  * Arguments/Returns follow uma_init specifications
1895  */
1896 static int
1897 zero_init(void *mem, int size, int flags)
1898 {
1899 	bzero(mem, size);
1900 	return (0);
1901 }
1902 
1903 #ifdef INVARIANTS
1904 struct noslabbits *
1905 slab_dbg_bits(uma_slab_t slab, uma_keg_t keg)
1906 {
1907 
1908 	return ((void *)((char *)&slab->us_free + BITSET_SIZE(keg->uk_ipers)));
1909 }
1910 #endif
1911 
1912 /*
1913  * Actual size of embedded struct slab (!OFFPAGE).
1914  */
1915 size_t
1916 slab_sizeof(int nitems)
1917 {
1918 	size_t s;
1919 
1920 	s = sizeof(struct uma_slab) + BITSET_SIZE(nitems) * SLAB_BITSETS;
1921 	return (roundup(s, UMA_ALIGN_PTR + 1));
1922 }
1923 
1924 /*
1925  * Size of memory for embedded slabs (!OFFPAGE).
1926  */
1927 size_t
1928 slab_space(int nitems)
1929 {
1930 	return (UMA_SLAB_SIZE - slab_sizeof(nitems));
1931 }
1932 
1933 #define	UMA_FIXPT_SHIFT	31
1934 #define	UMA_FRAC_FIXPT(n, d)						\
1935 	((uint32_t)(((uint64_t)(n) << UMA_FIXPT_SHIFT) / (d)))
1936 #define	UMA_FIXPT_PCT(f)						\
1937 	((u_int)(((uint64_t)100 * (f)) >> UMA_FIXPT_SHIFT))
1938 #define	UMA_PCT_FIXPT(pct)	UMA_FRAC_FIXPT((pct), 100)
1939 #define	UMA_MIN_EFF	UMA_PCT_FIXPT(100 - UMA_MAX_WASTE)
1940 
1941 /*
1942  * Compute the number of items that will fit in a slab.  If hdr is true, the
1943  * item count may be limited to provide space in the slab for an inline slab
1944  * header.  Otherwise, all slab space will be provided for item storage.
1945  */
1946 static u_int
1947 slab_ipers_hdr(u_int size, u_int rsize, u_int slabsize, bool hdr)
1948 {
1949 	u_int ipers;
1950 	u_int padpi;
1951 
1952 	/* The padding between items is not needed after the last item. */
1953 	padpi = rsize - size;
1954 
1955 	if (hdr) {
1956 		/*
1957 		 * Start with the maximum item count and remove items until
1958 		 * the slab header first alongside the allocatable memory.
1959 		 */
1960 		for (ipers = MIN(SLAB_MAX_SETSIZE,
1961 		    (slabsize + padpi - slab_sizeof(1)) / rsize);
1962 		    ipers > 0 &&
1963 		    ipers * rsize - padpi + slab_sizeof(ipers) > slabsize;
1964 		    ipers--)
1965 			continue;
1966 	} else {
1967 		ipers = MIN((slabsize + padpi) / rsize, SLAB_MAX_SETSIZE);
1968 	}
1969 
1970 	return (ipers);
1971 }
1972 
1973 /*
1974  * Compute the number of items that will fit in a slab for a startup zone.
1975  */
1976 int
1977 slab_ipers(size_t size, int align)
1978 {
1979 	int rsize;
1980 
1981 	rsize = roundup(size, align + 1); /* Assume no CACHESPREAD */
1982 	return (slab_ipers_hdr(size, rsize, UMA_SLAB_SIZE, true));
1983 }
1984 
1985 struct keg_layout_result {
1986 	u_int format;
1987 	u_int slabsize;
1988 	u_int ipers;
1989 	u_int eff;
1990 };
1991 
1992 static void
1993 keg_layout_one(uma_keg_t keg, u_int rsize, u_int slabsize, u_int fmt,
1994     struct keg_layout_result *kl)
1995 {
1996 	u_int total;
1997 
1998 	kl->format = fmt;
1999 	kl->slabsize = slabsize;
2000 
2001 	/* Handle INTERNAL as inline with an extra page. */
2002 	if ((fmt & UMA_ZFLAG_INTERNAL) != 0) {
2003 		kl->format &= ~UMA_ZFLAG_INTERNAL;
2004 		kl->slabsize += PAGE_SIZE;
2005 	}
2006 
2007 	kl->ipers = slab_ipers_hdr(keg->uk_size, rsize, kl->slabsize,
2008 	    (fmt & UMA_ZFLAG_OFFPAGE) == 0);
2009 
2010 	/* Account for memory used by an offpage slab header. */
2011 	total = kl->slabsize;
2012 	if ((fmt & UMA_ZFLAG_OFFPAGE) != 0)
2013 		total += slabzone(kl->ipers)->uz_keg->uk_rsize;
2014 
2015 	kl->eff = UMA_FRAC_FIXPT(kl->ipers * rsize, total);
2016 }
2017 
2018 /*
2019  * Determine the format of a uma keg.  This determines where the slab header
2020  * will be placed (inline or offpage) and calculates ipers, rsize, and ppera.
2021  *
2022  * Arguments
2023  *	keg  The zone we should initialize
2024  *
2025  * Returns
2026  *	Nothing
2027  */
2028 static void
2029 keg_layout(uma_keg_t keg)
2030 {
2031 	struct keg_layout_result kl = {}, kl_tmp;
2032 	u_int fmts[2];
2033 	u_int alignsize;
2034 	u_int nfmt;
2035 	u_int pages;
2036 	u_int rsize;
2037 	u_int slabsize;
2038 	u_int i, j;
2039 
2040 	KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
2041 	    (keg->uk_size <= UMA_PCPU_ALLOC_SIZE &&
2042 	     (keg->uk_flags & UMA_ZONE_CACHESPREAD) == 0),
2043 	    ("%s: cannot configure for PCPU: keg=%s, size=%u, flags=0x%b",
2044 	     __func__, keg->uk_name, keg->uk_size, keg->uk_flags,
2045 	     PRINT_UMA_ZFLAGS));
2046 	KASSERT((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) == 0 ||
2047 	    (keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0,
2048 	    ("%s: incompatible flags 0x%b", __func__, keg->uk_flags,
2049 	     PRINT_UMA_ZFLAGS));
2050 
2051 	alignsize = keg->uk_align + 1;
2052 
2053 	/*
2054 	 * Calculate the size of each allocation (rsize) according to
2055 	 * alignment.  If the requested size is smaller than we have
2056 	 * allocation bits for we round it up.
2057 	 */
2058 	rsize = MAX(keg->uk_size, UMA_SMALLEST_UNIT);
2059 	rsize = roundup2(rsize, alignsize);
2060 
2061 	if ((keg->uk_flags & UMA_ZONE_CACHESPREAD) != 0) {
2062 		/*
2063 		 * We want one item to start on every align boundary in a page.
2064 		 * To do this we will span pages.  We will also extend the item
2065 		 * by the size of align if it is an even multiple of align.
2066 		 * Otherwise, it would fall on the same boundary every time.
2067 		 */
2068 		if ((rsize & alignsize) == 0)
2069 			rsize += alignsize;
2070 		slabsize = rsize * (PAGE_SIZE / alignsize);
2071 		slabsize = MIN(slabsize, rsize * SLAB_MAX_SETSIZE);
2072 		slabsize = MIN(slabsize, UMA_CACHESPREAD_MAX_SIZE);
2073 		slabsize = round_page(slabsize);
2074 	} else {
2075 		/*
2076 		 * Start with a slab size of as many pages as it takes to
2077 		 * represent a single item.  We will try to fit as many
2078 		 * additional items into the slab as possible.
2079 		 */
2080 		slabsize = round_page(keg->uk_size);
2081 	}
2082 
2083 	/* Build a list of all of the available formats for this keg. */
2084 	nfmt = 0;
2085 
2086 	/* Evaluate an inline slab layout. */
2087 	if ((keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0)
2088 		fmts[nfmt++] = 0;
2089 
2090 	/* TODO: vm_page-embedded slab. */
2091 
2092 	/*
2093 	 * We can't do OFFPAGE if we're internal or if we've been
2094 	 * asked to not go to the VM for buckets.  If we do this we
2095 	 * may end up going to the VM for slabs which we do not want
2096 	 * to do if we're UMA_ZONE_VM, which clearly forbids it.
2097 	 * In those cases, evaluate a pseudo-format called INTERNAL
2098 	 * which has an inline slab header and one extra page to
2099 	 * guarantee that it fits.
2100 	 *
2101 	 * Otherwise, see if using an OFFPAGE slab will improve our
2102 	 * efficiency.
2103 	 */
2104 	if ((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) != 0)
2105 		fmts[nfmt++] = UMA_ZFLAG_INTERNAL;
2106 	else
2107 		fmts[nfmt++] = UMA_ZFLAG_OFFPAGE;
2108 
2109 	/*
2110 	 * Choose a slab size and format which satisfy the minimum efficiency.
2111 	 * Prefer the smallest slab size that meets the constraints.
2112 	 *
2113 	 * Start with a minimum slab size, to accommodate CACHESPREAD.  Then,
2114 	 * for small items (up to PAGE_SIZE), the iteration increment is one
2115 	 * page; and for large items, the increment is one item.
2116 	 */
2117 	i = (slabsize + rsize - keg->uk_size) / MAX(PAGE_SIZE, rsize);
2118 	KASSERT(i >= 1, ("keg %s(%p) flags=0x%b slabsize=%u, rsize=%u, i=%u",
2119 	    keg->uk_name, keg, keg->uk_flags, PRINT_UMA_ZFLAGS, slabsize,
2120 	    rsize, i));
2121 	for ( ; ; i++) {
2122 		slabsize = (rsize <= PAGE_SIZE) ? ptoa(i) :
2123 		    round_page(rsize * (i - 1) + keg->uk_size);
2124 
2125 		for (j = 0; j < nfmt; j++) {
2126 			/* Only if we have no viable format yet. */
2127 			if ((fmts[j] & UMA_ZFLAG_INTERNAL) != 0 &&
2128 			    kl.ipers > 0)
2129 				continue;
2130 
2131 			keg_layout_one(keg, rsize, slabsize, fmts[j], &kl_tmp);
2132 			if (kl_tmp.eff <= kl.eff)
2133 				continue;
2134 
2135 			kl = kl_tmp;
2136 
2137 			CTR6(KTR_UMA, "keg %s layout: format %#x "
2138 			    "(ipers %u * rsize %u) / slabsize %#x = %u%% eff",
2139 			    keg->uk_name, kl.format, kl.ipers, rsize,
2140 			    kl.slabsize, UMA_FIXPT_PCT(kl.eff));
2141 
2142 			/* Stop when we reach the minimum efficiency. */
2143 			if (kl.eff >= UMA_MIN_EFF)
2144 				break;
2145 		}
2146 
2147 		if (kl.eff >= UMA_MIN_EFF || !multipage_slabs ||
2148 		    slabsize >= SLAB_MAX_SETSIZE * rsize ||
2149 		    (keg->uk_flags & (UMA_ZONE_PCPU | UMA_ZONE_CONTIG)) != 0)
2150 			break;
2151 	}
2152 
2153 	pages = atop(kl.slabsize);
2154 	if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
2155 		pages *= mp_maxid + 1;
2156 
2157 	keg->uk_rsize = rsize;
2158 	keg->uk_ipers = kl.ipers;
2159 	keg->uk_ppera = pages;
2160 	keg->uk_flags |= kl.format;
2161 
2162 	/*
2163 	 * How do we find the slab header if it is offpage or if not all item
2164 	 * start addresses are in the same page?  We could solve the latter
2165 	 * case with vaddr alignment, but we don't.
2166 	 */
2167 	if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0 ||
2168 	    (keg->uk_ipers - 1) * rsize >= PAGE_SIZE) {
2169 		if ((keg->uk_flags & UMA_ZONE_NOTPAGE) != 0)
2170 			keg->uk_flags |= UMA_ZFLAG_HASH;
2171 		else
2172 			keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2173 	}
2174 
2175 	CTR6(KTR_UMA, "%s: keg=%s, flags=%#x, rsize=%u, ipers=%u, ppera=%u",
2176 	    __func__, keg->uk_name, keg->uk_flags, rsize, keg->uk_ipers,
2177 	    pages);
2178 	KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_MAX_SETSIZE,
2179 	    ("%s: keg=%s, flags=0x%b, rsize=%u, ipers=%u, ppera=%u", __func__,
2180 	     keg->uk_name, keg->uk_flags, PRINT_UMA_ZFLAGS, rsize,
2181 	     keg->uk_ipers, pages));
2182 }
2183 
2184 /*
2185  * Keg header ctor.  This initializes all fields, locks, etc.  And inserts
2186  * the keg onto the global keg list.
2187  *
2188  * Arguments/Returns follow uma_ctor specifications
2189  *	udata  Actually uma_kctor_args
2190  */
2191 static int
2192 keg_ctor(void *mem, int size, void *udata, int flags)
2193 {
2194 	struct uma_kctor_args *arg = udata;
2195 	uma_keg_t keg = mem;
2196 	uma_zone_t zone;
2197 	int i;
2198 
2199 	bzero(keg, size);
2200 	keg->uk_size = arg->size;
2201 	keg->uk_init = arg->uminit;
2202 	keg->uk_fini = arg->fini;
2203 	keg->uk_align = arg->align;
2204 	keg->uk_reserve = 0;
2205 	keg->uk_flags = arg->flags;
2206 
2207 	/*
2208 	 * We use a global round-robin policy by default.  Zones with
2209 	 * UMA_ZONE_FIRSTTOUCH set will use first-touch instead, in which
2210 	 * case the iterator is never run.
2211 	 */
2212 	keg->uk_dr.dr_policy = DOMAINSET_RR();
2213 	keg->uk_dr.dr_iter = 0;
2214 
2215 	/*
2216 	 * The master zone is passed to us at keg-creation time.
2217 	 */
2218 	zone = arg->zone;
2219 	keg->uk_name = zone->uz_name;
2220 
2221 	if (arg->flags & UMA_ZONE_ZINIT)
2222 		keg->uk_init = zero_init;
2223 
2224 	if (arg->flags & UMA_ZONE_MALLOC)
2225 		keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2226 
2227 #ifndef SMP
2228 	keg->uk_flags &= ~UMA_ZONE_PCPU;
2229 #endif
2230 
2231 	keg_layout(keg);
2232 
2233 	/*
2234 	 * Use a first-touch NUMA policy for kegs that pmap_extract() will
2235 	 * work on.  Use round-robin for everything else.
2236 	 *
2237 	 * Zones may override the default by specifying either.
2238 	 */
2239 #ifdef NUMA
2240 	if ((keg->uk_flags &
2241 	    (UMA_ZONE_ROUNDROBIN | UMA_ZFLAG_CACHE | UMA_ZONE_NOTPAGE)) == 0)
2242 		keg->uk_flags |= UMA_ZONE_FIRSTTOUCH;
2243 	else if ((keg->uk_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2244 		keg->uk_flags |= UMA_ZONE_ROUNDROBIN;
2245 #endif
2246 
2247 	/*
2248 	 * If we haven't booted yet we need allocations to go through the
2249 	 * startup cache until the vm is ready.
2250 	 */
2251 #ifdef UMA_MD_SMALL_ALLOC
2252 	if (keg->uk_ppera == 1)
2253 		keg->uk_allocf = uma_small_alloc;
2254 	else
2255 #endif
2256 	if (booted < BOOT_KVA)
2257 		keg->uk_allocf = startup_alloc;
2258 	else if (keg->uk_flags & UMA_ZONE_PCPU)
2259 		keg->uk_allocf = pcpu_page_alloc;
2260 	else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 && keg->uk_ppera > 1)
2261 		keg->uk_allocf = contig_alloc;
2262 	else
2263 		keg->uk_allocf = page_alloc;
2264 #ifdef UMA_MD_SMALL_ALLOC
2265 	if (keg->uk_ppera == 1)
2266 		keg->uk_freef = uma_small_free;
2267 	else
2268 #endif
2269 	if (keg->uk_flags & UMA_ZONE_PCPU)
2270 		keg->uk_freef = pcpu_page_free;
2271 	else
2272 		keg->uk_freef = page_free;
2273 
2274 	/*
2275 	 * Initialize keg's locks.
2276 	 */
2277 	for (i = 0; i < vm_ndomains; i++)
2278 		KEG_LOCK_INIT(keg, i, (arg->flags & UMA_ZONE_MTXCLASS));
2279 
2280 	/*
2281 	 * If we're putting the slab header in the actual page we need to
2282 	 * figure out where in each page it goes.  See slab_sizeof
2283 	 * definition.
2284 	 */
2285 	if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE)) {
2286 		size_t shsize;
2287 
2288 		shsize = slab_sizeof(keg->uk_ipers);
2289 		keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - shsize;
2290 		/*
2291 		 * The only way the following is possible is if with our
2292 		 * UMA_ALIGN_PTR adjustments we are now bigger than
2293 		 * UMA_SLAB_SIZE.  I haven't checked whether this is
2294 		 * mathematically possible for all cases, so we make
2295 		 * sure here anyway.
