xref: /freebsd/sys/vm/uma_core.c (revision 1e0ca65a3bb5798a80eccaae58d863f1f08b9ae8)
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/asan.h>
62 #include <sys/bitset.h>
63 #include <sys/domainset.h>
64 #include <sys/eventhandler.h>
65 #include <sys/kernel.h>
66 #include <sys/types.h>
67 #include <sys/limits.h>
68 #include <sys/queue.h>
69 #include <sys/malloc.h>
70 #include <sys/ktr.h>
71 #include <sys/lock.h>
72 #include <sys/msan.h>
73 #include <sys/mutex.h>
74 #include <sys/proc.h>
75 #include <sys/random.h>
76 #include <sys/rwlock.h>
77 #include <sys/sbuf.h>
78 #include <sys/sched.h>
79 #include <sys/sleepqueue.h>
80 #include <sys/smp.h>
81 #include <sys/smr.h>
82 #include <sys/sysctl.h>
83 #include <sys/taskqueue.h>
84 #include <sys/vmmeter.h>
85 
86 #include <vm/vm.h>
87 #include <vm/vm_param.h>
88 #include <vm/vm_domainset.h>
89 #include <vm/vm_object.h>
90 #include <vm/vm_page.h>
91 #include <vm/vm_pageout.h>
92 #include <vm/vm_phys.h>
93 #include <vm/vm_pagequeue.h>
94 #include <vm/vm_map.h>
95 #include <vm/vm_kern.h>
96 #include <vm/vm_extern.h>
97 #include <vm/vm_dumpset.h>
98 #include <vm/uma.h>
99 #include <vm/uma_int.h>
100 #include <vm/uma_dbg.h>
101 
102 #include <ddb/ddb.h>
103 
104 #ifdef DEBUG_MEMGUARD
105 #include <vm/memguard.h>
106 #endif
107 
108 #include <machine/md_var.h>
109 
110 #ifdef INVARIANTS
111 #define	UMA_ALWAYS_CTORDTOR	1
112 #else
113 #define	UMA_ALWAYS_CTORDTOR	0
114 #endif
115 
116 /*
117  * This is the zone and keg from which all zones are spawned.
118  */
119 static uma_zone_t kegs;
120 static uma_zone_t zones;
121 
122 /*
123  * On INVARIANTS builds, the slab contains a second bitset of the same size,
124  * "dbg_bits", which is laid out immediately after us_free.
125  */
126 #ifdef INVARIANTS
127 #define	SLAB_BITSETS	2
128 #else
129 #define	SLAB_BITSETS	1
130 #endif
131 
132 /*
133  * These are the two zones from which all offpage uma_slab_ts are allocated.
134  *
135  * One zone is for slab headers that can represent a larger number of items,
136  * making the slabs themselves more efficient, and the other zone is for
137  * headers that are smaller and represent fewer items, making the headers more
138  * efficient.
139  */
140 #define	SLABZONE_SIZE(setsize)					\
141     (sizeof(struct uma_hash_slab) + BITSET_SIZE(setsize) * SLAB_BITSETS)
142 #define	SLABZONE0_SETSIZE	(PAGE_SIZE / 16)
143 #define	SLABZONE1_SETSIZE	SLAB_MAX_SETSIZE
144 #define	SLABZONE0_SIZE	SLABZONE_SIZE(SLABZONE0_SETSIZE)
145 #define	SLABZONE1_SIZE	SLABZONE_SIZE(SLABZONE1_SETSIZE)
146 static uma_zone_t slabzones[2];
147 
148 /*
149  * The initial hash tables come out of this zone so they can be allocated
150  * prior to malloc coming up.
151  */
152 static uma_zone_t hashzone;
153 
154 /* The boot-time adjusted value for cache line alignment. */
155 int uma_align_cache = 64 - 1;
156 
157 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets");
158 static MALLOC_DEFINE(M_UMA, "UMA", "UMA Misc");
159 
160 /*
161  * Are we allowed to allocate buckets?
162  */
163 static int bucketdisable = 1;
164 
165 /* Linked list of all kegs in the system */
166 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs);
167 
168 /* Linked list of all cache-only zones in the system */
169 static LIST_HEAD(,uma_zone) uma_cachezones =
170     LIST_HEAD_INITIALIZER(uma_cachezones);
171 
172 /*
173  * Mutex for global lists: uma_kegs, uma_cachezones, and the per-keg list of
174  * zones.
175  */
176 static struct rwlock_padalign __exclusive_cache_line uma_rwlock;
177 
178 static struct sx uma_reclaim_lock;
179 
180 /*
181  * First available virual address for boot time allocations.
182  */
183 static vm_offset_t bootstart;
184 static vm_offset_t bootmem;
185 
186 /*
187  * kmem soft limit, initialized by uma_set_limit().  Ensure that early
188  * allocations don't trigger a wakeup of the reclaim thread.
189  */
190 unsigned long uma_kmem_limit = LONG_MAX;
191 SYSCTL_ULONG(_vm, OID_AUTO, uma_kmem_limit, CTLFLAG_RD, &uma_kmem_limit, 0,
192     "UMA kernel memory soft limit");
193 unsigned long uma_kmem_total;
194 SYSCTL_ULONG(_vm, OID_AUTO, uma_kmem_total, CTLFLAG_RD, &uma_kmem_total, 0,
195     "UMA kernel memory usage");
196 
197 /* Is the VM done starting up? */
198 static enum {
199 	BOOT_COLD,
200 	BOOT_KVA,
201 	BOOT_PCPU,
202 	BOOT_RUNNING,
203 	BOOT_SHUTDOWN,
204 } booted = BOOT_COLD;
205 
206 /*
207  * This is the handle used to schedule events that need to happen
208  * outside of the allocation fast path.
209  */
210 static struct callout uma_callout;
211 #define	UMA_TIMEOUT	20		/* Seconds for callout interval. */
212 
213 /*
214  * This structure is passed as the zone ctor arg so that I don't have to create
215  * a special allocation function just for zones.
216  */
217 struct uma_zctor_args {
218 	const char *name;
219 	size_t size;
220 	uma_ctor ctor;
221 	uma_dtor dtor;
222 	uma_init uminit;
223 	uma_fini fini;
224 	uma_import import;
225 	uma_release release;
226 	void *arg;
227 	uma_keg_t keg;
228 	int align;
229 	uint32_t flags;
230 };
231 
232 struct uma_kctor_args {
233 	uma_zone_t zone;
234 	size_t size;
235 	uma_init uminit;
236 	uma_fini fini;
237 	int align;
238 	uint32_t flags;
239 };
240 
241 struct uma_bucket_zone {
242 	uma_zone_t	ubz_zone;
243 	const char	*ubz_name;
244 	int		ubz_entries;	/* Number of items it can hold. */
245 	int		ubz_maxsize;	/* Maximum allocation size per-item. */
246 };
247 
248 /*
249  * Compute the actual number of bucket entries to pack them in power
250  * of two sizes for more efficient space utilization.
251  */
252 #define	BUCKET_SIZE(n)						\
253     (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *))
254 
255 #define	BUCKET_MAX	BUCKET_SIZE(256)
256 
257 struct uma_bucket_zone bucket_zones[] = {
258 	/* Literal bucket sizes. */
259 	{ NULL, "2 Bucket", 2, 4096 },
260 	{ NULL, "4 Bucket", 4, 3072 },
261 	{ NULL, "8 Bucket", 8, 2048 },
262 	{ NULL, "16 Bucket", 16, 1024 },
263 	/* Rounded down power of 2 sizes for efficiency. */
264 	{ NULL, "32 Bucket", BUCKET_SIZE(32), 512 },
265 	{ NULL, "64 Bucket", BUCKET_SIZE(64), 256 },
266 	{ NULL, "128 Bucket", BUCKET_SIZE(128), 128 },
267 	{ NULL, "256 Bucket", BUCKET_SIZE(256), 64 },
268 	{ NULL, NULL, 0}
269 };
270 
271 /*
272  * Flags and enumerations to be passed to internal functions.
273  */
274 enum zfreeskip {
275 	SKIP_NONE =	0,
276 	SKIP_CNT =	0x00000001,
277 	SKIP_DTOR =	0x00010000,
278 	SKIP_FINI =	0x00020000,
279 };
280 
281 /* Prototypes.. */
282 
283 void	uma_startup1(vm_offset_t);
284 void	uma_startup2(void);
285 
286 static void *noobj_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
287 static void *page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
288 static void *pcpu_page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
289 static void *startup_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
290 static void *contig_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
291 static void page_free(void *, vm_size_t, uint8_t);
292 static void pcpu_page_free(void *, vm_size_t, uint8_t);
293 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int, int, int);
294 static void cache_drain(uma_zone_t);
295 static void bucket_drain(uma_zone_t, uma_bucket_t);
296 static void bucket_cache_reclaim(uma_zone_t zone, bool, int);
297 static bool bucket_cache_reclaim_domain(uma_zone_t, bool, bool, int);
298 static int keg_ctor(void *, int, void *, int);
299 static void keg_dtor(void *, int, void *);
300 static void keg_drain(uma_keg_t keg, int domain);
301 static int zone_ctor(void *, int, void *, int);
302 static void zone_dtor(void *, int, void *);
303 static inline void item_dtor(uma_zone_t zone, void *item, int size,
304     void *udata, enum zfreeskip skip);
305 static int zero_init(void *, int, int);
306 static void zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata,
307     int itemdomain, bool ws);
308 static void zone_foreach(void (*zfunc)(uma_zone_t, void *), void *);
309 static void zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *), void *);
310 static void zone_timeout(uma_zone_t zone, void *);
311 static int hash_alloc(struct uma_hash *, u_int);
312 static int hash_expand(struct uma_hash *, struct uma_hash *);
313 static void hash_free(struct uma_hash *hash);
314 static void uma_timeout(void *);
315 static void uma_shutdown(void);
316 static void *zone_alloc_item(uma_zone_t, void *, int, int);
317 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip);
318 static int zone_alloc_limit(uma_zone_t zone, int count, int flags);
319 static void zone_free_limit(uma_zone_t zone, int count);
320 static void bucket_enable(void);
321 static void bucket_init(void);
322 static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int);
323 static void bucket_free(uma_zone_t zone, uma_bucket_t, void *);
324 static void bucket_zone_drain(int domain);
325 static uma_bucket_t zone_alloc_bucket(uma_zone_t, void *, int, int);
326 static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab);
327 static void slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item);
328 static size_t slab_sizeof(int nitems);
329 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit,
330     uma_fini fini, int align, uint32_t flags);
331 static int zone_import(void *, void **, int, int, int);
332 static void zone_release(void *, void **, int);
333 static bool cache_alloc(uma_zone_t, uma_cache_t, void *, int);
334 static bool cache_free(uma_zone_t, uma_cache_t, void *, int);
335 
336 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS);
337 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS);
338 static int sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS);
339 static int sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS);
340 static int sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS);
341 static int sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS);
342 static int sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS);
343 
344 static uint64_t uma_zone_get_allocs(uma_zone_t zone);
345 
346 static SYSCTL_NODE(_vm, OID_AUTO, debug, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
347     "Memory allocation debugging");
348 
349 #ifdef INVARIANTS
350 static uint64_t uma_keg_get_allocs(uma_keg_t zone);
351 static inline struct noslabbits *slab_dbg_bits(uma_slab_t slab, uma_keg_t keg);
352 
353 static bool uma_dbg_kskip(uma_keg_t keg, void *mem);
354 static bool uma_dbg_zskip(uma_zone_t zone, void *mem);
355 static void uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item);
356 static void uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item);
357 
358 static u_int dbg_divisor = 1;
359 SYSCTL_UINT(_vm_debug, OID_AUTO, divisor,
360     CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &dbg_divisor, 0,
361     "Debug & thrash every this item in memory allocator");
362 
363 static counter_u64_t uma_dbg_cnt = EARLY_COUNTER;
364 static counter_u64_t uma_skip_cnt = EARLY_COUNTER;
365 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, trashed, CTLFLAG_RD,
366     &uma_dbg_cnt, "memory items debugged");
367 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, skipped, CTLFLAG_RD,
368     &uma_skip_cnt, "memory items skipped, not debugged");
369 #endif
370 
371 SYSCTL_NODE(_vm, OID_AUTO, uma, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
372     "Universal Memory Allocator");
373 
374 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLFLAG_MPSAFE|CTLTYPE_INT,
375     0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones");
376 
377 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLFLAG_MPSAFE|CTLTYPE_STRUCT,
378     0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats");
379 
380 static int zone_warnings = 1;
381 SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RWTUN, &zone_warnings, 0,
382     "Warn when UMA zones becomes full");
383 
384 static int multipage_slabs = 1;
385 TUNABLE_INT("vm.debug.uma_multipage_slabs", &multipage_slabs);
386 SYSCTL_INT(_vm_debug, OID_AUTO, uma_multipage_slabs,
387     CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &multipage_slabs, 0,
388     "UMA may choose larger slab sizes for better efficiency");
389 
390 /*
391  * Select the slab zone for an offpage slab with the given maximum item count.
392  */
393 static inline uma_zone_t
394 slabzone(int ipers)
395 {
396 
397 	return (slabzones[ipers > SLABZONE0_SETSIZE]);
398 }
399 
400 /*
401  * This routine checks to see whether or not it's safe to enable buckets.
402  */
403 static void
404 bucket_enable(void)
405 {
406 
407 	KASSERT(booted >= BOOT_KVA, ("Bucket enable before init"));
408 	bucketdisable = vm_page_count_min();
409 }
410 
411 /*
412  * Initialize bucket_zones, the array of zones of buckets of various sizes.
413  *
414  * For each zone, calculate the memory required for each bucket, consisting
415  * of the header and an array of pointers.
416  */
417 static void
418 bucket_init(void)
419 {
420 	struct uma_bucket_zone *ubz;
421 	int size;
422 
423 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) {
424 		size = roundup(sizeof(struct uma_bucket), sizeof(void *));
425 		size += sizeof(void *) * ubz->ubz_entries;
426 		ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
427 		    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
428 		    UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET |
429 		    UMA_ZONE_FIRSTTOUCH);
430 	}
431 }
432 
433 /*
434  * Given a desired number of entries for a bucket, return the zone from which
435  * to allocate the bucket.
436  */
437 static struct uma_bucket_zone *
438 bucket_zone_lookup(int entries)
439 {
440 	struct uma_bucket_zone *ubz;
441 
442 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
443 		if (ubz->ubz_entries >= entries)
444 			return (ubz);
445 	ubz--;
446 	return (ubz);
447 }
448 
449 static int
450 bucket_select(int size)
451 {
452 	struct uma_bucket_zone *ubz;
453 
454 	ubz = &bucket_zones[0];
455 	if (size > ubz->ubz_maxsize)
456 		return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1);
457 
458 	for (; ubz->ubz_entries != 0; ubz++)
459 		if (ubz->ubz_maxsize < size)
460 			break;
461 	ubz--;
462 	return (ubz->ubz_entries);
463 }
464 
465 static uma_bucket_t
466 bucket_alloc(uma_zone_t zone, void *udata, int flags)
467 {
468 	struct uma_bucket_zone *ubz;
469 	uma_bucket_t bucket;
470 
471 	/*
472 	 * Don't allocate buckets early in boot.
473 	 */
474 	if (__predict_false(booted < BOOT_KVA))
475 		return (NULL);
476 
477 	/*
478 	 * To limit bucket recursion we store the original zone flags
479 	 * in a cookie passed via zalloc_arg/zfree_arg.  This allows the
480 	 * NOVM flag to persist even through deep recursions.  We also
481 	 * store ZFLAG_BUCKET once we have recursed attempting to allocate
482 	 * a bucket for a bucket zone so we do not allow infinite bucket
483 	 * recursion.  This cookie will even persist to frees of unused
484 	 * buckets via the allocation path or bucket allocations in the
485 	 * free path.
486 	 */
487 	if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
488 		udata = (void *)(uintptr_t)zone->uz_flags;
489 	else {
490 		if ((uintptr_t)udata & UMA_ZFLAG_BUCKET)
491 			return (NULL);
492 		udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET);
493 	}
494 	if (((uintptr_t)udata & UMA_ZONE_VM) != 0)
495 		flags |= M_NOVM;
496 	ubz = bucket_zone_lookup(atomic_load_16(&zone->uz_bucket_size));
497 	if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0)
498 		ubz++;
499 	bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags);
500 	if (bucket) {
501 #ifdef INVARIANTS
502 		bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
503 #endif
504 		bucket->ub_cnt = 0;
505 		bucket->ub_entries = min(ubz->ubz_entries,
506 		    zone->uz_bucket_size_max);
507 		bucket->ub_seq = SMR_SEQ_INVALID;
508 		CTR3(KTR_UMA, "bucket_alloc: zone %s(%p) allocated bucket %p",
509 		    zone->uz_name, zone, bucket);
510 	}
511 
512 	return (bucket);
513 }
514 
515 static void
516 bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata)
517 {
518 	struct uma_bucket_zone *ubz;
519 
520 	if (bucket->ub_cnt != 0)
521 		bucket_drain(zone, bucket);
522 
523 	KASSERT(bucket->ub_cnt == 0,
524 	    ("bucket_free: Freeing a non free bucket."));
525 	KASSERT(bucket->ub_seq == SMR_SEQ_INVALID,
526 	    ("bucket_free: Freeing an SMR bucket."));
527 	if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
528 		udata = (void *)(uintptr_t)zone->uz_flags;
529 	ubz = bucket_zone_lookup(bucket->ub_entries);
530 	uma_zfree_arg(ubz->ubz_zone, bucket, udata);
531 }
532 
533 static void
534 bucket_zone_drain(int domain)
535 {
536 	struct uma_bucket_zone *ubz;
537 
538 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
539 		uma_zone_reclaim_domain(ubz->ubz_zone, UMA_RECLAIM_DRAIN,
540 		    domain);
541 }
542 
543 #ifdef KASAN
544 _Static_assert(UMA_SMALLEST_UNIT % KASAN_SHADOW_SCALE == 0,
545     "Base UMA allocation size not a multiple of the KASAN scale factor");
546 
547 static void
548 kasan_mark_item_valid(uma_zone_t zone, void *item)
549 {
550 	void *pcpu_item;
551 	size_t sz, rsz;
552 	int i;
553 
554 	if ((zone->uz_flags & UMA_ZONE_NOKASAN) != 0)
555 		return;
556 
557 	sz = zone->uz_size;
558 	rsz = roundup2(sz, KASAN_SHADOW_SCALE);
559 	if ((zone->uz_flags & UMA_ZONE_PCPU) == 0) {
560 		kasan_mark(item, sz, rsz, KASAN_GENERIC_REDZONE);
561 	} else {
562 		pcpu_item = zpcpu_base_to_offset(item);
563 		for (i = 0; i <= mp_maxid; i++)
564 			kasan_mark(zpcpu_get_cpu(pcpu_item, i), sz, rsz,
565 			    KASAN_GENERIC_REDZONE);
566 	}
567 }
568 
569 static void
570 kasan_mark_item_invalid(uma_zone_t zone, void *item)
571 {
572 	void *pcpu_item;
573 	size_t sz;
574 	int i;
575 
576 	if ((zone->uz_flags & UMA_ZONE_NOKASAN) != 0)
577 		return;
578 
579 	sz = roundup2(zone->uz_size, KASAN_SHADOW_SCALE);
580 	if ((zone->uz_flags & UMA_ZONE_PCPU) == 0) {
581 		kasan_mark(item, 0, sz, KASAN_UMA_FREED);
582 	} else {
583 		pcpu_item = zpcpu_base_to_offset(item);
584 		for (i = 0; i <= mp_maxid; i++)
585 			kasan_mark(zpcpu_get_cpu(pcpu_item, i), 0, sz,
586 			    KASAN_UMA_FREED);
587 	}
588 }
589 
590 static void
591 kasan_mark_slab_valid(uma_keg_t keg, void *mem)
592 {
593 	size_t sz;
594 
595 	if ((keg->uk_flags & UMA_ZONE_NOKASAN) == 0) {
596 		sz = keg->uk_ppera * PAGE_SIZE;
597 		kasan_mark(mem, sz, sz, 0);
598 	}
599 }
600 
601 static void
602 kasan_mark_slab_invalid(uma_keg_t keg, void *mem)
603 {
604 	size_t sz;
605 
606 	if ((keg->uk_flags & UMA_ZONE_NOKASAN) == 0) {
607 		if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0)
608 			sz = keg->uk_ppera * PAGE_SIZE;
609 		else
610 			sz = keg->uk_pgoff;
611 		kasan_mark(mem, 0, sz, KASAN_UMA_FREED);
612 	}
613 }
614 #else /* !KASAN */
615 static void
616 kasan_mark_item_valid(uma_zone_t zone __unused, void *item __unused)
617 {
618 }
619 
620 static void
621 kasan_mark_item_invalid(uma_zone_t zone __unused, void *item __unused)
622 {
623 }
624 
625 static void
626 kasan_mark_slab_valid(uma_keg_t keg __unused, void *mem __unused)
627 {
628 }
629 
630 static void
631 kasan_mark_slab_invalid(uma_keg_t keg __unused, void *mem __unused)
632 {
633 }
634 #endif /* KASAN */
635 
636 #ifdef KMSAN
637 static inline void
638 kmsan_mark_item_uninitialized(uma_zone_t zone, void *item)
639 {
640 	void *pcpu_item;
641 	size_t sz;
642 	int i;
643 
644 	if ((zone->uz_flags &
645 	    (UMA_ZFLAG_CACHE | UMA_ZONE_SECONDARY | UMA_ZONE_MALLOC)) != 0) {
646 		/*
647 		 * Cache zones should not be instrumented by default, as UMA
648 		 * does not have enough information to do so correctly.
649 		 * Consumers can mark items themselves if it makes sense to do
650 		 * so.
651 		 *
652 		 * Items from secondary zones are initialized by the parent
653 		 * zone and thus cannot safely be marked by UMA.
654 		 *
655 		 * malloc zones are handled directly by malloc(9) and friends,
656 		 * since they can provide more precise origin tracking.
657 		 */
658 		return;
659 	}
660 	if (zone->uz_keg->uk_init != NULL) {
661 		/*
662 		 * By definition, initialized items cannot be marked.  The
663 		 * best we can do is mark items from these zones after they
664 		 * are freed to the keg.
665 		 */
666 		return;
667 	}
668 
669 	sz = zone->uz_size;
670 	if ((zone->uz_flags & UMA_ZONE_PCPU) == 0) {
671 		kmsan_orig(item, sz, KMSAN_TYPE_UMA, KMSAN_RET_ADDR);
672 		kmsan_mark(item, sz, KMSAN_STATE_UNINIT);
673 	} else {
674 		pcpu_item = zpcpu_base_to_offset(item);
675 		for (i = 0; i <= mp_maxid; i++) {
676 			kmsan_orig(zpcpu_get_cpu(pcpu_item, i), sz,
677 			    KMSAN_TYPE_UMA, KMSAN_RET_ADDR);
678 			kmsan_mark(zpcpu_get_cpu(pcpu_item, i), sz,
679 			    KMSAN_STATE_INITED);
680 		}
681 	}
682 }
683 #else /* !KMSAN */
684 static inline void
685 kmsan_mark_item_uninitialized(uma_zone_t zone __unused, void *item __unused)
686 {
687 }
688 #endif /* KMSAN */
689 
690 /*
691  * Acquire the domain lock and record contention.
692  */
693 static uma_zone_domain_t
694 zone_domain_lock(uma_zone_t zone, int domain)
695 {
696 	uma_zone_domain_t zdom;
697 	bool lockfail;
698 
699 	zdom = ZDOM_GET(zone, domain);
700 	lockfail = false;
701 	if (ZDOM_OWNED(zdom))
702 		lockfail = true;
703 	ZDOM_LOCK(zdom);
704 	/* This is unsynchronized.  The counter does not need to be precise. */
705 	if (lockfail && zone->uz_bucket_size < zone->uz_bucket_size_max)
706 		zone->uz_bucket_size++;
707 	return (zdom);
708 }
709 
710 /*
711  * Search for the domain with the least cached items and return it if it
712  * is out of balance with the preferred domain.
713  */
714 static __noinline int
715 zone_domain_lowest(uma_zone_t zone, int pref)
716 {
717 	long least, nitems, prefitems;
718 	int domain;
719 	int i;
720 
721 	prefitems = least = LONG_MAX;
722 	domain = 0;
723 	for (i = 0; i < vm_ndomains; i++) {
724 		nitems = ZDOM_GET(zone, i)->uzd_nitems;
725 		if (nitems < least) {
726 			domain = i;
727 			least = nitems;
728 		}
729 		if (domain == pref)
730 			prefitems = nitems;
731 	}
732 	if (prefitems < least * 2)
733 		return (pref);
734 
735 	return (domain);
736 }
737 
738 /*
739  * Search for the domain with the most cached items and return it or the
740  * preferred domain if it has enough to proceed.
741  */
742 static __noinline int
743 zone_domain_highest(uma_zone_t zone, int pref)
744 {
745 	long most, nitems;
746 	int domain;
747 	int i;
748 
749 	if (ZDOM_GET(zone, pref)->uzd_nitems > BUCKET_MAX)
750 		return (pref);
751 
752 	most = 0;
753 	domain = 0;
754 	for (i = 0; i < vm_ndomains; i++) {
755 		nitems = ZDOM_GET(zone, i)->uzd_nitems;
756 		if (nitems > most) {
757 			domain = i;
758 			most = nitems;
759 		}
760 	}
761 
762 	return (domain);
763 }
764 
765 /*
766  * Set the maximum imax value.
767  */
768 static void
769 zone_domain_imax_set(uma_zone_domain_t zdom, int nitems)
770 {
771 	long old;
772 
773 	old = zdom->uzd_imax;
774 	do {
775 		if (old >= nitems)
776 			return;
777 	} while (atomic_fcmpset_long(&zdom->uzd_imax, &old, nitems) == 0);
778 
779 	/*
780 	 * We are at new maximum, so do the last WSS update for the old
781 	 * bimin and prepare to measure next allocation batch.
782 	 */
783 	if (zdom->uzd_wss < old - zdom->uzd_bimin)
784 		zdom->uzd_wss = old - zdom->uzd_bimin;
785 	zdom->uzd_bimin = nitems;
786 }
787 
788 /*
789  * Attempt to satisfy an allocation by retrieving a full bucket from one of the
790  * zone's caches.  If a bucket is found the zone is not locked on return.
791  */
792 static uma_bucket_t
793 zone_fetch_bucket(uma_zone_t zone, uma_zone_domain_t zdom, bool reclaim)
794 {
795 	uma_bucket_t bucket;
796 	long cnt;
797 	int i;
798 	bool dtor = false;
799 
800 	ZDOM_LOCK_ASSERT(zdom);
801 
802 	if ((bucket = STAILQ_FIRST(&zdom->uzd_buckets)) == NULL)
803 		return (NULL);
804 
805 	/* SMR Buckets can not be re-used until readers expire. */
806 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
807 	    bucket->ub_seq != SMR_SEQ_INVALID) {
808 		if (!smr_poll(zone->uz_smr, bucket->ub_seq, false))
809 			return (NULL);
810 		bucket->ub_seq = SMR_SEQ_INVALID;
811 		dtor = (zone->uz_dtor != NULL) || UMA_ALWAYS_CTORDTOR;
812 		if (STAILQ_NEXT(bucket, ub_link) != NULL)
813 			zdom->uzd_seq = STAILQ_NEXT(bucket, ub_link)->ub_seq;
814 	}
815 	STAILQ_REMOVE_HEAD(&zdom->uzd_buckets, ub_link);
816 
817 	KASSERT(zdom->uzd_nitems >= bucket->ub_cnt,
818 	    ("%s: item count underflow (%ld, %d)",
819 	    __func__, zdom->uzd_nitems, bucket->ub_cnt));
820 	KASSERT(bucket->ub_cnt > 0,
821 	    ("%s: empty bucket in bucket cache", __func__));
822 	zdom->uzd_nitems -= bucket->ub_cnt;
823 
824 	if (reclaim) {
825 		/*
826 		 * Shift the bounds of the current WSS interval to avoid
827 		 * perturbing the estimates.
