xref: /freebsd/sys/vm/uma_core.c (revision 7be9a3b45356747f9fcb6d69a722c1c95f8060bf)
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 	for (int i = 0; i < vm_ndomains; i++) {
1227 		if (bucket_cache_reclaim_domain(zone, false, false, i) &&
1228 		    (zone->uz_flags & UMA_ZFLAG_CACHE) == 0)
1229 			keg_drain(zone->uz_keg, i);
1230 	}
1231 }
1232 
1233 /*
1234  * Allocate and zero fill the next sized hash table from the appropriate
1235  * backing store.
1236  *
1237  * Arguments:
1238  *	hash  A new hash structure with the old hash size in uh_hashsize
1239  *
1240  * Returns:
1241  *	1 on success and 0 on failure.
1242  */
1243 static int
1244 hash_alloc(struct uma_hash *hash, u_int size)
1245 {
1246 	size_t alloc;
1247 
1248 	KASSERT(powerof2(size), ("hash size must be power of 2"));
1249 	if (size > UMA_HASH_SIZE_INIT)  {
1250 		hash->uh_hashsize = size;
1251 		alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
1252 		hash->uh_slab_hash = malloc(alloc, M_UMAHASH, M_NOWAIT);
1253 	} else {
1254 		alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
1255 		hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
1256 		    UMA_ANYDOMAIN, M_WAITOK);
1257 		hash->uh_hashsize = UMA_HASH_SIZE_INIT;
1258 	}
1259 	if (hash->uh_slab_hash) {
1260 		bzero(hash->uh_slab_hash, alloc);
1261 		hash->uh_hashmask = hash->uh_hashsize - 1;
1262 		return (1);
1263 	}
1264 
1265 	return (0);
1266 }
1267 
1268 /*
1269  * Expands the hash table for HASH zones.  This is done from zone_timeout
1270  * to reduce collisions.  This must not be done in the regular allocation
1271  * path, otherwise, we can recurse on the vm while allocating pages.
1272  *
1273  * Arguments:
1274  *	oldhash  The hash you want to expand
1275  *	newhash  The hash structure for the new table
1276  *
1277  * Returns:
1278  *	Nothing
1279  *
1280  * Discussion:
1281  */
1282 static int
1283 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
1284 {
1285 	uma_hash_slab_t slab;
1286 	u_int hval;
1287 	u_int idx;
1288 
1289 	if (!newhash->uh_slab_hash)
1290 		return (0);
1291 
1292 	if (oldhash->uh_hashsize >= newhash->uh_hashsize)
1293 		return (0);
1294 
1295 	/*
1296 	 * I need to investigate hash algorithms for resizing without a
1297 	 * full rehash.
1298 	 */
1299 
1300 	for (idx = 0; idx < oldhash->uh_hashsize; idx++)
1301 		while (!LIST_EMPTY(&oldhash->uh_slab_hash[idx])) {
1302 			slab = LIST_FIRST(&oldhash->uh_slab_hash[idx]);
1303 			LIST_REMOVE(slab, uhs_hlink);
1304 			hval = UMA_HASH(newhash, slab->uhs_data);
1305 			LIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
1306 			    slab, uhs_hlink);
1307 		}
1308 
1309 	return (1);
1310 }
1311 
1312 /*
1313  * Free the hash bucket to the appropriate backing store.
1314  *
1315  * Arguments:
1316  *	slab_hash  The hash bucket we're freeing
1317  *	hashsize   The number of entries in that hash bucket
1318  *
1319  * Returns:
1320  *	Nothing
1321  */
1322 static void
1323 hash_free(struct uma_hash *hash)
1324 {
1325 	if (hash->uh_slab_hash == NULL)
1326 		return;
1327 	if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
1328 		zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE);
1329 	else
1330 		free(hash->uh_slab_hash, M_UMAHASH);
1331 }
1332 
1333 /*
1334  * Frees all outstanding items in a bucket
1335  *
1336  * Arguments:
1337  *	zone   The zone to free to, must be unlocked.
1338  *	bucket The free/alloc bucket with items.
1339  *
1340  * Returns:
1341  *	Nothing
1342  */
1343 static void
1344 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
1345 {
1346 	int i;
1347 
1348 	if (bucket->ub_cnt == 0)
1349 		return;
1350 
1351 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
1352 	    bucket->ub_seq != SMR_SEQ_INVALID) {
1353 		smr_wait(zone->uz_smr, bucket->ub_seq);
1354 		bucket->ub_seq = SMR_SEQ_INVALID;
1355 		for (i = 0; i < bucket->ub_cnt; i++)
1356 			item_dtor(zone, bucket->ub_bucket[i],
1357 			    zone->uz_size, NULL, SKIP_NONE);
1358 	}
1359 	if (zone->uz_fini)
1360 		for (i = 0; i < bucket->ub_cnt; i++) {
1361 			kasan_mark_item_valid(zone, bucket->ub_bucket[i]);
1362 			zone->uz_fini(bucket->ub_bucket[i], zone->uz_size);
1363 			kasan_mark_item_invalid(zone, bucket->ub_bucket[i]);
1364 		}
1365 	zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt);
1366 	if (zone->uz_max_items > 0)
1367 		zone_free_limit(zone, bucket->ub_cnt);
1368 #ifdef INVARIANTS
1369 	bzero(bucket->ub_bucket, sizeof(void *) * bucket->ub_cnt);
1370 #endif
1371 	bucket->ub_cnt = 0;
1372 }
1373 
1374 /*
1375  * Drains the per cpu caches for a zone.
1376  *
1377  * NOTE: This may only be called while the zone is being torn down, and not
1378  * during normal operation.  This is necessary in order that we do not have
1379  * to migrate CPUs to drain the per-CPU caches.
1380  *
1381  * Arguments:
1382  *	zone     The zone to drain, must be unlocked.
1383  *
1384  * Returns:
1385  *	Nothing
1386  */
1387 static void
1388 cache_drain(uma_zone_t zone)
1389 {
1390 	uma_cache_t cache;
1391 	uma_bucket_t bucket;
1392 	smr_seq_t seq;
1393 	int cpu;
1394 
1395 	/*
1396 	 * XXX: It is safe to not lock the per-CPU caches, because we're
1397 	 * tearing down the zone anyway.  I.e., there will be no further use
1398 	 * of the caches at this point.
1399 	 *
1400 	 * XXX: It would good to be able to assert that the zone is being
1401 	 * torn down to prevent improper use of cache_drain().
1402 	 */
1403 	seq = SMR_SEQ_INVALID;
1404 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
1405 		seq = smr_advance(zone->uz_smr);
1406 	CPU_FOREACH(cpu) {
1407 		cache = &zone->uz_cpu[cpu];
1408 		bucket = cache_bucket_unload_alloc(cache);
1409 		if (bucket != NULL)
1410 			bucket_free(zone, bucket, NULL);
1411 		bucket = cache_bucket_unload_free(cache);
1412 		if (bucket != NULL) {
1413 			bucket->ub_seq = seq;
1414 			bucket_free(zone, bucket, NULL);
1415 		}
1416 		bucket = cache_bucket_unload_cross(cache);
1417 		if (bucket != NULL) {
1418 			bucket->ub_seq = seq;
1419 			bucket_free(zone, bucket, NULL);
1420 		}
1421 	}
1422 	bucket_cache_reclaim(zone, true, UMA_ANYDOMAIN);
1423 }
1424 
1425 static void
1426 cache_shrink(uma_zone_t zone, void *unused)
1427 {
1428 
1429 	if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1430 		return;
1431 
1432 	ZONE_LOCK(zone);
1433 	zone->uz_bucket_size =
1434 	    (zone->uz_bucket_size_min + zone->uz_bucket_size) / 2;
1435 	ZONE_UNLOCK(zone);
1436 }
1437 
1438 static void
1439 cache_drain_safe_cpu(uma_zone_t zone, void *unused)
1440 {
1441 	uma_cache_t cache;
1442 	uma_bucket_t b1, b2, b3;
1443 	int domain;
1444 
1445 	if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1446 		return;
1447 
1448 	b1 = b2 = b3 = NULL;
1449 	critical_enter();
1450 	cache = &zone->uz_cpu[curcpu];
1451 	domain = PCPU_GET(domain);
1452 	b1 = cache_bucket_unload_alloc(cache);
1453 
1454 	/*
1455 	 * Don't flush SMR zone buckets.  This leaves the zone without a
1456 	 * bucket and forces every free to synchronize().
1457 	 */
1458 	if ((zone->uz_flags & UMA_ZONE_SMR) == 0) {
1459 		b2 = cache_bucket_unload_free(cache);
1460 		b3 = cache_bucket_unload_cross(cache);
1461 	}
1462 	critical_exit();
1463 
1464 	if (b1 != NULL)
1465 		zone_free_bucket(zone, b1, NULL, domain, false);
1466 	if (b2 != NULL)
1467 		zone_free_bucket(zone, b2, NULL, domain, false);
1468 	if (b3 != NULL) {
1469 		/* Adjust the domain so it goes to zone_free_cross. */
1470 		domain = (domain + 1) % vm_ndomains;
1471 		zone_free_bucket(zone, b3, NULL, domain, false);
1472 	}
1473 }
1474 
1475 /*
1476  * Safely drain per-CPU caches of a zone(s) to alloc bucket.
1477  * This is an expensive call because it needs to bind to all CPUs
1478  * one by one and enter a critical section on each of them in order
1479  * to safely access their cache buckets.
1480  * Zone lock must not be held on call this function.
1481  */
1482 static void
1483 pcpu_cache_drain_safe(uma_zone_t zone)
1484 {
1485 	int cpu;
1486 
1487 	/*
1488 	 * Polite bucket sizes shrinking was not enough, shrink aggressively.
1489 	 */
1490 	if (zone)
1491 		cache_shrink(zone, NULL);
1492 	else
1493 		zone_foreach(cache_shrink, NULL);
1494 
1495 	CPU_FOREACH(cpu) {
1496 		thread_lock(curthread);
1497 		sched_bind(curthread, cpu);
1498 		thread_unlock(curthread);
1499 
1500 		if (zone)
1501 			cache_drain_safe_cpu(zone, NULL);
1502 		else
1503 			zone_foreach(cache_drain_safe_cpu, NULL);
1504 	}
1505 	thread_lock(curthread);
1506 	sched_unbind(curthread);
1507 	thread_unlock(curthread);
1508 }
1509 
1510 /*
1511  * Reclaim cached buckets from a zone.  All buckets are reclaimed if the caller
1512  * requested a drain, otherwise the per-domain caches are trimmed to either
1513  * estimated working set size.
1514  */
1515 static bool
1516 bucket_cache_reclaim_domain(uma_zone_t zone, bool drain, bool trim, int domain)
1517 {
1518 	uma_zone_domain_t zdom;
1519 	uma_bucket_t bucket;
1520 	long target;
1521 	bool done = false;
1522 
1523 	/*
1524 	 * The cross bucket is partially filled and not part of
1525 	 * the item count.  Reclaim it individually here.
1526 	 */
1527 	zdom = ZDOM_GET(zone, domain);
1528 	if ((zone->uz_flags & UMA_ZONE_SMR) == 0 || drain) {
1529 		ZONE_CROSS_LOCK(zone);
1530 		bucket = zdom->uzd_cross;
1531 		zdom->uzd_cross = NULL;
1532 		ZONE_CROSS_UNLOCK(zone);
1533 		if (bucket != NULL)
1534 			bucket_free(zone, bucket, NULL);
1535 	}
1536 
1537 	/*
1538 	 * If we were asked to drain the zone, we are done only once
1539 	 * this bucket cache is empty.  If trim, we reclaim items in
1540 	 * excess of the zone's estimated working set size.  Multiple
1541 	 * consecutive calls will shrink the WSS and so reclaim more.
1542 	 * If neither drain nor trim, then voluntarily reclaim 1/4
1543 	 * (to reduce first spike) of items not used for a long time.
1544 	 */
1545 	ZDOM_LOCK(zdom);
1546 	zone_domain_update_wss(zdom);
1547 	if (drain)
1548 		target = 0;
1549 	else if (trim)
1550 		target = zdom->uzd_wss;
1551 	else if (zdom->uzd_timin > 900 / UMA_TIMEOUT)
1552 		target = zdom->uzd_nitems - zdom->uzd_limin / 4;
1553 	else {
1554 		ZDOM_UNLOCK(zdom);
1555 		return (done);
1556 	}
1557 	while ((bucket = STAILQ_FIRST(&zdom->uzd_buckets)) != NULL &&
1558 	    zdom->uzd_nitems >= target + bucket->ub_cnt) {
1559 		bucket = zone_fetch_bucket(zone, zdom, true);
1560 		if (bucket == NULL)
1561 			break;
1562 		bucket_free(zone, bucket, NULL);
1563 		done = true;
1564 		ZDOM_LOCK(zdom);
1565 	}
1566 	ZDOM_UNLOCK(zdom);
1567 	return (done);
1568 }
1569 
1570 static void
1571 bucket_cache_reclaim(uma_zone_t zone, bool drain, int domain)
1572 {
1573 	int i;
1574 
1575 	/*
1576 	 * Shrink the zone bucket size to ensure that the per-CPU caches
1577 	 * don't grow too large.
1578 	 */
1579 	if (zone->uz_bucket_size > zone->uz_bucket_size_min)
1580 		zone->uz_bucket_size--;
1581 
1582 	if (domain != UMA_ANYDOMAIN &&
1583 	    (zone->uz_flags & UMA_ZONE_ROUNDROBIN) == 0) {
1584 		bucket_cache_reclaim_domain(zone, drain, true, domain);
1585 	} else {
1586 		for (i = 0; i < vm_ndomains; i++)
1587 			bucket_cache_reclaim_domain(zone, drain, true, i);
1588 	}
1589 }
1590 
1591 static void
1592 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start)
1593 {
1594 	uint8_t *mem;
1595 	size_t size;
1596 	int i;
1597 	uint8_t flags;
1598 
1599 	CTR4(KTR_UMA, "keg_free_slab keg %s(%p) slab %p, returning %d bytes",
1600 	    keg->uk_name, keg, slab, PAGE_SIZE * keg->uk_ppera);
1601 
1602 	mem = slab_data(slab, keg);
1603 	size = PAGE_SIZE * keg->uk_ppera;
1604 
1605 	kasan_mark_slab_valid(keg, mem);
1606 	if (keg->uk_fini != NULL) {
1607 		for (i = start - 1; i > -1; i--)
1608 #ifdef INVARIANTS
1609 		/*
1610 		 * trash_fini implies that dtor was trash_dtor. trash_fini
1611 		 * would check that memory hasn't been modified since free,
1612 		 * which executed trash_dtor.
1613 		 * That's why we need to run uma_dbg_kskip() check here,
1614 		 * albeit we don't make skip check for other init/fini
1615 		 * invocations.
1616 		 */
1617 		if (!uma_dbg_kskip(keg, slab_item(slab, keg, i)) ||
1618 		    keg->uk_fini != trash_fini)
1619 #endif
1620 			keg->uk_fini(slab_item(slab, keg, i), keg->uk_size);
1621 	}
1622 	flags = slab->us_flags;
1623 	if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) {
1624 		zone_free_item(slabzone(keg->uk_ipers), slab_tohashslab(slab),
1625 		    NULL, SKIP_NONE);
1626 	}
1627 	keg->uk_freef(mem, size, flags);
1628 	uma_total_dec(size);
1629 }
1630 
1631 static void
1632 keg_drain_domain(uma_keg_t keg, int domain)
1633 {
1634 	struct slabhead freeslabs;
1635 	uma_domain_t dom;
1636 	uma_slab_t slab, tmp;
1637 	uint32_t i, stofree, stokeep, partial;
1638 
1639 	dom = &keg->uk_domain[domain];
1640 	LIST_INIT(&freeslabs);
1641 
1642 	CTR4(KTR_UMA, "keg_drain %s(%p) domain %d free items: %u",
1643 	    keg->uk_name, keg, domain, dom->ud_free_items);
1644 
1645 	KEG_LOCK(keg, domain);
1646 
1647 	/*
1648 	 * Are the free items in partially allocated slabs sufficient to meet
1649 	 * the reserve? If not, compute the number of fully free slabs that must
1650 	 * be kept.
1651 	 */
1652 	partial = dom->ud_free_items - dom->ud_free_slabs * keg->uk_ipers;
1653 	if (partial < keg->uk_reserve) {
1654 		stokeep = min(dom->ud_free_slabs,
1655 		    howmany(keg->uk_reserve - partial, keg->uk_ipers));
1656 	} else {
1657 		stokeep = 0;
1658 	}
1659 	stofree = dom->ud_free_slabs - stokeep;
1660 
1661 	/*
1662 	 * Partition the free slabs into two sets: those that must be kept in
1663 	 * order to maintain the reserve, and those that may be released back to
1664 	 * the system.  Since one set may be much larger than the other,
1665 	 * populate the smaller of the two sets and swap them if necessary.
1666 	 */
1667 	for (i = min(stofree, stokeep); i > 0; i--) {
1668 		slab = LIST_FIRST(&dom->ud_free_slab);
1669 		LIST_REMOVE(slab, us_link);
1670 		LIST_INSERT_HEAD(&freeslabs, slab, us_link);
1671 	}
1672 	if (stofree > stokeep)
1673 		LIST_SWAP(&freeslabs, &dom->ud_free_slab, uma_slab, us_link);
1674 
1675 	if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0) {
1676 		LIST_FOREACH(slab, &freeslabs, us_link)
1677 			UMA_HASH_REMOVE(&keg->uk_hash, slab);
1678 	}
1679 	dom->ud_free_items -= stofree * keg->uk_ipers;
1680 	dom->ud_free_slabs -= stofree;
1681 	dom->ud_pages -= stofree * keg->uk_ppera;
1682 	KEG_UNLOCK(keg, domain);
1683 
1684 	LIST_FOREACH_SAFE(slab, &freeslabs, us_link, tmp)
1685 		keg_free_slab(keg, slab, keg->uk_ipers);
1686 }
1687 
1688 /*
1689  * Frees pages from a keg back to the system.  This is done on demand from
1690  * the pageout daemon.
1691  *
1692  * Returns nothing.
1693  */
1694 static void
1695 keg_drain(uma_keg_t keg, int domain)
1696 {
1697 	int i;
1698 
1699 	if ((keg->uk_flags & UMA_ZONE_NOFREE) != 0)
1700 		return;
1701 	if (domain != UMA_ANYDOMAIN) {
1702 		keg_drain_domain(keg, domain);
1703 	} else {
1704 		for (i = 0; i < vm_ndomains; i++)
1705 			keg_drain_domain(keg, i);
1706 	}
1707 }
1708 
1709 static void
1710 zone_reclaim(uma_zone_t zone, int domain, int waitok, bool drain)
1711 {
1712 	/*
1713 	 * Count active reclaim operations in order to interlock with
1714 	 * zone_dtor(), which removes the zone from global lists before
1715 	 * attempting to reclaim items itself.
1716 	 *
1717 	 * The zone may be destroyed while sleeping, so only zone_dtor() should
1718 	 * specify M_WAITOK.
1719 	 */
1720 	ZONE_LOCK(zone);
1721 	if (waitok == M_WAITOK) {
1722 		while (zone->uz_reclaimers > 0)
1723 			msleep(zone, ZONE_LOCKPTR(zone), PVM, "zonedrain", 1);
1724 	}
1725 	zone->uz_reclaimers++;
1726 	ZONE_UNLOCK(zone);
1727 	bucket_cache_reclaim(zone, drain, domain);
1728 
1729 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0)
1730 		keg_drain(zone->uz_keg, domain);
1731 	ZONE_LOCK(zone);
1732 	zone->uz_reclaimers--;
1733 	if (zone->uz_reclaimers == 0)
1734 		wakeup(zone);
1735 	ZONE_UNLOCK(zone);
1736 }
1737 
1738 static void
1739 zone_drain(uma_zone_t zone, void *arg)
1740 {
1741 	int domain;
1742 
1743 	domain = (int)(uintptr_t)arg;
1744 	zone_reclaim(zone, domain, M_NOWAIT, true);
1745 }
1746 
1747 static void
1748 zone_trim(uma_zone_t zone, void *arg)
1749 {
1750 	int domain;
1751 
1752 	domain = (int)(uintptr_t)arg;
1753 	zone_reclaim(zone, domain, M_NOWAIT, false);
1754 }
1755 
1756 /*
1757  * Allocate a new slab for a keg and inserts it into the partial slab list.
1758  * The keg should be unlocked on entry.  If the allocation succeeds it will
1759  * be locked on return.
1760  *
1761  * Arguments:
1762  *	flags   Wait flags for the item initialization routine
1763  *	aflags  Wait flags for the slab allocation
1764  *
1765  * Returns:
1766  *	The slab that was allocated or NULL if there is no memory and the
1767  *	caller specified M_NOWAIT.
1768  */
1769 static uma_slab_t
1770 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int domain, int flags,
1771     int aflags)
1772 {
1773 	uma_domain_t dom;
1774 	uma_slab_t slab;
1775 	unsigned long size;
1776 	uint8_t *mem;
1777 	uint8_t sflags;
1778 	int i;
1779 
1780 	KASSERT(domain >= 0 && domain < vm_ndomains,
1781 	    ("keg_alloc_slab: domain %d out of range", domain));
1782 
1783 	slab = NULL;
1784 	mem = NULL;
1785 	if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) {
1786 		uma_hash_slab_t hslab;
1787 		hslab = zone_alloc_item(slabzone(keg->uk_ipers), NULL,
1788 		    domain, aflags);
1789 		if (hslab == NULL)
1790 			goto fail;
1791 		slab = &hslab->uhs_slab;
1792 	}
1793 
1794 	/*
1795 	 * This reproduces the old vm_zone behavior of zero filling pages the
1796 	 * first time they are added to a zone.
1797 	 *
1798 	 * Malloced items are zeroed in uma_zalloc.