2296 		 */
2297 		KASSERT(keg->uk_pgoff + shsize <= PAGE_SIZE * keg->uk_ppera,
2298 		    ("zone %s ipers %d rsize %d size %d slab won't fit",
2299 		    zone->uz_name, keg->uk_ipers, keg->uk_rsize, keg->uk_size));
2300 	}
2301 
2302 	if (keg->uk_flags & UMA_ZFLAG_HASH)
2303 		hash_alloc(&keg->uk_hash, 0);
2304 
2305 	CTR3(KTR_UMA, "keg_ctor %p zone %s(%p)", keg, zone->uz_name, zone);
2306 
2307 	LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
2308 
2309 	rw_wlock(&uma_rwlock);
2310 	LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
2311 	rw_wunlock(&uma_rwlock);
2312 	return (0);
2313 }
2314 
2315 static void
2316 zone_kva_available(uma_zone_t zone, void *unused)
2317 {
2318 	uma_keg_t keg;
2319 
2320 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
2321 		return;
2322 	KEG_GET(zone, keg);
2323 
2324 	if (keg->uk_allocf == startup_alloc) {
2325 		/* Switch to the real allocator. */
2326 		if (keg->uk_flags & UMA_ZONE_PCPU)
2327 			keg->uk_allocf = pcpu_page_alloc;
2328 		else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 &&
2329 		    keg->uk_ppera > 1)
2330 			keg->uk_allocf = contig_alloc;
2331 		else
2332 			keg->uk_allocf = page_alloc;
2333 	}
2334 }
2335 
2336 static void
2337 zone_alloc_counters(uma_zone_t zone, void *unused)
2338 {
2339 
2340 	zone->uz_allocs = counter_u64_alloc(M_WAITOK);
2341 	zone->uz_frees = counter_u64_alloc(M_WAITOK);
2342 	zone->uz_fails = counter_u64_alloc(M_WAITOK);
2343 	zone->uz_xdomain = counter_u64_alloc(M_WAITOK);
2344 }
2345 
2346 static void
2347 zone_alloc_sysctl(uma_zone_t zone, void *unused)
2348 {
2349 	uma_zone_domain_t zdom;
2350 	uma_domain_t dom;
2351 	uma_keg_t keg;
2352 	struct sysctl_oid *oid, *domainoid;
2353 	int domains, i, cnt;
2354 	static const char *nokeg = "cache zone";
2355 	char *c;
2356 
2357 	/*
2358 	 * Make a sysctl safe copy of the zone name by removing
2359 	 * any special characters and handling dups by appending
2360 	 * an index.
2361 	 */
2362 	if (zone->uz_namecnt != 0) {
2363 		/* Count the number of decimal digits and '_' separator. */
2364 		for (i = 1, cnt = zone->uz_namecnt; cnt != 0; i++)
2365 			cnt /= 10;
2366 		zone->uz_ctlname = malloc(strlen(zone->uz_name) + i + 1,
2367 		    M_UMA, M_WAITOK);
2368 		sprintf(zone->uz_ctlname, "%s_%d", zone->uz_name,
2369 		    zone->uz_namecnt);
2370 	} else
2371 		zone->uz_ctlname = strdup(zone->uz_name, M_UMA);
2372 	for (c = zone->uz_ctlname; *c != '\0'; c++)
2373 		if (strchr("./\\ -", *c) != NULL)
2374 			*c = '_';
2375 
2376 	/*
2377 	 * Basic parameters at the root.
2378 	 */
2379 	zone->uz_oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_vm_uma),
2380 	    OID_AUTO, zone->uz_ctlname, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2381 	oid = zone->uz_oid;
2382 	SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2383 	    "size", CTLFLAG_RD, &zone->uz_size, 0, "Allocation size");
2384 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2385 	    "flags", CTLFLAG_RD | CTLTYPE_STRING | CTLFLAG_MPSAFE,
2386 	    zone, 0, sysctl_handle_uma_zone_flags, "A",
2387 	    "Allocator configuration flags");
2388 	SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2389 	    "bucket_size", CTLFLAG_RD, &zone->uz_bucket_size, 0,
2390 	    "Desired per-cpu cache size");
2391 	SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2392 	    "bucket_size_max", CTLFLAG_RD, &zone->uz_bucket_size_max, 0,
2393 	    "Maximum allowed per-cpu cache size");
2394 
2395 	/*
2396 	 * keg if present.
2397 	 */
2398 	if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
2399 		domains = vm_ndomains;
2400 	else
2401 		domains = 1;
2402 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2403 	    "keg", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2404 	keg = zone->uz_keg;
2405 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0) {
2406 		SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2407 		    "name", CTLFLAG_RD, keg->uk_name, "Keg name");
2408 		SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2409 		    "rsize", CTLFLAG_RD, &keg->uk_rsize, 0,
2410 		    "Real object size with alignment");
2411 		SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2412 		    "ppera", CTLFLAG_RD, &keg->uk_ppera, 0,
2413 		    "pages per-slab allocation");
2414 		SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2415 		    "ipers", CTLFLAG_RD, &keg->uk_ipers, 0,
2416 		    "items available per-slab");
2417 		SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2418 		    "align", CTLFLAG_RD, &keg->uk_align, 0,
2419 		    "item alignment mask");
2420 		SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2421 		    "efficiency", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2422 		    keg, 0, sysctl_handle_uma_slab_efficiency, "I",
2423 		    "Slab utilization (100 - internal fragmentation %)");
2424 		domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(oid),
2425 		    OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2426 		for (i = 0; i < domains; i++) {
2427 			dom = &keg->uk_domain[i];
2428 			oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2429 			    OID_AUTO, VM_DOMAIN(i)->vmd_name,
2430 			    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2431 			SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2432 			    "pages", CTLFLAG_RD, &dom->ud_pages, 0,
2433 			    "Total pages currently allocated from VM");
2434 			SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2435 			    "free_items", CTLFLAG_RD, &dom->ud_free_items, 0,
2436 			    "items free in the slab layer");
2437 		}
2438 	} else
2439 		SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2440 		    "name", CTLFLAG_RD, nokeg, "Keg name");
2441 
2442 	/*
2443 	 * Information about zone limits.
2444 	 */
2445 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2446 	    "limit", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2447 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2448 	    "items", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2449 	    zone, 0, sysctl_handle_uma_zone_items, "QU",
2450 	    "current number of allocated items if limit is set");
2451 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2452 	    "max_items", CTLFLAG_RD, &zone->uz_max_items, 0,
2453 	    "Maximum number of cached items");
2454 	SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2455 	    "sleepers", CTLFLAG_RD, &zone->uz_sleepers, 0,
2456 	    "Number of threads sleeping at limit");
2457 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2458 	    "sleeps", CTLFLAG_RD, &zone->uz_sleeps, 0,
2459 	    "Total zone limit sleeps");
2460 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2461 	    "bucket_max", CTLFLAG_RD, &zone->uz_bucket_max, 0,
2462 	    "Maximum number of items in each domain's bucket cache");
2463 
2464 	/*
2465 	 * Per-domain zone information.
2466 	 */
2467 	domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid),
2468 	    OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2469 	for (i = 0; i < domains; i++) {
2470 		zdom = ZDOM_GET(zone, i);
2471 		oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2472 		    OID_AUTO, VM_DOMAIN(i)->vmd_name,
2473 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2474 		SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2475 		    "nitems", CTLFLAG_RD, &zdom->uzd_nitems,
2476 		    "number of items in this domain");
2477 		SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2478 		    "imax", CTLFLAG_RD, &zdom->uzd_imax,
2479 		    "maximum item count in this period");
2480 		SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2481 		    "imin", CTLFLAG_RD, &zdom->uzd_imin,
2482 		    "minimum item count in this period");
2483 		SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2484 		    "wss", CTLFLAG_RD, &zdom->uzd_wss,
2485 		    "Working set size");
2486 	}
2487 
2488 	/*
2489 	 * General statistics.
2490 	 */
2491 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2492 	    "stats", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2493 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2494 	    "current", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2495 	    zone, 1, sysctl_handle_uma_zone_cur, "I",
2496 	    "Current number of allocated items");
2497 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2498 	    "allocs", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2499 	    zone, 0, sysctl_handle_uma_zone_allocs, "QU",
2500 	    "Total allocation calls");
2501 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2502 	    "frees", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2503 	    zone, 0, sysctl_handle_uma_zone_frees, "QU",
2504 	    "Total free calls");
2505 	SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2506 	    "fails", CTLFLAG_RD, &zone->uz_fails,
2507 	    "Number of allocation failures");
2508 	SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2509 	    "xdomain", CTLFLAG_RD, &zone->uz_xdomain,
2510 	    "Free calls from the wrong domain");
2511 }
2512 
2513 struct uma_zone_count {
2514 	const char	*name;
2515 	int		count;
2516 };
2517 
2518 static void
2519 zone_count(uma_zone_t zone, void *arg)
2520 {
2521 	struct uma_zone_count *cnt;
2522 
2523 	cnt = arg;
2524 	/*
2525 	 * Some zones are rapidly created with identical names and
2526 	 * destroyed out of order.  This can lead to gaps in the count.
2527 	 * Use one greater than the maximum observed for this name.
2528 	 */
2529 	if (strcmp(zone->uz_name, cnt->name) == 0)
2530 		cnt->count = MAX(cnt->count,
2531 		    zone->uz_namecnt + 1);
2532 }
2533 
2534 static void
2535 zone_update_caches(uma_zone_t zone)
2536 {
2537 	int i;
2538 
2539 	for (i = 0; i <= mp_maxid; i++) {
2540 		cache_set_uz_size(&zone->uz_cpu[i], zone->uz_size);
2541 		cache_set_uz_flags(&zone->uz_cpu[i], zone->uz_flags);
2542 	}
2543 }
2544 
2545 /*
2546  * Zone header ctor.  This initializes all fields, locks, etc.
2547  *
2548  * Arguments/Returns follow uma_ctor specifications
2549  *	udata  Actually uma_zctor_args
2550  */
2551 static int
2552 zone_ctor(void *mem, int size, void *udata, int flags)
2553 {
2554 	struct uma_zone_count cnt;
2555 	struct uma_zctor_args *arg = udata;
2556 	uma_zone_domain_t zdom;
2557 	uma_zone_t zone = mem;
2558 	uma_zone_t z;
2559 	uma_keg_t keg;
2560 	int i;
2561 
2562 	bzero(zone, size);
2563 	zone->uz_name = arg->name;
2564 	zone->uz_ctor = arg->ctor;
2565 	zone->uz_dtor = arg->dtor;
2566 	zone->uz_init = NULL;
2567 	zone->uz_fini = NULL;
2568 	zone->uz_sleeps = 0;
2569 	zone->uz_bucket_size = 0;
2570 	zone->uz_bucket_size_min = 0;
2571 	zone->uz_bucket_size_max = BUCKET_MAX;
2572 	zone->uz_flags = (arg->flags & UMA_ZONE_SMR);
2573 	zone->uz_warning = NULL;
2574 	/* The domain structures follow the cpu structures. */
2575 	zone->uz_bucket_max = ULONG_MAX;
2576 	timevalclear(&zone->uz_ratecheck);
2577 
2578 	/* Count the number of duplicate names. */
2579 	cnt.name = arg->name;
2580 	cnt.count = 0;
2581 	zone_foreach(zone_count, &cnt);
2582 	zone->uz_namecnt = cnt.count;
2583 	ZONE_CROSS_LOCK_INIT(zone);
2584 
2585 	for (i = 0; i < vm_ndomains; i++) {
2586 		zdom = ZDOM_GET(zone, i);
2587 		ZDOM_LOCK_INIT(zone, zdom, (arg->flags & UMA_ZONE_MTXCLASS));
2588 		STAILQ_INIT(&zdom->uzd_buckets);
2589 	}
2590 
2591 #ifdef INVARIANTS
2592 	if (arg->uminit == trash_init && arg->fini == trash_fini)
2593 		zone->uz_flags |= UMA_ZFLAG_TRASH | UMA_ZFLAG_CTORDTOR;
2594 #endif
2595 
2596 	/*
2597 	 * This is a pure cache zone, no kegs.
2598 	 */
2599 	if (arg->import) {
2600 		KASSERT((arg->flags & UMA_ZFLAG_CACHE) != 0,
2601 		    ("zone_ctor: Import specified for non-cache zone."));
2602 		zone->uz_flags = arg->flags;
2603 		zone->uz_size = arg->size;
2604 		zone->uz_import = arg->import;
2605 		zone->uz_release = arg->release;
2606 		zone->uz_arg = arg->arg;
2607 #ifdef NUMA
2608 		/*
2609 		 * Cache zones are round-robin unless a policy is
2610 		 * specified because they may have incompatible
2611 		 * constraints.
2612 		 */
2613 		if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2614 			zone->uz_flags |= UMA_ZONE_ROUNDROBIN;
2615 #endif
2616 		rw_wlock(&uma_rwlock);
2617 		LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link);
2618 		rw_wunlock(&uma_rwlock);
2619 		goto out;
2620 	}
2621 
2622 	/*
2623 	 * Use the regular zone/keg/slab allocator.
2624 	 */
2625 	zone->uz_import = zone_import;
2626 	zone->uz_release = zone_release;
2627 	zone->uz_arg = zone;
2628 	keg = arg->keg;
2629 
2630 	if (arg->flags & UMA_ZONE_SECONDARY) {
2631 		KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
2632 		    ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
2633 		KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
2634 		zone->uz_init = arg->uminit;
2635 		zone->uz_fini = arg->fini;
2636 		zone->uz_flags |= UMA_ZONE_SECONDARY;
2637 		rw_wlock(&uma_rwlock);
2638 		ZONE_LOCK(zone);
2639 		LIST_FOREACH(z, &keg->uk_zones, uz_link) {
2640 			if (LIST_NEXT(z, uz_link) == NULL) {
2641 				LIST_INSERT_AFTER(z, zone, uz_link);
2642 				break;
2643 			}
2644 		}
2645 		ZONE_UNLOCK(zone);
2646 		rw_wunlock(&uma_rwlock);
2647 	} else if (keg == NULL) {
2648 		if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
2649 		    arg->align, arg->flags)) == NULL)
2650 			return (ENOMEM);
2651 	} else {
2652 		struct uma_kctor_args karg;
2653 		int error;
2654 
2655 		/* We should only be here from uma_startup() */
2656 		karg.size = arg->size;
2657 		karg.uminit = arg->uminit;
2658 		karg.fini = arg->fini;
2659 		karg.align = arg->align;
2660 		karg.flags = (arg->flags & ~UMA_ZONE_SMR);
2661 		karg.zone = zone;
2662 		error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
2663 		    flags);
2664 		if (error)
2665 			return (error);
2666 	}
2667 
2668 	/* Inherit properties from the keg. */
2669 	zone->uz_keg = keg;
2670 	zone->uz_size = keg->uk_size;
2671 	zone->uz_flags |= (keg->uk_flags &
2672 	    (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
2673 
2674 out:
2675 	if (__predict_true(booted >= BOOT_RUNNING)) {
2676 		zone_alloc_counters(zone, NULL);
2677 		zone_alloc_sysctl(zone, NULL);
2678 	} else {
2679 		zone->uz_allocs = EARLY_COUNTER;
2680 		zone->uz_frees = EARLY_COUNTER;
2681 		zone->uz_fails = EARLY_COUNTER;
2682 	}
2683 
2684 	/* Caller requests a private SMR context. */
2685 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
2686 		zone->uz_smr = smr_create(zone->uz_name, 0, 0);
2687 
2688 	KASSERT((arg->flags & (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET)) !=
2689 	    (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET),
2690 	    ("Invalid zone flag combination"));
2691 	if (arg->flags & UMA_ZFLAG_INTERNAL)
2692 		zone->uz_bucket_size_max = zone->uz_bucket_size = 0;
2693 	if ((arg->flags & UMA_ZONE_MAXBUCKET) != 0)
2694 		zone->uz_bucket_size = BUCKET_MAX;
2695 	else if ((arg->flags & UMA_ZONE_MINBUCKET) != 0)
2696 		zone->uz_bucket_size_max = zone->uz_bucket_size = BUCKET_MIN;
2697 	else if ((arg->flags & UMA_ZONE_NOBUCKET) != 0)
2698 		zone->uz_bucket_size = 0;
2699 	else
2700 		zone->uz_bucket_size = bucket_select(zone->uz_size);
2701 	zone->uz_bucket_size_min = zone->uz_bucket_size;
2702 	if (zone->uz_dtor != NULL || zone->uz_ctor != NULL)
2703 		zone->uz_flags |= UMA_ZFLAG_CTORDTOR;
2704 	zone_update_caches(zone);
2705 
2706 	return (0);
2707 }
2708 
2709 /*
2710  * Keg header dtor.  This frees all data, destroys locks, frees the hash
2711  * table and removes the keg from the global list.
