xref: /freebsd/sys/vm/uma_core.c (revision c1d255d3ffdbe447de3ab875bf4e7d7accc5bfc5)
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 *, 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 
1055 	/*
1056 	 * Avoid the lock if possible.
1057 	 */
1058 	zdom = ZDOM_GET(zone, domain);
1059 	if (zdom->uzd_nitems == 0)
1060 		return (NULL);
1061 
1062 	if ((cache_uz_flags(cache) & UMA_ZONE_SMR) != 0 &&
1063 	    !smr_poll(zone->uz_smr, zdom->uzd_seq, false))
1064 		return (NULL);
1065 
1066 	/*
1067 	 * Check the zone's cache of buckets.
1068 	 */
1069 	zdom = zone_domain_lock(zone, domain);
1070 	if ((bucket = zone_fetch_bucket(zone, zdom, false)) != NULL)
1071 		return (bucket);
1072 	ZDOM_UNLOCK(zdom);
1073 
1074 	return (NULL);
1075 }
1076 
1077 static void
1078 zone_log_warning(uma_zone_t zone)
1079 {
1080 	static const struct timeval warninterval = { 300, 0 };
1081 
1082 	if (!zone_warnings || zone->uz_warning == NULL)
1083 		return;
1084 
1085 	if (ratecheck(&zone->uz_ratecheck, &warninterval))
1086 		printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning);
1087 }
1088 
1089 static inline void
1090 zone_maxaction(uma_zone_t zone)
1091 {
1092 
1093 	if (zone->uz_maxaction.ta_func != NULL)
1094 		taskqueue_enqueue(taskqueue_thread, &zone->uz_maxaction);
1095 }
1096 
1097 /*
1098  * Routine called by timeout which is used to fire off some time interval
1099  * based calculations.  (stats, hash size, etc.)
1100  *
1101  * Arguments:
1102  *	arg   Unused
1103  *
1104  * Returns:
1105  *	Nothing
1106  */
1107 static void
1108 uma_timeout(void *unused)
1109 {
1110 	bucket_enable();
1111 	zone_foreach(zone_timeout, NULL);
1112 
1113 	/* Reschedule this event */
1114 	callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
1115 }
1116 
1117 /*
1118  * Update the working set size estimates for the zone's bucket cache.
1119  * The constants chosen here are somewhat arbitrary.
1120  */
1121 static void
1122 zone_domain_update_wss(uma_zone_domain_t zdom)
1123 {
1124 	long m;
1125 
1126 	ZDOM_LOCK_ASSERT(zdom);
1127 	MPASS(zdom->uzd_imax >= zdom->uzd_nitems);
1128 	MPASS(zdom->uzd_nitems >= zdom->uzd_bimin);
1129 	MPASS(zdom->uzd_bimin >= zdom->uzd_imin);
1130 
1131 	/*
1132 	 * Estimate WSS as modified moving average of biggest allocation
1133 	 * batches for each period over few minutes (UMA_TIMEOUT of 20s).
1134 	 */
1135 	zdom->uzd_wss = lmax(zdom->uzd_wss * 3 / 4,
1136 	    zdom->uzd_imax - zdom->uzd_bimin);
1137 
1138 	/*
1139 	 * Estimate longtime minimum item count as a combination of recent
1140 	 * minimum item count, adjusted by WSS for safety, and the modified
1141 	 * moving average over the last several hours (UMA_TIMEOUT of 20s).
1142 	 * timin measures time since limin tried to go negative, that means
1143 	 * we were dangerously close to or got out of cache.
1144 	 */
1145 	m = zdom->uzd_imin - zdom->uzd_wss;
1146 	if (m >= 0) {
1147 		if (zdom->uzd_limin >= m)
1148 			zdom->uzd_limin = m;
1149 		else
1150 			zdom->uzd_limin = (m + zdom->uzd_limin * 255) / 256;
1151 		zdom->uzd_timin++;
1152 	} else {
1153 		zdom->uzd_limin = 0;
1154 		zdom->uzd_timin = 0;
1155 	}
1156 
1157 	/* To reduce period edge effects on WSS keep half of the imax. */
1158 	atomic_subtract_long(&zdom->uzd_imax,
1159 	    (zdom->uzd_imax - zdom->uzd_nitems + 1) / 2);
1160 	zdom->uzd_imin = zdom->uzd_bimin = zdom->uzd_nitems;
1161 }
1162 
1163 /*
1164  * Routine to perform timeout driven calculations.  This expands the
1165  * hashes and does per cpu statistics aggregation.
1166  *
1167  *  Returns nothing.
1168  */
1169 static void
1170 zone_timeout(uma_zone_t zone, void *unused)
1171 {
1172 	uma_keg_t keg;
1173 	u_int slabs, pages;
1174 
1175 	if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
1176 		goto trim;
1177 
1178 	keg = zone->uz_keg;
1179 
1180 	/*
1181 	 * Hash zones are non-numa by definition so the first domain
1182 	 * is the only one present.
1183 	 */
1184 	KEG_LOCK(keg, 0);
1185 	pages = keg->uk_domain[0].ud_pages;
1186 
1187 	/*
1188 	 * Expand the keg hash table.
1189 	 *
1190 	 * This is done if the number of slabs is larger than the hash size.
1191 	 * What I'm trying to do here is completely reduce collisions.  This
1192 	 * may be a little aggressive.  Should I allow for two collisions max?
1193 	 */
1194 	if ((slabs = pages / keg->uk_ppera) > keg->uk_hash.uh_hashsize) {
1195 		struct uma_hash newhash;
1196 		struct uma_hash oldhash;
1197 		int ret;
1198 
1199 		/*
1200 		 * This is so involved because allocating and freeing
1201 		 * while the keg lock is held will lead to deadlock.
1202 		 * I have to do everything in stages and check for
1203 		 * races.
1204 		 */
1205 		KEG_UNLOCK(keg, 0);
1206 		ret = hash_alloc(&newhash, 1 << fls(slabs));
1207 		KEG_LOCK(keg, 0);
1208 		if (ret) {
1209 			if (hash_expand(&keg->uk_hash, &newhash)) {
1210 				oldhash = keg->uk_hash;
1211 				keg->uk_hash = newhash;
1212 			} else
1213 				oldhash = newhash;
1214 
1215 			KEG_UNLOCK(keg, 0);
1216 			hash_free(&oldhash);
1217 			goto trim;
1218 		}
1219 	}
1220 	KEG_UNLOCK(keg, 0);
1221 
1222 trim:
1223 	/* Trim caches not used for a long time. */
1224 	for (int i = 0; i < vm_ndomains; i++) {
1225 		if (bucket_cache_reclaim_domain(zone, false, false, i) &&
1226 		    (zone->uz_flags & UMA_ZFLAG_CACHE) == 0)
1227 			keg_drain(zone->uz_keg, i);
1228 	}
1229 }
1230 
1231 /*
1232  * Allocate and zero fill the next sized hash table from the appropriate
1233  * backing store.
1234  *
1235  * Arguments:
1236  *	hash  A new hash structure with the old hash size in uh_hashsize
1237  *
1238  * Returns:
1239  *	1 on success and 0 on failure.
1240  */
1241 static int
1242 hash_alloc(struct uma_hash *hash, u_int size)
1243 {
1244 	size_t alloc;
1245 
1246 	KASSERT(powerof2(size), ("hash size must be power of 2"));
1247 	if (size > UMA_HASH_SIZE_INIT)  {
1248 		hash->uh_hashsize = size;
1249 		alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
1250 		hash->uh_slab_hash = malloc(alloc, M_UMAHASH, M_NOWAIT);
1251 	} else {
1252 		alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
1253 		hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
1254 		    UMA_ANYDOMAIN, M_WAITOK);
1255 		hash->uh_hashsize = UMA_HASH_SIZE_INIT;
1256 	}
1257 	if (hash->uh_slab_hash) {
1258 		bzero(hash->uh_slab_hash, alloc);
1259 		hash->uh_hashmask = hash->uh_hashsize - 1;
1260 		return (1);
1261 	}
1262 
1263 	return (0);
1264 }
1265 
1266 /*
1267  * Expands the hash table for HASH zones.  This is done from zone_timeout
1268  * to reduce collisions.  This must not be done in the regular allocation
1269  * path, otherwise, we can recurse on the vm while allocating pages.
1270  *
1271  * Arguments:
1272  *	oldhash  The hash you want to expand
1273  *	newhash  The hash structure for the new table
1274  *
1275  * Returns:
1276  *	Nothing
1277  *
1278  * Discussion:
1279  */
1280 static int
1281 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
1282 {
1283 	uma_hash_slab_t slab;
1284 	u_int hval;
1285 	u_int idx;
1286 
1287 	if (!newhash->uh_slab_hash)
1288 		return (0);
1289 
1290 	if (oldhash->uh_hashsize >= newhash->uh_hashsize)
1291 		return (0);
1292 
1293 	/*
1294 	 * I need to investigate hash algorithms for resizing without a
1295 	 * full rehash.
1296 	 */
1297 
1298 	for (idx = 0; idx < oldhash->uh_hashsize; idx++)
1299 		while (!LIST_EMPTY(&oldhash->uh_slab_hash[idx])) {
1300 			slab = LIST_FIRST(&oldhash->uh_slab_hash[idx]);
1301 			LIST_REMOVE(slab, uhs_hlink);
1302 			hval = UMA_HASH(newhash, slab->uhs_data);
1303 			LIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
1304 			    slab, uhs_hlink);
1305 		}
1306 
1307 	return (1);
1308 }
1309 
1310 /*
1311  * Free the hash bucket to the appropriate backing store.
1312  *
1313  * Arguments:
1314  *	slab_hash  The hash bucket we're freeing
1315  *	hashsize   The number of entries in that hash bucket
1316  *
1317  * Returns:
1318  *	Nothing
1319  */
1320 static void
1321 hash_free(struct uma_hash *hash)
1322 {
1323 	if (hash->uh_slab_hash == NULL)
1324 		return;
1325 	if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
1326 		zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE);
1327 	else
1328 		free(hash->uh_slab_hash, M_UMAHASH);
1329 }
1330 
1331 /*
1332  * Frees all outstanding items in a bucket
1333  *
1334  * Arguments:
1335  *	zone   The zone to free to, must be unlocked.
1336  *	bucket The free/alloc bucket with items.
1337  *
1338  * Returns:
1339  *	Nothing
1340  */
1341 static void
1342 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
1343 {
1344 	int i;
1345 
1346 	if (bucket->ub_cnt == 0)
1347 		return;
1348 
1349 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
1350 	    bucket->ub_seq != SMR_SEQ_INVALID) {
1351 		smr_wait(zone->uz_smr, bucket->ub_seq);
1352 		bucket->ub_seq = SMR_SEQ_INVALID;
1353 		for (i = 0; i < bucket->ub_cnt; i++)
1354 			item_dtor(zone, bucket->ub_bucket[i],
1355 			    zone->uz_size, NULL, SKIP_NONE);
1356 	}
1357 	if (zone->uz_fini)
1358 		for (i = 0; i < bucket->ub_cnt; i++) {
1359 			kasan_mark_item_valid(zone, bucket->ub_bucket[i]);
1360 			zone->uz_fini(bucket->ub_bucket[i], zone->uz_size);
1361 			kasan_mark_item_invalid(zone, bucket->ub_bucket[i]);
1362 		}
1363 	zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt);
1364 	if (zone->uz_max_items > 0)
1365 		zone_free_limit(zone, bucket->ub_cnt);
1366 #ifdef INVARIANTS
1367 	bzero(bucket->ub_bucket, sizeof(void *) * bucket->ub_cnt);
1368 #endif
1369 	bucket->ub_cnt = 0;
1370 }
1371 
1372 /*
1373  * Drains the per cpu caches for a zone.
1374  *
1375  * NOTE: This may only be called while the zone is being torn down, and not
1376  * during normal operation.  This is necessary in order that we do not have
1377  * to migrate CPUs to drain the per-CPU caches.
1378  *
1379  * Arguments:
1380  *	zone     The zone to drain, must be unlocked.
1381  *
1382  * Returns:
1383  *	Nothing
1384  */
1385 static void
1386 cache_drain(uma_zone_t zone)
1387 {
1388 	uma_cache_t cache;
1389 	uma_bucket_t bucket;
1390 	smr_seq_t seq;
1391 	int cpu;
1392 
1393 	/*
1394 	 * XXX: It is safe to not lock the per-CPU caches, because we're
1395 	 * tearing down the zone anyway.  I.e., there will be no further use
1396 	 * of the caches at this point.
1397 	 *
1398 	 * XXX: It would good to be able to assert that the zone is being
1399 	 * torn down to prevent improper use of cache_drain().