1799 	 */
1800 
1801 	if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
1802 		aflags |= M_ZERO;
1803 	else
1804 		aflags &= ~M_ZERO;
1805 
1806 	if (keg->uk_flags & UMA_ZONE_NODUMP)
1807 		aflags |= M_NODUMP;
1808 
1809 	/* zone is passed for legacy reasons. */
1810 	size = keg->uk_ppera * PAGE_SIZE;
1811 	mem = keg->uk_allocf(zone, size, domain, &sflags, aflags);
1812 	if (mem == NULL) {
1813 		if (keg->uk_flags & UMA_ZFLAG_OFFPAGE)
1814 			zone_free_item(slabzone(keg->uk_ipers),
1815 			    slab_tohashslab(slab), NULL, SKIP_NONE);
1816 		goto fail;
1817 	}
1818 	uma_total_inc(size);
1819 
1820 	/* For HASH zones all pages go to the same uma_domain. */
1821 	if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
1822 		domain = 0;
1823 
1824 	/* Point the slab into the allocated memory */
1825 	if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE))
1826 		slab = (uma_slab_t)(mem + keg->uk_pgoff);
1827 	else
1828 		slab_tohashslab(slab)->uhs_data = mem;
1829 
1830 	if (keg->uk_flags & UMA_ZFLAG_VTOSLAB)
1831 		for (i = 0; i < keg->uk_ppera; i++)
1832 			vsetzoneslab((vm_offset_t)mem + (i * PAGE_SIZE),
1833 			    zone, slab);
1834 
1835 	slab->us_freecount = keg->uk_ipers;
1836 	slab->us_flags = sflags;
1837 	slab->us_domain = domain;
1838 
1839 	BIT_FILL(keg->uk_ipers, &slab->us_free);
1840 #ifdef INVARIANTS
1841 	BIT_ZERO(keg->uk_ipers, slab_dbg_bits(slab, keg));
1842 #endif
1843 
1844 	if (keg->uk_init != NULL) {
1845 		for (i = 0; i < keg->uk_ipers; i++)
1846 			if (keg->uk_init(slab_item(slab, keg, i),
1847 			    keg->uk_size, flags) != 0)
1848 				break;
1849 		if (i != keg->uk_ipers) {
1850 			keg_free_slab(keg, slab, i);
1851 			goto fail;
1852 		}
1853 	}
1854 	kasan_mark_slab_invalid(keg, mem);
1855 	KEG_LOCK(keg, domain);
1856 
1857 	CTR3(KTR_UMA, "keg_alloc_slab: allocated slab %p for %s(%p)",
1858 	    slab, keg->uk_name, keg);
1859 
1860 	if (keg->uk_flags & UMA_ZFLAG_HASH)
1861 		UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
1862 
1863 	/*
1864 	 * If we got a slab here it's safe to mark it partially used
1865 	 * and return.  We assume that the caller is going to remove
1866 	 * at least one item.
1867 	 */
1868 	dom = &keg->uk_domain[domain];
1869 	LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
1870 	dom->ud_pages += keg->uk_ppera;
1871 	dom->ud_free_items += keg->uk_ipers;
1872 
1873 	return (slab);
1874 
1875 fail:
1876 	return (NULL);
1877 }
1878 
1879 /*
1880  * This function is intended to be used early on in place of page_alloc().  It
1881  * performs contiguous physical memory allocations and uses a bump allocator for
1882  * KVA, so is usable before the kernel map is initialized.
1883  */
1884 static void *
1885 startup_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1886     int wait)
1887 {
1888 	vm_paddr_t pa;
1889 	vm_page_t m;
1890 	int i, pages;
1891 
1892 	pages = howmany(bytes, PAGE_SIZE);
1893 	KASSERT(pages > 0, ("%s can't reserve 0 pages", __func__));
1894 
1895 	*pflag = UMA_SLAB_BOOT;
1896 	m = vm_page_alloc_noobj_contig_domain(domain, malloc2vm_flags(wait) |
1897 	    VM_ALLOC_WIRED, pages, (vm_paddr_t)0, ~(vm_paddr_t)0, 1, 0,
1898 	    VM_MEMATTR_DEFAULT);
1899 	if (m == NULL)
1900 		return (NULL);
1901 
1902 	pa = VM_PAGE_TO_PHYS(m);
1903 	for (i = 0; i < pages; i++, pa += PAGE_SIZE) {
1904 #if defined(__aarch64__) || defined(__amd64__) || \
1905     defined(__riscv) || defined(__powerpc64__)
1906 		if ((wait & M_NODUMP) == 0)
1907 			dump_add_page(pa);
1908 #endif
1909 	}
1910 
1911 	/* Allocate KVA and indirectly advance bootmem. */
1912 	return ((void *)pmap_map(&bootmem, m->phys_addr,
1913 	    m->phys_addr + (pages * PAGE_SIZE), VM_PROT_READ | VM_PROT_WRITE));
1914 }
1915 
1916 static void
1917 startup_free(void *mem, vm_size_t bytes)
1918 {
1919 	vm_offset_t va;
1920 	vm_page_t m;
1921 
1922 	va = (vm_offset_t)mem;
1923 	m = PHYS_TO_VM_PAGE(pmap_kextract(va));
1924 
1925 	/*
1926 	 * startup_alloc() returns direct-mapped slabs on some platforms.  Avoid
1927 	 * unmapping ranges of the direct map.
1928 	 */
1929 	if (va >= bootstart && va + bytes <= bootmem)
1930 		pmap_remove(kernel_pmap, va, va + bytes);
1931 	for (; bytes != 0; bytes -= PAGE_SIZE, m++) {
1932 #if defined(__aarch64__) || defined(__amd64__) || \
1933     defined(__riscv) || defined(__powerpc64__)
1934 		dump_drop_page(VM_PAGE_TO_PHYS(m));
1935 #endif
1936 		vm_page_unwire_noq(m);
1937 		vm_page_free(m);
1938 	}
1939 }
1940 
1941 /*
1942  * Allocates a number of pages from the system
1943  *
1944  * Arguments:
1945  *	bytes  The number of bytes requested
1946  *	wait  Shall we wait?
1947  *
1948  * Returns:
1949  *	A pointer to the alloced memory or possibly
1950  *	NULL if M_NOWAIT is set.
1951  */
1952 static void *
1953 page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1954     int wait)
1955 {
1956 	void *p;	/* Returned page */
1957 
1958 	*pflag = UMA_SLAB_KERNEL;
1959 	p = (void *)kmem_malloc_domainset(DOMAINSET_FIXED(domain), bytes, wait);
1960 
1961 	return (p);
1962 }
1963 
1964 static void *
1965 pcpu_page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1966     int wait)
1967 {
1968 	struct pglist alloctail;
1969 	vm_offset_t addr, zkva;
1970 	int cpu, flags;
1971 	vm_page_t p, p_next;
1972 #ifdef NUMA
1973 	struct pcpu *pc;
1974 #endif
1975 
1976 	MPASS(bytes == (mp_maxid + 1) * PAGE_SIZE);
1977 
1978 	TAILQ_INIT(&alloctail);
1979 	flags = VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | malloc2vm_flags(wait);
1980 	*pflag = UMA_SLAB_KERNEL;
1981 	for (cpu = 0; cpu <= mp_maxid; cpu++) {
1982 		if (CPU_ABSENT(cpu)) {
1983 			p = vm_page_alloc_noobj(flags);
1984 		} else {
1985 #ifndef NUMA
1986 			p = vm_page_alloc_noobj(flags);
1987 #else
1988 			pc = pcpu_find(cpu);
1989 			if (__predict_false(VM_DOMAIN_EMPTY(pc->pc_domain)))
1990 				p = NULL;
1991 			else
1992 				p = vm_page_alloc_noobj_domain(pc->pc_domain,
1993 				    flags);
1994 			if (__predict_false(p == NULL))
1995 				p = vm_page_alloc_noobj(flags);
1996 #endif
1997 		}
1998 		if (__predict_false(p == NULL))
1999 			goto fail;
2000 		TAILQ_INSERT_TAIL(&alloctail, p, listq);
2001 	}
2002 	if ((addr = kva_alloc(bytes)) == 0)
2003 		goto fail;
2004 	zkva = addr;
2005 	TAILQ_FOREACH(p, &alloctail, listq) {
2006 		pmap_qenter(zkva, &p, 1);
2007 		zkva += PAGE_SIZE;
2008 	}
2009 	return ((void*)addr);
2010 fail:
2011 	TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
2012 		vm_page_unwire_noq(p);
2013 		vm_page_free(p);
2014 	}
2015 	return (NULL);
2016 }
2017 
2018 /*
2019  * Allocates a number of pages not belonging to a VM object
2020  *
2021  * Arguments:
2022  *	bytes  The number of bytes requested
2023  *	wait   Shall we wait?
2024  *
2025  * Returns:
2026  *	A pointer to the alloced memory or possibly
2027  *	NULL if M_NOWAIT is set.
2028  */
2029 static void *
2030 noobj_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags,
2031     int wait)
2032 {
2033 	TAILQ_HEAD(, vm_page) alloctail;
2034 	u_long npages;
2035 	vm_offset_t retkva, zkva;
2036 	vm_page_t p, p_next;
2037 	uma_keg_t keg;
2038 	int req;
2039 
2040 	TAILQ_INIT(&alloctail);
2041 	keg = zone->uz_keg;
2042 	req = VM_ALLOC_INTERRUPT | VM_ALLOC_WIRED;
2043 	if ((wait & M_WAITOK) != 0)
2044 		req |= VM_ALLOC_WAITOK;
2045 
2046 	npages = howmany(bytes, PAGE_SIZE);
2047 	while (npages > 0) {
2048 		p = vm_page_alloc_noobj_domain(domain, req);
2049 		if (p != NULL) {
2050 			/*
2051 			 * Since the page does not belong to an object, its
2052 			 * listq is unused.
2053 			 */
2054 			TAILQ_INSERT_TAIL(&alloctail, p, listq);
2055 			npages--;
2056 			continue;
2057 		}
2058 		/*
2059 		 * Page allocation failed, free intermediate pages and
2060 		 * exit.
2061 		 */
2062 		TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
2063 			vm_page_unwire_noq(p);
2064 			vm_page_free(p);
2065 		}
2066 		return (NULL);
2067 	}
2068 	*flags = UMA_SLAB_PRIV;
2069 	zkva = keg->uk_kva +
2070 	    atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
2071 	retkva = zkva;
2072 	TAILQ_FOREACH(p, &alloctail, listq) {
2073 		pmap_qenter(zkva, &p, 1);
2074 		zkva += PAGE_SIZE;
2075 	}
2076 
2077 	return ((void *)retkva);
2078 }
2079 
2080 /*
2081  * Allocate physically contiguous pages.
2082  */
2083 static void *
2084 contig_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
2085     int wait)
2086 {
2087 
2088 	*pflag = UMA_SLAB_KERNEL;
2089 	return ((void *)kmem_alloc_contig_domainset(DOMAINSET_FIXED(domain),
2090 	    bytes, wait, 0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT));
2091 }
2092 
2093 /*
2094  * Frees a number of pages to the system
2095  *
2096  * Arguments:
2097  *	mem   A pointer to the memory to be freed
2098  *	size  The size of the memory being freed
2099  *	flags The original p->us_flags field
2100  *
2101  * Returns:
2102  *	Nothing
2103  */
2104 static void
2105 page_free(void *mem, vm_size_t size, uint8_t flags)
2106 {
2107 
2108 	if ((flags & UMA_SLAB_BOOT) != 0) {
2109 		startup_free(mem, size);
2110 		return;
2111 	}
2112 
2113 	KASSERT((flags & UMA_SLAB_KERNEL) != 0,
2114 	    ("UMA: page_free used with invalid flags %x", flags));
2115 
2116 	kmem_free((vm_offset_t)mem, size);
2117 }
2118 
2119 /*
2120  * Frees pcpu zone allocations
2121  *
2122  * Arguments:
2123  *	mem   A pointer to the memory to be freed
2124  *	size  The size of the memory being freed
2125  *	flags The original p->us_flags field
2126  *
2127  * Returns:
2128  *	Nothing
2129  */
2130 static void
2131 pcpu_page_free(void *mem, vm_size_t size, uint8_t flags)
2132 {
2133 	vm_offset_t sva, curva;
2134 	vm_paddr_t paddr;
2135 	vm_page_t m;
2136 
2137 	MPASS(size == (mp_maxid+1)*PAGE_SIZE);
2138 
2139 	if ((flags & UMA_SLAB_BOOT) != 0) {
2140 		startup_free(mem, size);
2141 		return;
2142 	}
2143 
2144 	sva = (vm_offset_t)mem;
2145 	for (curva = sva; curva < sva + size; curva += PAGE_SIZE) {
2146 		paddr = pmap_kextract(curva);
2147 		m = PHYS_TO_VM_PAGE(paddr);
2148 		vm_page_unwire_noq(m);
2149 		vm_page_free(m);
2150 	}
2151 	pmap_qremove(sva, size >> PAGE_SHIFT);
2152 	kva_free(sva, size);
2153 }
2154 
2155 /*
2156  * Zero fill initializer
2157  *
2158  * Arguments/Returns follow uma_init specifications
2159  */
2160 static int
2161 zero_init(void *mem, int size, int flags)
2162 {
2163 	bzero(mem, size);
2164 	return (0);
2165 }
2166 
2167 #ifdef INVARIANTS
2168 static struct noslabbits *
2169 slab_dbg_bits(uma_slab_t slab, uma_keg_t keg)
2170 {
2171 
2172 	return ((void *)((char *)&slab->us_free + BITSET_SIZE(keg->uk_ipers)));
2173 }
2174 #endif
2175 
2176 /*
2177  * Actual size of embedded struct slab (!OFFPAGE).
2178  */
2179 static size_t
2180 slab_sizeof(int nitems)
2181 {
2182 	size_t s;
2183 
2184 	s = sizeof(struct uma_slab) + BITSET_SIZE(nitems) * SLAB_BITSETS;
2185 	return (roundup(s, UMA_ALIGN_PTR + 1));
2186 }
2187 
2188 #define	UMA_FIXPT_SHIFT	31
2189 #define	UMA_FRAC_FIXPT(n, d)						\
2190 	((uint32_t)(((uint64_t)(n) << UMA_FIXPT_SHIFT) / (d)))
2191 #define	UMA_FIXPT_PCT(f)						\
2192 	((u_int)(((uint64_t)100 * (f)) >> UMA_FIXPT_SHIFT))
2193 #define	UMA_PCT_FIXPT(pct)	UMA_FRAC_FIXPT((pct), 100)
2194 #define	UMA_MIN_EFF	UMA_PCT_FIXPT(100 - UMA_MAX_WASTE)
2195 
2196 /*
2197  * Compute the number of items that will fit in a slab.  If hdr is true, the
2198  * item count may be limited to provide space in the slab for an inline slab
2199  * header.  Otherwise, all slab space will be provided for item storage.
2200  */
2201 static u_int
2202 slab_ipers_hdr(u_int size, u_int rsize, u_int slabsize, bool hdr)
2203 {
2204 	u_int ipers;
2205 	u_int padpi;
2206 
2207 	/* The padding between items is not needed after the last item. */
2208 	padpi = rsize - size;
2209 
2210 	if (hdr) {
2211 		/*
2212 		 * Start with the maximum item count and remove items until
2213 		 * the slab header first alongside the allocatable memory.
2214 		 */
2215 		for (ipers = MIN(SLAB_MAX_SETSIZE,
2216 		    (slabsize + padpi - slab_sizeof(1)) / rsize);
2217 		    ipers > 0 &&
2218 		    ipers * rsize - padpi + slab_sizeof(ipers) > slabsize;
2219 		    ipers--)
2220 			continue;
2221 	} else {
2222 		ipers = MIN((slabsize + padpi) / rsize, SLAB_MAX_SETSIZE);
2223 	}
2224 
2225 	return (ipers);
2226 }
2227 
2228 struct keg_layout_result {
2229 	u_int format;
2230 	u_int slabsize;
2231 	u_int ipers;
2232 	u_int eff;
2233 };
2234 
2235 static void
2236 keg_layout_one(uma_keg_t keg, u_int rsize, u_int slabsize, u_int fmt,
2237     struct keg_layout_result *kl)
2238 {
2239 	u_int total;
2240 
2241 	kl->format = fmt;
2242 	kl->slabsize = slabsize;
2243 
2244 	/* Handle INTERNAL as inline with an extra page. */
2245 	if ((fmt & UMA_ZFLAG_INTERNAL) != 0) {
2246 		kl->format &= ~UMA_ZFLAG_INTERNAL;
2247 		kl->slabsize += PAGE_SIZE;
2248 	}
2249 
2250 	kl->ipers = slab_ipers_hdr(keg->uk_size, rsize, kl->slabsize,
2251 	    (fmt & UMA_ZFLAG_OFFPAGE) == 0);
2252 
2253 	/* Account for memory used by an offpage slab header. */
2254 	total = kl->slabsize;
2255 	if ((fmt & UMA_ZFLAG_OFFPAGE) != 0)
2256 		total += slabzone(kl->ipers)->uz_keg->uk_rsize;
2257 
2258 	kl->eff = UMA_FRAC_FIXPT(kl->ipers * rsize, total);
2259 }
2260 
2261 /*
2262  * Determine the format of a uma keg.  This determines where the slab header
2263  * will be placed (inline or offpage) and calculates ipers, rsize, and ppera.
2264  *
2265  * Arguments
2266  *	keg  The zone we should initialize
2267  *
2268  * Returns
2269  *	Nothing
2270  */
2271 static void
2272 keg_layout(uma_keg_t keg)
2273 {
2274 	struct keg_layout_result kl = {}, kl_tmp;
2275 	u_int fmts[2];
2276 	u_int alignsize;
2277 	u_int nfmt;
2278 	u_int pages;
2279 	u_int rsize;
2280 	u_int slabsize;
2281 	u_int i, j;
2282 
2283 	KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
2284 	    (keg->uk_size <= UMA_PCPU_ALLOC_SIZE &&
2285 	     (keg->uk_flags & UMA_ZONE_CACHESPREAD) == 0),
2286 	    ("%s: cannot configure for PCPU: keg=%s, size=%u, flags=0x%b",
2287 	     __func__, keg->uk_name, keg->uk_size, keg->uk_flags,
2288 	     PRINT_UMA_ZFLAGS));
2289 	KASSERT((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) == 0 ||
2290 	    (keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0,
2291 	    ("%s: incompatible flags 0x%b", __func__, keg->uk_flags,
2292 	     PRINT_UMA_ZFLAGS));
2293 
2294 	alignsize = keg->uk_align + 1;
2295 #ifdef KASAN
2296 	/*
2297 	 * ASAN requires that each allocation be aligned to the shadow map
2298 	 * scale factor.
2299 	 */
2300 	if (alignsize < KASAN_SHADOW_SCALE)
2301 		alignsize = KASAN_SHADOW_SCALE;
2302 #endif
2303 
2304 	/*
2305 	 * Calculate the size of each allocation (rsize) according to
2306 	 * alignment.  If the requested size is smaller than we have
2307 	 * allocation bits for we round it up.
2308 	 */
2309 	rsize = MAX(keg->uk_size, UMA_SMALLEST_UNIT);
2310 	rsize = roundup2(rsize, alignsize);
2311 
2312 	if ((keg->uk_flags & UMA_ZONE_CACHESPREAD) != 0) {
2313 		/*
2314 		 * We want one item to start on every align boundary in a page.
2315 		 * To do this we will span pages.  We will also extend the item
2316 		 * by the size of align if it is an even multiple of align.
2317 		 * Otherwise, it would fall on the same boundary every time.
2318 		 */
2319 		if ((rsize & alignsize) == 0)
2320 			rsize += alignsize;
2321 		slabsize = rsize * (PAGE_SIZE / alignsize);
2322 		slabsize = MIN(slabsize, rsize * SLAB_MAX_SETSIZE);
2323 		slabsize = MIN(slabsize, UMA_CACHESPREAD_MAX_SIZE);
2324 		slabsize = round_page(slabsize);
2325 	} else {
2326 		/*
2327 		 * Start with a slab size of as many pages as it takes to
2328 		 * represent a single item.  We will try to fit as many
2329 		 * additional items into the slab as possible.
2330 		 */
2331 		slabsize = round_page(keg->uk_size);
2332 	}
2333 
2334 	/* Build a list of all of the available formats for this keg. */
2335 	nfmt = 0;
2336 
2337 	/* Evaluate an inline slab layout. */
2338 	if ((keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0)
2339 		fmts[nfmt++] = 0;
2340 
2341 	/* TODO: vm_page-embedded slab. */
2342 
2343 	/*
2344 	 * We can't do OFFPAGE if we're internal or if we've been
2345 	 * asked to not go to the VM for buckets.  If we do this we
2346 	 * may end up going to the VM for slabs which we do not want
2347 	 * to do if we're UMA_ZONE_VM, which clearly forbids it.
2348 	 * In those cases, evaluate a pseudo-format called INTERNAL
2349 	 * which has an inline slab header and one extra page to
2350 	 * guarantee that it fits.
2351 	 *
2352 	 * Otherwise, see if using an OFFPAGE slab will improve our
2353 	 * efficiency.
2354 	 */
2355 	if ((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) != 0)
2356 		fmts[nfmt++] = UMA_ZFLAG_INTERNAL;
2357 	else
2358 		fmts[nfmt++] = UMA_ZFLAG_OFFPAGE;
2359 
2360 	/*
2361 	 * Choose a slab size and format which satisfy the minimum efficiency.
2362 	 * Prefer the smallest slab size that meets the constraints.
2363 	 *
2364 	 * Start with a minimum slab size, to accommodate CACHESPREAD.  Then,
2365 	 * for small items (up to PAGE_SIZE), the iteration increment is one
2366 	 * page; and for large items, the increment is one item.