2712  *
2713  * Arguments/Returns follow uma_dtor specifications
2714  *	udata  unused
2715  */
2716 static void
2717 keg_dtor(void *arg, int size, void *udata)
2718 {
2719 	uma_keg_t keg;
2720 	uint32_t free, pages;
2721 	int i;
2722 
2723 	keg = (uma_keg_t)arg;
2724 	free = pages = 0;
2725 	for (i = 0; i < vm_ndomains; i++) {
2726 		free += keg->uk_domain[i].ud_free_items;
2727 		pages += keg->uk_domain[i].ud_pages;
2728 		KEG_LOCK_FINI(keg, i);
2729 	}
2730 	if (pages != 0)
2731 		printf("Freed UMA keg (%s) was not empty (%u items). "
2732 		    " Lost %u pages of memory.\n",
2733 		    keg->uk_name ? keg->uk_name : "",
2734 		    pages / keg->uk_ppera * keg->uk_ipers - free, pages);
2735 
2736 	hash_free(&keg->uk_hash);
2737 }
2738 
2739 /*
2740  * Zone header dtor.
2741  *
2742  * Arguments/Returns follow uma_dtor specifications
2743  *	udata  unused
2744  */
2745 static void
2746 zone_dtor(void *arg, int size, void *udata)
2747 {
2748 	uma_zone_t zone;
2749 	uma_keg_t keg;
2750 	int i;
2751 
2752 	zone = (uma_zone_t)arg;
2753 
2754 	sysctl_remove_oid(zone->uz_oid, 1, 1);
2755 
2756 	if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
2757 		cache_drain(zone);
2758 
2759 	rw_wlock(&uma_rwlock);
2760 	LIST_REMOVE(zone, uz_link);
2761 	rw_wunlock(&uma_rwlock);
2762 	zone_reclaim(zone, M_WAITOK, true);
2763 
2764 	/*
2765 	 * We only destroy kegs from non secondary/non cache zones.
2766 	 */
2767 	if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) {
2768 		keg = zone->uz_keg;
2769 		rw_wlock(&uma_rwlock);
2770 		LIST_REMOVE(keg, uk_link);
2771 		rw_wunlock(&uma_rwlock);
2772 		zone_free_item(kegs, keg, NULL, SKIP_NONE);
2773 	}
2774 	counter_u64_free(zone->uz_allocs);
2775 	counter_u64_free(zone->uz_frees);
2776 	counter_u64_free(zone->uz_fails);
2777 	counter_u64_free(zone->uz_xdomain);
2778 	free(zone->uz_ctlname, M_UMA);
2779 	for (i = 0; i < vm_ndomains; i++)
2780 		ZDOM_LOCK_FINI(ZDOM_GET(zone, i));
2781 	ZONE_CROSS_LOCK_FINI(zone);
2782 }
2783 
2784 static void
2785 zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *arg), void *arg)
2786 {
2787 	uma_keg_t keg;
2788 	uma_zone_t zone;
2789 
2790 	LIST_FOREACH(keg, &uma_kegs, uk_link) {
2791 		LIST_FOREACH(zone, &keg->uk_zones, uz_link)
2792 			zfunc(zone, arg);
2793 	}
2794 	LIST_FOREACH(zone, &uma_cachezones, uz_link)
2795 		zfunc(zone, arg);
2796 }
2797 
2798 /*
2799  * Traverses every zone in the system and calls a callback
2800  *
2801  * Arguments:
2802  *	zfunc  A pointer to a function which accepts a zone
2803  *		as an argument.
2804  *
2805  * Returns:
2806  *	Nothing
2807  */
2808 static void
2809 zone_foreach(void (*zfunc)(uma_zone_t, void *arg), void *arg)
2810 {
2811 
2812 	rw_rlock(&uma_rwlock);
2813 	zone_foreach_unlocked(zfunc, arg);
2814 	rw_runlock(&uma_rwlock);
2815 }
2816 
2817 /*
2818  * Initialize the kernel memory allocator.  This is done after pages can be
2819  * allocated but before general KVA is available.
2820  */
2821 void
2822 uma_startup1(vm_offset_t virtual_avail)
2823 {
2824 	struct uma_zctor_args args;
2825 	size_t ksize, zsize, size;
2826 	uma_keg_t masterkeg;
2827 	uintptr_t m;
2828 	uint8_t pflag;
2829 
2830 	bootstart = bootmem = virtual_avail;
2831 
2832 	rw_init(&uma_rwlock, "UMA lock");
2833 	sx_init(&uma_reclaim_lock, "umareclaim");
2834 
2835 	ksize = sizeof(struct uma_keg) +
2836 	    (sizeof(struct uma_domain) * vm_ndomains);
2837 	ksize = roundup(ksize, UMA_SUPER_ALIGN);
2838 	zsize = sizeof(struct uma_zone) +
2839 	    (sizeof(struct uma_cache) * (mp_maxid + 1)) +
2840 	    (sizeof(struct uma_zone_domain) * vm_ndomains);
2841 	zsize = roundup(zsize, UMA_SUPER_ALIGN);
2842 
2843 	/* Allocate the zone of zones, zone of kegs, and zone of zones keg. */
2844 	size = (zsize * 2) + ksize;
2845 	m = (uintptr_t)startup_alloc(NULL, size, 0, &pflag, M_NOWAIT | M_ZERO);
2846 	zones = (uma_zone_t)m;
2847 	m += zsize;
2848 	kegs = (uma_zone_t)m;
2849 	m += zsize;
2850 	masterkeg = (uma_keg_t)m;
2851 
2852 	/* "manually" create the initial zone */
2853 	memset(&args, 0, sizeof(args));
2854 	args.name = "UMA Kegs";
2855 	args.size = ksize;
2856 	args.ctor = keg_ctor;
2857 	args.dtor = keg_dtor;
2858 	args.uminit = zero_init;
2859 	args.fini = NULL;
2860 	args.keg = masterkeg;
2861 	args.align = UMA_SUPER_ALIGN - 1;
2862 	args.flags = UMA_ZFLAG_INTERNAL;
2863 	zone_ctor(kegs, zsize, &args, M_WAITOK);
2864 
2865 	args.name = "UMA Zones";
2866 	args.size = zsize;
2867 	args.ctor = zone_ctor;
2868 	args.dtor = zone_dtor;
2869 	args.uminit = zero_init;
2870 	args.fini = NULL;
2871 	args.keg = NULL;
2872 	args.align = UMA_SUPER_ALIGN - 1;
2873 	args.flags = UMA_ZFLAG_INTERNAL;
2874 	zone_ctor(zones, zsize, &args, M_WAITOK);
2875 
2876 	/* Now make zones for slab headers */
2877 	slabzones[0] = uma_zcreate("UMA Slabs 0", SLABZONE0_SIZE,
2878 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2879 	slabzones[1] = uma_zcreate("UMA Slabs 1", SLABZONE1_SIZE,
2880 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2881 
2882 	hashzone = uma_zcreate("UMA Hash",
2883 	    sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
2884 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2885 
2886 	bucket_init();
2887 	smr_init();
2888 }
2889 
2890 #ifndef UMA_MD_SMALL_ALLOC
2891 extern void vm_radix_reserve_kva(void);
2892 #endif
2893 
2894 /*
2895  * Advertise the availability of normal kva allocations and switch to
2896  * the default back-end allocator.  Marks the KVA we consumed on startup
2897  * as used in the map.
2898  */
2899 void
2900 uma_startup2(void)
2901 {
2902 
2903 	if (bootstart != bootmem) {
2904 		vm_map_lock(kernel_map);
2905 		(void)vm_map_insert(kernel_map, NULL, 0, bootstart, bootmem,
2906 		    VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT);
2907 		vm_map_unlock(kernel_map);
2908 	}
2909 
2910 #ifndef UMA_MD_SMALL_ALLOC
2911 	/* Set up radix zone to use noobj_alloc. */
2912 	vm_radix_reserve_kva();
2913 #endif
2914 
2915 	booted = BOOT_KVA;
2916 	zone_foreach_unlocked(zone_kva_available, NULL);
2917 	bucket_enable();
2918 }
2919 
2920 /*
2921  * Finish our initialization steps.
2922  */
2923 static void
2924 uma_startup3(void)
2925 {
2926 
2927 #ifdef INVARIANTS
2928 	TUNABLE_INT_FETCH("vm.debug.divisor", &dbg_divisor);
2929 	uma_dbg_cnt = counter_u64_alloc(M_WAITOK);
2930 	uma_skip_cnt = counter_u64_alloc(M_WAITOK);
2931 #endif
2932 	zone_foreach_unlocked(zone_alloc_counters, NULL);
2933 	zone_foreach_unlocked(zone_alloc_sysctl, NULL);
2934 	callout_init(&uma_callout, 1);
2935 	callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
2936 	booted = BOOT_RUNNING;
2937 
2938 	EVENTHANDLER_REGISTER(shutdown_post_sync, uma_shutdown, NULL,
2939 	    EVENTHANDLER_PRI_FIRST);
2940 }
2941 
2942 static void
2943 uma_shutdown(void)
2944 {
2945 
2946 	booted = BOOT_SHUTDOWN;
2947 }
2948 
2949 static uma_keg_t
2950 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
2951 		int align, uint32_t flags)
2952 {
2953 	struct uma_kctor_args args;
2954 
2955 	args.size = size;
2956 	args.uminit = uminit;
2957 	args.fini = fini;
2958 	args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
2959 	args.flags = flags;
2960 	args.zone = zone;
2961 	return (zone_alloc_item(kegs, &args, UMA_ANYDOMAIN, M_WAITOK));
2962 }
2963 
2964 /* Public functions */
2965 /* See uma.h */
2966 void
2967 uma_set_align(int align)
2968 {
2969 
2970 	if (align != UMA_ALIGN_CACHE)
2971 		uma_align_cache = align;
2972 }
2973 
2974 /* See uma.h */
2975 uma_zone_t
2976 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
2977 		uma_init uminit, uma_fini fini, int align, uint32_t flags)
2978 
2979 {
2980 	struct uma_zctor_args args;
2981 	uma_zone_t res;
2982 
2983 	KASSERT(powerof2(align + 1), ("invalid zone alignment %d for \"%s\"",
2984 	    align, name));
2985 
2986 	/* This stuff is essential for the zone ctor */
2987 	memset(&args, 0, sizeof(args));
2988 	args.name = name;
2989 	args.size = size;
2990 	args.ctor = ctor;
2991 	args.dtor = dtor;
2992 	args.uminit = uminit;
2993 	args.fini = fini;
2994 #ifdef  INVARIANTS
2995 	/*
2996 	 * Inject procedures which check for memory use after free if we are
2997 	 * allowed to scramble the memory while it is not allocated.  This
2998 	 * requires that: UMA is actually able to access the memory, no init
2999 	 * or fini procedures, no dependency on the initial value of the
3000 	 * memory, and no (legitimate) use of the memory after free.  Note,
3001 	 * the ctor and dtor do not need to be empty.
3002 	 */
3003 	if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOTOUCH |
3004 	    UMA_ZONE_NOFREE))) && uminit == NULL && fini == NULL) {
3005 		args.uminit = trash_init;
3006 		args.fini = trash_fini;
3007 	}
3008 #endif
3009 	args.align = align;
3010 	args.flags = flags;
3011 	args.keg = NULL;
3012 
3013 	sx_slock(&uma_reclaim_lock);
3014 	res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3015 	sx_sunlock(&uma_reclaim_lock);
3016 
3017 	return (res);
3018 }
3019 
3020 /* See uma.h */
3021 uma_zone_t
3022 uma_zsecond_create(const char *name, uma_ctor ctor, uma_dtor dtor,
3023     uma_init zinit, uma_fini zfini, uma_zone_t master)
3024 {
3025 	struct uma_zctor_args args;
3026 	uma_keg_t keg;
3027 	uma_zone_t res;
3028 
3029 	keg = master->uz_keg;
3030 	memset(&args, 0, sizeof(args));
3031 	args.name = name;
3032 	args.size = keg->uk_size;
3033 	args.ctor = ctor;
3034 	args.dtor = dtor;
3035 	args.uminit = zinit;
3036 	args.fini = zfini;
3037 	args.align = keg->uk_align;
3038 	args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
3039 	args.keg = keg;
3040 
3041 	sx_slock(&uma_reclaim_lock);
3042 	res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3043 	sx_sunlock(&uma_reclaim_lock);
3044 
3045 	return (res);
3046 }
3047 
3048 /* See uma.h */
3049 uma_zone_t
3050 uma_zcache_create(const char *name, int size, uma_ctor ctor, uma_dtor dtor,
3051     uma_init zinit, uma_fini zfini, uma_import zimport, uma_release zrelease,
3052     void *arg, int flags)
3053 {
3054 	struct uma_zctor_args args;
3055 
3056 	memset(&args, 0, sizeof(args));
3057 	args.name = name;
3058 	args.size = size;
3059 	args.ctor = ctor;
3060 	args.dtor = dtor;
3061 	args.uminit = zinit;
3062 	args.fini = zfini;
3063 	args.import = zimport;
3064 	args.release = zrelease;
3065 	args.arg = arg;
3066 	args.align = 0;
3067 	args.flags = flags | UMA_ZFLAG_CACHE;
3068 
3069 	return (zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK));
3070 }
3071 
3072 /* See uma.h */
3073 void
3074 uma_zdestroy(uma_zone_t zone)
3075 {
3076 
3077 	/*
3078 	 * Large slabs are expensive to reclaim, so don't bother doing
3079 	 * unnecessary work if we're shutting down.
3080 	 */
3081 	if (booted == BOOT_SHUTDOWN &&
3082 	    zone->uz_fini == NULL && zone->uz_release == zone_release)
3083 		return;
3084 	sx_slock(&uma_reclaim_lock);
3085 	zone_free_item(zones, zone, NULL, SKIP_NONE);
3086 	sx_sunlock(&uma_reclaim_lock);
3087 }
3088 
3089 void
3090 uma_zwait(uma_zone_t zone)
3091 {
3092 
3093 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
3094 		uma_zfree_smr(zone, uma_zalloc_smr(zone, M_WAITOK));
3095 	else if ((zone->uz_flags & UMA_ZONE_PCPU) != 0)
3096 		uma_zfree_pcpu(zone, uma_zalloc_pcpu(zone, M_WAITOK));
3097 	else
3098 		uma_zfree(zone, uma_zalloc(zone, M_WAITOK));
3099 }
3100 
3101 void *
3102 uma_zalloc_pcpu_arg(uma_zone_t zone, void *udata, int flags)
3103 {
3104 	void *item, *pcpu_item;
3105 #ifdef SMP
3106 	int i;
3107 
3108 	MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3109 #endif
3110 	item = uma_zalloc_arg(zone, udata, flags & ~M_ZERO);
3111 	if (item == NULL)
3112 		return (NULL);
3113 	pcpu_item = zpcpu_base_to_offset(item);
3114 	if (flags & M_ZERO) {
3115 #ifdef SMP
3116 		for (i = 0; i <= mp_maxid; i++)
3117 			bzero(zpcpu_get_cpu(pcpu_item, i), zone->uz_size);
3118 #else
3119 		bzero(item, zone->uz_size);
3120 #endif
3121 	}
3122 	return (pcpu_item);
3123 }
3124 
3125 /*
3126  * A stub while both regular and pcpu cases are identical.