1400 	 */
1401 	seq = SMR_SEQ_INVALID;
1402 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
1403 		seq = smr_advance(zone->uz_smr);
1404 	CPU_FOREACH(cpu) {
1405 		cache = &zone->uz_cpu[cpu];
1406 		bucket = cache_bucket_unload_alloc(cache);
1407 		if (bucket != NULL)
1408 			bucket_free(zone, bucket, NULL);
1409 		bucket = cache_bucket_unload_free(cache);
1410 		if (bucket != NULL) {
1411 			bucket->ub_seq = seq;
1412 			bucket_free(zone, bucket, NULL);
1413 		}
1414 		bucket = cache_bucket_unload_cross(cache);
1415 		if (bucket != NULL) {
1416 			bucket->ub_seq = seq;
1417 			bucket_free(zone, bucket, NULL);
1418 		}
1419 	}
1420 	bucket_cache_reclaim(zone, true, UMA_ANYDOMAIN);
1421 }
1422 
1423 static void
1424 cache_shrink(uma_zone_t zone, void *unused)
1425 {
1426 
1427 	if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1428 		return;
1429 
1430 	ZONE_LOCK(zone);
1431 	zone->uz_bucket_size =
1432 	    (zone->uz_bucket_size_min + zone->uz_bucket_size) / 2;
1433 	ZONE_UNLOCK(zone);
1434 }
1435 
1436 static void
1437 cache_drain_safe_cpu(uma_zone_t zone, void *unused)
1438 {
1439 	uma_cache_t cache;
1440 	uma_bucket_t b1, b2, b3;
1441 	int domain;
1442 
1443 	if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1444 		return;
1445 
1446 	b1 = b2 = b3 = NULL;
1447 	critical_enter();
1448 	cache = &zone->uz_cpu[curcpu];
1449 	domain = PCPU_GET(domain);
1450 	b1 = cache_bucket_unload_alloc(cache);
1451 
1452 	/*
1453 	 * Don't flush SMR zone buckets.  This leaves the zone without a
1454 	 * bucket and forces every free to synchronize().
1455 	 */
1456 	if ((zone->uz_flags & UMA_ZONE_SMR) == 0) {
1457 		b2 = cache_bucket_unload_free(cache);
1458 		b3 = cache_bucket_unload_cross(cache);
1459 	}
1460 	critical_exit();
1461 
1462 	if (b1 != NULL)
1463 		zone_free_bucket(zone, b1, NULL, domain, false);
1464 	if (b2 != NULL)
1465 		zone_free_bucket(zone, b2, NULL, domain, false);
1466 	if (b3 != NULL) {
1467 		/* Adjust the domain so it goes to zone_free_cross. */
1468 		domain = (domain + 1) % vm_ndomains;
1469 		zone_free_bucket(zone, b3, NULL, domain, false);
1470 	}
1471 }
1472 
1473 /*
1474  * Safely drain per-CPU caches of a zone(s) to alloc bucket.
1475  * This is an expensive call because it needs to bind to all CPUs
1476  * one by one and enter a critical section on each of them in order
1477  * to safely access their cache buckets.
1478  * Zone lock must not be held on call this function.
1479  */
1480 static void
1481 pcpu_cache_drain_safe(uma_zone_t zone)
1482 {
1483 	int cpu;
1484 
1485 	/*
1486 	 * Polite bucket sizes shrinking was not enough, shrink aggressively.
1487 	 */
1488 	if (zone)
1489 		cache_shrink(zone, NULL);
1490 	else
1491 		zone_foreach(cache_shrink, NULL);
1492 
1493 	CPU_FOREACH(cpu) {
1494 		thread_lock(curthread);
1495 		sched_bind(curthread, cpu);
1496 		thread_unlock(curthread);
1497 
1498 		if (zone)
1499 			cache_drain_safe_cpu(zone, NULL);
1500 		else
1501 			zone_foreach(cache_drain_safe_cpu, NULL);
1502 	}
1503 	thread_lock(curthread);
1504 	sched_unbind(curthread);
1505 	thread_unlock(curthread);
1506 }
1507 
1508 /*
1509  * Reclaim cached buckets from a zone.  All buckets are reclaimed if the caller
1510  * requested a drain, otherwise the per-domain caches are trimmed to either
1511  * estimated working set size.
1512  */
1513 static bool
1514 bucket_cache_reclaim_domain(uma_zone_t zone, bool drain, bool trim, int domain)
1515 {
1516 	uma_zone_domain_t zdom;
1517 	uma_bucket_t bucket;
1518 	long target;
1519 	bool done = false;
1520 
1521 	/*
1522 	 * The cross bucket is partially filled and not part of
1523 	 * the item count.  Reclaim it individually here.
1524 	 */
1525 	zdom = ZDOM_GET(zone, domain);
1526 	if ((zone->uz_flags & UMA_ZONE_SMR) == 0 || drain) {
1527 		ZONE_CROSS_LOCK(zone);
1528 		bucket = zdom->uzd_cross;
1529 		zdom->uzd_cross = NULL;
1530 		ZONE_CROSS_UNLOCK(zone);
1531 		if (bucket != NULL)
1532 			bucket_free(zone, bucket, NULL);
1533 	}
1534 
1535 	/*
1536 	 * If we were asked to drain the zone, we are done only once
1537 	 * this bucket cache is empty.  If trim, we reclaim items in
1538 	 * excess of the zone's estimated working set size.  Multiple
1539 	 * consecutive calls will shrink the WSS and so reclaim more.
1540 	 * If neither drain nor trim, then voluntarily reclaim 1/4
1541 	 * (to reduce first spike) of items not used for a long time.
1542 	 */
1543 	ZDOM_LOCK(zdom);
1544 	zone_domain_update_wss(zdom);
1545 	if (drain)
1546 		target = 0;
1547 	else if (trim)
1548 		target = zdom->uzd_wss;
1549 	else if (zdom->uzd_timin > 900 / UMA_TIMEOUT)
1550 		target = zdom->uzd_nitems - zdom->uzd_limin / 4;
1551 	else {
1552 		ZDOM_UNLOCK(zdom);
1553 		return (done);
1554 	}
1555 	while ((bucket = STAILQ_FIRST(&zdom->uzd_buckets)) != NULL &&
1556 	    zdom->uzd_nitems >= target + bucket->ub_cnt) {
1557 		bucket = zone_fetch_bucket(zone, zdom, true);
1558 		if (bucket == NULL)
1559 			break;
1560 		bucket_free(zone, bucket, NULL);
1561 		done = true;
1562 		ZDOM_LOCK(zdom);
1563 	}
1564 	ZDOM_UNLOCK(zdom);
1565 	return (done);
1566 }
1567 
1568 static void
1569 bucket_cache_reclaim(uma_zone_t zone, bool drain, int domain)
1570 {
1571 	int i;
1572 
1573 	/*
1574 	 * Shrink the zone bucket size to ensure that the per-CPU caches
1575 	 * don't grow too large.
1576 	 */
1577 	if (zone->uz_bucket_size > zone->uz_bucket_size_min)
1578 		zone->uz_bucket_size--;
1579 
1580 	if (domain != UMA_ANYDOMAIN &&
1581 	    (zone->uz_flags & UMA_ZONE_ROUNDROBIN) == 0) {
1582 		bucket_cache_reclaim_domain(zone, drain, true, domain);
1583 	} else {
1584 		for (i = 0; i < vm_ndomains; i++)
1585 			bucket_cache_reclaim_domain(zone, drain, true, i);
1586 	}
1587 }
1588 
1589 static void
1590 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start)
1591 {
1592 	uint8_t *mem;
1593 	size_t size;
1594 	int i;
1595 	uint8_t flags;
1596 
1597 	CTR4(KTR_UMA, "keg_free_slab keg %s(%p) slab %p, returning %d bytes",
1598 	    keg->uk_name, keg, slab, PAGE_SIZE * keg->uk_ppera);
1599 
1600 	mem = slab_data(slab, keg);
1601 	size = PAGE_SIZE * keg->uk_ppera;
1602 
1603 	kasan_mark_slab_valid(keg, mem);
1604 	if (keg->uk_fini != NULL) {
1605 		for (i = start - 1; i > -1; i--)
1606 #ifdef INVARIANTS
1607 		/*
1608 		 * trash_fini implies that dtor was trash_dtor. trash_fini
1609 		 * would check that memory hasn't been modified since free,
1610 		 * which executed trash_dtor.
1611 		 * That's why we need to run uma_dbg_kskip() check here,
1612 		 * albeit we don't make skip check for other init/fini
1613 		 * invocations.
1614 		 */
1615 		if (!uma_dbg_kskip(keg, slab_item(slab, keg, i)) ||
1616 		    keg->uk_fini != trash_fini)
1617 #endif
1618 			keg->uk_fini(slab_item(slab, keg, i), keg->uk_size);
1619 	}
1620 	flags = slab->us_flags;
1621 	if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) {
1622 		zone_free_item(slabzone(keg->uk_ipers), slab_tohashslab(slab),
1623 		    NULL, SKIP_NONE);
1624 	}
1625 	keg->uk_freef(mem, size, flags);
1626 	uma_total_dec(size);
1627 }
1628 
1629 static void
1630 keg_drain_domain(uma_keg_t keg, int domain)
1631 {
1632 	struct slabhead freeslabs;
1633 	uma_domain_t dom;
1634 	uma_slab_t slab, tmp;
1635 	uint32_t i, stofree, stokeep, partial;
1636 
1637 	dom = &keg->uk_domain[domain];
1638 	LIST_INIT(&freeslabs);
1639 
1640 	CTR4(KTR_UMA, "keg_drain %s(%p) domain %d free items: %u",
1641 	    keg->uk_name, keg, domain, dom->ud_free_items);
1642 
1643 	KEG_LOCK(keg, domain);
1644 
1645 	/*
1646 	 * Are the free items in partially allocated slabs sufficient to meet
1647 	 * the reserve? If not, compute the number of fully free slabs that must
1648 	 * be kept.
1649 	 */
1650 	partial = dom->ud_free_items - dom->ud_free_slabs * keg->uk_ipers;
1651 	if (partial < keg->uk_reserve) {
1652 		stokeep = min(dom->ud_free_slabs,
1653 		    howmany(keg->uk_reserve - partial, keg->uk_ipers));
1654 	} else {
1655 		stokeep = 0;
1656 	}
1657 	stofree = dom->ud_free_slabs - stokeep;
1658 
1659 	/*
1660 	 * Partition the free slabs into two sets: those that must be kept in
1661 	 * order to maintain the reserve, and those that may be released back to
1662 	 * the system.  Since one set may be much larger than the other,
1663 	 * populate the smaller of the two sets and swap them if necessary.
1664 	 */
1665 	for (i = min(stofree, stokeep); i > 0; i--) {
1666 		slab = LIST_FIRST(&dom->ud_free_slab);
1667 		LIST_REMOVE(slab, us_link);
1668 		LIST_INSERT_HEAD(&freeslabs, slab, us_link);
1669 	}
1670 	if (stofree > stokeep)
1671 		LIST_SWAP(&freeslabs, &dom->ud_free_slab, uma_slab, us_link);
1672 
1673 	if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0) {
1674 		LIST_FOREACH(slab, &freeslabs, us_link)
1675 			UMA_HASH_REMOVE(&keg->uk_hash, slab);
1676 	}
1677 	dom->ud_free_items -= stofree * keg->uk_ipers;
1678 	dom->ud_free_slabs -= stofree;
1679 	dom->ud_pages -= stofree * keg->uk_ppera;
1680 	KEG_UNLOCK(keg, domain);
1681 
1682 	LIST_FOREACH_SAFE(slab, &freeslabs, us_link, tmp)
1683 		keg_free_slab(keg, slab, keg->uk_ipers);
1684 }
1685 
1686 /*
1687  * Frees pages from a keg back to the system.  This is done on demand from
1688  * the pageout daemon.
1689  *
1690  * Returns nothing.
1691  */
1692 static void
1693 keg_drain(uma_keg_t keg, int domain)
1694 {
1695 	int i;
1696 
1697 	if ((keg->uk_flags & UMA_ZONE_NOFREE) != 0)
1698 		return;
1699 	if (domain != UMA_ANYDOMAIN) {
1700 		keg_drain_domain(keg, domain);
1701 	} else {
1702 		for (i = 0; i < vm_ndomains; i++)
1703 			keg_drain_domain(keg, i);
1704 	}
1705 }
1706 
1707 static void
1708 zone_reclaim(uma_zone_t zone, int domain, int waitok, bool drain)
1709 {
1710 	/*
1711 	 * Count active reclaim operations in order to interlock with
1712 	 * zone_dtor(), which removes the zone from global lists before
1713 	 * attempting to reclaim items itself.
1714 	 *
1715 	 * The zone may be destroyed while sleeping, so only zone_dtor() should
1716 	 * specify M_WAITOK.