2367 	 */
2368 	i = (slabsize + rsize - keg->uk_size) / MAX(PAGE_SIZE, rsize);
2369 	KASSERT(i >= 1, ("keg %s(%p) flags=0x%b slabsize=%u, rsize=%u, i=%u",
2370 	    keg->uk_name, keg, keg->uk_flags, PRINT_UMA_ZFLAGS, slabsize,
2371 	    rsize, i));
2372 	for ( ; ; i++) {
2373 		slabsize = (rsize <= PAGE_SIZE) ? ptoa(i) :
2374 		    round_page(rsize * (i - 1) + keg->uk_size);
2375 
2376 		for (j = 0; j < nfmt; j++) {
2377 			/* Only if we have no viable format yet. */
2378 			if ((fmts[j] & UMA_ZFLAG_INTERNAL) != 0 &&
2379 			    kl.ipers > 0)
2380 				continue;
2381 
2382 			keg_layout_one(keg, rsize, slabsize, fmts[j], &kl_tmp);
2383 			if (kl_tmp.eff <= kl.eff)
2384 				continue;
2385 
2386 			kl = kl_tmp;
2387 
2388 			CTR6(KTR_UMA, "keg %s layout: format %#x "
2389 			    "(ipers %u * rsize %u) / slabsize %#x = %u%% eff",
2390 			    keg->uk_name, kl.format, kl.ipers, rsize,
2391 			    kl.slabsize, UMA_FIXPT_PCT(kl.eff));
2392 
2393 			/* Stop when we reach the minimum efficiency. */
2394 			if (kl.eff >= UMA_MIN_EFF)
2395 				break;
2396 		}
2397 
2398 		if (kl.eff >= UMA_MIN_EFF || !multipage_slabs ||
2399 		    slabsize >= SLAB_MAX_SETSIZE * rsize ||
2400 		    (keg->uk_flags & (UMA_ZONE_PCPU | UMA_ZONE_CONTIG)) != 0)
2401 			break;
2402 	}
2403 
2404 	pages = atop(kl.slabsize);
2405 	if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
2406 		pages *= mp_maxid + 1;
2407 
2408 	keg->uk_rsize = rsize;
2409 	keg->uk_ipers = kl.ipers;
2410 	keg->uk_ppera = pages;
2411 	keg->uk_flags |= kl.format;
2412 
2413 	/*
2414 	 * How do we find the slab header if it is offpage or if not all item
2415 	 * start addresses are in the same page?  We could solve the latter
2416 	 * case with vaddr alignment, but we don't.
2417 	 */
2418 	if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0 ||
2419 	    (keg->uk_ipers - 1) * rsize >= PAGE_SIZE) {
2420 		if ((keg->uk_flags & UMA_ZONE_NOTPAGE) != 0)
2421 			keg->uk_flags |= UMA_ZFLAG_HASH;
2422 		else
2423 			keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2424 	}
2425 
2426 	CTR6(KTR_UMA, "%s: keg=%s, flags=%#x, rsize=%u, ipers=%u, ppera=%u",
2427 	    __func__, keg->uk_name, keg->uk_flags, rsize, keg->uk_ipers,
2428 	    pages);
2429 	KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_MAX_SETSIZE,
2430 	    ("%s: keg=%s, flags=0x%b, rsize=%u, ipers=%u, ppera=%u", __func__,
2431 	     keg->uk_name, keg->uk_flags, PRINT_UMA_ZFLAGS, rsize,
2432 	     keg->uk_ipers, pages));
2433 }
2434 
2435 /*
2436  * Keg header ctor.  This initializes all fields, locks, etc.  And inserts
2437  * the keg onto the global keg list.
2438  *
2439  * Arguments/Returns follow uma_ctor specifications
2440  *	udata  Actually uma_kctor_args
2441  */
2442 static int
2443 keg_ctor(void *mem, int size, void *udata, int flags)
2444 {
2445 	struct uma_kctor_args *arg = udata;
2446 	uma_keg_t keg = mem;
2447 	uma_zone_t zone;
2448 	int i;
2449 
2450 	bzero(keg, size);
2451 	keg->uk_size = arg->size;
2452 	keg->uk_init = arg->uminit;
2453 	keg->uk_fini = arg->fini;
2454 	keg->uk_align = arg->align;
2455 	keg->uk_reserve = 0;
2456 	keg->uk_flags = arg->flags;
2457 
2458 	/*
2459 	 * We use a global round-robin policy by default.  Zones with
2460 	 * UMA_ZONE_FIRSTTOUCH set will use first-touch instead, in which
2461 	 * case the iterator is never run.
2462 	 */
2463 	keg->uk_dr.dr_policy = DOMAINSET_RR();
2464 	keg->uk_dr.dr_iter = 0;
2465 
2466 	/*
2467 	 * The primary zone is passed to us at keg-creation time.
2468 	 */
2469 	zone = arg->zone;
2470 	keg->uk_name = zone->uz_name;
2471 
2472 	if (arg->flags & UMA_ZONE_ZINIT)
2473 		keg->uk_init = zero_init;
2474 
2475 	if (arg->flags & UMA_ZONE_MALLOC)
2476 		keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2477 
2478 #ifndef SMP
2479 	keg->uk_flags &= ~UMA_ZONE_PCPU;
2480 #endif
2481 
2482 	keg_layout(keg);
2483 
2484 	/*
2485 	 * Use a first-touch NUMA policy for kegs that pmap_extract() will
2486 	 * work on.  Use round-robin for everything else.
2487 	 *
2488 	 * Zones may override the default by specifying either.
2489 	 */
2490 #ifdef NUMA
2491 	if ((keg->uk_flags &
2492 	    (UMA_ZONE_ROUNDROBIN | UMA_ZFLAG_CACHE | UMA_ZONE_NOTPAGE)) == 0)
2493 		keg->uk_flags |= UMA_ZONE_FIRSTTOUCH;
2494 	else if ((keg->uk_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2495 		keg->uk_flags |= UMA_ZONE_ROUNDROBIN;
2496 #endif
2497 
2498 	/*
2499 	 * If we haven't booted yet we need allocations to go through the
2500 	 * startup cache until the vm is ready.
2501 	 */
2502 #ifdef UMA_MD_SMALL_ALLOC
2503 	if (keg->uk_ppera == 1)
2504 		keg->uk_allocf = uma_small_alloc;
2505 	else
2506 #endif
2507 	if (booted < BOOT_KVA)
2508 		keg->uk_allocf = startup_alloc;
2509 	else if (keg->uk_flags & UMA_ZONE_PCPU)
2510 		keg->uk_allocf = pcpu_page_alloc;
2511 	else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 && keg->uk_ppera > 1)
2512 		keg->uk_allocf = contig_alloc;
2513 	else
2514 		keg->uk_allocf = page_alloc;
2515 #ifdef UMA_MD_SMALL_ALLOC
2516 	if (keg->uk_ppera == 1)
2517 		keg->uk_freef = uma_small_free;
2518 	else
2519 #endif
2520 	if (keg->uk_flags & UMA_ZONE_PCPU)
2521 		keg->uk_freef = pcpu_page_free;
2522 	else
2523 		keg->uk_freef = page_free;
2524 
2525 	/*
2526 	 * Initialize keg's locks.
2527 	 */
2528 	for (i = 0; i < vm_ndomains; i++)
2529 		KEG_LOCK_INIT(keg, i, (arg->flags & UMA_ZONE_MTXCLASS));
2530 
2531 	/*
2532 	 * If we're putting the slab header in the actual page we need to
2533 	 * figure out where in each page it goes.  See slab_sizeof
2534 	 * definition.
2535 	 */
2536 	if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE)) {
2537 		size_t shsize;
2538 
2539 		shsize = slab_sizeof(keg->uk_ipers);
2540 		keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - shsize;
2541 		/*
2542 		 * The only way the following is possible is if with our
2543 		 * UMA_ALIGN_PTR adjustments we are now bigger than
2544 		 * UMA_SLAB_SIZE.  I haven't checked whether this is
2545 		 * mathematically possible for all cases, so we make
2546 		 * sure here anyway.
2547 		 */
2548 		KASSERT(keg->uk_pgoff + shsize <= PAGE_SIZE * keg->uk_ppera,
2549 		    ("zone %s ipers %d rsize %d size %d slab won't fit",
2550 		    zone->uz_name, keg->uk_ipers, keg->uk_rsize, keg->uk_size));
2551 	}
2552 
2553 	if (keg->uk_flags & UMA_ZFLAG_HASH)
2554 		hash_alloc(&keg->uk_hash, 0);
2555 
2556 	CTR3(KTR_UMA, "keg_ctor %p zone %s(%p)", keg, zone->uz_name, zone);
2557 
2558 	LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
2559 
2560 	rw_wlock(&uma_rwlock);
2561 	LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
2562 	rw_wunlock(&uma_rwlock);
2563 	return (0);
2564 }
2565 
2566 static void
2567 zone_kva_available(uma_zone_t zone, void *unused)
2568 {
2569 	uma_keg_t keg;
2570 
2571 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
2572 		return;
2573 	KEG_GET(zone, keg);
2574 
2575 	if (keg->uk_allocf == startup_alloc) {
2576 		/* Switch to the real allocator. */
2577 		if (keg->uk_flags & UMA_ZONE_PCPU)
2578 			keg->uk_allocf = pcpu_page_alloc;
2579 		else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 &&
2580 		    keg->uk_ppera > 1)
2581 			keg->uk_allocf = contig_alloc;
2582 		else
2583 			keg->uk_allocf = page_alloc;
2584 	}
2585 }
2586 
2587 static void
2588 zone_alloc_counters(uma_zone_t zone, void *unused)
2589 {
2590 
2591 	zone->uz_allocs = counter_u64_alloc(M_WAITOK);
2592 	zone->uz_frees = counter_u64_alloc(M_WAITOK);
2593 	zone->uz_fails = counter_u64_alloc(M_WAITOK);
2594 	zone->uz_xdomain = counter_u64_alloc(M_WAITOK);
2595 }
2596 
2597 static void
2598 zone_alloc_sysctl(uma_zone_t zone, void *unused)
2599 {
2600 	uma_zone_domain_t zdom;
2601 	uma_domain_t dom;
2602 	uma_keg_t keg;
2603 	struct sysctl_oid *oid, *domainoid;
2604 	int domains, i, cnt;
2605 	static const char *nokeg = "cache zone";
2606 	char *c;
2607 
2608 	/*
2609 	 * Make a sysctl safe copy of the zone name by removing
2610 	 * any special characters and handling dups by appending
2611 	 * an index.
2612 	 */
2613 	if (zone->uz_namecnt != 0) {
2614 		/* Count the number of decimal digits and '_' separator. */
2615 		for (i = 1, cnt = zone->uz_namecnt; cnt != 0; i++)
2616 			cnt /= 10;
2617 		zone->uz_ctlname = malloc(strlen(zone->uz_name) + i + 1,
2618 		    M_UMA, M_WAITOK);
2619 		sprintf(zone->uz_ctlname, "%s_%d", zone->uz_name,
2620 		    zone->uz_namecnt);
2621 	} else
2622 		zone->uz_ctlname = strdup(zone->uz_name, M_UMA);
2623 	for (c = zone->uz_ctlname; *c != '\0'; c++)
2624 		if (strchr("./\\ -", *c) != NULL)
2625 			*c = '_';
2626 
2627 	/*
2628 	 * Basic parameters at the root.
2629 	 */
2630 	zone->uz_oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_vm_uma),
2631 	    OID_AUTO, zone->uz_ctlname, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2632 	oid = zone->uz_oid;
2633 	SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2634 	    "size", CTLFLAG_RD, &zone->uz_size, 0, "Allocation size");
2635 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2636 	    "flags", CTLFLAG_RD | CTLTYPE_STRING | CTLFLAG_MPSAFE,
2637 	    zone, 0, sysctl_handle_uma_zone_flags, "A",
2638 	    "Allocator configuration flags");
2639 	SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2640 	    "bucket_size", CTLFLAG_RD, &zone->uz_bucket_size, 0,
2641 	    "Desired per-cpu cache size");
2642 	SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2643 	    "bucket_size_max", CTLFLAG_RD, &zone->uz_bucket_size_max, 0,
2644 	    "Maximum allowed per-cpu cache size");
2645 
2646 	/*
2647 	 * keg if present.
2648 	 */
2649 	if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
2650 		domains = vm_ndomains;
2651 	else
2652 		domains = 1;
2653 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2654 	    "keg", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2655 	keg = zone->uz_keg;
2656 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0) {
2657 		SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2658 		    "name", CTLFLAG_RD, keg->uk_name, "Keg name");
2659 		SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2660 		    "rsize", CTLFLAG_RD, &keg->uk_rsize, 0,
2661 		    "Real object size with alignment");
2662 		SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2663 		    "ppera", CTLFLAG_RD, &keg->uk_ppera, 0,
2664 		    "pages per-slab allocation");
2665 		SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2666 		    "ipers", CTLFLAG_RD, &keg->uk_ipers, 0,
2667 		    "items available per-slab");
2668 		SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2669 		    "align", CTLFLAG_RD, &keg->uk_align, 0,
2670 		    "item alignment mask");
2671 		SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2672 		    "reserve", CTLFLAG_RD, &keg->uk_reserve, 0,
2673 		    "number of reserved items");
2674 		SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2675 		    "efficiency", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2676 		    keg, 0, sysctl_handle_uma_slab_efficiency, "I",
2677 		    "Slab utilization (100 - internal fragmentation %)");
2678 		domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(oid),
2679 		    OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2680 		for (i = 0; i < domains; i++) {
2681 			dom = &keg->uk_domain[i];
2682 			oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2683 			    OID_AUTO, VM_DOMAIN(i)->vmd_name,
2684 			    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2685 			SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2686 			    "pages", CTLFLAG_RD, &dom->ud_pages, 0,
2687 			    "Total pages currently allocated from VM");
2688 			SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2689 			    "free_items", CTLFLAG_RD, &dom->ud_free_items, 0,
2690 			    "Items free in the slab layer");
2691 			SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2692 			    "free_slabs", CTLFLAG_RD, &dom->ud_free_slabs, 0,
2693 			    "Unused slabs");
2694 		}
2695 	} else
2696 		SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2697 		    "name", CTLFLAG_RD, nokeg, "Keg name");
2698 
2699 	/*
2700 	 * Information about zone limits.
2701 	 */
2702 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2703 	    "limit", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2704 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2705 	    "items", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2706 	    zone, 0, sysctl_handle_uma_zone_items, "QU",
2707 	    "Current number of allocated items if limit is set");
2708 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2709 	    "max_items", CTLFLAG_RD, &zone->uz_max_items, 0,
2710 	    "Maximum number of allocated and cached items");
2711 	SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2712 	    "sleepers", CTLFLAG_RD, &zone->uz_sleepers, 0,
2713 	    "Number of threads sleeping at limit");
2714 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2715 	    "sleeps", CTLFLAG_RD, &zone->uz_sleeps, 0,
2716 	    "Total zone limit sleeps");
2717 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2718 	    "bucket_max", CTLFLAG_RD, &zone->uz_bucket_max, 0,
2719 	    "Maximum number of items in each domain's bucket cache");
2720 
2721 	/*
2722 	 * Per-domain zone information.
2723 	 */
2724 	domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid),
2725 	    OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2726 	for (i = 0; i < domains; i++) {
2727 		zdom = ZDOM_GET(zone, i);
2728 		oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2729 		    OID_AUTO, VM_DOMAIN(i)->vmd_name,
2730 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2731 		SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2732 		    "nitems", CTLFLAG_RD, &zdom->uzd_nitems,
2733 		    "number of items in this domain");
2734 		SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2735 		    "imax", CTLFLAG_RD, &zdom->uzd_imax,
2736 		    "maximum item count in this period");
2737 		SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2738 		    "imin", CTLFLAG_RD, &zdom->uzd_imin,
2739 		    "minimum item count in this period");
2740 		SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2741 		    "bimin", CTLFLAG_RD, &zdom->uzd_bimin,
2742 		    "Minimum item count in this batch");
2743 		SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2744 		    "wss", CTLFLAG_RD, &zdom->uzd_wss,
2745 		    "Working set size");
2746 		SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2747 		    "limin", CTLFLAG_RD, &zdom->uzd_limin,
2748 		    "Long time minimum item count");
2749 		SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2750 		    "timin", CTLFLAG_RD, &zdom->uzd_timin, 0,
2751 		    "Time since zero long time minimum item count");
2752 	}
2753 
2754 	/*
2755 	 * General statistics.
2756 	 */
2757 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2758 	    "stats", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2759 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2760 	    "current", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2761 	    zone, 1, sysctl_handle_uma_zone_cur, "I",
2762 	    "Current number of allocated items");
2763 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2764 	    "allocs", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2765 	    zone, 0, sysctl_handle_uma_zone_allocs, "QU",
2766 	    "Total allocation calls");
2767 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2768 	    "frees", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2769 	    zone, 0, sysctl_handle_uma_zone_frees, "QU",
2770 	    "Total free calls");
2771 	SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2772 	    "fails", CTLFLAG_RD, &zone->uz_fails,
2773 	    "Number of allocation failures");
2774 	SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2775 	    "xdomain", CTLFLAG_RD, &zone->uz_xdomain,
2776 	    "Free calls from the wrong domain");
2777 }
2778 
2779 struct uma_zone_count {
2780 	const char	*name;
2781 	int		count;
2782 };
2783 
2784 static void
2785 zone_count(uma_zone_t zone, void *arg)
2786 {
2787 	struct uma_zone_count *cnt;
2788 
2789 	cnt = arg;
2790 	/*
2791 	 * Some zones are rapidly created with identical names and
2792 	 * destroyed out of order.  This can lead to gaps in the count.
2793 	 * Use one greater than the maximum observed for this name.
2794 	 */
2795 	if (strcmp(zone->uz_name, cnt->name) == 0)
2796 		cnt->count = MAX(cnt->count,
2797 		    zone->uz_namecnt + 1);
2798 }
2799 
2800 static void
2801 zone_update_caches(uma_zone_t zone)
2802 {
2803 	int i;
2804 
2805 	for (i = 0; i <= mp_maxid; i++) {
2806 		cache_set_uz_size(&zone->uz_cpu[i], zone->uz_size);
2807 		cache_set_uz_flags(&zone->uz_cpu[i], zone->uz_flags);
2808 	}
2809 }
2810 
2811 /*
2812  * Zone header ctor.  This initializes all fields, locks, etc.
2813  *
2814  * Arguments/Returns follow uma_ctor specifications
2815  *	udata  Actually uma_zctor_args
2816  */
2817 static int
2818 zone_ctor(void *mem, int size, void *udata, int flags)
2819 {
2820 	struct uma_zone_count cnt;
2821 	struct uma_zctor_args *arg = udata;
2822 	uma_zone_domain_t zdom;
2823 	uma_zone_t zone = mem;
2824 	uma_zone_t z;
2825 	uma_keg_t keg;
2826 	int i;
2827 
2828 	bzero(zone, size);
2829 	zone->uz_name = arg->name;
2830 	zone->uz_ctor = arg->ctor;
2831 	zone->uz_dtor = arg->dtor;
2832 	zone->uz_init = NULL;
2833 	zone->uz_fini = NULL;
2834 	zone->uz_sleeps = 0;
2835 	zone->uz_bucket_size = 0;
2836 	zone->uz_bucket_size_min = 0;
2837 	zone->uz_bucket_size_max = BUCKET_MAX;
2838 	zone->uz_flags = (arg->flags & UMA_ZONE_SMR);
2839 	zone->uz_warning = NULL;
2840 	/* The domain structures follow the cpu structures. */
2841 	zone->uz_bucket_max = ULONG_MAX;
2842 	timevalclear(&zone->uz_ratecheck);
2843 
2844 	/* Count the number of duplicate names. */
2845 	cnt.name = arg->name;
2846 	cnt.count = 0;
2847 	zone_foreach(zone_count, &cnt);
2848 	zone->uz_namecnt = cnt.count;
2849 	ZONE_CROSS_LOCK_INIT(zone);
2850 
2851 	for (i = 0; i < vm_ndomains; i++) {
2852 		zdom = ZDOM_GET(zone, i);
2853 		ZDOM_LOCK_INIT(zone, zdom, (arg->flags & UMA_ZONE_MTXCLASS));
2854 		STAILQ_INIT(&zdom->uzd_buckets);
2855 	}
2856 
2857 #if defined(INVARIANTS) && !defined(KASAN) && !defined(KMSAN)
2858 	if (arg->uminit == trash_init && arg->fini == trash_fini)
2859 		zone->uz_flags |= UMA_ZFLAG_TRASH | UMA_ZFLAG_CTORDTOR;
2860 #elif defined(KASAN)
2861 	if ((arg->flags & (UMA_ZONE_NOFREE | UMA_ZFLAG_CACHE)) != 0)
2862 		arg->flags |= UMA_ZONE_NOKASAN;
2863 #endif
2864 
2865 	/*
2866 	 * This is a pure cache zone, no kegs.
2867 	 */
2868 	if (arg->import) {
2869 		KASSERT((arg->flags & UMA_ZFLAG_CACHE) != 0,
2870 		    ("zone_ctor: Import specified for non-cache zone."));
2871 		zone->uz_flags = arg->flags;
2872 		zone->uz_size = arg->size;
2873 		zone->uz_import = arg->import;
2874 		zone->uz_release = arg->release;
2875 		zone->uz_arg = arg->arg;
2876 #ifdef NUMA
2877 		/*
2878 		 * Cache zones are round-robin unless a policy is
2879 		 * specified because they may have incompatible
2880 		 * constraints.
2881 		 */
2882 		if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2883 			zone->uz_flags |= UMA_ZONE_ROUNDROBIN;
2884 #endif
2885 		rw_wlock(&uma_rwlock);
2886 		LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link);
2887 		rw_wunlock(&uma_rwlock);
2888 		goto out;
2889 	}
2890 
2891 	/*
2892 	 * Use the regular zone/keg/slab allocator.