828 		 */
829 		cnt = lmin(zdom->uzd_bimin, bucket->ub_cnt);
830 		atomic_subtract_long(&zdom->uzd_imax, cnt);
831 		zdom->uzd_bimin -= cnt;
832 		zdom->uzd_imin -= lmin(zdom->uzd_imin, bucket->ub_cnt);
833 		if (zdom->uzd_limin >= bucket->ub_cnt) {
834 			zdom->uzd_limin -= bucket->ub_cnt;
835 		} else {
836 			zdom->uzd_limin = 0;
837 			zdom->uzd_timin = 0;
838 		}
839 	} else if (zdom->uzd_bimin > zdom->uzd_nitems) {
840 		zdom->uzd_bimin = zdom->uzd_nitems;
841 		if (zdom->uzd_imin > zdom->uzd_nitems)
842 			zdom->uzd_imin = zdom->uzd_nitems;
843 	}
844 
845 	ZDOM_UNLOCK(zdom);
846 	if (dtor)
847 		for (i = 0; i < bucket->ub_cnt; i++)
848 			item_dtor(zone, bucket->ub_bucket[i], zone->uz_size,
849 			    NULL, SKIP_NONE);
850 
851 	return (bucket);
852 }
853 
854 /*
855  * Insert a full bucket into the specified cache.  The "ws" parameter indicates
856  * whether the bucket's contents should be counted as part of the zone's working
857  * set.  The bucket may be freed if it exceeds the bucket limit.
858  */
859 static void
860 zone_put_bucket(uma_zone_t zone, int domain, uma_bucket_t bucket, void *udata,
861     const bool ws)
862 {
863 	uma_zone_domain_t zdom;
864 
865 	/* We don't cache empty buckets.  This can happen after a reclaim. */
866 	if (bucket->ub_cnt == 0)
867 		goto out;
868 	zdom = zone_domain_lock(zone, domain);
869 
870 	/*
871 	 * Conditionally set the maximum number of items.
872 	 */
873 	zdom->uzd_nitems += bucket->ub_cnt;
874 	if (__predict_true(zdom->uzd_nitems < zone->uz_bucket_max)) {
875 		if (ws) {
876 			zone_domain_imax_set(zdom, zdom->uzd_nitems);
877 		} else {
878 			/*
879 			 * Shift the bounds of the current WSS interval to
880 			 * avoid perturbing the estimates.
881 			 */
882 			atomic_add_long(&zdom->uzd_imax, bucket->ub_cnt);
883 			zdom->uzd_imin += bucket->ub_cnt;
884 			zdom->uzd_bimin += bucket->ub_cnt;
885 			zdom->uzd_limin += bucket->ub_cnt;
886 		}
887 		if (STAILQ_EMPTY(&zdom->uzd_buckets))
888 			zdom->uzd_seq = bucket->ub_seq;
889 
890 		/*
891 		 * Try to promote reuse of recently used items.  For items
892 		 * protected by SMR, try to defer reuse to minimize polling.
893 		 */
894 		if (bucket->ub_seq == SMR_SEQ_INVALID)
895 			STAILQ_INSERT_HEAD(&zdom->uzd_buckets, bucket, ub_link);
896 		else
897 			STAILQ_INSERT_TAIL(&zdom->uzd_buckets, bucket, ub_link);
898 		ZDOM_UNLOCK(zdom);
899 		return;
900 	}
901 	zdom->uzd_nitems -= bucket->ub_cnt;
902 	ZDOM_UNLOCK(zdom);
903 out:
904 	bucket_free(zone, bucket, udata);
905 }
906 
907 /* Pops an item out of a per-cpu cache bucket. */
908 static inline void *
909 cache_bucket_pop(uma_cache_t cache, uma_cache_bucket_t bucket)
910 {
911 	void *item;
912 
913 	CRITICAL_ASSERT(curthread);
914 
915 	bucket->ucb_cnt--;
916 	item = bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt];
917 #ifdef INVARIANTS
918 	bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = NULL;
919 	KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled."));
920 #endif
921 	cache->uc_allocs++;
922 
923 	return (item);
924 }
925 
926 /* Pushes an item into a per-cpu cache bucket. */
927 static inline void
928 cache_bucket_push(uma_cache_t cache, uma_cache_bucket_t bucket, void *item)
929 {
930 
931 	CRITICAL_ASSERT(curthread);
932 	KASSERT(bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] == NULL,
933 	    ("uma_zfree: Freeing to non free bucket index."));
934 
935 	bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = item;
936 	bucket->ucb_cnt++;
937 	cache->uc_frees++;
938 }
939 
940 /*
941  * Unload a UMA bucket from a per-cpu cache.
942  */
943 static inline uma_bucket_t
944 cache_bucket_unload(uma_cache_bucket_t bucket)
945 {
946 	uma_bucket_t b;
947 
948 	b = bucket->ucb_bucket;
949 	if (b != NULL) {
950 		MPASS(b->ub_entries == bucket->ucb_entries);
951 		b->ub_cnt = bucket->ucb_cnt;
952 		bucket->ucb_bucket = NULL;
953 		bucket->ucb_entries = bucket->ucb_cnt = 0;
954 	}
955 
956 	return (b);
957 }
958 
959 static inline uma_bucket_t
960 cache_bucket_unload_alloc(uma_cache_t cache)
961 {
962 
963 	return (cache_bucket_unload(&cache->uc_allocbucket));
964 }
965 
966 static inline uma_bucket_t
967 cache_bucket_unload_free(uma_cache_t cache)
968 {
969 
970 	return (cache_bucket_unload(&cache->uc_freebucket));
971 }
972 
973 static inline uma_bucket_t
974 cache_bucket_unload_cross(uma_cache_t cache)
975 {
976 
977 	return (cache_bucket_unload(&cache->uc_crossbucket));
978 }
979 
980 /*
981  * Load a bucket into a per-cpu cache bucket.
982  */
983 static inline void
984 cache_bucket_load(uma_cache_bucket_t bucket, uma_bucket_t b)
985 {
986 
987 	CRITICAL_ASSERT(curthread);
988 	MPASS(bucket->ucb_bucket == NULL);
989 	MPASS(b->ub_seq == SMR_SEQ_INVALID);
990 
991 	bucket->ucb_bucket = b;
992 	bucket->ucb_cnt = b->ub_cnt;
993 	bucket->ucb_entries = b->ub_entries;
994 }
995 
996 static inline void
997 cache_bucket_load_alloc(uma_cache_t cache, uma_bucket_t b)
998 {
999 
1000 	cache_bucket_load(&cache->uc_allocbucket, b);
1001 }
1002 
1003 static inline void
1004 cache_bucket_load_free(uma_cache_t cache, uma_bucket_t b)
1005 {
1006 
1007 	cache_bucket_load(&cache->uc_freebucket, b);
1008 }
1009 
1010 #ifdef NUMA
1011 static inline void
1012 cache_bucket_load_cross(uma_cache_t cache, uma_bucket_t b)
1013 {
1014 
1015 	cache_bucket_load(&cache->uc_crossbucket, b);
1016 }
1017 #endif
1018 
1019 /*
1020  * Copy and preserve ucb_spare.
1021  */
1022 static inline void
1023 cache_bucket_copy(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
1024 {
1025 
1026 	b1->ucb_bucket = b2->ucb_bucket;
1027 	b1->ucb_entries = b2->ucb_entries;
1028 	b1->ucb_cnt = b2->ucb_cnt;
1029 }
1030 
1031 /*
1032  * Swap two cache buckets.
1033  */
1034 static inline void
1035 cache_bucket_swap(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
1036 {
1037 	struct uma_cache_bucket b3;
1038 
1039 	CRITICAL_ASSERT(curthread);
1040 
1041 	cache_bucket_copy(&b3, b1);
1042 	cache_bucket_copy(b1, b2);
1043 	cache_bucket_copy(b2, &b3);
1044 }
1045 
1046 /*
1047  * Attempt to fetch a bucket from a zone on behalf of the current cpu cache.
1048  */
1049 static uma_bucket_t
1050 cache_fetch_bucket(uma_zone_t zone, uma_cache_t cache, int domain)
1051 {
1052 	uma_zone_domain_t zdom;
1053 	uma_bucket_t bucket;
1054 	smr_seq_t seq;
1055 
1056 	/*
1057 	 * Avoid the lock if possible.
1058 	 */
1059 	zdom = ZDOM_GET(zone, domain);
1060 	if (zdom->uzd_nitems == 0)
1061 		return (NULL);
1062 
1063 	if ((cache_uz_flags(cache) & UMA_ZONE_SMR) != 0 &&
1064 	    (seq = atomic_load_32(&zdom->uzd_seq)) != SMR_SEQ_INVALID &&
1065 	    !smr_poll(zone->uz_smr, seq, false))
1066 		return (NULL);
1067 
1068 	/*
1069 	 * Check the zone's cache of buckets.
1070 	 */
1071 	zdom = zone_domain_lock(zone, domain);
1072 	if ((bucket = zone_fetch_bucket(zone, zdom, false)) != NULL)
1073 		return (bucket);
1074 	ZDOM_UNLOCK(zdom);
1075 
1076 	return (NULL);
1077 }
1078 
1079 static void
1080 zone_log_warning(uma_zone_t zone)
1081 {
1082 	static const struct timeval warninterval = { 300, 0 };
1083 
1084 	if (!zone_warnings || zone->uz_warning == NULL)
1085 		return;
1086 
1087 	if (ratecheck(&zone->uz_ratecheck, &warninterval))
1088 		printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning);
1089 }
1090 
1091 static inline void
1092 zone_maxaction(uma_zone_t zone)
1093 {
1094 
1095 	if (zone->uz_maxaction.ta_func != NULL)
1096 		taskqueue_enqueue(taskqueue_thread, &zone->uz_maxaction);
1097 }
1098 
1099 /*
1100  * Routine called by timeout which is used to fire off some time interval
1101  * based calculations.  (stats, hash size, etc.)
1102  *
1103  * Arguments:
1104  *	arg   Unused
1105  *
1106  * Returns:
1107  *	Nothing
1108  */
1109 static void
1110 uma_timeout(void *unused)
1111 {
1112 	bucket_enable();
1113 	zone_foreach(zone_timeout, NULL);
1114 
1115 	/* Reschedule this event */
1116 	callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
1117 }
1118 
1119 /*
1120  * Update the working set size estimates for the zone's bucket cache.
1121  * The constants chosen here are somewhat arbitrary.
1122  */
1123 static void
1124 zone_domain_update_wss(uma_zone_domain_t zdom)
1125 {
1126 	long m;
1127 
1128 	ZDOM_LOCK_ASSERT(zdom);
1129 	MPASS(zdom->uzd_imax >= zdom->uzd_nitems);
1130 	MPASS(zdom->uzd_nitems >= zdom->uzd_bimin);
1131 	MPASS(zdom->uzd_bimin >= zdom->uzd_imin);
1132 
1133 	/*
1134 	 * Estimate WSS as modified moving average of biggest allocation
1135 	 * batches for each period over few minutes (UMA_TIMEOUT of 20s).
1136 	 */
1137 	zdom->uzd_wss = lmax(zdom->uzd_wss * 3 / 4,
1138 	    zdom->uzd_imax - zdom->uzd_bimin);
1139 
1140 	/*
1141 	 * Estimate longtime minimum item count as a combination of recent
1142 	 * minimum item count, adjusted by WSS for safety, and the modified
1143 	 * moving average over the last several hours (UMA_TIMEOUT of 20s).
1144 	 * timin measures time since limin tried to go negative, that means
1145 	 * we were dangerously close to or got out of cache.
1146 	 */
1147 	m = zdom->uzd_imin - zdom->uzd_wss;
1148 	if (m >= 0) {
1149 		if (zdom->uzd_limin >= m)
1150 			zdom->uzd_limin = m;
1151 		else
1152 			zdom->uzd_limin = (m + zdom->uzd_limin * 255) / 256;
1153 		zdom->uzd_timin++;
1154 	} else {
1155 		zdom->uzd_limin = 0;
1156 		zdom->uzd_timin = 0;
1157 	}
1158 
1159 	/* To reduce period edge effects on WSS keep half of the imax. */
1160 	atomic_subtract_long(&zdom->uzd_imax,
1161 	    (zdom->uzd_imax - zdom->uzd_nitems + 1) / 2);
1162 	zdom->uzd_imin = zdom->uzd_bimin = zdom->uzd_nitems;
1163 }
1164 
1165 /*
1166  * Routine to perform timeout driven calculations.  This expands the
1167  * hashes and does per cpu statistics aggregation.
1168  *
1169  *  Returns nothing.
1170  */
1171 static void
1172 zone_timeout(uma_zone_t zone, void *unused)
1173 {
1174 	uma_keg_t keg;
1175 	u_int slabs, pages;
1176 
1177 	if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
1178 		goto trim;
1179 
1180 	keg = zone->uz_keg;
1181 
1182 	/*
1183 	 * Hash zones are non-numa by definition so the first domain
1184 	 * is the only one present.
1185 	 */
1186 	KEG_LOCK(keg, 0);
1187 	pages = keg->uk_domain[0].ud_pages;
1188 
1189 	/*
1190 	 * Expand the keg hash table.
1191 	 *
1192 	 * This is done if the number of slabs is larger than the hash size.
1193 	 * What I'm trying to do here is completely reduce collisions.  This
1194 	 * may be a little aggressive.  Should I allow for two collisions max?
1195 	 */
1196 	if ((slabs = pages / keg->uk_ppera) > keg->uk_hash.uh_hashsize) {
1197 		struct uma_hash newhash;
1198 		struct uma_hash oldhash;
1199 		int ret;
1200 
1201 		/*
1202 		 * This is so involved because allocating and freeing
1203 		 * while the keg lock is held will lead to deadlock.
1204 		 * I have to do everything in stages and check for
1205 		 * races.
1206 		 */
1207 		KEG_UNLOCK(keg, 0);
1208 		ret = hash_alloc(&newhash, 1 << fls(slabs));
1209 		KEG_LOCK(keg, 0);
1210 		if (ret) {
1211 			if (hash_expand(&keg->uk_hash, &newhash)) {
1212 				oldhash = keg->uk_hash;
1213 				keg->uk_hash = newhash;
1214 			} else
1215 				oldhash = newhash;
1216 
1217 			KEG_UNLOCK(keg, 0);
1218 			hash_free(&oldhash);
1219 			goto trim;
1220 		}
1221 	}
1222 	KEG_UNLOCK(keg, 0);
1223 
1224 trim:
1225 	/* Trim caches not used for a long time. */
1226 	if ((zone->uz_flags & UMA_ZONE_UNMANAGED) == 0) {
1227 		for (int i = 0; i < vm_ndomains; i++) {
1228 			if (bucket_cache_reclaim_domain(zone, false, false, i) &&
1229 			    (zone->uz_flags & UMA_ZFLAG_CACHE) == 0)
1230 				keg_drain(zone->uz_keg, i);
1231 		}
1232 	}
1233 }
1234 
1235 /*
1236  * Allocate and zero fill the next sized hash table from the appropriate
1237  * backing store.
1238  *
1239  * Arguments:
1240  *	hash  A new hash structure with the old hash size in uh_hashsize
1241  *
1242  * Returns:
1243  *	1 on success and 0 on failure.
1244  */
1245 static int
1246 hash_alloc(struct uma_hash *hash, u_int size)
1247 {
1248 	size_t alloc;
1249 
1250 	KASSERT(powerof2(size), ("hash size must be power of 2"));
1251 	if (size > UMA_HASH_SIZE_INIT)  {
1252 		hash->uh_hashsize = size;
1253 		alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
1254 		hash->uh_slab_hash = malloc(alloc, M_UMAHASH, M_NOWAIT);
1255 	} else {
1256 		alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
1257 		hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
1258 		    UMA_ANYDOMAIN, M_WAITOK);
1259 		hash->uh_hashsize = UMA_HASH_SIZE_INIT;
1260 	}
1261 	if (hash->uh_slab_hash) {
1262 		bzero(hash->uh_slab_hash, alloc);
1263 		hash->uh_hashmask = hash->uh_hashsize - 1;
1264 		return (1);
1265 	}
1266 
1267 	return (0);
1268 }
1269 
1270 /*
1271  * Expands the hash table for HASH zones.  This is done from zone_timeout
1272  * to reduce collisions.  This must not be done in the regular allocation
1273  * path, otherwise, we can recurse on the vm while allocating pages.
1274  *
1275  * Arguments:
1276  *	oldhash  The hash you want to expand
1277  *	newhash  The hash structure for the new table
1278  *
1279  * Returns:
1280  *	Nothing
1281  *
1282  * Discussion:
1283  */
1284 static int
1285 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
1286 {
1287 	uma_hash_slab_t slab;
1288 	u_int hval;
1289 	u_int idx;
1290 
1291 	if (!newhash->uh_slab_hash)
1292 		return (0);
1293 
1294 	if (oldhash->uh_hashsize >= newhash->uh_hashsize)
1295 		return (0);
1296 
1297 	/*
1298 	 * I need to investigate hash algorithms for resizing without a
1299 	 * full rehash.
1300 	 */
1301 
1302 	for (idx = 0; idx < oldhash->uh_hashsize; idx++)
1303 		while (!LIST_EMPTY(&oldhash->uh_slab_hash[idx])) {
1304 			slab = LIST_FIRST(&oldhash->uh_slab_hash[idx]);
1305 			LIST_REMOVE(slab, uhs_hlink);
1306 			hval = UMA_HASH(newhash, slab->uhs_data);
1307 			LIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
1308 			    slab, uhs_hlink);
1309 		}
1310 
1311 	return (1);
1312 }
1313 
1314 /*
1315  * Free the hash bucket to the appropriate backing store.
1316  *
1317  * Arguments:
1318  *	slab_hash  The hash bucket we're freeing
1319  *	hashsize   The number of entries in that hash bucket
1320  *
1321  * Returns:
1322  *	Nothing
1323  */
1324 static void
1325 hash_free(struct uma_hash *hash)
1326 {
1327 	if (hash->uh_slab_hash == NULL)
1328 		return;
1329 	if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
1330 		zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE);
1331 	else
1332 		free(hash->uh_slab_hash, M_UMAHASH);
1333 }
1334 
1335 /*
1336  * Frees all outstanding items in a bucket
1337  *
1338  * Arguments:
1339  *	zone   The zone to free to, must be unlocked.
1340  *	bucket The free/alloc bucket with items.
1341  *
1342  * Returns:
1343  *	Nothing
1344  */
1345 static void
1346 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
1347 {
1348 	int i;
1349 
1350 	if (bucket->ub_cnt == 0)
1351 		return;
1352 
1353 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
1354 	    bucket->ub_seq != SMR_SEQ_INVALID) {
1355 		smr_wait(zone->uz_smr, bucket->ub_seq);
1356 		bucket->ub_seq = SMR_SEQ_INVALID;
1357 		for (i = 0; i < bucket->ub_cnt; i++)
1358 			item_dtor(zone, bucket->ub_bucket[i],
1359 			    zone->uz_size, NULL, SKIP_NONE);
1360 	}
1361 	if (zone->uz_fini)
1362 		for (i = 0; i < bucket->ub_cnt; i++) {
1363 			kasan_mark_item_valid(zone, bucket->ub_bucket[i]);
1364 			zone->uz_fini(bucket->ub_bucket[i], zone->uz_size);
1365 			kasan_mark_item_invalid(zone, bucket->ub_bucket[i]);
1366 		}
1367 	zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt);
1368 	if (zone->uz_max_items > 0)
1369 		zone_free_limit(zone, bucket->ub_cnt);
1370 #ifdef INVARIANTS
1371 	bzero(bucket->ub_bucket, sizeof(void *) * bucket->ub_cnt);
1372 #endif
1373 	bucket->ub_cnt = 0;
1374 }
1375 
1376 /*
1377  * Drains the per cpu caches for a zone.
1378  *
1379  * NOTE: This may only be called while the zone is being torn down, and not
1380  * during normal operation.  This is necessary in order that we do not have
1381  * to migrate CPUs to drain the per-CPU caches.
1382  *
1383  * Arguments:
1384  *	zone     The zone to drain, must be unlocked.
1385  *
1386  * Returns:
1387  *	Nothing
1388  */
1389 static void
1390 cache_drain(uma_zone_t zone)
1391 {
1392 	uma_cache_t cache;
1393 	uma_bucket_t bucket;
1394 	smr_seq_t seq;
1395 	int cpu;
1396 
1397 	/*
1398 	 * XXX: It is safe to not lock the per-CPU caches, because we're
1399 	 * tearing down the zone anyway.  I.e., there will be no further use
1400 	 * of the caches at this point.
1401 	 *
1402 	 * XXX: It would good to be able to assert that the zone is being
1403 	 * torn down to prevent improper use of cache_drain().
1404 	 */
1405 	seq = SMR_SEQ_INVALID;
1406 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
1407 		seq = smr_advance(zone->uz_smr);
1408 	CPU_FOREACH(cpu) {
1409 		cache = &zone->uz_cpu[cpu];
1410 		bucket = cache_bucket_unload_alloc(cache);
1411 		if (bucket != NULL)
1412 			bucket_free(zone, bucket, NULL);
1413 		bucket = cache_bucket_unload_free(cache);
1414 		if (bucket != NULL) {
1415 			bucket->ub_seq = seq;
1416 			bucket_free(zone, bucket, NULL);
1417 		}
1418 		bucket = cache_bucket_unload_cross(cache);
1419 		if (bucket != NULL) {
1420 			bucket->ub_seq = seq;
1421 			bucket_free(zone, bucket, NULL);
1422 		}
1423 	}
1424 	bucket_cache_reclaim(zone, true, UMA_ANYDOMAIN);
1425 }
1426 
1427 static void
1428 cache_shrink(uma_zone_t zone, void *unused)
1429 {
1430 
1431 	if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1432 		return;
1433 
1434 	ZONE_LOCK(zone);
1435 	zone->uz_bucket_size =
1436 	    (zone->uz_bucket_size_min + zone->uz_bucket_size) / 2;
1437 	ZONE_UNLOCK(zone);
1438 }
1439 
1440 static void
1441 cache_drain_safe_cpu(uma_zone_t zone, void *unused)
1442 {
1443 	uma_cache_t cache;
1444 	uma_bucket_t b1, b2, b3;
1445 	int domain;
1446 
1447 	if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1448 		return;
1449 
1450 	b1 = b2 = b3 = NULL;
1451 	critical_enter();
1452 	cache = &zone->uz_cpu[curcpu];
1453 	domain = PCPU_GET(domain);
1454 	b1 = cache_bucket_unload_alloc(cache);
1455 
1456 	/*
1457 	 * Don't flush SMR zone buckets.  This leaves the zone without a
1458 	 * bucket and forces every free to synchronize().
1459 	 */
1460 	if ((zone->uz_flags & UMA_ZONE_SMR) == 0) {
1461 		b2 = cache_bucket_unload_free(cache);
1462 		b3 = cache_bucket_unload_cross(cache);
1463 	}
1464 	critical_exit();
1465 
1466 	if (b1 != NULL)
1467 		zone_free_bucket(zone, b1, NULL, domain, false);
1468 	if (b2 != NULL)
1469 		zone_free_bucket(zone, b2, NULL, domain, false);
1470 	if (b3 != NULL) {
1471 		/* Adjust the domain so it goes to zone_free_cross. */
1472 		domain = (domain + 1) % vm_ndomains;
1473 		zone_free_bucket(zone, b3, NULL, domain, false);
1474 	}
1475 }
1476 
1477 /*
1478  * Safely drain per-CPU caches of a zone(s) to alloc bucket.
1479  * This is an expensive call because it needs to bind to all CPUs
1480  * one by one and enter a critical section on each of them in order
1481  * to safely access their cache buckets.
1482  * Zone lock must not be held on call this function.
1483  */
1484 static void
1485 pcpu_cache_drain_safe(uma_zone_t zone)
1486 {
1487 	int cpu;
1488 
1489 	/*
1490 	 * Polite bucket sizes shrinking was not enough, shrink aggressively.
1491 	 */
1492 	if (zone)
1493 		cache_shrink(zone, NULL);
1494 	else
1495 		zone_foreach(cache_shrink, NULL);
1496 
1497 	CPU_FOREACH(cpu) {
1498 		thread_lock(curthread);
1499 		sched_bind(curthread, cpu);
1500 		thread_unlock(curthread);
1501 
1502 		if (zone)
1503 			cache_drain_safe_cpu(zone, NULL);
1504 		else
1505 			zone_foreach(cache_drain_safe_cpu, NULL);
1506 	}
1507 	thread_lock(curthread);
1508 	sched_unbind(curthread);
1509 	thread_unlock(curthread);
1510 }
1511 
1512 /*
1513  * Reclaim cached buckets from a zone.  All buckets are reclaimed if the caller
1514  * requested a drain, otherwise the per-domain caches are trimmed to either
1515  * estimated working set size.
1516  */
1517 static bool
1518 bucket_cache_reclaim_domain(uma_zone_t zone, bool drain, bool trim, int domain)
1519 {
1520 	uma_zone_domain_t zdom;
1521 	uma_bucket_t bucket;
1522 	long target;
1523 	bool done = false;
1524 
1525 	/*
1526 	 * The cross bucket is partially filled and not part of
1527 	 * the item count.  Reclaim it individually here.
1528 	 */
1529 	zdom = ZDOM_GET(zone, domain);
1530 	if ((zone->uz_flags & UMA_ZONE_SMR) == 0 || drain) {
1531 		ZONE_CROSS_LOCK(zone);
1532 		bucket = zdom->uzd_cross;
1533 		zdom->uzd_cross = NULL;
1534 		ZONE_CROSS_UNLOCK(zone);
1535 		if (bucket != NULL)
1536 			bucket_free(zone, bucket, NULL);
1537 	}
1538 
1539 	/*
1540 	 * If we were asked to drain the zone, we are done only once
1541 	 * this bucket cache is empty.  If trim, we reclaim items in
1542 	 * excess of the zone's estimated working set size.  Multiple
1543 	 * consecutive calls will shrink the WSS and so reclaim more.
1544 	 * If neither drain nor trim, then voluntarily reclaim 1/4
1545 	 * (to reduce first spike) of items not used for a long time.
1546 	 */
1547 	ZDOM_LOCK(zdom);
1548 	zone_domain_update_wss(zdom);
1549 	if (drain)
1550 		target = 0;
1551 	else if (trim)
1552 		target = zdom->uzd_wss;
1553 	else if (zdom->uzd_timin > 900 / UMA_TIMEOUT)
1554 		target = zdom->uzd_nitems - zdom->uzd_limin / 4;
1555 	else {
1556 		ZDOM_UNLOCK(zdom);
1557 		return (done);
1558 	}
1559 	while ((bucket = STAILQ_FIRST(&zdom->uzd_buckets)) != NULL &&
1560 	    zdom->uzd_nitems >= target + bucket->ub_cnt) {
1561 		bucket = zone_fetch_bucket(zone, zdom, true);
1562 		if (bucket == NULL)
1563 			break;
1564 		bucket_free(zone, bucket, NULL);
1565 		done = true;
1566 		ZDOM_LOCK(zdom);
1567 	}
1568 	ZDOM_UNLOCK(zdom);
1569 	return (done);
1570 }
1571 
1572 static void
1573 bucket_cache_reclaim(uma_zone_t zone, bool drain, int domain)
1574 {
1575 	int i;
1576 
1577 	/*
1578 	 * Shrink the zone bucket size to ensure that the per-CPU caches
1579 	 * don't grow too large.
