xref: /freebsd/sys/vm/uma_core.c (revision da759cfa320d5076b075d15ff3f00ab3ba5634fd)
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 	/*
720 	 * Conditionally set the maximum number of items.
721 	 */
722 	zdom->uzd_nitems += bucket->ub_cnt;
723 	if (__predict_true(zdom->uzd_nitems < zone->uz_bucket_max)) {
724 		if (ws)
725 			zone_domain_imax_set(zdom, zdom->uzd_nitems);
726 		if (STAILQ_EMPTY(&zdom->uzd_buckets))
727 			zdom->uzd_seq = bucket->ub_seq;
728 		STAILQ_INSERT_TAIL(&zdom->uzd_buckets, bucket, ub_link);
729 		ZDOM_UNLOCK(zdom);
730 		return;
731 	}
732 	zdom->uzd_nitems -= bucket->ub_cnt;
733 	ZDOM_UNLOCK(zdom);
734 out:
735 	bucket_free(zone, bucket, udata);
736 }
737 
738 /* Pops an item out of a per-cpu cache bucket. */
739 static inline void *
740 cache_bucket_pop(uma_cache_t cache, uma_cache_bucket_t bucket)
741 {
742 	void *item;
743 
744 	CRITICAL_ASSERT(curthread);
745 
746 	bucket->ucb_cnt--;
747 	item = bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt];
748 #ifdef INVARIANTS
749 	bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = NULL;
750 	KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled."));
751 #endif
752 	cache->uc_allocs++;
753 
754 	return (item);
755 }
756 
757 /* Pushes an item into a per-cpu cache bucket. */
758 static inline void
759 cache_bucket_push(uma_cache_t cache, uma_cache_bucket_t bucket, void *item)
760 {
761 
762 	CRITICAL_ASSERT(curthread);
763 	KASSERT(bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] == NULL,
764 	    ("uma_zfree: Freeing to non free bucket index."));
765 
766 	bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = item;
767 	bucket->ucb_cnt++;
768 	cache->uc_frees++;
769 }
770 
771 /*
772  * Unload a UMA bucket from a per-cpu cache.
773  */
774 static inline uma_bucket_t
775 cache_bucket_unload(uma_cache_bucket_t bucket)
776 {
777 	uma_bucket_t b;
778 
779 	b = bucket->ucb_bucket;
780 	if (b != NULL) {
781 		MPASS(b->ub_entries == bucket->ucb_entries);
782 		b->ub_cnt = bucket->ucb_cnt;
783 		bucket->ucb_bucket = NULL;
784 		bucket->ucb_entries = bucket->ucb_cnt = 0;
785 	}
786 
787 	return (b);
788 }
789 
790 static inline uma_bucket_t
791 cache_bucket_unload_alloc(uma_cache_t cache)
792 {
793 
794 	return (cache_bucket_unload(&cache->uc_allocbucket));
795 }
796 
797 static inline uma_bucket_t
798 cache_bucket_unload_free(uma_cache_t cache)
799 {
800 
801 	return (cache_bucket_unload(&cache->uc_freebucket));
802 }
803 
804 static inline uma_bucket_t
805 cache_bucket_unload_cross(uma_cache_t cache)
806 {
807 
808 	return (cache_bucket_unload(&cache->uc_crossbucket));
809 }
810 
811 /*
812  * Load a bucket into a per-cpu cache bucket.
813  */
814 static inline void
815 cache_bucket_load(uma_cache_bucket_t bucket, uma_bucket_t b)
816 {
817 
818 	CRITICAL_ASSERT(curthread);
819 	MPASS(bucket->ucb_bucket == NULL);
820 	MPASS(b->ub_seq == SMR_SEQ_INVALID);
821 
822 	bucket->ucb_bucket = b;
823 	bucket->ucb_cnt = b->ub_cnt;
824 	bucket->ucb_entries = b->ub_entries;
825 }
826 
827 static inline void
828 cache_bucket_load_alloc(uma_cache_t cache, uma_bucket_t b)
829 {
830 
831 	cache_bucket_load(&cache->uc_allocbucket, b);
832 }
833 
834 static inline void
835 cache_bucket_load_free(uma_cache_t cache, uma_bucket_t b)
836 {
837 
838 	cache_bucket_load(&cache->uc_freebucket, b);
839 }
840 
841 #ifdef NUMA
842 static inline void
843 cache_bucket_load_cross(uma_cache_t cache, uma_bucket_t b)
844 {
845 
846 	cache_bucket_load(&cache->uc_crossbucket, b);
847 }
848 #endif
849 
850 /*
851  * Copy and preserve ucb_spare.
852  */
853 static inline void
854 cache_bucket_copy(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
855 {
856 
857 	b1->ucb_bucket = b2->ucb_bucket;
858 	b1->ucb_entries = b2->ucb_entries;
859 	b1->ucb_cnt = b2->ucb_cnt;
860 }
861 
862 /*
863  * Swap two cache buckets.
864  */
865 static inline void
866 cache_bucket_swap(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
867 {
868 	struct uma_cache_bucket b3;
869 
870 	CRITICAL_ASSERT(curthread);
871 
872 	cache_bucket_copy(&b3, b1);
873 	cache_bucket_copy(b1, b2);
874 	cache_bucket_copy(b2, &b3);
875 }
876 
877 /*
878  * Attempt to fetch a bucket from a zone on behalf of the current cpu cache.
879  */
880 static uma_bucket_t
881 cache_fetch_bucket(uma_zone_t zone, uma_cache_t cache, int domain)
882 {
883 	uma_zone_domain_t zdom;
884 	uma_bucket_t bucket;
885 
886 	/*
887 	 * Avoid the lock if possible.
888 	 */
889 	zdom = ZDOM_GET(zone, domain);
890 	if (zdom->uzd_nitems == 0)
891 		return (NULL);
892 
893 	if ((cache_uz_flags(cache) & UMA_ZONE_SMR) != 0 &&
894 	    !smr_poll(zone->uz_smr, zdom->uzd_seq, false))
895 		return (NULL);
896 
897 	/*
898 	 * Check the zone's cache of buckets.
899 	 */
900 	zdom = zone_domain_lock(zone, domain);
901 	if ((bucket = zone_fetch_bucket(zone, zdom, false)) != NULL) {
902 		KASSERT(bucket->ub_cnt != 0,
903 		    ("cache_fetch_bucket: Returning an empty bucket."));
904 		return (bucket);
905 	}
906 	ZDOM_UNLOCK(zdom);
907 
908 	return (NULL);
909 }
910 
911 static void
912 zone_log_warning(uma_zone_t zone)
913 {
914 	static const struct timeval warninterval = { 300, 0 };
915 
916 	if (!zone_warnings || zone->uz_warning == NULL)
917 		return;
918 
919 	if (ratecheck(&zone->uz_ratecheck, &warninterval))
920 		printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning);
921 }
922 
923 static inline void
924 zone_maxaction(uma_zone_t zone)
925 {
926 
927 	if (zone->uz_maxaction.ta_func != NULL)
928 		taskqueue_enqueue(taskqueue_thread, &zone->uz_maxaction);
929 }
930 
931 /*
932  * Routine called by timeout which is used to fire off some time interval
933  * based calculations.  (stats, hash size, etc.)
934  *
935  * Arguments:
936  *	arg   Unused
937  *
938  * Returns:
939  *	Nothing
940  */
941 static void
942 uma_timeout(void *unused)
943 {
944 	bucket_enable();
945 	zone_foreach(zone_timeout, NULL);
946 
947 	/* Reschedule this event */
948 	callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
949 }
950 
951 /*
952  * Update the working set size estimate for the zone's bucket cache.
953  * The constants chosen here are somewhat arbitrary.  With an update period of
954  * 20s (UMA_TIMEOUT), this estimate is dominated by zone activity over the
955  * last 100s.
956  */
957 static void
958 zone_domain_update_wss(uma_zone_domain_t zdom)
959 {
960 	long wss;
961 
962 	ZDOM_LOCK(zdom);
963 	MPASS(zdom->uzd_imax >= zdom->uzd_imin);
964 	wss = zdom->uzd_imax - zdom->uzd_imin;
965 	zdom->uzd_imax = zdom->uzd_imin = zdom->uzd_nitems;
966 	zdom->uzd_wss = (4 * wss + zdom->uzd_wss) / 5;
967 	ZDOM_UNLOCK(zdom);
968 }
969 
970 /*
971  * Routine to perform timeout driven calculations.  This expands the
972  * hashes and does per cpu statistics aggregation.
973  *
974  *  Returns nothing.
975  */
976 static void
977 zone_timeout(uma_zone_t zone, void *unused)
978 {
979 	uma_keg_t keg;
980 	u_int slabs, pages;
981 
982 	if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
983 		goto update_wss;
984 
985 	keg = zone->uz_keg;
986 
987 	/*
988 	 * Hash zones are non-numa by definition so the first domain
989 	 * is the only one present.
990 	 */
991 	KEG_LOCK(keg, 0);
992 	pages = keg->uk_domain[0].ud_pages;
993 
994 	/*
995 	 * Expand the keg hash table.
996 	 *
997 	 * This is done if the number of slabs is larger than the hash size.
998 	 * What I'm trying to do here is completely reduce collisions.  This
999 	 * may be a little aggressive.  Should I allow for two collisions max?
1000 	 */
1001 	if ((slabs = pages / keg->uk_ppera) > keg->uk_hash.uh_hashsize) {
1002 		struct uma_hash newhash;
1003 		struct uma_hash oldhash;
1004 		int ret;
1005 
1006 		/*
1007 		 * This is so involved because allocating and freeing
1008 		 * while the keg lock is held will lead to deadlock.
1009 		 * I have to do everything in stages and check for
1010 		 * races.
1011 		 */
1012 		KEG_UNLOCK(keg, 0);
1013 		ret = hash_alloc(&newhash, 1 << fls(slabs));
1014 		KEG_LOCK(keg, 0);
1015 		if (ret) {
1016 			if (hash_expand(&keg->uk_hash, &newhash)) {
1017 				oldhash = keg->uk_hash;
1018 				keg->uk_hash = newhash;
1019 			} else
1020 				oldhash = newhash;
1021 
1022 			KEG_UNLOCK(keg, 0);
1023 			hash_free(&oldhash);
1024 			goto update_wss;
1025 		}
1026 	}
1027 	KEG_UNLOCK(keg, 0);
1028 
1029 update_wss:
1030 	for (int i = 0; i < vm_ndomains; i++)
1031 		zone_domain_update_wss(ZDOM_GET(zone, i));
1032 }
1033 
1034 /*
1035  * Allocate and zero fill the next sized hash table from the appropriate
1036  * backing store.
1037  *
1038  * Arguments:
1039  *	hash  A new hash structure with the old hash size in uh_hashsize
1040  *
1041  * Returns:
1042  *	1 on success and 0 on failure.
1043  */
1044 static int
1045 hash_alloc(struct uma_hash *hash, u_int size)
1046 {
1047 	size_t alloc;
1048 
1049 	KASSERT(powerof2(size), ("hash size must be power of 2"));
1050 	if (size > UMA_HASH_SIZE_INIT)  {
1051 		hash->uh_hashsize = size;
1052 		alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
1053 		hash->uh_slab_hash = malloc(alloc, M_UMAHASH, M_NOWAIT);
1054 	} else {
1055 		alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
1056 		hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
1057 		    UMA_ANYDOMAIN, M_WAITOK);
1058 		hash->uh_hashsize = UMA_HASH_SIZE_INIT;
1059 	}
1060 	if (hash->uh_slab_hash) {
1061 		bzero(hash->uh_slab_hash, alloc);
1062 		hash->uh_hashmask = hash->uh_hashsize - 1;
1063 		return (1);
1064 	}
1065 
1066 	return (0);
1067 }
1068 
1069 /*
1070  * Expands the hash table for HASH zones.  This is done from zone_timeout
1071  * to reduce collisions.  This must not be done in the regular allocation
1072  * path, otherwise, we can recurse on the vm while allocating pages.
1073  *
1074  * Arguments:
1075  *	oldhash  The hash you want to expand
1076  *	newhash  The hash structure for the new table
1077  *
1078  * Returns:
1079  *	Nothing
1080  *
1081  * Discussion:
1082  */
1083 static int
1084 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
1085 {
1086 	uma_hash_slab_t slab;
1087 	u_int hval;
1088 	u_int idx;
1089 
1090 	if (!newhash->uh_slab_hash)
1091 		return (0);
1092 
1093 	if (oldhash->uh_hashsize >= newhash->uh_hashsize)
1094 		return (0);
1095 
1096 	/*
1097 	 * I need to investigate hash algorithms for resizing without a
1098 	 * full rehash.
1099 	 */
1100 
1101 	for (idx = 0; idx < oldhash->uh_hashsize; idx++)
1102 		while (!LIST_EMPTY(&oldhash->uh_slab_hash[idx])) {
1103 			slab = LIST_FIRST(&oldhash->uh_slab_hash[idx]);
1104 			LIST_REMOVE(slab, uhs_hlink);
1105 			hval = UMA_HASH(newhash, slab->uhs_data);
1106 			LIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
1107 			    slab, uhs_hlink);
1108 		}
1109 
1110 	return (1);
1111 }
1112 
1113 /*
1114  * Free the hash bucket to the appropriate backing store.
1115  *
1116  * Arguments:
1117  *	slab_hash  The hash bucket we're freeing
1118  *	hashsize   The number of entries in that hash bucket
1119  *
1120  * Returns:
1121  *	Nothing
1122  */
1123 static void
1124 hash_free(struct uma_hash *hash)
1125 {
1126 	if (hash->uh_slab_hash == NULL)
1127 		return;
1128 	if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
1129 		zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE);
1130 	else
1131 		free(hash->uh_slab_hash, M_UMAHASH);
1132 }
1133 
1134 /*
1135  * Frees all outstanding items in a bucket
1136  *
1137  * Arguments:
1138  *	zone   The zone to free to, must be unlocked.
1139  *	bucket The free/alloc bucket with items.
1140  *
1141  * Returns:
1142  *	Nothing
1143  */
1144 static void
1145 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
1146 {
1147 	int i;
1148 
1149 	if (bucket->ub_cnt == 0)
1150 		return;
1151 
1152 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
1153 	    bucket->ub_seq != SMR_SEQ_INVALID) {
1154 		smr_wait(zone->uz_smr, bucket->ub_seq);
1155 		bucket->ub_seq = SMR_SEQ_INVALID;
1156 		for (i = 0; i < bucket->ub_cnt; i++)
1157 			item_dtor(zone, bucket->ub_bucket[i],
1158 			    zone->uz_size, NULL, SKIP_NONE);
1159 	}
1160 	if (zone->uz_fini)
1161 		for (i = 0; i < bucket->ub_cnt; i++)
1162 			zone->uz_fini(bucket->ub_bucket[i], zone->uz_size);
1163 	zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt);
1164 	if (zone->uz_max_items > 0)
1165 		zone_free_limit(zone, bucket->ub_cnt);
1166 #ifdef INVARIANTS
1167 	bzero(bucket->ub_bucket, sizeof(void *) * bucket->ub_cnt);
1168 #endif
1169 	bucket->ub_cnt = 0;
1170 }
1171 
1172 /*
1173  * Drains the per cpu caches for a zone.
1174  *
1175  * NOTE: This may only be called while the zone is being torn down, and not
1176  * during normal operation.  This is necessary in order that we do not have
1177  * to migrate CPUs to drain the per-CPU caches.
1178  *
1179  * Arguments:
1180  *	zone     The zone to drain, must be unlocked.
1181  *
1182  * Returns:
1183  *	Nothing
1184  */
1185 static void
1186 cache_drain(uma_zone_t zone)
1187 {
1188 	uma_cache_t cache;
1189 	uma_bucket_t bucket;
1190 	smr_seq_t seq;
1191 	int cpu;
1192 
1193 	/*
1194 	 * XXX: It is safe to not lock the per-CPU caches, because we're
1195 	 * tearing down the zone anyway.  I.e., there will be no further use
1196 	 * of the caches at this point.
1197 	 *
1198 	 * XXX: It would good to be able to assert that the zone is being
1199 	 * torn down to prevent improper use of cache_drain().
1200 	 */
1201 	seq = SMR_SEQ_INVALID;
1202 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
1203 		seq = smr_advance(zone->uz_smr);
1204 	CPU_FOREACH(cpu) {
1205 		cache = &zone->uz_cpu[cpu];
1206 		bucket = cache_bucket_unload_alloc(cache);
1207 		if (bucket != NULL)
1208 			bucket_free(zone, bucket, NULL);
1209 		bucket = cache_bucket_unload_free(cache);
1210 		if (bucket != NULL) {
1211 			bucket->ub_seq = seq;
1212 			bucket_free(zone, bucket, NULL);
1213 		}
1214 		bucket = cache_bucket_unload_cross(cache);
1215 		if (bucket != NULL) {
1216 			bucket->ub_seq = seq;
1217 			bucket_free(zone, bucket, NULL);
1218 		}
1219 	}
1220 	bucket_cache_reclaim(zone, true);
1221 }
1222 
1223 static void
1224 cache_shrink(uma_zone_t zone, void *unused)
1225 {
1226 
1227 	if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1228 		return;
1229 
1230 	zone->uz_bucket_size =
1231 	    (zone->uz_bucket_size_min + zone->uz_bucket_size) / 2;
1232 }
1233 
1234 static void
1235 cache_drain_safe_cpu(uma_zone_t zone, void *unused)
1236 {
1237 	uma_cache_t cache;
1238 	uma_bucket_t b1, b2, b3;
1239 	int domain;
1240 
1241 	if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1242 		return;
1243 
1244 	b1 = b2 = b3 = NULL;
1245 	critical_enter();
1246 	cache = &zone->uz_cpu[curcpu];
1247 	domain = PCPU_GET(domain);
1248 	b1 = cache_bucket_unload_alloc(cache);
1249 
1250 	/*
1251 	 * Don't flush SMR zone buckets.  This leaves the zone without a
1252 	 * bucket and forces every free to synchronize().
1253 	 */
1254 	if ((zone->uz_flags & UMA_ZONE_SMR) == 0) {
1255 		b2 = cache_bucket_unload_free(cache);
1256 		b3 = cache_bucket_unload_cross(cache);
1257 	}
1258 	critical_exit();
1259 
1260 	if (b1 != NULL)
1261 		zone_free_bucket(zone, b1, NULL, domain, false);
1262 	if (b2 != NULL)
1263 		zone_free_bucket(zone, b2, NULL, domain, false);
1264 	if (b3 != NULL) {
1265 		/* Adjust the domain so it goes to zone_free_cross. */
1266 		domain = (domain + 1) % vm_ndomains;
1267 		zone_free_bucket(zone, b3, NULL, domain, false);
1268 	}
1269 }
1270 
1271 /*
1272  * Safely drain per-CPU caches of a zone(s) to alloc bucket.
1273  * This is an expensive call because it needs to bind to all CPUs
1274  * one by one and enter a critical section on each of them in order
1275  * to safely access their cache buckets.
1276  * Zone lock must not be held on call this function.
1277  */
1278 static void
1279 pcpu_cache_drain_safe(uma_zone_t zone)
1280 {
1281 	int cpu;
1282 
1283 	/*
1284 	 * Polite bucket sizes shrinking was not enough, shrink aggressively.
1285 	 */
1286 	if (zone)
1287 		cache_shrink(zone, NULL);
1288 	else
1289 		zone_foreach(cache_shrink, NULL);
1290 
1291 	CPU_FOREACH(cpu) {
1292 		thread_lock(curthread);
1293 		sched_bind(curthread, cpu);
1294 		thread_unlock(curthread);
1295 
1296 		if (zone)
1297 			cache_drain_safe_cpu(zone, NULL);
1298 		else
1299 			zone_foreach(cache_drain_safe_cpu, NULL);
1300 	}
1301 	thread_lock(curthread);
1302 	sched_unbind(curthread);
1303 	thread_unlock(curthread);
1304 }
1305 
1306 /*
1307  * Reclaim cached buckets from a zone.  All buckets are reclaimed if the caller
1308  * requested a drain, otherwise the per-domain caches are trimmed to either
1309  * estimated working set size.
1310  */
1311 static void
1312 bucket_cache_reclaim(uma_zone_t zone, bool drain)
1313 {
1314 	uma_zone_domain_t zdom;
1315 	uma_bucket_t bucket;
1316 	long target;
1317 	int i;
1318 
1319 	/*
1320 	 * Shrink the zone bucket size to ensure that the per-CPU caches
1321 	 * don't grow too large.
1322 	 */
1323 	if (zone->uz_bucket_size > zone->uz_bucket_size_min)
1324 		zone->uz_bucket_size--;
1325 
1326 	for (i = 0; i < vm_ndomains; i++) {
1327 		/*
1328 		 * The cross bucket is partially filled and not part of
1329 		 * the item count.  Reclaim it individually here.
1330 		 */
1331 		zdom = ZDOM_GET(zone, i);
1332 		if ((zone->uz_flags & UMA_ZONE_SMR) == 0 || drain) {
1333 			ZONE_CROSS_LOCK(zone);
1334 			bucket = zdom->uzd_cross;
1335 			zdom->uzd_cross = NULL;
1336 			ZONE_CROSS_UNLOCK(zone);
1337 			if (bucket != NULL)
1338 				bucket_free(zone, bucket, NULL);
1339 		}
1340 
1341 		/*
1342 		 * If we were asked to drain the zone, we are done only once
1343 		 * this bucket cache is empty.  Otherwise, we reclaim items in
1344 		 * excess of the zone's estimated working set size.  If the
1345 		 * difference nitems - imin is larger than the WSS estimate,
1346 		 * then the estimate will grow at the end of this interval and
1347 		 * we ignore the historical average.