2893 	 */
2894 	zone->uz_import = zone_import;
2895 	zone->uz_release = zone_release;
2896 	zone->uz_arg = zone;
2897 	keg = arg->keg;
2898 
2899 	if (arg->flags & UMA_ZONE_SECONDARY) {
2900 		KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
2901 		    ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
2902 		KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
2903 		zone->uz_init = arg->uminit;
2904 		zone->uz_fini = arg->fini;
2905 		zone->uz_flags |= UMA_ZONE_SECONDARY;
2906 		rw_wlock(&uma_rwlock);
2907 		ZONE_LOCK(zone);
2908 		LIST_FOREACH(z, &keg->uk_zones, uz_link) {
2909 			if (LIST_NEXT(z, uz_link) == NULL) {
2910 				LIST_INSERT_AFTER(z, zone, uz_link);
2911 				break;
2912 			}
2913 		}
2914 		ZONE_UNLOCK(zone);
2915 		rw_wunlock(&uma_rwlock);
2916 	} else if (keg == NULL) {
2917 		if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
2918 		    arg->align, arg->flags)) == NULL)
2919 			return (ENOMEM);
2920 	} else {
2921 		struct uma_kctor_args karg;
2922 		int error;
2923 
2924 		/* We should only be here from uma_startup() */
2925 		karg.size = arg->size;
2926 		karg.uminit = arg->uminit;
2927 		karg.fini = arg->fini;
2928 		karg.align = arg->align;
2929 		karg.flags = (arg->flags & ~UMA_ZONE_SMR);
2930 		karg.zone = zone;
2931 		error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
2932 		    flags);
2933 		if (error)
2934 			return (error);
2935 	}
2936 
2937 	/* Inherit properties from the keg. */
2938 	zone->uz_keg = keg;
2939 	zone->uz_size = keg->uk_size;
2940 	zone->uz_flags |= (keg->uk_flags &
2941 	    (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
2942 
2943 out:
2944 	if (booted >= BOOT_PCPU) {
2945 		zone_alloc_counters(zone, NULL);
2946 		if (booted >= BOOT_RUNNING)
2947 			zone_alloc_sysctl(zone, NULL);
2948 	} else {
2949 		zone->uz_allocs = EARLY_COUNTER;
2950 		zone->uz_frees = EARLY_COUNTER;
2951 		zone->uz_fails = EARLY_COUNTER;
2952 	}
2953 
2954 	/* Caller requests a private SMR context. */
2955 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
2956 		zone->uz_smr = smr_create(zone->uz_name, 0, 0);
2957 
2958 	KASSERT((arg->flags & (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET)) !=
2959 	    (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET),
2960 	    ("Invalid zone flag combination"));
2961 	if (arg->flags & UMA_ZFLAG_INTERNAL)
2962 		zone->uz_bucket_size_max = zone->uz_bucket_size = 0;
2963 	if ((arg->flags & UMA_ZONE_MAXBUCKET) != 0)
2964 		zone->uz_bucket_size = BUCKET_MAX;
2965 	else if ((arg->flags & UMA_ZONE_NOBUCKET) != 0)
2966 		zone->uz_bucket_size = 0;
2967 	else
2968 		zone->uz_bucket_size = bucket_select(zone->uz_size);
2969 	zone->uz_bucket_size_min = zone->uz_bucket_size;
2970 	if (zone->uz_dtor != NULL || zone->uz_ctor != NULL)
2971 		zone->uz_flags |= UMA_ZFLAG_CTORDTOR;
2972 	zone_update_caches(zone);
2973 
2974 	return (0);
2975 }
2976 
2977 /*
2978  * Keg header dtor.  This frees all data, destroys locks, frees the hash
2979  * table and removes the keg from the global list.
2980  *
2981  * Arguments/Returns follow uma_dtor specifications
2982  *	udata  unused
2983  */
2984 static void
2985 keg_dtor(void *arg, int size, void *udata)
2986 {
2987 	uma_keg_t keg;
2988 	uint32_t free, pages;
2989 	int i;
2990 
2991 	keg = (uma_keg_t)arg;
2992 	free = pages = 0;
2993 	for (i = 0; i < vm_ndomains; i++) {
2994 		free += keg->uk_domain[i].ud_free_items;
2995 		pages += keg->uk_domain[i].ud_pages;
2996 		KEG_LOCK_FINI(keg, i);
2997 	}
2998 	if (pages != 0)
2999 		printf("Freed UMA keg (%s) was not empty (%u items). "
3000 		    " Lost %u pages of memory.\n",
3001 		    keg->uk_name ? keg->uk_name : "",
3002 		    pages / keg->uk_ppera * keg->uk_ipers - free, pages);
3003 
3004 	hash_free(&keg->uk_hash);
3005 }
3006 
3007 /*
3008  * Zone header dtor.
3009  *
3010  * Arguments/Returns follow uma_dtor specifications
3011  *	udata  unused
3012  */
3013 static void
3014 zone_dtor(void *arg, int size, void *udata)
3015 {
3016 	uma_zone_t zone;
3017 	uma_keg_t keg;
3018 	int i;
3019 
3020 	zone = (uma_zone_t)arg;
3021 
3022 	sysctl_remove_oid(zone->uz_oid, 1, 1);
3023 
3024 	if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
3025 		cache_drain(zone);
3026 
3027 	rw_wlock(&uma_rwlock);
3028 	LIST_REMOVE(zone, uz_link);
3029 	rw_wunlock(&uma_rwlock);
3030 	if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) {
3031 		keg = zone->uz_keg;
3032 		keg->uk_reserve = 0;
3033 	}
3034 	zone_reclaim(zone, UMA_ANYDOMAIN, M_WAITOK, true);
3035 
3036 	/*
3037 	 * We only destroy kegs from non secondary/non cache zones.
3038 	 */
3039 	if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) {
3040 		keg = zone->uz_keg;
3041 		rw_wlock(&uma_rwlock);
3042 		LIST_REMOVE(keg, uk_link);
3043 		rw_wunlock(&uma_rwlock);
3044 		zone_free_item(kegs, keg, NULL, SKIP_NONE);
3045 	}
3046 	counter_u64_free(zone->uz_allocs);
3047 	counter_u64_free(zone->uz_frees);
3048 	counter_u64_free(zone->uz_fails);
3049 	counter_u64_free(zone->uz_xdomain);
3050 	free(zone->uz_ctlname, M_UMA);
3051 	for (i = 0; i < vm_ndomains; i++)
3052 		ZDOM_LOCK_FINI(ZDOM_GET(zone, i));
3053 	ZONE_CROSS_LOCK_FINI(zone);
3054 }
3055 
3056 static void
3057 zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *arg), void *arg)
3058 {
3059 	uma_keg_t keg;
3060 	uma_zone_t zone;
3061 
3062 	LIST_FOREACH(keg, &uma_kegs, uk_link) {
3063 		LIST_FOREACH(zone, &keg->uk_zones, uz_link)
3064 			zfunc(zone, arg);
3065 	}
3066 	LIST_FOREACH(zone, &uma_cachezones, uz_link)
3067 		zfunc(zone, arg);
3068 }
3069 
3070 /*
3071  * Traverses every zone in the system and calls a callback
3072  *
3073  * Arguments:
3074  *	zfunc  A pointer to a function which accepts a zone
3075  *		as an argument.
3076  *
3077  * Returns:
3078  *	Nothing
3079  */
3080 static void
3081 zone_foreach(void (*zfunc)(uma_zone_t, void *arg), void *arg)
3082 {
3083 
3084 	rw_rlock(&uma_rwlock);
3085 	zone_foreach_unlocked(zfunc, arg);
3086 	rw_runlock(&uma_rwlock);
3087 }
3088 
3089 /*
3090  * Initialize the kernel memory allocator.  This is done after pages can be
3091  * allocated but before general KVA is available.
3092  */
3093 void
3094 uma_startup1(vm_offset_t virtual_avail)
3095 {
3096 	struct uma_zctor_args args;
3097 	size_t ksize, zsize, size;
3098 	uma_keg_t primarykeg;
3099 	uintptr_t m;
3100 	int domain;
3101 	uint8_t pflag;
3102 
3103 	bootstart = bootmem = virtual_avail;
3104 
3105 	rw_init(&uma_rwlock, "UMA lock");
3106 	sx_init(&uma_reclaim_lock, "umareclaim");
3107 
3108 	ksize = sizeof(struct uma_keg) +
3109 	    (sizeof(struct uma_domain) * vm_ndomains);
3110 	ksize = roundup(ksize, UMA_SUPER_ALIGN);
3111 	zsize = sizeof(struct uma_zone) +
3112 	    (sizeof(struct uma_cache) * (mp_maxid + 1)) +
3113 	    (sizeof(struct uma_zone_domain) * vm_ndomains);
3114 	zsize = roundup(zsize, UMA_SUPER_ALIGN);
3115 
3116 	/* Allocate the zone of zones, zone of kegs, and zone of zones keg. */
3117 	size = (zsize * 2) + ksize;
3118 	for (domain = 0; domain < vm_ndomains; domain++) {
3119 		m = (uintptr_t)startup_alloc(NULL, size, domain, &pflag,
3120 		    M_NOWAIT | M_ZERO);
3121 		if (m != 0)
3122 			break;
3123 	}
3124 	zones = (uma_zone_t)m;
3125 	m += zsize;
3126 	kegs = (uma_zone_t)m;
3127 	m += zsize;
3128 	primarykeg = (uma_keg_t)m;
3129 
3130 	/* "manually" create the initial zone */
3131 	memset(&args, 0, sizeof(args));
3132 	args.name = "UMA Kegs";
3133 	args.size = ksize;
3134 	args.ctor = keg_ctor;
3135 	args.dtor = keg_dtor;
3136 	args.uminit = zero_init;
3137 	args.fini = NULL;
3138 	args.keg = primarykeg;
3139 	args.align = UMA_SUPER_ALIGN - 1;
3140 	args.flags = UMA_ZFLAG_INTERNAL;
3141 	zone_ctor(kegs, zsize, &args, M_WAITOK);
3142 
3143 	args.name = "UMA Zones";
3144 	args.size = zsize;
3145 	args.ctor = zone_ctor;
3146 	args.dtor = zone_dtor;
3147 	args.uminit = zero_init;
3148 	args.fini = NULL;
3149 	args.keg = NULL;
3150 	args.align = UMA_SUPER_ALIGN - 1;
3151 	args.flags = UMA_ZFLAG_INTERNAL;
3152 	zone_ctor(zones, zsize, &args, M_WAITOK);
3153 
3154 	/* Now make zones for slab headers */
3155 	slabzones[0] = uma_zcreate("UMA Slabs 0", SLABZONE0_SIZE,
3156 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
3157 	slabzones[1] = uma_zcreate("UMA Slabs 1", SLABZONE1_SIZE,
3158 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
3159 
3160 	hashzone = uma_zcreate("UMA Hash",
3161 	    sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
3162 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
3163 
3164 	bucket_init();
3165 	smr_init();
3166 }
3167 
3168 #ifndef UMA_MD_SMALL_ALLOC
3169 extern void vm_radix_reserve_kva(void);
3170 #endif
3171 
3172 /*
3173  * Advertise the availability of normal kva allocations and switch to
3174  * the default back-end allocator.  Marks the KVA we consumed on startup
3175  * as used in the map.
3176  */
3177 void
3178 uma_startup2(void)
3179 {
3180 
3181 	if (bootstart != bootmem) {
3182 		vm_map_lock(kernel_map);
3183 		(void)vm_map_insert(kernel_map, NULL, 0, bootstart, bootmem,
3184 		    VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT);
3185 		vm_map_unlock(kernel_map);
3186 	}
3187 
3188 #ifndef UMA_MD_SMALL_ALLOC
3189 	/* Set up radix zone to use noobj_alloc. */
3190 	vm_radix_reserve_kva();
3191 #endif
3192 
3193 	booted = BOOT_KVA;
3194 	zone_foreach_unlocked(zone_kva_available, NULL);
3195 	bucket_enable();
3196 }
3197 
3198 /*
3199  * Allocate counters as early as possible so that boot-time allocations are
3200  * accounted more precisely.
3201  */
3202 static void
3203 uma_startup_pcpu(void *arg __unused)
3204 {
3205 
3206 	zone_foreach_unlocked(zone_alloc_counters, NULL);
3207 	booted = BOOT_PCPU;
3208 }
3209 SYSINIT(uma_startup_pcpu, SI_SUB_COUNTER, SI_ORDER_ANY, uma_startup_pcpu, NULL);
3210 
3211 /*
3212  * Finish our initialization steps.
3213  */
3214 static void
3215 uma_startup3(void *arg __unused)
3216 {
3217 
3218 #ifdef INVARIANTS
3219 	TUNABLE_INT_FETCH("vm.debug.divisor", &dbg_divisor);
3220 	uma_dbg_cnt = counter_u64_alloc(M_WAITOK);
3221 	uma_skip_cnt = counter_u64_alloc(M_WAITOK);
3222 #endif
3223 	zone_foreach_unlocked(zone_alloc_sysctl, NULL);
3224 	callout_init(&uma_callout, 1);
3225 	callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
3226 	booted = BOOT_RUNNING;
3227 
3228 	EVENTHANDLER_REGISTER(shutdown_post_sync, uma_shutdown, NULL,
3229 	    EVENTHANDLER_PRI_FIRST);
3230 }
3231 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
3232 
3233 static void
3234 uma_shutdown(void)
3235 {
3236 
3237 	booted = BOOT_SHUTDOWN;
3238 }
3239 
3240 static uma_keg_t
3241 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
3242 		int align, uint32_t flags)
3243 {
3244 	struct uma_kctor_args args;
3245 
3246 	args.size = size;
3247 	args.uminit = uminit;
3248 	args.fini = fini;
3249 	args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
3250 	args.flags = flags;
3251 	args.zone = zone;
3252 	return (zone_alloc_item(kegs, &args, UMA_ANYDOMAIN, M_WAITOK));
3253 }
3254 
3255 /* Public functions */
3256 /* See uma.h */
3257 void
3258 uma_set_align(int align)
3259 {
3260 
3261 	if (align != UMA_ALIGN_CACHE)
3262 		uma_align_cache = align;
3263 }
3264 
3265 /* See uma.h */
3266 uma_zone_t
3267 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
3268 		uma_init uminit, uma_fini fini, int align, uint32_t flags)
3269 
3270 {
3271 	struct uma_zctor_args args;
3272 	uma_zone_t res;
3273 
3274 	KASSERT(powerof2(align + 1), ("invalid zone alignment %d for \"%s\"",
3275 	    align, name));
3276 
3277 	/* This stuff is essential for the zone ctor */
3278 	memset(&args, 0, sizeof(args));
3279 	args.name = name;
3280 	args.size = size;
3281 	args.ctor = ctor;
3282 	args.dtor = dtor;
3283 	args.uminit = uminit;
3284 	args.fini = fini;
3285 #if defined(INVARIANTS) && !defined(KASAN) && !defined(KMSAN)
3286 	/*
3287 	 * Inject procedures which check for memory use after free if we are
3288 	 * allowed to scramble the memory while it is not allocated.  This
3289 	 * requires that: UMA is actually able to access the memory, no init
3290 	 * or fini procedures, no dependency on the initial value of the
3291 	 * memory, and no (legitimate) use of the memory after free.  Note,
3292 	 * the ctor and dtor do not need to be empty.
3293 	 */
3294 	if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOTOUCH |
3295 	    UMA_ZONE_NOFREE))) && uminit == NULL && fini == NULL) {
3296 		args.uminit = trash_init;
3297 		args.fini = trash_fini;
3298 	}
3299 #endif
3300 	args.align = align;
3301 	args.flags = flags;
3302 	args.keg = NULL;
3303 
3304 	sx_xlock(&uma_reclaim_lock);
3305 	res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3306 	sx_xunlock(&uma_reclaim_lock);
3307 
3308 	return (res);
3309 }
3310 
3311 /* See uma.h */
3312 uma_zone_t
3313 uma_zsecond_create(const char *name, uma_ctor ctor, uma_dtor dtor,
3314     uma_init zinit, uma_fini zfini, uma_zone_t primary)
3315 {
3316 	struct uma_zctor_args args;
3317 	uma_keg_t keg;
3318 	uma_zone_t res;
3319 
3320 	keg = primary->uz_keg;
3321 	memset(&args, 0, sizeof(args));
3322 	args.name = name;
3323 	args.size = keg->uk_size;
3324 	args.ctor = ctor;
3325 	args.dtor = dtor;
3326 	args.uminit = zinit;
3327 	args.fini = zfini;
3328 	args.align = keg->uk_align;
3329 	args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
3330 	args.keg = keg;
3331 
3332 	sx_xlock(&uma_reclaim_lock);
3333 	res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3334 	sx_xunlock(&uma_reclaim_lock);
3335 
3336 	return (res);
3337 }
3338 
3339 /* See uma.h */
3340 uma_zone_t
3341 uma_zcache_create(const char *name, int size, uma_ctor ctor, uma_dtor dtor,
3342     uma_init zinit, uma_fini zfini, uma_import zimport, uma_release zrelease,
3343     void *arg, int flags)
3344 {
3345 	struct uma_zctor_args args;
3346 
3347 	memset(&args, 0, sizeof(args));
3348 	args.name = name;
3349 	args.size = size;
3350 	args.ctor = ctor;
3351 	args.dtor = dtor;
3352 	args.uminit = zinit;
3353 	args.fini = zfini;
3354 	args.import = zimport;
3355 	args.release = zrelease;
3356 	args.arg = arg;
3357 	args.align = 0;
3358 	args.flags = flags | UMA_ZFLAG_CACHE;
3359 
3360 	return (zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK));
3361 }
3362 
3363 /* See uma.h */
3364 void
3365 uma_zdestroy(uma_zone_t zone)
3366 {
3367 
3368 	/*
3369 	 * Large slabs are expensive to reclaim, so don't bother doing
3370 	 * unnecessary work if we're shutting down.
3371 	 */
3372 	if (booted == BOOT_SHUTDOWN &&
3373 	    zone->uz_fini == NULL && zone->uz_release == zone_release)
3374 		return;
3375 	sx_xlock(&uma_reclaim_lock);
3376 	zone_free_item(zones, zone, NULL, SKIP_NONE);
3377 	sx_xunlock(&uma_reclaim_lock);
3378 }
3379 
3380 void
3381 uma_zwait(uma_zone_t zone)
3382 {
3383 
3384 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
3385 		uma_zfree_smr(zone, uma_zalloc_smr(zone, M_WAITOK));
3386 	else if ((zone->uz_flags & UMA_ZONE_PCPU) != 0)
3387 		uma_zfree_pcpu(zone, uma_zalloc_pcpu(zone, M_WAITOK));
3388 	else
3389 		uma_zfree(zone, uma_zalloc(zone, M_WAITOK));
3390 }
3391 
3392 void *
3393 uma_zalloc_pcpu_arg(uma_zone_t zone, void *udata, int flags)
3394 {
3395 	void *item, *pcpu_item;
3396 #ifdef SMP
3397 	int i;
3398 
3399 	MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3400 #endif
3401 	item = uma_zalloc_arg(zone, udata, flags & ~M_ZERO);
3402 	if (item == NULL)
3403 		return (NULL);
3404 	pcpu_item = zpcpu_base_to_offset(item);
3405 	if (flags & M_ZERO) {
3406 #ifdef SMP
3407 		for (i = 0; i <= mp_maxid; i++)
3408 			bzero(zpcpu_get_cpu(pcpu_item, i), zone->uz_size);
3409 #else
3410 		bzero(item, zone->uz_size);
3411 #endif
3412 	}
3413 	return (pcpu_item);
3414 }
3415 
3416 /*
3417  * A stub while both regular and pcpu cases are identical.