1717 	 */
1718 	ZONE_LOCK(zone);
1719 	if (waitok == M_WAITOK) {
1720 		while (zone->uz_reclaimers > 0)
1721 			msleep(zone, ZONE_LOCKPTR(zone), PVM, "zonedrain", 1);
1722 	}
1723 	zone->uz_reclaimers++;
1724 	ZONE_UNLOCK(zone);
1725 	bucket_cache_reclaim(zone, drain, domain);
1726 
1727 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0)
1728 		keg_drain(zone->uz_keg, domain);
1729 	ZONE_LOCK(zone);
1730 	zone->uz_reclaimers--;
1731 	if (zone->uz_reclaimers == 0)
1732 		wakeup(zone);
1733 	ZONE_UNLOCK(zone);
1734 }
1735 
1736 static void
1737 zone_drain(uma_zone_t zone, void *arg)
1738 {
1739 	int domain;
1740 
1741 	domain = (int)(uintptr_t)arg;
1742 	zone_reclaim(zone, domain, M_NOWAIT, true);
1743 }
1744 
1745 static void
1746 zone_trim(uma_zone_t zone, void *arg)
1747 {
1748 	int domain;
1749 
1750 	domain = (int)(uintptr_t)arg;
1751 	zone_reclaim(zone, domain, M_NOWAIT, false);
1752 }
1753 
1754 /*
1755  * Allocate a new slab for a keg and inserts it into the partial slab list.
1756  * The keg should be unlocked on entry.  If the allocation succeeds it will
1757  * be locked on return.
1758  *
1759  * Arguments:
1760  *	flags   Wait flags for the item initialization routine
1761  *	aflags  Wait flags for the slab allocation
1762  *
1763  * Returns:
1764  *	The slab that was allocated or NULL if there is no memory and the
1765  *	caller specified M_NOWAIT.
1766  */
1767 static uma_slab_t
1768 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int domain, int flags,
1769     int aflags)
1770 {
1771 	uma_domain_t dom;
1772 	uma_slab_t slab;
1773 	unsigned long size;
1774 	uint8_t *mem;
1775 	uint8_t sflags;
1776 	int i;
1777 
1778 	KASSERT(domain >= 0 && domain < vm_ndomains,
1779 	    ("keg_alloc_slab: domain %d out of range", domain));
1780 
1781 	slab = NULL;
1782 	mem = NULL;
1783 	if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) {
1784 		uma_hash_slab_t hslab;
1785 		hslab = zone_alloc_item(slabzone(keg->uk_ipers), NULL,
1786 		    domain, aflags);
1787 		if (hslab == NULL)
1788 			goto fail;
1789 		slab = &hslab->uhs_slab;
1790 	}
1791 
1792 	/*
1793 	 * This reproduces the old vm_zone behavior of zero filling pages the
1794 	 * first time they are added to a zone.
1795 	 *
1796 	 * Malloced items are zeroed in uma_zalloc.
1797 	 */
1798 
1799 	if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
1800 		aflags |= M_ZERO;
1801 	else
1802 		aflags &= ~M_ZERO;
1803 
1804 	if (keg->uk_flags & UMA_ZONE_NODUMP)
1805 		aflags |= M_NODUMP;
1806 
1807 	/* zone is passed for legacy reasons. */
1808 	size = keg->uk_ppera * PAGE_SIZE;
1809 	mem = keg->uk_allocf(zone, size, domain, &sflags, aflags);
1810 	if (mem == NULL) {
1811 		if (keg->uk_flags & UMA_ZFLAG_OFFPAGE)
1812 			zone_free_item(slabzone(keg->uk_ipers),
1813 			    slab_tohashslab(slab), NULL, SKIP_NONE);
1814 		goto fail;
1815 	}
1816 	uma_total_inc(size);
1817 
1818 	/* For HASH zones all pages go to the same uma_domain. */
1819 	if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
1820 		domain = 0;
1821 
1822 	/* Point the slab into the allocated memory */
1823 	if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE))
1824 		slab = (uma_slab_t)(mem + keg->uk_pgoff);
1825 	else
1826 		slab_tohashslab(slab)->uhs_data = mem;
1827 
1828 	if (keg->uk_flags & UMA_ZFLAG_VTOSLAB)
1829 		for (i = 0; i < keg->uk_ppera; i++)
1830 			vsetzoneslab((vm_offset_t)mem + (i * PAGE_SIZE),
1831 			    zone, slab);
1832 
1833 	slab->us_freecount = keg->uk_ipers;
1834 	slab->us_flags = sflags;
1835 	slab->us_domain = domain;
1836 
1837 	BIT_FILL(keg->uk_ipers, &slab->us_free);
1838 #ifdef INVARIANTS
1839 	BIT_ZERO(keg->uk_ipers, slab_dbg_bits(slab, keg));
1840 #endif
1841 
1842 	if (keg->uk_init != NULL) {
1843 		for (i = 0; i < keg->uk_ipers; i++)
1844 			if (keg->uk_init(slab_item(slab, keg, i),
1845 			    keg->uk_size, flags) != 0)
1846 				break;
1847 		if (i != keg->uk_ipers) {
1848 			keg_free_slab(keg, slab, i);
1849 			goto fail;
1850 		}
1851 	}
1852 	kasan_mark_slab_invalid(keg, mem);
1853 	KEG_LOCK(keg, domain);
1854 
1855 	CTR3(KTR_UMA, "keg_alloc_slab: allocated slab %p for %s(%p)",
1856 	    slab, keg->uk_name, keg);
1857 
1858 	if (keg->uk_flags & UMA_ZFLAG_HASH)
1859 		UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
1860 
1861 	/*
1862 	 * If we got a slab here it's safe to mark it partially used
1863 	 * and return.  We assume that the caller is going to remove
1864 	 * at least one item.
1865 	 */
1866 	dom = &keg->uk_domain[domain];
1867 	LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
1868 	dom->ud_pages += keg->uk_ppera;
1869 	dom->ud_free_items += keg->uk_ipers;
1870 
1871 	return (slab);
1872 
1873 fail:
1874 	return (NULL);
1875 }
1876 
1877 /*
1878  * This function is intended to be used early on in place of page_alloc().  It
1879  * performs contiguous physical memory allocations and uses a bump allocator for
1880  * KVA, so is usable before the kernel map is initialized.
1881  */
1882 static void *
1883 startup_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1884     int wait)
1885 {
1886 	vm_paddr_t pa;
1887 	vm_page_t m;
1888 	void *mem;
1889 	int pages;
1890 	int i;
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_contig_domain(NULL, 0, domain,
1897 	    malloc2vm_flags(wait) | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED, pages,
1898 	    (vm_paddr_t)0, ~(vm_paddr_t)0, 1, 0, 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__) || defined(__mips__) || \
1905     defined(__riscv) || defined(__powerpc64__)
1906 		if ((wait & M_NODUMP) == 0)
1907 			dump_add_page(pa);
1908 #endif
1909 	}
1910 	/* Allocate KVA and indirectly advance bootmem. */
1911 	mem = (void *)pmap_map(&bootmem, m->phys_addr,
1912 	    m->phys_addr + (pages * PAGE_SIZE), VM_PROT_READ | VM_PROT_WRITE);
1913         if ((wait & M_ZERO) != 0)
1914                 bzero(mem, pages * PAGE_SIZE);
1915 
1916         return (mem);
1917 }
1918 
1919 static void
1920 startup_free(void *mem, vm_size_t bytes)
1921 {
1922 	vm_offset_t va;
1923 	vm_page_t m;
1924 
1925 	va = (vm_offset_t)mem;
1926 	m = PHYS_TO_VM_PAGE(pmap_kextract(va));
1927 
1928 	/*
1929 	 * startup_alloc() returns direct-mapped slabs on some platforms.  Avoid
1930 	 * unmapping ranges of the direct map.
1931 	 */
1932 	if (va >= bootstart && va + bytes <= bootmem)
1933 		pmap_remove(kernel_pmap, va, va + bytes);
1934 	for (; bytes != 0; bytes -= PAGE_SIZE, m++) {
1935 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
1936     defined(__riscv) || defined(__powerpc64__)
1937 		dump_drop_page(VM_PAGE_TO_PHYS(m));
1938 #endif
1939 		vm_page_unwire_noq(m);
1940 		vm_page_free(m);
1941 	}
1942 }
1943 
1944 /*
1945  * Allocates a number of pages from the system
1946  *
1947  * Arguments:
1948  *	bytes  The number of bytes requested
1949  *	wait  Shall we wait?
1950  *
1951  * Returns:
1952  *	A pointer to the alloced memory or possibly
1953  *	NULL if M_NOWAIT is set.
1954  */
1955 static void *
1956 page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1957     int wait)
1958 {
1959 	void *p;	/* Returned page */
1960 
1961 	*pflag = UMA_SLAB_KERNEL;
1962 	p = (void *)kmem_malloc_domainset(DOMAINSET_FIXED(domain), bytes, wait);
1963 
1964 	return (p);
1965 }
1966 
1967 static void *
1968 pcpu_page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1969     int wait)
1970 {
1971 	struct pglist alloctail;
1972 	vm_offset_t addr, zkva;
1973 	int cpu, flags;
1974 	vm_page_t p, p_next;
1975 #ifdef NUMA
1976 	struct pcpu *pc;
1977 #endif
1978 
1979 	MPASS(bytes == (mp_maxid + 1) * PAGE_SIZE);
1980 
1981 	TAILQ_INIT(&alloctail);
1982 	flags = VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | VM_ALLOC_NOOBJ |
1983 	    malloc2vm_flags(wait);
1984 	*pflag = UMA_SLAB_KERNEL;
1985 	for (cpu = 0; cpu <= mp_maxid; cpu++) {
1986 		if (CPU_ABSENT(cpu)) {
1987 			p = vm_page_alloc(NULL, 0, flags);
1988 		} else {
1989 #ifndef NUMA
1990 			p = vm_page_alloc(NULL, 0, flags);
1991 #else
1992 			pc = pcpu_find(cpu);
1993 			if (__predict_false(VM_DOMAIN_EMPTY(pc->pc_domain)))
1994 				p = NULL;
1995 			else
1996 				p = vm_page_alloc_domain(NULL, 0,
1997 				    pc->pc_domain, flags);
1998 			if (__predict_false(p == NULL))
1999 				p = vm_page_alloc(NULL, 0, flags);
2000 #endif
2001 		}
2002 		if (__predict_false(p == NULL))
2003 			goto fail;
2004 		TAILQ_INSERT_TAIL(&alloctail, p, listq);
2005 	}
2006 	if ((addr = kva_alloc(bytes)) == 0)
2007 		goto fail;
2008 	zkva = addr;
2009 	TAILQ_FOREACH(p, &alloctail, listq) {
2010 		pmap_qenter(zkva, &p, 1);
2011 		zkva += PAGE_SIZE;
2012 	}
2013 	return ((void*)addr);
2014 fail:
2015 	TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
2016 		vm_page_unwire_noq(p);
2017 		vm_page_free(p);
2018 	}
2019 	return (NULL);
2020 }
2021 
2022 /*
2023  * Allocates a number of pages from within an object
2024  *
2025  * Arguments:
2026  *	bytes  The number of bytes requested
2027  *	wait   Shall we wait?
2028  *
2029  * Returns:
2030  *	A pointer to the alloced memory or possibly
2031  *	NULL if M_NOWAIT is set.
2032  */
2033 static void *
2034 noobj_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags,
2035     int wait)
2036 {
2037 	TAILQ_HEAD(, vm_page) alloctail;
2038 	u_long npages;
2039 	vm_offset_t retkva, zkva;
2040 	vm_page_t p, p_next;
2041 	uma_keg_t keg;
2042 
2043 	TAILQ_INIT(&alloctail);
2044 	keg = zone->uz_keg;
2045 
2046 	npages = howmany(bytes, PAGE_SIZE);
2047 	while (npages > 0) {
2048 		p = vm_page_alloc_domain(NULL, 0, domain, VM_ALLOC_INTERRUPT |
2049 		    VM_ALLOC_WIRED | VM_ALLOC_NOOBJ |
2050 		    ((wait & M_WAITOK) != 0 ? VM_ALLOC_WAITOK :
2051 		    VM_ALLOC_NOWAIT));
2052 		if (p != NULL) {
2053 			/*
2054 			 * Since the page does not belong to an object, its
2055 			 * listq is unused.
2056 			 */
2057 			TAILQ_INSERT_TAIL(&alloctail, p, listq);
2058 			npages--;
2059 			continue;
2060 		}
2061 		/*
2062 		 * Page allocation failed, free intermediate pages and
2063 		 * exit.
2064 		 */
2065 		TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
2066 			vm_page_unwire_noq(p);
2067 			vm_page_free(p);
2068 		}
2069 		return (NULL);
2070 	}
2071 	*flags = UMA_SLAB_PRIV;
2072 	zkva = keg->uk_kva +
2073 	    atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
2074 	retkva = zkva;
2075 	TAILQ_FOREACH(p, &alloctail, listq) {
2076 		pmap_qenter(zkva, &p, 1);
2077 		zkva += PAGE_SIZE;
2078 	}
2079 
2080 	return ((void *)retkva);
2081 }
2082 
2083 /*
2084  * Allocate physically contiguous pages.