1348 		 */
1349 		ZDOM_LOCK(zdom);
1350 		target = drain ? 0 : lmax(zdom->uzd_wss, zdom->uzd_nitems -
1351 		    zdom->uzd_imin);
1352 		while (zdom->uzd_nitems > target) {
1353 			bucket = zone_fetch_bucket(zone, zdom, true);
1354 			if (bucket == NULL)
1355 				break;
1356 			bucket_free(zone, bucket, NULL);
1357 			ZDOM_LOCK(zdom);
1358 		}
1359 		ZDOM_UNLOCK(zdom);
1360 	}
1361 }
1362 
1363 static void
1364 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start)
1365 {
1366 	uint8_t *mem;
1367 	int i;
1368 	uint8_t flags;
1369 
1370 	CTR4(KTR_UMA, "keg_free_slab keg %s(%p) slab %p, returning %d bytes",
1371 	    keg->uk_name, keg, slab, PAGE_SIZE * keg->uk_ppera);
1372 
1373 	mem = slab_data(slab, keg);
1374 	flags = slab->us_flags;
1375 	i = start;
1376 	if (keg->uk_fini != NULL) {
1377 		for (i--; i > -1; i--)
1378 #ifdef INVARIANTS
1379 		/*
1380 		 * trash_fini implies that dtor was trash_dtor. trash_fini
1381 		 * would check that memory hasn't been modified since free,
1382 		 * which executed trash_dtor.
1383 		 * That's why we need to run uma_dbg_kskip() check here,
1384 		 * albeit we don't make skip check for other init/fini
1385 		 * invocations.
1386 		 */
1387 		if (!uma_dbg_kskip(keg, slab_item(slab, keg, i)) ||
1388 		    keg->uk_fini != trash_fini)
1389 #endif
1390 			keg->uk_fini(slab_item(slab, keg, i), keg->uk_size);
1391 	}
1392 	if (keg->uk_flags & UMA_ZFLAG_OFFPAGE)
1393 		zone_free_item(slabzone(keg->uk_ipers), slab_tohashslab(slab),
1394 		    NULL, SKIP_NONE);
1395 	keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags);
1396 	uma_total_dec(PAGE_SIZE * keg->uk_ppera);
1397 }
1398 
1399 /*
1400  * Frees pages from a keg back to the system.  This is done on demand from
1401  * the pageout daemon.
1402  *
1403  * Returns nothing.
1404  */
1405 static void
1406 keg_drain(uma_keg_t keg)
1407 {
1408 	struct slabhead freeslabs;
1409 	uma_domain_t dom;
1410 	uma_slab_t slab, tmp;
1411 	int i, n;
1412 
1413 	if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL)
1414 		return;
1415 
1416 	for (i = 0; i < vm_ndomains; i++) {
1417 		CTR4(KTR_UMA, "keg_drain %s(%p) domain %d free items: %u",
1418 		    keg->uk_name, keg, i, dom->ud_free_items);
1419 		dom = &keg->uk_domain[i];
1420 		LIST_INIT(&freeslabs);
1421 
1422 		KEG_LOCK(keg, i);
1423 		if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0) {
1424 			LIST_FOREACH(slab, &dom->ud_free_slab, us_link)
1425 				UMA_HASH_REMOVE(&keg->uk_hash, slab);
1426 		}
1427 		n = dom->ud_free_slabs;
1428 		LIST_SWAP(&freeslabs, &dom->ud_free_slab, uma_slab, us_link);
1429 		dom->ud_free_slabs = 0;
1430 		dom->ud_free_items -= n * keg->uk_ipers;
1431 		dom->ud_pages -= n * keg->uk_ppera;
1432 		KEG_UNLOCK(keg, i);
1433 
1434 		LIST_FOREACH_SAFE(slab, &freeslabs, us_link, tmp)
1435 			keg_free_slab(keg, slab, keg->uk_ipers);
1436 	}
1437 }
1438 
1439 static void
1440 zone_reclaim(uma_zone_t zone, int waitok, bool drain)
1441 {
1442 
1443 	/*
1444 	 * Set draining to interlock with zone_dtor() so we can release our
1445 	 * locks as we go.  Only dtor() should do a WAITOK call since it
1446 	 * is the only call that knows the structure will still be available
1447 	 * when it wakes up.
1448 	 */
1449 	ZONE_LOCK(zone);
1450 	while (zone->uz_flags & UMA_ZFLAG_RECLAIMING) {
1451 		if (waitok == M_NOWAIT)
1452 			goto out;
1453 		msleep(zone, &ZDOM_GET(zone, 0)->uzd_lock, PVM, "zonedrain",
1454 		    1);
1455 	}
1456 	zone->uz_flags |= UMA_ZFLAG_RECLAIMING;
1457 	ZONE_UNLOCK(zone);
1458 	bucket_cache_reclaim(zone, drain);
1459 
1460 	/*
1461 	 * The DRAINING flag protects us from being freed while
1462 	 * we're running.  Normally the uma_rwlock would protect us but we
1463 	 * must be able to release and acquire the right lock for each keg.
1464 	 */
1465 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0)
1466 		keg_drain(zone->uz_keg);
1467 	ZONE_LOCK(zone);
1468 	zone->uz_flags &= ~UMA_ZFLAG_RECLAIMING;
1469 	wakeup(zone);
1470 out:
1471 	ZONE_UNLOCK(zone);
1472 }
1473 
1474 static void
1475 zone_drain(uma_zone_t zone, void *unused)
1476 {
1477 
1478 	zone_reclaim(zone, M_NOWAIT, true);
1479 }
1480 
1481 static void
1482 zone_trim(uma_zone_t zone, void *unused)
1483 {
1484 
1485 	zone_reclaim(zone, M_NOWAIT, false);
1486 }
1487 
1488 /*
1489  * Allocate a new slab for a keg and inserts it into the partial slab list.
1490  * The keg should be unlocked on entry.  If the allocation succeeds it will
1491  * be locked on return.
1492  *
1493  * Arguments:
1494  *	flags   Wait flags for the item initialization routine
1495  *	aflags  Wait flags for the slab allocation
1496  *
1497  * Returns:
1498  *	The slab that was allocated or NULL if there is no memory and the
1499  *	caller specified M_NOWAIT.
1500  */
1501 static uma_slab_t
1502 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int domain, int flags,
1503     int aflags)
1504 {
1505 	uma_domain_t dom;
1506 	uma_alloc allocf;
1507 	uma_slab_t slab;
1508 	unsigned long size;
1509 	uint8_t *mem;
1510 	uint8_t sflags;
1511 	int i;
1512 
1513 	KASSERT(domain >= 0 && domain < vm_ndomains,
1514 	    ("keg_alloc_slab: domain %d out of range", domain));
1515 
1516 	allocf = keg->uk_allocf;
1517 	slab = NULL;
1518 	mem = NULL;
1519 	if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) {
1520 		uma_hash_slab_t hslab;
1521 		hslab = zone_alloc_item(slabzone(keg->uk_ipers), NULL,
1522 		    domain, aflags);
1523 		if (hslab == NULL)
1524 			goto fail;
1525 		slab = &hslab->uhs_slab;
1526 	}
1527 
1528 	/*
1529 	 * This reproduces the old vm_zone behavior of zero filling pages the
1530 	 * first time they are added to a zone.
1531 	 *
1532 	 * Malloced items are zeroed in uma_zalloc.
1533 	 */
1534 
1535 	if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
1536 		aflags |= M_ZERO;
1537 	else
1538 		aflags &= ~M_ZERO;
1539 
1540 	if (keg->uk_flags & UMA_ZONE_NODUMP)
1541 		aflags |= M_NODUMP;
1542 
1543 	/* zone is passed for legacy reasons. */
1544 	size = keg->uk_ppera * PAGE_SIZE;
1545 	mem = allocf(zone, size, domain, &sflags, aflags);
1546 	if (mem == NULL) {
1547 		if (keg->uk_flags & UMA_ZFLAG_OFFPAGE)
1548 			zone_free_item(slabzone(keg->uk_ipers),
1549 			    slab_tohashslab(slab), NULL, SKIP_NONE);
1550 		goto fail;
1551 	}
1552 	uma_total_inc(size);
1553 
1554 	/* For HASH zones all pages go to the same uma_domain. */
1555 	if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
1556 		domain = 0;
1557 
1558 	/* Point the slab into the allocated memory */
1559 	if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE))
1560 		slab = (uma_slab_t )(mem + keg->uk_pgoff);
1561 	else
1562 		slab_tohashslab(slab)->uhs_data = mem;
1563 
1564 	if (keg->uk_flags & UMA_ZFLAG_VTOSLAB)
1565 		for (i = 0; i < keg->uk_ppera; i++)
1566 			vsetzoneslab((vm_offset_t)mem + (i * PAGE_SIZE),
1567 			    zone, slab);
1568 
1569 	slab->us_freecount = keg->uk_ipers;
1570 	slab->us_flags = sflags;
1571 	slab->us_domain = domain;
1572 
1573 	BIT_FILL(keg->uk_ipers, &slab->us_free);
1574 #ifdef INVARIANTS
1575 	BIT_ZERO(keg->uk_ipers, slab_dbg_bits(slab, keg));
1576 #endif
1577 
1578 	if (keg->uk_init != NULL) {
1579 		for (i = 0; i < keg->uk_ipers; i++)
1580 			if (keg->uk_init(slab_item(slab, keg, i),
1581 			    keg->uk_size, flags) != 0)
1582 				break;
1583 		if (i != keg->uk_ipers) {
1584 			keg_free_slab(keg, slab, i);
1585 			goto fail;
1586 		}
1587 	}
1588 	KEG_LOCK(keg, domain);
1589 
1590 	CTR3(KTR_UMA, "keg_alloc_slab: allocated slab %p for %s(%p)",
1591 	    slab, keg->uk_name, keg);
1592 
1593 	if (keg->uk_flags & UMA_ZFLAG_HASH)
1594 		UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
1595 
1596 	/*
1597 	 * If we got a slab here it's safe to mark it partially used
1598 	 * and return.  We assume that the caller is going to remove
1599 	 * at least one item.
1600 	 */
1601 	dom = &keg->uk_domain[domain];
1602 	LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
1603 	dom->ud_pages += keg->uk_ppera;
1604 	dom->ud_free_items += keg->uk_ipers;
1605 
1606 	return (slab);
1607 
1608 fail:
1609 	return (NULL);
1610 }
1611 
1612 /*
1613  * This function is intended to be used early on in place of page_alloc() so
1614  * that we may use the boot time page cache to satisfy allocations before
1615  * the VM is ready.
1616  */
1617 static void *
1618 startup_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1619     int wait)
1620 {
1621 	vm_paddr_t pa;
1622 	vm_page_t m;
1623 	void *mem;
1624 	int pages;
1625 	int i;
1626 
1627 	pages = howmany(bytes, PAGE_SIZE);
1628 	KASSERT(pages > 0, ("%s can't reserve 0 pages", __func__));
1629 
1630 	*pflag = UMA_SLAB_BOOT;
1631 	m = vm_page_alloc_contig_domain(NULL, 0, domain,
1632 	    malloc2vm_flags(wait) | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED, pages,
1633 	    (vm_paddr_t)0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT);
1634 	if (m == NULL)
1635 		return (NULL);
1636 
1637 	pa = VM_PAGE_TO_PHYS(m);
1638 	for (i = 0; i < pages; i++, pa += PAGE_SIZE) {
1639 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
1640     defined(__riscv) || defined(__powerpc64__)
1641 		if ((wait & M_NODUMP) == 0)
1642 			dump_add_page(pa);
1643 #endif
1644 	}
1645 	/* Allocate KVA and indirectly advance bootmem. */
1646 	mem = (void *)pmap_map(&bootmem, m->phys_addr,
1647 	    m->phys_addr + (pages * PAGE_SIZE), VM_PROT_READ | VM_PROT_WRITE);
1648         if ((wait & M_ZERO) != 0)
1649                 bzero(mem, pages * PAGE_SIZE);
1650 
1651         return (mem);
1652 }
1653 
1654 static void
1655 startup_free(void *mem, vm_size_t bytes)
1656 {
1657 	vm_offset_t va;
1658 	vm_page_t m;
1659 
1660 	va = (vm_offset_t)mem;
1661 	m = PHYS_TO_VM_PAGE(pmap_kextract(va));
1662 	pmap_remove(kernel_pmap, va, va + bytes);
1663 	for (; bytes != 0; bytes -= PAGE_SIZE, m++) {
1664 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
1665     defined(__riscv) || defined(__powerpc64__)
1666 		dump_drop_page(VM_PAGE_TO_PHYS(m));
1667 #endif
1668 		vm_page_unwire_noq(m);
1669 		vm_page_free(m);
1670 	}
1671 }
1672 
1673 /*
1674  * Allocates a number of pages from the system
1675  *
1676  * Arguments:
1677  *	bytes  The number of bytes requested
1678  *	wait  Shall we wait?
1679  *
1680  * Returns:
1681  *	A pointer to the alloced memory or possibly
1682  *	NULL if M_NOWAIT is set.
1683  */
1684 static void *
1685 page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1686     int wait)
1687 {
1688 	void *p;	/* Returned page */
1689 
1690 	*pflag = UMA_SLAB_KERNEL;
1691 	p = (void *)kmem_malloc_domainset(DOMAINSET_FIXED(domain), bytes, wait);
1692 
1693 	return (p);
1694 }
1695 
1696 static void *
1697 pcpu_page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1698     int wait)
1699 {
1700 	struct pglist alloctail;
1701 	vm_offset_t addr, zkva;
1702 	int cpu, flags;
1703 	vm_page_t p, p_next;
1704 #ifdef NUMA
1705 	struct pcpu *pc;
1706 #endif
1707 
1708 	MPASS(bytes == (mp_maxid + 1) * PAGE_SIZE);
1709 
1710 	TAILQ_INIT(&alloctail);
1711 	flags = VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | VM_ALLOC_NOOBJ |
1712 	    malloc2vm_flags(wait);
1713 	*pflag = UMA_SLAB_KERNEL;
1714 	for (cpu = 0; cpu <= mp_maxid; cpu++) {
1715 		if (CPU_ABSENT(cpu)) {
1716 			p = vm_page_alloc(NULL, 0, flags);
1717 		} else {
1718 #ifndef NUMA
1719 			p = vm_page_alloc(NULL, 0, flags);
1720 #else
1721 			pc = pcpu_find(cpu);
1722 			if (__predict_false(VM_DOMAIN_EMPTY(pc->pc_domain)))
1723 				p = NULL;
1724 			else
1725 				p = vm_page_alloc_domain(NULL, 0,
1726 				    pc->pc_domain, flags);
1727 			if (__predict_false(p == NULL))
1728 				p = vm_page_alloc(NULL, 0, flags);
1729 #endif
1730 		}
1731 		if (__predict_false(p == NULL))
1732 			goto fail;
1733 		TAILQ_INSERT_TAIL(&alloctail, p, listq);
1734 	}
1735 	if ((addr = kva_alloc(bytes)) == 0)
1736 		goto fail;
1737 	zkva = addr;
1738 	TAILQ_FOREACH(p, &alloctail, listq) {
1739 		pmap_qenter(zkva, &p, 1);
1740 		zkva += PAGE_SIZE;
1741 	}
1742 	return ((void*)addr);
1743 fail:
1744 	TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1745 		vm_page_unwire_noq(p);
1746 		vm_page_free(p);
1747 	}
1748 	return (NULL);
1749 }
1750 
1751 /*
1752  * Allocates a number of pages from within an object
1753  *
1754  * Arguments:
1755  *	bytes  The number of bytes requested
1756  *	wait   Shall we wait?
1757  *
1758  * Returns:
1759  *	A pointer to the alloced memory or possibly
1760  *	NULL if M_NOWAIT is set.
1761  */
1762 static void *
1763 noobj_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags,
1764     int wait)
1765 {
1766 	TAILQ_HEAD(, vm_page) alloctail;
1767 	u_long npages;
1768 	vm_offset_t retkva, zkva;
1769 	vm_page_t p, p_next;
1770 	uma_keg_t keg;
1771 
1772 	TAILQ_INIT(&alloctail);
1773 	keg = zone->uz_keg;
1774 
1775 	npages = howmany(bytes, PAGE_SIZE);
1776 	while (npages > 0) {
1777 		p = vm_page_alloc_domain(NULL, 0, domain, VM_ALLOC_INTERRUPT |
1778 		    VM_ALLOC_WIRED | VM_ALLOC_NOOBJ |
1779 		    ((wait & M_WAITOK) != 0 ? VM_ALLOC_WAITOK :
1780 		    VM_ALLOC_NOWAIT));
1781 		if (p != NULL) {
1782 			/*
1783 			 * Since the page does not belong to an object, its
1784 			 * listq is unused.
1785 			 */
1786 			TAILQ_INSERT_TAIL(&alloctail, p, listq);
1787 			npages--;
1788 			continue;
1789 		}
1790 		/*
1791 		 * Page allocation failed, free intermediate pages and
1792 		 * exit.
1793 		 */
1794 		TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1795 			vm_page_unwire_noq(p);
1796 			vm_page_free(p);
1797 		}
1798 		return (NULL);
1799 	}
1800 	*flags = UMA_SLAB_PRIV;
1801 	zkva = keg->uk_kva +
1802 	    atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
1803 	retkva = zkva;
1804 	TAILQ_FOREACH(p, &alloctail, listq) {
1805 		pmap_qenter(zkva, &p, 1);
1806 		zkva += PAGE_SIZE;
1807 	}
1808 
1809 	return ((void *)retkva);
1810 }
1811 
1812 /*
1813  * Allocate physically contiguous pages.
1814  */
1815 static void *
1816 contig_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1817     int wait)
1818 {
1819 
1820 	*pflag = UMA_SLAB_KERNEL;
1821 	return ((void *)kmem_alloc_contig_domainset(DOMAINSET_FIXED(domain),
1822 	    bytes, wait, 0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT));
1823 }
1824 
1825 /*
1826  * Frees a number of pages to the system
1827  *
1828  * Arguments:
1829  *	mem   A pointer to the memory to be freed
1830  *	size  The size of the memory being freed
1831  *	flags The original p->us_flags field
1832  *
1833  * Returns:
1834  *	Nothing
1835  */
1836 static void
1837 page_free(void *mem, vm_size_t size, uint8_t flags)
1838 {
1839 
1840 	if ((flags & UMA_SLAB_BOOT) != 0) {
1841 		startup_free(mem, size);
1842 		return;
1843 	}
1844 
1845 	KASSERT((flags & UMA_SLAB_KERNEL) != 0,
1846 	    ("UMA: page_free used with invalid flags %x", flags));
1847 
1848 	kmem_free((vm_offset_t)mem, size);
1849 }
1850 
1851 /*
1852  * Frees pcpu zone allocations
1853  *
1854  * Arguments:
1855  *	mem   A pointer to the memory to be freed
1856  *	size  The size of the memory being freed
1857  *	flags The original p->us_flags field
1858  *
1859  * Returns:
1860  *	Nothing
1861  */
1862 static void
1863 pcpu_page_free(void *mem, vm_size_t size, uint8_t flags)
1864 {
1865 	vm_offset_t sva, curva;
1866 	vm_paddr_t paddr;
1867 	vm_page_t m;
1868 
1869 	MPASS(size == (mp_maxid+1)*PAGE_SIZE);
1870 
1871 	if ((flags & UMA_SLAB_BOOT) != 0) {
1872 		startup_free(mem, size);
1873 		return;
1874 	}
1875 
1876 	sva = (vm_offset_t)mem;
1877 	for (curva = sva; curva < sva + size; curva += PAGE_SIZE) {
1878 		paddr = pmap_kextract(curva);
1879 		m = PHYS_TO_VM_PAGE(paddr);
1880 		vm_page_unwire_noq(m);
1881 		vm_page_free(m);
1882 	}
1883 	pmap_qremove(sva, size >> PAGE_SHIFT);
1884 	kva_free(sva, size);
1885 }
1886 
1887 
1888 /*
1889  * Zero fill initializer
1890  *
1891  * Arguments/Returns follow uma_init specifications
1892  */
1893 static int
1894 zero_init(void *mem, int size, int flags)
1895 {
1896 	bzero(mem, size);
1897 	return (0);
1898 }
1899 
1900 #ifdef INVARIANTS
1901 struct noslabbits *
1902 slab_dbg_bits(uma_slab_t slab, uma_keg_t keg)
1903 {
1904 
1905 	return ((void *)((char *)&slab->us_free + BITSET_SIZE(keg->uk_ipers)));
1906 }
1907 #endif
1908 
1909 /*
1910  * Actual size of embedded struct slab (!OFFPAGE).
1911  */
1912 size_t
1913 slab_sizeof(int nitems)
1914 {
1915 	size_t s;
1916 
1917 	s = sizeof(struct uma_slab) + BITSET_SIZE(nitems) * SLAB_BITSETS;
1918 	return (roundup(s, UMA_ALIGN_PTR + 1));
1919 }
1920 
1921 /*
1922  * Size of memory for embedded slabs (!OFFPAGE).
1923  */
1924 size_t
1925 slab_space(int nitems)
1926 {
1927 	return (UMA_SLAB_SIZE - slab_sizeof(nitems));
1928 }
1929 
1930 #define	UMA_FIXPT_SHIFT	31
1931 #define	UMA_FRAC_FIXPT(n, d)						\
1932 	((uint32_t)(((uint64_t)(n) << UMA_FIXPT_SHIFT) / (d)))
1933 #define	UMA_FIXPT_PCT(f)						\
1934 	((u_int)(((uint64_t)100 * (f)) >> UMA_FIXPT_SHIFT))
1935 #define	UMA_PCT_FIXPT(pct)	UMA_FRAC_FIXPT((pct), 100)
1936 #define	UMA_MIN_EFF	UMA_PCT_FIXPT(100 - UMA_MAX_WASTE)
1937 
1938 /*
1939  * Compute the number of items that will fit in a slab.  If hdr is true, the
1940  * item count may be limited to provide space in the slab for an inline slab
1941  * header.  Otherwise, all slab space will be provided for item storage.
1942  */
1943 static u_int
1944 slab_ipers_hdr(u_int size, u_int rsize, u_int slabsize, bool hdr)
1945 {
1946 	u_int ipers;
1947 	u_int padpi;
1948 
1949 	/* The padding between items is not needed after the last item. */
1950 	padpi = rsize - size;
1951 
1952 	if (hdr) {
1953 		/*
1954 		 * Start with the maximum item count and remove items until
1955 		 * the slab header first alongside the allocatable memory.