3418  */
3419 void
3420 uma_zfree_pcpu_arg(uma_zone_t zone, void *pcpu_item, void *udata)
3421 {
3422 	void *item;
3423 
3424 #ifdef SMP
3425 	MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3426 #endif
3427 
3428         /* uma_zfree_pcu_*(..., NULL) does nothing, to match free(9). */
3429         if (pcpu_item == NULL)
3430                 return;
3431 
3432 	item = zpcpu_offset_to_base(pcpu_item);
3433 	uma_zfree_arg(zone, item, udata);
3434 }
3435 
3436 static inline void *
3437 item_ctor(uma_zone_t zone, int uz_flags, int size, void *udata, int flags,
3438     void *item)
3439 {
3440 #ifdef INVARIANTS
3441 	bool skipdbg;
3442 #endif
3443 
3444 	kasan_mark_item_valid(zone, item);
3445 	kmsan_mark_item_uninitialized(zone, item);
3446 
3447 #ifdef INVARIANTS
3448 	skipdbg = uma_dbg_zskip(zone, item);
3449 	if (!skipdbg && (uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3450 	    zone->uz_ctor != trash_ctor)
3451 		trash_ctor(item, size, udata, flags);
3452 #endif
3453 
3454 	/* Check flags before loading ctor pointer. */
3455 	if (__predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0) &&
3456 	    __predict_false(zone->uz_ctor != NULL) &&
3457 	    zone->uz_ctor(item, size, udata, flags) != 0) {
3458 		counter_u64_add(zone->uz_fails, 1);
3459 		zone_free_item(zone, item, udata, SKIP_DTOR | SKIP_CNT);
3460 		return (NULL);
3461 	}
3462 #ifdef INVARIANTS
3463 	if (!skipdbg)
3464 		uma_dbg_alloc(zone, NULL, item);
3465 #endif
3466 	if (__predict_false(flags & M_ZERO))
3467 		return (memset(item, 0, size));
3468 
3469 	return (item);
3470 }
3471 
3472 static inline void
3473 item_dtor(uma_zone_t zone, void *item, int size, void *udata,
3474     enum zfreeskip skip)
3475 {
3476 #ifdef INVARIANTS
3477 	bool skipdbg;
3478 
3479 	skipdbg = uma_dbg_zskip(zone, item);
3480 	if (skip == SKIP_NONE && !skipdbg) {
3481 		if ((zone->uz_flags & UMA_ZONE_MALLOC) != 0)
3482 			uma_dbg_free(zone, udata, item);
3483 		else
3484 			uma_dbg_free(zone, NULL, item);
3485 	}
3486 #endif
3487 	if (__predict_true(skip < SKIP_DTOR)) {
3488 		if (zone->uz_dtor != NULL)
3489 			zone->uz_dtor(item, size, udata);
3490 #ifdef INVARIANTS
3491 		if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3492 		    zone->uz_dtor != trash_dtor)
3493 			trash_dtor(item, size, udata);
3494 #endif
3495 	}
3496 	kasan_mark_item_invalid(zone, item);
3497 }
3498 
3499 #ifdef NUMA
3500 static int
3501 item_domain(void *item)
3502 {
3503 	int domain;
3504 
3505 	domain = vm_phys_domain(vtophys(item));
3506 	KASSERT(domain >= 0 && domain < vm_ndomains,
3507 	    ("%s: unknown domain for item %p", __func__, item));
3508 	return (domain);
3509 }
3510 #endif
3511 
3512 #if defined(INVARIANTS) || defined(DEBUG_MEMGUARD) || defined(WITNESS)
3513 #define	UMA_ZALLOC_DEBUG
3514 static int
3515 uma_zalloc_debug(uma_zone_t zone, void **itemp, void *udata, int flags)
3516 {
3517 	int error;
3518 
3519 	error = 0;
3520 #ifdef WITNESS
3521 	if (flags & M_WAITOK) {
3522 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
3523 		    "uma_zalloc_debug: zone \"%s\"", zone->uz_name);
3524 	}
3525 #endif
3526 
3527 #ifdef INVARIANTS
3528 	KASSERT((flags & M_EXEC) == 0,
3529 	    ("uma_zalloc_debug: called with M_EXEC"));
3530 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3531 	    ("uma_zalloc_debug: called within spinlock or critical section"));
3532 	KASSERT((zone->uz_flags & UMA_ZONE_PCPU) == 0 || (flags & M_ZERO) == 0,
3533 	    ("uma_zalloc_debug: allocating from a pcpu zone with M_ZERO"));
3534 #endif
3535 
3536 #ifdef DEBUG_MEMGUARD
3537 	if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && memguard_cmp_zone(zone)) {
3538 		void *item;
3539 		item = memguard_alloc(zone->uz_size, flags);
3540 		if (item != NULL) {
3541 			error = EJUSTRETURN;
3542 			if (zone->uz_init != NULL &&
3543 			    zone->uz_init(item, zone->uz_size, flags) != 0) {
3544 				*itemp = NULL;
3545 				return (error);
3546 			}
3547 			if (zone->uz_ctor != NULL &&
3548 			    zone->uz_ctor(item, zone->uz_size, udata,
3549 			    flags) != 0) {
3550 				counter_u64_add(zone->uz_fails, 1);
3551 			    	zone->uz_fini(item, zone->uz_size);
3552 				*itemp = NULL;
3553 				return (error);
3554 			}
3555 			*itemp = item;
3556 			return (error);
3557 		}
3558 		/* This is unfortunate but should not be fatal. */
3559 	}
3560 #endif
3561 	return (error);
3562 }
3563 
3564 static int
3565 uma_zfree_debug(uma_zone_t zone, void *item, void *udata)
3566 {
3567 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3568 	    ("uma_zfree_debug: called with spinlock or critical section held"));
3569 
3570 #ifdef DEBUG_MEMGUARD
3571 	if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && is_memguard_addr(item)) {
3572 		if (zone->uz_dtor != NULL)
3573 			zone->uz_dtor(item, zone->uz_size, udata);
3574 		if (zone->uz_fini != NULL)
3575 			zone->uz_fini(item, zone->uz_size);
3576 		memguard_free(item);
3577 		return (EJUSTRETURN);
3578 	}
3579 #endif
3580 	return (0);
3581 }
3582 #endif
3583 
3584 static inline void *
3585 cache_alloc_item(uma_zone_t zone, uma_cache_t cache, uma_cache_bucket_t bucket,
3586     void *udata, int flags)
3587 {
3588 	void *item;
3589 	int size, uz_flags;
3590 
3591 	item = cache_bucket_pop(cache, bucket);
3592 	size = cache_uz_size(cache);
3593 	uz_flags = cache_uz_flags(cache);
3594 	critical_exit();
3595 	return (item_ctor(zone, uz_flags, size, udata, flags, item));
3596 }
3597 
3598 static __noinline void *
3599 cache_alloc_retry(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3600 {
3601 	uma_cache_bucket_t bucket;
3602 	int domain;
3603 
3604 	while (cache_alloc(zone, cache, udata, flags)) {
3605 		cache = &zone->uz_cpu[curcpu];
3606 		bucket = &cache->uc_allocbucket;
3607 		if (__predict_false(bucket->ucb_cnt == 0))
3608 			continue;
3609 		return (cache_alloc_item(zone, cache, bucket, udata, flags));
3610 	}
3611 	critical_exit();
3612 
3613 	/*
3614 	 * We can not get a bucket so try to return a single item.
3615 	 */
3616 	if (zone->uz_flags & UMA_ZONE_FIRSTTOUCH)
3617 		domain = PCPU_GET(domain);
3618 	else
3619 		domain = UMA_ANYDOMAIN;
3620 	return (zone_alloc_item(zone, udata, domain, flags));
3621 }
3622 
3623 /* See uma.h */
3624 void *
3625 uma_zalloc_smr(uma_zone_t zone, int flags)
3626 {
3627 	uma_cache_bucket_t bucket;
3628 	uma_cache_t cache;
3629 
3630 	CTR3(KTR_UMA, "uma_zalloc_smr zone %s(%p) flags %d", zone->uz_name,
3631 	    zone, flags);
3632 
3633 #ifdef UMA_ZALLOC_DEBUG
3634 	void *item;
3635 
3636 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
3637 	    ("uma_zalloc_arg: called with non-SMR zone."));
3638 	if (uma_zalloc_debug(zone, &item, NULL, flags) == EJUSTRETURN)
3639 		return (item);
3640 #endif
3641 
3642 	critical_enter();
3643 	cache = &zone->uz_cpu[curcpu];
3644 	bucket = &cache->uc_allocbucket;
3645 	if (__predict_false(bucket->ucb_cnt == 0))
3646 		return (cache_alloc_retry(zone, cache, NULL, flags));
3647 	return (cache_alloc_item(zone, cache, bucket, NULL, flags));
3648 }
3649 
3650 /* See uma.h */
3651 void *
3652 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
3653 {
3654 	uma_cache_bucket_t bucket;
3655 	uma_cache_t cache;
3656 
3657 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3658 	random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3659 
3660 	/* This is the fast path allocation */
3661 	CTR3(KTR_UMA, "uma_zalloc_arg zone %s(%p) flags %d", zone->uz_name,
3662 	    zone, flags);
3663 
3664 #ifdef UMA_ZALLOC_DEBUG
3665 	void *item;
3666 
3667 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3668 	    ("uma_zalloc_arg: called with SMR zone."));
3669 	if (uma_zalloc_debug(zone, &item, udata, flags) == EJUSTRETURN)
3670 		return (item);
3671 #endif
3672 
3673 	/*
3674 	 * If possible, allocate from the per-CPU cache.  There are two
3675 	 * requirements for safe access to the per-CPU cache: (1) the thread
3676 	 * accessing the cache must not be preempted or yield during access,
3677 	 * and (2) the thread must not migrate CPUs without switching which
3678 	 * cache it accesses.  We rely on a critical section to prevent
3679 	 * preemption and migration.  We release the critical section in
3680 	 * order to acquire the zone mutex if we are unable to allocate from
3681 	 * the current cache; when we re-acquire the critical section, we
3682 	 * must detect and handle migration if it has occurred.
3683 	 */
3684 	critical_enter();
3685 	cache = &zone->uz_cpu[curcpu];
3686 	bucket = &cache->uc_allocbucket;
3687 	if (__predict_false(bucket->ucb_cnt == 0))
3688 		return (cache_alloc_retry(zone, cache, udata, flags));
3689 	return (cache_alloc_item(zone, cache, bucket, udata, flags));
3690 }
3691 
3692 /*
3693  * Replenish an alloc bucket and possibly restore an old one.  Called in
3694  * a critical section.  Returns in a critical section.
3695  *
3696  * A false return value indicates an allocation failure.
3697  * A true return value indicates success and the caller should retry.
3698  */
3699 static __noinline bool
3700 cache_alloc(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3701 {
3702 	uma_bucket_t bucket;
3703 	int curdomain, domain;
3704 	bool new;
3705 
3706 	CRITICAL_ASSERT(curthread);
3707 
3708 	/*
3709 	 * If we have run out of items in our alloc bucket see
3710 	 * if we can switch with the free bucket.
3711 	 *
3712 	 * SMR Zones can't re-use the free bucket until the sequence has
3713 	 * expired.
3714 	 */
3715 	if ((cache_uz_flags(cache) & UMA_ZONE_SMR) == 0 &&
3716 	    cache->uc_freebucket.ucb_cnt != 0) {
3717 		cache_bucket_swap(&cache->uc_freebucket,
3718 		    &cache->uc_allocbucket);
3719 		return (true);
3720 	}
3721 
3722 	/*
3723 	 * Discard any empty allocation bucket while we hold no locks.
3724 	 */
3725 	bucket = cache_bucket_unload_alloc(cache);
3726 	critical_exit();
3727 
3728 	if (bucket != NULL) {
3729 		KASSERT(bucket->ub_cnt == 0,
3730 		    ("cache_alloc: Entered with non-empty alloc bucket."));
3731 		bucket_free(zone, bucket, udata);
3732 	}
3733 
3734 	/*
3735 	 * Attempt to retrieve the item from the per-CPU cache has failed, so
3736 	 * we must go back to the zone.  This requires the zdom lock, so we
3737 	 * must drop the critical section, then re-acquire it when we go back
3738 	 * to the cache.  Since the critical section is released, we may be
3739 	 * preempted or migrate.  As such, make sure not to maintain any
3740 	 * thread-local state specific to the cache from prior to releasing
3741 	 * the critical section.
3742 	 */
3743 	domain = PCPU_GET(domain);
3744 	if ((cache_uz_flags(cache) & UMA_ZONE_ROUNDROBIN) != 0 ||
3745 	    VM_DOMAIN_EMPTY(domain))
3746 		domain = zone_domain_highest(zone, domain);
3747 	bucket = cache_fetch_bucket(zone, cache, domain);
3748 	if (bucket == NULL && zone->uz_bucket_size != 0 && !bucketdisable) {
3749 		bucket = zone_alloc_bucket(zone, udata, domain, flags);
3750 		new = true;
3751 	} else {
3752 		new = false;
3753 	}
3754 
3755 	CTR3(KTR_UMA, "uma_zalloc: zone %s(%p) bucket zone returned %p",
3756 	    zone->uz_name, zone, bucket);
3757 	if (bucket == NULL) {
3758 		critical_enter();
3759 		return (false);
3760 	}
3761 
3762 	/*
3763 	 * See if we lost the race or were migrated.  Cache the
3764 	 * initialized bucket to make this less likely or claim
3765 	 * the memory directly.
3766 	 */
3767 	critical_enter();
3768 	cache = &zone->uz_cpu[curcpu];
3769 	if (cache->uc_allocbucket.ucb_bucket == NULL &&
3770 	    ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) == 0 ||
3771 	    (curdomain = PCPU_GET(domain)) == domain ||
3772 	    VM_DOMAIN_EMPTY(curdomain))) {
3773 		if (new)
3774 			atomic_add_long(&ZDOM_GET(zone, domain)->uzd_imax,
3775 			    bucket->ub_cnt);
3776 		cache_bucket_load_alloc(cache, bucket);
3777 		return (true);
3778 	}
3779 
3780 	/*
3781 	 * We lost the race, release this bucket and start over.
3782 	 */
3783 	critical_exit();
3784 	zone_put_bucket(zone, domain, bucket, udata, !new);
3785 	critical_enter();
3786 
3787 	return (true);
3788 }
3789 
3790 void *
3791 uma_zalloc_domain(uma_zone_t zone, void *udata, int domain, int flags)
3792 {
3793 #ifdef NUMA
3794 	uma_bucket_t bucket;
3795 	uma_zone_domain_t zdom;
3796 	void *item;
3797 #endif
3798 
3799 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3800 	random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3801 
3802 	/* This is the fast path allocation */
3803 	CTR4(KTR_UMA, "uma_zalloc_domain zone %s(%p) domain %d flags %d",
3804 	    zone->uz_name, zone, domain, flags);
3805 
3806 	if (flags & M_WAITOK) {
3807 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
3808 		    "uma_zalloc_domain: zone \"%s\"", zone->uz_name);
3809 	}
3810 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3811 	    ("uma_zalloc_domain: called with spinlock or critical section held"));
3812 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3813 	    ("uma_zalloc_domain: called with SMR zone."));
3814 #ifdef NUMA
3815 	KASSERT((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0,
3816 	    ("uma_zalloc_domain: called with non-FIRSTTOUCH zone."));
3817 
3818 	if (vm_ndomains == 1)
3819 		return (uma_zalloc_arg(zone, udata, flags));
3820 
3821 	/*
3822 	 * Try to allocate from the bucket cache before falling back to the keg.
3823 	 * We could try harder and attempt to allocate from per-CPU caches or
3824 	 * the per-domain cross-domain buckets, but the complexity is probably
3825 	 * not worth it.  It is more important that frees of previous
3826 	 * cross-domain allocations do not blow up the cache.
3827 	 */
3828 	zdom = zone_domain_lock(zone, domain);
3829 	if ((bucket = zone_fetch_bucket(zone, zdom, false)) != NULL) {
3830 		item = bucket->ub_bucket[bucket->ub_cnt - 1];
3831 #ifdef INVARIANTS
3832 		bucket->ub_bucket[bucket->ub_cnt - 1] = NULL;
3833 #endif
3834 		bucket->ub_cnt--;
3835 		zone_put_bucket(zone, domain, bucket, udata, true);
3836 		item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata,
3837 		    flags, item);
3838 		if (item != NULL) {
3839 			KASSERT(item_domain(item) == domain,
3840 			    ("%s: bucket cache item %p from wrong domain",
3841 			    __func__, item));
3842 			counter_u64_add(zone->uz_allocs, 1);
3843 		}
3844 		return (item);
3845 	}
3846 	ZDOM_UNLOCK(zdom);
3847 	return (zone_alloc_item(zone, udata, domain, flags));
3848 #else
3849 	return (uma_zalloc_arg(zone, udata, flags));
3850 #endif
3851 }
3852 
3853 /*
3854  * Find a slab with some space.  Prefer slabs that are partially used over those
3855  * that are totally full.  This helps to reduce fragmentation.
3856  *
3857  * If 'rr' is 1, search all domains starting from 'domain'.  Otherwise check
3858  * only 'domain'.
3859  */
3860 static uma_slab_t
3861 keg_first_slab(uma_keg_t keg, int domain, bool rr)
3862 {
3863 	uma_domain_t dom;
3864 	uma_slab_t slab;
3865 	int start;
3866 
3867 	KASSERT(domain >= 0 && domain < vm_ndomains,
3868 	    ("keg_first_slab: domain %d out of range", domain));
3869 	KEG_LOCK_ASSERT(keg, domain);
3870 
3871 	slab = NULL;
3872 	start = domain;
3873 	do {
3874 		dom = &keg->uk_domain[domain];
3875 		if ((slab = LIST_FIRST(&dom->ud_part_slab)) != NULL)
3876 			return (slab);
3877 		if ((slab = LIST_FIRST(&dom->ud_free_slab)) != NULL) {
3878 			LIST_REMOVE(slab, us_link);
3879 			dom->ud_free_slabs--;
3880 			LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
3881 			return (slab);
3882 		}
3883 		if (rr)
3884 			domain = (domain + 1) % vm_ndomains;
3885 	} while (domain != start);
3886 
3887 	return (NULL);
3888 }
3889 
3890 /*
3891  * Fetch an existing slab from a free or partial list.  Returns with the
3892  * keg domain lock held if a slab was found or unlocked if not.
3893  */
3894 static uma_slab_t
3895 keg_fetch_free_slab(uma_keg_t keg, int domain, bool rr, int flags)
3896 {
3897 	uma_slab_t slab;
3898 	uint32_t reserve;
3899 
3900 	/* HASH has a single free list. */
3901 	if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
3902 		domain = 0;
3903 
3904 	KEG_LOCK(keg, domain);
3905 	reserve = (flags & M_USE_RESERVE) != 0 ? 0 : keg->uk_reserve;
3906 	if (keg->uk_domain[domain].ud_free_items <= reserve ||
3907 	    (slab = keg_first_slab(keg, domain, rr)) == NULL) {
3908 		KEG_UNLOCK(keg, domain);
3909 		return (NULL);
3910 	}
3911 	return (slab);
3912 }
3913 
3914 static uma_slab_t
3915 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int rdomain, const int flags)
3916 {
3917 	struct vm_domainset_iter di;
3918 	uma_slab_t slab;
3919 	int aflags, domain;
3920 	bool rr;
3921 
3922 	KASSERT((flags & (M_WAITOK | M_NOVM)) != (M_WAITOK | M_NOVM),
3923 	    ("%s: invalid flags %#x", __func__, flags));
3924 
3925 restart:
3926 	/*
3927 	 * Use the keg's policy if upper layers haven't already specified a
3928 	 * domain (as happens with first-touch zones).
3929 	 *
3930 	 * To avoid races we run the iterator with the keg lock held, but that
3931 	 * means that we cannot allow the vm_domainset layer to sleep.  Thus,
3932 	 * clear M_WAITOK and handle low memory conditions locally.
3933 	 */
3934 	rr = rdomain == UMA_ANYDOMAIN;
3935 	if (rr) {
3936 		aflags = (flags & ~M_WAITOK) | M_NOWAIT;
3937 		vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
3938 		    &aflags);
3939 	} else {
3940 		aflags = flags;
3941 		domain = rdomain;
3942 	}
3943 
3944 	for (;;) {
3945 		slab = keg_fetch_free_slab(keg, domain, rr, flags);
3946 		if (slab != NULL)
3947 			return (slab);
3948 
3949 		/*
3950 		 * M_NOVM is used to break the recursion that can otherwise
3951 		 * occur if low-level memory management routines use UMA.
3952 		 */
3953 		if ((flags & M_NOVM) == 0) {
3954 			slab = keg_alloc_slab(keg, zone, domain, flags, aflags);
3955 			if (slab != NULL)
3956 				return (slab);
3957 		}
3958 
3959 		if (!rr) {
3960 			if ((flags & M_USE_RESERVE) != 0) {
3961 				/*
3962 				 * Drain reserves from other domains before
3963 				 * giving up or sleeping.  It may be useful to
3964 				 * support per-domain reserves eventually.
3965 				 */
3966 				rdomain = UMA_ANYDOMAIN;
3967 				goto restart;
3968 			}
3969 			if ((flags & M_WAITOK) == 0)
3970 				break;
3971 			vm_wait_domain(domain);
3972 		} else if (vm_domainset_iter_policy(&di, &domain) != 0) {
3973 			if ((flags & M_WAITOK) != 0) {
3974 				vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask, 0);
3975 				goto restart;
3976 			}
3977 			break;
3978 		}
3979 	}
3980 
3981 	/*
3982 	 * We might not have been able to get a slab but another cpu
3983 	 * could have while we were unlocked.  Check again before we
3984 	 * fail.
3985 	 */
3986 	if ((slab = keg_fetch_free_slab(keg, domain, rr, flags)) != NULL)
3987 		return (slab);
3988 
3989 	return (NULL);
3990 }
3991 
3992 static void *
3993 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
3994 {
3995 	uma_domain_t dom;
3996 	void *item;
3997 	int freei;
3998 
3999 	KEG_LOCK_ASSERT(keg, slab->us_domain);
4000 
4001 	dom = &keg->uk_domain[slab->us_domain];
4002 	freei = BIT_FFS(keg->uk_ipers, &slab->us_free) - 1;
4003 	BIT_CLR(keg->uk_ipers, freei, &slab->us_free);
4004 	item = slab_item(slab, keg, freei);
4005 	slab->us_freecount--;
4006 	dom->ud_free_items--;
4007 
4008 	/*
4009 	 * Move this slab to the full list.  It must be on the partial list, so
4010 	 * we do not need to update the free slab count.  In particular,
4011 	 * keg_fetch_slab() always returns slabs on the partial list.
4012 	 */
4013 	if (slab->us_freecount == 0) {
4014 		LIST_REMOVE(slab, us_link);
4015 		LIST_INSERT_HEAD(&dom->ud_full_slab, slab, us_link);
4016 	}
4017 
4018 	return (item);
4019 }
4020 
4021 static int
4022 zone_import(void *arg, void **bucket, int max, int domain, int flags)
4023 {
4024 	uma_domain_t dom;
4025 	uma_zone_t zone;
4026 	uma_slab_t slab;
4027 	uma_keg_t keg;
4028 #ifdef NUMA
4029 	int stripe;
4030 #endif
4031 	int i;
4032 
4033 	zone = arg;
4034 	slab = NULL;
4035 	keg = zone->uz_keg;
4036 	/* Try to keep the buckets totally full */
4037 	for (i = 0; i < max; ) {
4038 		if ((slab = keg_fetch_slab(keg, zone, domain, flags)) == NULL)
4039 			break;
4040 #ifdef NUMA
4041 		stripe = howmany(max, vm_ndomains);
4042 #endif
4043 		dom = &keg->uk_domain[slab->us_domain];
4044 		do {
4045 			bucket[i++] = slab_alloc_item(keg, slab);
4046 			if (keg->uk_reserve > 0 &&
4047 			    dom->ud_free_items <= keg->uk_reserve) {
4048 				/*
4049 				 * Avoid depleting the reserve after a
4050 				 * successful item allocation, even if
4051 				 * M_USE_RESERVE is specified.
4052 				 */
4053 				KEG_UNLOCK(keg, slab->us_domain);
4054 				goto out;
4055 			}
4056 #ifdef NUMA
4057 			/*
4058 			 * If the zone is striped we pick a new slab for every
4059 			 * N allocations.  Eliminating this conditional will
4060 			 * instead pick a new domain for each bucket rather
4061 			 * than stripe within each bucket.  The current option
4062 			 * produces more fragmentation and requires more cpu
4063 			 * time but yields better distribution.
4064 			 */
4065 			if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0 &&
4066 			    vm_ndomains > 1 && --stripe == 0)
4067 				break;
4068 #endif
4069 		} while (slab->us_freecount != 0 && i < max);
4070 		KEG_UNLOCK(keg, slab->us_domain);
4071 
4072 		/* Don't block if we allocated any successfully. */
4073 		flags &= ~M_WAITOK;
4074 		flags |= M_NOWAIT;
4075 	}
4076 out:
4077 	return i;
4078 }
4079 
4080 static int
4081 zone_alloc_limit_hard(uma_zone_t zone, int count, int flags)
4082 {
4083 	uint64_t old, new, total, max;
4084 
4085 	/*
4086 	 * The hard case.  We're going to sleep because there were existing
4087 	 * sleepers or because we ran out of items.  This routine enforces
4088 	 * fairness by keeping fifo order.