3127  */
3128 void
3129 uma_zfree_pcpu_arg(uma_zone_t zone, void *pcpu_item, void *udata)
3130 {
3131 	void *item;
3132 
3133 #ifdef SMP
3134 	MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3135 #endif
3136 	item = zpcpu_offset_to_base(pcpu_item);
3137 	uma_zfree_arg(zone, item, udata);
3138 }
3139 
3140 static inline void *
3141 item_ctor(uma_zone_t zone, int uz_flags, int size, void *udata, int flags,
3142     void *item)
3143 {
3144 #ifdef INVARIANTS
3145 	bool skipdbg;
3146 
3147 	skipdbg = uma_dbg_zskip(zone, item);
3148 	if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3149 	    zone->uz_ctor != trash_ctor)
3150 		trash_ctor(item, size, udata, flags);
3151 #endif
3152 	/* Check flags before loading ctor pointer. */
3153 	if (__predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0) &&
3154 	    __predict_false(zone->uz_ctor != NULL) &&
3155 	    zone->uz_ctor(item, size, udata, flags) != 0) {
3156 		counter_u64_add(zone->uz_fails, 1);
3157 		zone_free_item(zone, item, udata, SKIP_DTOR | SKIP_CNT);
3158 		return (NULL);
3159 	}
3160 #ifdef INVARIANTS
3161 	if (!skipdbg)
3162 		uma_dbg_alloc(zone, NULL, item);
3163 #endif
3164 	if (__predict_false(flags & M_ZERO))
3165 		return (memset(item, 0, size));
3166 
3167 	return (item);
3168 }
3169 
3170 static inline void
3171 item_dtor(uma_zone_t zone, void *item, int size, void *udata,
3172     enum zfreeskip skip)
3173 {
3174 #ifdef INVARIANTS
3175 	bool skipdbg;
3176 
3177 	skipdbg = uma_dbg_zskip(zone, item);
3178 	if (skip == SKIP_NONE && !skipdbg) {
3179 		if ((zone->uz_flags & UMA_ZONE_MALLOC) != 0)
3180 			uma_dbg_free(zone, udata, item);
3181 		else
3182 			uma_dbg_free(zone, NULL, item);
3183 	}
3184 #endif
3185 	if (__predict_true(skip < SKIP_DTOR)) {
3186 		if (zone->uz_dtor != NULL)
3187 			zone->uz_dtor(item, size, udata);
3188 #ifdef INVARIANTS
3189 		if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3190 		    zone->uz_dtor != trash_dtor)
3191 			trash_dtor(item, size, udata);
3192 #endif
3193 	}
3194 }
3195 
3196 #if defined(INVARIANTS) || defined(DEBUG_MEMGUARD) || defined(WITNESS)
3197 #define	UMA_ZALLOC_DEBUG
3198 static int
3199 uma_zalloc_debug(uma_zone_t zone, void **itemp, void *udata, int flags)
3200 {
3201 	int error;
3202 
3203 	error = 0;
3204 #ifdef WITNESS
3205 	if (flags & M_WAITOK) {
3206 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
3207 		    "uma_zalloc_debug: zone \"%s\"", zone->uz_name);
3208 	}
3209 #endif
3210 
3211 #ifdef INVARIANTS
3212 	KASSERT((flags & M_EXEC) == 0,
3213 	    ("uma_zalloc_debug: called with M_EXEC"));
3214 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3215 	    ("uma_zalloc_debug: called within spinlock or critical section"));
3216 	KASSERT((zone->uz_flags & UMA_ZONE_PCPU) == 0 || (flags & M_ZERO) == 0,
3217 	    ("uma_zalloc_debug: allocating from a pcpu zone with M_ZERO"));
3218 #endif
3219 
3220 #ifdef DEBUG_MEMGUARD
3221 	if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && memguard_cmp_zone(zone)) {
3222 		void *item;
3223 		item = memguard_alloc(zone->uz_size, flags);
3224 		if (item != NULL) {
3225 			error = EJUSTRETURN;
3226 			if (zone->uz_init != NULL &&
3227 			    zone->uz_init(item, zone->uz_size, flags) != 0) {
3228 				*itemp = NULL;
3229 				return (error);
3230 			}
3231 			if (zone->uz_ctor != NULL &&
3232 			    zone->uz_ctor(item, zone->uz_size, udata,
3233 			    flags) != 0) {
3234 				counter_u64_add(zone->uz_fails, 1);
3235 			    	zone->uz_fini(item, zone->uz_size);
3236 				*itemp = NULL;
3237 				return (error);
3238 			}
3239 			*itemp = item;
3240 			return (error);
3241 		}
3242 		/* This is unfortunate but should not be fatal. */
3243 	}
3244 #endif
3245 	return (error);
3246 }
3247 
3248 static int
3249 uma_zfree_debug(uma_zone_t zone, void *item, void *udata)
3250 {
3251 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3252 	    ("uma_zfree_debug: called with spinlock or critical section held"));
3253 
3254 #ifdef DEBUG_MEMGUARD
3255 	if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && is_memguard_addr(item)) {
3256 		if (zone->uz_dtor != NULL)
3257 			zone->uz_dtor(item, zone->uz_size, udata);
3258 		if (zone->uz_fini != NULL)
3259 			zone->uz_fini(item, zone->uz_size);
3260 		memguard_free(item);
3261 		return (EJUSTRETURN);
3262 	}
3263 #endif
3264 	return (0);
3265 }
3266 #endif
3267 
3268 static inline void *
3269 cache_alloc_item(uma_zone_t zone, uma_cache_t cache, uma_cache_bucket_t bucket,
3270     void *udata, int flags)
3271 {
3272 	void *item;
3273 	int size, uz_flags;
3274 
3275 	item = cache_bucket_pop(cache, bucket);
3276 	size = cache_uz_size(cache);
3277 	uz_flags = cache_uz_flags(cache);
3278 	critical_exit();
3279 	return (item_ctor(zone, uz_flags, size, udata, flags, item));
3280 }
3281 
3282 static __noinline void *
3283 cache_alloc_retry(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3284 {
3285 	uma_cache_bucket_t bucket;
3286 	int domain;
3287 
3288 	while (cache_alloc(zone, cache, udata, flags)) {
3289 		cache = &zone->uz_cpu[curcpu];
3290 		bucket = &cache->uc_allocbucket;
3291 		if (__predict_false(bucket->ucb_cnt == 0))
3292 			continue;
3293 		return (cache_alloc_item(zone, cache, bucket, udata, flags));
3294 	}
3295 	critical_exit();
3296 
3297 	/*
3298 	 * We can not get a bucket so try to return a single item.
3299 	 */
3300 	if (zone->uz_flags & UMA_ZONE_FIRSTTOUCH)
3301 		domain = PCPU_GET(domain);
3302 	else
3303 		domain = UMA_ANYDOMAIN;
3304 	return (zone_alloc_item(zone, udata, domain, flags));
3305 }
3306 
3307 /* See uma.h */
3308 void *
3309 uma_zalloc_smr(uma_zone_t zone, int flags)
3310 {
3311 	uma_cache_bucket_t bucket;
3312 	uma_cache_t cache;
3313 
3314 #ifdef UMA_ZALLOC_DEBUG
3315 	void *item;
3316 
3317 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
3318 	    ("uma_zalloc_arg: called with non-SMR zone.\n"));
3319 	if (uma_zalloc_debug(zone, &item, NULL, flags) == EJUSTRETURN)
3320 		return (item);
3321 #endif
3322 
3323 	critical_enter();
3324 	cache = &zone->uz_cpu[curcpu];
3325 	bucket = &cache->uc_allocbucket;
3326 	if (__predict_false(bucket->ucb_cnt == 0))
3327 		return (cache_alloc_retry(zone, cache, NULL, flags));
3328 	return (cache_alloc_item(zone, cache, bucket, NULL, flags));
3329 }
3330 
3331 /* See uma.h */
3332 void *
3333 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
3334 {
3335 	uma_cache_bucket_t bucket;
3336 	uma_cache_t cache;
3337 
3338 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3339 	random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3340 
3341 	/* This is the fast path allocation */
3342 	CTR3(KTR_UMA, "uma_zalloc_arg zone %s(%p) flags %d", zone->uz_name,
3343 	    zone, flags);
3344 
3345 #ifdef UMA_ZALLOC_DEBUG
3346 	void *item;
3347 
3348 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3349 	    ("uma_zalloc_arg: called with SMR zone.\n"));
3350 	if (uma_zalloc_debug(zone, &item, udata, flags) == EJUSTRETURN)
3351 		return (item);
3352 #endif
3353 
3354 	/*
3355 	 * If possible, allocate from the per-CPU cache.  There are two
3356 	 * requirements for safe access to the per-CPU cache: (1) the thread
3357 	 * accessing the cache must not be preempted or yield during access,
3358 	 * and (2) the thread must not migrate CPUs without switching which
3359 	 * cache it accesses.  We rely on a critical section to prevent
3360 	 * preemption and migration.  We release the critical section in
3361 	 * order to acquire the zone mutex if we are unable to allocate from
3362 	 * the current cache; when we re-acquire the critical section, we
3363 	 * must detect and handle migration if it has occurred.
3364 	 */
3365 	critical_enter();
3366 	cache = &zone->uz_cpu[curcpu];
3367 	bucket = &cache->uc_allocbucket;
3368 	if (__predict_false(bucket->ucb_cnt == 0))
3369 		return (cache_alloc_retry(zone, cache, udata, flags));
3370 	return (cache_alloc_item(zone, cache, bucket, udata, flags));
3371 }
3372 
3373 /*
3374  * Replenish an alloc bucket and possibly restore an old one.  Called in
3375  * a critical section.  Returns in a critical section.
3376  *
3377  * A false return value indicates an allocation failure.
3378  * A true return value indicates success and the caller should retry.
3379  */
3380 static __noinline bool
3381 cache_alloc(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3382 {
3383 	uma_bucket_t bucket;
3384 	int domain;
3385 	bool new;
3386 
3387 	CRITICAL_ASSERT(curthread);
3388 
3389 	/*
3390 	 * If we have run out of items in our alloc bucket see
3391 	 * if we can switch with the free bucket.
3392 	 *
3393 	 * SMR Zones can't re-use the free bucket until the sequence has
3394 	 * expired.
3395 	 */
3396 	if ((cache_uz_flags(cache) & UMA_ZONE_SMR) == 0 &&
3397 	    cache->uc_freebucket.ucb_cnt != 0) {
3398 		cache_bucket_swap(&cache->uc_freebucket,
3399 		    &cache->uc_allocbucket);
3400 		return (true);
3401 	}
3402 
3403 	/*
3404 	 * Discard any empty allocation bucket while we hold no locks.
3405 	 */
3406 	bucket = cache_bucket_unload_alloc(cache);
3407 	critical_exit();
3408 
3409 	if (bucket != NULL) {
3410 		KASSERT(bucket->ub_cnt == 0,
3411 		    ("cache_alloc: Entered with non-empty alloc bucket."));
3412 		bucket_free(zone, bucket, udata);
3413 	}
3414 
3415 	/* Short-circuit for zones without buckets and low memory. */
3416 	if (zone->uz_bucket_size == 0 || bucketdisable) {
3417 		critical_enter();
3418 		return (false);
3419 	}
3420 
3421 	/*
3422 	 * Attempt to retrieve the item from the per-CPU cache has failed, so
3423 	 * we must go back to the zone.  This requires the zdom lock, so we
3424 	 * must drop the critical section, then re-acquire it when we go back
3425 	 * to the cache.  Since the critical section is released, we may be
3426 	 * preempted or migrate.  As such, make sure not to maintain any
3427 	 * thread-local state specific to the cache from prior to releasing
3428 	 * the critical section.
3429 	 */
3430 	domain = PCPU_GET(domain);
3431 	if ((cache_uz_flags(cache) & UMA_ZONE_ROUNDROBIN) != 0)
3432 		domain = zone_domain_highest(zone, domain);
3433 	bucket = cache_fetch_bucket(zone, cache, domain);
3434 	if (bucket == NULL) {
3435 		bucket = zone_alloc_bucket(zone, udata, domain, flags);
3436 		new = true;
3437 	} else
3438 		new = false;
3439 
3440 	CTR3(KTR_UMA, "uma_zalloc: zone %s(%p) bucket zone returned %p",
3441 	    zone->uz_name, zone, bucket);
3442 	if (bucket == NULL) {
3443 		critical_enter();
3444 		return (false);
3445 	}
3446 
3447 	/*
3448 	 * See if we lost the race or were migrated.  Cache the
3449 	 * initialized bucket to make this less likely or claim
3450 	 * the memory directly.
3451 	 */
3452 	critical_enter();
3453 	cache = &zone->uz_cpu[curcpu];
3454 	if (cache->uc_allocbucket.ucb_bucket == NULL &&
3455 	    ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) == 0 ||
3456 	    domain == PCPU_GET(domain))) {
3457 		if (new)
3458 			atomic_add_long(&ZDOM_GET(zone, domain)->uzd_imax,
3459 			    bucket->ub_cnt);
3460 		cache_bucket_load_alloc(cache, bucket);
3461 		return (true);
3462 	}
3463 
3464 	/*
3465 	 * We lost the race, release this bucket and start over.
3466 	 */
3467 	critical_exit();
3468 	zone_put_bucket(zone, domain, bucket, udata, false);
3469 	critical_enter();
3470 
3471 	return (true);
3472 }
3473 
3474 void *
3475 uma_zalloc_domain(uma_zone_t zone, void *udata, int domain, int flags)
3476 {
3477 
3478 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3479 	random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3480 
3481 	/* This is the fast path allocation */
3482 	CTR4(KTR_UMA, "uma_zalloc_domain zone %s(%p) domain %d flags %d",
3483 	    zone->uz_name, zone, domain, flags);
3484 
3485 	if (flags & M_WAITOK) {
3486 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
3487 		    "uma_zalloc_domain: zone \"%s\"", zone->uz_name);
3488 	}
3489 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3490 	    ("uma_zalloc_domain: called with spinlock or critical section held"));
3491 
3492 	return (zone_alloc_item(zone, udata, domain, flags));
3493 }
3494 
3495 /*
3496  * Find a slab with some space.  Prefer slabs that are partially used over those
3497  * that are totally full.  This helps to reduce fragmentation.
3498  *
3499  * If 'rr' is 1, search all domains starting from 'domain'.  Otherwise check
3500  * only 'domain'.
3501  */
3502 static uma_slab_t
3503 keg_first_slab(uma_keg_t keg, int domain, bool rr)
3504 {
3505 	uma_domain_t dom;
3506 	uma_slab_t slab;
3507 	int start;
3508 
3509 	KASSERT(domain >= 0 && domain < vm_ndomains,
3510 	    ("keg_first_slab: domain %d out of range", domain));
3511 	KEG_LOCK_ASSERT(keg, domain);
3512 
3513 	slab = NULL;
3514 	start = domain;
3515 	do {
3516 		dom = &keg->uk_domain[domain];
3517 		if ((slab = LIST_FIRST(&dom->ud_part_slab)) != NULL)
3518 			return (slab);
3519 		if ((slab = LIST_FIRST(&dom->ud_free_slab)) != NULL) {
3520 			LIST_REMOVE(slab, us_link);
3521 			dom->ud_free_slabs--;
3522 			LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
3523 			return (slab);
3524 		}
3525 		if (rr)
3526 			domain = (domain + 1) % vm_ndomains;
3527 	} while (domain != start);
3528 
3529 	return (NULL);
3530 }
3531 
3532 /*
3533  * Fetch an existing slab from a free or partial list.  Returns with the
3534  * keg domain lock held if a slab was found or unlocked if not.
3535  */
3536 static uma_slab_t
3537 keg_fetch_free_slab(uma_keg_t keg, int domain, bool rr, int flags)
3538 {
3539 	uma_slab_t slab;
3540 	uint32_t reserve;
3541 
3542 	/* HASH has a single free list. */
3543 	if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
3544 		domain = 0;
3545 
3546 	KEG_LOCK(keg, domain);
3547 	reserve = (flags & M_USE_RESERVE) != 0 ? 0 : keg->uk_reserve;
3548 	if (keg->uk_domain[domain].ud_free_items <= reserve ||
3549 	    (slab = keg_first_slab(keg, domain, rr)) == NULL) {
3550 		KEG_UNLOCK(keg, domain);
3551 		return (NULL);
3552 	}
3553 	return (slab);
3554 }
3555 
3556 static uma_slab_t
3557 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int rdomain, const int flags)
3558 {
3559 	struct vm_domainset_iter di;
3560 	uma_slab_t slab;
3561 	int aflags, domain;
3562 	bool rr;
3563 
3564 restart:
3565 	/*
3566 	 * Use the keg's policy if upper layers haven't already specified a
3567 	 * domain (as happens with first-touch zones).
3568 	 *
3569 	 * To avoid races we run the iterator with the keg lock held, but that
3570 	 * means that we cannot allow the vm_domainset layer to sleep.  Thus,
3571 	 * clear M_WAITOK and handle low memory conditions locally.
3572 	 */
3573 	rr = rdomain == UMA_ANYDOMAIN;
3574 	if (rr) {
3575 		aflags = (flags & ~M_WAITOK) | M_NOWAIT;
3576 		vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
3577 		    &aflags);
3578 	} else {
3579 		aflags = flags;
3580 		domain = rdomain;
3581 	}
3582 
3583 	for (;;) {
3584 		slab = keg_fetch_free_slab(keg, domain, rr, flags);
3585 		if (slab != NULL)
3586 			return (slab);
3587 
3588 		/*
3589 		 * M_NOVM means don't ask at all!
3590 		 */
3591 		if (flags & M_NOVM)
3592 			break;
3593 
3594 		slab = keg_alloc_slab(keg, zone, domain, flags, aflags);
3595 		if (slab != NULL)
3596 			return (slab);
3597 		if (!rr && (flags & M_WAITOK) == 0)
3598 			break;
3599 		if (rr && vm_domainset_iter_policy(&di, &domain) != 0) {
3600 			if ((flags & M_WAITOK) != 0) {
3601 				vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask);
3602 				goto restart;
3603 			}
3604 			break;
3605 		}
3606 	}
3607 
3608 	/*
3609 	 * We might not have been able to get a slab but another cpu
3610 	 * could have while we were unlocked.  Check again before we
3611 	 * fail.