1956 		 */
1957 		for (ipers = MIN(SLAB_MAX_SETSIZE,
1958 		    (slabsize + padpi - slab_sizeof(1)) / rsize);
1959 		    ipers > 0 &&
1960 		    ipers * rsize - padpi + slab_sizeof(ipers) > slabsize;
1961 		    ipers--)
1962 			continue;
1963 	} else {
1964 		ipers = MIN((slabsize + padpi) / rsize, SLAB_MAX_SETSIZE);
1965 	}
1966 
1967 	return (ipers);
1968 }
1969 
1970 /*
1971  * Compute the number of items that will fit in a slab for a startup zone.
1972  */
1973 int
1974 slab_ipers(size_t size, int align)
1975 {
1976 	int rsize;
1977 
1978 	rsize = roundup(size, align + 1); /* Assume no CACHESPREAD */
1979 	return (slab_ipers_hdr(size, rsize, UMA_SLAB_SIZE, true));
1980 }
1981 
1982 struct keg_layout_result {
1983 	u_int format;
1984 	u_int slabsize;
1985 	u_int ipers;
1986 	u_int eff;
1987 };
1988 
1989 static void
1990 keg_layout_one(uma_keg_t keg, u_int rsize, u_int slabsize, u_int fmt,
1991     struct keg_layout_result *kl)
1992 {
1993 	u_int total;
1994 
1995 	kl->format = fmt;
1996 	kl->slabsize = slabsize;
1997 
1998 	/* Handle INTERNAL as inline with an extra page. */
1999 	if ((fmt & UMA_ZFLAG_INTERNAL) != 0) {
2000 		kl->format &= ~UMA_ZFLAG_INTERNAL;
2001 		kl->slabsize += PAGE_SIZE;
2002 	}
2003 
2004 	kl->ipers = slab_ipers_hdr(keg->uk_size, rsize, kl->slabsize,
2005 	    (fmt & UMA_ZFLAG_OFFPAGE) == 0);
2006 
2007 	/* Account for memory used by an offpage slab header. */
2008 	total = kl->slabsize;
2009 	if ((fmt & UMA_ZFLAG_OFFPAGE) != 0)
2010 		total += slabzone(kl->ipers)->uz_keg->uk_rsize;
2011 
2012 	kl->eff = UMA_FRAC_FIXPT(kl->ipers * rsize, total);
2013 }
2014 
2015 /*
2016  * Determine the format of a uma keg.  This determines where the slab header
2017  * will be placed (inline or offpage) and calculates ipers, rsize, and ppera.
2018  *
2019  * Arguments
2020  *	keg  The zone we should initialize
2021  *
2022  * Returns
2023  *	Nothing
2024  */
2025 static void
2026 keg_layout(uma_keg_t keg)
2027 {
2028 	struct keg_layout_result kl = {}, kl_tmp;
2029 	u_int fmts[2];
2030 	u_int alignsize;
2031 	u_int nfmt;
2032 	u_int pages;
2033 	u_int rsize;
2034 	u_int slabsize;
2035 	u_int i, j;
2036 
2037 	KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
2038 	    (keg->uk_size <= UMA_PCPU_ALLOC_SIZE &&
2039 	     (keg->uk_flags & UMA_ZONE_CACHESPREAD) == 0),
2040 	    ("%s: cannot configure for PCPU: keg=%s, size=%u, flags=0x%b",
2041 	     __func__, keg->uk_name, keg->uk_size, keg->uk_flags,
2042 	     PRINT_UMA_ZFLAGS));
2043 	KASSERT((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) == 0 ||
2044 	    (keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0,
2045 	    ("%s: incompatible flags 0x%b", __func__, keg->uk_flags,
2046 	     PRINT_UMA_ZFLAGS));
2047 
2048 	alignsize = keg->uk_align + 1;
2049 
2050 	/*
2051 	 * Calculate the size of each allocation (rsize) according to
2052 	 * alignment.  If the requested size is smaller than we have
2053 	 * allocation bits for we round it up.
2054 	 */
2055 	rsize = MAX(keg->uk_size, UMA_SMALLEST_UNIT);
2056 	rsize = roundup2(rsize, alignsize);
2057 
2058 	if ((keg->uk_flags & UMA_ZONE_CACHESPREAD) != 0) {
2059 		/*
2060 		 * We want one item to start on every align boundary in a page.
2061 		 * To do this we will span pages.  We will also extend the item
2062 		 * by the size of align if it is an even multiple of align.
2063 		 * Otherwise, it would fall on the same boundary every time.
2064 		 */
2065 		if ((rsize & alignsize) == 0)
2066 			rsize += alignsize;
2067 		slabsize = rsize * (PAGE_SIZE / alignsize);
2068 		slabsize = MIN(slabsize, rsize * SLAB_MAX_SETSIZE);
2069 		slabsize = MIN(slabsize, UMA_CACHESPREAD_MAX_SIZE);
2070 		slabsize = round_page(slabsize);
2071 	} else {
2072 		/*
2073 		 * Start with a slab size of as many pages as it takes to
2074 		 * represent a single item.  We will try to fit as many
2075 		 * additional items into the slab as possible.
2076 		 */
2077 		slabsize = round_page(keg->uk_size);
2078 	}
2079 
2080 	/* Build a list of all of the available formats for this keg. */
2081 	nfmt = 0;
2082 
2083 	/* Evaluate an inline slab layout. */
2084 	if ((keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0)
2085 		fmts[nfmt++] = 0;
2086 
2087 	/* TODO: vm_page-embedded slab. */
2088 
2089 	/*
2090 	 * We can't do OFFPAGE if we're internal or if we've been
2091 	 * asked to not go to the VM for buckets.  If we do this we
2092 	 * may end up going to the VM for slabs which we do not want
2093 	 * to do if we're UMA_ZONE_VM, which clearly forbids it.
2094 	 * In those cases, evaluate a pseudo-format called INTERNAL
2095 	 * which has an inline slab header and one extra page to
2096 	 * guarantee that it fits.
2097 	 *
2098 	 * Otherwise, see if using an OFFPAGE slab will improve our
2099 	 * efficiency.
2100 	 */
2101 	if ((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) != 0)
2102 		fmts[nfmt++] = UMA_ZFLAG_INTERNAL;
2103 	else
2104 		fmts[nfmt++] = UMA_ZFLAG_OFFPAGE;
2105 
2106 	/*
2107 	 * Choose a slab size and format which satisfy the minimum efficiency.
2108 	 * Prefer the smallest slab size that meets the constraints.
2109 	 *
2110 	 * Start with a minimum slab size, to accommodate CACHESPREAD.  Then,
2111 	 * for small items (up to PAGE_SIZE), the iteration increment is one
2112 	 * page; and for large items, the increment is one item.
2113 	 */
2114 	i = (slabsize + rsize - keg->uk_size) / MAX(PAGE_SIZE, rsize);
2115 	KASSERT(i >= 1, ("keg %s(%p) flags=0x%b slabsize=%u, rsize=%u, i=%u",
2116 	    keg->uk_name, keg, keg->uk_flags, PRINT_UMA_ZFLAGS, slabsize,
2117 	    rsize, i));
2118 	for ( ; ; i++) {
2119 		slabsize = (rsize <= PAGE_SIZE) ? ptoa(i) :
2120 		    round_page(rsize * (i - 1) + keg->uk_size);
2121 
2122 		for (j = 0; j < nfmt; j++) {
2123 			/* Only if we have no viable format yet. */
2124 			if ((fmts[j] & UMA_ZFLAG_INTERNAL) != 0 &&
2125 			    kl.ipers > 0)
2126 				continue;
2127 
2128 			keg_layout_one(keg, rsize, slabsize, fmts[j], &kl_tmp);
2129 			if (kl_tmp.eff <= kl.eff)
2130 				continue;
2131 
2132 			kl = kl_tmp;
2133 
2134 			CTR6(KTR_UMA, "keg %s layout: format %#x "
2135 			    "(ipers %u * rsize %u) / slabsize %#x = %u%% eff",
2136 			    keg->uk_name, kl.format, kl.ipers, rsize,
2137 			    kl.slabsize, UMA_FIXPT_PCT(kl.eff));
2138 
2139 			/* Stop when we reach the minimum efficiency. */
2140 			if (kl.eff >= UMA_MIN_EFF)
2141 				break;
2142 		}
2143 
2144 		if (kl.eff >= UMA_MIN_EFF || !multipage_slabs ||
2145 		    slabsize >= SLAB_MAX_SETSIZE * rsize ||
2146 		    (keg->uk_flags & (UMA_ZONE_PCPU | UMA_ZONE_CONTIG)) != 0)
2147 			break;
2148 	}
2149 
2150 	pages = atop(kl.slabsize);
2151 	if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
2152 		pages *= mp_maxid + 1;
2153 
2154 	keg->uk_rsize = rsize;
2155 	keg->uk_ipers = kl.ipers;
2156 	keg->uk_ppera = pages;
2157 	keg->uk_flags |= kl.format;
2158 
2159 	/*
2160 	 * How do we find the slab header if it is offpage or if not all item
2161 	 * start addresses are in the same page?  We could solve the latter
2162 	 * case with vaddr alignment, but we don't.
2163 	 */
2164 	if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0 ||
2165 	    (keg->uk_ipers - 1) * rsize >= PAGE_SIZE) {
2166 		if ((keg->uk_flags & UMA_ZONE_NOTPAGE) != 0)
2167 			keg->uk_flags |= UMA_ZFLAG_HASH;
2168 		else
2169 			keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2170 	}
2171 
2172 	CTR6(KTR_UMA, "%s: keg=%s, flags=%#x, rsize=%u, ipers=%u, ppera=%u",
2173 	    __func__, keg->uk_name, keg->uk_flags, rsize, keg->uk_ipers,
2174 	    pages);
2175 	KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_MAX_SETSIZE,
2176 	    ("%s: keg=%s, flags=0x%b, rsize=%u, ipers=%u, ppera=%u", __func__,
2177 	     keg->uk_name, keg->uk_flags, PRINT_UMA_ZFLAGS, rsize,
2178 	     keg->uk_ipers, pages));
2179 }
2180 
2181 /*
2182  * Keg header ctor.  This initializes all fields, locks, etc.  And inserts
2183  * the keg onto the global keg list.
2184  *
2185  * Arguments/Returns follow uma_ctor specifications
2186  *	udata  Actually uma_kctor_args
2187  */
2188 static int
2189 keg_ctor(void *mem, int size, void *udata, int flags)
2190 {
2191 	struct uma_kctor_args *arg = udata;
2192 	uma_keg_t keg = mem;
2193 	uma_zone_t zone;
2194 	int i;
2195 
2196 	bzero(keg, size);
2197 	keg->uk_size = arg->size;
2198 	keg->uk_init = arg->uminit;
2199 	keg->uk_fini = arg->fini;
2200 	keg->uk_align = arg->align;
2201 	keg->uk_reserve = 0;
2202 	keg->uk_flags = arg->flags;
2203 
2204 	/*
2205 	 * We use a global round-robin policy by default.  Zones with
2206 	 * UMA_ZONE_FIRSTTOUCH set will use first-touch instead, in which
2207 	 * case the iterator is never run.
2208 	 */
2209 	keg->uk_dr.dr_policy = DOMAINSET_RR();
2210 	keg->uk_dr.dr_iter = 0;
2211 
2212 	/*
2213 	 * The master zone is passed to us at keg-creation time.
2214 	 */
2215 	zone = arg->zone;
2216 	keg->uk_name = zone->uz_name;
2217 
2218 	if (arg->flags & UMA_ZONE_ZINIT)
2219 		keg->uk_init = zero_init;
2220 
2221 	if (arg->flags & UMA_ZONE_MALLOC)
2222 		keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2223 
2224 #ifndef SMP
2225 	keg->uk_flags &= ~UMA_ZONE_PCPU;
2226 #endif
2227 
2228 	keg_layout(keg);
2229 
2230 	/*
2231 	 * Use a first-touch NUMA policy for kegs that pmap_extract() will
2232 	 * work on.  Use round-robin for everything else.
2233 	 *
2234 	 * Zones may override the default by specifying either.
2235 	 */
2236 #ifdef NUMA
2237 	if ((keg->uk_flags &
2238 	    (UMA_ZONE_ROUNDROBIN | UMA_ZFLAG_CACHE | UMA_ZONE_NOTPAGE)) == 0)
2239 		keg->uk_flags |= UMA_ZONE_FIRSTTOUCH;
2240 	else if ((keg->uk_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2241 		keg->uk_flags |= UMA_ZONE_ROUNDROBIN;
2242 #endif
2243 
2244 	/*
2245 	 * If we haven't booted yet we need allocations to go through the
2246 	 * startup cache until the vm is ready.
2247 	 */
2248 #ifdef UMA_MD_SMALL_ALLOC
2249 	if (keg->uk_ppera == 1)
2250 		keg->uk_allocf = uma_small_alloc;
2251 	else
2252 #endif
2253 	if (booted < BOOT_KVA)
2254 		keg->uk_allocf = startup_alloc;
2255 	else if (keg->uk_flags & UMA_ZONE_PCPU)
2256 		keg->uk_allocf = pcpu_page_alloc;
2257 	else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 && keg->uk_ppera > 1)
2258 		keg->uk_allocf = contig_alloc;
2259 	else
2260 		keg->uk_allocf = page_alloc;
2261 #ifdef UMA_MD_SMALL_ALLOC
2262 	if (keg->uk_ppera == 1)
2263 		keg->uk_freef = uma_small_free;
2264 	else
2265 #endif
2266 	if (keg->uk_flags & UMA_ZONE_PCPU)
2267 		keg->uk_freef = pcpu_page_free;
2268 	else
2269 		keg->uk_freef = page_free;
2270 
2271 	/*
2272 	 * Initialize keg's locks.
2273 	 */
2274 	for (i = 0; i < vm_ndomains; i++)
2275 		KEG_LOCK_INIT(keg, i, (arg->flags & UMA_ZONE_MTXCLASS));
2276 
2277 	/*
2278 	 * If we're putting the slab header in the actual page we need to
2279 	 * figure out where in each page it goes.  See slab_sizeof
2280 	 * definition.
2281 	 */
2282 	if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE)) {
2283 		size_t shsize;
2284 
2285 		shsize = slab_sizeof(keg->uk_ipers);
2286 		keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - shsize;
2287 		/*
2288 		 * The only way the following is possible is if with our
2289 		 * UMA_ALIGN_PTR adjustments we are now bigger than
2290 		 * UMA_SLAB_SIZE.  I haven't checked whether this is
2291 		 * mathematically possible for all cases, so we make
2292 		 * sure here anyway.
2293 		 */
2294 		KASSERT(keg->uk_pgoff + shsize <= PAGE_SIZE * keg->uk_ppera,
2295 		    ("zone %s ipers %d rsize %d size %d slab won't fit",
2296 		    zone->uz_name, keg->uk_ipers, keg->uk_rsize, keg->uk_size));
2297 	}
2298 
2299 	if (keg->uk_flags & UMA_ZFLAG_HASH)
2300 		hash_alloc(&keg->uk_hash, 0);
2301 
2302 	CTR3(KTR_UMA, "keg_ctor %p zone %s(%p)", keg, zone->uz_name, zone);
2303 
2304 	LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
2305 
2306 	rw_wlock(&uma_rwlock);
2307 	LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
2308 	rw_wunlock(&uma_rwlock);
2309 	return (0);
2310 }
2311 
2312 static void
2313 zone_kva_available(uma_zone_t zone, void *unused)
2314 {
2315 	uma_keg_t keg;
2316 
2317 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
2318 		return;
2319 	KEG_GET(zone, keg);
2320 
2321 	if (keg->uk_allocf == startup_alloc) {
2322 		/* Switch to the real allocator. */
2323 		if (keg->uk_flags & UMA_ZONE_PCPU)
2324 			keg->uk_allocf = pcpu_page_alloc;
2325 		else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 &&
2326 		    keg->uk_ppera > 1)
2327 			keg->uk_allocf = contig_alloc;
2328 		else
2329 			keg->uk_allocf = page_alloc;
2330 	}
2331 }
2332 
2333 static void
2334 zone_alloc_counters(uma_zone_t zone, void *unused)
2335 {
2336 
2337 	zone->uz_allocs = counter_u64_alloc(M_WAITOK);
2338 	zone->uz_frees = counter_u64_alloc(M_WAITOK);
2339 	zone->uz_fails = counter_u64_alloc(M_WAITOK);
2340 	zone->uz_xdomain = counter_u64_alloc(M_WAITOK);
2341 }
2342 
2343 static void
2344 zone_alloc_sysctl(uma_zone_t zone, void *unused)
2345 {
2346 	uma_zone_domain_t zdom;
2347 	uma_domain_t dom;
2348 	uma_keg_t keg;
2349 	struct sysctl_oid *oid, *domainoid;
2350 	int domains, i, cnt;
2351 	static const char *nokeg = "cache zone";
2352 	char *c;
2353 
2354 	/*
2355 	 * Make a sysctl safe copy of the zone name by removing
2356 	 * any special characters and handling dups by appending
2357 	 * an index.
2358 	 */
2359 	if (zone->uz_namecnt != 0) {
2360 		/* Count the number of decimal digits and '_' separator. */
2361 		for (i = 1, cnt = zone->uz_namecnt; cnt != 0; i++)
2362 			cnt /= 10;
2363 		zone->uz_ctlname = malloc(strlen(zone->uz_name) + i + 1,
2364 		    M_UMA, M_WAITOK);
2365 		sprintf(zone->uz_ctlname, "%s_%d", zone->uz_name,
2366 		    zone->uz_namecnt);
2367 	} else
2368 		zone->uz_ctlname = strdup(zone->uz_name, M_UMA);
2369 	for (c = zone->uz_ctlname; *c != '\0'; c++)
2370 		if (strchr("./\\ -", *c) != NULL)
2371 			*c = '_';
2372 
2373 	/*
2374 	 * Basic parameters at the root.
2375 	 */
2376 	zone->uz_oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_vm_uma),
2377 	    OID_AUTO, zone->uz_ctlname, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2378 	oid = zone->uz_oid;
2379 	SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2380 	    "size", CTLFLAG_RD, &zone->uz_size, 0, "Allocation size");
2381 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2382 	    "flags", CTLFLAG_RD | CTLTYPE_STRING | CTLFLAG_MPSAFE,
2383 	    zone, 0, sysctl_handle_uma_zone_flags, "A",
2384 	    "Allocator configuration flags");
2385 	SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2386 	    "bucket_size", CTLFLAG_RD, &zone->uz_bucket_size, 0,
2387 	    "Desired per-cpu cache size");
2388 	SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2389 	    "bucket_size_max", CTLFLAG_RD, &zone->uz_bucket_size_max, 0,
2390 	    "Maximum allowed per-cpu cache size");
2391 
2392 	/*
2393 	 * keg if present.
2394 	 */
2395 	if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
2396 		domains = vm_ndomains;
2397 	else
2398 		domains = 1;
2399 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2400 	    "keg", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2401 	keg = zone->uz_keg;
2402 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0) {
2403 		SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2404 		    "name", CTLFLAG_RD, keg->uk_name, "Keg name");
2405 		SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2406 		    "rsize", CTLFLAG_RD, &keg->uk_rsize, 0,
2407 		    "Real object size with alignment");
2408 		SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2409 		    "ppera", CTLFLAG_RD, &keg->uk_ppera, 0,
2410 		    "pages per-slab allocation");
2411 		SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2412 		    "ipers", CTLFLAG_RD, &keg->uk_ipers, 0,
2413 		    "items available per-slab");
2414 		SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2415 		    "align", CTLFLAG_RD, &keg->uk_align, 0,
2416 		    "item alignment mask");
2417 		SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2418 		    "efficiency", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2419 		    keg, 0, sysctl_handle_uma_slab_efficiency, "I",
2420 		    "Slab utilization (100 - internal fragmentation %)");
2421 		domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(oid),
2422 		    OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2423 		for (i = 0; i < domains; i++) {
2424 			dom = &keg->uk_domain[i];
2425 			oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2426 			    OID_AUTO, VM_DOMAIN(i)->vmd_name,
2427 			    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2428 			SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2429 			    "pages", CTLFLAG_RD, &dom->ud_pages, 0,
2430 			    "Total pages currently allocated from VM");
2431 			SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2432 			    "free_items", CTLFLAG_RD, &dom->ud_free_items, 0,
2433 			    "items free in the slab layer");
2434 		}
2435 	} else
2436 		SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2437 		    "name", CTLFLAG_RD, nokeg, "Keg name");
2438 
2439 	/*
2440 	 * Information about zone limits.
2441 	 */
2442 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2443 	    "limit", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2444 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2445 	    "items", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2446 	    zone, 0, sysctl_handle_uma_zone_items, "QU",
2447 	    "current number of allocated items if limit is set");
2448 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2449 	    "max_items", CTLFLAG_RD, &zone->uz_max_items, 0,
2450 	    "Maximum number of cached items");
2451 	SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2452 	    "sleepers", CTLFLAG_RD, &zone->uz_sleepers, 0,
2453 	    "Number of threads sleeping at limit");
2454 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2455 	    "sleeps", CTLFLAG_RD, &zone->uz_sleeps, 0,
2456 	    "Total zone limit sleeps");
2457 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2458 	    "bucket_max", CTLFLAG_RD, &zone->uz_bucket_max, 0,
2459 	    "Maximum number of items in each domain's bucket cache");
2460 
2461 	/*
2462 	 * Per-domain zone information.
2463 	 */
2464 	domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid),
2465 	    OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2466 	for (i = 0; i < domains; i++) {
2467 		zdom = ZDOM_GET(zone, i);
2468 		oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2469 		    OID_AUTO, VM_DOMAIN(i)->vmd_name,
2470 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2471 		SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2472 		    "nitems", CTLFLAG_RD, &zdom->uzd_nitems,
2473 		    "number of items in this domain");
2474 		SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2475 		    "imax", CTLFLAG_RD, &zdom->uzd_imax,
2476 		    "maximum item count in this period");
2477 		SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2478 		    "imin", CTLFLAG_RD, &zdom->uzd_imin,
2479 		    "minimum item count in this period");
2480 		SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2481 		    "wss", CTLFLAG_RD, &zdom->uzd_wss,
2482 		    "Working set size");
2483 	}
2484 
2485 	/*
2486 	 * General statistics.