4089 	 *
4090 	 * First release our ill gotten gains and make some noise.
4091 	 */
4092 	for (;;) {
4093 		zone_free_limit(zone, count);
4094 		zone_log_warning(zone);
4095 		zone_maxaction(zone);
4096 		if (flags & M_NOWAIT)
4097 			return (0);
4098 
4099 		/*
4100 		 * We need to allocate an item or set ourself as a sleeper
4101 		 * while the sleepq lock is held to avoid wakeup races.  This
4102 		 * is essentially a home rolled semaphore.
4103 		 */
4104 		sleepq_lock(&zone->uz_max_items);
4105 		old = zone->uz_items;
4106 		do {
4107 			MPASS(UZ_ITEMS_SLEEPERS(old) < UZ_ITEMS_SLEEPERS_MAX);
4108 			/* Cache the max since we will evaluate twice. */
4109 			max = zone->uz_max_items;
4110 			if (UZ_ITEMS_SLEEPERS(old) != 0 ||
4111 			    UZ_ITEMS_COUNT(old) >= max)
4112 				new = old + UZ_ITEMS_SLEEPER;
4113 			else
4114 				new = old + MIN(count, max - old);
4115 		} while (atomic_fcmpset_64(&zone->uz_items, &old, new) == 0);
4116 
4117 		/* We may have successfully allocated under the sleepq lock. */
4118 		if (UZ_ITEMS_SLEEPERS(new) == 0) {
4119 			sleepq_release(&zone->uz_max_items);
4120 			return (new - old);
4121 		}
4122 
4123 		/*
4124 		 * This is in a different cacheline from uz_items so that we
4125 		 * don't constantly invalidate the fastpath cacheline when we
4126 		 * adjust item counts.  This could be limited to toggling on
4127 		 * transitions.
4128 		 */
4129 		atomic_add_32(&zone->uz_sleepers, 1);
4130 		atomic_add_64(&zone->uz_sleeps, 1);
4131 
4132 		/*
4133 		 * We have added ourselves as a sleeper.  The sleepq lock
4134 		 * protects us from wakeup races.  Sleep now and then retry.
4135 		 */
4136 		sleepq_add(&zone->uz_max_items, NULL, "zonelimit", 0, 0);
4137 		sleepq_wait(&zone->uz_max_items, PVM);
4138 
4139 		/*
4140 		 * After wakeup, remove ourselves as a sleeper and try
4141 		 * again.  We no longer have the sleepq lock for protection.
4142 		 *
4143 		 * Subract ourselves as a sleeper while attempting to add
4144 		 * our count.
4145 		 */
4146 		atomic_subtract_32(&zone->uz_sleepers, 1);
4147 		old = atomic_fetchadd_64(&zone->uz_items,
4148 		    -(UZ_ITEMS_SLEEPER - count));
4149 		/* We're no longer a sleeper. */
4150 		old -= UZ_ITEMS_SLEEPER;
4151 
4152 		/*
4153 		 * If we're still at the limit, restart.  Notably do not
4154 		 * block on other sleepers.  Cache the max value to protect
4155 		 * against changes via sysctl.
4156 		 */
4157 		total = UZ_ITEMS_COUNT(old);
4158 		max = zone->uz_max_items;
4159 		if (total >= max)
4160 			continue;
4161 		/* Truncate if necessary, otherwise wake other sleepers. */
4162 		if (total + count > max) {
4163 			zone_free_limit(zone, total + count - max);
4164 			count = max - total;
4165 		} else if (total + count < max && UZ_ITEMS_SLEEPERS(old) != 0)
4166 			wakeup_one(&zone->uz_max_items);
4167 
4168 		return (count);
4169 	}
4170 }
4171 
4172 /*
4173  * Allocate 'count' items from our max_items limit.  Returns the number
4174  * available.  If M_NOWAIT is not specified it will sleep until at least
4175  * one item can be allocated.
4176  */
4177 static int
4178 zone_alloc_limit(uma_zone_t zone, int count, int flags)
4179 {
4180 	uint64_t old;
4181 	uint64_t max;
4182 
4183 	max = zone->uz_max_items;
4184 	MPASS(max > 0);
4185 
4186 	/*
4187 	 * We expect normal allocations to succeed with a simple
4188 	 * fetchadd.
4189 	 */
4190 	old = atomic_fetchadd_64(&zone->uz_items, count);
4191 	if (__predict_true(old + count <= max))
4192 		return (count);
4193 
4194 	/*
4195 	 * If we had some items and no sleepers just return the
4196 	 * truncated value.  We have to release the excess space
4197 	 * though because that may wake sleepers who weren't woken
4198 	 * because we were temporarily over the limit.
4199 	 */
4200 	if (old < max) {
4201 		zone_free_limit(zone, (old + count) - max);
4202 		return (max - old);
4203 	}
4204 	return (zone_alloc_limit_hard(zone, count, flags));
4205 }
4206 
4207 /*
4208  * Free a number of items back to the limit.
4209  */
4210 static void
4211 zone_free_limit(uma_zone_t zone, int count)
4212 {
4213 	uint64_t old;
4214 
4215 	MPASS(count > 0);
4216 
4217 	/*
4218 	 * In the common case we either have no sleepers or
4219 	 * are still over the limit and can just return.
4220 	 */
4221 	old = atomic_fetchadd_64(&zone->uz_items, -count);
4222 	if (__predict_true(UZ_ITEMS_SLEEPERS(old) == 0 ||
4223 	   UZ_ITEMS_COUNT(old) - count >= zone->uz_max_items))
4224 		return;
4225 
4226 	/*
4227 	 * Moderate the rate of wakeups.  Sleepers will continue
4228 	 * to generate wakeups if necessary.
4229 	 */
4230 	wakeup_one(&zone->uz_max_items);
4231 }
4232 
4233 static uma_bucket_t
4234 zone_alloc_bucket(uma_zone_t zone, void *udata, int domain, int flags)
4235 {
4236 	uma_bucket_t bucket;
4237 	int error, maxbucket, cnt;
4238 
4239 	CTR3(KTR_UMA, "zone_alloc_bucket zone %s(%p) domain %d", zone->uz_name,
4240 	    zone, domain);
4241 
4242 	/* Avoid allocs targeting empty domains. */
4243 	if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
4244 		domain = UMA_ANYDOMAIN;
4245 	else if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
4246 		domain = UMA_ANYDOMAIN;
4247 
4248 	if (zone->uz_max_items > 0)
4249 		maxbucket = zone_alloc_limit(zone, zone->uz_bucket_size,
4250 		    M_NOWAIT);
4251 	else
4252 		maxbucket = zone->uz_bucket_size;
4253 	if (maxbucket == 0)
4254 		return (false);
4255 
4256 	/* Don't wait for buckets, preserve caller's NOVM setting. */
4257 	bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
4258 	if (bucket == NULL) {
4259 		cnt = 0;
4260 		goto out;
4261 	}
4262 
4263 	bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
4264 	    MIN(maxbucket, bucket->ub_entries), domain, flags);
4265 
4266 	/*
4267 	 * Initialize the memory if necessary.
4268 	 */
4269 	if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
4270 		int i;
4271 
4272 		for (i = 0; i < bucket->ub_cnt; i++) {
4273 			kasan_mark_item_valid(zone, bucket->ub_bucket[i]);
4274 			error = zone->uz_init(bucket->ub_bucket[i],
4275 			    zone->uz_size, flags);
4276 			kasan_mark_item_invalid(zone, bucket->ub_bucket[i]);
4277 			if (error != 0)
4278 				break;
4279 		}
4280 
4281 		/*
4282 		 * If we couldn't initialize the whole bucket, put the
4283 		 * rest back onto the freelist.
4284 		 */
4285 		if (i != bucket->ub_cnt) {
4286 			zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
4287 			    bucket->ub_cnt - i);
4288 #ifdef INVARIANTS
4289 			bzero(&bucket->ub_bucket[i],
4290 			    sizeof(void *) * (bucket->ub_cnt - i));
4291 #endif
4292 			bucket->ub_cnt = i;
4293 		}
4294 	}
4295 
4296 	cnt = bucket->ub_cnt;
4297 	if (bucket->ub_cnt == 0) {
4298 		bucket_free(zone, bucket, udata);
4299 		counter_u64_add(zone->uz_fails, 1);
4300 		bucket = NULL;
4301 	}
4302 out:
4303 	if (zone->uz_max_items > 0 && cnt < maxbucket)
4304 		zone_free_limit(zone, maxbucket - cnt);
4305 
4306 	return (bucket);
4307 }
4308 
4309 /*
4310  * Allocates a single item from a zone.
4311  *
4312  * Arguments
4313  *	zone   The zone to alloc for.
4314  *	udata  The data to be passed to the constructor.
4315  *	domain The domain to allocate from or UMA_ANYDOMAIN.
4316  *	flags  M_WAITOK, M_NOWAIT, M_ZERO.
4317  *
4318  * Returns
4319  *	NULL if there is no memory and M_NOWAIT is set
4320  *	An item if successful
4321  */
4322 
4323 static void *
4324 zone_alloc_item(uma_zone_t zone, void *udata, int domain, int flags)
4325 {
4326 	void *item;
4327 
4328 	if (zone->uz_max_items > 0 && zone_alloc_limit(zone, 1, flags) == 0) {
4329 		counter_u64_add(zone->uz_fails, 1);
4330 		return (NULL);
4331 	}
4332 
4333 	/* Avoid allocs targeting empty domains. */
4334 	if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
4335 		domain = UMA_ANYDOMAIN;
4336 
4337 	if (zone->uz_import(zone->uz_arg, &item, 1, domain, flags) != 1)
4338 		goto fail_cnt;
4339 
4340 	/*
4341 	 * We have to call both the zone's init (not the keg's init)
4342 	 * and the zone's ctor.  This is because the item is going from
4343 	 * a keg slab directly to the user, and the user is expecting it
4344 	 * to be both zone-init'd as well as zone-ctor'd.
4345 	 */
4346 	if (zone->uz_init != NULL) {
4347 		int error;
4348 
4349 		kasan_mark_item_valid(zone, item);
4350 		error = zone->uz_init(item, zone->uz_size, flags);
4351 		kasan_mark_item_invalid(zone, item);
4352 		if (error != 0) {
4353 			zone_free_item(zone, item, udata, SKIP_FINI | SKIP_CNT);
4354 			goto fail_cnt;
4355 		}
4356 	}
4357 	item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata, flags,
4358 	    item);
4359 	if (item == NULL)
4360 		goto fail;
4361 
4362 	counter_u64_add(zone->uz_allocs, 1);
4363 	CTR3(KTR_UMA, "zone_alloc_item item %p from %s(%p)", item,
4364 	    zone->uz_name, zone);
4365 
4366 	return (item);
4367 
4368 fail_cnt:
4369 	counter_u64_add(zone->uz_fails, 1);
4370 fail:
4371 	if (zone->uz_max_items > 0)
4372 		zone_free_limit(zone, 1);
4373 	CTR2(KTR_UMA, "zone_alloc_item failed from %s(%p)",
4374 	    zone->uz_name, zone);
4375 
4376 	return (NULL);
4377 }
4378 
4379 /* See uma.h */
4380 void
4381 uma_zfree_smr(uma_zone_t zone, void *item)
4382 {
4383 	uma_cache_t cache;
4384 	uma_cache_bucket_t bucket;
4385 	int itemdomain, uz_flags;
4386 
4387 	CTR3(KTR_UMA, "uma_zfree_smr zone %s(%p) item %p",
4388 	    zone->uz_name, zone, item);
4389 
4390 #ifdef UMA_ZALLOC_DEBUG
4391 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
4392 	    ("uma_zfree_smr: called with non-SMR zone."));
4393 	KASSERT(item != NULL, ("uma_zfree_smr: Called with NULL pointer."));
4394 	SMR_ASSERT_NOT_ENTERED(zone->uz_smr);
4395 	if (uma_zfree_debug(zone, item, NULL) == EJUSTRETURN)
4396 		return;
4397 #endif
4398 	cache = &zone->uz_cpu[curcpu];
4399 	uz_flags = cache_uz_flags(cache);
4400 	itemdomain = 0;
4401 #ifdef NUMA
4402 	if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4403 		itemdomain = item_domain(item);
4404 #endif
4405 	critical_enter();
4406 	do {
4407 		cache = &zone->uz_cpu[curcpu];
4408 		/* SMR Zones must free to the free bucket. */
4409 		bucket = &cache->uc_freebucket;
4410 #ifdef NUMA
4411 		if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4412 		    PCPU_GET(domain) != itemdomain) {
4413 			bucket = &cache->uc_crossbucket;
4414 		}
4415 #endif
4416 		if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4417 			cache_bucket_push(cache, bucket, item);
4418 			critical_exit();
4419 			return;
4420 		}
4421 	} while (cache_free(zone, cache, NULL, itemdomain));
4422 	critical_exit();
4423 
4424 	/*
4425 	 * If nothing else caught this, we'll just do an internal free.
4426 	 */
4427 	zone_free_item(zone, item, NULL, SKIP_NONE);
4428 }
4429 
4430 /* See uma.h */
4431 void
4432 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
4433 {
4434 	uma_cache_t cache;
4435 	uma_cache_bucket_t bucket;
4436 	int itemdomain, uz_flags;
4437 
4438 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
4439 	random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
4440 
4441 	CTR3(KTR_UMA, "uma_zfree_arg zone %s(%p) item %p",
4442 	    zone->uz_name, zone, item);
4443 
4444 #ifdef UMA_ZALLOC_DEBUG
4445 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
4446 	    ("uma_zfree_arg: called with SMR zone."));
4447 	if (uma_zfree_debug(zone, item, udata) == EJUSTRETURN)
4448 		return;
4449 #endif
4450         /* uma_zfree(..., NULL) does nothing, to match free(9). */
4451         if (item == NULL)
4452                 return;
4453 
4454 	/*
4455 	 * We are accessing the per-cpu cache without a critical section to
4456 	 * fetch size and flags.  This is acceptable, if we are preempted we
4457 	 * will simply read another cpu's line.
4458 	 */
4459 	cache = &zone->uz_cpu[curcpu];
4460 	uz_flags = cache_uz_flags(cache);
4461 	if (UMA_ALWAYS_CTORDTOR ||
4462 	    __predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0))
4463 		item_dtor(zone, item, cache_uz_size(cache), udata, SKIP_NONE);
4464 
4465 	/*
4466 	 * The race here is acceptable.  If we miss it we'll just have to wait
4467 	 * a little longer for the limits to be reset.
4468 	 */
4469 	if (__predict_false(uz_flags & UMA_ZFLAG_LIMIT)) {
4470 		if (atomic_load_32(&zone->uz_sleepers) > 0)
4471 			goto zfree_item;
4472 	}
4473 
4474 	/*
4475 	 * If possible, free to the per-CPU cache.  There are two
4476 	 * requirements for safe access to the per-CPU cache: (1) the thread
4477 	 * accessing the cache must not be preempted or yield during access,
4478 	 * and (2) the thread must not migrate CPUs without switching which
4479 	 * cache it accesses.  We rely on a critical section to prevent
4480 	 * preemption and migration.  We release the critical section in
4481 	 * order to acquire the zone mutex if we are unable to free to the
4482 	 * current cache; when we re-acquire the critical section, we must
4483 	 * detect and handle migration if it has occurred.
4484 	 */
4485 	itemdomain = 0;
4486 #ifdef NUMA
4487 	if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4488 		itemdomain = item_domain(item);
4489 #endif
4490 	critical_enter();
4491 	do {
4492 		cache = &zone->uz_cpu[curcpu];
4493 		/*
4494 		 * Try to free into the allocbucket first to give LIFO
4495 		 * ordering for cache-hot datastructures.  Spill over
4496 		 * into the freebucket if necessary.  Alloc will swap
4497 		 * them if one runs dry.
4498 		 */
4499 		bucket = &cache->uc_allocbucket;
4500 #ifdef NUMA
4501 		if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4502 		    PCPU_GET(domain) != itemdomain) {
4503 			bucket = &cache->uc_crossbucket;
4504 		} else
4505 #endif
4506 		if (bucket->ucb_cnt == bucket->ucb_entries &&
4507 		   cache->uc_freebucket.ucb_cnt <
4508 		   cache->uc_freebucket.ucb_entries)
4509 			cache_bucket_swap(&cache->uc_freebucket,
4510 			    &cache->uc_allocbucket);
4511 		if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4512 			cache_bucket_push(cache, bucket, item);
4513 			critical_exit();
4514 			return;
4515 		}
4516 	} while (cache_free(zone, cache, udata, itemdomain));
4517 	critical_exit();
4518 
4519 	/*
4520 	 * If nothing else caught this, we'll just do an internal free.
4521 	 */
4522 zfree_item:
4523 	zone_free_item(zone, item, udata, SKIP_DTOR);
4524 }
4525 
4526 #ifdef NUMA
4527 /*
4528  * sort crossdomain free buckets to domain correct buckets and cache
4529  * them.
4530  */
4531 static void
4532 zone_free_cross(uma_zone_t zone, uma_bucket_t bucket, void *udata)
4533 {
4534 	struct uma_bucketlist emptybuckets, fullbuckets;
4535 	uma_zone_domain_t zdom;
4536 	uma_bucket_t b;
4537 	smr_seq_t seq;
4538 	void *item;
4539 	int domain;
4540 
4541 	CTR3(KTR_UMA,
4542 	    "uma_zfree: zone %s(%p) draining cross bucket %p",
4543 	    zone->uz_name, zone, bucket);
4544 
4545 	/*
4546 	 * It is possible for buckets to arrive here out of order so we fetch
4547 	 * the current smr seq rather than accepting the bucket's.
4548 	 */
4549 	seq = SMR_SEQ_INVALID;
4550 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
4551 		seq = smr_advance(zone->uz_smr);
4552 
4553 	/*
4554 	 * To avoid having ndomain * ndomain buckets for sorting we have a
4555 	 * lock on the current crossfree bucket.  A full matrix with
4556 	 * per-domain locking could be used if necessary.
4557 	 */
4558 	STAILQ_INIT(&emptybuckets);
4559 	STAILQ_INIT(&fullbuckets);
4560 	ZONE_CROSS_LOCK(zone);
4561 	for (; bucket->ub_cnt > 0; bucket->ub_cnt--) {
4562 		item = bucket->ub_bucket[bucket->ub_cnt - 1];
4563 		domain = item_domain(item);
4564 		zdom = ZDOM_GET(zone, domain);
4565 		if (zdom->uzd_cross == NULL) {
4566 			if ((b = STAILQ_FIRST(&emptybuckets)) != NULL) {
4567 				STAILQ_REMOVE_HEAD(&emptybuckets, ub_link);
4568 				zdom->uzd_cross = b;
4569 			} else {
4570 				/*
4571 				 * Avoid allocating a bucket with the cross lock
4572 				 * held, since allocation can trigger a
4573 				 * cross-domain free and bucket zones may
4574 				 * allocate from each other.
4575 				 */
4576 				ZONE_CROSS_UNLOCK(zone);
4577 				b = bucket_alloc(zone, udata, M_NOWAIT);
4578 				if (b == NULL)
4579 					goto out;
4580 				ZONE_CROSS_LOCK(zone);
4581 				if (zdom->uzd_cross != NULL) {
4582 					STAILQ_INSERT_HEAD(&emptybuckets, b,
4583 					    ub_link);
4584 				} else {
4585 					zdom->uzd_cross = b;
4586 				}
4587 			}
4588 		}
4589 		b = zdom->uzd_cross;
4590 		b->ub_bucket[b->ub_cnt++] = item;
4591 		b->ub_seq = seq;
4592 		if (b->ub_cnt == b->ub_entries) {
4593 			STAILQ_INSERT_HEAD(&fullbuckets, b, ub_link);
4594 			if ((b = STAILQ_FIRST(&emptybuckets)) != NULL)
4595 				STAILQ_REMOVE_HEAD(&emptybuckets, ub_link);
4596 			zdom->uzd_cross = b;
4597 		}
4598 	}
4599 	ZONE_CROSS_UNLOCK(zone);
4600 out:
4601 	if (bucket->ub_cnt == 0)
4602 		bucket->ub_seq = SMR_SEQ_INVALID;
4603 	bucket_free(zone, bucket, udata);
4604 
4605 	while ((b = STAILQ_FIRST(&emptybuckets)) != NULL) {
4606 		STAILQ_REMOVE_HEAD(&emptybuckets, ub_link);
4607 		bucket_free(zone, b, udata);
4608 	}
4609 	while ((b = STAILQ_FIRST(&fullbuckets)) != NULL) {
4610 		STAILQ_REMOVE_HEAD(&fullbuckets, ub_link);
4611 		domain = item_domain(b->ub_bucket[0]);
4612 		zone_put_bucket(zone, domain, b, udata, true);
4613 	}
4614 }
4615 #endif
4616 
4617 static void
4618 zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata,
4619     int itemdomain, bool ws)
4620 {
4621 
4622 #ifdef NUMA
4623 	/*
4624 	 * Buckets coming from the wrong domain will be entirely for the
4625 	 * only other domain on two domain systems.  In this case we can
4626 	 * simply cache them.  Otherwise we need to sort them back to
4627 	 * correct domains.
4628 	 */
4629 	if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4630 	    vm_ndomains > 2 && PCPU_GET(domain) != itemdomain) {
4631 		zone_free_cross(zone, bucket, udata);
4632 		return;
4633 	}
4634 #endif
4635 
4636 	/*
4637 	 * Attempt to save the bucket in the zone's domain bucket cache.