3612 	 */
3613 	if ((slab = keg_fetch_free_slab(keg, domain, rr, flags)) != NULL)
3614 		return (slab);
3615 
3616 	return (NULL);
3617 }
3618 
3619 static void *
3620 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
3621 {
3622 	uma_domain_t dom;
3623 	void *item;
3624 	int freei;
3625 
3626 	KEG_LOCK_ASSERT(keg, slab->us_domain);
3627 
3628 	dom = &keg->uk_domain[slab->us_domain];
3629 	freei = BIT_FFS(keg->uk_ipers, &slab->us_free) - 1;
3630 	BIT_CLR(keg->uk_ipers, freei, &slab->us_free);
3631 	item = slab_item(slab, keg, freei);
3632 	slab->us_freecount--;
3633 	dom->ud_free_items--;
3634 
3635 	/*
3636 	 * Move this slab to the full list.  It must be on the partial list, so
3637 	 * we do not need to update the free slab count.  In particular,
3638 	 * keg_fetch_slab() always returns slabs on the partial list.
3639 	 */
3640 	if (slab->us_freecount == 0) {
3641 		LIST_REMOVE(slab, us_link);
3642 		LIST_INSERT_HEAD(&dom->ud_full_slab, slab, us_link);
3643 	}
3644 
3645 	return (item);
3646 }
3647 
3648 static int
3649 zone_import(void *arg, void **bucket, int max, int domain, int flags)
3650 {
3651 	uma_domain_t dom;
3652 	uma_zone_t zone;
3653 	uma_slab_t slab;
3654 	uma_keg_t keg;
3655 #ifdef NUMA
3656 	int stripe;
3657 #endif
3658 	int i;
3659 
3660 	zone = arg;
3661 	slab = NULL;
3662 	keg = zone->uz_keg;
3663 	/* Try to keep the buckets totally full */
3664 	for (i = 0; i < max; ) {
3665 		if ((slab = keg_fetch_slab(keg, zone, domain, flags)) == NULL)
3666 			break;
3667 #ifdef NUMA
3668 		stripe = howmany(max, vm_ndomains);
3669 #endif
3670 		dom = &keg->uk_domain[slab->us_domain];
3671 		while (slab->us_freecount && i < max) {
3672 			bucket[i++] = slab_alloc_item(keg, slab);
3673 			if (dom->ud_free_items <= keg->uk_reserve)
3674 				break;
3675 #ifdef NUMA
3676 			/*
3677 			 * If the zone is striped we pick a new slab for every
3678 			 * N allocations.  Eliminating this conditional will
3679 			 * instead pick a new domain for each bucket rather
3680 			 * than stripe within each bucket.  The current option
3681 			 * produces more fragmentation and requires more cpu
3682 			 * time but yields better distribution.
3683 			 */
3684 			if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0 &&
3685 			    vm_ndomains > 1 && --stripe == 0)
3686 				break;
3687 #endif
3688 		}
3689 		KEG_UNLOCK(keg, slab->us_domain);
3690 		/* Don't block if we allocated any successfully. */
3691 		flags &= ~M_WAITOK;
3692 		flags |= M_NOWAIT;
3693 	}
3694 
3695 	return i;
3696 }
3697 
3698 static int
3699 zone_alloc_limit_hard(uma_zone_t zone, int count, int flags)
3700 {
3701 	uint64_t old, new, total, max;
3702 
3703 	/*
3704 	 * The hard case.  We're going to sleep because there were existing
3705 	 * sleepers or because we ran out of items.  This routine enforces
3706 	 * fairness by keeping fifo order.
3707 	 *
3708 	 * First release our ill gotten gains and make some noise.
3709 	 */
3710 	for (;;) {
3711 		zone_free_limit(zone, count);
3712 		zone_log_warning(zone);
3713 		zone_maxaction(zone);
3714 		if (flags & M_NOWAIT)
3715 			return (0);
3716 
3717 		/*
3718 		 * We need to allocate an item or set ourself as a sleeper
3719 		 * while the sleepq lock is held to avoid wakeup races.  This
3720 		 * is essentially a home rolled semaphore.
3721 		 */
3722 		sleepq_lock(&zone->uz_max_items);
3723 		old = zone->uz_items;
3724 		do {
3725 			MPASS(UZ_ITEMS_SLEEPERS(old) < UZ_ITEMS_SLEEPERS_MAX);
3726 			/* Cache the max since we will evaluate twice. */
3727 			max = zone->uz_max_items;
3728 			if (UZ_ITEMS_SLEEPERS(old) != 0 ||
3729 			    UZ_ITEMS_COUNT(old) >= max)
3730 				new = old + UZ_ITEMS_SLEEPER;
3731 			else
3732 				new = old + MIN(count, max - old);
3733 		} while (atomic_fcmpset_64(&zone->uz_items, &old, new) == 0);
3734 
3735 		/* We may have successfully allocated under the sleepq lock. */
3736 		if (UZ_ITEMS_SLEEPERS(new) == 0) {
3737 			sleepq_release(&zone->uz_max_items);
3738 			return (new - old);
3739 		}
3740 
3741 		/*
3742 		 * This is in a different cacheline from uz_items so that we
3743 		 * don't constantly invalidate the fastpath cacheline when we
3744 		 * adjust item counts.  This could be limited to toggling on
3745 		 * transitions.
3746 		 */
3747 		atomic_add_32(&zone->uz_sleepers, 1);
3748 		atomic_add_64(&zone->uz_sleeps, 1);
3749 
3750 		/*
3751 		 * We have added ourselves as a sleeper.  The sleepq lock
3752 		 * protects us from wakeup races.  Sleep now and then retry.
3753 		 */
3754 		sleepq_add(&zone->uz_max_items, NULL, "zonelimit", 0, 0);
3755 		sleepq_wait(&zone->uz_max_items, PVM);
3756 
3757 		/*
3758 		 * After wakeup, remove ourselves as a sleeper and try
3759 		 * again.  We no longer have the sleepq lock for protection.
3760 		 *
3761 		 * Subract ourselves as a sleeper while attempting to add
3762 		 * our count.
3763 		 */
3764 		atomic_subtract_32(&zone->uz_sleepers, 1);
3765 		old = atomic_fetchadd_64(&zone->uz_items,
3766 		    -(UZ_ITEMS_SLEEPER - count));
3767 		/* We're no longer a sleeper. */
3768 		old -= UZ_ITEMS_SLEEPER;
3769 
3770 		/*
3771 		 * If we're still at the limit, restart.  Notably do not
3772 		 * block on other sleepers.  Cache the max value to protect
3773 		 * against changes via sysctl.
3774 		 */
3775 		total = UZ_ITEMS_COUNT(old);
3776 		max = zone->uz_max_items;
3777 		if (total >= max)
3778 			continue;
3779 		/* Truncate if necessary, otherwise wake other sleepers. */
3780 		if (total + count > max) {
3781 			zone_free_limit(zone, total + count - max);
3782 			count = max - total;
3783 		} else if (total + count < max && UZ_ITEMS_SLEEPERS(old) != 0)
3784 			wakeup_one(&zone->uz_max_items);
3785 
3786 		return (count);
3787 	}
3788 }
3789 
3790 /*
3791  * Allocate 'count' items from our max_items limit.  Returns the number
3792  * available.  If M_NOWAIT is not specified it will sleep until at least
3793  * one item can be allocated.
3794  */
3795 static int
3796 zone_alloc_limit(uma_zone_t zone, int count, int flags)
3797 {
3798 	uint64_t old;
3799 	uint64_t max;
3800 
3801 	max = zone->uz_max_items;
3802 	MPASS(max > 0);
3803 
3804 	/*
3805 	 * We expect normal allocations to succeed with a simple
3806 	 * fetchadd.
3807 	 */
3808 	old = atomic_fetchadd_64(&zone->uz_items, count);
3809 	if (__predict_true(old + count <= max))
3810 		return (count);
3811 
3812 	/*
3813 	 * If we had some items and no sleepers just return the
3814 	 * truncated value.  We have to release the excess space
3815 	 * though because that may wake sleepers who weren't woken
3816 	 * because we were temporarily over the limit.
3817 	 */
3818 	if (old < max) {
3819 		zone_free_limit(zone, (old + count) - max);
3820 		return (max - old);
3821 	}
3822 	return (zone_alloc_limit_hard(zone, count, flags));
3823 }
3824 
3825 /*
3826  * Free a number of items back to the limit.
3827  */
3828 static void
3829 zone_free_limit(uma_zone_t zone, int count)
3830 {
3831 	uint64_t old;
3832 
3833 	MPASS(count > 0);
3834 
3835 	/*
3836 	 * In the common case we either have no sleepers or
3837 	 * are still over the limit and can just return.
3838 	 */
3839 	old = atomic_fetchadd_64(&zone->uz_items, -count);
3840 	if (__predict_true(UZ_ITEMS_SLEEPERS(old) == 0 ||
3841 	   UZ_ITEMS_COUNT(old) - count >= zone->uz_max_items))
3842 		return;
3843 
3844 	/*
3845 	 * Moderate the rate of wakeups.  Sleepers will continue
3846 	 * to generate wakeups if necessary.
3847 	 */
3848 	wakeup_one(&zone->uz_max_items);
3849 }
3850 
3851 static uma_bucket_t
3852 zone_alloc_bucket(uma_zone_t zone, void *udata, int domain, int flags)
3853 {
3854 	uma_bucket_t bucket;
3855 	int maxbucket, cnt;
3856 
3857 	CTR3(KTR_UMA, "zone_alloc_bucket zone %s(%p) domain %d", zone->uz_name,
3858 	    zone, domain);
3859 
3860 	/* Avoid allocs targeting empty domains. */
3861 	if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
3862 		domain = UMA_ANYDOMAIN;
3863 	if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
3864 		domain = UMA_ANYDOMAIN;
3865 
3866 	if (zone->uz_max_items > 0)
3867 		maxbucket = zone_alloc_limit(zone, zone->uz_bucket_size,
3868 		    M_NOWAIT);
3869 	else
3870 		maxbucket = zone->uz_bucket_size;
3871 	if (maxbucket == 0)
3872 		return (false);
3873 
3874 	/* Don't wait for buckets, preserve caller's NOVM setting. */
3875 	bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
3876 	if (bucket == NULL) {
3877 		cnt = 0;
3878 		goto out;
3879 	}
3880 
3881 	bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
3882 	    MIN(maxbucket, bucket->ub_entries), domain, flags);
3883 
3884 	/*
3885 	 * Initialize the memory if necessary.
3886 	 */
3887 	if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
3888 		int i;
3889 
3890 		for (i = 0; i < bucket->ub_cnt; i++)
3891 			if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size,
3892 			    flags) != 0)
3893 				break;
3894 		/*
3895 		 * If we couldn't initialize the whole bucket, put the
3896 		 * rest back onto the freelist.
3897 		 */
3898 		if (i != bucket->ub_cnt) {
3899 			zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
3900 			    bucket->ub_cnt - i);
3901 #ifdef INVARIANTS
3902 			bzero(&bucket->ub_bucket[i],
3903 			    sizeof(void *) * (bucket->ub_cnt - i));
3904 #endif
3905 			bucket->ub_cnt = i;
3906 		}
3907 	}
3908 
3909 	cnt = bucket->ub_cnt;
3910 	if (bucket->ub_cnt == 0) {
3911 		bucket_free(zone, bucket, udata);
3912 		counter_u64_add(zone->uz_fails, 1);
3913 		bucket = NULL;
3914 	}
3915 out:
3916 	if (zone->uz_max_items > 0 && cnt < maxbucket)
3917 		zone_free_limit(zone, maxbucket - cnt);
3918 
3919 	return (bucket);
3920 }
3921 
3922 /*
3923  * Allocates a single item from a zone.
3924  *
3925  * Arguments
3926  *	zone   The zone to alloc for.
3927  *	udata  The data to be passed to the constructor.
3928  *	domain The domain to allocate from or UMA_ANYDOMAIN.
3929  *	flags  M_WAITOK, M_NOWAIT, M_ZERO.
3930  *
3931  * Returns
3932  *	NULL if there is no memory and M_NOWAIT is set
3933  *	An item if successful
3934  */
3935 
3936 static void *
3937 zone_alloc_item(uma_zone_t zone, void *udata, int domain, int flags)
3938 {
3939 	void *item;
3940 
3941 	if (zone->uz_max_items > 0 && zone_alloc_limit(zone, 1, flags) == 0)
3942 		return (NULL);
3943 
3944 	/* Avoid allocs targeting empty domains. */
3945 	if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
3946 		domain = UMA_ANYDOMAIN;
3947 
3948 	if (zone->uz_import(zone->uz_arg, &item, 1, domain, flags) != 1)
3949 		goto fail_cnt;
3950 
3951 	/*
3952 	 * We have to call both the zone's init (not the keg's init)
3953 	 * and the zone's ctor.  This is because the item is going from
3954 	 * a keg slab directly to the user, and the user is expecting it
3955 	 * to be both zone-init'd as well as zone-ctor'd.
3956 	 */
3957 	if (zone->uz_init != NULL) {
3958 		if (zone->uz_init(item, zone->uz_size, flags) != 0) {
3959 			zone_free_item(zone, item, udata, SKIP_FINI | SKIP_CNT);
3960 			goto fail_cnt;
3961 		}
3962 	}
3963 	item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata, flags,
3964 	    item);
3965 	if (item == NULL)
3966 		goto fail;
3967 
3968 	counter_u64_add(zone->uz_allocs, 1);
3969 	CTR3(KTR_UMA, "zone_alloc_item item %p from %s(%p)", item,
3970 	    zone->uz_name, zone);
3971 
3972 	return (item);
3973 
3974 fail_cnt:
3975 	counter_u64_add(zone->uz_fails, 1);
3976 fail:
3977 	if (zone->uz_max_items > 0)
3978 		zone_free_limit(zone, 1);
3979 	CTR2(KTR_UMA, "zone_alloc_item failed from %s(%p)",
3980 	    zone->uz_name, zone);
3981 
3982 	return (NULL);
3983 }
3984 
3985 /* See uma.h */
3986 void
3987 uma_zfree_smr(uma_zone_t zone, void *item)
3988 {
3989 	uma_cache_t cache;
3990 	uma_cache_bucket_t bucket;
3991 	int itemdomain, uz_flags;
3992 
3993 #ifdef UMA_ZALLOC_DEBUG
3994 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
3995 	    ("uma_zfree_smr: called with non-SMR zone.\n"));
3996 	KASSERT(item != NULL, ("uma_zfree_smr: Called with NULL pointer."));
3997 	SMR_ASSERT_NOT_ENTERED(zone->uz_smr);
3998 	if (uma_zfree_debug(zone, item, NULL) == EJUSTRETURN)
3999 		return;
4000 #endif
4001 	cache = &zone->uz_cpu[curcpu];
4002 	uz_flags = cache_uz_flags(cache);
4003 	itemdomain = 0;
4004 #ifdef NUMA
4005 	if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4006 		itemdomain = _vm_phys_domain(pmap_kextract((vm_offset_t)item));
4007 #endif
4008 	critical_enter();
4009 	do {
4010 		cache = &zone->uz_cpu[curcpu];
4011 		/* SMR Zones must free to the free bucket. */
4012 		bucket = &cache->uc_freebucket;
4013 #ifdef NUMA
4014 		if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4015 		    PCPU_GET(domain) != itemdomain) {
4016 			bucket = &cache->uc_crossbucket;
4017 		}
4018 #endif
4019 		if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4020 			cache_bucket_push(cache, bucket, item);
4021 			critical_exit();
4022 			return;
4023 		}
4024 	} while (cache_free(zone, cache, NULL, item, itemdomain));
4025 	critical_exit();
4026 
4027 	/*
4028 	 * If nothing else caught this, we'll just do an internal free.
4029 	 */
4030 	zone_free_item(zone, item, NULL, SKIP_NONE);
4031 }
4032 
4033 /* See uma.h */
4034 void
4035 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
4036 {
4037 	uma_cache_t cache;
4038 	uma_cache_bucket_t bucket;
4039 	int itemdomain, uz_flags;
4040 
4041 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
4042 	random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
4043 
4044 	CTR2(KTR_UMA, "uma_zfree_arg zone %s(%p)", zone->uz_name, zone);
4045 
4046 #ifdef UMA_ZALLOC_DEBUG
4047 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
4048 	    ("uma_zfree_arg: called with SMR zone.\n"));
4049 	if (uma_zfree_debug(zone, item, udata) == EJUSTRETURN)
4050 		return;
4051 #endif
4052         /* uma_zfree(..., NULL) does nothing, to match free(9). */
4053         if (item == NULL)
4054                 return;
4055 
4056 	/*
4057 	 * We are accessing the per-cpu cache without a critical section to
4058 	 * fetch size and flags.  This is acceptable, if we are preempted we
4059 	 * will simply read another cpu's line.
4060 	 */
4061 	cache = &zone->uz_cpu[curcpu];
4062 	uz_flags = cache_uz_flags(cache);
4063 	if (UMA_ALWAYS_CTORDTOR ||
4064 	    __predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0))
4065 		item_dtor(zone, item, cache_uz_size(cache), udata, SKIP_NONE);
4066 
4067 	/*
4068 	 * The race here is acceptable.  If we miss it we'll just have to wait
4069 	 * a little longer for the limits to be reset.