1580 	 */
1581 	if (zone->uz_bucket_size > zone->uz_bucket_size_min)
1582 		zone->uz_bucket_size--;
1583 
1584 	if (domain != UMA_ANYDOMAIN &&
1585 	    (zone->uz_flags & UMA_ZONE_ROUNDROBIN) == 0) {
1586 		bucket_cache_reclaim_domain(zone, drain, true, domain);
1587 	} else {
1588 		for (i = 0; i < vm_ndomains; i++)
1589 			bucket_cache_reclaim_domain(zone, drain, true, i);
1590 	}
1591 }
1592 
1593 static void
1594 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start)
1595 {
1596 	uint8_t *mem;
1597 	size_t size;
1598 	int i;
1599 	uint8_t flags;
1600 
1601 	CTR4(KTR_UMA, "keg_free_slab keg %s(%p) slab %p, returning %d bytes",
1602 	    keg->uk_name, keg, slab, PAGE_SIZE * keg->uk_ppera);
1603 
1604 	mem = slab_data(slab, keg);
1605 	size = PAGE_SIZE * keg->uk_ppera;
1606 
1607 	kasan_mark_slab_valid(keg, mem);
1608 	if (keg->uk_fini != NULL) {
1609 		for (i = start - 1; i > -1; i--)
1610 #ifdef INVARIANTS
1611 		/*
1612 		 * trash_fini implies that dtor was trash_dtor. trash_fini
1613 		 * would check that memory hasn't been modified since free,
1614 		 * which executed trash_dtor.
1615 		 * That's why we need to run uma_dbg_kskip() check here,
1616 		 * albeit we don't make skip check for other init/fini
1617 		 * invocations.
1618 		 */
1619 		if (!uma_dbg_kskip(keg, slab_item(slab, keg, i)) ||
1620 		    keg->uk_fini != trash_fini)
1621 #endif
1622 			keg->uk_fini(slab_item(slab, keg, i), keg->uk_size);
1623 	}
1624 	flags = slab->us_flags;
1625 	if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) {
1626 		zone_free_item(slabzone(keg->uk_ipers), slab_tohashslab(slab),
1627 		    NULL, SKIP_NONE);
1628 	}
1629 	keg->uk_freef(mem, size, flags);
1630 	uma_total_dec(size);
1631 }
1632 
1633 static void
1634 keg_drain_domain(uma_keg_t keg, int domain)
1635 {
1636 	struct slabhead freeslabs;
1637 	uma_domain_t dom;
1638 	uma_slab_t slab, tmp;
1639 	uint32_t i, stofree, stokeep, partial;
1640 
1641 	dom = &keg->uk_domain[domain];
1642 	LIST_INIT(&freeslabs);
1643 
1644 	CTR4(KTR_UMA, "keg_drain %s(%p) domain %d free items: %u",
1645 	    keg->uk_name, keg, domain, dom->ud_free_items);
1646 
1647 	KEG_LOCK(keg, domain);
1648 
1649 	/*
1650 	 * Are the free items in partially allocated slabs sufficient to meet
1651 	 * the reserve? If not, compute the number of fully free slabs that must
1652 	 * be kept.
1653 	 */
1654 	partial = dom->ud_free_items - dom->ud_free_slabs * keg->uk_ipers;
1655 	if (partial < keg->uk_reserve) {
1656 		stokeep = min(dom->ud_free_slabs,
1657 		    howmany(keg->uk_reserve - partial, keg->uk_ipers));
1658 	} else {
1659 		stokeep = 0;
1660 	}
1661 	stofree = dom->ud_free_slabs - stokeep;
1662 
1663 	/*
1664 	 * Partition the free slabs into two sets: those that must be kept in
1665 	 * order to maintain the reserve, and those that may be released back to
1666 	 * the system.  Since one set may be much larger than the other,
1667 	 * populate the smaller of the two sets and swap them if necessary.
1668 	 */
1669 	for (i = min(stofree, stokeep); i > 0; i--) {
1670 		slab = LIST_FIRST(&dom->ud_free_slab);
1671 		LIST_REMOVE(slab, us_link);
1672 		LIST_INSERT_HEAD(&freeslabs, slab, us_link);
1673 	}
1674 	if (stofree > stokeep)
1675 		LIST_SWAP(&freeslabs, &dom->ud_free_slab, uma_slab, us_link);
1676 
1677 	if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0) {
1678 		LIST_FOREACH(slab, &freeslabs, us_link)
1679 			UMA_HASH_REMOVE(&keg->uk_hash, slab);
1680 	}
1681 	dom->ud_free_items -= stofree * keg->uk_ipers;
1682 	dom->ud_free_slabs -= stofree;
1683 	dom->ud_pages -= stofree * keg->uk_ppera;
1684 	KEG_UNLOCK(keg, domain);
1685 
1686 	LIST_FOREACH_SAFE(slab, &freeslabs, us_link, tmp)
1687 		keg_free_slab(keg, slab, keg->uk_ipers);
1688 }
1689 
1690 /*
1691  * Frees pages from a keg back to the system.  This is done on demand from
1692  * the pageout daemon.
1693  *
1694  * Returns nothing.
1695  */
1696 static void
1697 keg_drain(uma_keg_t keg, int domain)
1698 {
1699 	int i;
1700 
1701 	if ((keg->uk_flags & UMA_ZONE_NOFREE) != 0)
1702 		return;
1703 	if (domain != UMA_ANYDOMAIN) {
1704 		keg_drain_domain(keg, domain);
1705 	} else {
1706 		for (i = 0; i < vm_ndomains; i++)
1707 			keg_drain_domain(keg, i);
1708 	}
1709 }
1710 
1711 static void
1712 zone_reclaim(uma_zone_t zone, int domain, int waitok, bool drain)
1713 {
1714 	/*
1715 	 * Count active reclaim operations in order to interlock with
1716 	 * zone_dtor(), which removes the zone from global lists before
1717 	 * attempting to reclaim items itself.
1718 	 *
1719 	 * The zone may be destroyed while sleeping, so only zone_dtor() should
1720 	 * specify M_WAITOK.
1721 	 */
1722 	ZONE_LOCK(zone);
1723 	if (waitok == M_WAITOK) {
1724 		while (zone->uz_reclaimers > 0)
1725 			msleep(zone, ZONE_LOCKPTR(zone), PVM, "zonedrain", 1);
1726 	}
1727 	zone->uz_reclaimers++;
1728 	ZONE_UNLOCK(zone);
1729 	bucket_cache_reclaim(zone, drain, domain);
1730 
1731 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0)
1732 		keg_drain(zone->uz_keg, domain);
1733 	ZONE_LOCK(zone);
1734 	zone->uz_reclaimers--;
1735 	if (zone->uz_reclaimers == 0)
1736 		wakeup(zone);
1737 	ZONE_UNLOCK(zone);
1738 }
1739 
1740 /*
1741  * Allocate a new slab for a keg and inserts it into the partial slab list.
1742  * The keg should be unlocked on entry.  If the allocation succeeds it will
1743  * be locked on return.
1744  *
1745  * Arguments:
1746  *	flags   Wait flags for the item initialization routine
1747  *	aflags  Wait flags for the slab allocation
1748  *
1749  * Returns:
1750  *	The slab that was allocated or NULL if there is no memory and the
1751  *	caller specified M_NOWAIT.
1752  */
1753 static uma_slab_t
1754 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int domain, int flags,
1755     int aflags)
1756 {
1757 	uma_domain_t dom;
1758 	uma_slab_t slab;
1759 	unsigned long size;
1760 	uint8_t *mem;
1761 	uint8_t sflags;
1762 	int i;
1763 
1764 	KASSERT(domain >= 0 && domain < vm_ndomains,
1765 	    ("keg_alloc_slab: domain %d out of range", domain));
1766 
1767 	slab = NULL;
1768 	mem = NULL;
1769 	if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) {
1770 		uma_hash_slab_t hslab;
1771 		hslab = zone_alloc_item(slabzone(keg->uk_ipers), NULL,
1772 		    domain, aflags);
1773 		if (hslab == NULL)
1774 			goto fail;
1775 		slab = &hslab->uhs_slab;
1776 	}
1777 
1778 	/*
1779 	 * This reproduces the old vm_zone behavior of zero filling pages the
1780 	 * first time they are added to a zone.
1781 	 *
1782 	 * Malloced items are zeroed in uma_zalloc.
1783 	 */
1784 
1785 	if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
1786 		aflags |= M_ZERO;
1787 	else
1788 		aflags &= ~M_ZERO;
1789 
1790 	if (keg->uk_flags & UMA_ZONE_NODUMP)
1791 		aflags |= M_NODUMP;
1792 
1793 	/* zone is passed for legacy reasons. */
1794 	size = keg->uk_ppera * PAGE_SIZE;
1795 	mem = keg->uk_allocf(zone, size, domain, &sflags, aflags);
1796 	if (mem == NULL) {
1797 		if (keg->uk_flags & UMA_ZFLAG_OFFPAGE)
1798 			zone_free_item(slabzone(keg->uk_ipers),
1799 			    slab_tohashslab(slab), NULL, SKIP_NONE);
1800 		goto fail;
1801 	}
1802 	uma_total_inc(size);
1803 
1804 	/* For HASH zones all pages go to the same uma_domain. */
1805 	if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
1806 		domain = 0;
1807 
1808 	/* Point the slab into the allocated memory */
1809 	if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE))
1810 		slab = (uma_slab_t)(mem + keg->uk_pgoff);
1811 	else
1812 		slab_tohashslab(slab)->uhs_data = mem;
1813 
1814 	if (keg->uk_flags & UMA_ZFLAG_VTOSLAB)
1815 		for (i = 0; i < keg->uk_ppera; i++)
1816 			vsetzoneslab((vm_offset_t)mem + (i * PAGE_SIZE),
1817 			    zone, slab);
1818 
1819 	slab->us_freecount = keg->uk_ipers;
1820 	slab->us_flags = sflags;
1821 	slab->us_domain = domain;
1822 
1823 	BIT_FILL(keg->uk_ipers, &slab->us_free);
1824 #ifdef INVARIANTS
1825 	BIT_ZERO(keg->uk_ipers, slab_dbg_bits(slab, keg));
1826 #endif
1827 
1828 	if (keg->uk_init != NULL) {
1829 		for (i = 0; i < keg->uk_ipers; i++)
1830 			if (keg->uk_init(slab_item(slab, keg, i),
1831 			    keg->uk_size, flags) != 0)
1832 				break;
1833 		if (i != keg->uk_ipers) {
1834 			keg_free_slab(keg, slab, i);
1835 			goto fail;
1836 		}
1837 	}
1838 	kasan_mark_slab_invalid(keg, mem);
1839 	KEG_LOCK(keg, domain);
1840 
1841 	CTR3(KTR_UMA, "keg_alloc_slab: allocated slab %p for %s(%p)",
1842 	    slab, keg->uk_name, keg);
1843 
1844 	if (keg->uk_flags & UMA_ZFLAG_HASH)
1845 		UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
1846 
1847 	/*
1848 	 * If we got a slab here it's safe to mark it partially used
1849 	 * and return.  We assume that the caller is going to remove
1850 	 * at least one item.
1851 	 */
1852 	dom = &keg->uk_domain[domain];
1853 	LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
1854 	dom->ud_pages += keg->uk_ppera;
1855 	dom->ud_free_items += keg->uk_ipers;
1856 
1857 	return (slab);
1858 
1859 fail:
1860 	return (NULL);
1861 }
1862 
1863 /*
1864  * This function is intended to be used early on in place of page_alloc().  It
1865  * performs contiguous physical memory allocations and uses a bump allocator for
1866  * KVA, so is usable before the kernel map is initialized.
1867  */
1868 static void *
1869 startup_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1870     int wait)
1871 {
1872 	vm_paddr_t pa;
1873 	vm_page_t m;
1874 	int i, pages;
1875 
1876 	pages = howmany(bytes, PAGE_SIZE);
1877 	KASSERT(pages > 0, ("%s can't reserve 0 pages", __func__));
1878 
1879 	*pflag = UMA_SLAB_BOOT;
1880 	m = vm_page_alloc_noobj_contig_domain(domain, malloc2vm_flags(wait) |
1881 	    VM_ALLOC_WIRED, pages, (vm_paddr_t)0, ~(vm_paddr_t)0, 1, 0,
1882 	    VM_MEMATTR_DEFAULT);
1883 	if (m == NULL)
1884 		return (NULL);
1885 
1886 	pa = VM_PAGE_TO_PHYS(m);
1887 	for (i = 0; i < pages; i++, pa += PAGE_SIZE) {
1888 #if defined(__aarch64__) || defined(__amd64__) || \
1889     defined(__riscv) || defined(__powerpc64__)
1890 		if ((wait & M_NODUMP) == 0)
1891 			dump_add_page(pa);
1892 #endif
1893 	}
1894 
1895 	/* Allocate KVA and indirectly advance bootmem. */
1896 	return ((void *)pmap_map(&bootmem, m->phys_addr,
1897 	    m->phys_addr + (pages * PAGE_SIZE), VM_PROT_READ | VM_PROT_WRITE));
1898 }
1899 
1900 static void
1901 startup_free(void *mem, vm_size_t bytes)
1902 {
1903 	vm_offset_t va;
1904 	vm_page_t m;
1905 
1906 	va = (vm_offset_t)mem;
1907 	m = PHYS_TO_VM_PAGE(pmap_kextract(va));
1908 
1909 	/*
1910 	 * startup_alloc() returns direct-mapped slabs on some platforms.  Avoid
1911 	 * unmapping ranges of the direct map.
1912 	 */
1913 	if (va >= bootstart && va + bytes <= bootmem)
1914 		pmap_remove(kernel_pmap, va, va + bytes);
1915 	for (; bytes != 0; bytes -= PAGE_SIZE, m++) {
1916 #if defined(__aarch64__) || defined(__amd64__) || \
1917     defined(__riscv) || defined(__powerpc64__)
1918 		dump_drop_page(VM_PAGE_TO_PHYS(m));
1919 #endif
1920 		vm_page_unwire_noq(m);
1921 		vm_page_free(m);
1922 	}
1923 }
1924 
1925 /*
1926  * Allocates a number of pages from the system
1927  *
1928  * Arguments:
1929  *	bytes  The number of bytes requested
1930  *	wait  Shall we wait?
1931  *
1932  * Returns:
1933  *	A pointer to the alloced memory or possibly
1934  *	NULL if M_NOWAIT is set.
1935  */
1936 static void *
1937 page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1938     int wait)
1939 {
1940 	void *p;	/* Returned page */
1941 
1942 	*pflag = UMA_SLAB_KERNEL;
1943 	p = (void *)kmem_malloc_domainset(DOMAINSET_FIXED(domain), bytes, wait);
1944 
1945 	return (p);
1946 }
1947 
1948 static void *
1949 pcpu_page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1950     int wait)
1951 {
1952 	struct pglist alloctail;
1953 	vm_offset_t addr, zkva;
1954 	int cpu, flags;
1955 	vm_page_t p, p_next;
1956 #ifdef NUMA
1957 	struct pcpu *pc;
1958 #endif
1959 
1960 	MPASS(bytes == (mp_maxid + 1) * PAGE_SIZE);
1961 
1962 	TAILQ_INIT(&alloctail);
1963 	flags = VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | malloc2vm_flags(wait);
1964 	*pflag = UMA_SLAB_KERNEL;
1965 	for (cpu = 0; cpu <= mp_maxid; cpu++) {
1966 		if (CPU_ABSENT(cpu)) {
1967 			p = vm_page_alloc_noobj(flags);
1968 		} else {
1969 #ifndef NUMA
1970 			p = vm_page_alloc_noobj(flags);
1971 #else
1972 			pc = pcpu_find(cpu);
1973 			if (__predict_false(VM_DOMAIN_EMPTY(pc->pc_domain)))
1974 				p = NULL;
1975 			else
1976 				p = vm_page_alloc_noobj_domain(pc->pc_domain,
1977 				    flags);
1978 			if (__predict_false(p == NULL))
1979 				p = vm_page_alloc_noobj(flags);
1980 #endif
1981 		}
1982 		if (__predict_false(p == NULL))
1983 			goto fail;
1984 		TAILQ_INSERT_TAIL(&alloctail, p, listq);
1985 	}
1986 	if ((addr = kva_alloc(bytes)) == 0)
1987 		goto fail;
1988 	zkva = addr;
1989 	TAILQ_FOREACH(p, &alloctail, listq) {
1990 		pmap_qenter(zkva, &p, 1);
1991 		zkva += PAGE_SIZE;
1992 	}
1993 	return ((void*)addr);
1994 fail:
1995 	TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1996 		vm_page_unwire_noq(p);
1997 		vm_page_free(p);
1998 	}
1999 	return (NULL);
2000 }
2001 
2002 /*
2003  * Allocates a number of pages not belonging to a VM object
2004  *
2005  * Arguments:
2006  *	bytes  The number of bytes requested
2007  *	wait   Shall we wait?
2008  *
2009  * Returns:
2010  *	A pointer to the alloced memory or possibly
2011  *	NULL if M_NOWAIT is set.
2012  */
2013 static void *
2014 noobj_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags,
2015     int wait)
2016 {
2017 	TAILQ_HEAD(, vm_page) alloctail;
2018 	u_long npages;
2019 	vm_offset_t retkva, zkva;
2020 	vm_page_t p, p_next;
2021 	uma_keg_t keg;
2022 	int req;
2023 
2024 	TAILQ_INIT(&alloctail);
2025 	keg = zone->uz_keg;
2026 	req = VM_ALLOC_INTERRUPT | VM_ALLOC_WIRED;
2027 	if ((wait & M_WAITOK) != 0)
2028 		req |= VM_ALLOC_WAITOK;
2029 
2030 	npages = howmany(bytes, PAGE_SIZE);
2031 	while (npages > 0) {
2032 		p = vm_page_alloc_noobj_domain(domain, req);
2033 		if (p != NULL) {
2034 			/*
2035 			 * Since the page does not belong to an object, its
2036 			 * listq is unused.
2037 			 */
2038 			TAILQ_INSERT_TAIL(&alloctail, p, listq);
2039 			npages--;
2040 			continue;
2041 		}
2042 		/*
2043 		 * Page allocation failed, free intermediate pages and
2044 		 * exit.
2045 		 */
2046 		TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
2047 			vm_page_unwire_noq(p);
2048 			vm_page_free(p);
2049 		}
2050 		return (NULL);
2051 	}
2052 	*flags = UMA_SLAB_PRIV;
2053 	zkva = keg->uk_kva +
2054 	    atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
2055 	retkva = zkva;
2056 	TAILQ_FOREACH(p, &alloctail, listq) {
2057 		pmap_qenter(zkva, &p, 1);
2058 		zkva += PAGE_SIZE;
2059 	}
2060 
2061 	return ((void *)retkva);
2062 }
2063 
2064 /*
2065  * Allocate physically contiguous pages.
2066  */
2067 static void *
2068 contig_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
2069     int wait)
2070 {
2071 
2072 	*pflag = UMA_SLAB_KERNEL;
2073 	return ((void *)kmem_alloc_contig_domainset(DOMAINSET_FIXED(domain),
2074 	    bytes, wait, 0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT));
2075 }
2076 
2077 /*
2078  * Frees a number of pages to the system
2079  *
2080  * Arguments:
2081  *	mem   A pointer to the memory to be freed
2082  *	size  The size of the memory being freed
2083  *	flags The original p->us_flags field
2084  *
2085  * Returns:
2086  *	Nothing
2087  */
2088 static void
2089 page_free(void *mem, vm_size_t size, uint8_t flags)
2090 {
2091 
2092 	if ((flags & UMA_SLAB_BOOT) != 0) {
2093 		startup_free(mem, size);
2094 		return;
2095 	}
2096 
2097 	KASSERT((flags & UMA_SLAB_KERNEL) != 0,
2098 	    ("UMA: page_free used with invalid flags %x", flags));
2099 
2100 	kmem_free((vm_offset_t)mem, size);
2101 }
2102 
2103 /*
2104  * Frees pcpu zone allocations
2105  *
2106  * Arguments:
2107  *	mem   A pointer to the memory to be freed
2108  *	size  The size of the memory being freed
2109  *	flags The original p->us_flags field
2110  *
2111  * Returns:
2112  *	Nothing
2113  */
2114 static void
2115 pcpu_page_free(void *mem, vm_size_t size, uint8_t flags)
2116 {
2117 	vm_offset_t sva, curva;
2118 	vm_paddr_t paddr;
2119 	vm_page_t m;
2120 
2121 	MPASS(size == (mp_maxid+1)*PAGE_SIZE);
2122 
2123 	if ((flags & UMA_SLAB_BOOT) != 0) {
2124 		startup_free(mem, size);
2125 		return;
2126 	}
2127 
2128 	sva = (vm_offset_t)mem;
2129 	for (curva = sva; curva < sva + size; curva += PAGE_SIZE) {
2130 		paddr = pmap_kextract(curva);
2131 		m = PHYS_TO_VM_PAGE(paddr);
2132 		vm_page_unwire_noq(m);
2133 		vm_page_free(m);
2134 	}
2135 	pmap_qremove(sva, size >> PAGE_SHIFT);
2136 	kva_free(sva, size);
2137 }
2138 
2139 /*
2140  * Zero fill initializer
2141  *
2142  * Arguments/Returns follow uma_init specifications
2143  */
2144 static int
2145 zero_init(void *mem, int size, int flags)
2146 {
2147 	bzero(mem, size);
2148 	return (0);
2149 }
2150 
2151 #ifdef INVARIANTS
2152 static struct noslabbits *
2153 slab_dbg_bits(uma_slab_t slab, uma_keg_t keg)
2154 {
2155 
2156 	return ((void *)((char *)&slab->us_free + BITSET_SIZE(keg->uk_ipers)));
2157 }
2158 #endif
2159 
2160 /*
2161  * Actual size of embedded struct slab (!OFFPAGE).
2162  */
2163 static size_t
2164 slab_sizeof(int nitems)
2165 {
2166 	size_t s;
2167 
2168 	s = sizeof(struct uma_slab) + BITSET_SIZE(nitems) * SLAB_BITSETS;
2169 	return (roundup(s, UMA_ALIGN_PTR + 1));
2170 }
2171 
2172 #define	UMA_FIXPT_SHIFT	31
2173 #define	UMA_FRAC_FIXPT(n, d)						\
2174 	((uint32_t)(((uint64_t)(n) << UMA_FIXPT_SHIFT) / (d)))
2175 #define	UMA_FIXPT_PCT(f)						\
2176 	((u_int)(((uint64_t)100 * (f)) >> UMA_FIXPT_SHIFT))
2177 #define	UMA_PCT_FIXPT(pct)	UMA_FRAC_FIXPT((pct), 100)
2178 #define	UMA_MIN_EFF	UMA_PCT_FIXPT(100 - UMA_MAX_WASTE)
2179 
2180 /*
2181  * Compute the number of items that will fit in a slab.  If hdr is true, the
2182  * item count may be limited to provide space in the slab for an inline slab
2183  * header.  Otherwise, all slab space will be provided for item storage.
2184  */
2185 static u_int
2186 slab_ipers_hdr(u_int size, u_int rsize, u_int slabsize, bool hdr)
2187 {
2188 	u_int ipers;
2189 	u_int padpi;
2190 
2191 	/* The padding between items is not needed after the last item. */
2192 	padpi = rsize - size;
2193 
2194 	if (hdr) {
2195 		/*
2196 		 * Start with the maximum item count and remove items until
2197 		 * the slab header first alongside the allocatable memory.
2198 		 */
2199 		for (ipers = MIN(SLAB_MAX_SETSIZE,
2200 		    (slabsize + padpi - slab_sizeof(1)) / rsize);
2201 		    ipers > 0 &&
2202 		    ipers * rsize - padpi + slab_sizeof(ipers) > slabsize;
2203 		    ipers--)
2204 			continue;
2205 	} else {
2206 		ipers = MIN((slabsize + padpi) / rsize, SLAB_MAX_SETSIZE);
2207 	}
2208 
2209 	return (ipers);
2210 }
2211 
2212 struct keg_layout_result {
2213 	u_int format;
2214 	u_int slabsize;
2215 	u_int ipers;
2216 	u_int eff;
2217 };
2218 
2219 static void
2220 keg_layout_one(uma_keg_t keg, u_int rsize, u_int slabsize, u_int fmt,
2221     struct keg_layout_result *kl)
2222 {
2223 	u_int total;
2224 
2225 	kl->format = fmt;
2226 	kl->slabsize = slabsize;
2227 
2228 	/* Handle INTERNAL as inline with an extra page. */
2229 	if ((fmt & UMA_ZFLAG_INTERNAL) != 0) {
2230 		kl->format &= ~UMA_ZFLAG_INTERNAL;
2231 		kl->slabsize += PAGE_SIZE;
2232 	}
2233 
2234 	kl->ipers = slab_ipers_hdr(keg->uk_size, rsize, kl->slabsize,
2235 	    (fmt & UMA_ZFLAG_OFFPAGE) == 0);
2236 
2237 	/* Account for memory used by an offpage slab header. */
2238 	total = kl->slabsize;
2239 	if ((fmt & UMA_ZFLAG_OFFPAGE) != 0)
2240 		total += slabzone(kl->ipers)->uz_keg->uk_rsize;
2241 
2242 	kl->eff = UMA_FRAC_FIXPT(kl->ipers * rsize, total);
2243 }
2244 
2245 /*
2246  * Determine the format of a uma keg.  This determines where the slab header
2247  * will be placed (inline or offpage) and calculates ipers, rsize, and ppera.
2248  *
2249  * Arguments
2250  *	keg  The zone we should initialize
2251  *
2252  * Returns
2253  *	Nothing
2254  */
2255 static void
2256 keg_layout(uma_keg_t keg)
2257 {
2258 	struct keg_layout_result kl = {}, kl_tmp;
2259 	u_int fmts[2];
2260 	u_int alignsize;
2261 	u_int nfmt;
2262 	u_int pages;
2263 	u_int rsize;
2264 	u_int slabsize;
2265 	u_int i, j;
2266 
2267 	KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
2268 	    (keg->uk_size <= UMA_PCPU_ALLOC_SIZE &&
2269 	     (keg->uk_flags & UMA_ZONE_CACHESPREAD) == 0),
2270 	    ("%s: cannot configure for PCPU: keg=%s, size=%u, flags=0x%b",
2271 	     __func__, keg->uk_name, keg->uk_size, keg->uk_flags,
2272 	     PRINT_UMA_ZFLAGS));
2273 	KASSERT((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) == 0 ||
2274 	    (keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0,
2275 	    ("%s: incompatible flags 0x%b", __func__, keg->uk_flags,
2276 	     PRINT_UMA_ZFLAGS));
2277 
2278 	alignsize = keg->uk_align + 1;
2279 #ifdef KASAN
2280 	/*
2281 	 * ASAN requires that each allocation be aligned to the shadow map
2282 	 * scale factor.
2283 	 */
2284 	if (alignsize < KASAN_SHADOW_SCALE)
2285 		alignsize = KASAN_SHADOW_SCALE;
2286 #endif
2287 
2288 	/*
2289 	 * Calculate the size of each allocation (rsize) according to
2290 	 * alignment.  If the requested size is smaller than we have
2291 	 * allocation bits for we round it up.