2487 	 */
2488 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2489 	    "stats", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2490 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2491 	    "current", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2492 	    zone, 1, sysctl_handle_uma_zone_cur, "I",
2493 	    "Current number of allocated items");
2494 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2495 	    "allocs", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2496 	    zone, 0, sysctl_handle_uma_zone_allocs, "QU",
2497 	    "Total allocation calls");
2498 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2499 	    "frees", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2500 	    zone, 0, sysctl_handle_uma_zone_frees, "QU",
2501 	    "Total free calls");
2502 	SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2503 	    "fails", CTLFLAG_RD, &zone->uz_fails,
2504 	    "Number of allocation failures");
2505 	SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2506 	    "xdomain", CTLFLAG_RD, &zone->uz_xdomain,
2507 	    "Free calls from the wrong domain");
2508 }
2509 
2510 struct uma_zone_count {
2511 	const char	*name;
2512 	int		count;
2513 };
2514 
2515 static void
2516 zone_count(uma_zone_t zone, void *arg)
2517 {
2518 	struct uma_zone_count *cnt;
2519 
2520 	cnt = arg;
2521 	/*
2522 	 * Some zones are rapidly created with identical names and
2523 	 * destroyed out of order.  This can lead to gaps in the count.
2524 	 * Use one greater than the maximum observed for this name.
2525 	 */
2526 	if (strcmp(zone->uz_name, cnt->name) == 0)
2527 		cnt->count = MAX(cnt->count,
2528 		    zone->uz_namecnt + 1);
2529 }
2530 
2531 static void
2532 zone_update_caches(uma_zone_t zone)
2533 {
2534 	int i;
2535 
2536 	for (i = 0; i <= mp_maxid; i++) {
2537 		cache_set_uz_size(&zone->uz_cpu[i], zone->uz_size);
2538 		cache_set_uz_flags(&zone->uz_cpu[i], zone->uz_flags);
2539 	}
2540 }
2541 
2542 /*
2543  * Zone header ctor.  This initializes all fields, locks, etc.
2544  *
2545  * Arguments/Returns follow uma_ctor specifications
2546  *	udata  Actually uma_zctor_args
2547  */
2548 static int
2549 zone_ctor(void *mem, int size, void *udata, int flags)
2550 {
2551 	struct uma_zone_count cnt;
2552 	struct uma_zctor_args *arg = udata;
2553 	uma_zone_domain_t zdom;
2554 	uma_zone_t zone = mem;
2555 	uma_zone_t z;
2556 	uma_keg_t keg;
2557 	int i;
2558 
2559 	bzero(zone, size);
2560 	zone->uz_name = arg->name;
2561 	zone->uz_ctor = arg->ctor;
2562 	zone->uz_dtor = arg->dtor;
2563 	zone->uz_init = NULL;
2564 	zone->uz_fini = NULL;
2565 	zone->uz_sleeps = 0;
2566 	zone->uz_bucket_size = 0;
2567 	zone->uz_bucket_size_min = 0;
2568 	zone->uz_bucket_size_max = BUCKET_MAX;
2569 	zone->uz_flags = (arg->flags & UMA_ZONE_SMR);
2570 	zone->uz_warning = NULL;
2571 	/* The domain structures follow the cpu structures. */
2572 	zone->uz_bucket_max = ULONG_MAX;
2573 	timevalclear(&zone->uz_ratecheck);
2574 
2575 	/* Count the number of duplicate names. */
2576 	cnt.name = arg->name;
2577 	cnt.count = 0;
2578 	zone_foreach(zone_count, &cnt);
2579 	zone->uz_namecnt = cnt.count;
2580 	ZONE_CROSS_LOCK_INIT(zone);
2581 
2582 	for (i = 0; i < vm_ndomains; i++) {
2583 		zdom = ZDOM_GET(zone, i);
2584 		ZDOM_LOCK_INIT(zone, zdom, (arg->flags & UMA_ZONE_MTXCLASS));
2585 		STAILQ_INIT(&zdom->uzd_buckets);
2586 	}
2587 
2588 #ifdef INVARIANTS
2589 	if (arg->uminit == trash_init && arg->fini == trash_fini)
2590 		zone->uz_flags |= UMA_ZFLAG_TRASH | UMA_ZFLAG_CTORDTOR;
2591 #endif
2592 
2593 	/*
2594 	 * This is a pure cache zone, no kegs.
2595 	 */
2596 	if (arg->import) {
2597 		KASSERT((arg->flags & UMA_ZFLAG_CACHE) != 0,
2598 		    ("zone_ctor: Import specified for non-cache zone."));
2599 		zone->uz_flags = arg->flags;
2600 		zone->uz_size = arg->size;
2601 		zone->uz_import = arg->import;
2602 		zone->uz_release = arg->release;
2603 		zone->uz_arg = arg->arg;
2604 #ifdef NUMA
2605 		/*
2606 		 * Cache zones are round-robin unless a policy is
2607 		 * specified because they may have incompatible
2608 		 * constraints.
2609 		 */
2610 		if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2611 			zone->uz_flags |= UMA_ZONE_ROUNDROBIN;
2612 #endif
2613 		rw_wlock(&uma_rwlock);
2614 		LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link);
2615 		rw_wunlock(&uma_rwlock);
2616 		goto out;
2617 	}
2618 
2619 	/*
2620 	 * Use the regular zone/keg/slab allocator.
2621 	 */
2622 	zone->uz_import = zone_import;
2623 	zone->uz_release = zone_release;
2624 	zone->uz_arg = zone;
2625 	keg = arg->keg;
2626 
2627 	if (arg->flags & UMA_ZONE_SECONDARY) {
2628 		KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
2629 		    ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
2630 		KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
2631 		zone->uz_init = arg->uminit;
2632 		zone->uz_fini = arg->fini;
2633 		zone->uz_flags |= UMA_ZONE_SECONDARY;
2634 		rw_wlock(&uma_rwlock);
2635 		ZONE_LOCK(zone);
2636 		LIST_FOREACH(z, &keg->uk_zones, uz_link) {
2637 			if (LIST_NEXT(z, uz_link) == NULL) {
2638 				LIST_INSERT_AFTER(z, zone, uz_link);
2639 				break;
2640 			}
2641 		}
2642 		ZONE_UNLOCK(zone);
2643 		rw_wunlock(&uma_rwlock);
2644 	} else if (keg == NULL) {
2645 		if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
2646 		    arg->align, arg->flags)) == NULL)
2647 			return (ENOMEM);
2648 	} else {
2649 		struct uma_kctor_args karg;
2650 		int error;
2651 
2652 		/* We should only be here from uma_startup() */
2653 		karg.size = arg->size;
2654 		karg.uminit = arg->uminit;
2655 		karg.fini = arg->fini;
2656 		karg.align = arg->align;
2657 		karg.flags = (arg->flags & ~UMA_ZONE_SMR);
2658 		karg.zone = zone;
2659 		error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
2660 		    flags);
2661 		if (error)
2662 			return (error);
2663 	}
2664 
2665 	/* Inherit properties from the keg. */
2666 	zone->uz_keg = keg;
2667 	zone->uz_size = keg->uk_size;
2668 	zone->uz_flags |= (keg->uk_flags &
2669 	    (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
2670 
2671 out:
2672 	if (__predict_true(booted >= BOOT_RUNNING)) {
2673 		zone_alloc_counters(zone, NULL);
2674 		zone_alloc_sysctl(zone, NULL);
2675 	} else {
2676 		zone->uz_allocs = EARLY_COUNTER;
2677 		zone->uz_frees = EARLY_COUNTER;
2678 		zone->uz_fails = EARLY_COUNTER;
2679 	}
2680 
2681 	/* Caller requests a private SMR context. */
2682 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
2683 		zone->uz_smr = smr_create(zone->uz_name, 0, 0);
2684 
2685 	KASSERT((arg->flags & (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET)) !=
2686 	    (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET),
2687 	    ("Invalid zone flag combination"));
2688 	if (arg->flags & UMA_ZFLAG_INTERNAL)
2689 		zone->uz_bucket_size_max = zone->uz_bucket_size = 0;
2690 	if ((arg->flags & UMA_ZONE_MAXBUCKET) != 0)
2691 		zone->uz_bucket_size = BUCKET_MAX;
2692 	else if ((arg->flags & UMA_ZONE_MINBUCKET) != 0)
2693 		zone->uz_bucket_size_max = zone->uz_bucket_size = BUCKET_MIN;
2694 	else if ((arg->flags & UMA_ZONE_NOBUCKET) != 0)
2695 		zone->uz_bucket_size = 0;
2696 	else
2697 		zone->uz_bucket_size = bucket_select(zone->uz_size);
2698 	zone->uz_bucket_size_min = zone->uz_bucket_size;
2699 	if (zone->uz_dtor != NULL || zone->uz_ctor != NULL)
2700 		zone->uz_flags |= UMA_ZFLAG_CTORDTOR;
2701 	zone_update_caches(zone);
2702 
2703 	return (0);
2704 }
2705 
2706 /*
2707  * Keg header dtor.  This frees all data, destroys locks, frees the hash
2708  * table and removes the keg from the global list.
2709  *
2710  * Arguments/Returns follow uma_dtor specifications
2711  *	udata  unused
2712  */
2713 static void
2714 keg_dtor(void *arg, int size, void *udata)
2715 {
2716 	uma_keg_t keg;
2717 	uint32_t free, pages;
2718 	int i;
2719 
2720 	keg = (uma_keg_t)arg;
2721 	free = pages = 0;
2722 	for (i = 0; i < vm_ndomains; i++) {
2723 		free += keg->uk_domain[i].ud_free_items;
2724 		pages += keg->uk_domain[i].ud_pages;
2725 		KEG_LOCK_FINI(keg, i);
2726 	}
2727 	if (pages != 0)
2728 		printf("Freed UMA keg (%s) was not empty (%u items). "
2729 		    " Lost %u pages of memory.\n",
2730 		    keg->uk_name ? keg->uk_name : "",
2731 		    pages / keg->uk_ppera * keg->uk_ipers - free, pages);
2732 
2733 	hash_free(&keg->uk_hash);
2734 }
2735 
2736 /*
2737  * Zone header dtor.
2738  *
2739  * Arguments/Returns follow uma_dtor specifications
2740  *	udata  unused
2741  */
2742 static void
2743 zone_dtor(void *arg, int size, void *udata)
2744 {
2745 	uma_zone_t zone;
2746 	uma_keg_t keg;
2747 	int i;
2748 
2749 	zone = (uma_zone_t)arg;
2750 
2751 	sysctl_remove_oid(zone->uz_oid, 1, 1);
2752 
2753 	if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
2754 		cache_drain(zone);
2755 
2756 	rw_wlock(&uma_rwlock);
2757 	LIST_REMOVE(zone, uz_link);
2758 	rw_wunlock(&uma_rwlock);
2759 	zone_reclaim(zone, M_WAITOK, true);
2760 
2761 	/*
2762 	 * We only destroy kegs from non secondary/non cache zones.
2763 	 */
2764 	if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) {
2765 		keg = zone->uz_keg;
2766 		rw_wlock(&uma_rwlock);
2767 		LIST_REMOVE(keg, uk_link);
2768 		rw_wunlock(&uma_rwlock);
2769 		zone_free_item(kegs, keg, NULL, SKIP_NONE);
2770 	}
2771 	counter_u64_free(zone->uz_allocs);
2772 	counter_u64_free(zone->uz_frees);
2773 	counter_u64_free(zone->uz_fails);
2774 	counter_u64_free(zone->uz_xdomain);
2775 	free(zone->uz_ctlname, M_UMA);
2776 	for (i = 0; i < vm_ndomains; i++)
2777 		ZDOM_LOCK_FINI(ZDOM_GET(zone, i));
2778 	ZONE_CROSS_LOCK_FINI(zone);
2779 }
2780 
2781 static void
2782 zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *arg), void *arg)
2783 {
2784 	uma_keg_t keg;
2785 	uma_zone_t zone;
2786 
2787 	LIST_FOREACH(keg, &uma_kegs, uk_link) {
2788 		LIST_FOREACH(zone, &keg->uk_zones, uz_link)
2789 			zfunc(zone, arg);
2790 	}
2791 	LIST_FOREACH(zone, &uma_cachezones, uz_link)
2792 		zfunc(zone, arg);
2793 }
2794 
2795 /*
2796  * Traverses every zone in the system and calls a callback
2797  *
2798  * Arguments:
2799  *	zfunc  A pointer to a function which accepts a zone
2800  *		as an argument.
2801  *
2802  * Returns:
2803  *	Nothing
2804  */
2805 static void
2806 zone_foreach(void (*zfunc)(uma_zone_t, void *arg), void *arg)
2807 {
2808 
2809 	rw_rlock(&uma_rwlock);
2810 	zone_foreach_unlocked(zfunc, arg);
2811 	rw_runlock(&uma_rwlock);
2812 }
2813 
2814 /*
2815  * Initialize the kernel memory allocator.  This is done after pages can be
2816  * allocated but before general KVA is available.
2817  */
2818 void
2819 uma_startup1(vm_offset_t virtual_avail)
2820 {
2821 	struct uma_zctor_args args;
2822 	size_t ksize, zsize, size;
2823 	uma_keg_t masterkeg;
2824 	uintptr_t m;
2825 	uint8_t pflag;
2826 
2827 	bootstart = bootmem = virtual_avail;
2828 
2829 	rw_init(&uma_rwlock, "UMA lock");
2830 	sx_init(&uma_reclaim_lock, "umareclaim");
2831 
2832 	ksize = sizeof(struct uma_keg) +
2833 	    (sizeof(struct uma_domain) * vm_ndomains);
2834 	ksize = roundup(ksize, UMA_SUPER_ALIGN);
2835 	zsize = sizeof(struct uma_zone) +
2836 	    (sizeof(struct uma_cache) * (mp_maxid + 1)) +
2837 	    (sizeof(struct uma_zone_domain) * vm_ndomains);
2838 	zsize = roundup(zsize, UMA_SUPER_ALIGN);
2839 
2840 	/* Allocate the zone of zones, zone of kegs, and zone of zones keg. */
2841 	size = (zsize * 2) + ksize;
2842 	m = (uintptr_t)startup_alloc(NULL, size, 0, &pflag, M_NOWAIT | M_ZERO);
2843 	zones = (uma_zone_t)m;
2844 	m += zsize;
2845 	kegs = (uma_zone_t)m;
2846 	m += zsize;
2847 	masterkeg = (uma_keg_t)m;
2848 
2849 	/* "manually" create the initial zone */
2850 	memset(&args, 0, sizeof(args));
2851 	args.name = "UMA Kegs";
2852 	args.size = ksize;
2853 	args.ctor = keg_ctor;
2854 	args.dtor = keg_dtor;
2855 	args.uminit = zero_init;
2856 	args.fini = NULL;
2857 	args.keg = masterkeg;
2858 	args.align = UMA_SUPER_ALIGN - 1;
2859 	args.flags = UMA_ZFLAG_INTERNAL;
2860 	zone_ctor(kegs, zsize, &args, M_WAITOK);
2861 
2862 	args.name = "UMA Zones";
2863 	args.size = zsize;
2864 	args.ctor = zone_ctor;
2865 	args.dtor = zone_dtor;
2866 	args.uminit = zero_init;
2867 	args.fini = NULL;
2868 	args.keg = NULL;
2869 	args.align = UMA_SUPER_ALIGN - 1;
2870 	args.flags = UMA_ZFLAG_INTERNAL;
2871 	zone_ctor(zones, zsize, &args, M_WAITOK);
2872 
2873 	/* Now make zones for slab headers */
2874 	slabzones[0] = uma_zcreate("UMA Slabs 0", SLABZONE0_SIZE,
2875 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2876 	slabzones[1] = uma_zcreate("UMA Slabs 1", SLABZONE1_SIZE,
2877 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2878 
2879 	hashzone = uma_zcreate("UMA Hash",
2880 	    sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
2881 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2882 
2883 	bucket_init();
2884 	smr_init();
2885 }
2886 
2887 #ifndef UMA_MD_SMALL_ALLOC
2888 extern void vm_radix_reserve_kva(void);
2889 #endif
2890 
2891 /*
2892  * Advertise the availability of normal kva allocations and switch to
2893  * the default back-end allocator.  Marks the KVA we consumed on startup
2894  * as used in the map.
2895  */
2896 void
2897 uma_startup2(void)
2898 {
2899 
2900 	if (bootstart != bootmem) {
2901 		vm_map_lock(kernel_map);
2902 		(void)vm_map_insert(kernel_map, NULL, 0, bootstart, bootmem,
2903 		    VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT);
2904 		vm_map_unlock(kernel_map);
2905 	}
2906 
2907 #ifndef UMA_MD_SMALL_ALLOC
2908 	/* Set up radix zone to use noobj_alloc. */
2909 	vm_radix_reserve_kva();
2910 #endif
2911 
2912 	booted = BOOT_KVA;
2913 	zone_foreach_unlocked(zone_kva_available, NULL);
2914 	bucket_enable();
2915 }
2916 
2917 /*
2918  * Finish our initialization steps.
2919  */
2920 static void
2921 uma_startup3(void)
2922 {
2923 
2924 #ifdef INVARIANTS
2925 	TUNABLE_INT_FETCH("vm.debug.divisor", &dbg_divisor);
2926 	uma_dbg_cnt = counter_u64_alloc(M_WAITOK);
2927 	uma_skip_cnt = counter_u64_alloc(M_WAITOK);
2928 #endif
2929 	zone_foreach_unlocked(zone_alloc_counters, NULL);
2930 	zone_foreach_unlocked(zone_alloc_sysctl, NULL);
2931 	callout_init(&uma_callout, 1);
2932 	callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
2933 	booted = BOOT_RUNNING;
2934 
2935 	EVENTHANDLER_REGISTER(shutdown_post_sync, uma_shutdown, NULL,
2936 	    EVENTHANDLER_PRI_FIRST);
2937 }
2938 
2939 static void
2940 uma_shutdown(void)
2941 {
2942 
2943 	booted = BOOT_SHUTDOWN;
2944 }
2945 
2946 static uma_keg_t
2947 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
2948 		int align, uint32_t flags)
2949 {
2950 	struct uma_kctor_args args;
2951 
2952 	args.size = size;
2953 	args.uminit = uminit;
2954 	args.fini = fini;
2955 	args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
2956 	args.flags = flags;
2957 	args.zone = zone;
2958 	return (zone_alloc_item(kegs, &args, UMA_ANYDOMAIN, M_WAITOK));
2959 }
2960 
2961 /* Public functions */
2962 /* See uma.h */
2963 void
2964 uma_set_align(int align)
2965 {
2966 
2967 	if (align != UMA_ALIGN_CACHE)
2968 		uma_align_cache = align;
2969 }
2970 
2971 /* See uma.h */
2972 uma_zone_t
2973 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
2974 		uma_init uminit, uma_fini fini, int align, uint32_t flags)
2975 
2976 {
2977 	struct uma_zctor_args args;
2978 	uma_zone_t res;
2979 
2980 	KASSERT(powerof2(align + 1), ("invalid zone alignment %d for \"%s\"",
2981 	    align, name));
2982 
2983 	/* This stuff is essential for the zone ctor */
2984 	memset(&args, 0, sizeof(args));
2985 	args.name = name;
2986 	args.size = size;
2987 	args.ctor = ctor;
2988 	args.dtor = dtor;
2989 	args.uminit = uminit;
2990 	args.fini = fini;
2991 #ifdef  INVARIANTS
2992 	/*
2993 	 * Inject procedures which check for memory use after free if we are
2994 	 * allowed to scramble the memory while it is not allocated.  This
2995 	 * requires that: UMA is actually able to access the memory, no init
2996 	 * or fini procedures, no dependency on the initial value of the
2997 	 * memory, and no (legitimate) use of the memory after free.  Note,
2998 	 * the ctor and dtor do not need to be empty.
2999 	 */
3000 	if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOTOUCH |
3001 	    UMA_ZONE_NOFREE))) && uminit == NULL && fini == NULL) {
3002 		args.uminit = trash_init;
3003 		args.fini = trash_fini;
3004 	}
3005 #endif
3006 	args.align = align;
3007 	args.flags = flags;
3008 	args.keg = NULL;
3009 
3010 	sx_slock(&uma_reclaim_lock);
3011 	res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3012 	sx_sunlock(&uma_reclaim_lock);
3013 
3014 	return (res);
3015 }
3016 
3017 /* See uma.h */
3018 uma_zone_t
3019 uma_zsecond_create(const char *name, uma_ctor ctor, uma_dtor dtor,
3020     uma_init zinit, uma_fini zfini, uma_zone_t master)
3021 {
3022 	struct uma_zctor_args args;
3023 	uma_keg_t keg;
3024 	uma_zone_t res;
3025 
3026 	keg = master->uz_keg;
3027 	memset(&args, 0, sizeof(args));
3028 	args.name = name;
3029 	args.size = keg->uk_size;
3030 	args.ctor = ctor;
3031 	args.dtor = dtor;
3032 	args.uminit = zinit;
3033 	args.fini = zfini;
3034 	args.align = keg->uk_align;
3035 	args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
3036 	args.keg = keg;
3037 
3038 	sx_slock(&uma_reclaim_lock);
3039 	res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3040 	sx_sunlock(&uma_reclaim_lock);
3041 
3042 	return (res);
3043 }
3044 
3045 /* See uma.h */
3046 uma_zone_t
3047 uma_zcache_create(const char *name, int size, uma_ctor ctor, uma_dtor dtor,
3048     uma_init zinit, uma_fini zfini, uma_import zimport, uma_release zrelease,
3049     void *arg, int flags)
3050 {
3051 	struct uma_zctor_args args;
3052 
3053 	memset(&args, 0, sizeof(args));
3054 	args.name = name;
3055 	args.size = size;
3056 	args.ctor = ctor;
3057 	args.dtor = dtor;
3058 	args.uminit = zinit;
3059 	args.fini = zfini;
3060 	args.import = zimport;
3061 	args.release = zrelease;
3062 	args.arg = arg;
3063 	args.align = 0;
3064 	args.flags = flags | UMA_ZFLAG_CACHE;
3065 
3066 	return (zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK));
3067 }
3068 
3069 /* See uma.h */
3070 void
3071 uma_zdestroy(uma_zone_t zone)
3072 {
3073 
3074 	/*
3075 	 * Large slabs are expensive to reclaim, so don't bother doing
3076 	 * unnecessary work if we're shutting down.