4070 	 */
4071 	if (__predict_false(uz_flags & UMA_ZFLAG_LIMIT)) {
4072 		if (zone->uz_sleepers > 0)
4073 			goto zfree_item;
4074 	}
4075 
4076 	/*
4077 	 * If possible, free to the per-CPU cache.  There are two
4078 	 * requirements for safe access to the per-CPU cache: (1) the thread
4079 	 * accessing the cache must not be preempted or yield during access,
4080 	 * and (2) the thread must not migrate CPUs without switching which
4081 	 * cache it accesses.  We rely on a critical section to prevent
4082 	 * preemption and migration.  We release the critical section in
4083 	 * order to acquire the zone mutex if we are unable to free to the
4084 	 * current cache; when we re-acquire the critical section, we must
4085 	 * detect and handle migration if it has occurred.
4086 	 */
4087 	itemdomain = 0;
4088 #ifdef NUMA
4089 	if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4090 		itemdomain = _vm_phys_domain(pmap_kextract((vm_offset_t)item));
4091 #endif
4092 	critical_enter();
4093 	do {
4094 		cache = &zone->uz_cpu[curcpu];
4095 		/*
4096 		 * Try to free into the allocbucket first to give LIFO
4097 		 * ordering for cache-hot datastructures.  Spill over
4098 		 * into the freebucket if necessary.  Alloc will swap
4099 		 * them if one runs dry.
4100 		 */
4101 		bucket = &cache->uc_allocbucket;
4102 #ifdef NUMA
4103 		if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4104 		    PCPU_GET(domain) != itemdomain) {
4105 			bucket = &cache->uc_crossbucket;
4106 		} else
4107 #endif
4108 		if (bucket->ucb_cnt == bucket->ucb_entries &&
4109 		   cache->uc_freebucket.ucb_cnt <
4110 		   cache->uc_freebucket.ucb_entries)
4111 			cache_bucket_swap(&cache->uc_freebucket,
4112 			    &cache->uc_allocbucket);
4113 		if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4114 			cache_bucket_push(cache, bucket, item);
4115 			critical_exit();
4116 			return;
4117 		}
4118 	} while (cache_free(zone, cache, udata, item, itemdomain));
4119 	critical_exit();
4120 
4121 	/*
4122 	 * If nothing else caught this, we'll just do an internal free.
4123 	 */
4124 zfree_item:
4125 	zone_free_item(zone, item, udata, SKIP_DTOR);
4126 }
4127 
4128 #ifdef NUMA
4129 /*
4130  * sort crossdomain free buckets to domain correct buckets and cache
4131  * them.
4132  */
4133 static void
4134 zone_free_cross(uma_zone_t zone, uma_bucket_t bucket, void *udata)
4135 {
4136 	struct uma_bucketlist fullbuckets;
4137 	uma_zone_domain_t zdom;
4138 	uma_bucket_t b;
4139 	smr_seq_t seq;
4140 	void *item;
4141 	int domain;
4142 
4143 	CTR3(KTR_UMA,
4144 	    "uma_zfree: zone %s(%p) draining cross bucket %p",
4145 	    zone->uz_name, zone, bucket);
4146 
4147 	/*
4148 	 * It is possible for buckets to arrive here out of order so we fetch
4149 	 * the current smr seq rather than accepting the bucket's.
4150 	 */
4151 	seq = SMR_SEQ_INVALID;
4152 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
4153 		seq = smr_advance(zone->uz_smr);
4154 
4155 	/*
4156 	 * To avoid having ndomain * ndomain buckets for sorting we have a
4157 	 * lock on the current crossfree bucket.  A full matrix with
4158 	 * per-domain locking could be used if necessary.
4159 	 */
4160 	STAILQ_INIT(&fullbuckets);
4161 	ZONE_CROSS_LOCK(zone);
4162 	while (bucket->ub_cnt > 0) {
4163 		item = bucket->ub_bucket[bucket->ub_cnt - 1];
4164 		domain = _vm_phys_domain(pmap_kextract((vm_offset_t)item));
4165 		zdom = ZDOM_GET(zone, domain);
4166 		if (zdom->uzd_cross == NULL) {
4167 			zdom->uzd_cross = bucket_alloc(zone, udata, M_NOWAIT);
4168 			if (zdom->uzd_cross == NULL)
4169 				break;
4170 		}
4171 		b = zdom->uzd_cross;
4172 		b->ub_bucket[b->ub_cnt++] = item;
4173 		b->ub_seq = seq;
4174 		if (b->ub_cnt == b->ub_entries) {
4175 			STAILQ_INSERT_HEAD(&fullbuckets, b, ub_link);
4176 			zdom->uzd_cross = NULL;
4177 		}
4178 		bucket->ub_cnt--;
4179 	}
4180 	ZONE_CROSS_UNLOCK(zone);
4181 	if (bucket->ub_cnt == 0)
4182 		bucket->ub_seq = SMR_SEQ_INVALID;
4183 	bucket_free(zone, bucket, udata);
4184 
4185 	while ((b = STAILQ_FIRST(&fullbuckets)) != NULL) {
4186 		STAILQ_REMOVE_HEAD(&fullbuckets, ub_link);
4187 		domain = _vm_phys_domain(pmap_kextract(
4188 		    (vm_offset_t)b->ub_bucket[0]));
4189 		zone_put_bucket(zone, domain, b, udata, true);
4190 	}
4191 }
4192 #endif
4193 
4194 static void
4195 zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata,
4196     int itemdomain, bool ws)
4197 {
4198 
4199 #ifdef NUMA
4200 	/*
4201 	 * Buckets coming from the wrong domain will be entirely for the
4202 	 * only other domain on two domain systems.  In this case we can
4203 	 * simply cache them.  Otherwise we need to sort them back to
4204 	 * correct domains.
4205 	 */
4206 	if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4207 	    vm_ndomains > 2 && PCPU_GET(domain) != itemdomain) {
4208 		zone_free_cross(zone, bucket, udata);
4209 		return;
4210 	}
4211 #endif
4212 
4213 	/*
4214 	 * Attempt to save the bucket in the zone's domain bucket cache.
4215 	 */
4216 	CTR3(KTR_UMA,
4217 	    "uma_zfree: zone %s(%p) putting bucket %p on free list",
4218 	    zone->uz_name, zone, bucket);
4219 	/* ub_cnt is pointing to the last free item */
4220 	if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
4221 		itemdomain = zone_domain_lowest(zone, itemdomain);
4222 	zone_put_bucket(zone, itemdomain, bucket, udata, ws);
4223 }
4224 
4225 /*
4226  * Populate a free or cross bucket for the current cpu cache.  Free any
4227  * existing full bucket either to the zone cache or back to the slab layer.
4228  *
4229  * Enters and returns in a critical section.  false return indicates that
4230  * we can not satisfy this free in the cache layer.  true indicates that
4231  * the caller should retry.
4232  */
4233 static __noinline bool
4234 cache_free(uma_zone_t zone, uma_cache_t cache, void *udata, void *item,
4235     int itemdomain)
4236 {
4237 	uma_cache_bucket_t cbucket;
4238 	uma_bucket_t newbucket, bucket;
4239 
4240 	CRITICAL_ASSERT(curthread);
4241 
4242 	if (zone->uz_bucket_size == 0)
4243 		return false;
4244 
4245 	cache = &zone->uz_cpu[curcpu];
4246 	newbucket = NULL;
4247 
4248 	/*
4249 	 * FIRSTTOUCH domains need to free to the correct zdom.  When
4250 	 * enabled this is the zdom of the item.   The bucket is the
4251 	 * cross bucket if the current domain and itemdomain do not match.
4252 	 */
4253 	cbucket = &cache->uc_freebucket;
4254 #ifdef NUMA
4255 	if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4256 		if (PCPU_GET(domain) != itemdomain) {
4257 			cbucket = &cache->uc_crossbucket;
4258 			if (cbucket->ucb_cnt != 0)
4259 				counter_u64_add(zone->uz_xdomain,
4260 				    cbucket->ucb_cnt);
4261 		}
4262 	}
4263 #endif
4264 	bucket = cache_bucket_unload(cbucket);
4265 	KASSERT(bucket == NULL || bucket->ub_cnt == bucket->ub_entries,
4266 	    ("cache_free: Entered with non-full free bucket."));
4267 
4268 	/* We are no longer associated with this CPU. */
4269 	critical_exit();
4270 
4271 	/*
4272 	 * Don't let SMR zones operate without a free bucket.  Force
4273 	 * a synchronize and re-use this one.  We will only degrade
4274 	 * to a synchronize every bucket_size items rather than every
4275 	 * item if we fail to allocate a bucket.
4276 	 */
4277 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0) {
4278 		if (bucket != NULL)
4279 			bucket->ub_seq = smr_advance(zone->uz_smr);
4280 		newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4281 		if (newbucket == NULL && bucket != NULL) {
4282 			bucket_drain(zone, bucket);
4283 			newbucket = bucket;
4284 			bucket = NULL;
4285 		}
4286 	} else if (!bucketdisable)
4287 		newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4288 
4289 	if (bucket != NULL)
4290 		zone_free_bucket(zone, bucket, udata, itemdomain, true);
4291 
4292 	critical_enter();
4293 	if ((bucket = newbucket) == NULL)
4294 		return (false);
4295 	cache = &zone->uz_cpu[curcpu];
4296 #ifdef NUMA
4297 	/*
4298 	 * Check to see if we should be populating the cross bucket.  If it
4299 	 * is already populated we will fall through and attempt to populate
4300 	 * the free bucket.
4301 	 */
4302 	if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4303 		if (PCPU_GET(domain) != itemdomain &&
4304 		    cache->uc_crossbucket.ucb_bucket == NULL) {
4305 			cache_bucket_load_cross(cache, bucket);
4306 			return (true);
4307 		}
4308 	}
4309 #endif
4310 	/*
4311 	 * We may have lost the race to fill the bucket or switched CPUs.
4312 	 */
4313 	if (cache->uc_freebucket.ucb_bucket != NULL) {
4314 		critical_exit();
4315 		bucket_free(zone, bucket, udata);
4316 		critical_enter();
4317 	} else
4318 		cache_bucket_load_free(cache, bucket);
4319 
4320 	return (true);
4321 }
4322 
4323 void
4324 uma_zfree_domain(uma_zone_t zone, void *item, void *udata)
4325 {
4326 
4327 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
4328 	random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
4329 
4330 	CTR2(KTR_UMA, "uma_zfree_domain zone %s(%p)", zone->uz_name, zone);
4331 
4332 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
4333 	    ("uma_zfree_domain: called with spinlock or critical section held"));
4334 
4335         /* uma_zfree(..., NULL) does nothing, to match free(9). */
4336         if (item == NULL)
4337                 return;
4338 	zone_free_item(zone, item, udata, SKIP_NONE);
4339 }
4340 
4341 static void
4342 slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item)
4343 {
4344 	uma_keg_t keg;
4345 	uma_domain_t dom;
4346 	int freei;
4347 
4348 	keg = zone->uz_keg;
4349 	KEG_LOCK_ASSERT(keg, slab->us_domain);
4350 
4351 	/* Do we need to remove from any lists? */
4352 	dom = &keg->uk_domain[slab->us_domain];
4353 	if (slab->us_freecount + 1 == keg->uk_ipers) {
4354 		LIST_REMOVE(slab, us_link);
4355 		LIST_INSERT_HEAD(&dom->ud_free_slab, slab, us_link);
4356 		dom->ud_free_slabs++;
4357 	} else if (slab->us_freecount == 0) {
4358 		LIST_REMOVE(slab, us_link);
4359 		LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
4360 	}
4361 
4362 	/* Slab management. */
4363 	freei = slab_item_index(slab, keg, item);
4364 	BIT_SET(keg->uk_ipers, freei, &slab->us_free);
4365 	slab->us_freecount++;
4366 
4367 	/* Keg statistics. */
4368 	dom->ud_free_items++;
4369 }
4370 
4371 static void
4372 zone_release(void *arg, void **bucket, int cnt)
4373 {
4374 	struct mtx *lock;
4375 	uma_zone_t zone;
4376 	uma_slab_t slab;
4377 	uma_keg_t keg;
4378 	uint8_t *mem;
4379 	void *item;
4380 	int i;
4381 
4382 	zone = arg;
4383 	keg = zone->uz_keg;
4384 	lock = NULL;
4385 	if (__predict_false((zone->uz_flags & UMA_ZFLAG_HASH) != 0))
4386 		lock = KEG_LOCK(keg, 0);
4387 	for (i = 0; i < cnt; i++) {
4388 		item = bucket[i];
4389 		if (__predict_true((zone->uz_flags & UMA_ZFLAG_VTOSLAB) != 0)) {
4390 			slab = vtoslab((vm_offset_t)item);
4391 		} else {
4392 			mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
4393 			if ((zone->uz_flags & UMA_ZFLAG_HASH) != 0)
4394 				slab = hash_sfind(&keg->uk_hash, mem);
4395 			else
4396 				slab = (uma_slab_t)(mem + keg->uk_pgoff);
4397 		}
4398 		if (lock != KEG_LOCKPTR(keg, slab->us_domain)) {
4399 			if (lock != NULL)
4400 				mtx_unlock(lock);
4401 			lock = KEG_LOCK(keg, slab->us_domain);
4402 		}
4403 		slab_free_item(zone, slab, item);
4404 	}
4405 	if (lock != NULL)
4406 		mtx_unlock(lock);
4407 }
4408 
4409 /*
4410  * Frees a single item to any zone.
4411  *
4412  * Arguments:
4413  *	zone   The zone to free to
4414  *	item   The item we're freeing
4415  *	udata  User supplied data for the dtor
4416  *	skip   Skip dtors and finis
4417  */
4418 static __noinline void
4419 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
4420 {
4421 
4422 	/*
4423 	 * If a free is sent directly to an SMR zone we have to
4424 	 * synchronize immediately because the item can instantly
4425 	 * be reallocated. This should only happen in degenerate
4426 	 * cases when no memory is available for per-cpu caches.
4427 	 */
4428 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0 && skip == SKIP_NONE)
4429 		smr_synchronize(zone->uz_smr);
4430 
4431 	item_dtor(zone, item, zone->uz_size, udata, skip);
4432 
4433 	if (skip < SKIP_FINI && zone->uz_fini)
4434 		zone->uz_fini(item, zone->uz_size);
4435 
4436 	zone->uz_release(zone->uz_arg, &item, 1);
4437 
4438 	if (skip & SKIP_CNT)
4439 		return;
4440 
4441 	counter_u64_add(zone->uz_frees, 1);
4442 
4443 	if (zone->uz_max_items > 0)
4444 		zone_free_limit(zone, 1);
4445 }
4446 
4447 /* See uma.h */
4448 int
4449 uma_zone_set_max(uma_zone_t zone, int nitems)
4450 {
4451 	struct uma_bucket_zone *ubz;
4452 	int count;
4453 
4454 	/*
4455 	 * XXX This can misbehave if the zone has any allocations with
4456 	 * no limit and a limit is imposed.  There is currently no
4457 	 * way to clear a limit.