2292 	 */
2293 	rsize = MAX(keg->uk_size, UMA_SMALLEST_UNIT);
2294 	rsize = roundup2(rsize, alignsize);
2295 
2296 	if ((keg->uk_flags & UMA_ZONE_CACHESPREAD) != 0) {
2297 		/*
2298 		 * We want one item to start on every align boundary in a page.
2299 		 * To do this we will span pages.  We will also extend the item
2300 		 * by the size of align if it is an even multiple of align.
2301 		 * Otherwise, it would fall on the same boundary every time.
2302 		 */
2303 		if ((rsize & alignsize) == 0)
2304 			rsize += alignsize;
2305 		slabsize = rsize * (PAGE_SIZE / alignsize);
2306 		slabsize = MIN(slabsize, rsize * SLAB_MAX_SETSIZE);
2307 		slabsize = MIN(slabsize, UMA_CACHESPREAD_MAX_SIZE);
2308 		slabsize = round_page(slabsize);
2309 	} else {
2310 		/*
2311 		 * Start with a slab size of as many pages as it takes to
2312 		 * represent a single item.  We will try to fit as many
2313 		 * additional items into the slab as possible.
2314 		 */
2315 		slabsize = round_page(keg->uk_size);
2316 	}
2317 
2318 	/* Build a list of all of the available formats for this keg. */
2319 	nfmt = 0;
2320 
2321 	/* Evaluate an inline slab layout. */
2322 	if ((keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0)
2323 		fmts[nfmt++] = 0;
2324 
2325 	/* TODO: vm_page-embedded slab. */
2326 
2327 	/*
2328 	 * We can't do OFFPAGE if we're internal or if we've been
2329 	 * asked to not go to the VM for buckets.  If we do this we
2330 	 * may end up going to the VM for slabs which we do not want
2331 	 * to do if we're UMA_ZONE_VM, which clearly forbids it.
2332 	 * In those cases, evaluate a pseudo-format called INTERNAL
2333 	 * which has an inline slab header and one extra page to
2334 	 * guarantee that it fits.
2335 	 *
2336 	 * Otherwise, see if using an OFFPAGE slab will improve our
2337 	 * efficiency.
2338 	 */
2339 	if ((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) != 0)
2340 		fmts[nfmt++] = UMA_ZFLAG_INTERNAL;
2341 	else
2342 		fmts[nfmt++] = UMA_ZFLAG_OFFPAGE;
2343 
2344 	/*
2345 	 * Choose a slab size and format which satisfy the minimum efficiency.
2346 	 * Prefer the smallest slab size that meets the constraints.
2347 	 *
2348 	 * Start with a minimum slab size, to accommodate CACHESPREAD.  Then,
2349 	 * for small items (up to PAGE_SIZE), the iteration increment is one
2350 	 * page; and for large items, the increment is one item.
2351 	 */
2352 	i = (slabsize + rsize - keg->uk_size) / MAX(PAGE_SIZE, rsize);
2353 	KASSERT(i >= 1, ("keg %s(%p) flags=0x%b slabsize=%u, rsize=%u, i=%u",
2354 	    keg->uk_name, keg, keg->uk_flags, PRINT_UMA_ZFLAGS, slabsize,
2355 	    rsize, i));
2356 	for ( ; ; i++) {
2357 		slabsize = (rsize <= PAGE_SIZE) ? ptoa(i) :
2358 		    round_page(rsize * (i - 1) + keg->uk_size);
2359 
2360 		for (j = 0; j < nfmt; j++) {
2361 			/* Only if we have no viable format yet. */
2362 			if ((fmts[j] & UMA_ZFLAG_INTERNAL) != 0 &&
2363 			    kl.ipers > 0)
2364 				continue;
2365 
2366 			keg_layout_one(keg, rsize, slabsize, fmts[j], &kl_tmp);
2367 			if (kl_tmp.eff <= kl.eff)
2368 				continue;
2369 
2370 			kl = kl_tmp;
2371 
2372 			CTR6(KTR_UMA, "keg %s layout: format %#x "
2373 			    "(ipers %u * rsize %u) / slabsize %#x = %u%% eff",
2374 			    keg->uk_name, kl.format, kl.ipers, rsize,
2375 			    kl.slabsize, UMA_FIXPT_PCT(kl.eff));
2376 
2377 			/* Stop when we reach the minimum efficiency. */
2378 			if (kl.eff >= UMA_MIN_EFF)
2379 				break;
2380 		}
2381 
2382 		if (kl.eff >= UMA_MIN_EFF || !multipage_slabs ||
2383 		    slabsize >= SLAB_MAX_SETSIZE * rsize ||
2384 		    (keg->uk_flags & (UMA_ZONE_PCPU | UMA_ZONE_CONTIG)) != 0)
2385 			break;
2386 	}
2387 
2388 	pages = atop(kl.slabsize);
2389 	if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
2390 		pages *= mp_maxid + 1;
2391 
2392 	keg->uk_rsize = rsize;
2393 	keg->uk_ipers = kl.ipers;
2394 	keg->uk_ppera = pages;
2395 	keg->uk_flags |= kl.format;
2396 
2397 	/*
2398 	 * How do we find the slab header if it is offpage or if not all item
2399 	 * start addresses are in the same page?  We could solve the latter
2400 	 * case with vaddr alignment, but we don't.
2401 	 */
2402 	if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0 ||
2403 	    (keg->uk_ipers - 1) * rsize >= PAGE_SIZE) {
2404 		if ((keg->uk_flags & UMA_ZONE_NOTPAGE) != 0)
2405 			keg->uk_flags |= UMA_ZFLAG_HASH;
2406 		else
2407 			keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2408 	}
2409 
2410 	CTR6(KTR_UMA, "%s: keg=%s, flags=%#x, rsize=%u, ipers=%u, ppera=%u",
2411 	    __func__, keg->uk_name, keg->uk_flags, rsize, keg->uk_ipers,
2412 	    pages);
2413 	KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_MAX_SETSIZE,
2414 	    ("%s: keg=%s, flags=0x%b, rsize=%u, ipers=%u, ppera=%u", __func__,
2415 	     keg->uk_name, keg->uk_flags, PRINT_UMA_ZFLAGS, rsize,
2416 	     keg->uk_ipers, pages));
2417 }
2418 
2419 /*
2420  * Keg header ctor.  This initializes all fields, locks, etc.  And inserts
2421  * the keg onto the global keg list.
2422  *
2423  * Arguments/Returns follow uma_ctor specifications
2424  *	udata  Actually uma_kctor_args
2425  */
2426 static int
2427 keg_ctor(void *mem, int size, void *udata, int flags)
2428 {
2429 	struct uma_kctor_args *arg = udata;
2430 	uma_keg_t keg = mem;
2431 	uma_zone_t zone;
2432 	int i;
2433 
2434 	bzero(keg, size);
2435 	keg->uk_size = arg->size;
2436 	keg->uk_init = arg->uminit;
2437 	keg->uk_fini = arg->fini;
2438 	keg->uk_align = arg->align;
2439 	keg->uk_reserve = 0;
2440 	keg->uk_flags = arg->flags;
2441 
2442 	/*
2443 	 * We use a global round-robin policy by default.  Zones with
2444 	 * UMA_ZONE_FIRSTTOUCH set will use first-touch instead, in which
2445 	 * case the iterator is never run.
2446 	 */
2447 	keg->uk_dr.dr_policy = DOMAINSET_RR();
2448 	keg->uk_dr.dr_iter = 0;
2449 
2450 	/*
2451 	 * The primary zone is passed to us at keg-creation time.
2452 	 */
2453 	zone = arg->zone;
2454 	keg->uk_name = zone->uz_name;
2455 
2456 	if (arg->flags & UMA_ZONE_ZINIT)
2457 		keg->uk_init = zero_init;
2458 
2459 	if (arg->flags & UMA_ZONE_MALLOC)
2460 		keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2461 
2462 #ifndef SMP
2463 	keg->uk_flags &= ~UMA_ZONE_PCPU;
2464 #endif
2465 
2466 	keg_layout(keg);
2467 
2468 	/*
2469 	 * Use a first-touch NUMA policy for kegs that pmap_extract() will
2470 	 * work on.  Use round-robin for everything else.
2471 	 *
2472 	 * Zones may override the default by specifying either.
2473 	 */
2474 #ifdef NUMA
2475 	if ((keg->uk_flags &
2476 	    (UMA_ZONE_ROUNDROBIN | UMA_ZFLAG_CACHE | UMA_ZONE_NOTPAGE)) == 0)
2477 		keg->uk_flags |= UMA_ZONE_FIRSTTOUCH;
2478 	else if ((keg->uk_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2479 		keg->uk_flags |= UMA_ZONE_ROUNDROBIN;
2480 #endif
2481 
2482 	/*
2483 	 * If we haven't booted yet we need allocations to go through the
2484 	 * startup cache until the vm is ready.
2485 	 */
2486 #ifdef UMA_MD_SMALL_ALLOC
2487 	if (keg->uk_ppera == 1)
2488 		keg->uk_allocf = uma_small_alloc;
2489 	else
2490 #endif
2491 	if (booted < BOOT_KVA)
2492 		keg->uk_allocf = startup_alloc;
2493 	else if (keg->uk_flags & UMA_ZONE_PCPU)
2494 		keg->uk_allocf = pcpu_page_alloc;
2495 	else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 && keg->uk_ppera > 1)
2496 		keg->uk_allocf = contig_alloc;
2497 	else
2498 		keg->uk_allocf = page_alloc;
2499 #ifdef UMA_MD_SMALL_ALLOC
2500 	if (keg->uk_ppera == 1)
2501 		keg->uk_freef = uma_small_free;
2502 	else
2503 #endif
2504 	if (keg->uk_flags & UMA_ZONE_PCPU)
2505 		keg->uk_freef = pcpu_page_free;
2506 	else
2507 		keg->uk_freef = page_free;
2508 
2509 	/*
2510 	 * Initialize keg's locks.
2511 	 */
2512 	for (i = 0; i < vm_ndomains; i++)
2513 		KEG_LOCK_INIT(keg, i, (arg->flags & UMA_ZONE_MTXCLASS));
2514 
2515 	/*
2516 	 * If we're putting the slab header in the actual page we need to
2517 	 * figure out where in each page it goes.  See slab_sizeof
2518 	 * definition.
2519 	 */
2520 	if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE)) {
2521 		size_t shsize;
2522 
2523 		shsize = slab_sizeof(keg->uk_ipers);
2524 		keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - shsize;
2525 		/*
2526 		 * The only way the following is possible is if with our
2527 		 * UMA_ALIGN_PTR adjustments we are now bigger than
2528 		 * UMA_SLAB_SIZE.  I haven't checked whether this is
2529 		 * mathematically possible for all cases, so we make
2530 		 * sure here anyway.
2531 		 */
2532 		KASSERT(keg->uk_pgoff + shsize <= PAGE_SIZE * keg->uk_ppera,
2533 		    ("zone %s ipers %d rsize %d size %d slab won't fit",
2534 		    zone->uz_name, keg->uk_ipers, keg->uk_rsize, keg->uk_size));
2535 	}
2536 
2537 	if (keg->uk_flags & UMA_ZFLAG_HASH)
2538 		hash_alloc(&keg->uk_hash, 0);
2539 
2540 	CTR3(KTR_UMA, "keg_ctor %p zone %s(%p)", keg, zone->uz_name, zone);
2541 
2542 	LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
2543 
2544 	rw_wlock(&uma_rwlock);
2545 	LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
2546 	rw_wunlock(&uma_rwlock);
2547 	return (0);
2548 }
2549 
2550 static void
2551 zone_kva_available(uma_zone_t zone, void *unused)
2552 {
2553 	uma_keg_t keg;
2554 
2555 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
2556 		return;
2557 	KEG_GET(zone, keg);
2558 
2559 	if (keg->uk_allocf == startup_alloc) {
2560 		/* Switch to the real allocator. */
2561 		if (keg->uk_flags & UMA_ZONE_PCPU)
2562 			keg->uk_allocf = pcpu_page_alloc;
2563 		else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 &&
2564 		    keg->uk_ppera > 1)
2565 			keg->uk_allocf = contig_alloc;
2566 		else
2567 			keg->uk_allocf = page_alloc;
2568 	}
2569 }
2570 
2571 static void
2572 zone_alloc_counters(uma_zone_t zone, void *unused)
2573 {
2574 
2575 	zone->uz_allocs = counter_u64_alloc(M_WAITOK);
2576 	zone->uz_frees = counter_u64_alloc(M_WAITOK);
2577 	zone->uz_fails = counter_u64_alloc(M_WAITOK);
2578 	zone->uz_xdomain = counter_u64_alloc(M_WAITOK);
2579 }
2580 
2581 static void
2582 zone_alloc_sysctl(uma_zone_t zone, void *unused)
2583 {
2584 	uma_zone_domain_t zdom;
2585 	uma_domain_t dom;
2586 	uma_keg_t keg;
2587 	struct sysctl_oid *oid, *domainoid;
2588 	int domains, i, cnt;
2589 	static const char *nokeg = "cache zone";
2590 	char *c;
2591 
2592 	/*
2593 	 * Make a sysctl safe copy of the zone name by removing
2594 	 * any special characters and handling dups by appending
2595 	 * an index.
2596 	 */
2597 	if (zone->uz_namecnt != 0) {
2598 		/* Count the number of decimal digits and '_' separator. */
2599 		for (i = 1, cnt = zone->uz_namecnt; cnt != 0; i++)
2600 			cnt /= 10;
2601 		zone->uz_ctlname = malloc(strlen(zone->uz_name) + i + 1,
2602 		    M_UMA, M_WAITOK);
2603 		sprintf(zone->uz_ctlname, "%s_%d", zone->uz_name,
2604 		    zone->uz_namecnt);
2605 	} else
2606 		zone->uz_ctlname = strdup(zone->uz_name, M_UMA);
2607 	for (c = zone->uz_ctlname; *c != '\0'; c++)
2608 		if (strchr("./\\ -", *c) != NULL)
2609 			*c = '_';
2610 
2611 	/*
2612 	 * Basic parameters at the root.
2613 	 */
2614 	zone->uz_oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_vm_uma),
2615 	    OID_AUTO, zone->uz_ctlname, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2616 	oid = zone->uz_oid;
2617 	SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2618 	    "size", CTLFLAG_RD, &zone->uz_size, 0, "Allocation size");
2619 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2620 	    "flags", CTLFLAG_RD | CTLTYPE_STRING | CTLFLAG_MPSAFE,
2621 	    zone, 0, sysctl_handle_uma_zone_flags, "A",
2622 	    "Allocator configuration flags");
2623 	SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2624 	    "bucket_size", CTLFLAG_RD, &zone->uz_bucket_size, 0,
2625 	    "Desired per-cpu cache size");
2626 	SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2627 	    "bucket_size_max", CTLFLAG_RD, &zone->uz_bucket_size_max, 0,
2628 	    "Maximum allowed per-cpu cache size");
2629 
2630 	/*
2631 	 * keg if present.
2632 	 */
2633 	if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
2634 		domains = vm_ndomains;
2635 	else
2636 		domains = 1;
2637 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2638 	    "keg", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2639 	keg = zone->uz_keg;
2640 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0) {
2641 		SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2642 		    "name", CTLFLAG_RD, keg->uk_name, "Keg name");
2643 		SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2644 		    "rsize", CTLFLAG_RD, &keg->uk_rsize, 0,
2645 		    "Real object size with alignment");
2646 		SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2647 		    "ppera", CTLFLAG_RD, &keg->uk_ppera, 0,
2648 		    "pages per-slab allocation");
2649 		SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2650 		    "ipers", CTLFLAG_RD, &keg->uk_ipers, 0,
2651 		    "items available per-slab");
2652 		SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2653 		    "align", CTLFLAG_RD, &keg->uk_align, 0,
2654 		    "item alignment mask");
2655 		SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2656 		    "reserve", CTLFLAG_RD, &keg->uk_reserve, 0,
2657 		    "number of reserved items");
2658 		SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2659 		    "efficiency", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2660 		    keg, 0, sysctl_handle_uma_slab_efficiency, "I",
2661 		    "Slab utilization (100 - internal fragmentation %)");
2662 		domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(oid),
2663 		    OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2664 		for (i = 0; i < domains; i++) {
2665 			dom = &keg->uk_domain[i];
2666 			oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2667 			    OID_AUTO, VM_DOMAIN(i)->vmd_name,
2668 			    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2669 			SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2670 			    "pages", CTLFLAG_RD, &dom->ud_pages, 0,
2671 			    "Total pages currently allocated from VM");
2672 			SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2673 			    "free_items", CTLFLAG_RD, &dom->ud_free_items, 0,
2674 			    "Items free in the slab layer");
2675 			SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2676 			    "free_slabs", CTLFLAG_RD, &dom->ud_free_slabs, 0,
2677 			    "Unused slabs");
2678 		}
2679 	} else
2680 		SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2681 		    "name", CTLFLAG_RD, nokeg, "Keg name");
2682 
2683 	/*
2684 	 * Information about zone limits.
2685 	 */
2686 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2687 	    "limit", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2688 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2689 	    "items", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2690 	    zone, 0, sysctl_handle_uma_zone_items, "QU",
2691 	    "Current number of allocated items if limit is set");
2692 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2693 	    "max_items", CTLFLAG_RD, &zone->uz_max_items, 0,
2694 	    "Maximum number of allocated and cached items");
2695 	SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2696 	    "sleepers", CTLFLAG_RD, &zone->uz_sleepers, 0,
2697 	    "Number of threads sleeping at limit");
2698 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2699 	    "sleeps", CTLFLAG_RD, &zone->uz_sleeps, 0,
2700 	    "Total zone limit sleeps");
2701 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2702 	    "bucket_max", CTLFLAG_RD, &zone->uz_bucket_max, 0,
2703 	    "Maximum number of items in each domain's bucket cache");
2704 
2705 	/*
2706 	 * Per-domain zone information.
2707 	 */
2708 	domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid),
2709 	    OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2710 	for (i = 0; i < domains; i++) {
2711 		zdom = ZDOM_GET(zone, i);
2712 		oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2713 		    OID_AUTO, VM_DOMAIN(i)->vmd_name,
2714 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2715 		SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2716 		    "nitems", CTLFLAG_RD, &zdom->uzd_nitems,
2717 		    "number of items in this domain");
2718 		SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2719 		    "imax", CTLFLAG_RD, &zdom->uzd_imax,
2720 		    "maximum item count in this period");
2721 		SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2722 		    "imin", CTLFLAG_RD, &zdom->uzd_imin,
2723 		    "minimum item count in this period");
2724 		SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2725 		    "bimin", CTLFLAG_RD, &zdom->uzd_bimin,
2726 		    "Minimum item count in this batch");
2727 		SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2728 		    "wss", CTLFLAG_RD, &zdom->uzd_wss,
2729 		    "Working set size");
2730 		SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2731 		    "limin", CTLFLAG_RD, &zdom->uzd_limin,
2732 		    "Long time minimum item count");
2733 		SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2734 		    "timin", CTLFLAG_RD, &zdom->uzd_timin, 0,
2735 		    "Time since zero long time minimum item count");
2736 	}
2737 
2738 	/*
2739 	 * General statistics.
2740 	 */
2741 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2742 	    "stats", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2743 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2744 	    "current", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2745 	    zone, 1, sysctl_handle_uma_zone_cur, "I",
2746 	    "Current number of allocated items");
2747 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2748 	    "allocs", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2749 	    zone, 0, sysctl_handle_uma_zone_allocs, "QU",
2750 	    "Total allocation calls");
2751 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2752 	    "frees", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2753 	    zone, 0, sysctl_handle_uma_zone_frees, "QU",
2754 	    "Total free calls");
2755 	SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2756 	    "fails", CTLFLAG_RD, &zone->uz_fails,
2757 	    "Number of allocation failures");
2758 	SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2759 	    "xdomain", CTLFLAG_RD, &zone->uz_xdomain,
2760 	    "Free calls from the wrong domain");
2761 }
2762 
2763 struct uma_zone_count {
2764 	const char	*name;
2765 	int		count;
2766 };
2767 
2768 static void
2769 zone_count(uma_zone_t zone, void *arg)
2770 {
2771 	struct uma_zone_count *cnt;
2772 
2773 	cnt = arg;
2774 	/*
2775 	 * Some zones are rapidly created with identical names and
2776 	 * destroyed out of order.  This can lead to gaps in the count.
2777 	 * Use one greater than the maximum observed for this name.
2778 	 */
2779 	if (strcmp(zone->uz_name, cnt->name) == 0)
2780 		cnt->count = MAX(cnt->count,
2781 		    zone->uz_namecnt + 1);
2782 }
2783 
2784 static void
2785 zone_update_caches(uma_zone_t zone)
2786 {
2787 	int i;
2788 
2789 	for (i = 0; i <= mp_maxid; i++) {
2790 		cache_set_uz_size(&zone->uz_cpu[i], zone->uz_size);
2791 		cache_set_uz_flags(&zone->uz_cpu[i], zone->uz_flags);
2792 	}
2793 }
2794 
2795 /*
2796  * Zone header ctor.  This initializes all fields, locks, etc.
2797  *
2798  * Arguments/Returns follow uma_ctor specifications
2799  *	udata  Actually uma_zctor_args
2800  */
2801 static int
2802 zone_ctor(void *mem, int size, void *udata, int flags)
2803 {
2804 	struct uma_zone_count cnt;
2805 	struct uma_zctor_args *arg = udata;
2806 	uma_zone_domain_t zdom;
2807 	uma_zone_t zone = mem;
2808 	uma_zone_t z;
2809 	uma_keg_t keg;
2810 	int i;
2811 
2812 	bzero(zone, size);
2813 	zone->uz_name = arg->name;
2814 	zone->uz_ctor = arg->ctor;
2815 	zone->uz_dtor = arg->dtor;
2816 	zone->uz_init = NULL;
2817 	zone->uz_fini = NULL;
2818 	zone->uz_sleeps = 0;
2819 	zone->uz_bucket_size = 0;
2820 	zone->uz_bucket_size_min = 0;
2821 	zone->uz_bucket_size_max = BUCKET_MAX;
2822 	zone->uz_flags = (arg->flags & UMA_ZONE_SMR);
2823 	zone->uz_warning = NULL;
2824 	/* The domain structures follow the cpu structures. */
2825 	zone->uz_bucket_max = ULONG_MAX;
2826 	timevalclear(&zone->uz_ratecheck);
2827 
2828 	/* Count the number of duplicate names. */
2829 	cnt.name = arg->name;
2830 	cnt.count = 0;
2831 	zone_foreach(zone_count, &cnt);
2832 	zone->uz_namecnt = cnt.count;
2833 	ZONE_CROSS_LOCK_INIT(zone);
2834 
2835 	for (i = 0; i < vm_ndomains; i++) {
2836 		zdom = ZDOM_GET(zone, i);
2837 		ZDOM_LOCK_INIT(zone, zdom, (arg->flags & UMA_ZONE_MTXCLASS));
2838 		STAILQ_INIT(&zdom->uzd_buckets);
2839 	}
2840 
2841 #if defined(INVARIANTS) && !defined(KASAN) && !defined(KMSAN)
2842 	if (arg->uminit == trash_init && arg->fini == trash_fini)
2843 		zone->uz_flags |= UMA_ZFLAG_TRASH | UMA_ZFLAG_CTORDTOR;
2844 #elif defined(KASAN)
2845 	if ((arg->flags & (UMA_ZONE_NOFREE | UMA_ZFLAG_CACHE)) != 0)
2846 		arg->flags |= UMA_ZONE_NOKASAN;
2847 #endif
2848 
2849 	/*
2850 	 * This is a pure cache zone, no kegs.
2851 	 */
2852 	if (arg->import) {
2853 		KASSERT((arg->flags & UMA_ZFLAG_CACHE) != 0,
2854 		    ("zone_ctor: Import specified for non-cache zone."));
2855 		zone->uz_flags = arg->flags;
2856 		zone->uz_size = arg->size;
2857 		zone->uz_import = arg->import;
2858 		zone->uz_release = arg->release;
2859 		zone->uz_arg = arg->arg;
2860 #ifdef NUMA
2861 		/*
2862 		 * Cache zones are round-robin unless a policy is
2863 		 * specified because they may have incompatible
2864 		 * constraints.
2865 		 */
2866 		if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2867 			zone->uz_flags |= UMA_ZONE_ROUNDROBIN;
2868 #endif
2869 		rw_wlock(&uma_rwlock);
2870 		LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link);
2871 		rw_wunlock(&uma_rwlock);
2872 		goto out;
2873 	}
2874 
2875 	/*
2876 	 * Use the regular zone/keg/slab allocator.
2877 	 */
2878 	zone->uz_import = zone_import;
2879 	zone->uz_release = zone_release;
2880 	zone->uz_arg = zone;
2881 	keg = arg->keg;
2882 
2883 	if (arg->flags & UMA_ZONE_SECONDARY) {
2884 		KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
2885 		    ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
2886 		KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
2887 		zone->uz_init = arg->uminit;
2888 		zone->uz_fini = arg->fini;
2889 		zone->uz_flags |= UMA_ZONE_SECONDARY;
2890 		rw_wlock(&uma_rwlock);
2891 		ZONE_LOCK(zone);
2892 		LIST_FOREACH(z, &keg->uk_zones, uz_link) {
2893 			if (LIST_NEXT(z, uz_link) == NULL) {
2894 				LIST_INSERT_AFTER(z, zone, uz_link);
2895 				break;
2896 			}
2897 		}
2898 		ZONE_UNLOCK(zone);
2899 		rw_wunlock(&uma_rwlock);
2900 	} else if (keg == NULL) {
2901 		if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
2902 		    arg->align, arg->flags)) == NULL)
2903 			return (ENOMEM);
2904 	} else {
2905 		struct uma_kctor_args karg;
2906 		int error;
2907 
2908 		/* We should only be here from uma_startup() */
2909 		karg.size = arg->size;
2910 		karg.uminit = arg->uminit;
2911 		karg.fini = arg->fini;
2912 		karg.align = arg->align;
2913 		karg.flags = (arg->flags & ~UMA_ZONE_SMR);
2914 		karg.zone = zone;
2915 		error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
2916 		    flags);
2917 		if (error)
2918 			return (error);
2919 	}
2920 
2921 	/* Inherit properties from the keg. */
2922 	zone->uz_keg = keg;
2923 	zone->uz_size = keg->uk_size;
2924 	zone->uz_flags |= (keg->uk_flags &
2925 	    (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
2926 
2927 out:
2928 	if (booted >= BOOT_PCPU) {
2929 		zone_alloc_counters(zone, NULL);
2930 		if (booted >= BOOT_RUNNING)
2931 			zone_alloc_sysctl(zone, NULL);
2932 	} else {
2933 		zone->uz_allocs = EARLY_COUNTER;
2934 		zone->uz_frees = EARLY_COUNTER;
2935 		zone->uz_fails = EARLY_COUNTER;
2936 	}
2937 
2938 	/* Caller requests a private SMR context. */
2939 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
2940 		zone->uz_smr = smr_create(zone->uz_name, 0, 0);
2941 
2942 	KASSERT((arg->flags & (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET)) !=
2943 	    (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET),
2944 	    ("Invalid zone flag combination"));
2945 	if (arg->flags & UMA_ZFLAG_INTERNAL)
2946 		zone->uz_bucket_size_max = zone->uz_bucket_size = 0;
2947 	if ((arg->flags & UMA_ZONE_MAXBUCKET) != 0)
2948 		zone->uz_bucket_size = BUCKET_MAX;
2949 	else if ((arg->flags & UMA_ZONE_NOBUCKET) != 0)
2950 		zone->uz_bucket_size = 0;
2951 	else
2952 		zone->uz_bucket_size = bucket_select(zone->uz_size);
2953 	zone->uz_bucket_size_min = zone->uz_bucket_size;
2954 	if (zone->uz_dtor != NULL || zone->uz_ctor != NULL)
2955 		zone->uz_flags |= UMA_ZFLAG_CTORDTOR;
2956 	zone_update_caches(zone);
2957 
2958 	return (0);
2959 }
2960 
2961 /*
2962  * Keg header dtor.  This frees all data, destroys locks, frees the hash
2963  * table and removes the keg from the global list.