2085  */
2086 static void *
2087 contig_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
2088     int wait)
2089 {
2090 
2091 	*pflag = UMA_SLAB_KERNEL;
2092 	return ((void *)kmem_alloc_contig_domainset(DOMAINSET_FIXED(domain),
2093 	    bytes, wait, 0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT));
2094 }
2095 
2096 /*
2097  * Frees a number of pages to the system
2098  *
2099  * Arguments:
2100  *	mem   A pointer to the memory to be freed
2101  *	size  The size of the memory being freed
2102  *	flags The original p->us_flags field
2103  *
2104  * Returns:
2105  *	Nothing
2106  */
2107 static void
2108 page_free(void *mem, vm_size_t size, uint8_t flags)
2109 {
2110 
2111 	if ((flags & UMA_SLAB_BOOT) != 0) {
2112 		startup_free(mem, size);
2113 		return;
2114 	}
2115 
2116 	KASSERT((flags & UMA_SLAB_KERNEL) != 0,
2117 	    ("UMA: page_free used with invalid flags %x", flags));
2118 
2119 	kmem_free((vm_offset_t)mem, size);
2120 }
2121 
2122 /*
2123  * Frees pcpu zone allocations
2124  *
2125  * Arguments:
2126  *	mem   A pointer to the memory to be freed
2127  *	size  The size of the memory being freed
2128  *	flags The original p->us_flags field
2129  *
2130  * Returns:
2131  *	Nothing
2132  */
2133 static void
2134 pcpu_page_free(void *mem, vm_size_t size, uint8_t flags)
2135 {
2136 	vm_offset_t sva, curva;
2137 	vm_paddr_t paddr;
2138 	vm_page_t m;
2139 
2140 	MPASS(size == (mp_maxid+1)*PAGE_SIZE);
2141 
2142 	if ((flags & UMA_SLAB_BOOT) != 0) {
2143 		startup_free(mem, size);
2144 		return;
2145 	}
2146 
2147 	sva = (vm_offset_t)mem;
2148 	for (curva = sva; curva < sva + size; curva += PAGE_SIZE) {
2149 		paddr = pmap_kextract(curva);
2150 		m = PHYS_TO_VM_PAGE(paddr);
2151 		vm_page_unwire_noq(m);
2152 		vm_page_free(m);
2153 	}
2154 	pmap_qremove(sva, size >> PAGE_SHIFT);
2155 	kva_free(sva, size);
2156 }
2157 
2158 /*
2159  * Zero fill initializer
2160  *
2161  * Arguments/Returns follow uma_init specifications
2162  */
2163 static int
2164 zero_init(void *mem, int size, int flags)
2165 {
2166 	bzero(mem, size);
2167 	return (0);
2168 }
2169 
2170 #ifdef INVARIANTS
2171 static struct noslabbits *
2172 slab_dbg_bits(uma_slab_t slab, uma_keg_t keg)
2173 {
2174 
2175 	return ((void *)((char *)&slab->us_free + BITSET_SIZE(keg->uk_ipers)));
2176 }
2177 #endif
2178 
2179 /*
2180  * Actual size of embedded struct slab (!OFFPAGE).
2181  */
2182 static size_t
2183 slab_sizeof(int nitems)
2184 {
2185 	size_t s;
2186 
2187 	s = sizeof(struct uma_slab) + BITSET_SIZE(nitems) * SLAB_BITSETS;
2188 	return (roundup(s, UMA_ALIGN_PTR + 1));
2189 }
2190 
2191 #define	UMA_FIXPT_SHIFT	31
2192 #define	UMA_FRAC_FIXPT(n, d)						\
2193 	((uint32_t)(((uint64_t)(n) << UMA_FIXPT_SHIFT) / (d)))
2194 #define	UMA_FIXPT_PCT(f)						\
2195 	((u_int)(((uint64_t)100 * (f)) >> UMA_FIXPT_SHIFT))
2196 #define	UMA_PCT_FIXPT(pct)	UMA_FRAC_FIXPT((pct), 100)
2197 #define	UMA_MIN_EFF	UMA_PCT_FIXPT(100 - UMA_MAX_WASTE)
2198 
2199 /*
2200  * Compute the number of items that will fit in a slab.  If hdr is true, the
2201  * item count may be limited to provide space in the slab for an inline slab
2202  * header.  Otherwise, all slab space will be provided for item storage.
2203  */
2204 static u_int
2205 slab_ipers_hdr(u_int size, u_int rsize, u_int slabsize, bool hdr)
2206 {
2207 	u_int ipers;
2208 	u_int padpi;
2209 
2210 	/* The padding between items is not needed after the last item. */
2211 	padpi = rsize - size;
2212 
2213 	if (hdr) {
2214 		/*
2215 		 * Start with the maximum item count and remove items until
2216 		 * the slab header first alongside the allocatable memory.
2217 		 */
2218 		for (ipers = MIN(SLAB_MAX_SETSIZE,
2219 		    (slabsize + padpi - slab_sizeof(1)) / rsize);
2220 		    ipers > 0 &&
2221 		    ipers * rsize - padpi + slab_sizeof(ipers) > slabsize;
2222 		    ipers--)
2223 			continue;
2224 	} else {
2225 		ipers = MIN((slabsize + padpi) / rsize, SLAB_MAX_SETSIZE);
2226 	}
2227 
2228 	return (ipers);
2229 }
2230 
2231 struct keg_layout_result {
2232 	u_int format;
2233 	u_int slabsize;
2234 	u_int ipers;
2235 	u_int eff;
2236 };
2237 
2238 static void
2239 keg_layout_one(uma_keg_t keg, u_int rsize, u_int slabsize, u_int fmt,
2240     struct keg_layout_result *kl)
2241 {
2242 	u_int total;
2243 
2244 	kl->format = fmt;
2245 	kl->slabsize = slabsize;
2246 
2247 	/* Handle INTERNAL as inline with an extra page. */
2248 	if ((fmt & UMA_ZFLAG_INTERNAL) != 0) {
2249 		kl->format &= ~UMA_ZFLAG_INTERNAL;
2250 		kl->slabsize += PAGE_SIZE;
2251 	}
2252 
2253 	kl->ipers = slab_ipers_hdr(keg->uk_size, rsize, kl->slabsize,
2254 	    (fmt & UMA_ZFLAG_OFFPAGE) == 0);
2255 
2256 	/* Account for memory used by an offpage slab header. */
2257 	total = kl->slabsize;
2258 	if ((fmt & UMA_ZFLAG_OFFPAGE) != 0)
2259 		total += slabzone(kl->ipers)->uz_keg->uk_rsize;
2260 
2261 	kl->eff = UMA_FRAC_FIXPT(kl->ipers * rsize, total);
2262 }
2263 
2264 /*
2265  * Determine the format of a uma keg.  This determines where the slab header
2266  * will be placed (inline or offpage) and calculates ipers, rsize, and ppera.
2267  *
2268  * Arguments
2269  *	keg  The zone we should initialize
2270  *
2271  * Returns
2272  *	Nothing
2273  */
2274 static void
2275 keg_layout(uma_keg_t keg)
2276 {
2277 	struct keg_layout_result kl = {}, kl_tmp;
2278 	u_int fmts[2];
2279 	u_int alignsize;
2280 	u_int nfmt;
2281 	u_int pages;
2282 	u_int rsize;
2283 	u_int slabsize;
2284 	u_int i, j;
2285 
2286 	KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
2287 	    (keg->uk_size <= UMA_PCPU_ALLOC_SIZE &&
2288 	     (keg->uk_flags & UMA_ZONE_CACHESPREAD) == 0),
2289 	    ("%s: cannot configure for PCPU: keg=%s, size=%u, flags=0x%b",
2290 	     __func__, keg->uk_name, keg->uk_size, keg->uk_flags,
2291 	     PRINT_UMA_ZFLAGS));
2292 	KASSERT((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) == 0 ||
2293 	    (keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0,
2294 	    ("%s: incompatible flags 0x%b", __func__, keg->uk_flags,
2295 	     PRINT_UMA_ZFLAGS));
2296 
2297 	alignsize = keg->uk_align + 1;
2298 #ifdef KASAN
2299 	/*
2300 	 * ASAN requires that each allocation be aligned to the shadow map
2301 	 * scale factor.
2302 	 */
2303 	if (alignsize < KASAN_SHADOW_SCALE)
2304 		alignsize = KASAN_SHADOW_SCALE;
2305 #endif
2306 
2307 	/*
2308 	 * Calculate the size of each allocation (rsize) according to
2309 	 * alignment.  If the requested size is smaller than we have
2310 	 * allocation bits for we round it up.
2311 	 */
2312 	rsize = MAX(keg->uk_size, UMA_SMALLEST_UNIT);
2313 	rsize = roundup2(rsize, alignsize);
2314 
2315 	if ((keg->uk_flags & UMA_ZONE_CACHESPREAD) != 0) {
2316 		/*
2317 		 * We want one item to start on every align boundary in a page.
2318 		 * To do this we will span pages.  We will also extend the item
2319 		 * by the size of align if it is an even multiple of align.
2320 		 * Otherwise, it would fall on the same boundary every time.
2321 		 */
2322 		if ((rsize & alignsize) == 0)
2323 			rsize += alignsize;
2324 		slabsize = rsize * (PAGE_SIZE / alignsize);
2325 		slabsize = MIN(slabsize, rsize * SLAB_MAX_SETSIZE);
2326 		slabsize = MIN(slabsize, UMA_CACHESPREAD_MAX_SIZE);
2327 		slabsize = round_page(slabsize);
2328 	} else {
2329 		/*
2330 		 * Start with a slab size of as many pages as it takes to
2331 		 * represent a single item.  We will try to fit as many
2332 		 * additional items into the slab as possible.
2333 		 */
2334 		slabsize = round_page(keg->uk_size);
2335 	}
2336 
2337 	/* Build a list of all of the available formats for this keg. */
2338 	nfmt = 0;
2339 
2340 	/* Evaluate an inline slab layout. */
2341 	if ((keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0)
2342 		fmts[nfmt++] = 0;
2343 
2344 	/* TODO: vm_page-embedded slab. */
2345 
2346 	/*
2347 	 * We can't do OFFPAGE if we're internal or if we've been
2348 	 * asked to not go to the VM for buckets.  If we do this we
2349 	 * may end up going to the VM for slabs which we do not want
2350 	 * to do if we're UMA_ZONE_VM, which clearly forbids it.
2351 	 * In those cases, evaluate a pseudo-format called INTERNAL
2352 	 * which has an inline slab header and one extra page to
2353 	 * guarantee that it fits.
2354 	 *
2355 	 * Otherwise, see if using an OFFPAGE slab will improve our
2356 	 * efficiency.
2357 	 */
2358 	if ((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) != 0)
2359 		fmts[nfmt++] = UMA_ZFLAG_INTERNAL;
2360 	else
2361 		fmts[nfmt++] = UMA_ZFLAG_OFFPAGE;
2362 
2363 	/*
2364 	 * Choose a slab size and format which satisfy the minimum efficiency.
2365 	 * Prefer the smallest slab size that meets the constraints.
2366 	 *
2367 	 * Start with a minimum slab size, to accommodate CACHESPREAD.  Then,
2368 	 * for small items (up to PAGE_SIZE), the iteration increment is one
2369 	 * page; and for large items, the increment is one item.
2370 	 */
2371 	i = (slabsize + rsize - keg->uk_size) / MAX(PAGE_SIZE, rsize);
2372 	KASSERT(i >= 1, ("keg %s(%p) flags=0x%b slabsize=%u, rsize=%u, i=%u",
2373 	    keg->uk_name, keg, keg->uk_flags, PRINT_UMA_ZFLAGS, slabsize,
2374 	    rsize, i));
2375 	for ( ; ; i++) {
2376 		slabsize = (rsize <= PAGE_SIZE) ? ptoa(i) :
2377 		    round_page(rsize * (i - 1) + keg->uk_size);
2378 
2379 		for (j = 0; j < nfmt; j++) {
2380 			/* Only if we have no viable format yet. */
2381 			if ((fmts[j] & UMA_ZFLAG_INTERNAL) != 0 &&
2382 			    kl.ipers > 0)
2383 				continue;
2384 
2385 			keg_layout_one(keg, rsize, slabsize, fmts[j], &kl_tmp);
2386 			if (kl_tmp.eff <= kl.eff)
2387 				continue;
2388 
2389 			kl = kl_tmp;
2390 
2391 			CTR6(KTR_UMA, "keg %s layout: format %#x "
2392 			    "(ipers %u * rsize %u) / slabsize %#x = %u%% eff",
2393 			    keg->uk_name, kl.format, kl.ipers, rsize,
2394 			    kl.slabsize, UMA_FIXPT_PCT(kl.eff));
2395 
2396 			/* Stop when we reach the minimum efficiency. */
2397 			if (kl.eff >= UMA_MIN_EFF)
2398 				break;
2399 		}
2400 
2401 		if (kl.eff >= UMA_MIN_EFF || !multipage_slabs ||
2402 		    slabsize >= SLAB_MAX_SETSIZE * rsize ||
2403 		    (keg->uk_flags & (UMA_ZONE_PCPU | UMA_ZONE_CONTIG)) != 0)
2404 			break;
2405 	}
2406 
2407 	pages = atop(kl.slabsize);
2408 	if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
2409 		pages *= mp_maxid + 1;
2410 
2411 	keg->uk_rsize = rsize;
2412 	keg->uk_ipers = kl.ipers;
2413 	keg->uk_ppera = pages;
2414 	keg->uk_flags |= kl.format;
2415 
2416 	/*
2417 	 * How do we find the slab header if it is offpage or if not all item
2418 	 * start addresses are in the same page?  We could solve the latter
2419 	 * case with vaddr alignment, but we don't.