4458 	 */
4459 	ZONE_LOCK(zone);
4460 	ubz = bucket_zone_max(zone, nitems);
4461 	count = ubz != NULL ? ubz->ubz_entries : 0;
4462 	zone->uz_bucket_size_max = zone->uz_bucket_size = count;
4463 	if (zone->uz_bucket_size_min > zone->uz_bucket_size_max)
4464 		zone->uz_bucket_size_min = zone->uz_bucket_size_max;
4465 	zone->uz_max_items = nitems;
4466 	zone->uz_flags |= UMA_ZFLAG_LIMIT;
4467 	zone_update_caches(zone);
4468 	/* We may need to wake waiters. */
4469 	wakeup(&zone->uz_max_items);
4470 	ZONE_UNLOCK(zone);
4471 
4472 	return (nitems);
4473 }
4474 
4475 /* See uma.h */
4476 void
4477 uma_zone_set_maxcache(uma_zone_t zone, int nitems)
4478 {
4479 	struct uma_bucket_zone *ubz;
4480 	int bpcpu;
4481 
4482 	ZONE_LOCK(zone);
4483 	ubz = bucket_zone_max(zone, nitems);
4484 	if (ubz != NULL) {
4485 		bpcpu = 2;
4486 		if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4487 			/* Count the cross-domain bucket. */
4488 			bpcpu++;
4489 		nitems -= ubz->ubz_entries * bpcpu * mp_ncpus;
4490 		zone->uz_bucket_size_max = ubz->ubz_entries;
4491 	} else {
4492 		zone->uz_bucket_size_max = zone->uz_bucket_size = 0;
4493 	}
4494 	if (zone->uz_bucket_size_min > zone->uz_bucket_size_max)
4495 		zone->uz_bucket_size_min = zone->uz_bucket_size_max;
4496 	zone->uz_bucket_max = nitems / vm_ndomains;
4497 	ZONE_UNLOCK(zone);
4498 }
4499 
4500 /* See uma.h */
4501 int
4502 uma_zone_get_max(uma_zone_t zone)
4503 {
4504 	int nitems;
4505 
4506 	nitems = atomic_load_64(&zone->uz_max_items);
4507 
4508 	return (nitems);
4509 }
4510 
4511 /* See uma.h */
4512 void
4513 uma_zone_set_warning(uma_zone_t zone, const char *warning)
4514 {
4515 
4516 	ZONE_ASSERT_COLD(zone);
4517 	zone->uz_warning = warning;
4518 }
4519 
4520 /* See uma.h */
4521 void
4522 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction)
4523 {
4524 
4525 	ZONE_ASSERT_COLD(zone);
4526 	TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone);
4527 }
4528 
4529 /* See uma.h */
4530 int
4531 uma_zone_get_cur(uma_zone_t zone)
4532 {
4533 	int64_t nitems;
4534 	u_int i;
4535 
4536 	nitems = 0;
4537 	if (zone->uz_allocs != EARLY_COUNTER && zone->uz_frees != EARLY_COUNTER)
4538 		nitems = counter_u64_fetch(zone->uz_allocs) -
4539 		    counter_u64_fetch(zone->uz_frees);
4540 	CPU_FOREACH(i)
4541 		nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs) -
4542 		    atomic_load_64(&zone->uz_cpu[i].uc_frees);
4543 
4544 	return (nitems < 0 ? 0 : nitems);
4545 }
4546 
4547 static uint64_t
4548 uma_zone_get_allocs(uma_zone_t zone)
4549 {
4550 	uint64_t nitems;
4551 	u_int i;
4552 
4553 	nitems = 0;
4554 	if (zone->uz_allocs != EARLY_COUNTER)
4555 		nitems = counter_u64_fetch(zone->uz_allocs);
4556 	CPU_FOREACH(i)
4557 		nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs);
4558 
4559 	return (nitems);
4560 }
4561 
4562 static uint64_t
4563 uma_zone_get_frees(uma_zone_t zone)
4564 {
4565 	uint64_t nitems;
4566 	u_int i;
4567 
4568 	nitems = 0;
4569 	if (zone->uz_frees != EARLY_COUNTER)
4570 		nitems = counter_u64_fetch(zone->uz_frees);
4571 	CPU_FOREACH(i)
4572 		nitems += atomic_load_64(&zone->uz_cpu[i].uc_frees);
4573 
4574 	return (nitems);
4575 }
4576 
4577 #ifdef INVARIANTS
4578 /* Used only for KEG_ASSERT_COLD(). */
4579 static uint64_t
4580 uma_keg_get_allocs(uma_keg_t keg)
4581 {
4582 	uma_zone_t z;
4583 	uint64_t nitems;
4584 
4585 	nitems = 0;
4586 	LIST_FOREACH(z, &keg->uk_zones, uz_link)
4587 		nitems += uma_zone_get_allocs(z);
4588 
4589 	return (nitems);
4590 }
4591 #endif
4592 
4593 /* See uma.h */
4594 void
4595 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
4596 {
4597 	uma_keg_t keg;
4598 
4599 	KEG_GET(zone, keg);
4600 	KEG_ASSERT_COLD(keg);
4601 	keg->uk_init = uminit;
4602 }
4603 
4604 /* See uma.h */
4605 void
4606 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
4607 {
4608 	uma_keg_t keg;
4609 
4610 	KEG_GET(zone, keg);
4611 	KEG_ASSERT_COLD(keg);
4612 	keg->uk_fini = fini;
4613 }
4614 
4615 /* See uma.h */
4616 void
4617 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
4618 {
4619 
4620 	ZONE_ASSERT_COLD(zone);
4621 	zone->uz_init = zinit;
4622 }
4623 
4624 /* See uma.h */
4625 void
4626 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
4627 {
4628 
4629 	ZONE_ASSERT_COLD(zone);
4630 	zone->uz_fini = zfini;
4631 }
4632 
4633 /* See uma.h */
4634 void
4635 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
4636 {
4637 	uma_keg_t keg;
4638 
4639 	KEG_GET(zone, keg);
4640 	KEG_ASSERT_COLD(keg);
4641 	keg->uk_freef = freef;
4642 }
4643 
4644 /* See uma.h */
4645 void
4646 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
4647 {
4648 	uma_keg_t keg;
4649 
4650 	KEG_GET(zone, keg);
4651 	KEG_ASSERT_COLD(keg);
4652 	keg->uk_allocf = allocf;
4653 }
4654 
4655 /* See uma.h */
4656 void
4657 uma_zone_set_smr(uma_zone_t zone, smr_t smr)
4658 {
4659 
4660 	ZONE_ASSERT_COLD(zone);
4661 
4662 	KASSERT(smr != NULL, ("Got NULL smr"));
4663 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
4664 	    ("zone %p (%s) already uses SMR", zone, zone->uz_name));
4665 	zone->uz_flags |= UMA_ZONE_SMR;
4666 	zone->uz_smr = smr;
4667 	zone_update_caches(zone);
4668 }
4669 
4670 smr_t
4671 uma_zone_get_smr(uma_zone_t zone)
4672 {
4673 
4674 	return (zone->uz_smr);
4675 }
4676 
4677 /* See uma.h */
4678 void
4679 uma_zone_reserve(uma_zone_t zone, int items)
4680 {
4681 	uma_keg_t keg;
4682 
4683 	KEG_GET(zone, keg);
4684 	KEG_ASSERT_COLD(keg);
4685 	keg->uk_reserve = items;
4686 }
4687 
4688 /* See uma.h */
4689 int
4690 uma_zone_reserve_kva(uma_zone_t zone, int count)
4691 {
4692 	uma_keg_t keg;
4693 	vm_offset_t kva;
4694 	u_int pages;
4695 
4696 	KEG_GET(zone, keg);
4697 	KEG_ASSERT_COLD(keg);
4698 	ZONE_ASSERT_COLD(zone);
4699 
4700 	pages = howmany(count, keg->uk_ipers) * keg->uk_ppera;
4701 
4702 #ifdef UMA_MD_SMALL_ALLOC
4703 	if (keg->uk_ppera > 1) {
4704 #else
4705 	if (1) {
4706 #endif
4707 		kva = kva_alloc((vm_size_t)pages * PAGE_SIZE);
4708 		if (kva == 0)
4709 			return (0);
4710 	} else
4711 		kva = 0;
4712 
4713 	MPASS(keg->uk_kva == 0);
4714 	keg->uk_kva = kva;
4715 	keg->uk_offset = 0;
4716 	zone->uz_max_items = pages * keg->uk_ipers;
4717 #ifdef UMA_MD_SMALL_ALLOC
4718 	keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
4719 #else
4720 	keg->uk_allocf = noobj_alloc;
4721 #endif
4722 	keg->uk_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
4723 	zone->uz_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
4724 	zone_update_caches(zone);
4725 
4726 	return (1);
4727 }
4728 
4729 /* See uma.h */
4730 void
4731 uma_prealloc(uma_zone_t zone, int items)
4732 {
4733 	struct vm_domainset_iter di;
4734 	uma_domain_t dom;
4735 	uma_slab_t slab;
4736 	uma_keg_t keg;
4737 	int aflags, domain, slabs;
4738 
4739 	KEG_GET(zone, keg);
4740 	slabs = howmany(items, keg->uk_ipers);
4741 	while (slabs-- > 0) {
4742 		aflags = M_NOWAIT;
4743 		vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
4744 		    &aflags);
4745 		for (;;) {
4746 			slab = keg_alloc_slab(keg, zone, domain, M_WAITOK,
4747 			    aflags);
4748 			if (slab != NULL) {
4749 				dom = &keg->uk_domain[slab->us_domain];
4750 				/*
4751 				 * keg_alloc_slab() always returns a slab on the
4752 				 * partial list.
4753 				 */
4754 				LIST_REMOVE(slab, us_link);
4755 				LIST_INSERT_HEAD(&dom->ud_free_slab, slab,
4756 				    us_link);
4757 				dom->ud_free_slabs++;
4758 				KEG_UNLOCK(keg, slab->us_domain);
4759 				break;
4760 			}
4761 			if (vm_domainset_iter_policy(&di, &domain) != 0)
4762 				vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask);
4763 		}
4764 	}
4765 }
4766 
4767 /*
4768  * Returns a snapshot of memory consumption in bytes.
4769  */
4770 size_t
4771 uma_zone_memory(uma_zone_t zone)
4772 {
4773 	size_t sz;
4774 	int i;
4775 
4776 	sz = 0;
4777 	if (zone->uz_flags & UMA_ZFLAG_CACHE) {
4778 		for (i = 0; i < vm_ndomains; i++)
4779 			sz += ZDOM_GET(zone, i)->uzd_nitems;
4780 		return (sz * zone->uz_size);
4781 	}
4782 	for (i = 0; i < vm_ndomains; i++)
4783 		sz += zone->uz_keg->uk_domain[i].ud_pages;
4784 
4785 	return (sz * PAGE_SIZE);
4786 }
4787 
4788 /* See uma.h */
4789 void
4790 uma_reclaim(int req)
4791 {
4792 
4793 	CTR0(KTR_UMA, "UMA: vm asked us to release pages!");
4794 	sx_xlock(&uma_reclaim_lock);
4795 	bucket_enable();
4796 
4797 	switch (req) {
4798 	case UMA_RECLAIM_TRIM:
4799 		zone_foreach(zone_trim, NULL);
4800 		break;
4801 	case UMA_RECLAIM_DRAIN:
4802 	case UMA_RECLAIM_DRAIN_CPU:
4803 		zone_foreach(zone_drain, NULL);
4804 		if (req == UMA_RECLAIM_DRAIN_CPU) {
4805 			pcpu_cache_drain_safe(NULL);
4806 			zone_foreach(zone_drain, NULL);
4807 		}
4808 		break;
4809 	default:
4810 		panic("unhandled reclamation request %d", req);
4811 	}
4812 
4813 	/*
4814 	 * Some slabs may have been freed but this zone will be visited early
4815 	 * we visit again so that we can free pages that are empty once other
4816 	 * zones are drained.  We have to do the same for buckets.
4817 	 */
4818 	zone_drain(slabzones[0], NULL);
4819 	zone_drain(slabzones[1], NULL);
4820 	bucket_zone_drain();
4821 	sx_xunlock(&uma_reclaim_lock);
4822 }
4823 
4824 static volatile int uma_reclaim_needed;
4825 
4826 void
4827 uma_reclaim_wakeup(void)
4828 {
4829 
4830 	if (atomic_fetchadd_int(&uma_reclaim_needed, 1) == 0)
4831 		wakeup(uma_reclaim);
4832 }
4833 
4834 void
4835 uma_reclaim_worker(void *arg __unused)
4836 {
4837 
4838 	for (;;) {
4839 		sx_xlock(&uma_reclaim_lock);
4840 		while (atomic_load_int(&uma_reclaim_needed) == 0)
4841 			sx_sleep(uma_reclaim, &uma_reclaim_lock, PVM, "umarcl",
4842 			    hz);
4843 		sx_xunlock(&uma_reclaim_lock);
4844 		EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM);
4845 		uma_reclaim(UMA_RECLAIM_DRAIN_CPU);
4846 		atomic_store_int(&uma_reclaim_needed, 0);
4847 		/* Don't fire more than once per-second. */
4848 		pause("umarclslp", hz);
4849 	}
4850 }
4851 
4852 /* See uma.h */
4853 void
4854 uma_zone_reclaim(uma_zone_t zone, int req)
4855 {
4856 
4857 	switch (req) {
4858 	case UMA_RECLAIM_TRIM:
4859 		zone_trim(zone, NULL);
4860 		break;
4861 	case UMA_RECLAIM_DRAIN:
4862 		zone_drain(zone, NULL);
4863 		break;
4864 	case UMA_RECLAIM_DRAIN_CPU:
4865 		pcpu_cache_drain_safe(zone);
4866 		zone_drain(zone, NULL);
4867 		break;
4868 	default:
4869 		panic("unhandled reclamation request %d", req);
4870 	}
4871 }
4872 
4873 /* See uma.h */
4874 int
4875 uma_zone_exhausted(uma_zone_t zone)
4876 {
4877 
4878 	return (atomic_load_32(&zone->uz_sleepers) > 0);
4879 }
4880 
4881 unsigned long
4882 uma_limit(void)
4883 {
4884 
4885 	return (uma_kmem_limit);
4886 }
4887 
4888 void
4889 uma_set_limit(unsigned long limit)
4890 {
4891 
4892 	uma_kmem_limit = limit;
4893 }
4894 
4895 unsigned long
4896 uma_size(void)
4897 {
4898 
4899 	return (atomic_load_long(&uma_kmem_total));
4900 }
4901 
4902 long
4903 uma_avail(void)
4904 {
4905 
4906 	return (uma_kmem_limit - uma_size());
4907 }
4908 
4909 #ifdef DDB
4910 /*
4911  * Generate statistics across both the zone and its per-cpu cache's.  Return
4912  * desired statistics if the pointer is non-NULL for that statistic.
4913  *
4914  * Note: does not update the zone statistics, as it can't safely clear the
4915  * per-CPU cache statistic.
4916  *
4917  */
4918 static void
4919 uma_zone_sumstat(uma_zone_t z, long *cachefreep, uint64_t *allocsp,
4920     uint64_t *freesp, uint64_t *sleepsp, uint64_t *xdomainp)
4921 {
4922 	uma_cache_t cache;
4923 	uint64_t allocs, frees, sleeps, xdomain;
4924 	int cachefree, cpu;
4925 
4926 	allocs = frees = sleeps = xdomain = 0;
4927 	cachefree = 0;
4928 	CPU_FOREACH(cpu) {
4929 		cache = &z->uz_cpu[cpu];
4930 		cachefree += cache->uc_allocbucket.ucb_cnt;
4931 		cachefree += cache->uc_freebucket.ucb_cnt;
4932 		xdomain += cache->uc_crossbucket.ucb_cnt;
4933 		cachefree += cache->uc_crossbucket.ucb_cnt;
4934 		allocs += cache->uc_allocs;
4935 		frees += cache->uc_frees;
4936 	}
4937 	allocs += counter_u64_fetch(z->uz_allocs);
4938 	frees += counter_u64_fetch(z->uz_frees);
4939 	xdomain += counter_u64_fetch(z->uz_xdomain);
4940 	sleeps += z->uz_sleeps;
4941 	if (cachefreep != NULL)
4942 		*cachefreep = cachefree;
4943 	if (allocsp != NULL)
4944 		*allocsp = allocs;
4945 	if (freesp != NULL)
4946 		*freesp = frees;
4947 	if (sleepsp != NULL)
4948 		*sleepsp = sleeps;
4949 	if (xdomainp != NULL)
4950 		*xdomainp = xdomain;
4951 }
4952 #endif /* DDB */
4953 
4954 static int
4955 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
4956 {
4957 	uma_keg_t kz;
4958 	uma_zone_t z;
4959 	int count;
4960 
4961 	count = 0;
4962 	rw_rlock(&uma_rwlock);
4963 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
4964 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
4965 			count++;
4966 	}
4967 	LIST_FOREACH(z, &uma_cachezones, uz_link)
4968 		count++;
4969 
4970 	rw_runlock(&uma_rwlock);
4971 	return (sysctl_handle_int(oidp, &count, 0, req));
4972 }
4973 
4974 static void
4975 uma_vm_zone_stats(struct uma_type_header *uth, uma_zone_t z, struct sbuf *sbuf,
4976     struct uma_percpu_stat *ups, bool internal)
4977 {
4978 	uma_zone_domain_t zdom;
4979 	uma_cache_t cache;
4980 	int i;
4981 
4982 
4983 	for (i = 0; i < vm_ndomains; i++) {
4984 		zdom = ZDOM_GET(z, i);
4985 		uth->uth_zone_free += zdom->uzd_nitems;
4986 	}
4987 	uth->uth_allocs = counter_u64_fetch(z->uz_allocs);
4988 	uth->uth_frees = counter_u64_fetch(z->uz_frees);
4989 	uth->uth_fails = counter_u64_fetch(z->uz_fails);
4990 	uth->uth_xdomain = counter_u64_fetch(z->uz_xdomain);
4991 	uth->uth_sleeps = z->uz_sleeps;
4992 
4993 	for (i = 0; i < mp_maxid + 1; i++) {
4994 		bzero(&ups[i], sizeof(*ups));
4995 		if (internal || CPU_ABSENT(i))
4996 			continue;
4997 		cache = &z->uz_cpu[i];
4998 		ups[i].ups_cache_free += cache->uc_allocbucket.ucb_cnt;
4999 		ups[i].ups_cache_free += cache->uc_freebucket.ucb_cnt;
5000 		ups[i].ups_cache_free += cache->uc_crossbucket.ucb_cnt;
5001 		ups[i].ups_allocs = cache->uc_allocs;
5002 		ups[i].ups_frees = cache->uc_frees;
5003 	}
5004 }
5005 
5006 static int
5007 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
5008 {
5009 	struct uma_stream_header ush;
5010 	struct uma_type_header uth;
5011 	struct uma_percpu_stat *ups;
5012 	struct sbuf sbuf;
5013 	uma_keg_t kz;
5014 	uma_zone_t z;
5015 	uint64_t items;
5016 	uint32_t kfree, pages;
5017 	int count, error, i;
5018 
5019 	error = sysctl_wire_old_buffer(req, 0);
5020 	if (error != 0)
5021 		return (error);
5022 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
5023 	sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL);
5024 	ups = malloc((mp_maxid + 1) * sizeof(*ups), M_TEMP, M_WAITOK);
5025 
5026 	count = 0;
5027 	rw_rlock(&uma_rwlock);
5028 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
5029 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
5030 			count++;
5031 	}
5032 
5033 	LIST_FOREACH(z, &uma_cachezones, uz_link)
5034 		count++;
5035 
5036 	/*
5037 	 * Insert stream header.