2964  *
2965  * Arguments/Returns follow uma_dtor specifications
2966  *	udata  unused
2967  */
2968 static void
2969 keg_dtor(void *arg, int size, void *udata)
2970 {
2971 	uma_keg_t keg;
2972 	uint32_t free, pages;
2973 	int i;
2974 
2975 	keg = (uma_keg_t)arg;
2976 	free = pages = 0;
2977 	for (i = 0; i < vm_ndomains; i++) {
2978 		free += keg->uk_domain[i].ud_free_items;
2979 		pages += keg->uk_domain[i].ud_pages;
2980 		KEG_LOCK_FINI(keg, i);
2981 	}
2982 	if (pages != 0)
2983 		printf("Freed UMA keg (%s) was not empty (%u items). "
2984 		    " Lost %u pages of memory.\n",
2985 		    keg->uk_name ? keg->uk_name : "",
2986 		    pages / keg->uk_ppera * keg->uk_ipers - free, pages);
2987 
2988 	hash_free(&keg->uk_hash);
2989 }
2990 
2991 /*
2992  * Zone header dtor.
2993  *
2994  * Arguments/Returns follow uma_dtor specifications
2995  *	udata  unused
2996  */
2997 static void
2998 zone_dtor(void *arg, int size, void *udata)
2999 {
3000 	uma_zone_t zone;
3001 	uma_keg_t keg;
3002 	int i;
3003 
3004 	zone = (uma_zone_t)arg;
3005 
3006 	sysctl_remove_oid(zone->uz_oid, 1, 1);
3007 
3008 	if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
3009 		cache_drain(zone);
3010 
3011 	rw_wlock(&uma_rwlock);
3012 	LIST_REMOVE(zone, uz_link);
3013 	rw_wunlock(&uma_rwlock);
3014 	if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) {
3015 		keg = zone->uz_keg;
3016 		keg->uk_reserve = 0;
3017 	}
3018 	zone_reclaim(zone, UMA_ANYDOMAIN, M_WAITOK, true);
3019 
3020 	/*
3021 	 * We only destroy kegs from non secondary/non cache zones.
3022 	 */
3023 	if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) {
3024 		keg = zone->uz_keg;
3025 		rw_wlock(&uma_rwlock);
3026 		LIST_REMOVE(keg, uk_link);
3027 		rw_wunlock(&uma_rwlock);
3028 		zone_free_item(kegs, keg, NULL, SKIP_NONE);
3029 	}
3030 	counter_u64_free(zone->uz_allocs);
3031 	counter_u64_free(zone->uz_frees);
3032 	counter_u64_free(zone->uz_fails);
3033 	counter_u64_free(zone->uz_xdomain);
3034 	free(zone->uz_ctlname, M_UMA);
3035 	for (i = 0; i < vm_ndomains; i++)
3036 		ZDOM_LOCK_FINI(ZDOM_GET(zone, i));
3037 	ZONE_CROSS_LOCK_FINI(zone);
3038 }
3039 
3040 static void
3041 zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *arg), void *arg)
3042 {
3043 	uma_keg_t keg;
3044 	uma_zone_t zone;
3045 
3046 	LIST_FOREACH(keg, &uma_kegs, uk_link) {
3047 		LIST_FOREACH(zone, &keg->uk_zones, uz_link)
3048 			zfunc(zone, arg);
3049 	}
3050 	LIST_FOREACH(zone, &uma_cachezones, uz_link)
3051 		zfunc(zone, arg);
3052 }
3053 
3054 /*
3055  * Traverses every zone in the system and calls a callback
3056  *
3057  * Arguments:
3058  *	zfunc  A pointer to a function which accepts a zone
3059  *		as an argument.
3060  *
3061  * Returns:
3062  *	Nothing
3063  */
3064 static void
3065 zone_foreach(void (*zfunc)(uma_zone_t, void *arg), void *arg)
3066 {
3067 
3068 	rw_rlock(&uma_rwlock);
3069 	zone_foreach_unlocked(zfunc, arg);
3070 	rw_runlock(&uma_rwlock);
3071 }
3072 
3073 /*
3074  * Initialize the kernel memory allocator.  This is done after pages can be
3075  * allocated but before general KVA is available.
3076  */
3077 void
3078 uma_startup1(vm_offset_t virtual_avail)
3079 {
3080 	struct uma_zctor_args args;
3081 	size_t ksize, zsize, size;
3082 	uma_keg_t primarykeg;
3083 	uintptr_t m;
3084 	int domain;
3085 	uint8_t pflag;
3086 
3087 	bootstart = bootmem = virtual_avail;
3088 
3089 	rw_init(&uma_rwlock, "UMA lock");
3090 	sx_init(&uma_reclaim_lock, "umareclaim");
3091 
3092 	ksize = sizeof(struct uma_keg) +
3093 	    (sizeof(struct uma_domain) * vm_ndomains);
3094 	ksize = roundup(ksize, UMA_SUPER_ALIGN);
3095 	zsize = sizeof(struct uma_zone) +
3096 	    (sizeof(struct uma_cache) * (mp_maxid + 1)) +
3097 	    (sizeof(struct uma_zone_domain) * vm_ndomains);
3098 	zsize = roundup(zsize, UMA_SUPER_ALIGN);
3099 
3100 	/* Allocate the zone of zones, zone of kegs, and zone of zones keg. */
3101 	size = (zsize * 2) + ksize;
3102 	for (domain = 0; domain < vm_ndomains; domain++) {
3103 		m = (uintptr_t)startup_alloc(NULL, size, domain, &pflag,
3104 		    M_NOWAIT | M_ZERO);
3105 		if (m != 0)
3106 			break;
3107 	}
3108 	zones = (uma_zone_t)m;
3109 	m += zsize;
3110 	kegs = (uma_zone_t)m;
3111 	m += zsize;
3112 	primarykeg = (uma_keg_t)m;
3113 
3114 	/* "manually" create the initial zone */
3115 	memset(&args, 0, sizeof(args));
3116 	args.name = "UMA Kegs";
3117 	args.size = ksize;
3118 	args.ctor = keg_ctor;
3119 	args.dtor = keg_dtor;
3120 	args.uminit = zero_init;
3121 	args.fini = NULL;
3122 	args.keg = primarykeg;
3123 	args.align = UMA_SUPER_ALIGN - 1;
3124 	args.flags = UMA_ZFLAG_INTERNAL;
3125 	zone_ctor(kegs, zsize, &args, M_WAITOK);
3126 
3127 	args.name = "UMA Zones";
3128 	args.size = zsize;
3129 	args.ctor = zone_ctor;
3130 	args.dtor = zone_dtor;
3131 	args.uminit = zero_init;
3132 	args.fini = NULL;
3133 	args.keg = NULL;
3134 	args.align = UMA_SUPER_ALIGN - 1;
3135 	args.flags = UMA_ZFLAG_INTERNAL;
3136 	zone_ctor(zones, zsize, &args, M_WAITOK);
3137 
3138 	/* Now make zones for slab headers */
3139 	slabzones[0] = uma_zcreate("UMA Slabs 0", SLABZONE0_SIZE,
3140 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
3141 	slabzones[1] = uma_zcreate("UMA Slabs 1", SLABZONE1_SIZE,
3142 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
3143 
3144 	hashzone = uma_zcreate("UMA Hash",
3145 	    sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
3146 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
3147 
3148 	bucket_init();
3149 	smr_init();
3150 }
3151 
3152 #ifndef UMA_MD_SMALL_ALLOC
3153 extern void vm_radix_reserve_kva(void);
3154 #endif
3155 
3156 /*
3157  * Advertise the availability of normal kva allocations and switch to
3158  * the default back-end allocator.  Marks the KVA we consumed on startup
3159  * as used in the map.
3160  */
3161 void
3162 uma_startup2(void)
3163 {
3164 
3165 	if (bootstart != bootmem) {
3166 		vm_map_lock(kernel_map);
3167 		(void)vm_map_insert(kernel_map, NULL, 0, bootstart, bootmem,
3168 		    VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT);
3169 		vm_map_unlock(kernel_map);
3170 	}
3171 
3172 #ifndef UMA_MD_SMALL_ALLOC
3173 	/* Set up radix zone to use noobj_alloc. */
3174 	vm_radix_reserve_kva();
3175 #endif
3176 
3177 	booted = BOOT_KVA;
3178 	zone_foreach_unlocked(zone_kva_available, NULL);
3179 	bucket_enable();
3180 }
3181 
3182 /*
3183  * Allocate counters as early as possible so that boot-time allocations are
3184  * accounted more precisely.
3185  */
3186 static void
3187 uma_startup_pcpu(void *arg __unused)
3188 {
3189 
3190 	zone_foreach_unlocked(zone_alloc_counters, NULL);
3191 	booted = BOOT_PCPU;
3192 }
3193 SYSINIT(uma_startup_pcpu, SI_SUB_COUNTER, SI_ORDER_ANY, uma_startup_pcpu, NULL);
3194 
3195 /*
3196  * Finish our initialization steps.
3197  */
3198 static void
3199 uma_startup3(void *arg __unused)
3200 {
3201 
3202 #ifdef INVARIANTS
3203 	TUNABLE_INT_FETCH("vm.debug.divisor", &dbg_divisor);
3204 	uma_dbg_cnt = counter_u64_alloc(M_WAITOK);
3205 	uma_skip_cnt = counter_u64_alloc(M_WAITOK);
3206 #endif
3207 	zone_foreach_unlocked(zone_alloc_sysctl, NULL);
3208 	callout_init(&uma_callout, 1);
3209 	callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
3210 	booted = BOOT_RUNNING;
3211 
3212 	EVENTHANDLER_REGISTER(shutdown_post_sync, uma_shutdown, NULL,
3213 	    EVENTHANDLER_PRI_FIRST);
3214 }
3215 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
3216 
3217 static void
3218 uma_shutdown(void)
3219 {
3220 
3221 	booted = BOOT_SHUTDOWN;
3222 }
3223 
3224 static uma_keg_t
3225 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
3226 		int align, uint32_t flags)
3227 {
3228 	struct uma_kctor_args args;
3229 
3230 	args.size = size;
3231 	args.uminit = uminit;
3232 	args.fini = fini;
3233 	args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
3234 	args.flags = flags;
3235 	args.zone = zone;
3236 	return (zone_alloc_item(kegs, &args, UMA_ANYDOMAIN, M_WAITOK));
3237 }
3238 
3239 /* Public functions */
3240 /* See uma.h */
3241 void
3242 uma_set_align(int align)
3243 {
3244 
3245 	if (align != UMA_ALIGN_CACHE)
3246 		uma_align_cache = align;
3247 }
3248 
3249 /* See uma.h */
3250 uma_zone_t
3251 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
3252 		uma_init uminit, uma_fini fini, int align, uint32_t flags)
3253 
3254 {
3255 	struct uma_zctor_args args;
3256 	uma_zone_t res;
3257 
3258 	KASSERT(powerof2(align + 1), ("invalid zone alignment %d for \"%s\"",
3259 	    align, name));
3260 
3261 	/* This stuff is essential for the zone ctor */
3262 	memset(&args, 0, sizeof(args));
3263 	args.name = name;
3264 	args.size = size;
3265 	args.ctor = ctor;
3266 	args.dtor = dtor;
3267 	args.uminit = uminit;
3268 	args.fini = fini;
3269 #if defined(INVARIANTS) && !defined(KASAN) && !defined(KMSAN)
3270 	/*
3271 	 * Inject procedures which check for memory use after free if we are
3272 	 * allowed to scramble the memory while it is not allocated.  This
3273 	 * requires that: UMA is actually able to access the memory, no init
3274 	 * or fini procedures, no dependency on the initial value of the
3275 	 * memory, and no (legitimate) use of the memory after free.  Note,
3276 	 * the ctor and dtor do not need to be empty.
3277 	 */
3278 	if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOTOUCH |
3279 	    UMA_ZONE_NOFREE))) && uminit == NULL && fini == NULL) {
3280 		args.uminit = trash_init;
3281 		args.fini = trash_fini;
3282 	}
3283 #endif
3284 	args.align = align;
3285 	args.flags = flags;
3286 	args.keg = NULL;
3287 
3288 	sx_xlock(&uma_reclaim_lock);
3289 	res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3290 	sx_xunlock(&uma_reclaim_lock);
3291 
3292 	return (res);
3293 }
3294 
3295 /* See uma.h */
3296 uma_zone_t
3297 uma_zsecond_create(const char *name, uma_ctor ctor, uma_dtor dtor,
3298     uma_init zinit, uma_fini zfini, uma_zone_t primary)
3299 {
3300 	struct uma_zctor_args args;
3301 	uma_keg_t keg;
3302 	uma_zone_t res;
3303 
3304 	keg = primary->uz_keg;
3305 	memset(&args, 0, sizeof(args));
3306 	args.name = name;
3307 	args.size = keg->uk_size;
3308 	args.ctor = ctor;
3309 	args.dtor = dtor;
3310 	args.uminit = zinit;
3311 	args.fini = zfini;
3312 	args.align = keg->uk_align;
3313 	args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
3314 	args.keg = keg;
3315 
3316 	sx_xlock(&uma_reclaim_lock);
3317 	res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3318 	sx_xunlock(&uma_reclaim_lock);
3319 
3320 	return (res);
3321 }
3322 
3323 /* See uma.h */
3324 uma_zone_t
3325 uma_zcache_create(const char *name, int size, uma_ctor ctor, uma_dtor dtor,
3326     uma_init zinit, uma_fini zfini, uma_import zimport, uma_release zrelease,
3327     void *arg, int flags)
3328 {
3329 	struct uma_zctor_args args;
3330 
3331 	memset(&args, 0, sizeof(args));
3332 	args.name = name;
3333 	args.size = size;
3334 	args.ctor = ctor;
3335 	args.dtor = dtor;
3336 	args.uminit = zinit;
3337 	args.fini = zfini;
3338 	args.import = zimport;
3339 	args.release = zrelease;
3340 	args.arg = arg;
3341 	args.align = 0;
3342 	args.flags = flags | UMA_ZFLAG_CACHE;
3343 
3344 	return (zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK));
3345 }
3346 
3347 /* See uma.h */
3348 void
3349 uma_zdestroy(uma_zone_t zone)
3350 {
3351 
3352 	/*
3353 	 * Large slabs are expensive to reclaim, so don't bother doing
3354 	 * unnecessary work if we're shutting down.
3355 	 */
3356 	if (booted == BOOT_SHUTDOWN &&
3357 	    zone->uz_fini == NULL && zone->uz_release == zone_release)
3358 		return;
3359 	sx_xlock(&uma_reclaim_lock);
3360 	zone_free_item(zones, zone, NULL, SKIP_NONE);
3361 	sx_xunlock(&uma_reclaim_lock);
3362 }
3363 
3364 void
3365 uma_zwait(uma_zone_t zone)
3366 {
3367 
3368 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
3369 		uma_zfree_smr(zone, uma_zalloc_smr(zone, M_WAITOK));
3370 	else if ((zone->uz_flags & UMA_ZONE_PCPU) != 0)
3371 		uma_zfree_pcpu(zone, uma_zalloc_pcpu(zone, M_WAITOK));
3372 	else
3373 		uma_zfree(zone, uma_zalloc(zone, M_WAITOK));
3374 }
3375 
3376 void *
3377 uma_zalloc_pcpu_arg(uma_zone_t zone, void *udata, int flags)
3378 {
3379 	void *item, *pcpu_item;
3380 #ifdef SMP
3381 	int i;
3382 
3383 	MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3384 #endif
3385 	item = uma_zalloc_arg(zone, udata, flags & ~M_ZERO);
3386 	if (item == NULL)
3387 		return (NULL);
3388 	pcpu_item = zpcpu_base_to_offset(item);
3389 	if (flags & M_ZERO) {
3390 #ifdef SMP
3391 		for (i = 0; i <= mp_maxid; i++)
3392 			bzero(zpcpu_get_cpu(pcpu_item, i), zone->uz_size);
3393 #else
3394 		bzero(item, zone->uz_size);
3395 #endif
3396 	}
3397 	return (pcpu_item);
3398 }
3399 
3400 /*
3401  * A stub while both regular and pcpu cases are identical.
3402  */
3403 void
3404 uma_zfree_pcpu_arg(uma_zone_t zone, void *pcpu_item, void *udata)
3405 {
3406 	void *item;
3407 
3408 #ifdef SMP
3409 	MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3410 #endif
3411 
3412         /* uma_zfree_pcu_*(..., NULL) does nothing, to match free(9). */
3413         if (pcpu_item == NULL)
3414                 return;
3415 
3416 	item = zpcpu_offset_to_base(pcpu_item);
3417 	uma_zfree_arg(zone, item, udata);
3418 }
3419 
3420 static inline void *
3421 item_ctor(uma_zone_t zone, int uz_flags, int size, void *udata, int flags,
3422     void *item)
3423 {
3424 #ifdef INVARIANTS
3425 	bool skipdbg;
3426 #endif
3427 
3428 	kasan_mark_item_valid(zone, item);
3429 	kmsan_mark_item_uninitialized(zone, item);
3430 
3431 #ifdef INVARIANTS
3432 	skipdbg = uma_dbg_zskip(zone, item);
3433 	if (!skipdbg && (uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3434 	    zone->uz_ctor != trash_ctor)
3435 		trash_ctor(item, size, udata, flags);
3436 #endif
3437 
3438 	/* Check flags before loading ctor pointer. */
3439 	if (__predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0) &&
3440 	    __predict_false(zone->uz_ctor != NULL) &&
3441 	    zone->uz_ctor(item, size, udata, flags) != 0) {
3442 		counter_u64_add(zone->uz_fails, 1);
3443 		zone_free_item(zone, item, udata, SKIP_DTOR | SKIP_CNT);
3444 		return (NULL);
3445 	}
3446 #ifdef INVARIANTS
3447 	if (!skipdbg)
3448 		uma_dbg_alloc(zone, NULL, item);
3449 #endif
3450 	if (__predict_false(flags & M_ZERO))
3451 		return (memset(item, 0, size));
3452 
3453 	return (item);
3454 }
3455 
3456 static inline void
3457 item_dtor(uma_zone_t zone, void *item, int size, void *udata,
3458     enum zfreeskip skip)
3459 {
3460 #ifdef INVARIANTS
3461 	bool skipdbg;
3462 
3463 	skipdbg = uma_dbg_zskip(zone, item);
3464 	if (skip == SKIP_NONE && !skipdbg) {
3465 		if ((zone->uz_flags & UMA_ZONE_MALLOC) != 0)
3466 			uma_dbg_free(zone, udata, item);
3467 		else
3468 			uma_dbg_free(zone, NULL, item);
3469 	}
3470 #endif
3471 	if (__predict_true(skip < SKIP_DTOR)) {
3472 		if (zone->uz_dtor != NULL)
3473 			zone->uz_dtor(item, size, udata);
3474 #ifdef INVARIANTS
3475 		if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3476 		    zone->uz_dtor != trash_dtor)
3477 			trash_dtor(item, size, udata);
3478 #endif
3479 	}
3480 	kasan_mark_item_invalid(zone, item);
3481 }
3482 
3483 #ifdef NUMA
3484 static int
3485 item_domain(void *item)
3486 {
3487 	int domain;
3488 
3489 	domain = vm_phys_domain(vtophys(item));
3490 	KASSERT(domain >= 0 && domain < vm_ndomains,
3491 	    ("%s: unknown domain for item %p", __func__, item));
3492 	return (domain);
3493 }
3494 #endif
3495 
3496 #if defined(INVARIANTS) || defined(DEBUG_MEMGUARD) || defined(WITNESS)
3497 #if defined(INVARIANTS) && (defined(DDB) || defined(STACK))
3498 #include <sys/stack.h>
3499 #endif
3500 #define	UMA_ZALLOC_DEBUG
3501 static int
3502 uma_zalloc_debug(uma_zone_t zone, void **itemp, void *udata, int flags)
3503 {
3504 	int error;
3505 
3506 	error = 0;
3507 #ifdef WITNESS
3508 	if (flags & M_WAITOK) {
3509 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
3510 		    "uma_zalloc_debug: zone \"%s\"", zone->uz_name);
3511 	}
3512 #endif
3513 
3514 #ifdef INVARIANTS
3515 	KASSERT((flags & M_EXEC) == 0,
3516 	    ("uma_zalloc_debug: called with M_EXEC"));
3517 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3518 	    ("uma_zalloc_debug: called within spinlock or critical section"));
3519 	KASSERT((zone->uz_flags & UMA_ZONE_PCPU) == 0 || (flags & M_ZERO) == 0,
3520 	    ("uma_zalloc_debug: allocating from a pcpu zone with M_ZERO"));
3521 
3522 	_Static_assert(M_NOWAIT != 0 && M_WAITOK != 0,
3523 	    "M_NOWAIT and M_WAITOK must be non-zero for this assertion:");
3524 #if 0
3525 	/*
3526 	 * Give the #elif clause time to find problems, then remove it
3527 	 * and enable this.  (Remove <sys/stack.h> above, too.)
3528 	 */
3529 	KASSERT((flags & (M_NOWAIT|M_WAITOK)) == M_NOWAIT ||
3530 	    (flags & (M_NOWAIT|M_WAITOK)) == M_WAITOK,
3531 	    ("uma_zalloc_debug: must pass one of M_NOWAIT or M_WAITOK"));
3532 #elif defined(DDB) || defined(STACK)
3533 	if (__predict_false((flags & (M_NOWAIT|M_WAITOK)) != M_NOWAIT &&
3534 	    (flags & (M_NOWAIT|M_WAITOK)) != M_WAITOK)) {
3535 		static int stack_count;
3536 		struct stack st;
3537 
3538 		if (stack_count < 10) {
3539 			++stack_count;
3540 			printf("uma_zalloc* called with bad WAIT flags:\n");
3541 			stack_save(&st);
3542 			stack_print(&st);
3543 		}
3544 	}
3545 #endif
3546 #endif
3547 
3548 #ifdef DEBUG_MEMGUARD
3549 	if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && memguard_cmp_zone(zone)) {
3550 		void *item;
3551 		item = memguard_alloc(zone->uz_size, flags);
3552 		if (item != NULL) {
3553 			error = EJUSTRETURN;
3554 			if (zone->uz_init != NULL &&
3555 			    zone->uz_init(item, zone->uz_size, flags) != 0) {
3556 				*itemp = NULL;
3557 				return (error);
3558 			}
3559 			if (zone->uz_ctor != NULL &&
3560 			    zone->uz_ctor(item, zone->uz_size, udata,
3561 			    flags) != 0) {
3562 				counter_u64_add(zone->uz_fails, 1);
3563 				if (zone->uz_fini != NULL)
3564 					zone->uz_fini(item, zone->uz_size);
3565 				*itemp = NULL;
3566 				return (error);
3567 			}
3568 			*itemp = item;
3569 			return (error);
3570 		}
3571 		/* This is unfortunate but should not be fatal. */
3572 	}
3573 #endif
3574 	return (error);
3575 }
3576 
3577 static int
3578 uma_zfree_debug(uma_zone_t zone, void *item, void *udata)
3579 {
3580 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3581 	    ("uma_zfree_debug: called with spinlock or critical section held"));
3582 
3583 #ifdef DEBUG_MEMGUARD
3584 	if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && is_memguard_addr(item)) {
3585 		if (zone->uz_dtor != NULL)
3586 			zone->uz_dtor(item, zone->uz_size, udata);
3587 		if (zone->uz_fini != NULL)
3588 			zone->uz_fini(item, zone->uz_size);
3589 		memguard_free(item);
3590 		return (EJUSTRETURN);
3591 	}
3592 #endif
3593 	return (0);
3594 }
3595 #endif
3596 
3597 static inline void *
3598 cache_alloc_item(uma_zone_t zone, uma_cache_t cache, uma_cache_bucket_t bucket,
3599     void *udata, int flags)
3600 {
3601 	void *item;
3602 	int size, uz_flags;
3603 
3604 	item = cache_bucket_pop(cache, bucket);
3605 	size = cache_uz_size(cache);
3606 	uz_flags = cache_uz_flags(cache);
3607 	critical_exit();
3608 	return (item_ctor(zone, uz_flags, size, udata, flags, item));
3609 }
3610 
3611 static __noinline void *
3612 cache_alloc_retry(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3613 {
3614 	uma_cache_bucket_t bucket;
3615 	int domain;
3616 
3617 	while (cache_alloc(zone, cache, udata, flags)) {
3618 		cache = &zone->uz_cpu[curcpu];
3619 		bucket = &cache->uc_allocbucket;
3620 		if (__predict_false(bucket->ucb_cnt == 0))
3621 			continue;
3622 		return (cache_alloc_item(zone, cache, bucket, udata, flags));
3623 	}
3624 	critical_exit();
3625 
3626 	/*
3627 	 * We can not get a bucket so try to return a single item.
3628 	 */
3629 	if (zone->uz_flags & UMA_ZONE_FIRSTTOUCH)
3630 		domain = PCPU_GET(domain);
3631 	else
3632 		domain = UMA_ANYDOMAIN;
3633 	return (zone_alloc_item(zone, udata, domain, flags));
3634 }
3635 
3636 /* See uma.h */
3637 void *
3638 uma_zalloc_smr(uma_zone_t zone, int flags)
3639 {
3640 	uma_cache_bucket_t bucket;
3641 	uma_cache_t cache;
3642 
3643 	CTR3(KTR_UMA, "uma_zalloc_smr zone %s(%p) flags %d", zone->uz_name,
3644 	    zone, flags);
3645 
3646 #ifdef UMA_ZALLOC_DEBUG
3647 	void *item;
3648 
3649 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
3650 	    ("uma_zalloc_arg: called with non-SMR zone."));
3651 	if (uma_zalloc_debug(zone, &item, NULL, flags) == EJUSTRETURN)
3652 		return (item);
3653 #endif
3654 
3655 	critical_enter();
3656 	cache = &zone->uz_cpu[curcpu];
3657 	bucket = &cache->uc_allocbucket;
3658 	if (__predict_false(bucket->ucb_cnt == 0))
3659 		return (cache_alloc_retry(zone, cache, NULL, flags));
3660 	return (cache_alloc_item(zone, cache, bucket, NULL, flags));
3661 }
3662 
3663 /* See uma.h */
3664 void *
3665 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
3666 {
3667 	uma_cache_bucket_t bucket;
3668 	uma_cache_t cache;
3669 
3670 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3671 	random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3672 
3673 	/* This is the fast path allocation */
3674 	CTR3(KTR_UMA, "uma_zalloc_arg zone %s(%p) flags %d", zone->uz_name,
3675 	    zone, flags);
3676 
3677 #ifdef UMA_ZALLOC_DEBUG
3678 	void *item;
3679 
3680 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3681 	    ("uma_zalloc_arg: called with SMR zone."));
3682 	if (uma_zalloc_debug(zone, &item, udata, flags) == EJUSTRETURN)
3683 		return (item);
3684 #endif
3685 
3686 	/*
3687 	 * If possible, allocate from the per-CPU cache.  There are two
3688 	 * requirements for safe access to the per-CPU cache: (1) the thread
3689 	 * accessing the cache must not be preempted or yield during access,
3690 	 * and (2) the thread must not migrate CPUs without switching which
3691 	 * cache it accesses.  We rely on a critical section to prevent
3692 	 * preemption and migration.  We release the critical section in
3693 	 * order to acquire the zone mutex if we are unable to allocate from
3694 	 * the current cache; when we re-acquire the critical section, we
3695 	 * must detect and handle migration if it has occurred.