2420 	 */
2421 	if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0 ||
2422 	    (keg->uk_ipers - 1) * rsize >= PAGE_SIZE) {
2423 		if ((keg->uk_flags & UMA_ZONE_NOTPAGE) != 0)
2424 			keg->uk_flags |= UMA_ZFLAG_HASH;
2425 		else
2426 			keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2427 	}
2428 
2429 	CTR6(KTR_UMA, "%s: keg=%s, flags=%#x, rsize=%u, ipers=%u, ppera=%u",
2430 	    __func__, keg->uk_name, keg->uk_flags, rsize, keg->uk_ipers,
2431 	    pages);
2432 	KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_MAX_SETSIZE,
2433 	    ("%s: keg=%s, flags=0x%b, rsize=%u, ipers=%u, ppera=%u", __func__,
2434 	     keg->uk_name, keg->uk_flags, PRINT_UMA_ZFLAGS, rsize,
2435 	     keg->uk_ipers, pages));
2436 }
2437 
2438 /*
2439  * Keg header ctor.  This initializes all fields, locks, etc.  And inserts
2440  * the keg onto the global keg list.
2441  *
2442  * Arguments/Returns follow uma_ctor specifications
2443  *	udata  Actually uma_kctor_args
2444  */
2445 static int
2446 keg_ctor(void *mem, int size, void *udata, int flags)
2447 {
2448 	struct uma_kctor_args *arg = udata;
2449 	uma_keg_t keg = mem;
2450 	uma_zone_t zone;
2451 	int i;
2452 
2453 	bzero(keg, size);
2454 	keg->uk_size = arg->size;
2455 	keg->uk_init = arg->uminit;
2456 	keg->uk_fini = arg->fini;
2457 	keg->uk_align = arg->align;
2458 	keg->uk_reserve = 0;
2459 	keg->uk_flags = arg->flags;
2460 
2461 	/*
2462 	 * We use a global round-robin policy by default.  Zones with
2463 	 * UMA_ZONE_FIRSTTOUCH set will use first-touch instead, in which
2464 	 * case the iterator is never run.
2465 	 */
2466 	keg->uk_dr.dr_policy = DOMAINSET_RR();
2467 	keg->uk_dr.dr_iter = 0;
2468 
2469 	/*
2470 	 * The primary zone is passed to us at keg-creation time.
2471 	 */
2472 	zone = arg->zone;
2473 	keg->uk_name = zone->uz_name;
2474 
2475 	if (arg->flags & UMA_ZONE_ZINIT)
2476 		keg->uk_init = zero_init;
2477 
2478 	if (arg->flags & UMA_ZONE_MALLOC)
2479 		keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2480 
2481 #ifndef SMP
2482 	keg->uk_flags &= ~UMA_ZONE_PCPU;
2483 #endif
2484 
2485 	keg_layout(keg);
2486 
2487 	/*
2488 	 * Use a first-touch NUMA policy for kegs that pmap_extract() will
2489 	 * work on.  Use round-robin for everything else.
2490 	 *
2491 	 * Zones may override the default by specifying either.
2492 	 */
2493 #ifdef NUMA
2494 	if ((keg->uk_flags &
2495 	    (UMA_ZONE_ROUNDROBIN | UMA_ZFLAG_CACHE | UMA_ZONE_NOTPAGE)) == 0)
2496 		keg->uk_flags |= UMA_ZONE_FIRSTTOUCH;
2497 	else if ((keg->uk_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2498 		keg->uk_flags |= UMA_ZONE_ROUNDROBIN;
2499 #endif
2500 
2501 	/*
2502 	 * If we haven't booted yet we need allocations to go through the
2503 	 * startup cache until the vm is ready.
2504 	 */
2505 #ifdef UMA_MD_SMALL_ALLOC
2506 	if (keg->uk_ppera == 1)
2507 		keg->uk_allocf = uma_small_alloc;
2508 	else
2509 #endif
2510 	if (booted < BOOT_KVA)
2511 		keg->uk_allocf = startup_alloc;
2512 	else if (keg->uk_flags & UMA_ZONE_PCPU)
2513 		keg->uk_allocf = pcpu_page_alloc;
2514 	else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 && keg->uk_ppera > 1)
2515 		keg->uk_allocf = contig_alloc;
2516 	else
2517 		keg->uk_allocf = page_alloc;
2518 #ifdef UMA_MD_SMALL_ALLOC
2519 	if (keg->uk_ppera == 1)
2520 		keg->uk_freef = uma_small_free;
2521 	else
2522 #endif
2523 	if (keg->uk_flags & UMA_ZONE_PCPU)
2524 		keg->uk_freef = pcpu_page_free;
2525 	else
2526 		keg->uk_freef = page_free;
2527 
2528 	/*
2529 	 * Initialize keg's locks.
2530 	 */
2531 	for (i = 0; i < vm_ndomains; i++)
2532 		KEG_LOCK_INIT(keg, i, (arg->flags & UMA_ZONE_MTXCLASS));
2533 
2534 	/*
2535 	 * If we're putting the slab header in the actual page we need to
2536 	 * figure out where in each page it goes.  See slab_sizeof
2537 	 * definition.
2538 	 */
2539 	if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE)) {
2540 		size_t shsize;
2541 
2542 		shsize = slab_sizeof(keg->uk_ipers);
2543 		keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - shsize;
2544 		/*
2545 		 * The only way the following is possible is if with our
2546 		 * UMA_ALIGN_PTR adjustments we are now bigger than
2547 		 * UMA_SLAB_SIZE.  I haven't checked whether this is
2548 		 * mathematically possible for all cases, so we make
2549 		 * sure here anyway.
2550 		 */
2551 		KASSERT(keg->uk_pgoff + shsize <= PAGE_SIZE * keg->uk_ppera,
2552 		    ("zone %s ipers %d rsize %d size %d slab won't fit",
2553 		    zone->uz_name, keg->uk_ipers, keg->uk_rsize, keg->uk_size));
2554 	}
2555 
2556 	if (keg->uk_flags & UMA_ZFLAG_HASH)
2557 		hash_alloc(&keg->uk_hash, 0);
2558 
2559 	CTR3(KTR_UMA, "keg_ctor %p zone %s(%p)", keg, zone->uz_name, zone);
2560 
2561 	LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
2562 
2563 	rw_wlock(&uma_rwlock);
2564 	LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
2565 	rw_wunlock(&uma_rwlock);
2566 	return (0);
2567 }
2568 
2569 static void
2570 zone_kva_available(uma_zone_t zone, void *unused)
2571 {
2572 	uma_keg_t keg;
2573 
2574 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
2575 		return;
2576 	KEG_GET(zone, keg);
2577 
2578 	if (keg->uk_allocf == startup_alloc) {
2579 		/* Switch to the real allocator. */
2580 		if (keg->uk_flags & UMA_ZONE_PCPU)
2581 			keg->uk_allocf = pcpu_page_alloc;
2582 		else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 &&
2583 		    keg->uk_ppera > 1)
2584 			keg->uk_allocf = contig_alloc;
2585 		else
2586 			keg->uk_allocf = page_alloc;
2587 	}
2588 }
2589 
2590 static void
2591 zone_alloc_counters(uma_zone_t zone, void *unused)
2592 {
2593 
2594 	zone->uz_allocs = counter_u64_alloc(M_WAITOK);
2595 	zone->uz_frees = counter_u64_alloc(M_WAITOK);
2596 	zone->uz_fails = counter_u64_alloc(M_WAITOK);
2597 	zone->uz_xdomain = counter_u64_alloc(M_WAITOK);
2598 }
2599 
2600 static void
2601 zone_alloc_sysctl(uma_zone_t zone, void *unused)
2602 {
2603 	uma_zone_domain_t zdom;
2604 	uma_domain_t dom;
2605 	uma_keg_t keg;
2606 	struct sysctl_oid *oid, *domainoid;
2607 	int domains, i, cnt;
2608 	static const char *nokeg = "cache zone";
2609 	char *c;
2610 
2611 	/*
2612 	 * Make a sysctl safe copy of the zone name by removing
2613 	 * any special characters and handling dups by appending
2614 	 * an index.
2615 	 */
2616 	if (zone->uz_namecnt != 0) {
2617 		/* Count the number of decimal digits and '_' separator. */
2618 		for (i = 1, cnt = zone->uz_namecnt; cnt != 0; i++)
2619 			cnt /= 10;
2620 		zone->uz_ctlname = malloc(strlen(zone->uz_name) + i + 1,
2621 		    M_UMA, M_WAITOK);
2622 		sprintf(zone->uz_ctlname, "%s_%d", zone->uz_name,
2623 		    zone->uz_namecnt);
2624 	} else
2625 		zone->uz_ctlname = strdup(zone->uz_name, M_UMA);
2626 	for (c = zone->uz_ctlname; *c != '\0'; c++)
2627 		if (strchr("./\\ -", *c) != NULL)
2628 			*c = '_';
2629 
2630 	/*
2631 	 * Basic parameters at the root.
2632 	 */
2633 	zone->uz_oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_vm_uma),
2634 	    OID_AUTO, zone->uz_ctlname, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2635 	oid = zone->uz_oid;
2636 	SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2637 	    "size", CTLFLAG_RD, &zone->uz_size, 0, "Allocation size");
2638 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2639 	    "flags", CTLFLAG_RD | CTLTYPE_STRING | CTLFLAG_MPSAFE,
2640 	    zone, 0, sysctl_handle_uma_zone_flags, "A",
2641 	    "Allocator configuration flags");
2642 	SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2643 	    "bucket_size", CTLFLAG_RD, &zone->uz_bucket_size, 0,
2644 	    "Desired per-cpu cache size");
2645 	SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2646 	    "bucket_size_max", CTLFLAG_RD, &zone->uz_bucket_size_max, 0,
2647 	    "Maximum allowed per-cpu cache size");
2648 
2649 	/*
2650 	 * keg if present.
2651 	 */
2652 	if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
2653 		domains = vm_ndomains;
2654 	else
2655 		domains = 1;
2656 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2657 	    "keg", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2658 	keg = zone->uz_keg;
2659 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0) {
2660 		SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2661 		    "name", CTLFLAG_RD, keg->uk_name, "Keg name");
2662 		SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2663 		    "rsize", CTLFLAG_RD, &keg->uk_rsize, 0,
2664 		    "Real object size with alignment");
2665 		SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2666 		    "ppera", CTLFLAG_RD, &keg->uk_ppera, 0,
2667 		    "pages per-slab allocation");
2668 		SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2669 		    "ipers", CTLFLAG_RD, &keg->uk_ipers, 0,
2670 		    "items available per-slab");
2671 		SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2672 		    "align", CTLFLAG_RD, &keg->uk_align, 0,
2673 		    "item alignment mask");
2674 		SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2675 		    "reserve", CTLFLAG_RD, &keg->uk_reserve, 0,
2676 		    "number of reserved items");
2677 		SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2678 		    "efficiency", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2679 		    keg, 0, sysctl_handle_uma_slab_efficiency, "I",
2680 		    "Slab utilization (100 - internal fragmentation %)");
2681 		domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(oid),
2682 		    OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2683 		for (i = 0; i < domains; i++) {
2684 			dom = &keg->uk_domain[i];
2685 			oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2686 			    OID_AUTO, VM_DOMAIN(i)->vmd_name,
2687 			    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2688 			SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2689 			    "pages", CTLFLAG_RD, &dom->ud_pages, 0,
2690 			    "Total pages currently allocated from VM");
2691 			SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2692 			    "free_items", CTLFLAG_RD, &dom->ud_free_items, 0,
2693 			    "items free in the slab layer");
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 #ifdef UMA_ZALLOC_DEBUG
3631 	void *item;
3632 
3633 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
3634 	    ("uma_zalloc_arg: called with non-SMR zone."));
3635 	if (uma_zalloc_debug(zone, &item, NULL, flags) == EJUSTRETURN)
3636 		return (item);
3637 #endif
3638 
3639 	critical_enter();
3640 	cache = &zone->uz_cpu[curcpu];
3641 	bucket = &cache->uc_allocbucket;
3642 	if (__predict_false(bucket->ucb_cnt == 0))
3643 		return (cache_alloc_retry(zone, cache, NULL, flags));
3644 	return (cache_alloc_item(zone, cache, bucket, NULL, flags));
3645 }
3646 
3647 /* See uma.h */
3648 void *
3649 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
3650 {
3651 	uma_cache_bucket_t bucket;
3652 	uma_cache_t cache;
3653 
3654 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3655 	random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3656 
3657 	/* This is the fast path allocation */
3658 	CTR3(KTR_UMA, "uma_zalloc_arg zone %s(%p) flags %d", zone->uz_name,
3659 	    zone, flags);
3660 
3661 #ifdef UMA_ZALLOC_DEBUG
3662 	void *item;
3663 
3664 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3665 	    ("uma_zalloc_arg: called with SMR zone."));
3666 	if (uma_zalloc_debug(zone, &item, udata, flags) == EJUSTRETURN)
3667 		return (item);
3668 #endif
3669 
3670 	/*
3671 	 * If possible, allocate from the per-CPU cache.  There are two
3672 	 * requirements for safe access to the per-CPU cache: (1) the thread
3673 	 * accessing the cache must not be preempted or yield during access,
3674 	 * and (2) the thread must not migrate CPUs without switching which
3675 	 * cache it accesses.  We rely on a critical section to prevent
3676 	 * preemption and migration.  We release the critical section in
3677 	 * order to acquire the zone mutex if we are unable to allocate from
3678 	 * the current cache; when we re-acquire the critical section, we
3679 	 * must detect and handle migration if it has occurred.