4638 	 */
4639 	CTR3(KTR_UMA,
4640 	    "uma_zfree: zone %s(%p) putting bucket %p on free list",
4641 	    zone->uz_name, zone, bucket);
4642 	/* ub_cnt is pointing to the last free item */
4643 	if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
4644 		itemdomain = zone_domain_lowest(zone, itemdomain);
4645 	zone_put_bucket(zone, itemdomain, bucket, udata, ws);
4646 }
4647 
4648 /*
4649  * Populate a free or cross bucket for the current cpu cache.  Free any
4650  * existing full bucket either to the zone cache or back to the slab layer.
4651  *
4652  * Enters and returns in a critical section.  false return indicates that
4653  * we can not satisfy this free in the cache layer.  true indicates that
4654  * the caller should retry.
4655  */
4656 static __noinline bool
4657 cache_free(uma_zone_t zone, uma_cache_t cache, void *udata, int itemdomain)
4658 {
4659 	uma_cache_bucket_t cbucket;
4660 	uma_bucket_t newbucket, bucket;
4661 
4662 	CRITICAL_ASSERT(curthread);
4663 
4664 	if (zone->uz_bucket_size == 0)
4665 		return false;
4666 
4667 	cache = &zone->uz_cpu[curcpu];
4668 	newbucket = NULL;
4669 
4670 	/*
4671 	 * FIRSTTOUCH domains need to free to the correct zdom.  When
4672 	 * enabled this is the zdom of the item.   The bucket is the
4673 	 * cross bucket if the current domain and itemdomain do not match.
4674 	 */
4675 	cbucket = &cache->uc_freebucket;
4676 #ifdef NUMA
4677 	if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4678 		if (PCPU_GET(domain) != itemdomain) {
4679 			cbucket = &cache->uc_crossbucket;
4680 			if (cbucket->ucb_cnt != 0)
4681 				counter_u64_add(zone->uz_xdomain,
4682 				    cbucket->ucb_cnt);
4683 		}
4684 	}
4685 #endif
4686 	bucket = cache_bucket_unload(cbucket);
4687 	KASSERT(bucket == NULL || bucket->ub_cnt == bucket->ub_entries,
4688 	    ("cache_free: Entered with non-full free bucket."));
4689 
4690 	/* We are no longer associated with this CPU. */
4691 	critical_exit();
4692 
4693 	/*
4694 	 * Don't let SMR zones operate without a free bucket.  Force
4695 	 * a synchronize and re-use this one.  We will only degrade
4696 	 * to a synchronize every bucket_size items rather than every
4697 	 * item if we fail to allocate a bucket.
4698 	 */
4699 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0) {
4700 		if (bucket != NULL)
4701 			bucket->ub_seq = smr_advance(zone->uz_smr);
4702 		newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4703 		if (newbucket == NULL && bucket != NULL) {
4704 			bucket_drain(zone, bucket);
4705 			newbucket = bucket;
4706 			bucket = NULL;
4707 		}
4708 	} else if (!bucketdisable)
4709 		newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4710 
4711 	if (bucket != NULL)
4712 		zone_free_bucket(zone, bucket, udata, itemdomain, true);
4713 
4714 	critical_enter();
4715 	if ((bucket = newbucket) == NULL)
4716 		return (false);
4717 	cache = &zone->uz_cpu[curcpu];
4718 #ifdef NUMA
4719 	/*
4720 	 * Check to see if we should be populating the cross bucket.  If it
4721 	 * is already populated we will fall through and attempt to populate
4722 	 * the free bucket.
4723 	 */
4724 	if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4725 		if (PCPU_GET(domain) != itemdomain &&
4726 		    cache->uc_crossbucket.ucb_bucket == NULL) {
4727 			cache_bucket_load_cross(cache, bucket);
4728 			return (true);
4729 		}
4730 	}
4731 #endif
4732 	/*
4733 	 * We may have lost the race to fill the bucket or switched CPUs.
4734 	 */
4735 	if (cache->uc_freebucket.ucb_bucket != NULL) {
4736 		critical_exit();
4737 		bucket_free(zone, bucket, udata);
4738 		critical_enter();
4739 	} else
4740 		cache_bucket_load_free(cache, bucket);
4741 
4742 	return (true);
4743 }
4744 
4745 static void
4746 slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item)
4747 {
4748 	uma_keg_t keg;
4749 	uma_domain_t dom;
4750 	int freei;
4751 
4752 	keg = zone->uz_keg;
4753 	KEG_LOCK_ASSERT(keg, slab->us_domain);
4754 
4755 	/* Do we need to remove from any lists? */
4756 	dom = &keg->uk_domain[slab->us_domain];
4757 	if (slab->us_freecount + 1 == keg->uk_ipers) {
4758 		LIST_REMOVE(slab, us_link);
4759 		LIST_INSERT_HEAD(&dom->ud_free_slab, slab, us_link);
4760 		dom->ud_free_slabs++;
4761 	} else if (slab->us_freecount == 0) {
4762 		LIST_REMOVE(slab, us_link);
4763 		LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
4764 	}
4765 
4766 	/* Slab management. */
4767 	freei = slab_item_index(slab, keg, item);
4768 	BIT_SET(keg->uk_ipers, freei, &slab->us_free);
4769 	slab->us_freecount++;
4770 
4771 	/* Keg statistics. */
4772 	dom->ud_free_items++;
4773 }
4774 
4775 static void
4776 zone_release(void *arg, void **bucket, int cnt)
4777 {
4778 	struct mtx *lock;
4779 	uma_zone_t zone;
4780 	uma_slab_t slab;
4781 	uma_keg_t keg;
4782 	uint8_t *mem;
4783 	void *item;
4784 	int i;
4785 
4786 	zone = arg;
4787 	keg = zone->uz_keg;
4788 	lock = NULL;
4789 	if (__predict_false((zone->uz_flags & UMA_ZFLAG_HASH) != 0))
4790 		lock = KEG_LOCK(keg, 0);
4791 	for (i = 0; i < cnt; i++) {
4792 		item = bucket[i];
4793 		if (__predict_true((zone->uz_flags & UMA_ZFLAG_VTOSLAB) != 0)) {
4794 			slab = vtoslab((vm_offset_t)item);
4795 		} else {
4796 			mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
4797 			if ((zone->uz_flags & UMA_ZFLAG_HASH) != 0)
4798 				slab = hash_sfind(&keg->uk_hash, mem);
4799 			else
4800 				slab = (uma_slab_t)(mem + keg->uk_pgoff);
4801 		}
4802 		if (lock != KEG_LOCKPTR(keg, slab->us_domain)) {
4803 			if (lock != NULL)
4804 				mtx_unlock(lock);
4805 			lock = KEG_LOCK(keg, slab->us_domain);
4806 		}
4807 		slab_free_item(zone, slab, item);
4808 	}
4809 	if (lock != NULL)
4810 		mtx_unlock(lock);
4811 }
4812 
4813 /*
4814  * Frees a single item to any zone.
4815  *
4816  * Arguments:
4817  *	zone   The zone to free to
4818  *	item   The item we're freeing
4819  *	udata  User supplied data for the dtor
4820  *	skip   Skip dtors and finis
4821  */
4822 static __noinline void
4823 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
4824 {
4825 
4826 	/*
4827 	 * If a free is sent directly to an SMR zone we have to
4828 	 * synchronize immediately because the item can instantly
4829 	 * be reallocated. This should only happen in degenerate
4830 	 * cases when no memory is available for per-cpu caches.
4831 	 */
4832 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0 && skip == SKIP_NONE)
4833 		smr_synchronize(zone->uz_smr);
4834 
4835 	item_dtor(zone, item, zone->uz_size, udata, skip);
4836 
4837 	if (skip < SKIP_FINI && zone->uz_fini) {
4838 		kasan_mark_item_valid(zone, item);
4839 		zone->uz_fini(item, zone->uz_size);
4840 		kasan_mark_item_invalid(zone, item);
4841 	}
4842 
4843 	zone->uz_release(zone->uz_arg, &item, 1);
4844 
4845 	if (skip & SKIP_CNT)
4846 		return;
4847 
4848 	counter_u64_add(zone->uz_frees, 1);
4849 
4850 	if (zone->uz_max_items > 0)
4851 		zone_free_limit(zone, 1);
4852 }
4853 
4854 /* See uma.h */
4855 int
4856 uma_zone_set_max(uma_zone_t zone, int nitems)
4857 {
4858 
4859 	/*
4860 	 * If the limit is small, we may need to constrain the maximum per-CPU
4861 	 * cache size, or disable caching entirely.
4862 	 */
4863 	uma_zone_set_maxcache(zone, nitems);
4864 
4865 	/*
4866 	 * XXX This can misbehave if the zone has any allocations with
4867 	 * no limit and a limit is imposed.  There is currently no
4868 	 * way to clear a limit.
4869 	 */
4870 	ZONE_LOCK(zone);
4871 	zone->uz_max_items = nitems;
4872 	zone->uz_flags |= UMA_ZFLAG_LIMIT;
4873 	zone_update_caches(zone);
4874 	/* We may need to wake waiters. */
4875 	wakeup(&zone->uz_max_items);
4876 	ZONE_UNLOCK(zone);
4877 
4878 	return (nitems);
4879 }
4880 
4881 /* See uma.h */
4882 void
4883 uma_zone_set_maxcache(uma_zone_t zone, int nitems)
4884 {
4885 	int bpcpu, bpdom, bsize, nb;
4886 
4887 	ZONE_LOCK(zone);
4888 
4889 	/*
4890 	 * Compute a lower bound on the number of items that may be cached in
4891 	 * the zone.  Each CPU gets at least two buckets, and for cross-domain
4892 	 * frees we use an additional bucket per CPU and per domain.  Select the
4893 	 * largest bucket size that does not exceed half of the requested limit,
4894 	 * with the left over space given to the full bucket cache.
4895 	 */
4896 	bpdom = 0;
4897 	bpcpu = 2;
4898 #ifdef NUMA
4899 	if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 && vm_ndomains > 1) {
4900 		bpcpu++;
4901 		bpdom++;
4902 	}
4903 #endif
4904 	nb = bpcpu * mp_ncpus + bpdom * vm_ndomains;
4905 	bsize = nitems / nb / 2;
4906 	if (bsize > BUCKET_MAX)
4907 		bsize = BUCKET_MAX;
4908 	else if (bsize == 0 && nitems / nb > 0)
4909 		bsize = 1;
4910 	zone->uz_bucket_size_max = zone->uz_bucket_size = bsize;
4911 	if (zone->uz_bucket_size_min > zone->uz_bucket_size_max)
4912 		zone->uz_bucket_size_min = zone->uz_bucket_size_max;
4913 	zone->uz_bucket_max = nitems - nb * bsize;
4914 	ZONE_UNLOCK(zone);
4915 }
4916 
4917 /* See uma.h */
4918 int
4919 uma_zone_get_max(uma_zone_t zone)
4920 {
4921 	int nitems;
4922 
4923 	nitems = atomic_load_64(&zone->uz_max_items);
4924 
4925 	return (nitems);
4926 }
4927 
4928 /* See uma.h */
4929 void
4930 uma_zone_set_warning(uma_zone_t zone, const char *warning)
4931 {
4932 
4933 	ZONE_ASSERT_COLD(zone);
4934 	zone->uz_warning = warning;
4935 }
4936 
4937 /* See uma.h */
4938 void
4939 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction)
4940 {
4941 
4942 	ZONE_ASSERT_COLD(zone);
4943 	TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone);
4944 }
4945 
4946 /* See uma.h */
4947 int
4948 uma_zone_get_cur(uma_zone_t zone)
4949 {
4950 	int64_t nitems;
4951 	u_int i;
4952 
4953 	nitems = 0;
4954 	if (zone->uz_allocs != EARLY_COUNTER && zone->uz_frees != EARLY_COUNTER)
4955 		nitems = counter_u64_fetch(zone->uz_allocs) -
4956 		    counter_u64_fetch(zone->uz_frees);
4957 	CPU_FOREACH(i)
4958 		nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs) -
4959 		    atomic_load_64(&zone->uz_cpu[i].uc_frees);
4960 
4961 	return (nitems < 0 ? 0 : nitems);
4962 }
4963 
4964 static uint64_t
4965 uma_zone_get_allocs(uma_zone_t zone)
4966 {
4967 	uint64_t nitems;
4968 	u_int i;
4969 
4970 	nitems = 0;
4971 	if (zone->uz_allocs != EARLY_COUNTER)
4972 		nitems = counter_u64_fetch(zone->uz_allocs);
4973 	CPU_FOREACH(i)
4974 		nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs);
4975 
4976 	return (nitems);
4977 }
4978 
4979 static uint64_t
4980 uma_zone_get_frees(uma_zone_t zone)
4981 {
4982 	uint64_t nitems;
4983 	u_int i;
4984 
4985 	nitems = 0;
4986 	if (zone->uz_frees != EARLY_COUNTER)
4987 		nitems = counter_u64_fetch(zone->uz_frees);
4988 	CPU_FOREACH(i)
4989 		nitems += atomic_load_64(&zone->uz_cpu[i].uc_frees);
4990 
4991 	return (nitems);
4992 }
4993 
4994 #ifdef INVARIANTS
4995 /* Used only for KEG_ASSERT_COLD(). */
4996 static uint64_t
4997 uma_keg_get_allocs(uma_keg_t keg)
4998 {
4999 	uma_zone_t z;
5000 	uint64_t nitems;
5001 
5002 	nitems = 0;
5003 	LIST_FOREACH(z, &keg->uk_zones, uz_link)
5004 		nitems += uma_zone_get_allocs(z);
5005 
5006 	return (nitems);
5007 }
5008 #endif
5009 
5010 /* See uma.h */
5011 void
5012 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
5013 {
5014 	uma_keg_t keg;
5015 
5016 	KEG_GET(zone, keg);
5017 	KEG_ASSERT_COLD(keg);
5018 	keg->uk_init = uminit;
5019 }
5020 
5021 /* See uma.h */
5022 void
5023 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
5024 {
5025 	uma_keg_t keg;
5026 
5027 	KEG_GET(zone, keg);
5028 	KEG_ASSERT_COLD(keg);
5029 	keg->uk_fini = fini;
5030 }
5031 
5032 /* See uma.h */
5033 void
5034 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
5035 {
5036 
5037 	ZONE_ASSERT_COLD(zone);
5038 	zone->uz_init = zinit;
5039 }
5040 
5041 /* See uma.h */
5042 void
5043 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
5044 {
5045 
5046 	ZONE_ASSERT_COLD(zone);
5047 	zone->uz_fini = zfini;
5048 }
5049 
5050 /* See uma.h */
5051 void
5052 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
5053 {
5054 	uma_keg_t keg;
5055 
5056 	KEG_GET(zone, keg);
5057 	KEG_ASSERT_COLD(keg);
5058 	keg->uk_freef = freef;
5059 }
5060 
5061 /* See uma.h */
5062 void
5063 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
5064 {
5065 	uma_keg_t keg;
5066 
5067 	KEG_GET(zone, keg);
5068 	KEG_ASSERT_COLD(keg);
5069 	keg->uk_allocf = allocf;
5070 }
5071 
5072 /* See uma.h */
5073 void
5074 uma_zone_set_smr(uma_zone_t zone, smr_t smr)
5075 {
5076 
5077 	ZONE_ASSERT_COLD(zone);
5078 
5079 	KASSERT(smr != NULL, ("Got NULL smr"));
5080 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
5081 	    ("zone %p (%s) already uses SMR", zone, zone->uz_name));
5082 	zone->uz_flags |= UMA_ZONE_SMR;
5083 	zone->uz_smr = smr;
5084 	zone_update_caches(zone);
5085 }
5086 
5087 smr_t
5088 uma_zone_get_smr(uma_zone_t zone)
5089 {
5090 
5091 	return (zone->uz_smr);
5092 }
5093 
5094 /* See uma.h */
5095 void
5096 uma_zone_reserve(uma_zone_t zone, int items)
5097 {
5098 	uma_keg_t keg;
5099 
5100 	KEG_GET(zone, keg);
5101 	KEG_ASSERT_COLD(keg);
5102 	keg->uk_reserve = items;
5103 }
5104 
5105 /* See uma.h */
5106 int
5107 uma_zone_reserve_kva(uma_zone_t zone, int count)
5108 {
5109 	uma_keg_t keg;
5110 	vm_offset_t kva;
5111 	u_int pages;
5112 
5113 	KEG_GET(zone, keg);
5114 	KEG_ASSERT_COLD(keg);
5115 	ZONE_ASSERT_COLD(zone);
5116 
5117 	pages = howmany(count, keg->uk_ipers) * keg->uk_ppera;
5118 
5119 #ifdef UMA_MD_SMALL_ALLOC
5120 	if (keg->uk_ppera > 1) {
5121 #else
5122 	if (1) {
5123 #endif
5124 		kva = kva_alloc((vm_size_t)pages * PAGE_SIZE);
5125 		if (kva == 0)
5126 			return (0);
5127 	} else
5128 		kva = 0;
5129 
5130 	MPASS(keg->uk_kva == 0);
5131 	keg->uk_kva = kva;
5132 	keg->uk_offset = 0;
5133 	zone->uz_max_items = pages * keg->uk_ipers;
5134 #ifdef UMA_MD_SMALL_ALLOC
5135 	keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
5136 #else
5137 	keg->uk_allocf = noobj_alloc;
5138 #endif
5139 	keg->uk_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
5140 	zone->uz_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
5141 	zone_update_caches(zone);
5142 
5143 	return (1);
5144 }
5145 
5146 /* See uma.h */
5147 void
5148 uma_prealloc(uma_zone_t zone, int items)
5149 {
5150 	struct vm_domainset_iter di;
5151 	uma_domain_t dom;
5152 	uma_slab_t slab;
5153 	uma_keg_t keg;
5154 	int aflags, domain, slabs;
5155 
5156 	KEG_GET(zone, keg);
5157 	slabs = howmany(items, keg->uk_ipers);
5158 	while (slabs-- > 0) {
5159 		aflags = M_NOWAIT;
5160 		vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
5161 		    &aflags);
5162 		for (;;) {
5163 			slab = keg_alloc_slab(keg, zone, domain, M_WAITOK,
5164 			    aflags);
5165 			if (slab != NULL) {
5166 				dom = &keg->uk_domain[slab->us_domain];
5167 				/*
5168 				 * keg_alloc_slab() always returns a slab on the
5169 				 * partial list.
5170 				 */
5171 				LIST_REMOVE(slab, us_link);
5172 				LIST_INSERT_HEAD(&dom->ud_free_slab, slab,
5173 				    us_link);
5174 				dom->ud_free_slabs++;
5175 				KEG_UNLOCK(keg, slab->us_domain);
5176 				break;
5177 			}
5178 			if (vm_domainset_iter_policy(&di, &domain) != 0)
5179 				vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask, 0);
5180 		}
5181 	}
5182 }
5183 
5184 /*
5185  * Returns a snapshot of memory consumption in bytes.
5186  */
5187 size_t
5188 uma_zone_memory(uma_zone_t zone)
5189 {
5190 	size_t sz;
5191 	int i;
5192 
5193 	sz = 0;
5194 	if (zone->uz_flags & UMA_ZFLAG_CACHE) {
5195 		for (i = 0; i < vm_ndomains; i++)
5196 			sz += ZDOM_GET(zone, i)->uzd_nitems;
5197 		return (sz * zone->uz_size);
5198 	}
5199 	for (i = 0; i < vm_ndomains; i++)
5200 		sz += zone->uz_keg->uk_domain[i].ud_pages;
5201 
5202 	return (sz * PAGE_SIZE);
5203 }
5204 
5205 /* See uma.h */
5206 void
5207 uma_reclaim(int req)
5208 {
5209 	uma_reclaim_domain(req, UMA_ANYDOMAIN);
5210 }
5211 
5212 void
5213 uma_reclaim_domain(int req, int domain)
5214 {
5215 	void *arg;
5216 
5217 	bucket_enable();
5218 
5219 	arg = (void *)(uintptr_t)domain;
5220 	sx_slock(&uma_reclaim_lock);
5221 	switch (req) {
5222 	case UMA_RECLAIM_TRIM:
5223 		zone_foreach(zone_trim, arg);
5224 		break;
5225 	case UMA_RECLAIM_DRAIN:
5226 		zone_foreach(zone_drain, arg);
5227 		break;
5228 	case UMA_RECLAIM_DRAIN_CPU:
5229 		zone_foreach(zone_drain, arg);
5230 		pcpu_cache_drain_safe(NULL);
5231 		zone_foreach(zone_drain, arg);
5232 		break;
5233 	default:
5234 		panic("unhandled reclamation request %d", req);
5235 	}
5236 
5237 	/*
5238 	 * Some slabs may have been freed but this zone will be visited early
5239 	 * we visit again so that we can free pages that are empty once other
5240 	 * zones are drained.  We have to do the same for buckets.