3696 	 */
3697 	critical_enter();
3698 	cache = &zone->uz_cpu[curcpu];
3699 	bucket = &cache->uc_allocbucket;
3700 	if (__predict_false(bucket->ucb_cnt == 0))
3701 		return (cache_alloc_retry(zone, cache, udata, flags));
3702 	return (cache_alloc_item(zone, cache, bucket, udata, flags));
3703 }
3704 
3705 /*
3706  * Replenish an alloc bucket and possibly restore an old one.  Called in
3707  * a critical section.  Returns in a critical section.
3708  *
3709  * A false return value indicates an allocation failure.
3710  * A true return value indicates success and the caller should retry.
3711  */
3712 static __noinline bool
3713 cache_alloc(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3714 {
3715 	uma_bucket_t bucket;
3716 	int curdomain, domain;
3717 	bool new;
3718 
3719 	CRITICAL_ASSERT(curthread);
3720 
3721 	/*
3722 	 * If we have run out of items in our alloc bucket see
3723 	 * if we can switch with the free bucket.
3724 	 *
3725 	 * SMR Zones can't re-use the free bucket until the sequence has
3726 	 * expired.
3727 	 */
3728 	if ((cache_uz_flags(cache) & UMA_ZONE_SMR) == 0 &&
3729 	    cache->uc_freebucket.ucb_cnt != 0) {
3730 		cache_bucket_swap(&cache->uc_freebucket,
3731 		    &cache->uc_allocbucket);
3732 		return (true);
3733 	}
3734 
3735 	/*
3736 	 * Discard any empty allocation bucket while we hold no locks.
3737 	 */
3738 	bucket = cache_bucket_unload_alloc(cache);
3739 	critical_exit();
3740 
3741 	if (bucket != NULL) {
3742 		KASSERT(bucket->ub_cnt == 0,
3743 		    ("cache_alloc: Entered with non-empty alloc bucket."));
3744 		bucket_free(zone, bucket, udata);
3745 	}
3746 
3747 	/*
3748 	 * Attempt to retrieve the item from the per-CPU cache has failed, so
3749 	 * we must go back to the zone.  This requires the zdom lock, so we
3750 	 * must drop the critical section, then re-acquire it when we go back
3751 	 * to the cache.  Since the critical section is released, we may be
3752 	 * preempted or migrate.  As such, make sure not to maintain any
3753 	 * thread-local state specific to the cache from prior to releasing
3754 	 * the critical section.
3755 	 */
3756 	domain = PCPU_GET(domain);
3757 	if ((cache_uz_flags(cache) & UMA_ZONE_ROUNDROBIN) != 0 ||
3758 	    VM_DOMAIN_EMPTY(domain))
3759 		domain = zone_domain_highest(zone, domain);
3760 	bucket = cache_fetch_bucket(zone, cache, domain);
3761 	if (bucket == NULL && zone->uz_bucket_size != 0 && !bucketdisable) {
3762 		bucket = zone_alloc_bucket(zone, udata, domain, flags);
3763 		new = true;
3764 	} else {
3765 		new = false;
3766 	}
3767 
3768 	CTR3(KTR_UMA, "uma_zalloc: zone %s(%p) bucket zone returned %p",
3769 	    zone->uz_name, zone, bucket);
3770 	if (bucket == NULL) {
3771 		critical_enter();
3772 		return (false);
3773 	}
3774 
3775 	/*
3776 	 * See if we lost the race or were migrated.  Cache the
3777 	 * initialized bucket to make this less likely or claim
3778 	 * the memory directly.
3779 	 */
3780 	critical_enter();
3781 	cache = &zone->uz_cpu[curcpu];
3782 	if (cache->uc_allocbucket.ucb_bucket == NULL &&
3783 	    ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) == 0 ||
3784 	    (curdomain = PCPU_GET(domain)) == domain ||
3785 	    VM_DOMAIN_EMPTY(curdomain))) {
3786 		if (new)
3787 			atomic_add_long(&ZDOM_GET(zone, domain)->uzd_imax,
3788 			    bucket->ub_cnt);
3789 		cache_bucket_load_alloc(cache, bucket);
3790 		return (true);
3791 	}
3792 
3793 	/*
3794 	 * We lost the race, release this bucket and start over.
3795 	 */
3796 	critical_exit();
3797 	zone_put_bucket(zone, domain, bucket, udata, !new);
3798 	critical_enter();
3799 
3800 	return (true);
3801 }
3802 
3803 void *
3804 uma_zalloc_domain(uma_zone_t zone, void *udata, int domain, int flags)
3805 {
3806 #ifdef NUMA
3807 	uma_bucket_t bucket;
3808 	uma_zone_domain_t zdom;
3809 	void *item;
3810 #endif
3811 
3812 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3813 	random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3814 
3815 	/* This is the fast path allocation */
3816 	CTR4(KTR_UMA, "uma_zalloc_domain zone %s(%p) domain %d flags %d",
3817 	    zone->uz_name, zone, domain, flags);
3818 
3819 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3820 	    ("uma_zalloc_domain: called with SMR zone."));
3821 #ifdef NUMA
3822 	KASSERT((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0,
3823 	    ("uma_zalloc_domain: called with non-FIRSTTOUCH zone."));
3824 
3825 	if (vm_ndomains == 1)
3826 		return (uma_zalloc_arg(zone, udata, flags));
3827 
3828 #ifdef UMA_ZALLOC_DEBUG
3829 	if (uma_zalloc_debug(zone, &item, udata, flags) == EJUSTRETURN)
3830 		return (item);
3831 #endif
3832 
3833 	/*
3834 	 * Try to allocate from the bucket cache before falling back to the keg.
3835 	 * We could try harder and attempt to allocate from per-CPU caches or
3836 	 * the per-domain cross-domain buckets, but the complexity is probably
3837 	 * not worth it.  It is more important that frees of previous
3838 	 * cross-domain allocations do not blow up the cache.
3839 	 */
3840 	zdom = zone_domain_lock(zone, domain);
3841 	if ((bucket = zone_fetch_bucket(zone, zdom, false)) != NULL) {
3842 		item = bucket->ub_bucket[bucket->ub_cnt - 1];
3843 #ifdef INVARIANTS
3844 		bucket->ub_bucket[bucket->ub_cnt - 1] = NULL;
3845 #endif
3846 		bucket->ub_cnt--;
3847 		zone_put_bucket(zone, domain, bucket, udata, true);
3848 		item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata,
3849 		    flags, item);
3850 		if (item != NULL) {
3851 			KASSERT(item_domain(item) == domain,
3852 			    ("%s: bucket cache item %p from wrong domain",
3853 			    __func__, item));
3854 			counter_u64_add(zone->uz_allocs, 1);
3855 		}
3856 		return (item);
3857 	}
3858 	ZDOM_UNLOCK(zdom);
3859 	return (zone_alloc_item(zone, udata, domain, flags));
3860 #else
3861 	return (uma_zalloc_arg(zone, udata, flags));
3862 #endif
3863 }
3864 
3865 /*
3866  * Find a slab with some space.  Prefer slabs that are partially used over those
3867  * that are totally full.  This helps to reduce fragmentation.
3868  *
3869  * If 'rr' is 1, search all domains starting from 'domain'.  Otherwise check
3870  * only 'domain'.
3871  */
3872 static uma_slab_t
3873 keg_first_slab(uma_keg_t keg, int domain, bool rr)
3874 {
3875 	uma_domain_t dom;
3876 	uma_slab_t slab;
3877 	int start;
3878 
3879 	KASSERT(domain >= 0 && domain < vm_ndomains,
3880 	    ("keg_first_slab: domain %d out of range", domain));
3881 	KEG_LOCK_ASSERT(keg, domain);
3882 
3883 	slab = NULL;
3884 	start = domain;
3885 	do {
3886 		dom = &keg->uk_domain[domain];
3887 		if ((slab = LIST_FIRST(&dom->ud_part_slab)) != NULL)
3888 			return (slab);
3889 		if ((slab = LIST_FIRST(&dom->ud_free_slab)) != NULL) {
3890 			LIST_REMOVE(slab, us_link);
3891 			dom->ud_free_slabs--;
3892 			LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
3893 			return (slab);
3894 		}
3895 		if (rr)
3896 			domain = (domain + 1) % vm_ndomains;
3897 	} while (domain != start);
3898 
3899 	return (NULL);
3900 }
3901 
3902 /*
3903  * Fetch an existing slab from a free or partial list.  Returns with the
3904  * keg domain lock held if a slab was found or unlocked if not.
3905  */
3906 static uma_slab_t
3907 keg_fetch_free_slab(uma_keg_t keg, int domain, bool rr, int flags)
3908 {
3909 	uma_slab_t slab;
3910 	uint32_t reserve;
3911 
3912 	/* HASH has a single free list. */
3913 	if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
3914 		domain = 0;
3915 
3916 	KEG_LOCK(keg, domain);
3917 	reserve = (flags & M_USE_RESERVE) != 0 ? 0 : keg->uk_reserve;
3918 	if (keg->uk_domain[domain].ud_free_items <= reserve ||
3919 	    (slab = keg_first_slab(keg, domain, rr)) == NULL) {
3920 		KEG_UNLOCK(keg, domain);
3921 		return (NULL);
3922 	}
3923 	return (slab);
3924 }
3925 
3926 static uma_slab_t
3927 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int rdomain, const int flags)
3928 {
3929 	struct vm_domainset_iter di;
3930 	uma_slab_t slab;
3931 	int aflags, domain;
3932 	bool rr;
3933 
3934 	KASSERT((flags & (M_WAITOK | M_NOVM)) != (M_WAITOK | M_NOVM),
3935 	    ("%s: invalid flags %#x", __func__, flags));
3936 
3937 restart:
3938 	/*
3939 	 * Use the keg's policy if upper layers haven't already specified a
3940 	 * domain (as happens with first-touch zones).
3941 	 *
3942 	 * To avoid races we run the iterator with the keg lock held, but that
3943 	 * means that we cannot allow the vm_domainset layer to sleep.  Thus,
3944 	 * clear M_WAITOK and handle low memory conditions locally.
3945 	 */
3946 	rr = rdomain == UMA_ANYDOMAIN;
3947 	if (rr) {
3948 		aflags = (flags & ~M_WAITOK) | M_NOWAIT;
3949 		vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
3950 		    &aflags);
3951 	} else {
3952 		aflags = flags;
3953 		domain = rdomain;
3954 	}
3955 
3956 	for (;;) {
3957 		slab = keg_fetch_free_slab(keg, domain, rr, flags);
3958 		if (slab != NULL)
3959 			return (slab);
3960 
3961 		/*
3962 		 * M_NOVM is used to break the recursion that can otherwise
3963 		 * occur if low-level memory management routines use UMA.
3964 		 */
3965 		if ((flags & M_NOVM) == 0) {
3966 			slab = keg_alloc_slab(keg, zone, domain, flags, aflags);
3967 			if (slab != NULL)
3968 				return (slab);
3969 		}
3970 
3971 		if (!rr) {
3972 			if ((flags & M_USE_RESERVE) != 0) {
3973 				/*
3974 				 * Drain reserves from other domains before
3975 				 * giving up or sleeping.  It may be useful to
3976 				 * support per-domain reserves eventually.
3977 				 */
3978 				rdomain = UMA_ANYDOMAIN;
3979 				goto restart;
3980 			}
3981 			if ((flags & M_WAITOK) == 0)
3982 				break;
3983 			vm_wait_domain(domain);
3984 		} else if (vm_domainset_iter_policy(&di, &domain) != 0) {
3985 			if ((flags & M_WAITOK) != 0) {
3986 				vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask, 0);
3987 				goto restart;
3988 			}
3989 			break;
3990 		}
3991 	}
3992 
3993 	/*
3994 	 * We might not have been able to get a slab but another cpu
3995 	 * could have while we were unlocked.  Check again before we
3996 	 * fail.
3997 	 */
3998 	if ((slab = keg_fetch_free_slab(keg, domain, rr, flags)) != NULL)
3999 		return (slab);
4000 
4001 	return (NULL);
4002 }
4003 
4004 static void *
4005 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
4006 {
4007 	uma_domain_t dom;
4008 	void *item;
4009 	int freei;
4010 
4011 	KEG_LOCK_ASSERT(keg, slab->us_domain);
4012 
4013 	dom = &keg->uk_domain[slab->us_domain];
4014 	freei = BIT_FFS(keg->uk_ipers, &slab->us_free) - 1;
4015 	BIT_CLR(keg->uk_ipers, freei, &slab->us_free);
4016 	item = slab_item(slab, keg, freei);
4017 	slab->us_freecount--;
4018 	dom->ud_free_items--;
4019 
4020 	/*
4021 	 * Move this slab to the full list.  It must be on the partial list, so
4022 	 * we do not need to update the free slab count.  In particular,
4023 	 * keg_fetch_slab() always returns slabs on the partial list.
4024 	 */
4025 	if (slab->us_freecount == 0) {
4026 		LIST_REMOVE(slab, us_link);
4027 		LIST_INSERT_HEAD(&dom->ud_full_slab, slab, us_link);
4028 	}
4029 
4030 	return (item);
4031 }
4032 
4033 static int
4034 zone_import(void *arg, void **bucket, int max, int domain, int flags)
4035 {
4036 	uma_domain_t dom;
4037 	uma_zone_t zone;
4038 	uma_slab_t slab;
4039 	uma_keg_t keg;
4040 #ifdef NUMA
4041 	int stripe;
4042 #endif
4043 	int i;
4044 
4045 	zone = arg;
4046 	slab = NULL;
4047 	keg = zone->uz_keg;
4048 	/* Try to keep the buckets totally full */
4049 	for (i = 0; i < max; ) {
4050 		if ((slab = keg_fetch_slab(keg, zone, domain, flags)) == NULL)
4051 			break;
4052 #ifdef NUMA
4053 		stripe = howmany(max, vm_ndomains);
4054 #endif
4055 		dom = &keg->uk_domain[slab->us_domain];
4056 		do {
4057 			bucket[i++] = slab_alloc_item(keg, slab);
4058 			if (keg->uk_reserve > 0 &&
4059 			    dom->ud_free_items <= keg->uk_reserve) {
4060 				/*
4061 				 * Avoid depleting the reserve after a
4062 				 * successful item allocation, even if
4063 				 * M_USE_RESERVE is specified.
4064 				 */
4065 				KEG_UNLOCK(keg, slab->us_domain);
4066 				goto out;
4067 			}
4068 #ifdef NUMA
4069 			/*
4070 			 * If the zone is striped we pick a new slab for every
4071 			 * N allocations.  Eliminating this conditional will
4072 			 * instead pick a new domain for each bucket rather
4073 			 * than stripe within each bucket.  The current option
4074 			 * produces more fragmentation and requires more cpu
4075 			 * time but yields better distribution.
4076 			 */
4077 			if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0 &&
4078 			    vm_ndomains > 1 && --stripe == 0)
4079 				break;
4080 #endif
4081 		} while (slab->us_freecount != 0 && i < max);
4082 		KEG_UNLOCK(keg, slab->us_domain);
4083 
4084 		/* Don't block if we allocated any successfully. */
4085 		flags &= ~M_WAITOK;
4086 		flags |= M_NOWAIT;
4087 	}
4088 out:
4089 	return i;
4090 }
4091 
4092 static int
4093 zone_alloc_limit_hard(uma_zone_t zone, int count, int flags)
4094 {
4095 	uint64_t old, new, total, max;
4096 
4097 	/*
4098 	 * The hard case.  We're going to sleep because there were existing
4099 	 * sleepers or because we ran out of items.  This routine enforces
4100 	 * fairness by keeping fifo order.
4101 	 *
4102 	 * First release our ill gotten gains and make some noise.
4103 	 */
4104 	for (;;) {
4105 		zone_free_limit(zone, count);
4106 		zone_log_warning(zone);
4107 		zone_maxaction(zone);
4108 		if (flags & M_NOWAIT)
4109 			return (0);
4110 
4111 		/*
4112 		 * We need to allocate an item or set ourself as a sleeper
4113 		 * while the sleepq lock is held to avoid wakeup races.  This
4114 		 * is essentially a home rolled semaphore.
4115 		 */
4116 		sleepq_lock(&zone->uz_max_items);
4117 		old = zone->uz_items;
4118 		do {
4119 			MPASS(UZ_ITEMS_SLEEPERS(old) < UZ_ITEMS_SLEEPERS_MAX);
4120 			/* Cache the max since we will evaluate twice. */
4121 			max = zone->uz_max_items;
4122 			if (UZ_ITEMS_SLEEPERS(old) != 0 ||
4123 			    UZ_ITEMS_COUNT(old) >= max)
4124 				new = old + UZ_ITEMS_SLEEPER;
4125 			else
4126 				new = old + MIN(count, max - old);
4127 		} while (atomic_fcmpset_64(&zone->uz_items, &old, new) == 0);
4128 
4129 		/* We may have successfully allocated under the sleepq lock. */
4130 		if (UZ_ITEMS_SLEEPERS(new) == 0) {
4131 			sleepq_release(&zone->uz_max_items);
4132 			return (new - old);
4133 		}
4134 
4135 		/*
4136 		 * This is in a different cacheline from uz_items so that we
4137 		 * don't constantly invalidate the fastpath cacheline when we
4138 		 * adjust item counts.  This could be limited to toggling on
4139 		 * transitions.
4140 		 */
4141 		atomic_add_32(&zone->uz_sleepers, 1);
4142 		atomic_add_64(&zone->uz_sleeps, 1);
4143 
4144 		/*
4145 		 * We have added ourselves as a sleeper.  The sleepq lock
4146 		 * protects us from wakeup races.  Sleep now and then retry.
4147 		 */
4148 		sleepq_add(&zone->uz_max_items, NULL, "zonelimit", 0, 0);
4149 		sleepq_wait(&zone->uz_max_items, PVM);
4150 
4151 		/*
4152 		 * After wakeup, remove ourselves as a sleeper and try
4153 		 * again.  We no longer have the sleepq lock for protection.
4154 		 *
4155 		 * Subract ourselves as a sleeper while attempting to add
4156 		 * our count.
4157 		 */
4158 		atomic_subtract_32(&zone->uz_sleepers, 1);
4159 		old = atomic_fetchadd_64(&zone->uz_items,
4160 		    -(UZ_ITEMS_SLEEPER - count));
4161 		/* We're no longer a sleeper. */
4162 		old -= UZ_ITEMS_SLEEPER;
4163 
4164 		/*
4165 		 * If we're still at the limit, restart.  Notably do not
4166 		 * block on other sleepers.  Cache the max value to protect
4167 		 * against changes via sysctl.
4168 		 */
4169 		total = UZ_ITEMS_COUNT(old);
4170 		max = zone->uz_max_items;
4171 		if (total >= max)
4172 			continue;
4173 		/* Truncate if necessary, otherwise wake other sleepers. */
4174 		if (total + count > max) {
4175 			zone_free_limit(zone, total + count - max);
4176 			count = max - total;
4177 		} else if (total + count < max && UZ_ITEMS_SLEEPERS(old) != 0)
4178 			wakeup_one(&zone->uz_max_items);
4179 
4180 		return (count);
4181 	}
4182 }
4183 
4184 /*
4185  * Allocate 'count' items from our max_items limit.  Returns the number
4186  * available.  If M_NOWAIT is not specified it will sleep until at least
4187  * one item can be allocated.
4188  */
4189 static int
4190 zone_alloc_limit(uma_zone_t zone, int count, int flags)
4191 {
4192 	uint64_t old;
4193 	uint64_t max;
4194 
4195 	max = zone->uz_max_items;
4196 	MPASS(max > 0);
4197 
4198 	/*
4199 	 * We expect normal allocations to succeed with a simple
4200 	 * fetchadd.
4201 	 */
4202 	old = atomic_fetchadd_64(&zone->uz_items, count);
4203 	if (__predict_true(old + count <= max))
4204 		return (count);
4205 
4206 	/*
4207 	 * If we had some items and no sleepers just return the
4208 	 * truncated value.  We have to release the excess space
4209 	 * though because that may wake sleepers who weren't woken
4210 	 * because we were temporarily over the limit.
4211 	 */
4212 	if (old < max) {
4213 		zone_free_limit(zone, (old + count) - max);
4214 		return (max - old);
4215 	}
4216 	return (zone_alloc_limit_hard(zone, count, flags));
4217 }
4218 
4219 /*
4220  * Free a number of items back to the limit.
4221  */
4222 static void
4223 zone_free_limit(uma_zone_t zone, int count)
4224 {
4225 	uint64_t old;
4226 
4227 	MPASS(count > 0);
4228 
4229 	/*
4230 	 * In the common case we either have no sleepers or
4231 	 * are still over the limit and can just return.
4232 	 */
4233 	old = atomic_fetchadd_64(&zone->uz_items, -count);
4234 	if (__predict_true(UZ_ITEMS_SLEEPERS(old) == 0 ||
4235 	   UZ_ITEMS_COUNT(old) - count >= zone->uz_max_items))
4236 		return;
4237 
4238 	/*
4239 	 * Moderate the rate of wakeups.  Sleepers will continue
4240 	 * to generate wakeups if necessary.
4241 	 */
4242 	wakeup_one(&zone->uz_max_items);
4243 }
4244 
4245 static uma_bucket_t
4246 zone_alloc_bucket(uma_zone_t zone, void *udata, int domain, int flags)
4247 {
4248 	uma_bucket_t bucket;
4249 	int error, maxbucket, cnt;
4250 
4251 	CTR3(KTR_UMA, "zone_alloc_bucket zone %s(%p) domain %d", zone->uz_name,
4252 	    zone, domain);
4253 
4254 	/* Avoid allocs targeting empty domains. */
4255 	if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
4256 		domain = UMA_ANYDOMAIN;
4257 	else if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
4258 		domain = UMA_ANYDOMAIN;
4259 
4260 	if (zone->uz_max_items > 0)
4261 		maxbucket = zone_alloc_limit(zone, zone->uz_bucket_size,
4262 		    M_NOWAIT);
4263 	else
4264 		maxbucket = zone->uz_bucket_size;
4265 	if (maxbucket == 0)
4266 		return (NULL);
4267 
4268 	/* Don't wait for buckets, preserve caller's NOVM setting. */
4269 	bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
4270 	if (bucket == NULL) {
4271 		cnt = 0;
4272 		goto out;
4273 	}
4274 
4275 	bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
4276 	    MIN(maxbucket, bucket->ub_entries), domain, flags);
4277 
4278 	/*
4279 	 * Initialize the memory if necessary.
4280 	 */
4281 	if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
4282 		int i;
4283 
4284 		for (i = 0; i < bucket->ub_cnt; i++) {
4285 			kasan_mark_item_valid(zone, bucket->ub_bucket[i]);
4286 			error = zone->uz_init(bucket->ub_bucket[i],
4287 			    zone->uz_size, flags);
4288 			kasan_mark_item_invalid(zone, bucket->ub_bucket[i]);
4289 			if (error != 0)
4290 				break;
4291 		}
4292 
4293 		/*
4294 		 * If we couldn't initialize the whole bucket, put the
4295 		 * rest back onto the freelist.
4296 		 */
4297 		if (i != bucket->ub_cnt) {
4298 			zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
4299 			    bucket->ub_cnt - i);
4300 #ifdef INVARIANTS
4301 			bzero(&bucket->ub_bucket[i],
4302 			    sizeof(void *) * (bucket->ub_cnt - i));
4303 #endif
4304 			bucket->ub_cnt = i;
4305 		}
4306 	}
4307 
4308 	cnt = bucket->ub_cnt;
4309 	if (bucket->ub_cnt == 0) {
4310 		bucket_free(zone, bucket, udata);
4311 		counter_u64_add(zone->uz_fails, 1);
4312 		bucket = NULL;
4313 	}
4314 out:
4315 	if (zone->uz_max_items > 0 && cnt < maxbucket)
4316 		zone_free_limit(zone, maxbucket - cnt);
4317 
4318 	return (bucket);
4319 }
4320 
4321 /*
4322  * Allocates a single item from a zone.
4323  *
4324  * Arguments
4325  *	zone   The zone to alloc for.
4326  *	udata  The data to be passed to the constructor.
4327  *	domain The domain to allocate from or UMA_ANYDOMAIN.
4328  *	flags  M_WAITOK, M_NOWAIT, M_ZERO.
4329  *
4330  * Returns
4331  *	NULL if there is no memory and M_NOWAIT is set
4332  *	An item if successful
4333  */
4334 
4335 static void *
4336 zone_alloc_item(uma_zone_t zone, void *udata, int domain, int flags)
4337 {
4338 	void *item;
4339 
4340 	if (zone->uz_max_items > 0 && zone_alloc_limit(zone, 1, flags) == 0) {
4341 		counter_u64_add(zone->uz_fails, 1);
4342 		return (NULL);
4343 	}
4344 
4345 	/* Avoid allocs targeting empty domains. */
4346 	if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
4347 		domain = UMA_ANYDOMAIN;
4348 
4349 	if (zone->uz_import(zone->uz_arg, &item, 1, domain, flags) != 1)
4350 		goto fail_cnt;
4351 
4352 	/*
4353 	 * We have to call both the zone's init (not the keg's init)
4354 	 * and the zone's ctor.  This is because the item is going from
4355 	 * a keg slab directly to the user, and the user is expecting it
4356 	 * to be both zone-init'd as well as zone-ctor'd.
4357 	 */
4358 	if (zone->uz_init != NULL) {
4359 		int error;
4360 
4361 		kasan_mark_item_valid(zone, item);
4362 		error = zone->uz_init(item, zone->uz_size, flags);
4363 		kasan_mark_item_invalid(zone, item);
4364 		if (error != 0) {
4365 			zone_free_item(zone, item, udata, SKIP_FINI | SKIP_CNT);
4366 			goto fail_cnt;
4367 		}
4368 	}
4369 	item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata, flags,
4370 	    item);
4371 	if (item == NULL)
4372 		goto fail;
4373 
4374 	counter_u64_add(zone->uz_allocs, 1);
4375 	CTR3(KTR_UMA, "zone_alloc_item item %p from %s(%p)", item,
4376 	    zone->uz_name, zone);
4377 
4378 	return (item);
4379 
4380 fail_cnt:
4381 	counter_u64_add(zone->uz_fails, 1);
4382 fail:
4383 	if (zone->uz_max_items > 0)
4384 		zone_free_limit(zone, 1);
4385 	CTR2(KTR_UMA, "zone_alloc_item failed from %s(%p)",
4386 	    zone->uz_name, zone);
4387 
4388 	return (NULL);
4389 }
4390 
4391 /* See uma.h */
4392 void
4393 uma_zfree_smr(uma_zone_t zone, void *item)
4394 {
4395 	uma_cache_t cache;
4396 	uma_cache_bucket_t bucket;
4397 	int itemdomain;
4398 #ifdef NUMA
4399 	int uz_flags;
4400 #endif
4401 
4402 	CTR3(KTR_UMA, "uma_zfree_smr zone %s(%p) item %p",
4403 	    zone->uz_name, zone, item);
4404 
4405 #ifdef UMA_ZALLOC_DEBUG
4406 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
4407 	    ("uma_zfree_smr: called with non-SMR zone."));
4408 	KASSERT(item != NULL, ("uma_zfree_smr: Called with NULL pointer."));
4409 	SMR_ASSERT_NOT_ENTERED(zone->uz_smr);
4410 	if (uma_zfree_debug(zone, item, NULL) == EJUSTRETURN)
4411 		return;
4412 #endif
4413 	cache = &zone->uz_cpu[curcpu];
4414 	itemdomain = 0;
4415 #ifdef NUMA
4416 	uz_flags = cache_uz_flags(cache);
4417 	if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4418 		itemdomain = item_domain(item);
4419 #endif
4420 	critical_enter();
4421 	do {
4422 		cache = &zone->uz_cpu[curcpu];
4423 		/* SMR Zones must free to the free bucket. */
4424 		bucket = &cache->uc_freebucket;
4425 #ifdef NUMA
4426 		if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4427 		    PCPU_GET(domain) != itemdomain) {
4428 			bucket = &cache->uc_crossbucket;
4429 		}
4430 #endif
4431 		if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4432 			cache_bucket_push(cache, bucket, item);
4433 			critical_exit();
4434 			return;
4435 		}
4436 	} while (cache_free(zone, cache, NULL, itemdomain));
4437 	critical_exit();
4438 
4439 	/*
4440 	 * If nothing else caught this, we'll just do an internal free.