3077 	 */
3078 	if (booted == BOOT_SHUTDOWN &&
3079 	    zone->uz_fini == NULL && zone->uz_release == zone_release)
3080 		return;
3081 	sx_slock(&uma_reclaim_lock);
3082 	zone_free_item(zones, zone, NULL, SKIP_NONE);
3083 	sx_sunlock(&uma_reclaim_lock);
3084 }
3085 
3086 void
3087 uma_zwait(uma_zone_t zone)
3088 {
3089 
3090 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
3091 		uma_zfree_smr(zone, uma_zalloc_smr(zone, M_WAITOK));
3092 	else if ((zone->uz_flags & UMA_ZONE_PCPU) != 0)
3093 		uma_zfree_pcpu(zone, uma_zalloc_pcpu(zone, M_WAITOK));
3094 	else
3095 		uma_zfree(zone, uma_zalloc(zone, M_WAITOK));
3096 }
3097 
3098 void *
3099 uma_zalloc_pcpu_arg(uma_zone_t zone, void *udata, int flags)
3100 {
3101 	void *item, *pcpu_item;
3102 #ifdef SMP
3103 	int i;
3104 
3105 	MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3106 #endif
3107 	item = uma_zalloc_arg(zone, udata, flags & ~M_ZERO);
3108 	if (item == NULL)
3109 		return (NULL);
3110 	pcpu_item = zpcpu_base_to_offset(item);
3111 	if (flags & M_ZERO) {
3112 #ifdef SMP
3113 		for (i = 0; i <= mp_maxid; i++)
3114 			bzero(zpcpu_get_cpu(pcpu_item, i), zone->uz_size);
3115 #else
3116 		bzero(item, zone->uz_size);
3117 #endif
3118 	}
3119 	return (pcpu_item);
3120 }
3121 
3122 /*
3123  * A stub while both regular and pcpu cases are identical.
3124  */
3125 void
3126 uma_zfree_pcpu_arg(uma_zone_t zone, void *pcpu_item, void *udata)
3127 {
3128 	void *item;
3129 
3130 #ifdef SMP
3131 	MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3132 #endif
3133 	item = zpcpu_offset_to_base(pcpu_item);
3134 	uma_zfree_arg(zone, item, udata);
3135 }
3136 
3137 static inline void *
3138 item_ctor(uma_zone_t zone, int uz_flags, int size, void *udata, int flags,
3139     void *item)
3140 {
3141 #ifdef INVARIANTS
3142 	bool skipdbg;
3143 
3144 	skipdbg = uma_dbg_zskip(zone, item);
3145 	if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3146 	    zone->uz_ctor != trash_ctor)
3147 		trash_ctor(item, size, udata, flags);
3148 #endif
3149 	/* Check flags before loading ctor pointer. */
3150 	if (__predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0) &&
3151 	    __predict_false(zone->uz_ctor != NULL) &&
3152 	    zone->uz_ctor(item, size, udata, flags) != 0) {
3153 		counter_u64_add(zone->uz_fails, 1);
3154 		zone_free_item(zone, item, udata, SKIP_DTOR | SKIP_CNT);
3155 		return (NULL);
3156 	}
3157 #ifdef INVARIANTS
3158 	if (!skipdbg)
3159 		uma_dbg_alloc(zone, NULL, item);
3160 #endif
3161 	if (__predict_false(flags & M_ZERO))
3162 		return (memset(item, 0, size));
3163 
3164 	return (item);
3165 }
3166 
3167 static inline void
3168 item_dtor(uma_zone_t zone, void *item, int size, void *udata,
3169     enum zfreeskip skip)
3170 {
3171 #ifdef INVARIANTS
3172 	bool skipdbg;
3173 
3174 	skipdbg = uma_dbg_zskip(zone, item);
3175 	if (skip == SKIP_NONE && !skipdbg) {
3176 		if ((zone->uz_flags & UMA_ZONE_MALLOC) != 0)
3177 			uma_dbg_free(zone, udata, item);
3178 		else
3179 			uma_dbg_free(zone, NULL, item);
3180 	}
3181 #endif
3182 	if (__predict_true(skip < SKIP_DTOR)) {
3183 		if (zone->uz_dtor != NULL)
3184 			zone->uz_dtor(item, size, udata);
3185 #ifdef INVARIANTS
3186 		if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3187 		    zone->uz_dtor != trash_dtor)
3188 			trash_dtor(item, size, udata);
3189 #endif
3190 	}
3191 }
3192 
3193 #if defined(INVARIANTS) || defined(DEBUG_MEMGUARD) || defined(WITNESS)
3194 #define	UMA_ZALLOC_DEBUG
3195 static int
3196 uma_zalloc_debug(uma_zone_t zone, void **itemp, void *udata, int flags)
3197 {
3198 	int error;
3199 
3200 	error = 0;
3201 #ifdef WITNESS
3202 	if (flags & M_WAITOK) {
3203 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
3204 		    "uma_zalloc_debug: zone \"%s\"", zone->uz_name);
3205 	}
3206 #endif
3207 
3208 #ifdef INVARIANTS
3209 	KASSERT((flags & M_EXEC) == 0,
3210 	    ("uma_zalloc_debug: called with M_EXEC"));
3211 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3212 	    ("uma_zalloc_debug: called within spinlock or critical section"));
3213 	KASSERT((zone->uz_flags & UMA_ZONE_PCPU) == 0 || (flags & M_ZERO) == 0,
3214 	    ("uma_zalloc_debug: allocating from a pcpu zone with M_ZERO"));
3215 #endif
3216 
3217 #ifdef DEBUG_MEMGUARD
3218 	if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && memguard_cmp_zone(zone)) {
3219 		void *item;
3220 		item = memguard_alloc(zone->uz_size, flags);
3221 		if (item != NULL) {
3222 			error = EJUSTRETURN;
3223 			if (zone->uz_init != NULL &&
3224 			    zone->uz_init(item, zone->uz_size, flags) != 0) {
3225 				*itemp = NULL;
3226 				return (error);
3227 			}
3228 			if (zone->uz_ctor != NULL &&
3229 			    zone->uz_ctor(item, zone->uz_size, udata,
3230 			    flags) != 0) {
3231 				counter_u64_add(zone->uz_fails, 1);
3232 			    	zone->uz_fini(item, zone->uz_size);
3233 				*itemp = NULL;
3234 				return (error);
3235 			}
3236 			*itemp = item;
3237 			return (error);
3238 		}
3239 		/* This is unfortunate but should not be fatal. */
3240 	}
3241 #endif
3242 	return (error);
3243 }
3244 
3245 static int
3246 uma_zfree_debug(uma_zone_t zone, void *item, void *udata)
3247 {
3248 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3249 	    ("uma_zfree_debug: called with spinlock or critical section held"));
3250 
3251 #ifdef DEBUG_MEMGUARD
3252 	if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && is_memguard_addr(item)) {
3253 		if (zone->uz_dtor != NULL)
3254 			zone->uz_dtor(item, zone->uz_size, udata);
3255 		if (zone->uz_fini != NULL)
3256 			zone->uz_fini(item, zone->uz_size);
3257 		memguard_free(item);
3258 		return (EJUSTRETURN);
3259 	}
3260 #endif
3261 	return (0);
3262 }
3263 #endif
3264 
3265 static inline void *
3266 cache_alloc_item(uma_zone_t zone, uma_cache_t cache, uma_cache_bucket_t bucket,
3267     void *udata, int flags)
3268 {
3269 	void *item;
3270 	int size, uz_flags;
3271 
3272 	item = cache_bucket_pop(cache, bucket);
3273 	size = cache_uz_size(cache);
3274 	uz_flags = cache_uz_flags(cache);
3275 	critical_exit();
3276 	return (item_ctor(zone, uz_flags, size, udata, flags, item));
3277 }
3278 
3279 static __noinline void *
3280 cache_alloc_retry(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3281 {
3282 	uma_cache_bucket_t bucket;
3283 	int domain;
3284 
3285 	while (cache_alloc(zone, cache, udata, flags)) {
3286 		cache = &zone->uz_cpu[curcpu];
3287 		bucket = &cache->uc_allocbucket;
3288 		if (__predict_false(bucket->ucb_cnt == 0))
3289 			continue;
3290 		return (cache_alloc_item(zone, cache, bucket, udata, flags));
3291 	}
3292 	critical_exit();
3293 
3294 	/*
3295 	 * We can not get a bucket so try to return a single item.
3296 	 */
3297 	if (zone->uz_flags & UMA_ZONE_FIRSTTOUCH)
3298 		domain = PCPU_GET(domain);
3299 	else
3300 		domain = UMA_ANYDOMAIN;
3301 	return (zone_alloc_item(zone, udata, domain, flags));
3302 }
3303 
3304 /* See uma.h */
3305 void *
3306 uma_zalloc_smr(uma_zone_t zone, int flags)
3307 {
3308 	uma_cache_bucket_t bucket;
3309 	uma_cache_t cache;
3310 
3311 #ifdef UMA_ZALLOC_DEBUG
3312 	void *item;
3313 
3314 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
3315 	    ("uma_zalloc_arg: called with non-SMR zone.\n"));
3316 	if (uma_zalloc_debug(zone, &item, NULL, flags) == EJUSTRETURN)
3317 		return (item);
3318 #endif
3319 
3320 	critical_enter();
3321 	cache = &zone->uz_cpu[curcpu];
3322 	bucket = &cache->uc_allocbucket;
3323 	if (__predict_false(bucket->ucb_cnt == 0))
3324 		return (cache_alloc_retry(zone, cache, NULL, flags));
3325 	return (cache_alloc_item(zone, cache, bucket, NULL, flags));
3326 }
3327 
3328 /* See uma.h */
3329 void *
3330 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
3331 {
3332 	uma_cache_bucket_t bucket;
3333 	uma_cache_t cache;
3334 
3335 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3336 	random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3337 
3338 	/* This is the fast path allocation */
3339 	CTR3(KTR_UMA, "uma_zalloc_arg zone %s(%p) flags %d", zone->uz_name,
3340 	    zone, flags);
3341 
3342 #ifdef UMA_ZALLOC_DEBUG
3343 	void *item;
3344 
3345 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3346 	    ("uma_zalloc_arg: called with SMR zone.\n"));
3347 	if (uma_zalloc_debug(zone, &item, udata, flags) == EJUSTRETURN)
3348 		return (item);
3349 #endif
3350 
3351 	/*
3352 	 * If possible, allocate from the per-CPU cache.  There are two
3353 	 * requirements for safe access to the per-CPU cache: (1) the thread
3354 	 * accessing the cache must not be preempted or yield during access,
3355 	 * and (2) the thread must not migrate CPUs without switching which
3356 	 * cache it accesses.  We rely on a critical section to prevent
3357 	 * preemption and migration.  We release the critical section in
3358 	 * order to acquire the zone mutex if we are unable to allocate from
3359 	 * the current cache; when we re-acquire the critical section, we
3360 	 * must detect and handle migration if it has occurred.
3361 	 */
3362 	critical_enter();
3363 	cache = &zone->uz_cpu[curcpu];
3364 	bucket = &cache->uc_allocbucket;
3365 	if (__predict_false(bucket->ucb_cnt == 0))
3366 		return (cache_alloc_retry(zone, cache, udata, flags));
3367 	return (cache_alloc_item(zone, cache, bucket, udata, flags));
3368 }
3369 
3370 /*
3371  * Replenish an alloc bucket and possibly restore an old one.  Called in
3372  * a critical section.  Returns in a critical section.
3373  *
3374  * A false return value indicates an allocation failure.
3375  * A true return value indicates success and the caller should retry.
3376  */
3377 static __noinline bool
3378 cache_alloc(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3379 {
3380 	uma_bucket_t bucket;
3381 	int domain;
3382 	bool new;
3383 
3384 	CRITICAL_ASSERT(curthread);
3385 
3386 	/*
3387 	 * If we have run out of items in our alloc bucket see
3388 	 * if we can switch with the free bucket.
3389 	 *
3390 	 * SMR Zones can't re-use the free bucket until the sequence has
3391 	 * expired.
3392 	 */
3393 	if ((cache_uz_flags(cache) & UMA_ZONE_SMR) == 0 &&
3394 	    cache->uc_freebucket.ucb_cnt != 0) {
3395 		cache_bucket_swap(&cache->uc_freebucket,
3396 		    &cache->uc_allocbucket);
3397 		return (true);
3398 	}
3399 
3400 	/*
3401 	 * Discard any empty allocation bucket while we hold no locks.
3402 	 */
3403 	bucket = cache_bucket_unload_alloc(cache);
3404 	critical_exit();
3405 
3406 	if (bucket != NULL) {
3407 		KASSERT(bucket->ub_cnt == 0,
3408 		    ("cache_alloc: Entered with non-empty alloc bucket."));
3409 		bucket_free(zone, bucket, udata);
3410 	}
3411 
3412 	/* Short-circuit for zones without buckets and low memory. */
3413 	if (zone->uz_bucket_size == 0 || bucketdisable) {
3414 		critical_enter();
3415 		return (false);
3416 	}
3417 
3418 	/*
3419 	 * Attempt to retrieve the item from the per-CPU cache has failed, so
3420 	 * we must go back to the zone.  This requires the zdom lock, so we
3421 	 * must drop the critical section, then re-acquire it when we go back
3422 	 * to the cache.  Since the critical section is released, we may be
3423 	 * preempted or migrate.  As such, make sure not to maintain any
3424 	 * thread-local state specific to the cache from prior to releasing
3425 	 * the critical section.
3426 	 */
3427 	domain = PCPU_GET(domain);
3428 	if ((cache_uz_flags(cache) & UMA_ZONE_ROUNDROBIN) != 0)
3429 		domain = zone_domain_highest(zone, domain);
3430 	bucket = cache_fetch_bucket(zone, cache, domain);
3431 	if (bucket == NULL) {
3432 		bucket = zone_alloc_bucket(zone, udata, domain, flags);
3433 		new = true;
3434 	} else
3435 		new = false;
3436 
3437 	CTR3(KTR_UMA, "uma_zalloc: zone %s(%p) bucket zone returned %p",
3438 	    zone->uz_name, zone, bucket);
3439 	if (bucket == NULL) {
3440 		critical_enter();
3441 		return (false);
3442 	}
3443 
3444 	/*
3445 	 * See if we lost the race or were migrated.  Cache the
3446 	 * initialized bucket to make this less likely or claim
3447 	 * the memory directly.
3448 	 */
3449 	critical_enter();
3450 	cache = &zone->uz_cpu[curcpu];
3451 	if (cache->uc_allocbucket.ucb_bucket == NULL &&
3452 	    ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) == 0 ||
3453 	    domain == PCPU_GET(domain))) {
3454 		if (new)
3455 			atomic_add_long(&ZDOM_GET(zone, domain)->uzd_imax,
3456 			    bucket->ub_cnt);
3457 		cache_bucket_load_alloc(cache, bucket);
3458 		return (true);
3459 	}
3460 
3461 	/*
3462 	 * We lost the race, release this bucket and start over.
3463 	 */
3464 	critical_exit();
3465 	zone_put_bucket(zone, domain, bucket, udata, false);
3466 	critical_enter();
3467 
3468 	return (true);
3469 }
3470 
3471 void *
3472 uma_zalloc_domain(uma_zone_t zone, void *udata, int domain, int flags)
3473 {
3474 
3475 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3476 	random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3477 
3478 	/* This is the fast path allocation */
3479 	CTR4(KTR_UMA, "uma_zalloc_domain zone %s(%p) domain %d flags %d",
3480 	    zone->uz_name, zone, domain, flags);
3481 
3482 	if (flags & M_WAITOK) {
3483 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
3484 		    "uma_zalloc_domain: zone \"%s\"", zone->uz_name);
3485 	}
3486 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3487 	    ("uma_zalloc_domain: called with spinlock or critical section held"));
3488 
3489 	return (zone_alloc_item(zone, udata, domain, flags));
3490 }
3491 
3492 /*
3493  * Find a slab with some space.  Prefer slabs that are partially used over those
3494  * that are totally full.  This helps to reduce fragmentation.
3495  *
3496  * If 'rr' is 1, search all domains starting from 'domain'.  Otherwise check
3497  * only 'domain'.
3498  */
3499 static uma_slab_t
3500 keg_first_slab(uma_keg_t keg, int domain, bool rr)
3501 {
3502 	uma_domain_t dom;
3503 	uma_slab_t slab;
3504 	int start;
3505 
3506 	KASSERT(domain >= 0 && domain < vm_ndomains,
3507 	    ("keg_first_slab: domain %d out of range", domain));
3508 	KEG_LOCK_ASSERT(keg, domain);
3509 
3510 	slab = NULL;
3511 	start = domain;
3512 	do {
3513 		dom = &keg->uk_domain[domain];
3514 		if ((slab = LIST_FIRST(&dom->ud_part_slab)) != NULL)
3515 			return (slab);
3516 		if ((slab = LIST_FIRST(&dom->ud_free_slab)) != NULL) {
3517 			LIST_REMOVE(slab, us_link);
3518 			dom->ud_free_slabs--;
3519 			LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
3520 			return (slab);
3521 		}
3522 		if (rr)
3523 			domain = (domain + 1) % vm_ndomains;
3524 	} while (domain != start);
3525 
3526 	return (NULL);
3527 }
3528 
3529 /*
3530  * Fetch an existing slab from a free or partial list.  Returns with the
3531  * keg domain lock held if a slab was found or unlocked if not.
3532  */
3533 static uma_slab_t
3534 keg_fetch_free_slab(uma_keg_t keg, int domain, bool rr, int flags)
3535 {
3536 	uma_slab_t slab;
3537 	uint32_t reserve;
3538 
3539 	/* HASH has a single free list. */
3540 	if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
3541 		domain = 0;
3542 
3543 	KEG_LOCK(keg, domain);
3544 	reserve = (flags & M_USE_RESERVE) != 0 ? 0 : keg->uk_reserve;
3545 	if (keg->uk_domain[domain].ud_free_items <= reserve ||
3546 	    (slab = keg_first_slab(keg, domain, rr)) == NULL) {
3547 		KEG_UNLOCK(keg, domain);
3548 		return (NULL);
3549 	}
3550 	return (slab);
3551 }
3552 
3553 static uma_slab_t
3554 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int rdomain, const int flags)
3555 {
3556 	struct vm_domainset_iter di;
3557 	uma_slab_t slab;
3558 	int aflags, domain;
3559 	bool rr;
3560 
3561 restart:
3562 	/*
3563 	 * Use the keg's policy if upper layers haven't already specified a
3564 	 * domain (as happens with first-touch zones).
3565 	 *
3566 	 * To avoid races we run the iterator with the keg lock held, but that
3567 	 * means that we cannot allow the vm_domainset layer to sleep.  Thus,
3568 	 * clear M_WAITOK and handle low memory conditions locally.
3569 	 */
3570 	rr = rdomain == UMA_ANYDOMAIN;
3571 	if (rr) {
3572 		aflags = (flags & ~M_WAITOK) | M_NOWAIT;
3573 		vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
3574 		    &aflags);
3575 	} else {
3576 		aflags = flags;
3577 		domain = rdomain;
3578 	}
3579 
3580 	for (;;) {
3581 		slab = keg_fetch_free_slab(keg, domain, rr, flags);
3582 		if (slab != NULL)
3583 			return (slab);
3584 
3585 		/*
3586 		 * M_NOVM means don't ask at all!
3587 		 */
3588 		if (flags & M_NOVM)
3589 			break;
3590 
3591 		slab = keg_alloc_slab(keg, zone, domain, flags, aflags);
3592 		if (slab != NULL)
3593 			return (slab);
3594 		if (!rr && (flags & M_WAITOK) == 0)
3595 			break;
3596 		if (rr && vm_domainset_iter_policy(&di, &domain) != 0) {
3597 			if ((flags & M_WAITOK) != 0) {
3598 				vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask);
3599 				goto restart;
3600 			}
3601 			break;
3602 		}
3603 	}
3604 
3605 	/*
3606 	 * We might not have been able to get a slab but another cpu
3607 	 * could have while we were unlocked.  Check again before we
3608 	 * fail.
3609 	 */
3610 	if ((slab = keg_fetch_free_slab(keg, domain, rr, flags)) != NULL)
3611 		return (slab);
3612 
3613 	return (NULL);
3614 }
3615 
3616 static void *
3617 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
3618 {
3619 	uma_domain_t dom;
3620 	void *item;
3621 	int freei;
3622 
3623 	KEG_LOCK_ASSERT(keg, slab->us_domain);
3624 
3625 	dom = &keg->uk_domain[slab->us_domain];
3626 	freei = BIT_FFS(keg->uk_ipers, &slab->us_free) - 1;
3627 	BIT_CLR(keg->uk_ipers, freei, &slab->us_free);
3628 	item = slab_item(slab, keg, freei);
3629 	slab->us_freecount--;
3630 	dom->ud_free_items--;
3631 
3632 	/*
3633 	 * Move this slab to the full list.  It must be on the partial list, so
3634 	 * we do not need to update the free slab count.  In particular,
3635 	 * keg_fetch_slab() always returns slabs on the partial list.