5038 	 */
5039 	bzero(&ush, sizeof(ush));
5040 	ush.ush_version = UMA_STREAM_VERSION;
5041 	ush.ush_maxcpus = (mp_maxid + 1);
5042 	ush.ush_count = count;
5043 	(void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
5044 
5045 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
5046 		kfree = pages = 0;
5047 		for (i = 0; i < vm_ndomains; i++) {
5048 			kfree += kz->uk_domain[i].ud_free_items;
5049 			pages += kz->uk_domain[i].ud_pages;
5050 		}
5051 		LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5052 			bzero(&uth, sizeof(uth));
5053 			strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5054 			uth.uth_align = kz->uk_align;
5055 			uth.uth_size = kz->uk_size;
5056 			uth.uth_rsize = kz->uk_rsize;
5057 			if (z->uz_max_items > 0) {
5058 				items = UZ_ITEMS_COUNT(z->uz_items);
5059 				uth.uth_pages = (items / kz->uk_ipers) *
5060 					kz->uk_ppera;
5061 			} else
5062 				uth.uth_pages = pages;
5063 			uth.uth_maxpages = (z->uz_max_items / kz->uk_ipers) *
5064 			    kz->uk_ppera;
5065 			uth.uth_limit = z->uz_max_items;
5066 			uth.uth_keg_free = kfree;
5067 
5068 			/*
5069 			 * A zone is secondary is it is not the first entry
5070 			 * on the keg's zone list.
5071 			 */
5072 			if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
5073 			    (LIST_FIRST(&kz->uk_zones) != z))
5074 				uth.uth_zone_flags = UTH_ZONE_SECONDARY;
5075 			uma_vm_zone_stats(&uth, z, &sbuf, ups,
5076 			    kz->uk_flags & UMA_ZFLAG_INTERNAL);
5077 			(void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5078 			for (i = 0; i < mp_maxid + 1; i++)
5079 				(void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5080 		}
5081 	}
5082 	LIST_FOREACH(z, &uma_cachezones, uz_link) {
5083 		bzero(&uth, sizeof(uth));
5084 		strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5085 		uth.uth_size = z->uz_size;
5086 		uma_vm_zone_stats(&uth, z, &sbuf, ups, false);
5087 		(void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5088 		for (i = 0; i < mp_maxid + 1; i++)
5089 			(void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5090 	}
5091 
5092 	rw_runlock(&uma_rwlock);
5093 	error = sbuf_finish(&sbuf);
5094 	sbuf_delete(&sbuf);
5095 	free(ups, M_TEMP);
5096 	return (error);
5097 }
5098 
5099 int
5100 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS)
5101 {
5102 	uma_zone_t zone = *(uma_zone_t *)arg1;
5103 	int error, max;
5104 
5105 	max = uma_zone_get_max(zone);
5106 	error = sysctl_handle_int(oidp, &max, 0, req);
5107 	if (error || !req->newptr)
5108 		return (error);
5109 
5110 	uma_zone_set_max(zone, max);
5111 
5112 	return (0);
5113 }
5114 
5115 int
5116 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS)
5117 {
5118 	uma_zone_t zone;
5119 	int cur;
5120 
5121 	/*
5122 	 * Some callers want to add sysctls for global zones that
5123 	 * may not yet exist so they pass a pointer to a pointer.
5124 	 */
5125 	if (arg2 == 0)
5126 		zone = *(uma_zone_t *)arg1;
5127 	else
5128 		zone = arg1;
5129 	cur = uma_zone_get_cur(zone);
5130 	return (sysctl_handle_int(oidp, &cur, 0, req));
5131 }
5132 
5133 static int
5134 sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS)
5135 {
5136 	uma_zone_t zone = arg1;
5137 	uint64_t cur;
5138 
5139 	cur = uma_zone_get_allocs(zone);
5140 	return (sysctl_handle_64(oidp, &cur, 0, req));
5141 }
5142 
5143 static int
5144 sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS)
5145 {
5146 	uma_zone_t zone = arg1;
5147 	uint64_t cur;
5148 
5149 	cur = uma_zone_get_frees(zone);
5150 	return (sysctl_handle_64(oidp, &cur, 0, req));
5151 }
5152 
5153 static int
5154 sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS)
5155 {
5156 	struct sbuf sbuf;
5157 	uma_zone_t zone = arg1;
5158 	int error;
5159 
5160 	sbuf_new_for_sysctl(&sbuf, NULL, 0, req);
5161 	if (zone->uz_flags != 0)
5162 		sbuf_printf(&sbuf, "0x%b", zone->uz_flags, PRINT_UMA_ZFLAGS);
5163 	else
5164 		sbuf_printf(&sbuf, "0");
5165 	error = sbuf_finish(&sbuf);
5166 	sbuf_delete(&sbuf);
5167 
5168 	return (error);
5169 }
5170 
5171 static int
5172 sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS)
5173 {
5174 	uma_keg_t keg = arg1;
5175 	int avail, effpct, total;
5176 
5177 	total = keg->uk_ppera * PAGE_SIZE;
5178 	if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0)
5179 		total += slabzone(keg->uk_ipers)->uz_keg->uk_rsize;
5180 	/*
5181 	 * We consider the client's requested size and alignment here, not the
5182 	 * real size determination uk_rsize, because we also adjust the real
5183 	 * size for internal implementation reasons (max bitset size).
5184 	 */
5185 	avail = keg->uk_ipers * roundup2(keg->uk_size, keg->uk_align + 1);
5186 	if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
5187 		avail *= mp_maxid + 1;
5188 	effpct = 100 * avail / total;
5189 	return (sysctl_handle_int(oidp, &effpct, 0, req));
5190 }
5191 
5192 static int
5193 sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS)
5194 {
5195 	uma_zone_t zone = arg1;
5196 	uint64_t cur;
5197 
5198 	cur = UZ_ITEMS_COUNT(atomic_load_64(&zone->uz_items));
5199 	return (sysctl_handle_64(oidp, &cur, 0, req));
5200 }
5201 
5202 #ifdef INVARIANTS
5203 static uma_slab_t
5204 uma_dbg_getslab(uma_zone_t zone, void *item)
5205 {
5206 	uma_slab_t slab;
5207 	uma_keg_t keg;
5208 	uint8_t *mem;
5209 
5210 	/*
5211 	 * It is safe to return the slab here even though the
5212 	 * zone is unlocked because the item's allocation state
5213 	 * essentially holds a reference.
5214 	 */
5215 	mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
5216 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5217 		return (NULL);
5218 	if (zone->uz_flags & UMA_ZFLAG_VTOSLAB)
5219 		return (vtoslab((vm_offset_t)mem));
5220 	keg = zone->uz_keg;
5221 	if ((keg->uk_flags & UMA_ZFLAG_HASH) == 0)
5222 		return ((uma_slab_t)(mem + keg->uk_pgoff));
5223 	KEG_LOCK(keg, 0);
5224 	slab = hash_sfind(&keg->uk_hash, mem);
5225 	KEG_UNLOCK(keg, 0);
5226 
5227 	return (slab);
5228 }
5229 
5230 static bool
5231 uma_dbg_zskip(uma_zone_t zone, void *mem)
5232 {
5233 
5234 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5235 		return (true);
5236 
5237 	return (uma_dbg_kskip(zone->uz_keg, mem));
5238 }
5239 
5240 static bool
5241 uma_dbg_kskip(uma_keg_t keg, void *mem)
5242 {
5243 	uintptr_t idx;
5244 
5245 	if (dbg_divisor == 0)
5246 		return (true);
5247 
5248 	if (dbg_divisor == 1)
5249 		return (false);
5250 
5251 	idx = (uintptr_t)mem >> PAGE_SHIFT;
5252 	if (keg->uk_ipers > 1) {
5253 		idx *= keg->uk_ipers;
5254 		idx += ((uintptr_t)mem & PAGE_MASK) / keg->uk_rsize;
5255 	}
5256 
5257 	if ((idx / dbg_divisor) * dbg_divisor != idx) {
5258 		counter_u64_add(uma_skip_cnt, 1);
5259 		return (true);
5260 	}
5261 	counter_u64_add(uma_dbg_cnt, 1);
5262 
5263 	return (false);
5264 }
5265 
5266 /*
5267  * Set up the slab's freei data such that uma_dbg_free can function.
5268  *
5269  */
5270 static void
5271 uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item)
5272 {
5273 	uma_keg_t keg;
5274 	int freei;
5275 
5276 	if (slab == NULL) {
5277 		slab = uma_dbg_getslab(zone, item);
5278 		if (slab == NULL)
5279 			panic("uma: item %p did not belong to zone %s\n",
5280 			    item, zone->uz_name);
5281 	}
5282 	keg = zone->uz_keg;
5283 	freei = slab_item_index(slab, keg, item);
5284 
5285 	if (BIT_ISSET(keg->uk_ipers, freei, slab_dbg_bits(slab, keg)))
5286 		panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)\n",
5287 		    item, zone, zone->uz_name, slab, freei);
5288 	BIT_SET_ATOMIC(keg->uk_ipers, freei, slab_dbg_bits(slab, keg));
5289 }
5290 
5291 /*
5292  * Verifies freed addresses.  Checks for alignment, valid slab membership
5293  * and duplicate frees.
5294  *
5295  */
5296 static void
5297 uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item)
5298 {
5299 	uma_keg_t keg;
5300 	int freei;
5301 
5302 	if (slab == NULL) {
5303 		slab = uma_dbg_getslab(zone, item);
5304 		if (slab == NULL)
5305 			panic("uma: Freed item %p did not belong to zone %s\n",
5306 			    item, zone->uz_name);
5307 	}
5308 	keg = zone->uz_keg;
5309 	freei = slab_item_index(slab, keg, item);
5310 
5311 	if (freei >= keg->uk_ipers)
5312 		panic("Invalid free of %p from zone %p(%s) slab %p(%d)\n",
5313 		    item, zone, zone->uz_name, slab, freei);
5314 
5315 	if (slab_item(slab, keg, freei) != item)
5316 		panic("Unaligned free of %p from zone %p(%s) slab %p(%d)\n",
5317 		    item, zone, zone->uz_name, slab, freei);
5318 
5319 	if (!BIT_ISSET(keg->uk_ipers, freei, slab_dbg_bits(slab, keg)))
5320 		panic("Duplicate free of %p from zone %p(%s) slab %p(%d)\n",
5321 		    item, zone, zone->uz_name, slab, freei);
5322 
5323 	BIT_CLR_ATOMIC(keg->uk_ipers, freei, slab_dbg_bits(slab, keg));
5324 }
5325 #endif /* INVARIANTS */
5326 
5327 #ifdef DDB
5328 static int64_t
5329 get_uma_stats(uma_keg_t kz, uma_zone_t z, uint64_t *allocs, uint64_t *used,
5330     uint64_t *sleeps, long *cachefree, uint64_t *xdomain)
5331 {
5332 	uint64_t frees;
5333 	int i;
5334 
5335 	if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
5336 		*allocs = counter_u64_fetch(z->uz_allocs);
5337 		frees = counter_u64_fetch(z->uz_frees);
5338 		*sleeps = z->uz_sleeps;
5339 		*cachefree = 0;
5340 		*xdomain = 0;
5341 	} else
5342 		uma_zone_sumstat(z, cachefree, allocs, &frees, sleeps,
5343 		    xdomain);
5344 	for (i = 0; i < vm_ndomains; i++) {
5345 		*cachefree += ZDOM_GET(z, i)->uzd_nitems;
5346 		if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
5347 		    (LIST_FIRST(&kz->uk_zones) != z)))
5348 			*cachefree += kz->uk_domain[i].ud_free_items;
5349 	}
5350 	*used = *allocs - frees;
5351 	return (((int64_t)*used + *cachefree) * kz->uk_size);
5352 }
5353 
5354 DB_SHOW_COMMAND(uma, db_show_uma)
5355 {
5356 	const char *fmt_hdr, *fmt_entry;
5357 	uma_keg_t kz;
5358 	uma_zone_t z;
5359 	uint64_t allocs, used, sleeps, xdomain;
5360 	long cachefree;
5361 	/* variables for sorting */
5362 	uma_keg_t cur_keg;
5363 	uma_zone_t cur_zone, last_zone;
5364 	int64_t cur_size, last_size, size;
5365 	int ties;
5366 
5367 	/* /i option produces machine-parseable CSV output */
5368 	if (modif[0] == 'i') {
5369 		fmt_hdr = "%s,%s,%s,%s,%s,%s,%s,%s,%s\n";
5370 		fmt_entry = "\"%s\",%ju,%jd,%ld,%ju,%ju,%u,%jd,%ju\n";
5371 	} else {
5372 		fmt_hdr = "%18s %6s %7s %7s %11s %7s %7s %10s %8s\n";
5373 		fmt_entry = "%18s %6ju %7jd %7ld %11ju %7ju %7u %10jd %8ju\n";
5374 	}
5375 
5376 	db_printf(fmt_hdr, "Zone", "Size", "Used", "Free", "Requests",
5377 	    "Sleeps", "Bucket", "Total Mem", "XFree");
5378 
5379 	/* Sort the zones with largest size first. */
5380 	last_zone = NULL;
5381 	last_size = INT64_MAX;
5382 	for (;;) {
5383 		cur_zone = NULL;
5384 		cur_size = -1;
5385 		ties = 0;
5386 		LIST_FOREACH(kz, &uma_kegs, uk_link) {
5387 			LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5388 				/*
5389 				 * In the case of size ties, print out zones
5390 				 * in the order they are encountered.  That is,
5391 				 * when we encounter the most recently output
5392 				 * zone, we have already printed all preceding
5393 				 * ties, and we must print all following ties.
5394 				 */
5395 				if (z == last_zone) {
5396 					ties = 1;
5397 					continue;
5398 				}
5399 				size = get_uma_stats(kz, z, &allocs, &used,
5400 				    &sleeps, &cachefree, &xdomain);
5401 				if (size > cur_size && size < last_size + ties)
5402 				{
5403 					cur_size = size;
5404 					cur_zone = z;
5405 					cur_keg = kz;
5406 				}
5407 			}
5408 		}
5409 		if (cur_zone == NULL)
5410 			break;
5411 
5412 		size = get_uma_stats(cur_keg, cur_zone, &allocs, &used,
5413 		    &sleeps, &cachefree, &xdomain);
5414 		db_printf(fmt_entry, cur_zone->uz_name,
5415 		    (uintmax_t)cur_keg->uk_size, (intmax_t)used, cachefree,
5416 		    (uintmax_t)allocs, (uintmax_t)sleeps,
5417 		    (unsigned)cur_zone->uz_bucket_size, (intmax_t)size,
5418 		    xdomain);
5419 
5420 		if (db_pager_quit)
5421 			return;
5422 		last_zone = cur_zone;
5423 		last_size = cur_size;
5424 	}
5425 }
5426 
5427 DB_SHOW_COMMAND(umacache, db_show_umacache)
5428 {
5429 	uma_zone_t z;
5430 	uint64_t allocs, frees;
5431 	long cachefree;
5432 	int i;
5433 
5434 	db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
5435 	    "Requests", "Bucket");
5436 	LIST_FOREACH(z, &uma_cachezones, uz_link) {
5437 		uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL, NULL);
5438 		for (i = 0; i < vm_ndomains; i++)
5439 			cachefree += ZDOM_GET(z, i)->uzd_nitems;
5440 		db_printf("%18s %8ju %8jd %8ld %12ju %8u\n",
5441 		    z->uz_name, (uintmax_t)z->uz_size,
5442 		    (intmax_t)(allocs - frees), cachefree,
5443 		    (uintmax_t)allocs, z->uz_bucket_size);
5444 		if (db_pager_quit)
5445 			return;
5446 	}
5447 }
5448 #endif	/* DDB */
5449