4441 	 */
4442 	zone_free_item(zone, item, NULL, SKIP_NONE);
4443 }
4444 
4445 /* See uma.h */
4446 void
4447 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
4448 {
4449 	uma_cache_t cache;
4450 	uma_cache_bucket_t bucket;
4451 	int itemdomain, uz_flags;
4452 
4453 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
4454 	random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
4455 
4456 	CTR3(KTR_UMA, "uma_zfree_arg zone %s(%p) item %p",
4457 	    zone->uz_name, zone, item);
4458 
4459 #ifdef UMA_ZALLOC_DEBUG
4460 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
4461 	    ("uma_zfree_arg: called with SMR zone."));
4462 	if (uma_zfree_debug(zone, item, udata) == EJUSTRETURN)
4463 		return;
4464 #endif
4465         /* uma_zfree(..., NULL) does nothing, to match free(9). */
4466         if (item == NULL)
4467                 return;
4468 
4469 	/*
4470 	 * We are accessing the per-cpu cache without a critical section to
4471 	 * fetch size and flags.  This is acceptable, if we are preempted we
4472 	 * will simply read another cpu's line.
4473 	 */
4474 	cache = &zone->uz_cpu[curcpu];
4475 	uz_flags = cache_uz_flags(cache);
4476 	if (UMA_ALWAYS_CTORDTOR ||
4477 	    __predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0))
4478 		item_dtor(zone, item, cache_uz_size(cache), udata, SKIP_NONE);
4479 
4480 	/*
4481 	 * The race here is acceptable.  If we miss it we'll just have to wait
4482 	 * a little longer for the limits to be reset.
4483 	 */
4484 	if (__predict_false(uz_flags & UMA_ZFLAG_LIMIT)) {
4485 		if (atomic_load_32(&zone->uz_sleepers) > 0)
4486 			goto zfree_item;
4487 	}
4488 
4489 	/*
4490 	 * If possible, free to the per-CPU cache.  There are two
4491 	 * requirements for safe access to the per-CPU cache: (1) the thread
4492 	 * accessing the cache must not be preempted or yield during access,
4493 	 * and (2) the thread must not migrate CPUs without switching which
4494 	 * cache it accesses.  We rely on a critical section to prevent
4495 	 * preemption and migration.  We release the critical section in
4496 	 * order to acquire the zone mutex if we are unable to free to the
4497 	 * current cache; when we re-acquire the critical section, we must
4498 	 * detect and handle migration if it has occurred.
4499 	 */
4500 	itemdomain = 0;
4501 #ifdef NUMA
4502 	if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4503 		itemdomain = item_domain(item);
4504 #endif
4505 	critical_enter();
4506 	do {
4507 		cache = &zone->uz_cpu[curcpu];
4508 		/*
4509 		 * Try to free into the allocbucket first to give LIFO
4510 		 * ordering for cache-hot datastructures.  Spill over
4511 		 * into the freebucket if necessary.  Alloc will swap
4512 		 * them if one runs dry.
4513 		 */
4514 		bucket = &cache->uc_allocbucket;
4515 #ifdef NUMA
4516 		if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4517 		    PCPU_GET(domain) != itemdomain) {
4518 			bucket = &cache->uc_crossbucket;
4519 		} else
4520 #endif
4521 		if (bucket->ucb_cnt == bucket->ucb_entries &&
4522 		   cache->uc_freebucket.ucb_cnt <
4523 		   cache->uc_freebucket.ucb_entries)
4524 			cache_bucket_swap(&cache->uc_freebucket,
4525 			    &cache->uc_allocbucket);
4526 		if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4527 			cache_bucket_push(cache, bucket, item);
4528 			critical_exit();
4529 			return;
4530 		}
4531 	} while (cache_free(zone, cache, udata, itemdomain));
4532 	critical_exit();
4533 
4534 	/*
4535 	 * If nothing else caught this, we'll just do an internal free.
4536 	 */
4537 zfree_item:
4538 	zone_free_item(zone, item, udata, SKIP_DTOR);
4539 }
4540 
4541 #ifdef NUMA
4542 /*
4543  * sort crossdomain free buckets to domain correct buckets and cache
4544  * them.
4545  */
4546 static void
4547 zone_free_cross(uma_zone_t zone, uma_bucket_t bucket, void *udata)
4548 {
4549 	struct uma_bucketlist emptybuckets, fullbuckets;
4550 	uma_zone_domain_t zdom;
4551 	uma_bucket_t b;
4552 	smr_seq_t seq;
4553 	void *item;
4554 	int domain;
4555 
4556 	CTR3(KTR_UMA,
4557 	    "uma_zfree: zone %s(%p) draining cross bucket %p",
4558 	    zone->uz_name, zone, bucket);
4559 
4560 	/*
4561 	 * It is possible for buckets to arrive here out of order so we fetch
4562 	 * the current smr seq rather than accepting the bucket's.
4563 	 */
4564 	seq = SMR_SEQ_INVALID;
4565 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
4566 		seq = smr_advance(zone->uz_smr);
4567 
4568 	/*
4569 	 * To avoid having ndomain * ndomain buckets for sorting we have a
4570 	 * lock on the current crossfree bucket.  A full matrix with
4571 	 * per-domain locking could be used if necessary.
4572 	 */
4573 	STAILQ_INIT(&emptybuckets);
4574 	STAILQ_INIT(&fullbuckets);
4575 	ZONE_CROSS_LOCK(zone);
4576 	for (; bucket->ub_cnt > 0; bucket->ub_cnt--) {
4577 		item = bucket->ub_bucket[bucket->ub_cnt - 1];
4578 		domain = item_domain(item);
4579 		zdom = ZDOM_GET(zone, domain);
4580 		if (zdom->uzd_cross == NULL) {
4581 			if ((b = STAILQ_FIRST(&emptybuckets)) != NULL) {
4582 				STAILQ_REMOVE_HEAD(&emptybuckets, ub_link);
4583 				zdom->uzd_cross = b;
4584 			} else {
4585 				/*
4586 				 * Avoid allocating a bucket with the cross lock
4587 				 * held, since allocation can trigger a
4588 				 * cross-domain free and bucket zones may
4589 				 * allocate from each other.
4590 				 */
4591 				ZONE_CROSS_UNLOCK(zone);
4592 				b = bucket_alloc(zone, udata, M_NOWAIT);
4593 				if (b == NULL)
4594 					goto out;
4595 				ZONE_CROSS_LOCK(zone);
4596 				if (zdom->uzd_cross != NULL) {
4597 					STAILQ_INSERT_HEAD(&emptybuckets, b,
4598 					    ub_link);
4599 				} else {
4600 					zdom->uzd_cross = b;
4601 				}
4602 			}
4603 		}
4604 		b = zdom->uzd_cross;
4605 		b->ub_bucket[b->ub_cnt++] = item;
4606 		b->ub_seq = seq;
4607 		if (b->ub_cnt == b->ub_entries) {
4608 			STAILQ_INSERT_HEAD(&fullbuckets, b, ub_link);
4609 			if ((b = STAILQ_FIRST(&emptybuckets)) != NULL)
4610 				STAILQ_REMOVE_HEAD(&emptybuckets, ub_link);
4611 			zdom->uzd_cross = b;
4612 		}
4613 	}
4614 	ZONE_CROSS_UNLOCK(zone);
4615 out:
4616 	if (bucket->ub_cnt == 0)
4617 		bucket->ub_seq = SMR_SEQ_INVALID;
4618 	bucket_free(zone, bucket, udata);
4619 
4620 	while ((b = STAILQ_FIRST(&emptybuckets)) != NULL) {
4621 		STAILQ_REMOVE_HEAD(&emptybuckets, ub_link);
4622 		bucket_free(zone, b, udata);
4623 	}
4624 	while ((b = STAILQ_FIRST(&fullbuckets)) != NULL) {
4625 		STAILQ_REMOVE_HEAD(&fullbuckets, ub_link);
4626 		domain = item_domain(b->ub_bucket[0]);
4627 		zone_put_bucket(zone, domain, b, udata, true);
4628 	}
4629 }
4630 #endif
4631 
4632 static void
4633 zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata,
4634     int itemdomain, bool ws)
4635 {
4636 
4637 #ifdef NUMA
4638 	/*
4639 	 * Buckets coming from the wrong domain will be entirely for the
4640 	 * only other domain on two domain systems.  In this case we can
4641 	 * simply cache them.  Otherwise we need to sort them back to
4642 	 * correct domains.
4643 	 */
4644 	if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4645 	    vm_ndomains > 2 && PCPU_GET(domain) != itemdomain) {
4646 		zone_free_cross(zone, bucket, udata);
4647 		return;
4648 	}
4649 #endif
4650 
4651 	/*
4652 	 * Attempt to save the bucket in the zone's domain bucket cache.
4653 	 */
4654 	CTR3(KTR_UMA,
4655 	    "uma_zfree: zone %s(%p) putting bucket %p on free list",
4656 	    zone->uz_name, zone, bucket);
4657 	/* ub_cnt is pointing to the last free item */
4658 	if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
4659 		itemdomain = zone_domain_lowest(zone, itemdomain);
4660 	zone_put_bucket(zone, itemdomain, bucket, udata, ws);
4661 }
4662 
4663 /*
4664  * Populate a free or cross bucket for the current cpu cache.  Free any
4665  * existing full bucket either to the zone cache or back to the slab layer.
4666  *
4667  * Enters and returns in a critical section.  false return indicates that
4668  * we can not satisfy this free in the cache layer.  true indicates that
4669  * the caller should retry.
4670  */
4671 static __noinline bool
4672 cache_free(uma_zone_t zone, uma_cache_t cache, void *udata, int itemdomain)
4673 {
4674 	uma_cache_bucket_t cbucket;
4675 	uma_bucket_t newbucket, bucket;
4676 
4677 	CRITICAL_ASSERT(curthread);
4678 
4679 	if (zone->uz_bucket_size == 0)
4680 		return false;
4681 
4682 	cache = &zone->uz_cpu[curcpu];
4683 	newbucket = NULL;
4684 
4685 	/*
4686 	 * FIRSTTOUCH domains need to free to the correct zdom.  When
4687 	 * enabled this is the zdom of the item.   The bucket is the
4688 	 * cross bucket if the current domain and itemdomain do not match.
4689 	 */
4690 	cbucket = &cache->uc_freebucket;
4691 #ifdef NUMA
4692 	if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4693 		if (PCPU_GET(domain) != itemdomain) {
4694 			cbucket = &cache->uc_crossbucket;
4695 			if (cbucket->ucb_cnt != 0)
4696 				counter_u64_add(zone->uz_xdomain,
4697 				    cbucket->ucb_cnt);
4698 		}
4699 	}
4700 #endif
4701 	bucket = cache_bucket_unload(cbucket);
4702 	KASSERT(bucket == NULL || bucket->ub_cnt == bucket->ub_entries,
4703 	    ("cache_free: Entered with non-full free bucket."));
4704 
4705 	/* We are no longer associated with this CPU. */
4706 	critical_exit();
4707 
4708 	/*
4709 	 * Don't let SMR zones operate without a free bucket.  Force
4710 	 * a synchronize and re-use this one.  We will only degrade
4711 	 * to a synchronize every bucket_size items rather than every
4712 	 * item if we fail to allocate a bucket.
4713 	 */
4714 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0) {
4715 		if (bucket != NULL)
4716 			bucket->ub_seq = smr_advance(zone->uz_smr);
4717 		newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4718 		if (newbucket == NULL && bucket != NULL) {
4719 			bucket_drain(zone, bucket);
4720 			newbucket = bucket;
4721 			bucket = NULL;
4722 		}
4723 	} else if (!bucketdisable)
4724 		newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4725 
4726 	if (bucket != NULL)
4727 		zone_free_bucket(zone, bucket, udata, itemdomain, true);
4728 
4729 	critical_enter();
4730 	if ((bucket = newbucket) == NULL)
4731 		return (false);
4732 	cache = &zone->uz_cpu[curcpu];
4733 #ifdef NUMA
4734 	/*
4735 	 * Check to see if we should be populating the cross bucket.  If it
4736 	 * is already populated we will fall through and attempt to populate
4737 	 * the free bucket.
4738 	 */
4739 	if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4740 		if (PCPU_GET(domain) != itemdomain &&
4741 		    cache->uc_crossbucket.ucb_bucket == NULL) {
4742 			cache_bucket_load_cross(cache, bucket);
4743 			return (true);
4744 		}
4745 	}
4746 #endif
4747 	/*
4748 	 * We may have lost the race to fill the bucket or switched CPUs.
4749 	 */
4750 	if (cache->uc_freebucket.ucb_bucket != NULL) {
4751 		critical_exit();
4752 		bucket_free(zone, bucket, udata);
4753 		critical_enter();
4754 	} else
4755 		cache_bucket_load_free(cache, bucket);
4756 
4757 	return (true);
4758 }
4759 
4760 static void
4761 slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item)
4762 {
4763 	uma_keg_t keg;
4764 	uma_domain_t dom;
4765 	int freei;
4766 
4767 	keg = zone->uz_keg;
4768 	KEG_LOCK_ASSERT(keg, slab->us_domain);
4769 
4770 	/* Do we need to remove from any lists? */
4771 	dom = &keg->uk_domain[slab->us_domain];
4772 	if (slab->us_freecount + 1 == keg->uk_ipers) {
4773 		LIST_REMOVE(slab, us_link);
4774 		LIST_INSERT_HEAD(&dom->ud_free_slab, slab, us_link);
4775 		dom->ud_free_slabs++;
4776 	} else if (slab->us_freecount == 0) {
4777 		LIST_REMOVE(slab, us_link);
4778 		LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
4779 	}
4780 
4781 	/* Slab management. */
4782 	freei = slab_item_index(slab, keg, item);
4783 	BIT_SET(keg->uk_ipers, freei, &slab->us_free);
4784 	slab->us_freecount++;
4785 
4786 	/* Keg statistics. */
4787 	dom->ud_free_items++;
4788 }
4789 
4790 static void
4791 zone_release(void *arg, void **bucket, int cnt)
4792 {
4793 	struct mtx *lock;
4794 	uma_zone_t zone;
4795 	uma_slab_t slab;
4796 	uma_keg_t keg;
4797 	uint8_t *mem;
4798 	void *item;
4799 	int i;
4800 
4801 	zone = arg;
4802 	keg = zone->uz_keg;
4803 	lock = NULL;
4804 	if (__predict_false((zone->uz_flags & UMA_ZFLAG_HASH) != 0))
4805 		lock = KEG_LOCK(keg, 0);
4806 	for (i = 0; i < cnt; i++) {
4807 		item = bucket[i];
4808 		if (__predict_true((zone->uz_flags & UMA_ZFLAG_VTOSLAB) != 0)) {
4809 			slab = vtoslab((vm_offset_t)item);
4810 		} else {
4811 			mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
4812 			if ((zone->uz_flags & UMA_ZFLAG_HASH) != 0)
4813 				slab = hash_sfind(&keg->uk_hash, mem);
4814 			else
4815 				slab = (uma_slab_t)(mem + keg->uk_pgoff);
4816 		}
4817 		if (lock != KEG_LOCKPTR(keg, slab->us_domain)) {
4818 			if (lock != NULL)
4819 				mtx_unlock(lock);
4820 			lock = KEG_LOCK(keg, slab->us_domain);
4821 		}
4822 		slab_free_item(zone, slab, item);
4823 	}
4824 	if (lock != NULL)
4825 		mtx_unlock(lock);
4826 }
4827 
4828 /*
4829  * Frees a single item to any zone.
4830  *
4831  * Arguments:
4832  *	zone   The zone to free to
4833  *	item   The item we're freeing
4834  *	udata  User supplied data for the dtor
4835  *	skip   Skip dtors and finis
4836  */
4837 static __noinline void
4838 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
4839 {
4840 
4841 	/*
4842 	 * If a free is sent directly to an SMR zone we have to
4843 	 * synchronize immediately because the item can instantly
4844 	 * be reallocated. This should only happen in degenerate
4845 	 * cases when no memory is available for per-cpu caches.
4846 	 */
4847 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0 && skip == SKIP_NONE)
4848 		smr_synchronize(zone->uz_smr);
4849 
4850 	item_dtor(zone, item, zone->uz_size, udata, skip);
4851 
4852 	if (skip < SKIP_FINI && zone->uz_fini) {
4853 		kasan_mark_item_valid(zone, item);
4854 		zone->uz_fini(item, zone->uz_size);
4855 		kasan_mark_item_invalid(zone, item);
4856 	}
4857 
4858 	zone->uz_release(zone->uz_arg, &item, 1);
4859 
4860 	if (skip & SKIP_CNT)
4861 		return;
4862 
4863 	counter_u64_add(zone->uz_frees, 1);
4864 
4865 	if (zone->uz_max_items > 0)
4866 		zone_free_limit(zone, 1);
4867 }
4868 
4869 /* See uma.h */
4870 int
4871 uma_zone_set_max(uma_zone_t zone, int nitems)
4872 {
4873 
4874 	/*
4875 	 * If the limit is small, we may need to constrain the maximum per-CPU
4876 	 * cache size, or disable caching entirely.
4877 	 */
4878 	uma_zone_set_maxcache(zone, nitems);
4879 
4880 	/*
4881 	 * XXX This can misbehave if the zone has any allocations with
4882 	 * no limit and a limit is imposed.  There is currently no
4883 	 * way to clear a limit.
4884 	 */
4885 	ZONE_LOCK(zone);
4886 	if (zone->uz_max_items == 0)
4887 		ZONE_ASSERT_COLD(zone);
4888 	zone->uz_max_items = nitems;
4889 	zone->uz_flags |= UMA_ZFLAG_LIMIT;
4890 	zone_update_caches(zone);
4891 	/* We may need to wake waiters. */
4892 	wakeup(&zone->uz_max_items);
4893 	ZONE_UNLOCK(zone);
4894 
4895 	return (nitems);
4896 }
4897 
4898 /* See uma.h */
4899 void
4900 uma_zone_set_maxcache(uma_zone_t zone, int nitems)
4901 {
4902 	int bpcpu, bpdom, bsize, nb;
4903 
4904 	ZONE_LOCK(zone);
4905 
4906 	/*
4907 	 * Compute a lower bound on the number of items that may be cached in
4908 	 * the zone.  Each CPU gets at least two buckets, and for cross-domain
4909 	 * frees we use an additional bucket per CPU and per domain.  Select the
4910 	 * largest bucket size that does not exceed half of the requested limit,
4911 	 * with the left over space given to the full bucket cache.
4912 	 */
4913 	bpdom = 0;
4914 	bpcpu = 2;
4915 #ifdef NUMA
4916 	if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 && vm_ndomains > 1) {
4917 		bpcpu++;
4918 		bpdom++;
4919 	}
4920 #endif
4921 	nb = bpcpu * mp_ncpus + bpdom * vm_ndomains;
4922 	bsize = nitems / nb / 2;
4923 	if (bsize > BUCKET_MAX)
4924 		bsize = BUCKET_MAX;
4925 	else if (bsize == 0 && nitems / nb > 0)
4926 		bsize = 1;
4927 	zone->uz_bucket_size_max = zone->uz_bucket_size = bsize;
4928 	if (zone->uz_bucket_size_min > zone->uz_bucket_size_max)
4929 		zone->uz_bucket_size_min = zone->uz_bucket_size_max;
4930 	zone->uz_bucket_max = nitems - nb * bsize;
4931 	ZONE_UNLOCK(zone);
4932 }
4933 
4934 /* See uma.h */
4935 int
4936 uma_zone_get_max(uma_zone_t zone)
4937 {
4938 	int nitems;
4939 
4940 	nitems = atomic_load_64(&zone->uz_max_items);
4941 
4942 	return (nitems);
4943 }
4944 
4945 /* See uma.h */
4946 void
4947 uma_zone_set_warning(uma_zone_t zone, const char *warning)
4948 {
4949 
4950 	ZONE_ASSERT_COLD(zone);
4951 	zone->uz_warning = warning;
4952 }
4953 
4954 /* See uma.h */
4955 void
4956 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction)
4957 {
4958 
4959 	ZONE_ASSERT_COLD(zone);
4960 	TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone);
4961 }
4962 
4963 /* See uma.h */
4964 int
4965 uma_zone_get_cur(uma_zone_t zone)
4966 {
4967 	int64_t nitems;
4968 	u_int i;
4969 
4970 	nitems = 0;
4971 	if (zone->uz_allocs != EARLY_COUNTER && zone->uz_frees != EARLY_COUNTER)
4972 		nitems = counter_u64_fetch(zone->uz_allocs) -
4973 		    counter_u64_fetch(zone->uz_frees);
4974 	CPU_FOREACH(i)
4975 		nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs) -
4976 		    atomic_load_64(&zone->uz_cpu[i].uc_frees);
4977 
4978 	return (nitems < 0 ? 0 : nitems);
4979 }
4980 
4981 static uint64_t
4982 uma_zone_get_allocs(uma_zone_t zone)
4983 {
4984 	uint64_t nitems;
4985 	u_int i;
4986 
4987 	nitems = 0;
4988 	if (zone->uz_allocs != EARLY_COUNTER)
4989 		nitems = counter_u64_fetch(zone->uz_allocs);
4990 	CPU_FOREACH(i)
4991 		nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs);
4992 
4993 	return (nitems);
4994 }
4995 
4996 static uint64_t
4997 uma_zone_get_frees(uma_zone_t zone)
4998 {
4999 	uint64_t nitems;
5000 	u_int i;
5001 
5002 	nitems = 0;
5003 	if (zone->uz_frees != EARLY_COUNTER)
5004 		nitems = counter_u64_fetch(zone->uz_frees);
5005 	CPU_FOREACH(i)
5006 		nitems += atomic_load_64(&zone->uz_cpu[i].uc_frees);
5007 
5008 	return (nitems);
5009 }
5010 
5011 #ifdef INVARIANTS
5012 /* Used only for KEG_ASSERT_COLD(). */
5013 static uint64_t
5014 uma_keg_get_allocs(uma_keg_t keg)
5015 {
5016 	uma_zone_t z;
5017 	uint64_t nitems;
5018 
5019 	nitems = 0;
5020 	LIST_FOREACH(z, &keg->uk_zones, uz_link)
5021 		nitems += uma_zone_get_allocs(z);
5022 
5023 	return (nitems);
5024 }
5025 #endif
5026 
5027 /* See uma.h */
5028 void
5029 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
5030 {
5031 	uma_keg_t keg;
5032 
5033 	KEG_GET(zone, keg);
5034 	KEG_ASSERT_COLD(keg);
5035 	keg->uk_init = uminit;
5036 }
5037 
5038 /* See uma.h */
5039 void
5040 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
5041 {
5042 	uma_keg_t keg;
5043 
5044 	KEG_GET(zone, keg);
5045 	KEG_ASSERT_COLD(keg);
5046 	keg->uk_fini = fini;
5047 }
5048 
5049 /* See uma.h */
5050 void
5051 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
5052 {
5053 
5054 	ZONE_ASSERT_COLD(zone);
5055 	zone->uz_init = zinit;
5056 }
5057 
5058 /* See uma.h */
5059 void
5060 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
5061 {
5062 
5063 	ZONE_ASSERT_COLD(zone);
5064 	zone->uz_fini = zfini;
5065 }
5066 
5067 /* See uma.h */
5068 void
5069 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
5070 {
5071 	uma_keg_t keg;
5072 
5073 	KEG_GET(zone, keg);
5074 	KEG_ASSERT_COLD(keg);
5075 	keg->uk_freef = freef;
5076 }
5077 
5078 /* See uma.h */
5079 void
5080 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
5081 {
5082 	uma_keg_t keg;
5083 
5084 	KEG_GET(zone, keg);
5085 	KEG_ASSERT_COLD(keg);
5086 	keg->uk_allocf = allocf;
5087 }
5088 
5089 /* See uma.h */
5090 void
5091 uma_zone_set_smr(uma_zone_t zone, smr_t smr)
5092 {
5093 
5094 	ZONE_ASSERT_COLD(zone);
5095 
5096 	KASSERT(smr != NULL, ("Got NULL smr"));
5097 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
5098 	    ("zone %p (%s) already uses SMR", zone, zone->uz_name));
5099 	zone->uz_flags |= UMA_ZONE_SMR;
5100 	zone->uz_smr = smr;
5101 	zone_update_caches(zone);
5102 }
5103 
5104 smr_t
5105 uma_zone_get_smr(uma_zone_t zone)
5106 {
5107 
5108 	return (zone->uz_smr);
5109 }
5110 
5111 /* See uma.h */
5112 void
5113 uma_zone_reserve(uma_zone_t zone, int items)
5114 {
5115 	uma_keg_t keg;
5116 
5117 	KEG_GET(zone, keg);
5118 	KEG_ASSERT_COLD(keg);
5119 	keg->uk_reserve = items;
5120 }
5121 
5122 /* See uma.h */
5123 int
5124 uma_zone_reserve_kva(uma_zone_t zone, int count)
5125 {
5126 	uma_keg_t keg;
5127 	vm_offset_t kva;
5128 	u_int pages;
5129 
5130 	KEG_GET(zone, keg);
5131 	KEG_ASSERT_COLD(keg);
5132 	ZONE_ASSERT_COLD(zone);
5133 
5134 	pages = howmany(count, keg->uk_ipers) * keg->uk_ppera;
5135 
5136 #ifdef UMA_MD_SMALL_ALLOC
5137 	if (keg->uk_ppera > 1) {
5138 #else
5139 	if (1) {
5140 #endif
5141 		kva = kva_alloc((vm_size_t)pages * PAGE_SIZE);
5142 		if (kva == 0)
5143 			return (0);
5144 	} else
5145 		kva = 0;
5146 
5147 	MPASS(keg->uk_kva == 0);
5148 	keg->uk_kva = kva;
5149 	keg->uk_offset = 0;
5150 	zone->uz_max_items = pages * keg->uk_ipers;
5151 #ifdef UMA_MD_SMALL_ALLOC
5152 	keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
5153 #else
5154 	keg->uk_allocf = noobj_alloc;
5155 #endif
5156 	keg->uk_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
5157 	zone->uz_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
5158 	zone_update_caches(zone);
5159 
5160 	return (1);
5161 }
5162 
5163 /* See uma.h */
5164 void
5165 uma_prealloc(uma_zone_t zone, int items)
5166 {
5167 	struct vm_domainset_iter di;
5168 	uma_domain_t dom;
5169 	uma_slab_t slab;
5170 	uma_keg_t keg;
5171 	int aflags, domain, slabs;
5172 
5173 	KEG_GET(zone, keg);
5174 	slabs = howmany(items, keg->uk_ipers);
5175 	while (slabs-- > 0) {
5176 		aflags = M_NOWAIT;
5177 		vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
5178 		    &aflags);
5179 		for (;;) {
5180 			slab = keg_alloc_slab(keg, zone, domain, M_WAITOK,
5181 			    aflags);
5182 			if (slab != NULL) {
5183 				dom = &keg->uk_domain[slab->us_domain];
5184 				/*
5185 				 * keg_alloc_slab() always returns a slab on the
5186 				 * partial list.