3636 	 */
3637 	if (slab->us_freecount == 0) {
3638 		LIST_REMOVE(slab, us_link);
3639 		LIST_INSERT_HEAD(&dom->ud_full_slab, slab, us_link);
3640 	}
3641 
3642 	return (item);
3643 }
3644 
3645 static int
3646 zone_import(void *arg, void **bucket, int max, int domain, int flags)
3647 {
3648 	uma_domain_t dom;
3649 	uma_zone_t zone;
3650 	uma_slab_t slab;
3651 	uma_keg_t keg;
3652 #ifdef NUMA
3653 	int stripe;
3654 #endif
3655 	int i;
3656 
3657 	zone = arg;
3658 	slab = NULL;
3659 	keg = zone->uz_keg;
3660 	/* Try to keep the buckets totally full */
3661 	for (i = 0; i < max; ) {
3662 		if ((slab = keg_fetch_slab(keg, zone, domain, flags)) == NULL)
3663 			break;
3664 #ifdef NUMA
3665 		stripe = howmany(max, vm_ndomains);
3666 #endif
3667 		dom = &keg->uk_domain[slab->us_domain];
3668 		while (slab->us_freecount && i < max) {
3669 			bucket[i++] = slab_alloc_item(keg, slab);
3670 			if (dom->ud_free_items <= keg->uk_reserve)
3671 				break;
3672 #ifdef NUMA
3673 			/*
3674 			 * If the zone is striped we pick a new slab for every
3675 			 * N allocations.  Eliminating this conditional will
3676 			 * instead pick a new domain for each bucket rather
3677 			 * than stripe within each bucket.  The current option
3678 			 * produces more fragmentation and requires more cpu
3679 			 * time but yields better distribution.
3680 			 */
3681 			if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0 &&
3682 			    vm_ndomains > 1 && --stripe == 0)
3683 				break;
3684 #endif
3685 		}
3686 		KEG_UNLOCK(keg, slab->us_domain);
3687 		/* Don't block if we allocated any successfully. */
3688 		flags &= ~M_WAITOK;
3689 		flags |= M_NOWAIT;
3690 	}
3691 
3692 	return i;
3693 }
3694 
3695 static int
3696 zone_alloc_limit_hard(uma_zone_t zone, int count, int flags)
3697 {
3698 	uint64_t old, new, total, max;
3699 
3700 	/*
3701 	 * The hard case.  We're going to sleep because there were existing
3702 	 * sleepers or because we ran out of items.  This routine enforces
3703 	 * fairness by keeping fifo order.
3704 	 *
3705 	 * First release our ill gotten gains and make some noise.
3706 	 */
3707 	for (;;) {
3708 		zone_free_limit(zone, count);
3709 		zone_log_warning(zone);
3710 		zone_maxaction(zone);
3711 		if (flags & M_NOWAIT)
3712 			return (0);
3713 
3714 		/*
3715 		 * We need to allocate an item or set ourself as a sleeper
3716 		 * while the sleepq lock is held to avoid wakeup races.  This
3717 		 * is essentially a home rolled semaphore.
3718 		 */
3719 		sleepq_lock(&zone->uz_max_items);
3720 		old = zone->uz_items;
3721 		do {
3722 			MPASS(UZ_ITEMS_SLEEPERS(old) < UZ_ITEMS_SLEEPERS_MAX);
3723 			/* Cache the max since we will evaluate twice. */
3724 			max = zone->uz_max_items;
3725 			if (UZ_ITEMS_SLEEPERS(old) != 0 ||
3726 			    UZ_ITEMS_COUNT(old) >= max)
3727 				new = old + UZ_ITEMS_SLEEPER;
3728 			else
3729 				new = old + MIN(count, max - old);
3730 		} while (atomic_fcmpset_64(&zone->uz_items, &old, new) == 0);
3731 
3732 		/* We may have successfully allocated under the sleepq lock. */
3733 		if (UZ_ITEMS_SLEEPERS(new) == 0) {
3734 			sleepq_release(&zone->uz_max_items);
3735 			return (new - old);
3736 		}
3737 
3738 		/*
3739 		 * This is in a different cacheline from uz_items so that we
3740 		 * don't constantly invalidate the fastpath cacheline when we
3741 		 * adjust item counts.  This could be limited to toggling on
3742 		 * transitions.
3743 		 */
3744 		atomic_add_32(&zone->uz_sleepers, 1);
3745 		atomic_add_64(&zone->uz_sleeps, 1);
3746 
3747 		/*
3748 		 * We have added ourselves as a sleeper.  The sleepq lock
3749 		 * protects us from wakeup races.  Sleep now and then retry.
3750 		 */
3751 		sleepq_add(&zone->uz_max_items, NULL, "zonelimit", 0, 0);
3752 		sleepq_wait(&zone->uz_max_items, PVM);
3753 
3754 		/*
3755 		 * After wakeup, remove ourselves as a sleeper and try
3756 		 * again.  We no longer have the sleepq lock for protection.
3757 		 *
3758 		 * Subract ourselves as a sleeper while attempting to add
3759 		 * our count.
3760 		 */
3761 		atomic_subtract_32(&zone->uz_sleepers, 1);
3762 		old = atomic_fetchadd_64(&zone->uz_items,
3763 		    -(UZ_ITEMS_SLEEPER - count));
3764 		/* We're no longer a sleeper. */
3765 		old -= UZ_ITEMS_SLEEPER;
3766 
3767 		/*
3768 		 * If we're still at the limit, restart.  Notably do not
3769 		 * block on other sleepers.  Cache the max value to protect
3770 		 * against changes via sysctl.
3771 		 */
3772 		total = UZ_ITEMS_COUNT(old);
3773 		max = zone->uz_max_items;
3774 		if (total >= max)
3775 			continue;
3776 		/* Truncate if necessary, otherwise wake other sleepers. */
3777 		if (total + count > max) {
3778 			zone_free_limit(zone, total + count - max);
3779 			count = max - total;
3780 		} else if (total + count < max && UZ_ITEMS_SLEEPERS(old) != 0)
3781 			wakeup_one(&zone->uz_max_items);
3782 
3783 		return (count);
3784 	}
3785 }
3786 
3787 /*
3788  * Allocate 'count' items from our max_items limit.  Returns the number
3789  * available.  If M_NOWAIT is not specified it will sleep until at least
3790  * one item can be allocated.
3791  */
3792 static int
3793 zone_alloc_limit(uma_zone_t zone, int count, int flags)
3794 {
3795 	uint64_t old;
3796 	uint64_t max;
3797 
3798 	max = zone->uz_max_items;
3799 	MPASS(max > 0);
3800 
3801 	/*
3802 	 * We expect normal allocations to succeed with a simple
3803 	 * fetchadd.
3804 	 */
3805 	old = atomic_fetchadd_64(&zone->uz_items, count);
3806 	if (__predict_true(old + count <= max))
3807 		return (count);
3808 
3809 	/*
3810 	 * If we had some items and no sleepers just return the
3811 	 * truncated value.  We have to release the excess space
3812 	 * though because that may wake sleepers who weren't woken
3813 	 * because we were temporarily over the limit.
3814 	 */
3815 	if (old < max) {
3816 		zone_free_limit(zone, (old + count) - max);
3817 		return (max - old);
3818 	}
3819 	return (zone_alloc_limit_hard(zone, count, flags));
3820 }
3821 
3822 /*
3823  * Free a number of items back to the limit.
3824  */
3825 static void
3826 zone_free_limit(uma_zone_t zone, int count)
3827 {
3828 	uint64_t old;
3829 
3830 	MPASS(count > 0);
3831 
3832 	/*
3833 	 * In the common case we either have no sleepers or
3834 	 * are still over the limit and can just return.
3835 	 */
3836 	old = atomic_fetchadd_64(&zone->uz_items, -count);
3837 	if (__predict_true(UZ_ITEMS_SLEEPERS(old) == 0 ||
3838 	   UZ_ITEMS_COUNT(old) - count >= zone->uz_max_items))
3839 		return;
3840 
3841 	/*
3842 	 * Moderate the rate of wakeups.  Sleepers will continue
3843 	 * to generate wakeups if necessary.
3844 	 */
3845 	wakeup_one(&zone->uz_max_items);
3846 }
3847 
3848 static uma_bucket_t
3849 zone_alloc_bucket(uma_zone_t zone, void *udata, int domain, int flags)
3850 {
3851 	uma_bucket_t bucket;
3852 	int maxbucket, cnt;
3853 
3854 	CTR3(KTR_UMA, "zone_alloc_bucket zone %s(%p) domain %d", zone->uz_name,
3855 	    zone, domain);
3856 
3857 	/* Avoid allocs targeting empty domains. */
3858 	if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
3859 		domain = UMA_ANYDOMAIN;
3860 	if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
3861 		domain = UMA_ANYDOMAIN;
3862 
3863 	if (zone->uz_max_items > 0)
3864 		maxbucket = zone_alloc_limit(zone, zone->uz_bucket_size,
3865 		    M_NOWAIT);
3866 	else
3867 		maxbucket = zone->uz_bucket_size;
3868 	if (maxbucket == 0)
3869 		return (false);
3870 
3871 	/* Don't wait for buckets, preserve caller's NOVM setting. */
3872 	bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
3873 	if (bucket == NULL) {
3874 		cnt = 0;
3875 		goto out;
3876 	}
3877 
3878 	bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
3879 	    MIN(maxbucket, bucket->ub_entries), domain, flags);
3880 
3881 	/*
3882 	 * Initialize the memory if necessary.
3883 	 */
3884 	if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
3885 		int i;
3886 
3887 		for (i = 0; i < bucket->ub_cnt; i++)
3888 			if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size,
3889 			    flags) != 0)
3890 				break;
3891 		/*
3892 		 * If we couldn't initialize the whole bucket, put the
3893 		 * rest back onto the freelist.
3894 		 */
3895 		if (i != bucket->ub_cnt) {
3896 			zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
3897 			    bucket->ub_cnt - i);
3898 #ifdef INVARIANTS
3899 			bzero(&bucket->ub_bucket[i],
3900 			    sizeof(void *) * (bucket->ub_cnt - i));
3901 #endif
3902 			bucket->ub_cnt = i;
3903 		}
3904 	}
3905 
3906 	cnt = bucket->ub_cnt;
3907 	if (bucket->ub_cnt == 0) {
3908 		bucket_free(zone, bucket, udata);
3909 		counter_u64_add(zone->uz_fails, 1);
3910 		bucket = NULL;
3911 	}
3912 out:
3913 	if (zone->uz_max_items > 0 && cnt < maxbucket)
3914 		zone_free_limit(zone, maxbucket - cnt);
3915 
3916 	return (bucket);
3917 }
3918 
3919 /*
3920  * Allocates a single item from a zone.
3921  *
3922  * Arguments
3923  *	zone   The zone to alloc for.
3924  *	udata  The data to be passed to the constructor.
3925  *	domain The domain to allocate from or UMA_ANYDOMAIN.
3926  *	flags  M_WAITOK, M_NOWAIT, M_ZERO.
3927  *
3928  * Returns
3929  *	NULL if there is no memory and M_NOWAIT is set
3930  *	An item if successful
3931  */
3932 
3933 static void *
3934 zone_alloc_item(uma_zone_t zone, void *udata, int domain, int flags)
3935 {
3936 	void *item;
3937 
3938 	if (zone->uz_max_items > 0 && zone_alloc_limit(zone, 1, flags) == 0)
3939 		return (NULL);
3940 
3941 	/* Avoid allocs targeting empty domains. */
3942 	if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
3943 		domain = UMA_ANYDOMAIN;
3944 
3945 	if (zone->uz_import(zone->uz_arg, &item, 1, domain, flags) != 1)
3946 		goto fail_cnt;
3947 
3948 	/*
3949 	 * We have to call both the zone's init (not the keg's init)
3950 	 * and the zone's ctor.  This is because the item is going from
3951 	 * a keg slab directly to the user, and the user is expecting it
3952 	 * to be both zone-init'd as well as zone-ctor'd.
3953 	 */
3954 	if (zone->uz_init != NULL) {
3955 		if (zone->uz_init(item, zone->uz_size, flags) != 0) {
3956 			zone_free_item(zone, item, udata, SKIP_FINI | SKIP_CNT);
3957 			goto fail_cnt;
3958 		}
3959 	}
3960 	item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata, flags,
3961 	    item);
3962 	if (item == NULL)
3963 		goto fail;
3964 
3965 	counter_u64_add(zone->uz_allocs, 1);
3966 	CTR3(KTR_UMA, "zone_alloc_item item %p from %s(%p)", item,
3967 	    zone->uz_name, zone);
3968 
3969 	return (item);
3970 
3971 fail_cnt:
3972 	counter_u64_add(zone->uz_fails, 1);
3973 fail:
3974 	if (zone->uz_max_items > 0)
3975 		zone_free_limit(zone, 1);
3976 	CTR2(KTR_UMA, "zone_alloc_item failed from %s(%p)",
3977 	    zone->uz_name, zone);
3978 
3979 	return (NULL);
3980 }
3981 
3982 /* See uma.h */
3983 void
3984 uma_zfree_smr(uma_zone_t zone, void *item)
3985 {
3986 	uma_cache_t cache;
3987 	uma_cache_bucket_t bucket;
3988 	int itemdomain, uz_flags;
3989 
3990 #ifdef UMA_ZALLOC_DEBUG
3991 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
3992 	    ("uma_zfree_smr: called with non-SMR zone.\n"));
3993 	KASSERT(item != NULL, ("uma_zfree_smr: Called with NULL pointer."));
3994 	SMR_ASSERT_NOT_ENTERED(zone->uz_smr);
3995 	if (uma_zfree_debug(zone, item, NULL) == EJUSTRETURN)
3996 		return;
3997 #endif
3998 	cache = &zone->uz_cpu[curcpu];
3999 	uz_flags = cache_uz_flags(cache);
4000 	itemdomain = 0;
4001 #ifdef NUMA
4002 	if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4003 		itemdomain = _vm_phys_domain(pmap_kextract((vm_offset_t)item));
4004 #endif
4005 	critical_enter();
4006 	do {
4007 		cache = &zone->uz_cpu[curcpu];
4008 		/* SMR Zones must free to the free bucket. */
4009 		bucket = &cache->uc_freebucket;
4010 #ifdef NUMA
4011 		if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4012 		    PCPU_GET(domain) != itemdomain) {
4013 			bucket = &cache->uc_crossbucket;
4014 		}
4015 #endif
4016 		if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4017 			cache_bucket_push(cache, bucket, item);
4018 			critical_exit();
4019 			return;
4020 		}
4021 	} while (cache_free(zone, cache, NULL, item, itemdomain));
4022 	critical_exit();
4023 
4024 	/*
4025 	 * If nothing else caught this, we'll just do an internal free.
4026 	 */
4027 	zone_free_item(zone, item, NULL, SKIP_NONE);
4028 }
4029 
4030 /* See uma.h */
4031 void
4032 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
4033 {
4034 	uma_cache_t cache;
4035 	uma_cache_bucket_t bucket;
4036 	int itemdomain, uz_flags;
4037 
4038 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
4039 	random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
4040 
4041 	CTR2(KTR_UMA, "uma_zfree_arg zone %s(%p)", zone->uz_name, zone);
4042 
4043 #ifdef UMA_ZALLOC_DEBUG
4044 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
4045 	    ("uma_zfree_arg: called with SMR zone.\n"));
4046 	if (uma_zfree_debug(zone, item, udata) == EJUSTRETURN)
4047 		return;
4048 #endif
4049         /* uma_zfree(..., NULL) does nothing, to match free(9). */
4050         if (item == NULL)
4051                 return;
4052 
4053 	/*
4054 	 * We are accessing the per-cpu cache without a critical section to
4055 	 * fetch size and flags.  This is acceptable, if we are preempted we
4056 	 * will simply read another cpu's line.
4057 	 */
4058 	cache = &zone->uz_cpu[curcpu];
4059 	uz_flags = cache_uz_flags(cache);
4060 	if (UMA_ALWAYS_CTORDTOR ||
4061 	    __predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0))
4062 		item_dtor(zone, item, cache_uz_size(cache), udata, SKIP_NONE);
4063 
4064 	/*
4065 	 * The race here is acceptable.  If we miss it we'll just have to wait
4066 	 * a little longer for the limits to be reset.
4067 	 */
4068 	if (__predict_false(uz_flags & UMA_ZFLAG_LIMIT)) {
4069 		if (zone->uz_sleepers > 0)
4070 			goto zfree_item;
4071 	}
4072 
4073 	/*
4074 	 * If possible, free to the per-CPU cache.  There are two
4075 	 * requirements for safe access to the per-CPU cache: (1) the thread
4076 	 * accessing the cache must not be preempted or yield during access,
4077 	 * and (2) the thread must not migrate CPUs without switching which
4078 	 * cache it accesses.  We rely on a critical section to prevent
4079 	 * preemption and migration.  We release the critical section in
4080 	 * order to acquire the zone mutex if we are unable to free to the
4081 	 * current cache; when we re-acquire the critical section, we must
4082 	 * detect and handle migration if it has occurred.
4083 	 */
4084 	itemdomain = 0;
4085 #ifdef NUMA
4086 	if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4087 		itemdomain = _vm_phys_domain(pmap_kextract((vm_offset_t)item));
4088 #endif
4089 	critical_enter();
4090 	do {
4091 		cache = &zone->uz_cpu[curcpu];
4092 		/*
4093 		 * Try to free into the allocbucket first to give LIFO
4094 		 * ordering for cache-hot datastructures.  Spill over
4095 		 * into the freebucket if necessary.  Alloc will swap
4096 		 * them if one runs dry.
4097 		 */
4098 		bucket = &cache->uc_allocbucket;
4099 #ifdef NUMA
4100 		if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4101 		    PCPU_GET(domain) != itemdomain) {
4102 			bucket = &cache->uc_crossbucket;
4103 		} else
4104 #endif
4105 		if (bucket->ucb_cnt == bucket->ucb_entries &&
4106 		   cache->uc_freebucket.ucb_cnt <
4107 		   cache->uc_freebucket.ucb_entries)
4108 			cache_bucket_swap(&cache->uc_freebucket,
4109 			    &cache->uc_allocbucket);
4110 		if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4111 			cache_bucket_push(cache, bucket, item);
4112 			critical_exit();
4113 			return;
4114 		}
4115 	} while (cache_free(zone, cache, udata, item, itemdomain));
4116 	critical_exit();
4117 
4118 	/*
4119 	 * If nothing else caught this, we'll just do an internal free.
4120 	 */
4121 zfree_item:
4122 	zone_free_item(zone, item, udata, SKIP_DTOR);
4123 }
4124 
4125 #ifdef NUMA
4126 /*
4127  * sort crossdomain free buckets to domain correct buckets and cache
4128  * them.
4129  */
4130 static void
4131 zone_free_cross(uma_zone_t zone, uma_bucket_t bucket, void *udata)
4132 {
4133 	struct uma_bucketlist fullbuckets;
4134 	uma_zone_domain_t zdom;
4135 	uma_bucket_t b;
4136 	smr_seq_t seq;
4137 	void *item;
4138 	int domain;
4139 
4140 	CTR3(KTR_UMA,
4141 	    "uma_zfree: zone %s(%p) draining cross bucket %p",
4142 	    zone->uz_name, zone, bucket);
4143 
4144 	/*
4145 	 * It is possible for buckets to arrive here out of order so we fetch
4146 	 * the current smr seq rather than accepting the bucket's.
4147 	 */
4148 	seq = SMR_SEQ_INVALID;
4149 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
4150 		seq = smr_advance(zone->uz_smr);
4151 
4152 	/*
4153 	 * To avoid having ndomain * ndomain buckets for sorting we have a
4154 	 * lock on the current crossfree bucket.  A full matrix with
4155 	 * per-domain locking could be used if necessary.
4156 	 */
4157 	STAILQ_INIT(&fullbuckets);
4158 	ZONE_CROSS_LOCK(zone);
4159 	while (bucket->ub_cnt > 0) {
4160 		item = bucket->ub_bucket[bucket->ub_cnt - 1];
4161 		domain = _vm_phys_domain(pmap_kextract((vm_offset_t)item));
4162 		zdom = ZDOM_GET(zone, domain);
4163 		if (zdom->uzd_cross == NULL) {
4164 			zdom->uzd_cross = bucket_alloc(zone, udata, M_NOWAIT);
4165 			if (zdom->uzd_cross == NULL)
4166 				break;
4167 		}
4168 		b = zdom->uzd_cross;
4169 		b->ub_bucket[b->ub_cnt++] = item;
4170 		b->ub_seq = seq;
4171 		if (b->ub_cnt == b->ub_entries) {
4172 			STAILQ_INSERT_HEAD(&fullbuckets, b, ub_link);
4173 			zdom->uzd_cross = NULL;
4174 		}
4175 		bucket->ub_cnt--;
4176 	}
4177 	ZONE_CROSS_UNLOCK(zone);
4178 	if (bucket->ub_cnt == 0)
4179 		bucket->ub_seq = SMR_SEQ_INVALID;
4180 	bucket_free(zone, bucket, udata);
4181 
4182 	while ((b = STAILQ_FIRST(&fullbuckets)) != NULL) {
4183 		STAILQ_REMOVE_HEAD(&fullbuckets, ub_link);
4184 		domain = _vm_phys_domain(pmap_kextract(
4185 		    (vm_offset_t)b->ub_bucket[0]));
4186 		zone_put_bucket(zone, domain, b, udata, true);
4187 	}
4188 }
4189 #endif
4190 
4191 static void
4192 zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata,
4193     int itemdomain, bool ws)
4194 {
4195 
4196 #ifdef NUMA
4197 	/*
4198 	 * Buckets coming from the wrong domain will be entirely for the
4199 	 * only other domain on two domain systems.  In this case we can
4200 	 * simply cache them.  Otherwise we need to sort them back to
4201 	 * correct domains.
4202 	 */
4203 	if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4204 	    vm_ndomains > 2 && PCPU_GET(domain) != itemdomain) {
4205 		zone_free_cross(zone, bucket, udata);
4206 		return;
4207 	}
4208 #endif
4209 
4210 	/*
4211 	 * Attempt to save the bucket in the zone's domain bucket cache.