3680 	 */
3681 	critical_enter();
3682 	cache = &zone->uz_cpu[curcpu];
3683 	bucket = &cache->uc_allocbucket;
3684 	if (__predict_false(bucket->ucb_cnt == 0))
3685 		return (cache_alloc_retry(zone, cache, udata, flags));
3686 	return (cache_alloc_item(zone, cache, bucket, udata, flags));
3687 }
3688 
3689 /*
3690  * Replenish an alloc bucket and possibly restore an old one.  Called in
3691  * a critical section.  Returns in a critical section.
3692  *
3693  * A false return value indicates an allocation failure.
3694  * A true return value indicates success and the caller should retry.
3695  */
3696 static __noinline bool
3697 cache_alloc(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3698 {
3699 	uma_bucket_t bucket;
3700 	int curdomain, domain;
3701 	bool new;
3702 
3703 	CRITICAL_ASSERT(curthread);
3704 
3705 	/*
3706 	 * If we have run out of items in our alloc bucket see
3707 	 * if we can switch with the free bucket.
3708 	 *
3709 	 * SMR Zones can't re-use the free bucket until the sequence has
3710 	 * expired.
3711 	 */
3712 	if ((cache_uz_flags(cache) & UMA_ZONE_SMR) == 0 &&
3713 	    cache->uc_freebucket.ucb_cnt != 0) {
3714 		cache_bucket_swap(&cache->uc_freebucket,
3715 		    &cache->uc_allocbucket);
3716 		return (true);
3717 	}
3718 
3719 	/*
3720 	 * Discard any empty allocation bucket while we hold no locks.
3721 	 */
3722 	bucket = cache_bucket_unload_alloc(cache);
3723 	critical_exit();
3724 
3725 	if (bucket != NULL) {
3726 		KASSERT(bucket->ub_cnt == 0,
3727 		    ("cache_alloc: Entered with non-empty alloc bucket."));
3728 		bucket_free(zone, bucket, udata);
3729 	}
3730 
3731 	/*
3732 	 * Attempt to retrieve the item from the per-CPU cache has failed, so
3733 	 * we must go back to the zone.  This requires the zdom lock, so we
3734 	 * must drop the critical section, then re-acquire it when we go back
3735 	 * to the cache.  Since the critical section is released, we may be
3736 	 * preempted or migrate.  As such, make sure not to maintain any
3737 	 * thread-local state specific to the cache from prior to releasing
3738 	 * the critical section.
3739 	 */
3740 	domain = PCPU_GET(domain);
3741 	if ((cache_uz_flags(cache) & UMA_ZONE_ROUNDROBIN) != 0 ||
3742 	    VM_DOMAIN_EMPTY(domain))
3743 		domain = zone_domain_highest(zone, domain);
3744 	bucket = cache_fetch_bucket(zone, cache, domain);
3745 	if (bucket == NULL && zone->uz_bucket_size != 0 && !bucketdisable) {
3746 		bucket = zone_alloc_bucket(zone, udata, domain, flags);
3747 		new = true;
3748 	} else {
3749 		new = false;
3750 	}
3751 
3752 	CTR3(KTR_UMA, "uma_zalloc: zone %s(%p) bucket zone returned %p",
3753 	    zone->uz_name, zone, bucket);
3754 	if (bucket == NULL) {
3755 		critical_enter();
3756 		return (false);
3757 	}
3758 
3759 	/*
3760 	 * See if we lost the race or were migrated.  Cache the
3761 	 * initialized bucket to make this less likely or claim
3762 	 * the memory directly.
3763 	 */
3764 	critical_enter();
3765 	cache = &zone->uz_cpu[curcpu];
3766 	if (cache->uc_allocbucket.ucb_bucket == NULL &&
3767 	    ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) == 0 ||
3768 	    (curdomain = PCPU_GET(domain)) == domain ||
3769 	    VM_DOMAIN_EMPTY(curdomain))) {
3770 		if (new)
3771 			atomic_add_long(&ZDOM_GET(zone, domain)->uzd_imax,
3772 			    bucket->ub_cnt);
3773 		cache_bucket_load_alloc(cache, bucket);
3774 		return (true);
3775 	}
3776 
3777 	/*
3778 	 * We lost the race, release this bucket and start over.
3779 	 */
3780 	critical_exit();
3781 	zone_put_bucket(zone, domain, bucket, udata, !new);
3782 	critical_enter();
3783 
3784 	return (true);
3785 }
3786 
3787 void *
3788 uma_zalloc_domain(uma_zone_t zone, void *udata, int domain, int flags)
3789 {
3790 #ifdef NUMA
3791 	uma_bucket_t bucket;
3792 	uma_zone_domain_t zdom;
3793 	void *item;
3794 #endif
3795 
3796 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3797 	random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3798 
3799 	/* This is the fast path allocation */
3800 	CTR4(KTR_UMA, "uma_zalloc_domain zone %s(%p) domain %d flags %d",
3801 	    zone->uz_name, zone, domain, flags);
3802 
3803 	if (flags & M_WAITOK) {
3804 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
3805 		    "uma_zalloc_domain: zone \"%s\"", zone->uz_name);
3806 	}
3807 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3808 	    ("uma_zalloc_domain: called with spinlock or critical section held"));
3809 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3810 	    ("uma_zalloc_domain: called with SMR zone."));
3811 #ifdef NUMA
3812 	KASSERT((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0,
3813 	    ("uma_zalloc_domain: called with non-FIRSTTOUCH zone."));
3814 
3815 	if (vm_ndomains == 1)
3816 		return (uma_zalloc_arg(zone, udata, flags));
3817 
3818 	/*
3819 	 * Try to allocate from the bucket cache before falling back to the keg.
3820 	 * We could try harder and attempt to allocate from per-CPU caches or
3821 	 * the per-domain cross-domain buckets, but the complexity is probably
3822 	 * not worth it.  It is more important that frees of previous
3823 	 * cross-domain allocations do not blow up the cache.
3824 	 */
3825 	zdom = zone_domain_lock(zone, domain);
3826 	if ((bucket = zone_fetch_bucket(zone, zdom, false)) != NULL) {
3827 		item = bucket->ub_bucket[bucket->ub_cnt - 1];
3828 #ifdef INVARIANTS
3829 		bucket->ub_bucket[bucket->ub_cnt - 1] = NULL;
3830 #endif
3831 		bucket->ub_cnt--;
3832 		zone_put_bucket(zone, domain, bucket, udata, true);
3833 		item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata,
3834 		    flags, item);
3835 		if (item != NULL) {
3836 			KASSERT(item_domain(item) == domain,
3837 			    ("%s: bucket cache item %p from wrong domain",
3838 			    __func__, item));
3839 			counter_u64_add(zone->uz_allocs, 1);
3840 		}
3841 		return (item);
3842 	}
3843 	ZDOM_UNLOCK(zdom);
3844 	return (zone_alloc_item(zone, udata, domain, flags));
3845 #else
3846 	return (uma_zalloc_arg(zone, udata, flags));
3847 #endif
3848 }
3849 
3850 /*
3851  * Find a slab with some space.  Prefer slabs that are partially used over those
3852  * that are totally full.  This helps to reduce fragmentation.
3853  *
3854  * If 'rr' is 1, search all domains starting from 'domain'.  Otherwise check
3855  * only 'domain'.
3856  */
3857 static uma_slab_t
3858 keg_first_slab(uma_keg_t keg, int domain, bool rr)
3859 {
3860 	uma_domain_t dom;
3861 	uma_slab_t slab;
3862 	int start;
3863 
3864 	KASSERT(domain >= 0 && domain < vm_ndomains,
3865 	    ("keg_first_slab: domain %d out of range", domain));
3866 	KEG_LOCK_ASSERT(keg, domain);
3867 
3868 	slab = NULL;
3869 	start = domain;
3870 	do {
3871 		dom = &keg->uk_domain[domain];
3872 		if ((slab = LIST_FIRST(&dom->ud_part_slab)) != NULL)
3873 			return (slab);
3874 		if ((slab = LIST_FIRST(&dom->ud_free_slab)) != NULL) {
3875 			LIST_REMOVE(slab, us_link);
3876 			dom->ud_free_slabs--;
3877 			LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
3878 			return (slab);
3879 		}
3880 		if (rr)
3881 			domain = (domain + 1) % vm_ndomains;
3882 	} while (domain != start);
3883 
3884 	return (NULL);
3885 }
3886 
3887 /*
3888  * Fetch an existing slab from a free or partial list.  Returns with the
3889  * keg domain lock held if a slab was found or unlocked if not.
3890  */
3891 static uma_slab_t
3892 keg_fetch_free_slab(uma_keg_t keg, int domain, bool rr, int flags)
3893 {
3894 	uma_slab_t slab;
3895 	uint32_t reserve;
3896 
3897 	/* HASH has a single free list. */
3898 	if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
3899 		domain = 0;
3900 
3901 	KEG_LOCK(keg, domain);
3902 	reserve = (flags & M_USE_RESERVE) != 0 ? 0 : keg->uk_reserve;
3903 	if (keg->uk_domain[domain].ud_free_items <= reserve ||
3904 	    (slab = keg_first_slab(keg, domain, rr)) == NULL) {
3905 		KEG_UNLOCK(keg, domain);
3906 		return (NULL);
3907 	}
3908 	return (slab);
3909 }
3910 
3911 static uma_slab_t
3912 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int rdomain, const int flags)
3913 {
3914 	struct vm_domainset_iter di;
3915 	uma_slab_t slab;
3916 	int aflags, domain;
3917 	bool rr;
3918 
3919 restart:
3920 	/*
3921 	 * Use the keg's policy if upper layers haven't already specified a
3922 	 * domain (as happens with first-touch zones).
3923 	 *
3924 	 * To avoid races we run the iterator with the keg lock held, but that
3925 	 * means that we cannot allow the vm_domainset layer to sleep.  Thus,
3926 	 * clear M_WAITOK and handle low memory conditions locally.
3927 	 */
3928 	rr = rdomain == UMA_ANYDOMAIN;
3929 	if (rr) {
3930 		aflags = (flags & ~M_WAITOK) | M_NOWAIT;
3931 		vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
3932 		    &aflags);
3933 	} else {
3934 		aflags = flags;
3935 		domain = rdomain;
3936 	}
3937 
3938 	for (;;) {
3939 		slab = keg_fetch_free_slab(keg, domain, rr, flags);
3940 		if (slab != NULL)
3941 			return (slab);
3942 
3943 		/*
3944 		 * M_NOVM means don't ask at all!
3945 		 */
3946 		if (flags & M_NOVM)
3947 			break;
3948 
3949 		slab = keg_alloc_slab(keg, zone, domain, flags, aflags);
3950 		if (slab != NULL)
3951 			return (slab);
3952 		if (!rr && (flags & M_WAITOK) == 0)
3953 			break;
3954 		if (rr && vm_domainset_iter_policy(&di, &domain) != 0) {
3955 			if ((flags & M_WAITOK) != 0) {
3956 				vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask, 0);
3957 				goto restart;
3958 			}
3959 			break;
3960 		}
3961 	}
3962 
3963 	/*
3964 	 * We might not have been able to get a slab but another cpu
3965 	 * could have while we were unlocked.  Check again before we
3966 	 * fail.
3967 	 */
3968 	if ((slab = keg_fetch_free_slab(keg, domain, rr, flags)) != NULL)
3969 		return (slab);
3970 
3971 	return (NULL);
3972 }
3973 
3974 static void *
3975 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
3976 {
3977 	uma_domain_t dom;
3978 	void *item;
3979 	int freei;
3980 
3981 	KEG_LOCK_ASSERT(keg, slab->us_domain);
3982 
3983 	dom = &keg->uk_domain[slab->us_domain];
3984 	freei = BIT_FFS(keg->uk_ipers, &slab->us_free) - 1;
3985 	BIT_CLR(keg->uk_ipers, freei, &slab->us_free);
3986 	item = slab_item(slab, keg, freei);
3987 	slab->us_freecount--;
3988 	dom->ud_free_items--;
3989 
3990 	/*
3991 	 * Move this slab to the full list.  It must be on the partial list, so
3992 	 * we do not need to update the free slab count.  In particular,
3993 	 * keg_fetch_slab() always returns slabs on the partial list.