5187 				 */
5188 				LIST_REMOVE(slab, us_link);
5189 				LIST_INSERT_HEAD(&dom->ud_free_slab, slab,
5190 				    us_link);
5191 				dom->ud_free_slabs++;
5192 				KEG_UNLOCK(keg, slab->us_domain);
5193 				break;
5194 			}
5195 			if (vm_domainset_iter_policy(&di, &domain) != 0)
5196 				vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask, 0);
5197 		}
5198 	}
5199 }
5200 
5201 /*
5202  * Returns a snapshot of memory consumption in bytes.
5203  */
5204 size_t
5205 uma_zone_memory(uma_zone_t zone)
5206 {
5207 	size_t sz;
5208 	int i;
5209 
5210 	sz = 0;
5211 	if (zone->uz_flags & UMA_ZFLAG_CACHE) {
5212 		for (i = 0; i < vm_ndomains; i++)
5213 			sz += ZDOM_GET(zone, i)->uzd_nitems;
5214 		return (sz * zone->uz_size);
5215 	}
5216 	for (i = 0; i < vm_ndomains; i++)
5217 		sz += zone->uz_keg->uk_domain[i].ud_pages;
5218 
5219 	return (sz * PAGE_SIZE);
5220 }
5221 
5222 struct uma_reclaim_args {
5223 	int	domain;
5224 	int	req;
5225 };
5226 
5227 static void
5228 uma_reclaim_domain_cb(uma_zone_t zone, void *arg)
5229 {
5230 	struct uma_reclaim_args *args;
5231 
5232 	args = arg;
5233 	if ((zone->uz_flags & UMA_ZONE_UNMANAGED) == 0)
5234 		uma_zone_reclaim_domain(zone, args->req, args->domain);
5235 }
5236 
5237 /* See uma.h */
5238 void
5239 uma_reclaim(int req)
5240 {
5241 	uma_reclaim_domain(req, UMA_ANYDOMAIN);
5242 }
5243 
5244 void
5245 uma_reclaim_domain(int req, int domain)
5246 {
5247 	struct uma_reclaim_args args;
5248 
5249 	bucket_enable();
5250 
5251 	args.domain = domain;
5252 	args.req = req;
5253 
5254 	sx_slock(&uma_reclaim_lock);
5255 	switch (req) {
5256 	case UMA_RECLAIM_TRIM:
5257 	case UMA_RECLAIM_DRAIN:
5258 		zone_foreach(uma_reclaim_domain_cb, &args);
5259 		break;
5260 	case UMA_RECLAIM_DRAIN_CPU:
5261 		zone_foreach(uma_reclaim_domain_cb, &args);
5262 		pcpu_cache_drain_safe(NULL);
5263 		zone_foreach(uma_reclaim_domain_cb, &args);
5264 		break;
5265 	default:
5266 		panic("unhandled reclamation request %d", req);
5267 	}
5268 
5269 	/*
5270 	 * Some slabs may have been freed but this zone will be visited early
5271 	 * we visit again so that we can free pages that are empty once other
5272 	 * zones are drained.  We have to do the same for buckets.
5273 	 */
5274 	uma_zone_reclaim_domain(slabzones[0], UMA_RECLAIM_DRAIN, domain);
5275 	uma_zone_reclaim_domain(slabzones[1], UMA_RECLAIM_DRAIN, domain);
5276 	bucket_zone_drain(domain);
5277 	sx_sunlock(&uma_reclaim_lock);
5278 }
5279 
5280 static volatile int uma_reclaim_needed;
5281 
5282 void
5283 uma_reclaim_wakeup(void)
5284 {
5285 
5286 	if (atomic_fetchadd_int(&uma_reclaim_needed, 1) == 0)
5287 		wakeup(uma_reclaim);
5288 }
5289 
5290 void
5291 uma_reclaim_worker(void *arg __unused)
5292 {
5293 
5294 	for (;;) {
5295 		sx_xlock(&uma_reclaim_lock);
5296 		while (atomic_load_int(&uma_reclaim_needed) == 0)
5297 			sx_sleep(uma_reclaim, &uma_reclaim_lock, PVM, "umarcl",
5298 			    hz);
5299 		sx_xunlock(&uma_reclaim_lock);
5300 		EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM);
5301 		uma_reclaim(UMA_RECLAIM_DRAIN_CPU);
5302 		atomic_store_int(&uma_reclaim_needed, 0);
5303 		/* Don't fire more than once per-second. */
5304 		pause("umarclslp", hz);
5305 	}
5306 }
5307 
5308 /* See uma.h */
5309 void
5310 uma_zone_reclaim(uma_zone_t zone, int req)
5311 {
5312 	uma_zone_reclaim_domain(zone, req, UMA_ANYDOMAIN);
5313 }
5314 
5315 void
5316 uma_zone_reclaim_domain(uma_zone_t zone, int req, int domain)
5317 {
5318 	switch (req) {
5319 	case UMA_RECLAIM_TRIM:
5320 		zone_reclaim(zone, domain, M_NOWAIT, false);
5321 		break;
5322 	case UMA_RECLAIM_DRAIN:
5323 		zone_reclaim(zone, domain, M_NOWAIT, true);
5324 		break;
5325 	case UMA_RECLAIM_DRAIN_CPU:
5326 		pcpu_cache_drain_safe(zone);
5327 		zone_reclaim(zone, domain, M_NOWAIT, true);
5328 		break;
5329 	default:
5330 		panic("unhandled reclamation request %d", req);
5331 	}
5332 }
5333 
5334 /* See uma.h */
5335 int
5336 uma_zone_exhausted(uma_zone_t zone)
5337 {
5338 
5339 	return (atomic_load_32(&zone->uz_sleepers) > 0);
5340 }
5341 
5342 unsigned long
5343 uma_limit(void)
5344 {
5345 
5346 	return (uma_kmem_limit);
5347 }
5348 
5349 void
5350 uma_set_limit(unsigned long limit)
5351 {
5352 
5353 	uma_kmem_limit = limit;
5354 }
5355 
5356 unsigned long
5357 uma_size(void)
5358 {
5359 
5360 	return (atomic_load_long(&uma_kmem_total));
5361 }
5362 
5363 long
5364 uma_avail(void)
5365 {
5366 
5367 	return (uma_kmem_limit - uma_size());
5368 }
5369 
5370 #ifdef DDB
5371 /*
5372  * Generate statistics across both the zone and its per-cpu cache's.  Return
5373  * desired statistics if the pointer is non-NULL for that statistic.
5374  *
5375  * Note: does not update the zone statistics, as it can't safely clear the
5376  * per-CPU cache statistic.
5377  *
5378  */
5379 static void
5380 uma_zone_sumstat(uma_zone_t z, long *cachefreep, uint64_t *allocsp,
5381     uint64_t *freesp, uint64_t *sleepsp, uint64_t *xdomainp)
5382 {
5383 	uma_cache_t cache;
5384 	uint64_t allocs, frees, sleeps, xdomain;
5385 	int cachefree, cpu;
5386 
5387 	allocs = frees = sleeps = xdomain = 0;
5388 	cachefree = 0;
5389 	CPU_FOREACH(cpu) {
5390 		cache = &z->uz_cpu[cpu];
5391 		cachefree += cache->uc_allocbucket.ucb_cnt;
5392 		cachefree += cache->uc_freebucket.ucb_cnt;
5393 		xdomain += cache->uc_crossbucket.ucb_cnt;
5394 		cachefree += cache->uc_crossbucket.ucb_cnt;
5395 		allocs += cache->uc_allocs;
5396 		frees += cache->uc_frees;
5397 	}
5398 	allocs += counter_u64_fetch(z->uz_allocs);
5399 	frees += counter_u64_fetch(z->uz_frees);
5400 	xdomain += counter_u64_fetch(z->uz_xdomain);
5401 	sleeps += z->uz_sleeps;
5402 	if (cachefreep != NULL)
5403 		*cachefreep = cachefree;
5404 	if (allocsp != NULL)
5405 		*allocsp = allocs;
5406 	if (freesp != NULL)
5407 		*freesp = frees;
5408 	if (sleepsp != NULL)
5409 		*sleepsp = sleeps;
5410 	if (xdomainp != NULL)
5411 		*xdomainp = xdomain;
5412 }
5413 #endif /* DDB */
5414 
5415 static int
5416 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
5417 {
5418 	uma_keg_t kz;
5419 	uma_zone_t z;
5420 	int count;
5421 
5422 	count = 0;
5423 	rw_rlock(&uma_rwlock);
5424 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
5425 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
5426 			count++;
5427 	}
5428 	LIST_FOREACH(z, &uma_cachezones, uz_link)
5429 		count++;
5430 
5431 	rw_runlock(&uma_rwlock);
5432 	return (sysctl_handle_int(oidp, &count, 0, req));
5433 }
5434 
5435 static void
5436 uma_vm_zone_stats(struct uma_type_header *uth, uma_zone_t z, struct sbuf *sbuf,
5437     struct uma_percpu_stat *ups, bool internal)
5438 {
5439 	uma_zone_domain_t zdom;
5440 	uma_cache_t cache;
5441 	int i;
5442 
5443 	for (i = 0; i < vm_ndomains; i++) {
5444 		zdom = ZDOM_GET(z, i);
5445 		uth->uth_zone_free += zdom->uzd_nitems;
5446 	}
5447 	uth->uth_allocs = counter_u64_fetch(z->uz_allocs);
5448 	uth->uth_frees = counter_u64_fetch(z->uz_frees);
5449 	uth->uth_fails = counter_u64_fetch(z->uz_fails);
5450 	uth->uth_xdomain = counter_u64_fetch(z->uz_xdomain);
5451 	uth->uth_sleeps = z->uz_sleeps;
5452 
5453 	for (i = 0; i < mp_maxid + 1; i++) {
5454 		bzero(&ups[i], sizeof(*ups));
5455 		if (internal || CPU_ABSENT(i))
5456 			continue;
5457 		cache = &z->uz_cpu[i];
5458 		ups[i].ups_cache_free += cache->uc_allocbucket.ucb_cnt;
5459 		ups[i].ups_cache_free += cache->uc_freebucket.ucb_cnt;
5460 		ups[i].ups_cache_free += cache->uc_crossbucket.ucb_cnt;
5461 		ups[i].ups_allocs = cache->uc_allocs;
5462 		ups[i].ups_frees = cache->uc_frees;
5463 	}
5464 }
5465 
5466 static int
5467 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
5468 {
5469 	struct uma_stream_header ush;
5470 	struct uma_type_header uth;
5471 	struct uma_percpu_stat *ups;
5472 	struct sbuf sbuf;
5473 	uma_keg_t kz;
5474 	uma_zone_t z;
5475 	uint64_t items;
5476 	uint32_t kfree, pages;
5477 	int count, error, i;
5478 
5479 	error = sysctl_wire_old_buffer(req, 0);
5480 	if (error != 0)
5481 		return (error);
5482 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
5483 	sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL);
5484 	ups = malloc((mp_maxid + 1) * sizeof(*ups), M_TEMP, M_WAITOK);
5485 
5486 	count = 0;
5487 	rw_rlock(&uma_rwlock);
5488 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
5489 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
5490 			count++;
5491 	}
5492 
5493 	LIST_FOREACH(z, &uma_cachezones, uz_link)
5494 		count++;
5495 
5496 	/*
5497 	 * Insert stream header.
5498 	 */
5499 	bzero(&ush, sizeof(ush));
5500 	ush.ush_version = UMA_STREAM_VERSION;
5501 	ush.ush_maxcpus = (mp_maxid + 1);
5502 	ush.ush_count = count;
5503 	(void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
5504 
5505 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
5506 		kfree = pages = 0;
5507 		for (i = 0; i < vm_ndomains; i++) {
5508 			kfree += kz->uk_domain[i].ud_free_items;
5509 			pages += kz->uk_domain[i].ud_pages;
5510 		}
5511 		LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5512 			bzero(&uth, sizeof(uth));
5513 			strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5514 			uth.uth_align = kz->uk_align;
5515 			uth.uth_size = kz->uk_size;
5516 			uth.uth_rsize = kz->uk_rsize;
5517 			if (z->uz_max_items > 0) {
5518 				items = UZ_ITEMS_COUNT(z->uz_items);
5519 				uth.uth_pages = (items / kz->uk_ipers) *
5520 					kz->uk_ppera;
5521 			} else
5522 				uth.uth_pages = pages;
5523 			uth.uth_maxpages = (z->uz_max_items / kz->uk_ipers) *
5524 			    kz->uk_ppera;
5525 			uth.uth_limit = z->uz_max_items;
5526 			uth.uth_keg_free = kfree;
5527 
5528 			/*
5529 			 * A zone is secondary is it is not the first entry
5530 			 * on the keg's zone list.
5531 			 */
5532 			if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
5533 			    (LIST_FIRST(&kz->uk_zones) != z))
5534 				uth.uth_zone_flags = UTH_ZONE_SECONDARY;
5535 			uma_vm_zone_stats(&uth, z, &sbuf, ups,
5536 			    kz->uk_flags & UMA_ZFLAG_INTERNAL);
5537 			(void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5538 			for (i = 0; i < mp_maxid + 1; i++)
5539 				(void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5540 		}
5541 	}
5542 	LIST_FOREACH(z, &uma_cachezones, uz_link) {
5543 		bzero(&uth, sizeof(uth));
5544 		strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5545 		uth.uth_size = z->uz_size;
5546 		uma_vm_zone_stats(&uth, z, &sbuf, ups, false);
5547 		(void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5548 		for (i = 0; i < mp_maxid + 1; i++)
5549 			(void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5550 	}
5551 
5552 	rw_runlock(&uma_rwlock);
5553 	error = sbuf_finish(&sbuf);
5554 	sbuf_delete(&sbuf);
5555 	free(ups, M_TEMP);
5556 	return (error);
5557 }
5558 
5559 int
5560 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS)
5561 {
5562 	uma_zone_t zone = *(uma_zone_t *)arg1;
5563 	int error, max;
5564 
5565 	max = uma_zone_get_max(zone);
5566 	error = sysctl_handle_int(oidp, &max, 0, req);
5567 	if (error || !req->newptr)
5568 		return (error);
5569 
5570 	uma_zone_set_max(zone, max);
5571 
5572 	return (0);
5573 }
5574 
5575 int
5576 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS)
5577 {
5578 	uma_zone_t zone;
5579 	int cur;
5580 
5581 	/*
5582 	 * Some callers want to add sysctls for global zones that
5583 	 * may not yet exist so they pass a pointer to a pointer.
5584 	 */
5585 	if (arg2 == 0)
5586 		zone = *(uma_zone_t *)arg1;
5587 	else
5588 		zone = arg1;
5589 	cur = uma_zone_get_cur(zone);
5590 	return (sysctl_handle_int(oidp, &cur, 0, req));
5591 }
5592 
5593 static int
5594 sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS)
5595 {
5596 	uma_zone_t zone = arg1;
5597 	uint64_t cur;
5598 
5599 	cur = uma_zone_get_allocs(zone);
5600 	return (sysctl_handle_64(oidp, &cur, 0, req));
5601 }
5602 
5603 static int
5604 sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS)
5605 {
5606 	uma_zone_t zone = arg1;
5607 	uint64_t cur;
5608 
5609 	cur = uma_zone_get_frees(zone);
5610 	return (sysctl_handle_64(oidp, &cur, 0, req));
5611 }
5612 
5613 static int
5614 sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS)
5615 {
5616 	struct sbuf sbuf;
5617 	uma_zone_t zone = arg1;
5618 	int error;
5619 
5620 	sbuf_new_for_sysctl(&sbuf, NULL, 0, req);
5621 	if (zone->uz_flags != 0)
5622 		sbuf_printf(&sbuf, "0x%b", zone->uz_flags, PRINT_UMA_ZFLAGS);
5623 	else
5624 		sbuf_printf(&sbuf, "0");
5625 	error = sbuf_finish(&sbuf);
5626 	sbuf_delete(&sbuf);
5627 
5628 	return (error);
5629 }
5630 
5631 static int
5632 sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS)
5633 {
5634 	uma_keg_t keg = arg1;
5635 	int avail, effpct, total;
5636 
5637 	total = keg->uk_ppera * PAGE_SIZE;
5638 	if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0)
5639 		total += slabzone(keg->uk_ipers)->uz_keg->uk_rsize;
5640 	/*
5641 	 * We consider the client's requested size and alignment here, not the
5642 	 * real size determination uk_rsize, because we also adjust the real
5643 	 * size for internal implementation reasons (max bitset size).
5644 	 */
5645 	avail = keg->uk_ipers * roundup2(keg->uk_size, keg->uk_align + 1);
5646 	if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
5647 		avail *= mp_maxid + 1;
5648 	effpct = 100 * avail / total;
5649 	return (sysctl_handle_int(oidp, &effpct, 0, req));
5650 }
5651 
5652 static int
5653 sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS)
5654 {
5655 	uma_zone_t zone = arg1;
5656 	uint64_t cur;
5657 
5658 	cur = UZ_ITEMS_COUNT(atomic_load_64(&zone->uz_items));
5659 	return (sysctl_handle_64(oidp, &cur, 0, req));
5660 }
5661 
5662 #ifdef INVARIANTS
5663 static uma_slab_t
5664 uma_dbg_getslab(uma_zone_t zone, void *item)
5665 {
5666 	uma_slab_t slab;
5667 	uma_keg_t keg;
5668 	uint8_t *mem;
5669 
5670 	/*
5671 	 * It is safe to return the slab here even though the
5672 	 * zone is unlocked because the item's allocation state
5673 	 * essentially holds a reference.
5674 	 */
5675 	mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
5676 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5677 		return (NULL);
5678 	if (zone->uz_flags & UMA_ZFLAG_VTOSLAB)
5679 		return (vtoslab((vm_offset_t)mem));
5680 	keg = zone->uz_keg;
5681 	if ((keg->uk_flags & UMA_ZFLAG_HASH) == 0)
5682 		return ((uma_slab_t)(mem + keg->uk_pgoff));
5683 	KEG_LOCK(keg, 0);
5684 	slab = hash_sfind(&keg->uk_hash, mem);
5685 	KEG_UNLOCK(keg, 0);
5686 
5687 	return (slab);
5688 }
5689 
5690 static bool
5691 uma_dbg_zskip(uma_zone_t zone, void *mem)
5692 {
5693 
5694 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5695 		return (true);
5696 
5697 	return (uma_dbg_kskip(zone->uz_keg, mem));
5698 }
5699 
5700 static bool
5701 uma_dbg_kskip(uma_keg_t keg, void *mem)
5702 {
5703 	uintptr_t idx;
5704 
5705 	if (dbg_divisor == 0)
5706 		return (true);
5707 
5708 	if (dbg_divisor == 1)
5709 		return (false);
5710 
5711 	idx = (uintptr_t)mem >> PAGE_SHIFT;
5712 	if (keg->uk_ipers > 1) {
5713 		idx *= keg->uk_ipers;
5714 		idx += ((uintptr_t)mem & PAGE_MASK) / keg->uk_rsize;
5715 	}
5716 
5717 	if ((idx / dbg_divisor) * dbg_divisor != idx) {
5718 		counter_u64_add(uma_skip_cnt, 1);
5719 		return (true);
5720 	}
5721 	counter_u64_add(uma_dbg_cnt, 1);
5722 
5723 	return (false);
5724 }
5725 
5726 /*
5727  * Set up the slab's freei data such that uma_dbg_free can function.
5728  *
5729  */
5730 static void
5731 uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item)
5732 {
5733 	uma_keg_t keg;
5734 	int freei;
5735 
5736 	if (slab == NULL) {
5737 		slab = uma_dbg_getslab(zone, item);
5738 		if (slab == NULL)
5739 			panic("uma: item %p did not belong to zone %s",
5740 			    item, zone->uz_name);
5741 	}
5742 	keg = zone->uz_keg;
5743 	freei = slab_item_index(slab, keg, item);
5744 
5745 	if (BIT_TEST_SET_ATOMIC(keg->uk_ipers, freei,
5746 	    slab_dbg_bits(slab, keg)))
5747 		panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)",
5748 		    item, zone, zone->uz_name, slab, freei);
5749 }
5750 
5751 /*
5752  * Verifies freed addresses.  Checks for alignment, valid slab membership
5753  * and duplicate frees.
5754  *
5755  */
5756 static void
5757 uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item)
5758 {
5759 	uma_keg_t keg;
5760 	int freei;
5761 
5762 	if (slab == NULL) {
5763 		slab = uma_dbg_getslab(zone, item);
5764 		if (slab == NULL)
5765 			panic("uma: Freed item %p did not belong to zone %s",
5766 			    item, zone->uz_name);
5767 	}
5768 	keg = zone->uz_keg;
5769 	freei = slab_item_index(slab, keg, item);
5770 
5771 	if (freei >= keg->uk_ipers)
5772 		panic("Invalid free of %p from zone %p(%s) slab %p(%d)",
5773 		    item, zone, zone->uz_name, slab, freei);
5774 
5775 	if (slab_item(slab, keg, freei) != item)
5776 		panic("Unaligned free of %p from zone %p(%s) slab %p(%d)",
5777 		    item, zone, zone->uz_name, slab, freei);
5778 
5779 	if (!BIT_TEST_CLR_ATOMIC(keg->uk_ipers, freei,
5780 	    slab_dbg_bits(slab, keg)))
5781 		panic("Duplicate free of %p from zone %p(%s) slab %p(%d)",
5782 		    item, zone, zone->uz_name, slab, freei);
5783 }
5784 #endif /* INVARIANTS */
5785 
5786 #ifdef DDB
5787 static int64_t
5788 get_uma_stats(uma_keg_t kz, uma_zone_t z, uint64_t *allocs, uint64_t *used,
5789     uint64_t *sleeps, long *cachefree, uint64_t *xdomain)
5790 {
5791 	uint64_t frees;
5792 	int i;
5793 
5794 	if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
5795 		*allocs = counter_u64_fetch(z->uz_allocs);
5796 		frees = counter_u64_fetch(z->uz_frees);
5797 		*sleeps = z->uz_sleeps;
5798 		*cachefree = 0;
5799 		*xdomain = 0;
5800 	} else
5801 		uma_zone_sumstat(z, cachefree, allocs, &frees, sleeps,
5802 		    xdomain);
5803 	for (i = 0; i < vm_ndomains; i++) {
5804 		*cachefree += ZDOM_GET(z, i)->uzd_nitems;
5805 		if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
5806 		    (LIST_FIRST(&kz->uk_zones) != z)))
5807 			*cachefree += kz->uk_domain[i].ud_free_items;
5808 	}
5809 	*used = *allocs - frees;
5810 	return (((int64_t)*used + *cachefree) * kz->uk_size);
5811 }
5812 
5813 DB_SHOW_COMMAND(uma, db_show_uma)
5814 {
5815 	const char *fmt_hdr, *fmt_entry;
5816 	uma_keg_t kz;
5817 	uma_zone_t z;
5818 	uint64_t allocs, used, sleeps, xdomain;
5819 	long cachefree;
5820 	/* variables for sorting */
5821 	uma_keg_t cur_keg;
5822 	uma_zone_t cur_zone, last_zone;
5823 	int64_t cur_size, last_size, size;
5824 	int ties;
5825 
5826 	/* /i option produces machine-parseable CSV output */
5827 	if (modif[0] == 'i') {
5828 		fmt_hdr = "%s,%s,%s,%s,%s,%s,%s,%s,%s\n";
5829 		fmt_entry = "\"%s\",%ju,%jd,%ld,%ju,%ju,%u,%jd,%ju\n";
5830 	} else {
5831 		fmt_hdr = "%18s %6s %7s %7s %11s %7s %7s %10s %8s\n";
5832 		fmt_entry = "%18s %6ju %7jd %7ld %11ju %7ju %7u %10jd %8ju\n";
5833 	}
5834 
5835 	db_printf(fmt_hdr, "Zone", "Size", "Used", "Free", "Requests",
5836 	    "Sleeps", "Bucket", "Total Mem", "XFree");
5837 
5838 	/* Sort the zones with largest size first. */
5839 	last_zone = NULL;
5840 	last_size = INT64_MAX;
5841 	for (;;) {
5842 		cur_zone = NULL;
5843 		cur_size = -1;
5844 		ties = 0;
5845 		LIST_FOREACH(kz, &uma_kegs, uk_link) {
5846 			LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5847 				/*
5848 				 * In the case of size ties, print out zones
5849 				 * in the order they are encountered.  That is,
5850 				 * when we encounter the most recently output
5851 				 * zone, we have already printed all preceding
5852 				 * ties, and we must print all following ties.
5853 				 */
5854 				if (z == last_zone) {
5855 					ties = 1;
5856 					continue;
5857 				}
5858 				size = get_uma_stats(kz, z, &allocs, &used,
5859 				    &sleeps, &cachefree, &xdomain);
5860 				if (size > cur_size && size < last_size + ties)
5861 				{
5862 					cur_size = size;
5863 					cur_zone = z;
5864 					cur_keg = kz;
5865 				}
5866 			}
5867 		}
5868 		if (cur_zone == NULL)
5869 			break;
5870 
5871 		size = get_uma_stats(cur_keg, cur_zone, &allocs, &used,
5872 		    &sleeps, &cachefree, &xdomain);
5873 		db_printf(fmt_entry, cur_zone->uz_name,
5874 		    (uintmax_t)cur_keg->uk_size, (intmax_t)used, cachefree,
5875 		    (uintmax_t)allocs, (uintmax_t)sleeps,
5876 		    (unsigned)cur_zone->uz_bucket_size, (intmax_t)size,
5877 		    xdomain);
5878 
5879 		if (db_pager_quit)
5880 			return;
5881 		last_zone = cur_zone;
5882 		last_size = cur_size;
5883 	}
5884 }
5885 
5886 DB_SHOW_COMMAND(umacache, db_show_umacache)
5887 {
5888 	uma_zone_t z;
5889 	uint64_t allocs, frees;
5890 	long cachefree;
5891 	int i;
5892 
5893 	db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
5894 	    "Requests", "Bucket");
5895 	LIST_FOREACH(z, &uma_cachezones, uz_link) {
5896 		uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL, NULL);
5897 		for (i = 0; i < vm_ndomains; i++)
5898 			cachefree += ZDOM_GET(z, i)->uzd_nitems;
5899 		db_printf("%18s %8ju %8jd %8ld %12ju %8u\n",
5900 		    z->uz_name, (uintmax_t)z->uz_size,
5901 		    (intmax_t)(allocs - frees), cachefree,
5902 		    (uintmax_t)allocs, z->uz_bucket_size);
5903 		if (db_pager_quit)
5904 			return;
5905 	}
5906 }
5907 #endif	/* DDB */
5908