4212 	 */
4213 	CTR3(KTR_UMA,
4214 	    "uma_zfree: zone %s(%p) putting bucket %p on free list",
4215 	    zone->uz_name, zone, bucket);
4216 	/* ub_cnt is pointing to the last free item */
4217 	if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
4218 		itemdomain = zone_domain_lowest(zone, itemdomain);
4219 	zone_put_bucket(zone, itemdomain, bucket, udata, ws);
4220 }
4221 
4222 /*
4223  * Populate a free or cross bucket for the current cpu cache.  Free any
4224  * existing full bucket either to the zone cache or back to the slab layer.
4225  *
4226  * Enters and returns in a critical section.  false return indicates that
4227  * we can not satisfy this free in the cache layer.  true indicates that
4228  * the caller should retry.
4229  */
4230 static __noinline bool
4231 cache_free(uma_zone_t zone, uma_cache_t cache, void *udata, void *item,
4232     int itemdomain)
4233 {
4234 	uma_cache_bucket_t cbucket;
4235 	uma_bucket_t newbucket, bucket;
4236 
4237 	CRITICAL_ASSERT(curthread);
4238 
4239 	if (zone->uz_bucket_size == 0)
4240 		return false;
4241 
4242 	cache = &zone->uz_cpu[curcpu];
4243 	newbucket = NULL;
4244 
4245 	/*
4246 	 * FIRSTTOUCH domains need to free to the correct zdom.  When
4247 	 * enabled this is the zdom of the item.   The bucket is the
4248 	 * cross bucket if the current domain and itemdomain do not match.
4249 	 */
4250 	cbucket = &cache->uc_freebucket;
4251 #ifdef NUMA
4252 	if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4253 		if (PCPU_GET(domain) != itemdomain) {
4254 			cbucket = &cache->uc_crossbucket;
4255 			if (cbucket->ucb_cnt != 0)
4256 				counter_u64_add(zone->uz_xdomain,
4257 				    cbucket->ucb_cnt);
4258 		}
4259 	}
4260 #endif
4261 	bucket = cache_bucket_unload(cbucket);
4262 	KASSERT(bucket == NULL || bucket->ub_cnt == bucket->ub_entries,
4263 	    ("cache_free: Entered with non-full free bucket."));
4264 
4265 	/* We are no longer associated with this CPU. */
4266 	critical_exit();
4267 
4268 	/*
4269 	 * Don't let SMR zones operate without a free bucket.  Force
4270 	 * a synchronize and re-use this one.  We will only degrade
4271 	 * to a synchronize every bucket_size items rather than every
4272 	 * item if we fail to allocate a bucket.
4273 	 */
4274 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0) {
4275 		if (bucket != NULL)
4276 			bucket->ub_seq = smr_advance(zone->uz_smr);
4277 		newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4278 		if (newbucket == NULL && bucket != NULL) {
4279 			bucket_drain(zone, bucket);
4280 			newbucket = bucket;
4281 			bucket = NULL;
4282 		}
4283 	} else if (!bucketdisable)
4284 		newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4285 
4286 	if (bucket != NULL)
4287 		zone_free_bucket(zone, bucket, udata, itemdomain, true);
4288 
4289 	critical_enter();
4290 	if ((bucket = newbucket) == NULL)
4291 		return (false);
4292 	cache = &zone->uz_cpu[curcpu];
4293 #ifdef NUMA
4294 	/*
4295 	 * Check to see if we should be populating the cross bucket.  If it
4296 	 * is already populated we will fall through and attempt to populate
4297 	 * the free bucket.
4298 	 */
4299 	if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4300 		if (PCPU_GET(domain) != itemdomain &&
4301 		    cache->uc_crossbucket.ucb_bucket == NULL) {
4302 			cache_bucket_load_cross(cache, bucket);
4303 			return (true);
4304 		}
4305 	}
4306 #endif
4307 	/*
4308 	 * We may have lost the race to fill the bucket or switched CPUs.
4309 	 */
4310 	if (cache->uc_freebucket.ucb_bucket != NULL) {
4311 		critical_exit();
4312 		bucket_free(zone, bucket, udata);
4313 		critical_enter();
4314 	} else
4315 		cache_bucket_load_free(cache, bucket);
4316 
4317 	return (true);
4318 }
4319 
4320 void
4321 uma_zfree_domain(uma_zone_t zone, void *item, void *udata)
4322 {
4323 
4324 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
4325 	random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
4326 
4327 	CTR2(KTR_UMA, "uma_zfree_domain zone %s(%p)", zone->uz_name, zone);
4328 
4329 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
4330 	    ("uma_zfree_domain: called with spinlock or critical section held"));
4331 
4332         /* uma_zfree(..., NULL) does nothing, to match free(9). */
4333         if (item == NULL)
4334                 return;
4335 	zone_free_item(zone, item, udata, SKIP_NONE);
4336 }
4337 
4338 static void
4339 slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item)
4340 {
4341 	uma_keg_t keg;
4342 	uma_domain_t dom;
4343 	int freei;
4344 
4345 	keg = zone->uz_keg;
4346 	KEG_LOCK_ASSERT(keg, slab->us_domain);
4347 
4348 	/* Do we need to remove from any lists? */
4349 	dom = &keg->uk_domain[slab->us_domain];
4350 	if (slab->us_freecount + 1 == keg->uk_ipers) {
4351 		LIST_REMOVE(slab, us_link);
4352 		LIST_INSERT_HEAD(&dom->ud_free_slab, slab, us_link);
4353 		dom->ud_free_slabs++;
4354 	} else if (slab->us_freecount == 0) {
4355 		LIST_REMOVE(slab, us_link);
4356 		LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
4357 	}
4358 
4359 	/* Slab management. */
4360 	freei = slab_item_index(slab, keg, item);
4361 	BIT_SET(keg->uk_ipers, freei, &slab->us_free);
4362 	slab->us_freecount++;
4363 
4364 	/* Keg statistics. */
4365 	dom->ud_free_items++;
4366 }
4367 
4368 static void
4369 zone_release(void *arg, void **bucket, int cnt)
4370 {
4371 	struct mtx *lock;
4372 	uma_zone_t zone;
4373 	uma_slab_t slab;
4374 	uma_keg_t keg;
4375 	uint8_t *mem;
4376 	void *item;
4377 	int i;
4378 
4379 	zone = arg;
4380 	keg = zone->uz_keg;
4381 	lock = NULL;
4382 	if (__predict_false((zone->uz_flags & UMA_ZFLAG_HASH) != 0))
4383 		lock = KEG_LOCK(keg, 0);
4384 	for (i = 0; i < cnt; i++) {
4385 		item = bucket[i];
4386 		if (__predict_true((zone->uz_flags & UMA_ZFLAG_VTOSLAB) != 0)) {
4387 			slab = vtoslab((vm_offset_t)item);
4388 		} else {
4389 			mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
4390 			if ((zone->uz_flags & UMA_ZFLAG_HASH) != 0)
4391 				slab = hash_sfind(&keg->uk_hash, mem);
4392 			else
4393 				slab = (uma_slab_t)(mem + keg->uk_pgoff);
4394 		}
4395 		if (lock != KEG_LOCKPTR(keg, slab->us_domain)) {
4396 			if (lock != NULL)
4397 				mtx_unlock(lock);
4398 			lock = KEG_LOCK(keg, slab->us_domain);
4399 		}
4400 		slab_free_item(zone, slab, item);
4401 	}
4402 	if (lock != NULL)
4403 		mtx_unlock(lock);
4404 }
4405 
4406 /*
4407  * Frees a single item to any zone.
4408  *
4409  * Arguments:
4410  *	zone   The zone to free to
4411  *	item   The item we're freeing
4412  *	udata  User supplied data for the dtor
4413  *	skip   Skip dtors and finis
4414  */
4415 static __noinline void
4416 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
4417 {
4418 
4419 	/*
4420 	 * If a free is sent directly to an SMR zone we have to
4421 	 * synchronize immediately because the item can instantly
4422 	 * be reallocated. This should only happen in degenerate
4423 	 * cases when no memory is available for per-cpu caches.
4424 	 */
4425 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0 && skip == SKIP_NONE)
4426 		smr_synchronize(zone->uz_smr);
4427 
4428 	item_dtor(zone, item, zone->uz_size, udata, skip);
4429 
4430 	if (skip < SKIP_FINI && zone->uz_fini)
4431 		zone->uz_fini(item, zone->uz_size);
4432 
4433 	zone->uz_release(zone->uz_arg, &item, 1);
4434 
4435 	if (skip & SKIP_CNT)
4436 		return;
4437 
4438 	counter_u64_add(zone->uz_frees, 1);
4439 
4440 	if (zone->uz_max_items > 0)
4441 		zone_free_limit(zone, 1);
4442 }
4443 
4444 /* See uma.h */
4445 int
4446 uma_zone_set_max(uma_zone_t zone, int nitems)
4447 {
4448 	struct uma_bucket_zone *ubz;
4449 	int count;
4450 
4451 	/*
4452 	 * XXX This can misbehave if the zone has any allocations with
4453 	 * no limit and a limit is imposed.  There is currently no
4454 	 * way to clear a limit.
4455 	 */
4456 	ZONE_LOCK(zone);
4457 	ubz = bucket_zone_max(zone, nitems);
4458 	count = ubz != NULL ? ubz->ubz_entries : 0;
4459 	zone->uz_bucket_size_max = zone->uz_bucket_size = count;
4460 	if (zone->uz_bucket_size_min > zone->uz_bucket_size_max)
4461 		zone->uz_bucket_size_min = zone->uz_bucket_size_max;
4462 	zone->uz_max_items = nitems;
4463 	zone->uz_flags |= UMA_ZFLAG_LIMIT;
4464 	zone_update_caches(zone);
4465 	/* We may need to wake waiters. */
4466 	wakeup(&zone->uz_max_items);
4467 	ZONE_UNLOCK(zone);
4468 
4469 	return (nitems);
4470 }
4471 
4472 /* See uma.h */
4473 void
4474 uma_zone_set_maxcache(uma_zone_t zone, int nitems)
4475 {
4476 	struct uma_bucket_zone *ubz;
4477 	int bpcpu;
4478 
4479 	ZONE_LOCK(zone);
4480 	ubz = bucket_zone_max(zone, nitems);
4481 	if (ubz != NULL) {
4482 		bpcpu = 2;
4483 		if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4484 			/* Count the cross-domain bucket. */
4485 			bpcpu++;
4486 		nitems -= ubz->ubz_entries * bpcpu * mp_ncpus;
4487 		zone->uz_bucket_size_max = ubz->ubz_entries;
4488 	} else {
4489 		zone->uz_bucket_size_max = zone->uz_bucket_size = 0;
4490 	}
4491 	if (zone->uz_bucket_size_min > zone->uz_bucket_size_max)
4492 		zone->uz_bucket_size_min = zone->uz_bucket_size_max;
4493 	zone->uz_bucket_max = nitems / vm_ndomains;
4494 	ZONE_UNLOCK(zone);
4495 }
4496 
4497 /* See uma.h */
4498 int
4499 uma_zone_get_max(uma_zone_t zone)
4500 {
4501 	int nitems;
4502 
4503 	nitems = atomic_load_64(&zone->uz_max_items);
4504 
4505 	return (nitems);
4506 }
4507 
4508 /* See uma.h */
4509 void
4510 uma_zone_set_warning(uma_zone_t zone, const char *warning)
4511 {
4512 
4513 	ZONE_ASSERT_COLD(zone);
4514 	zone->uz_warning = warning;
4515 }
4516 
4517 /* See uma.h */
4518 void
4519 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction)
4520 {
4521 
4522 	ZONE_ASSERT_COLD(zone);
4523 	TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone);
4524 }
4525 
4526 /* See uma.h */
4527 int
4528 uma_zone_get_cur(uma_zone_t zone)
4529 {
4530 	int64_t nitems;
4531 	u_int i;
4532 
4533 	nitems = 0;
4534 	if (zone->uz_allocs != EARLY_COUNTER && zone->uz_frees != EARLY_COUNTER)
4535 		nitems = counter_u64_fetch(zone->uz_allocs) -
4536 		    counter_u64_fetch(zone->uz_frees);
4537 	CPU_FOREACH(i)
4538 		nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs) -
4539 		    atomic_load_64(&zone->uz_cpu[i].uc_frees);
4540 
4541 	return (nitems < 0 ? 0 : nitems);
4542 }
4543 
4544 static uint64_t
4545 uma_zone_get_allocs(uma_zone_t zone)
4546 {
4547 	uint64_t nitems;
4548 	u_int i;
4549 
4550 	nitems = 0;
4551 	if (zone->uz_allocs != EARLY_COUNTER)
4552 		nitems = counter_u64_fetch(zone->uz_allocs);
4553 	CPU_FOREACH(i)
4554 		nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs);
4555 
4556 	return (nitems);
4557 }
4558 
4559 static uint64_t
4560 uma_zone_get_frees(uma_zone_t zone)
4561 {
4562 	uint64_t nitems;
4563 	u_int i;
4564 
4565 	nitems = 0;
4566 	if (zone->uz_frees != EARLY_COUNTER)
4567 		nitems = counter_u64_fetch(zone->uz_frees);
4568 	CPU_FOREACH(i)
4569 		nitems += atomic_load_64(&zone->uz_cpu[i].uc_frees);
4570 
4571 	return (nitems);
4572 }
4573 
4574 #ifdef INVARIANTS
4575 /* Used only for KEG_ASSERT_COLD(). */
4576 static uint64_t
4577 uma_keg_get_allocs(uma_keg_t keg)
4578 {
4579 	uma_zone_t z;
4580 	uint64_t nitems;
4581 
4582 	nitems = 0;
4583 	LIST_FOREACH(z, &keg->uk_zones, uz_link)
4584 		nitems += uma_zone_get_allocs(z);
4585 
4586 	return (nitems);
4587 }
4588 #endif
4589 
4590 /* See uma.h */
4591 void
4592 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
4593 {
4594 	uma_keg_t keg;
4595 
4596 	KEG_GET(zone, keg);
4597 	KEG_ASSERT_COLD(keg);
4598 	keg->uk_init = uminit;
4599 }
4600 
4601 /* See uma.h */
4602 void
4603 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
4604 {
4605 	uma_keg_t keg;
4606 
4607 	KEG_GET(zone, keg);
4608 	KEG_ASSERT_COLD(keg);
4609 	keg->uk_fini = fini;
4610 }
4611 
4612 /* See uma.h */
4613 void
4614 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
4615 {
4616 
4617 	ZONE_ASSERT_COLD(zone);
4618 	zone->uz_init = zinit;
4619 }
4620 
4621 /* See uma.h */
4622 void
4623 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
4624 {
4625 
4626 	ZONE_ASSERT_COLD(zone);
4627 	zone->uz_fini = zfini;
4628 }
4629 
4630 /* See uma.h */
4631 void
4632 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
4633 {
4634 	uma_keg_t keg;
4635 
4636 	KEG_GET(zone, keg);
4637 	KEG_ASSERT_COLD(keg);
4638 	keg->uk_freef = freef;
4639 }
4640 
4641 /* See uma.h */
4642 void
4643 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
4644 {
4645 	uma_keg_t keg;
4646 
4647 	KEG_GET(zone, keg);
4648 	KEG_ASSERT_COLD(keg);
4649 	keg->uk_allocf = allocf;
4650 }
4651 
4652 /* See uma.h */
4653 void
4654 uma_zone_set_smr(uma_zone_t zone, smr_t smr)
4655 {
4656 
4657 	ZONE_ASSERT_COLD(zone);
4658 
4659 	KASSERT(smr != NULL, ("Got NULL smr"));
4660 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
4661 	    ("zone %p (%s) already uses SMR", zone, zone->uz_name));
4662 	zone->uz_flags |= UMA_ZONE_SMR;
4663 	zone->uz_smr = smr;
4664 	zone_update_caches(zone);
4665 }
4666 
4667 smr_t
4668 uma_zone_get_smr(uma_zone_t zone)
4669 {
4670 
4671 	return (zone->uz_smr);
4672 }
4673 
4674 /* See uma.h */
4675 void
4676 uma_zone_reserve(uma_zone_t zone, int items)
4677 {
4678 	uma_keg_t keg;
4679 
4680 	KEG_GET(zone, keg);
4681 	KEG_ASSERT_COLD(keg);
4682 	keg->uk_reserve = items;
4683 }
4684 
4685 /* See uma.h */
4686 int
4687 uma_zone_reserve_kva(uma_zone_t zone, int count)
4688 {
4689 	uma_keg_t keg;
4690 	vm_offset_t kva;
4691 	u_int pages;
4692 
4693 	KEG_GET(zone, keg);
4694 	KEG_ASSERT_COLD(keg);
4695 	ZONE_ASSERT_COLD(zone);
4696 
4697 	pages = howmany(count, keg->uk_ipers) * keg->uk_ppera;
4698 
4699 #ifdef UMA_MD_SMALL_ALLOC
4700 	if (keg->uk_ppera > 1) {
4701 #else
4702 	if (1) {
4703 #endif
4704 		kva = kva_alloc((vm_size_t)pages * PAGE_SIZE);
4705 		if (kva == 0)
4706 			return (0);
4707 	} else
4708 		kva = 0;
4709 
4710 	MPASS(keg->uk_kva == 0);
4711 	keg->uk_kva = kva;
4712 	keg->uk_offset = 0;
4713 	zone->uz_max_items = pages * keg->uk_ipers;
4714 #ifdef UMA_MD_SMALL_ALLOC
4715 	keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
4716 #else
4717 	keg->uk_allocf = noobj_alloc;
4718 #endif
4719 	keg->uk_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
4720 	zone->uz_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
4721 	zone_update_caches(zone);
4722 
4723 	return (1);
4724 }
4725 
4726 /* See uma.h */
4727 void
4728 uma_prealloc(uma_zone_t zone, int items)
4729 {
4730 	struct vm_domainset_iter di;
4731 	uma_domain_t dom;
4732 	uma_slab_t slab;
4733 	uma_keg_t keg;
4734 	int aflags, domain, slabs;
4735 
4736 	KEG_GET(zone, keg);
4737 	slabs = howmany(items, keg->uk_ipers);
4738 	while (slabs-- > 0) {
4739 		aflags = M_NOWAIT;
4740 		vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
4741 		    &aflags);
4742 		for (;;) {
4743 			slab = keg_alloc_slab(keg, zone, domain, M_WAITOK,
4744 			    aflags);
4745 			if (slab != NULL) {
4746 				dom = &keg->uk_domain[slab->us_domain];
4747 				/*
4748 				 * keg_alloc_slab() always returns a slab on the
4749 				 * partial list.
4750 				 */
4751 				LIST_REMOVE(slab, us_link);
4752 				LIST_INSERT_HEAD(&dom->ud_free_slab, slab,
4753 				    us_link);
4754 				dom->ud_free_slabs++;
4755 				KEG_UNLOCK(keg, slab->us_domain);
4756 				break;
4757 			}
4758 			if (vm_domainset_iter_policy(&di, &domain) != 0)
4759 				vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask);
4760 		}
4761 	}
4762 }
4763 
4764 /*
4765  * Returns a snapshot of memory consumption in bytes.
4766  */
4767 size_t
4768 uma_zone_memory(uma_zone_t zone)
4769 {
4770 	size_t sz;
4771 	int i;
4772 
4773 	sz = 0;
4774 	if (zone->uz_flags & UMA_ZFLAG_CACHE) {
4775 		for (i = 0; i < vm_ndomains; i++)
4776 			sz += ZDOM_GET(zone, i)->uzd_nitems;
4777 		return (sz * zone->uz_size);
4778 	}
4779 	for (i = 0; i < vm_ndomains; i++)
4780 		sz += zone->uz_keg->uk_domain[i].ud_pages;
4781 
4782 	return (sz * PAGE_SIZE);
4783 }
4784 
4785 /* See uma.h */
4786 void
4787 uma_reclaim(int req)
4788 {
4789 
4790 	CTR0(KTR_UMA, "UMA: vm asked us to release pages!");
4791 	sx_xlock(&uma_reclaim_lock);
4792 	bucket_enable();
4793 
4794 	switch (req) {
4795 	case UMA_RECLAIM_TRIM:
4796 		zone_foreach(zone_trim, NULL);
4797 		break;
4798 	case UMA_RECLAIM_DRAIN:
4799 	case UMA_RECLAIM_DRAIN_CPU:
4800 		zone_foreach(zone_drain, NULL);
4801 		if (req == UMA_RECLAIM_DRAIN_CPU) {
4802 			pcpu_cache_drain_safe(NULL);
4803 			zone_foreach(zone_drain, NULL);
4804 		}
4805 		break;
4806 	default:
4807 		panic("unhandled reclamation request %d", req);
4808 	}
4809 
4810 	/*
4811 	 * Some slabs may have been freed but this zone will be visited early
4812 	 * we visit again so that we can free pages that are empty once other
4813 	 * zones are drained.  We have to do the same for buckets.