3994 	 */
3995 	if (slab->us_freecount == 0) {
3996 		LIST_REMOVE(slab, us_link);
3997 		LIST_INSERT_HEAD(&dom->ud_full_slab, slab, us_link);
3998 	}
3999 
4000 	return (item);
4001 }
4002 
4003 static int
4004 zone_import(void *arg, void **bucket, int max, int domain, int flags)
4005 {
4006 	uma_domain_t dom;
4007 	uma_zone_t zone;
4008 	uma_slab_t slab;
4009 	uma_keg_t keg;
4010 #ifdef NUMA
4011 	int stripe;
4012 #endif
4013 	int i;
4014 
4015 	zone = arg;
4016 	slab = NULL;
4017 	keg = zone->uz_keg;
4018 	/* Try to keep the buckets totally full */
4019 	for (i = 0; i < max; ) {
4020 		if ((slab = keg_fetch_slab(keg, zone, domain, flags)) == NULL)
4021 			break;
4022 #ifdef NUMA
4023 		stripe = howmany(max, vm_ndomains);
4024 #endif
4025 		dom = &keg->uk_domain[slab->us_domain];
4026 		do {
4027 			bucket[i++] = slab_alloc_item(keg, slab);
4028 			if (dom->ud_free_items <= keg->uk_reserve) {
4029 				/*
4030 				 * Avoid depleting the reserve after a
4031 				 * successful item allocation, even if
4032 				 * M_USE_RESERVE is specified.
4033 				 */
4034 				KEG_UNLOCK(keg, slab->us_domain);
4035 				goto out;
4036 			}
4037 #ifdef NUMA
4038 			/*
4039 			 * If the zone is striped we pick a new slab for every
4040 			 * N allocations.  Eliminating this conditional will
4041 			 * instead pick a new domain for each bucket rather
4042 			 * than stripe within each bucket.  The current option
4043 			 * produces more fragmentation and requires more cpu
4044 			 * time but yields better distribution.
4045 			 */
4046 			if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0 &&
4047 			    vm_ndomains > 1 && --stripe == 0)
4048 				break;
4049 #endif
4050 		} while (slab->us_freecount != 0 && i < max);
4051 		KEG_UNLOCK(keg, slab->us_domain);
4052 
4053 		/* Don't block if we allocated any successfully. */
4054 		flags &= ~M_WAITOK;
4055 		flags |= M_NOWAIT;
4056 	}
4057 out:
4058 	return i;
4059 }
4060 
4061 static int
4062 zone_alloc_limit_hard(uma_zone_t zone, int count, int flags)
4063 {
4064 	uint64_t old, new, total, max;
4065 
4066 	/*
4067 	 * The hard case.  We're going to sleep because there were existing
4068 	 * sleepers or because we ran out of items.  This routine enforces
4069 	 * fairness by keeping fifo order.
4070 	 *
4071 	 * First release our ill gotten gains and make some noise.
4072 	 */
4073 	for (;;) {
4074 		zone_free_limit(zone, count);
4075 		zone_log_warning(zone);
4076 		zone_maxaction(zone);
4077 		if (flags & M_NOWAIT)
4078 			return (0);
4079 
4080 		/*
4081 		 * We need to allocate an item or set ourself as a sleeper
4082 		 * while the sleepq lock is held to avoid wakeup races.  This
4083 		 * is essentially a home rolled semaphore.
4084 		 */
4085 		sleepq_lock(&zone->uz_max_items);
4086 		old = zone->uz_items;
4087 		do {
4088 			MPASS(UZ_ITEMS_SLEEPERS(old) < UZ_ITEMS_SLEEPERS_MAX);
4089 			/* Cache the max since we will evaluate twice. */
4090 			max = zone->uz_max_items;
4091 			if (UZ_ITEMS_SLEEPERS(old) != 0 ||
4092 			    UZ_ITEMS_COUNT(old) >= max)
4093 				new = old + UZ_ITEMS_SLEEPER;
4094 			else
4095 				new = old + MIN(count, max - old);
4096 		} while (atomic_fcmpset_64(&zone->uz_items, &old, new) == 0);
4097 
4098 		/* We may have successfully allocated under the sleepq lock. */
4099 		if (UZ_ITEMS_SLEEPERS(new) == 0) {
4100 			sleepq_release(&zone->uz_max_items);
4101 			return (new - old);
4102 		}
4103 
4104 		/*
4105 		 * This is in a different cacheline from uz_items so that we
4106 		 * don't constantly invalidate the fastpath cacheline when we
4107 		 * adjust item counts.  This could be limited to toggling on
4108 		 * transitions.
4109 		 */
4110 		atomic_add_32(&zone->uz_sleepers, 1);
4111 		atomic_add_64(&zone->uz_sleeps, 1);
4112 
4113 		/*
4114 		 * We have added ourselves as a sleeper.  The sleepq lock
4115 		 * protects us from wakeup races.  Sleep now and then retry.
4116 		 */
4117 		sleepq_add(&zone->uz_max_items, NULL, "zonelimit", 0, 0);
4118 		sleepq_wait(&zone->uz_max_items, PVM);
4119 
4120 		/*
4121 		 * After wakeup, remove ourselves as a sleeper and try
4122 		 * again.  We no longer have the sleepq lock for protection.
4123 		 *
4124 		 * Subract ourselves as a sleeper while attempting to add
4125 		 * our count.
4126 		 */
4127 		atomic_subtract_32(&zone->uz_sleepers, 1);
4128 		old = atomic_fetchadd_64(&zone->uz_items,
4129 		    -(UZ_ITEMS_SLEEPER - count));
4130 		/* We're no longer a sleeper. */
4131 		old -= UZ_ITEMS_SLEEPER;
4132 
4133 		/*
4134 		 * If we're still at the limit, restart.  Notably do not
4135 		 * block on other sleepers.  Cache the max value to protect
4136 		 * against changes via sysctl.
4137 		 */
4138 		total = UZ_ITEMS_COUNT(old);
4139 		max = zone->uz_max_items;
4140 		if (total >= max)
4141 			continue;
4142 		/* Truncate if necessary, otherwise wake other sleepers. */
4143 		if (total + count > max) {
4144 			zone_free_limit(zone, total + count - max);
4145 			count = max - total;
4146 		} else if (total + count < max && UZ_ITEMS_SLEEPERS(old) != 0)
4147 			wakeup_one(&zone->uz_max_items);
4148 
4149 		return (count);
4150 	}
4151 }
4152 
4153 /*
4154  * Allocate 'count' items from our max_items limit.  Returns the number
4155  * available.  If M_NOWAIT is not specified it will sleep until at least
4156  * one item can be allocated.
4157  */
4158 static int
4159 zone_alloc_limit(uma_zone_t zone, int count, int flags)
4160 {
4161 	uint64_t old;
4162 	uint64_t max;
4163 
4164 	max = zone->uz_max_items;
4165 	MPASS(max > 0);
4166 
4167 	/*
4168 	 * We expect normal allocations to succeed with a simple
4169 	 * fetchadd.
4170 	 */
4171 	old = atomic_fetchadd_64(&zone->uz_items, count);
4172 	if (__predict_true(old + count <= max))
4173 		return (count);
4174 
4175 	/*
4176 	 * If we had some items and no sleepers just return the
4177 	 * truncated value.  We have to release the excess space
4178 	 * though because that may wake sleepers who weren't woken
4179 	 * because we were temporarily over the limit.
4180 	 */
4181 	if (old < max) {
4182 		zone_free_limit(zone, (old + count) - max);
4183 		return (max - old);
4184 	}
4185 	return (zone_alloc_limit_hard(zone, count, flags));
4186 }
4187 
4188 /*
4189  * Free a number of items back to the limit.
4190  */
4191 static void
4192 zone_free_limit(uma_zone_t zone, int count)
4193 {
4194 	uint64_t old;
4195 
4196 	MPASS(count > 0);
4197 
4198 	/*
4199 	 * In the common case we either have no sleepers or
4200 	 * are still over the limit and can just return.
4201 	 */
4202 	old = atomic_fetchadd_64(&zone->uz_items, -count);
4203 	if (__predict_true(UZ_ITEMS_SLEEPERS(old) == 0 ||
4204 	   UZ_ITEMS_COUNT(old) - count >= zone->uz_max_items))
4205 		return;
4206 
4207 	/*
4208 	 * Moderate the rate of wakeups.  Sleepers will continue
4209 	 * to generate wakeups if necessary.
4210 	 */
4211 	wakeup_one(&zone->uz_max_items);
4212 }
4213 
4214 static uma_bucket_t
4215 zone_alloc_bucket(uma_zone_t zone, void *udata, int domain, int flags)
4216 {
4217 	uma_bucket_t bucket;
4218 	int error, maxbucket, cnt;
4219 
4220 	CTR3(KTR_UMA, "zone_alloc_bucket zone %s(%p) domain %d", zone->uz_name,
4221 	    zone, domain);
4222 
4223 	/* Avoid allocs targeting empty domains. */
4224 	if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
4225 		domain = UMA_ANYDOMAIN;
4226 	else if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
4227 		domain = UMA_ANYDOMAIN;
4228 
4229 	if (zone->uz_max_items > 0)
4230 		maxbucket = zone_alloc_limit(zone, zone->uz_bucket_size,
4231 		    M_NOWAIT);
4232 	else
4233 		maxbucket = zone->uz_bucket_size;
4234 	if (maxbucket == 0)
4235 		return (false);
4236 
4237 	/* Don't wait for buckets, preserve caller's NOVM setting. */
4238 	bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
4239 	if (bucket == NULL) {
4240 		cnt = 0;
4241 		goto out;
4242 	}
4243 
4244 	bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
4245 	    MIN(maxbucket, bucket->ub_entries), domain, flags);
4246 
4247 	/*
4248 	 * Initialize the memory if necessary.
4249 	 */
4250 	if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
4251 		int i;
4252 
4253 		for (i = 0; i < bucket->ub_cnt; i++) {
4254 			kasan_mark_item_valid(zone, bucket->ub_bucket[i]);
4255 			error = zone->uz_init(bucket->ub_bucket[i],
4256 			    zone->uz_size, flags);
4257 			kasan_mark_item_invalid(zone, bucket->ub_bucket[i]);
4258 			if (error != 0)
4259 				break;
4260 		}
4261 
4262 		/*
4263 		 * If we couldn't initialize the whole bucket, put the
4264 		 * rest back onto the freelist.
4265 		 */
4266 		if (i != bucket->ub_cnt) {
4267 			zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
4268 			    bucket->ub_cnt - i);
4269 #ifdef INVARIANTS
4270 			bzero(&bucket->ub_bucket[i],
4271 			    sizeof(void *) * (bucket->ub_cnt - i));
4272 #endif
4273 			bucket->ub_cnt = i;
4274 		}
4275 	}
4276 
4277 	cnt = bucket->ub_cnt;
4278 	if (bucket->ub_cnt == 0) {
4279 		bucket_free(zone, bucket, udata);
4280 		counter_u64_add(zone->uz_fails, 1);
4281 		bucket = NULL;
4282 	}
4283 out:
4284 	if (zone->uz_max_items > 0 && cnt < maxbucket)
4285 		zone_free_limit(zone, maxbucket - cnt);
4286 
4287 	return (bucket);
4288 }
4289 
4290 /*
4291  * Allocates a single item from a zone.
4292  *
4293  * Arguments
4294  *	zone   The zone to alloc for.
4295  *	udata  The data to be passed to the constructor.
4296  *	domain The domain to allocate from or UMA_ANYDOMAIN.
4297  *	flags  M_WAITOK, M_NOWAIT, M_ZERO.
4298  *
4299  * Returns
4300  *	NULL if there is no memory and M_NOWAIT is set
4301  *	An item if successful
4302  */
4303 
4304 static void *
4305 zone_alloc_item(uma_zone_t zone, void *udata, int domain, int flags)
4306 {
4307 	void *item;
4308 
4309 	if (zone->uz_max_items > 0 && zone_alloc_limit(zone, 1, flags) == 0) {
4310 		counter_u64_add(zone->uz_fails, 1);
4311 		return (NULL);
4312 	}
4313 
4314 	/* Avoid allocs targeting empty domains. */
4315 	if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
4316 		domain = UMA_ANYDOMAIN;
4317 
4318 	if (zone->uz_import(zone->uz_arg, &item, 1, domain, flags) != 1)
4319 		goto fail_cnt;
4320 
4321 	/*
4322 	 * We have to call both the zone's init (not the keg's init)
4323 	 * and the zone's ctor.  This is because the item is going from
4324 	 * a keg slab directly to the user, and the user is expecting it
4325 	 * to be both zone-init'd as well as zone-ctor'd.