5241 	 */
5242 	zone_drain(slabzones[0], arg);
5243 	zone_drain(slabzones[1], arg);
5244 	bucket_zone_drain(domain);
5245 	sx_sunlock(&uma_reclaim_lock);
5246 }
5247 
5248 static volatile int uma_reclaim_needed;
5249 
5250 void
5251 uma_reclaim_wakeup(void)
5252 {
5253 
5254 	if (atomic_fetchadd_int(&uma_reclaim_needed, 1) == 0)
5255 		wakeup(uma_reclaim);
5256 }
5257 
5258 void
5259 uma_reclaim_worker(void *arg __unused)
5260 {
5261 
5262 	for (;;) {
5263 		sx_xlock(&uma_reclaim_lock);
5264 		while (atomic_load_int(&uma_reclaim_needed) == 0)
5265 			sx_sleep(uma_reclaim, &uma_reclaim_lock, PVM, "umarcl",
5266 			    hz);
5267 		sx_xunlock(&uma_reclaim_lock);
5268 		EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM);
5269 		uma_reclaim(UMA_RECLAIM_DRAIN_CPU);
5270 		atomic_store_int(&uma_reclaim_needed, 0);
5271 		/* Don't fire more than once per-second. */
5272 		pause("umarclslp", hz);
5273 	}
5274 }
5275 
5276 /* See uma.h */
5277 void
5278 uma_zone_reclaim(uma_zone_t zone, int req)
5279 {
5280 	uma_zone_reclaim_domain(zone, req, UMA_ANYDOMAIN);
5281 }
5282 
5283 void
5284 uma_zone_reclaim_domain(uma_zone_t zone, int req, int domain)
5285 {
5286 	void *arg;
5287 
5288 	arg = (void *)(uintptr_t)domain;
5289 	switch (req) {
5290 	case UMA_RECLAIM_TRIM:
5291 		zone_trim(zone, arg);
5292 		break;
5293 	case UMA_RECLAIM_DRAIN:
5294 		zone_drain(zone, arg);
5295 		break;
5296 	case UMA_RECLAIM_DRAIN_CPU:
5297 		pcpu_cache_drain_safe(zone);
5298 		zone_drain(zone, arg);
5299 		break;
5300 	default:
5301 		panic("unhandled reclamation request %d", req);
5302 	}
5303 }
5304 
5305 /* See uma.h */
5306 int
5307 uma_zone_exhausted(uma_zone_t zone)
5308 {
5309 
5310 	return (atomic_load_32(&zone->uz_sleepers) > 0);
5311 }
5312 
5313 unsigned long
5314 uma_limit(void)
5315 {
5316 
5317 	return (uma_kmem_limit);
5318 }
5319 
5320 void
5321 uma_set_limit(unsigned long limit)
5322 {
5323 
5324 	uma_kmem_limit = limit;
5325 }
5326 
5327 unsigned long
5328 uma_size(void)
5329 {
5330 
5331 	return (atomic_load_long(&uma_kmem_total));
5332 }
5333 
5334 long
5335 uma_avail(void)
5336 {
5337 
5338 	return (uma_kmem_limit - uma_size());
5339 }
5340 
5341 #ifdef DDB
5342 /*
5343  * Generate statistics across both the zone and its per-cpu cache's.  Return
5344  * desired statistics if the pointer is non-NULL for that statistic.
5345  *
5346  * Note: does not update the zone statistics, as it can't safely clear the
5347  * per-CPU cache statistic.
5348  *
5349  */
5350 static void
5351 uma_zone_sumstat(uma_zone_t z, long *cachefreep, uint64_t *allocsp,
5352     uint64_t *freesp, uint64_t *sleepsp, uint64_t *xdomainp)
5353 {
5354 	uma_cache_t cache;
5355 	uint64_t allocs, frees, sleeps, xdomain;
5356 	int cachefree, cpu;
5357 
5358 	allocs = frees = sleeps = xdomain = 0;
5359 	cachefree = 0;
5360 	CPU_FOREACH(cpu) {
5361 		cache = &z->uz_cpu[cpu];
5362 		cachefree += cache->uc_allocbucket.ucb_cnt;
5363 		cachefree += cache->uc_freebucket.ucb_cnt;
5364 		xdomain += cache->uc_crossbucket.ucb_cnt;
5365 		cachefree += cache->uc_crossbucket.ucb_cnt;
5366 		allocs += cache->uc_allocs;
5367 		frees += cache->uc_frees;
5368 	}
5369 	allocs += counter_u64_fetch(z->uz_allocs);
5370 	frees += counter_u64_fetch(z->uz_frees);
5371 	xdomain += counter_u64_fetch(z->uz_xdomain);
5372 	sleeps += z->uz_sleeps;
5373 	if (cachefreep != NULL)
5374 		*cachefreep = cachefree;
5375 	if (allocsp != NULL)
5376 		*allocsp = allocs;
5377 	if (freesp != NULL)
5378 		*freesp = frees;
5379 	if (sleepsp != NULL)
5380 		*sleepsp = sleeps;
5381 	if (xdomainp != NULL)
5382 		*xdomainp = xdomain;
5383 }
5384 #endif /* DDB */
5385 
5386 static int
5387 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
5388 {
5389 	uma_keg_t kz;
5390 	uma_zone_t z;
5391 	int count;
5392 
5393 	count = 0;
5394 	rw_rlock(&uma_rwlock);
5395 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
5396 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
5397 			count++;
5398 	}
5399 	LIST_FOREACH(z, &uma_cachezones, uz_link)
5400 		count++;
5401 
5402 	rw_runlock(&uma_rwlock);
5403 	return (sysctl_handle_int(oidp, &count, 0, req));
5404 }
5405 
5406 static void
5407 uma_vm_zone_stats(struct uma_type_header *uth, uma_zone_t z, struct sbuf *sbuf,
5408     struct uma_percpu_stat *ups, bool internal)
5409 {
5410 	uma_zone_domain_t zdom;
5411 	uma_cache_t cache;
5412 	int i;
5413 
5414 	for (i = 0; i < vm_ndomains; i++) {
5415 		zdom = ZDOM_GET(z, i);
5416 		uth->uth_zone_free += zdom->uzd_nitems;
5417 	}
5418 	uth->uth_allocs = counter_u64_fetch(z->uz_allocs);
5419 	uth->uth_frees = counter_u64_fetch(z->uz_frees);
5420 	uth->uth_fails = counter_u64_fetch(z->uz_fails);
5421 	uth->uth_xdomain = counter_u64_fetch(z->uz_xdomain);
5422 	uth->uth_sleeps = z->uz_sleeps;
5423 
5424 	for (i = 0; i < mp_maxid + 1; i++) {
5425 		bzero(&ups[i], sizeof(*ups));
5426 		if (internal || CPU_ABSENT(i))
5427 			continue;
5428 		cache = &z->uz_cpu[i];
5429 		ups[i].ups_cache_free += cache->uc_allocbucket.ucb_cnt;
5430 		ups[i].ups_cache_free += cache->uc_freebucket.ucb_cnt;
5431 		ups[i].ups_cache_free += cache->uc_crossbucket.ucb_cnt;
5432 		ups[i].ups_allocs = cache->uc_allocs;
5433 		ups[i].ups_frees = cache->uc_frees;
5434 	}
5435 }
5436 
5437 static int
5438 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
5439 {
5440 	struct uma_stream_header ush;
5441 	struct uma_type_header uth;
5442 	struct uma_percpu_stat *ups;
5443 	struct sbuf sbuf;
5444 	uma_keg_t kz;
5445 	uma_zone_t z;
5446 	uint64_t items;
5447 	uint32_t kfree, pages;
5448 	int count, error, i;
5449 
5450 	error = sysctl_wire_old_buffer(req, 0);
5451 	if (error != 0)
5452 		return (error);
5453 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
5454 	sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL);
5455 	ups = malloc((mp_maxid + 1) * sizeof(*ups), M_TEMP, M_WAITOK);
5456 
5457 	count = 0;
5458 	rw_rlock(&uma_rwlock);
5459 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
5460 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
5461 			count++;
5462 	}
5463 
5464 	LIST_FOREACH(z, &uma_cachezones, uz_link)
5465 		count++;
5466 
5467 	/*
5468 	 * Insert stream header.
5469 	 */
5470 	bzero(&ush, sizeof(ush));
5471 	ush.ush_version = UMA_STREAM_VERSION;
5472 	ush.ush_maxcpus = (mp_maxid + 1);
5473 	ush.ush_count = count;
5474 	(void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
5475 
5476 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
5477 		kfree = pages = 0;
5478 		for (i = 0; i < vm_ndomains; i++) {
5479 			kfree += kz->uk_domain[i].ud_free_items;
5480 			pages += kz->uk_domain[i].ud_pages;
5481 		}
5482 		LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5483 			bzero(&uth, sizeof(uth));
5484 			strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5485 			uth.uth_align = kz->uk_align;
5486 			uth.uth_size = kz->uk_size;
5487 			uth.uth_rsize = kz->uk_rsize;
5488 			if (z->uz_max_items > 0) {
5489 				items = UZ_ITEMS_COUNT(z->uz_items);
5490 				uth.uth_pages = (items / kz->uk_ipers) *
5491 					kz->uk_ppera;
5492 			} else
5493 				uth.uth_pages = pages;
5494 			uth.uth_maxpages = (z->uz_max_items / kz->uk_ipers) *
5495 			    kz->uk_ppera;
5496 			uth.uth_limit = z->uz_max_items;
5497 			uth.uth_keg_free = kfree;
5498 
5499 			/*
5500 			 * A zone is secondary is it is not the first entry
5501 			 * on the keg's zone list.
5502 			 */
5503 			if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
5504 			    (LIST_FIRST(&kz->uk_zones) != z))
5505 				uth.uth_zone_flags = UTH_ZONE_SECONDARY;
5506 			uma_vm_zone_stats(&uth, z, &sbuf, ups,
5507 			    kz->uk_flags & UMA_ZFLAG_INTERNAL);
5508 			(void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5509 			for (i = 0; i < mp_maxid + 1; i++)
5510 				(void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5511 		}
5512 	}
5513 	LIST_FOREACH(z, &uma_cachezones, uz_link) {
5514 		bzero(&uth, sizeof(uth));
5515 		strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5516 		uth.uth_size = z->uz_size;
5517 		uma_vm_zone_stats(&uth, z, &sbuf, ups, false);
5518 		(void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5519 		for (i = 0; i < mp_maxid + 1; i++)
5520 			(void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5521 	}
5522 
5523 	rw_runlock(&uma_rwlock);
5524 	error = sbuf_finish(&sbuf);
5525 	sbuf_delete(&sbuf);
5526 	free(ups, M_TEMP);
5527 	return (error);
5528 }
5529 
5530 int
5531 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS)
5532 {
5533 	uma_zone_t zone = *(uma_zone_t *)arg1;
5534 	int error, max;
5535 
5536 	max = uma_zone_get_max(zone);
5537 	error = sysctl_handle_int(oidp, &max, 0, req);
5538 	if (error || !req->newptr)
5539 		return (error);
5540 
5541 	uma_zone_set_max(zone, max);
5542 
5543 	return (0);
5544 }
5545 
5546 int
5547 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS)
5548 {
5549 	uma_zone_t zone;
5550 	int cur;
5551 
5552 	/*
5553 	 * Some callers want to add sysctls for global zones that
5554 	 * may not yet exist so they pass a pointer to a pointer.
5555 	 */
5556 	if (arg2 == 0)
5557 		zone = *(uma_zone_t *)arg1;
5558 	else
5559 		zone = arg1;
5560 	cur = uma_zone_get_cur(zone);
5561 	return (sysctl_handle_int(oidp, &cur, 0, req));
5562 }
5563 
5564 static int
5565 sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS)
5566 {
5567 	uma_zone_t zone = arg1;
5568 	uint64_t cur;
5569 
5570 	cur = uma_zone_get_allocs(zone);
5571 	return (sysctl_handle_64(oidp, &cur, 0, req));
5572 }
5573 
5574 static int
5575 sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS)
5576 {
5577 	uma_zone_t zone = arg1;
5578 	uint64_t cur;
5579 
5580 	cur = uma_zone_get_frees(zone);
5581 	return (sysctl_handle_64(oidp, &cur, 0, req));
5582 }
5583 
5584 static int
5585 sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS)
5586 {
5587 	struct sbuf sbuf;
5588 	uma_zone_t zone = arg1;
5589 	int error;
5590 
5591 	sbuf_new_for_sysctl(&sbuf, NULL, 0, req);
5592 	if (zone->uz_flags != 0)
5593 		sbuf_printf(&sbuf, "0x%b", zone->uz_flags, PRINT_UMA_ZFLAGS);
5594 	else
5595 		sbuf_printf(&sbuf, "0");
5596 	error = sbuf_finish(&sbuf);
5597 	sbuf_delete(&sbuf);
5598 
5599 	return (error);
5600 }
5601 
5602 static int
5603 sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS)
5604 {
5605 	uma_keg_t keg = arg1;
5606 	int avail, effpct, total;
5607 
5608 	total = keg->uk_ppera * PAGE_SIZE;
5609 	if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0)
5610 		total += slabzone(keg->uk_ipers)->uz_keg->uk_rsize;
5611 	/*
5612 	 * We consider the client's requested size and alignment here, not the
5613 	 * real size determination uk_rsize, because we also adjust the real
5614 	 * size for internal implementation reasons (max bitset size).
5615 	 */
5616 	avail = keg->uk_ipers * roundup2(keg->uk_size, keg->uk_align + 1);
5617 	if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
5618 		avail *= mp_maxid + 1;
5619 	effpct = 100 * avail / total;
5620 	return (sysctl_handle_int(oidp, &effpct, 0, req));
5621 }
5622 
5623 static int
5624 sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS)
5625 {
5626 	uma_zone_t zone = arg1;
5627 	uint64_t cur;
5628 
5629 	cur = UZ_ITEMS_COUNT(atomic_load_64(&zone->uz_items));
5630 	return (sysctl_handle_64(oidp, &cur, 0, req));
5631 }
5632 
5633 #ifdef INVARIANTS
5634 static uma_slab_t
5635 uma_dbg_getslab(uma_zone_t zone, void *item)
5636 {
5637 	uma_slab_t slab;
5638 	uma_keg_t keg;
5639 	uint8_t *mem;
5640 
5641 	/*
5642 	 * It is safe to return the slab here even though the
5643 	 * zone is unlocked because the item's allocation state
5644 	 * essentially holds a reference.
5645 	 */
5646 	mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
5647 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5648 		return (NULL);
5649 	if (zone->uz_flags & UMA_ZFLAG_VTOSLAB)
5650 		return (vtoslab((vm_offset_t)mem));
5651 	keg = zone->uz_keg;
5652 	if ((keg->uk_flags & UMA_ZFLAG_HASH) == 0)
5653 		return ((uma_slab_t)(mem + keg->uk_pgoff));
5654 	KEG_LOCK(keg, 0);
5655 	slab = hash_sfind(&keg->uk_hash, mem);
5656 	KEG_UNLOCK(keg, 0);
5657 
5658 	return (slab);
5659 }
5660 
5661 static bool
5662 uma_dbg_zskip(uma_zone_t zone, void *mem)
5663 {
5664 
5665 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5666 		return (true);
5667 
5668 	return (uma_dbg_kskip(zone->uz_keg, mem));
5669 }
5670 
5671 static bool
5672 uma_dbg_kskip(uma_keg_t keg, void *mem)
5673 {
5674 	uintptr_t idx;
5675 
5676 	if (dbg_divisor == 0)
5677 		return (true);
5678 
5679 	if (dbg_divisor == 1)
5680 		return (false);
5681 
5682 	idx = (uintptr_t)mem >> PAGE_SHIFT;
5683 	if (keg->uk_ipers > 1) {
5684 		idx *= keg->uk_ipers;
5685 		idx += ((uintptr_t)mem & PAGE_MASK) / keg->uk_rsize;
5686 	}
5687 
5688 	if ((idx / dbg_divisor) * dbg_divisor != idx) {
5689 		counter_u64_add(uma_skip_cnt, 1);
5690 		return (true);
5691 	}
5692 	counter_u64_add(uma_dbg_cnt, 1);
5693 
5694 	return (false);
5695 }
5696 
5697 /*
5698  * Set up the slab's freei data such that uma_dbg_free can function.
5699  *
5700  */
5701 static void
5702 uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item)
5703 {
5704 	uma_keg_t keg;
5705 	int freei;
5706 
5707 	if (slab == NULL) {
5708 		slab = uma_dbg_getslab(zone, item);
5709 		if (slab == NULL)
5710 			panic("uma: item %p did not belong to zone %s",
5711 			    item, zone->uz_name);
5712 	}
5713 	keg = zone->uz_keg;
5714 	freei = slab_item_index(slab, keg, item);
5715 
5716 	if (BIT_TEST_SET_ATOMIC(keg->uk_ipers, freei,
5717 	    slab_dbg_bits(slab, keg)))
5718 		panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)",
5719 		    item, zone, zone->uz_name, slab, freei);
5720 }
5721 
5722 /*
5723  * Verifies freed addresses.  Checks for alignment, valid slab membership
5724  * and duplicate frees.
5725  *
5726  */
5727 static void
5728 uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item)
5729 {
5730 	uma_keg_t keg;
5731 	int freei;
5732 
5733 	if (slab == NULL) {
5734 		slab = uma_dbg_getslab(zone, item);
5735 		if (slab == NULL)
5736 			panic("uma: Freed item %p did not belong to zone %s",
5737 			    item, zone->uz_name);
5738 	}
5739 	keg = zone->uz_keg;
5740 	freei = slab_item_index(slab, keg, item);
5741 
5742 	if (freei >= keg->uk_ipers)
5743 		panic("Invalid free of %p from zone %p(%s) slab %p(%d)",
5744 		    item, zone, zone->uz_name, slab, freei);
5745 
5746 	if (slab_item(slab, keg, freei) != item)
5747 		panic("Unaligned free of %p from zone %p(%s) slab %p(%d)",
5748 		    item, zone, zone->uz_name, slab, freei);
5749 
5750 	if (!BIT_TEST_CLR_ATOMIC(keg->uk_ipers, freei,
5751 	    slab_dbg_bits(slab, keg)))
5752 		panic("Duplicate free of %p from zone %p(%s) slab %p(%d)",
5753 		    item, zone, zone->uz_name, slab, freei);
5754 }
5755 #endif /* INVARIANTS */
5756 
5757 #ifdef DDB
5758 static int64_t
5759 get_uma_stats(uma_keg_t kz, uma_zone_t z, uint64_t *allocs, uint64_t *used,
5760     uint64_t *sleeps, long *cachefree, uint64_t *xdomain)
5761 {
5762 	uint64_t frees;
5763 	int i;
5764 
5765 	if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
5766 		*allocs = counter_u64_fetch(z->uz_allocs);
5767 		frees = counter_u64_fetch(z->uz_frees);
5768 		*sleeps = z->uz_sleeps;
5769 		*cachefree = 0;
5770 		*xdomain = 0;
5771 	} else
5772 		uma_zone_sumstat(z, cachefree, allocs, &frees, sleeps,
5773 		    xdomain);
5774 	for (i = 0; i < vm_ndomains; i++) {
5775 		*cachefree += ZDOM_GET(z, i)->uzd_nitems;
5776 		if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
5777 		    (LIST_FIRST(&kz->uk_zones) != z)))
5778 			*cachefree += kz->uk_domain[i].ud_free_items;
5779 	}
5780 	*used = *allocs - frees;
5781 	return (((int64_t)*used + *cachefree) * kz->uk_size);
5782 }
5783 
5784 DB_SHOW_COMMAND(uma, db_show_uma)
5785 {
5786 	const char *fmt_hdr, *fmt_entry;
5787 	uma_keg_t kz;
5788 	uma_zone_t z;
5789 	uint64_t allocs, used, sleeps, xdomain;
5790 	long cachefree;
5791 	/* variables for sorting */
5792 	uma_keg_t cur_keg;
5793 	uma_zone_t cur_zone, last_zone;
5794 	int64_t cur_size, last_size, size;
5795 	int ties;
5796 
5797 	/* /i option produces machine-parseable CSV output */
5798 	if (modif[0] == 'i') {
5799 		fmt_hdr = "%s,%s,%s,%s,%s,%s,%s,%s,%s\n";
5800 		fmt_entry = "\"%s\",%ju,%jd,%ld,%ju,%ju,%u,%jd,%ju\n";
5801 	} else {
5802 		fmt_hdr = "%18s %6s %7s %7s %11s %7s %7s %10s %8s\n";
5803 		fmt_entry = "%18s %6ju %7jd %7ld %11ju %7ju %7u %10jd %8ju\n";
5804 	}
5805 
5806 	db_printf(fmt_hdr, "Zone", "Size", "Used", "Free", "Requests",
5807 	    "Sleeps", "Bucket", "Total Mem", "XFree");
5808 
5809 	/* Sort the zones with largest size first. */
5810 	last_zone = NULL;
5811 	last_size = INT64_MAX;
5812 	for (;;) {
5813 		cur_zone = NULL;
5814 		cur_size = -1;
5815 		ties = 0;
5816 		LIST_FOREACH(kz, &uma_kegs, uk_link) {
5817 			LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5818 				/*
5819 				 * In the case of size ties, print out zones
5820 				 * in the order they are encountered.  That is,
5821 				 * when we encounter the most recently output
5822 				 * zone, we have already printed all preceding
5823 				 * ties, and we must print all following ties.
5824 				 */
5825 				if (z == last_zone) {
5826 					ties = 1;
5827 					continue;
5828 				}
5829 				size = get_uma_stats(kz, z, &allocs, &used,
5830 				    &sleeps, &cachefree, &xdomain);
5831 				if (size > cur_size && size < last_size + ties)
5832 				{
5833 					cur_size = size;
5834 					cur_zone = z;
5835 					cur_keg = kz;
5836 				}
5837 			}
5838 		}
5839 		if (cur_zone == NULL)
5840 			break;
5841 
5842 		size = get_uma_stats(cur_keg, cur_zone, &allocs, &used,
5843 		    &sleeps, &cachefree, &xdomain);
5844 		db_printf(fmt_entry, cur_zone->uz_name,
5845 		    (uintmax_t)cur_keg->uk_size, (intmax_t)used, cachefree,
5846 		    (uintmax_t)allocs, (uintmax_t)sleeps,
5847 		    (unsigned)cur_zone->uz_bucket_size, (intmax_t)size,
5848 		    xdomain);
5849 
5850 		if (db_pager_quit)
5851 			return;
5852 		last_zone = cur_zone;
5853 		last_size = cur_size;
5854 	}
5855 }
5856 
5857 DB_SHOW_COMMAND(umacache, db_show_umacache)
5858 {
5859 	uma_zone_t z;
5860 	uint64_t allocs, frees;
5861 	long cachefree;
5862 	int i;
5863 
5864 	db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
5865 	    "Requests", "Bucket");
5866 	LIST_FOREACH(z, &uma_cachezones, uz_link) {
5867 		uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL, NULL);
5868 		for (i = 0; i < vm_ndomains; i++)
5869 			cachefree += ZDOM_GET(z, i)->uzd_nitems;
5870 		db_printf("%18s %8ju %8jd %8ld %12ju %8u\n",
5871 		    z->uz_name, (uintmax_t)z->uz_size,
5872 		    (intmax_t)(allocs - frees), cachefree,
5873 		    (uintmax_t)allocs, z->uz_bucket_size);
5874 		if (db_pager_quit)
5875 			return;
5876 	}
5877 }
5878 #endif	/* DDB */
5879