4814 	 */
4815 	zone_drain(slabzones[0], NULL);
4816 	zone_drain(slabzones[1], NULL);
4817 	bucket_zone_drain();
4818 	sx_xunlock(&uma_reclaim_lock);
4819 }
4820 
4821 static volatile int uma_reclaim_needed;
4822 
4823 void
4824 uma_reclaim_wakeup(void)
4825 {
4826 
4827 	if (atomic_fetchadd_int(&uma_reclaim_needed, 1) == 0)
4828 		wakeup(uma_reclaim);
4829 }
4830 
4831 void
4832 uma_reclaim_worker(void *arg __unused)
4833 {
4834 
4835 	for (;;) {
4836 		sx_xlock(&uma_reclaim_lock);
4837 		while (atomic_load_int(&uma_reclaim_needed) == 0)
4838 			sx_sleep(uma_reclaim, &uma_reclaim_lock, PVM, "umarcl",
4839 			    hz);
4840 		sx_xunlock(&uma_reclaim_lock);
4841 		EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM);
4842 		uma_reclaim(UMA_RECLAIM_DRAIN_CPU);
4843 		atomic_store_int(&uma_reclaim_needed, 0);
4844 		/* Don't fire more than once per-second. */
4845 		pause("umarclslp", hz);
4846 	}
4847 }
4848 
4849 /* See uma.h */
4850 void
4851 uma_zone_reclaim(uma_zone_t zone, int req)
4852 {
4853 
4854 	switch (req) {
4855 	case UMA_RECLAIM_TRIM:
4856 		zone_trim(zone, NULL);
4857 		break;
4858 	case UMA_RECLAIM_DRAIN:
4859 		zone_drain(zone, NULL);
4860 		break;
4861 	case UMA_RECLAIM_DRAIN_CPU:
4862 		pcpu_cache_drain_safe(zone);
4863 		zone_drain(zone, NULL);
4864 		break;
4865 	default:
4866 		panic("unhandled reclamation request %d", req);
4867 	}
4868 }
4869 
4870 /* See uma.h */
4871 int
4872 uma_zone_exhausted(uma_zone_t zone)
4873 {
4874 
4875 	return (atomic_load_32(&zone->uz_sleepers) > 0);
4876 }
4877 
4878 unsigned long
4879 uma_limit(void)
4880 {
4881 
4882 	return (uma_kmem_limit);
4883 }
4884 
4885 void
4886 uma_set_limit(unsigned long limit)
4887 {
4888 
4889 	uma_kmem_limit = limit;
4890 }
4891 
4892 unsigned long
4893 uma_size(void)
4894 {
4895 
4896 	return (atomic_load_long(&uma_kmem_total));
4897 }
4898 
4899 long
4900 uma_avail(void)
4901 {
4902 
4903 	return (uma_kmem_limit - uma_size());
4904 }
4905 
4906 #ifdef DDB
4907 /*
4908  * Generate statistics across both the zone and its per-cpu cache's.  Return
4909  * desired statistics if the pointer is non-NULL for that statistic.
4910  *
4911  * Note: does not update the zone statistics, as it can't safely clear the
4912  * per-CPU cache statistic.
4913  *
4914  */
4915 static void
4916 uma_zone_sumstat(uma_zone_t z, long *cachefreep, uint64_t *allocsp,
4917     uint64_t *freesp, uint64_t *sleepsp, uint64_t *xdomainp)
4918 {
4919 	uma_cache_t cache;
4920 	uint64_t allocs, frees, sleeps, xdomain;
4921 	int cachefree, cpu;
4922 
4923 	allocs = frees = sleeps = xdomain = 0;
4924 	cachefree = 0;
4925 	CPU_FOREACH(cpu) {
4926 		cache = &z->uz_cpu[cpu];
4927 		cachefree += cache->uc_allocbucket.ucb_cnt;
4928 		cachefree += cache->uc_freebucket.ucb_cnt;
4929 		xdomain += cache->uc_crossbucket.ucb_cnt;
4930 		cachefree += cache->uc_crossbucket.ucb_cnt;
4931 		allocs += cache->uc_allocs;
4932 		frees += cache->uc_frees;
4933 	}
4934 	allocs += counter_u64_fetch(z->uz_allocs);
4935 	frees += counter_u64_fetch(z->uz_frees);
4936 	xdomain += counter_u64_fetch(z->uz_xdomain);
4937 	sleeps += z->uz_sleeps;
4938 	if (cachefreep != NULL)
4939 		*cachefreep = cachefree;
4940 	if (allocsp != NULL)
4941 		*allocsp = allocs;
4942 	if (freesp != NULL)
4943 		*freesp = frees;
4944 	if (sleepsp != NULL)
4945 		*sleepsp = sleeps;
4946 	if (xdomainp != NULL)
4947 		*xdomainp = xdomain;
4948 }
4949 #endif /* DDB */
4950 
4951 static int
4952 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
4953 {
4954 	uma_keg_t kz;
4955 	uma_zone_t z;
4956 	int count;
4957 
4958 	count = 0;
4959 	rw_rlock(&uma_rwlock);
4960 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
4961 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
4962 			count++;
4963 	}
4964 	LIST_FOREACH(z, &uma_cachezones, uz_link)
4965 		count++;
4966 
4967 	rw_runlock(&uma_rwlock);
4968 	return (sysctl_handle_int(oidp, &count, 0, req));
4969 }
4970 
4971 static void
4972 uma_vm_zone_stats(struct uma_type_header *uth, uma_zone_t z, struct sbuf *sbuf,
4973     struct uma_percpu_stat *ups, bool internal)
4974 {
4975 	uma_zone_domain_t zdom;
4976 	uma_cache_t cache;
4977 	int i;
4978 
4979 
4980 	for (i = 0; i < vm_ndomains; i++) {
4981 		zdom = ZDOM_GET(z, i);
4982 		uth->uth_zone_free += zdom->uzd_nitems;
4983 	}
4984 	uth->uth_allocs = counter_u64_fetch(z->uz_allocs);
4985 	uth->uth_frees = counter_u64_fetch(z->uz_frees);
4986 	uth->uth_fails = counter_u64_fetch(z->uz_fails);
4987 	uth->uth_xdomain = counter_u64_fetch(z->uz_xdomain);
4988 	uth->uth_sleeps = z->uz_sleeps;
4989 
4990 	for (i = 0; i < mp_maxid + 1; i++) {
4991 		bzero(&ups[i], sizeof(*ups));
4992 		if (internal || CPU_ABSENT(i))
4993 			continue;
4994 		cache = &z->uz_cpu[i];
4995 		ups[i].ups_cache_free += cache->uc_allocbucket.ucb_cnt;
4996 		ups[i].ups_cache_free += cache->uc_freebucket.ucb_cnt;
4997 		ups[i].ups_cache_free += cache->uc_crossbucket.ucb_cnt;
4998 		ups[i].ups_allocs = cache->uc_allocs;
4999 		ups[i].ups_frees = cache->uc_frees;
5000 	}
5001 }
5002 
5003 static int
5004 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
5005 {
5006 	struct uma_stream_header ush;
5007 	struct uma_type_header uth;
5008 	struct uma_percpu_stat *ups;
5009 	struct sbuf sbuf;
5010 	uma_keg_t kz;
5011 	uma_zone_t z;
5012 	uint64_t items;
5013 	uint32_t kfree, pages;
5014 	int count, error, i;
5015 
5016 	error = sysctl_wire_old_buffer(req, 0);
5017 	if (error != 0)
5018 		return (error);
5019 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
5020 	sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL);
5021 	ups = malloc((mp_maxid + 1) * sizeof(*ups), M_TEMP, M_WAITOK);
5022 
5023 	count = 0;
5024 	rw_rlock(&uma_rwlock);
5025 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
5026 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
5027 			count++;
5028 	}
5029 
5030 	LIST_FOREACH(z, &uma_cachezones, uz_link)
5031 		count++;
5032 
5033 	/*
5034 	 * Insert stream header.
5035 	 */
5036 	bzero(&ush, sizeof(ush));
5037 	ush.ush_version = UMA_STREAM_VERSION;
5038 	ush.ush_maxcpus = (mp_maxid + 1);
5039 	ush.ush_count = count;
5040 	(void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
5041 
5042 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
5043 		kfree = pages = 0;
5044 		for (i = 0; i < vm_ndomains; i++) {
5045 			kfree += kz->uk_domain[i].ud_free_items;
5046 			pages += kz->uk_domain[i].ud_pages;
5047 		}
5048 		LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5049 			bzero(&uth, sizeof(uth));
5050 			strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5051 			uth.uth_align = kz->uk_align;
5052 			uth.uth_size = kz->uk_size;
5053 			uth.uth_rsize = kz->uk_rsize;
5054 			if (z->uz_max_items > 0) {
5055 				items = UZ_ITEMS_COUNT(z->uz_items);
5056 				uth.uth_pages = (items / kz->uk_ipers) *
5057 					kz->uk_ppera;
5058 			} else
5059 				uth.uth_pages = pages;
5060 			uth.uth_maxpages = (z->uz_max_items / kz->uk_ipers) *
5061 			    kz->uk_ppera;
5062 			uth.uth_limit = z->uz_max_items;
5063 			uth.uth_keg_free = kfree;
5064 
5065 			/*
5066 			 * A zone is secondary is it is not the first entry
5067 			 * on the keg's zone list.
5068 			 */
5069 			if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
5070 			    (LIST_FIRST(&kz->uk_zones) != z))
5071 				uth.uth_zone_flags = UTH_ZONE_SECONDARY;
5072 			uma_vm_zone_stats(&uth, z, &sbuf, ups,
5073 			    kz->uk_flags & UMA_ZFLAG_INTERNAL);
5074 			(void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5075 			for (i = 0; i < mp_maxid + 1; i++)
5076 				(void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5077 		}
5078 	}
5079 	LIST_FOREACH(z, &uma_cachezones, uz_link) {
5080 		bzero(&uth, sizeof(uth));
5081 		strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5082 		uth.uth_size = z->uz_size;
5083 		uma_vm_zone_stats(&uth, z, &sbuf, ups, false);
5084 		(void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5085 		for (i = 0; i < mp_maxid + 1; i++)
5086 			(void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5087 	}
5088 
5089 	rw_runlock(&uma_rwlock);
5090 	error = sbuf_finish(&sbuf);
5091 	sbuf_delete(&sbuf);
5092 	free(ups, M_TEMP);
5093 	return (error);
5094 }
5095 
5096 int
5097 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS)
5098 {
5099 	uma_zone_t zone = *(uma_zone_t *)arg1;
5100 	int error, max;
5101 
5102 	max = uma_zone_get_max(zone);
5103 	error = sysctl_handle_int(oidp, &max, 0, req);
5104 	if (error || !req->newptr)
5105 		return (error);
5106 
5107 	uma_zone_set_max(zone, max);
5108 
5109 	return (0);
5110 }
5111 
5112 int
5113 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS)
5114 {
5115 	uma_zone_t zone;
5116 	int cur;
5117 
5118 	/*
5119 	 * Some callers want to add sysctls for global zones that
5120 	 * may not yet exist so they pass a pointer to a pointer.
5121 	 */
5122 	if (arg2 == 0)
5123 		zone = *(uma_zone_t *)arg1;
5124 	else
5125 		zone = arg1;
5126 	cur = uma_zone_get_cur(zone);
5127 	return (sysctl_handle_int(oidp, &cur, 0, req));
5128 }
5129 
5130 static int
5131 sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS)
5132 {
5133 	uma_zone_t zone = arg1;
5134 	uint64_t cur;
5135 
5136 	cur = uma_zone_get_allocs(zone);
5137 	return (sysctl_handle_64(oidp, &cur, 0, req));
5138 }
5139 
5140 static int
5141 sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS)
5142 {
5143 	uma_zone_t zone = arg1;
5144 	uint64_t cur;
5145 
5146 	cur = uma_zone_get_frees(zone);
5147 	return (sysctl_handle_64(oidp, &cur, 0, req));
5148 }
5149 
5150 static int
5151 sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS)
5152 {
5153 	struct sbuf sbuf;
5154 	uma_zone_t zone = arg1;
5155 	int error;
5156 
5157 	sbuf_new_for_sysctl(&sbuf, NULL, 0, req);
5158 	if (zone->uz_flags != 0)
5159 		sbuf_printf(&sbuf, "0x%b", zone->uz_flags, PRINT_UMA_ZFLAGS);
5160 	else
5161 		sbuf_printf(&sbuf, "0");
5162 	error = sbuf_finish(&sbuf);
5163 	sbuf_delete(&sbuf);
5164 
5165 	return (error);
5166 }
5167 
5168 static int
5169 sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS)
5170 {
5171 	uma_keg_t keg = arg1;
5172 	int avail, effpct, total;
5173 
5174 	total = keg->uk_ppera * PAGE_SIZE;
5175 	if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0)
5176 		total += slabzone(keg->uk_ipers)->uz_keg->uk_rsize;
5177 	/*
5178 	 * We consider the client's requested size and alignment here, not the
5179 	 * real size determination uk_rsize, because we also adjust the real
5180 	 * size for internal implementation reasons (max bitset size).
5181 	 */
5182 	avail = keg->uk_ipers * roundup2(keg->uk_size, keg->uk_align + 1);
5183 	if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
5184 		avail *= mp_maxid + 1;
5185 	effpct = 100 * avail / total;
5186 	return (sysctl_handle_int(oidp, &effpct, 0, req));
5187 }
5188 
5189 static int
5190 sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS)
5191 {
5192 	uma_zone_t zone = arg1;
5193 	uint64_t cur;
5194 
5195 	cur = UZ_ITEMS_COUNT(atomic_load_64(&zone->uz_items));
5196 	return (sysctl_handle_64(oidp, &cur, 0, req));
5197 }
5198 
5199 #ifdef INVARIANTS
5200 static uma_slab_t
5201 uma_dbg_getslab(uma_zone_t zone, void *item)
5202 {
5203 	uma_slab_t slab;
5204 	uma_keg_t keg;
5205 	uint8_t *mem;
5206 
5207 	/*
5208 	 * It is safe to return the slab here even though the
5209 	 * zone is unlocked because the item's allocation state
5210 	 * essentially holds a reference.
5211 	 */
5212 	mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
5213 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5214 		return (NULL);
5215 	if (zone->uz_flags & UMA_ZFLAG_VTOSLAB)
5216 		return (vtoslab((vm_offset_t)mem));
5217 	keg = zone->uz_keg;
5218 	if ((keg->uk_flags & UMA_ZFLAG_HASH) == 0)
5219 		return ((uma_slab_t)(mem + keg->uk_pgoff));
5220 	KEG_LOCK(keg, 0);
5221 	slab = hash_sfind(&keg->uk_hash, mem);
5222 	KEG_UNLOCK(keg, 0);
5223 
5224 	return (slab);
5225 }
5226 
5227 static bool
5228 uma_dbg_zskip(uma_zone_t zone, void *mem)
5229 {
5230 
5231 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5232 		return (true);
5233 
5234 	return (uma_dbg_kskip(zone->uz_keg, mem));
5235 }
5236 
5237 static bool
5238 uma_dbg_kskip(uma_keg_t keg, void *mem)
5239 {
5240 	uintptr_t idx;
5241 
5242 	if (dbg_divisor == 0)
5243 		return (true);
5244 
5245 	if (dbg_divisor == 1)
5246 		return (false);
5247 
5248 	idx = (uintptr_t)mem >> PAGE_SHIFT;
5249 	if (keg->uk_ipers > 1) {
5250 		idx *= keg->uk_ipers;
5251 		idx += ((uintptr_t)mem & PAGE_MASK) / keg->uk_rsize;
5252 	}
5253 
5254 	if ((idx / dbg_divisor) * dbg_divisor != idx) {
5255 		counter_u64_add(uma_skip_cnt, 1);
5256 		return (true);
5257 	}
5258 	counter_u64_add(uma_dbg_cnt, 1);
5259 
5260 	return (false);
5261 }
5262 
5263 /*
5264  * Set up the slab's freei data such that uma_dbg_free can function.
5265  *
5266  */
5267 static void
5268 uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item)
5269 {
5270 	uma_keg_t keg;
5271 	int freei;
5272 
5273 	if (slab == NULL) {
5274 		slab = uma_dbg_getslab(zone, item);
5275 		if (slab == NULL)
5276 			panic("uma: item %p did not belong to zone %s\n",
5277 			    item, zone->uz_name);
5278 	}
5279 	keg = zone->uz_keg;
5280 	freei = slab_item_index(slab, keg, item);
5281 
5282 	if (BIT_ISSET(keg->uk_ipers, freei, slab_dbg_bits(slab, keg)))
5283 		panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)\n",
5284 		    item, zone, zone->uz_name, slab, freei);
5285 	BIT_SET_ATOMIC(keg->uk_ipers, freei, slab_dbg_bits(slab, keg));
5286 }
5287 
5288 /*
5289  * Verifies freed addresses.  Checks for alignment, valid slab membership
5290  * and duplicate frees.
5291  *
5292  */
5293 static void
5294 uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item)
5295 {
5296 	uma_keg_t keg;
5297 	int freei;
5298 
5299 	if (slab == NULL) {
5300 		slab = uma_dbg_getslab(zone, item);
5301 		if (slab == NULL)
5302 			panic("uma: Freed item %p did not belong to zone %s\n",
5303 			    item, zone->uz_name);
5304 	}
5305 	keg = zone->uz_keg;
5306 	freei = slab_item_index(slab, keg, item);
5307 
5308 	if (freei >= keg->uk_ipers)
5309 		panic("Invalid free of %p from zone %p(%s) slab %p(%d)\n",
5310 		    item, zone, zone->uz_name, slab, freei);
5311 
5312 	if (slab_item(slab, keg, freei) != item)
5313 		panic("Unaligned free of %p from zone %p(%s) slab %p(%d)\n",
5314 		    item, zone, zone->uz_name, slab, freei);
5315 
5316 	if (!BIT_ISSET(keg->uk_ipers, freei, slab_dbg_bits(slab, keg)))
5317 		panic("Duplicate free of %p from zone %p(%s) slab %p(%d)\n",
5318 		    item, zone, zone->uz_name, slab, freei);
5319 
5320 	BIT_CLR_ATOMIC(keg->uk_ipers, freei, slab_dbg_bits(slab, keg));
5321 }
5322 #endif /* INVARIANTS */
5323 
5324 #ifdef DDB
5325 static int64_t
5326 get_uma_stats(uma_keg_t kz, uma_zone_t z, uint64_t *allocs, uint64_t *used,
5327     uint64_t *sleeps, long *cachefree, uint64_t *xdomain)
5328 {
5329 	uint64_t frees;
5330 	int i;
5331 
5332 	if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
5333 		*allocs = counter_u64_fetch(z->uz_allocs);
5334 		frees = counter_u64_fetch(z->uz_frees);
5335 		*sleeps = z->uz_sleeps;
5336 		*cachefree = 0;
5337 		*xdomain = 0;
5338 	} else
5339 		uma_zone_sumstat(z, cachefree, allocs, &frees, sleeps,
5340 		    xdomain);
5341 	for (i = 0; i < vm_ndomains; i++) {
5342 		*cachefree += ZDOM_GET(z, i)->uzd_nitems;
5343 		if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
5344 		    (LIST_FIRST(&kz->uk_zones) != z)))
5345 			*cachefree += kz->uk_domain[i].ud_free_items;
5346 	}
5347 	*used = *allocs - frees;
5348 	return (((int64_t)*used + *cachefree) * kz->uk_size);
5349 }
5350 
5351 DB_SHOW_COMMAND(uma, db_show_uma)
5352 {
5353 	const char *fmt_hdr, *fmt_entry;
5354 	uma_keg_t kz;
5355 	uma_zone_t z;
5356 	uint64_t allocs, used, sleeps, xdomain;
5357 	long cachefree;
5358 	/* variables for sorting */
5359 	uma_keg_t cur_keg;
5360 	uma_zone_t cur_zone, last_zone;
5361 	int64_t cur_size, last_size, size;
5362 	int ties;
5363 
5364 	/* /i option produces machine-parseable CSV output */
5365 	if (modif[0] == 'i') {
5366 		fmt_hdr = "%s,%s,%s,%s,%s,%s,%s,%s,%s\n";
5367 		fmt_entry = "\"%s\",%ju,%jd,%ld,%ju,%ju,%u,%jd,%ju\n";
5368 	} else {
5369 		fmt_hdr = "%18s %6s %7s %7s %11s %7s %7s %10s %8s\n";
5370 		fmt_entry = "%18s %6ju %7jd %7ld %11ju %7ju %7u %10jd %8ju\n";
5371 	}
5372 
5373 	db_printf(fmt_hdr, "Zone", "Size", "Used", "Free", "Requests",
5374 	    "Sleeps", "Bucket", "Total Mem", "XFree");
5375 
5376 	/* Sort the zones with largest size first. */
5377 	last_zone = NULL;
5378 	last_size = INT64_MAX;
5379 	for (;;) {
5380 		cur_zone = NULL;
5381 		cur_size = -1;
5382 		ties = 0;
5383 		LIST_FOREACH(kz, &uma_kegs, uk_link) {
5384 			LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5385 				/*
5386 				 * In the case of size ties, print out zones
5387 				 * in the order they are encountered.  That is,
5388 				 * when we encounter the most recently output
5389 				 * zone, we have already printed all preceding
5390 				 * ties, and we must print all following ties.
5391 				 */
5392 				if (z == last_zone) {
5393 					ties = 1;
5394 					continue;
5395 				}
5396 				size = get_uma_stats(kz, z, &allocs, &used,
5397 				    &sleeps, &cachefree, &xdomain);
5398 				if (size > cur_size && size < last_size + ties)
5399 				{
5400 					cur_size = size;
5401 					cur_zone = z;
5402 					cur_keg = kz;
5403 				}
5404 			}
5405 		}
5406 		if (cur_zone == NULL)
5407 			break;
5408 
5409 		size = get_uma_stats(cur_keg, cur_zone, &allocs, &used,
5410 		    &sleeps, &cachefree, &xdomain);
5411 		db_printf(fmt_entry, cur_zone->uz_name,
5412 		    (uintmax_t)cur_keg->uk_size, (intmax_t)used, cachefree,
5413 		    (uintmax_t)allocs, (uintmax_t)sleeps,
5414 		    (unsigned)cur_zone->uz_bucket_size, (intmax_t)size,
5415 		    xdomain);
5416 
5417 		if (db_pager_quit)
5418 			return;
5419 		last_zone = cur_zone;
5420 		last_size = cur_size;
5421 	}
5422 }
5423 
5424 DB_SHOW_COMMAND(umacache, db_show_umacache)
5425 {
5426 	uma_zone_t z;
5427 	uint64_t allocs, frees;
5428 	long cachefree;
5429 	int i;
5430 
5431 	db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
5432 	    "Requests", "Bucket");
5433 	LIST_FOREACH(z, &uma_cachezones, uz_link) {
5434 		uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL, NULL);
5435 		for (i = 0; i < vm_ndomains; i++)
5436 			cachefree += ZDOM_GET(z, i)->uzd_nitems;
5437 		db_printf("%18s %8ju %8jd %8ld %12ju %8u\n",
5438 		    z->uz_name, (uintmax_t)z->uz_size,
5439 		    (intmax_t)(allocs - frees), cachefree,
5440 		    (uintmax_t)allocs, z->uz_bucket_size);
5441 		if (db_pager_quit)
5442 			return;
5443 	}
5444 }
5445 #endif	/* DDB */
5446