4326 	 */
4327 	if (zone->uz_init != NULL) {
4328 		int error;
4329 
4330 		kasan_mark_item_valid(zone, item);
4331 		error = zone->uz_init(item, zone->uz_size, flags);
4332 		kasan_mark_item_invalid(zone, item);
4333 		if (error != 0) {
4334 			zone_free_item(zone, item, udata, SKIP_FINI | SKIP_CNT);
4335 			goto fail_cnt;
4336 		}
4337 	}
4338 	item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata, flags,
4339 	    item);
4340 	if (item == NULL)
4341 		goto fail;
4342 
4343 	counter_u64_add(zone->uz_allocs, 1);
4344 	CTR3(KTR_UMA, "zone_alloc_item item %p from %s(%p)", item,
4345 	    zone->uz_name, zone);
4346 
4347 	return (item);
4348 
4349 fail_cnt:
4350 	counter_u64_add(zone->uz_fails, 1);
4351 fail:
4352 	if (zone->uz_max_items > 0)
4353 		zone_free_limit(zone, 1);
4354 	CTR2(KTR_UMA, "zone_alloc_item failed from %s(%p)",
4355 	    zone->uz_name, zone);
4356 
4357 	return (NULL);
4358 }
4359 
4360 /* See uma.h */
4361 void
4362 uma_zfree_smr(uma_zone_t zone, void *item)
4363 {
4364 	uma_cache_t cache;
4365 	uma_cache_bucket_t bucket;
4366 	int itemdomain, uz_flags;
4367 
4368 #ifdef UMA_ZALLOC_DEBUG
4369 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
4370 	    ("uma_zfree_smr: called with non-SMR zone."));
4371 	KASSERT(item != NULL, ("uma_zfree_smr: Called with NULL pointer."));
4372 	SMR_ASSERT_NOT_ENTERED(zone->uz_smr);
4373 	if (uma_zfree_debug(zone, item, NULL) == EJUSTRETURN)
4374 		return;
4375 #endif
4376 	cache = &zone->uz_cpu[curcpu];
4377 	uz_flags = cache_uz_flags(cache);
4378 	itemdomain = 0;
4379 #ifdef NUMA
4380 	if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4381 		itemdomain = item_domain(item);
4382 #endif
4383 	critical_enter();
4384 	do {
4385 		cache = &zone->uz_cpu[curcpu];
4386 		/* SMR Zones must free to the free bucket. */
4387 		bucket = &cache->uc_freebucket;
4388 #ifdef NUMA
4389 		if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4390 		    PCPU_GET(domain) != itemdomain) {
4391 			bucket = &cache->uc_crossbucket;
4392 		}
4393 #endif
4394 		if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4395 			cache_bucket_push(cache, bucket, item);
4396 			critical_exit();
4397 			return;
4398 		}
4399 	} while (cache_free(zone, cache, NULL, item, itemdomain));
4400 	critical_exit();
4401 
4402 	/*
4403 	 * If nothing else caught this, we'll just do an internal free.
4404 	 */
4405 	zone_free_item(zone, item, NULL, SKIP_NONE);
4406 }
4407 
4408 /* See uma.h */
4409 void
4410 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
4411 {
4412 	uma_cache_t cache;
4413 	uma_cache_bucket_t bucket;
4414 	int itemdomain, uz_flags;
4415 
4416 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
4417 	random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
4418 
4419 	CTR2(KTR_UMA, "uma_zfree_arg zone %s(%p)", zone->uz_name, zone);
4420 
4421 #ifdef UMA_ZALLOC_DEBUG
4422 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
4423 	    ("uma_zfree_arg: called with SMR zone."));
4424 	if (uma_zfree_debug(zone, item, udata) == EJUSTRETURN)
4425 		return;
4426 #endif
4427         /* uma_zfree(..., NULL) does nothing, to match free(9). */
4428         if (item == NULL)
4429                 return;
4430 
4431 	/*
4432 	 * We are accessing the per-cpu cache without a critical section to
4433 	 * fetch size and flags.  This is acceptable, if we are preempted we
4434 	 * will simply read another cpu's line.
4435 	 */
4436 	cache = &zone->uz_cpu[curcpu];
4437 	uz_flags = cache_uz_flags(cache);
4438 	if (UMA_ALWAYS_CTORDTOR ||
4439 	    __predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0))
4440 		item_dtor(zone, item, cache_uz_size(cache), udata, SKIP_NONE);
4441 
4442 	/*
4443 	 * The race here is acceptable.  If we miss it we'll just have to wait
4444 	 * a little longer for the limits to be reset.
4445 	 */
4446 	if (__predict_false(uz_flags & UMA_ZFLAG_LIMIT)) {
4447 		if (atomic_load_32(&zone->uz_sleepers) > 0)
4448 			goto zfree_item;
4449 	}
4450 
4451 	/*
4452 	 * If possible, free to the per-CPU cache.  There are two
4453 	 * requirements for safe access to the per-CPU cache: (1) the thread
4454 	 * accessing the cache must not be preempted or yield during access,
4455 	 * and (2) the thread must not migrate CPUs without switching which
4456 	 * cache it accesses.  We rely on a critical section to prevent
4457 	 * preemption and migration.  We release the critical section in
4458 	 * order to acquire the zone mutex if we are unable to free to the
4459 	 * current cache; when we re-acquire the critical section, we must
4460 	 * detect and handle migration if it has occurred.
4461 	 */
4462 	itemdomain = 0;
4463 #ifdef NUMA
4464 	if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4465 		itemdomain = item_domain(item);
4466 #endif
4467 	critical_enter();
4468 	do {
4469 		cache = &zone->uz_cpu[curcpu];
4470 		/*
4471 		 * Try to free into the allocbucket first to give LIFO
4472 		 * ordering for cache-hot datastructures.  Spill over
4473 		 * into the freebucket if necessary.  Alloc will swap
4474 		 * them if one runs dry.
4475 		 */
4476 		bucket = &cache->uc_allocbucket;
4477 #ifdef NUMA
4478 		if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4479 		    PCPU_GET(domain) != itemdomain) {
4480 			bucket = &cache->uc_crossbucket;
4481 		} else
4482 #endif
4483 		if (bucket->ucb_cnt == bucket->ucb_entries &&
4484 		   cache->uc_freebucket.ucb_cnt <
4485 		   cache->uc_freebucket.ucb_entries)
4486 			cache_bucket_swap(&cache->uc_freebucket,
4487 			    &cache->uc_allocbucket);
4488 		if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4489 			cache_bucket_push(cache, bucket, item);
4490 			critical_exit();
4491 			return;
4492 		}
4493 	} while (cache_free(zone, cache, udata, item, itemdomain));
4494 	critical_exit();
4495 
4496 	/*
4497 	 * If nothing else caught this, we'll just do an internal free.
4498 	 */
4499 zfree_item:
4500 	zone_free_item(zone, item, udata, SKIP_DTOR);
4501 }
4502 
4503 #ifdef NUMA
4504 /*
4505  * sort crossdomain free buckets to domain correct buckets and cache
4506  * them.
4507  */
4508 static void
4509 zone_free_cross(uma_zone_t zone, uma_bucket_t bucket, void *udata)
4510 {
4511 	struct uma_bucketlist emptybuckets, fullbuckets;
4512 	uma_zone_domain_t zdom;
4513 	uma_bucket_t b;
4514 	smr_seq_t seq;
4515 	void *item;
4516 	int domain;
4517 
4518 	CTR3(KTR_UMA,
4519 	    "uma_zfree: zone %s(%p) draining cross bucket %p",
4520 	    zone->uz_name, zone, bucket);
4521 
4522 	/*
4523 	 * It is possible for buckets to arrive here out of order so we fetch
4524 	 * the current smr seq rather than accepting the bucket's.
4525 	 */
4526 	seq = SMR_SEQ_INVALID;
4527 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
4528 		seq = smr_advance(zone->uz_smr);
4529 
4530 	/*
4531 	 * To avoid having ndomain * ndomain buckets for sorting we have a
4532 	 * lock on the current crossfree bucket.  A full matrix with
4533 	 * per-domain locking could be used if necessary.
4534 	 */
4535 	STAILQ_INIT(&emptybuckets);
4536 	STAILQ_INIT(&fullbuckets);
4537 	ZONE_CROSS_LOCK(zone);
4538 	for (; bucket->ub_cnt > 0; bucket->ub_cnt--) {
4539 		item = bucket->ub_bucket[bucket->ub_cnt - 1];
4540 		domain = item_domain(item);
4541 		zdom = ZDOM_GET(zone, domain);
4542 		if (zdom->uzd_cross == NULL) {
4543 			if ((b = STAILQ_FIRST(&emptybuckets)) != NULL) {
4544 				STAILQ_REMOVE_HEAD(&emptybuckets, ub_link);
4545 				zdom->uzd_cross = b;
4546 			} else {
4547 				/*
4548 				 * Avoid allocating a bucket with the cross lock
4549 				 * held, since allocation can trigger a
4550 				 * cross-domain free and bucket zones may
4551 				 * allocate from each other.
4552 				 */
4553 				ZONE_CROSS_UNLOCK(zone);
4554 				b = bucket_alloc(zone, udata, M_NOWAIT);
4555 				if (b == NULL)
4556 					goto out;
4557 				ZONE_CROSS_LOCK(zone);
4558 				if (zdom->uzd_cross != NULL) {
4559 					STAILQ_INSERT_HEAD(&emptybuckets, b,
4560 					    ub_link);
4561 				} else {
4562 					zdom->uzd_cross = b;
4563 				}
4564 			}
4565 		}
4566 		b = zdom->uzd_cross;
4567 		b->ub_bucket[b->ub_cnt++] = item;
4568 		b->ub_seq = seq;
4569 		if (b->ub_cnt == b->ub_entries) {
4570 			STAILQ_INSERT_HEAD(&fullbuckets, b, ub_link);
4571 			if ((b = STAILQ_FIRST(&emptybuckets)) != NULL)
4572 				STAILQ_REMOVE_HEAD(&emptybuckets, ub_link);
4573 			zdom->uzd_cross = b;
4574 		}
4575 	}
4576 	ZONE_CROSS_UNLOCK(zone);
4577 out:
4578 	if (bucket->ub_cnt == 0)
4579 		bucket->ub_seq = SMR_SEQ_INVALID;
4580 	bucket_free(zone, bucket, udata);
4581 
4582 	while ((b = STAILQ_FIRST(&emptybuckets)) != NULL) {
4583 		STAILQ_REMOVE_HEAD(&emptybuckets, ub_link);
4584 		bucket_free(zone, b, udata);
4585 	}
4586 	while ((b = STAILQ_FIRST(&fullbuckets)) != NULL) {
4587 		STAILQ_REMOVE_HEAD(&fullbuckets, ub_link);
4588 		domain = item_domain(b->ub_bucket[0]);
4589 		zone_put_bucket(zone, domain, b, udata, true);
4590 	}
4591 }
4592 #endif
4593 
4594 static void
4595 zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata,
4596     int itemdomain, bool ws)
4597 {
4598 
4599 #ifdef NUMA
4600 	/*
4601 	 * Buckets coming from the wrong domain will be entirely for the
4602 	 * only other domain on two domain systems.  In this case we can
4603 	 * simply cache them.  Otherwise we need to sort them back to
4604 	 * correct domains.
4605 	 */
4606 	if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4607 	    vm_ndomains > 2 && PCPU_GET(domain) != itemdomain) {
4608 		zone_free_cross(zone, bucket, udata);
4609 		return;
4610 	}
4611 #endif
4612 
4613 	/*
4614 	 * Attempt to save the bucket in the zone's domain bucket cache.
4615 	 */
4616 	CTR3(KTR_UMA,
4617 	    "uma_zfree: zone %s(%p) putting bucket %p on free list",
4618 	    zone->uz_name, zone, bucket);
4619 	/* ub_cnt is pointing to the last free item */
4620 	if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
4621 		itemdomain = zone_domain_lowest(zone, itemdomain);
4622 	zone_put_bucket(zone, itemdomain, bucket, udata, ws);
4623 }
4624 
4625 /*
4626  * Populate a free or cross bucket for the current cpu cache.  Free any
4627  * existing full bucket either to the zone cache or back to the slab layer.
4628  *
4629  * Enters and returns in a critical section.  false return indicates that
4630  * we can not satisfy this free in the cache layer.  true indicates that
4631  * the caller should retry.
4632  */
4633 static __noinline bool
4634 cache_free(uma_zone_t zone, uma_cache_t cache, void *udata, void *item,
4635     int itemdomain)
4636 {
4637 	uma_cache_bucket_t cbucket;
4638 	uma_bucket_t newbucket, bucket;
4639 
4640 	CRITICAL_ASSERT(curthread);
4641 
4642 	if (zone->uz_bucket_size == 0)
4643 		return false;
4644 
4645 	cache = &zone->uz_cpu[curcpu];
4646 	newbucket = NULL;
4647 
4648 	/*
4649 	 * FIRSTTOUCH domains need to free to the correct zdom.  When
4650 	 * enabled this is the zdom of the item.   The bucket is the
4651 	 * cross bucket if the current domain and itemdomain do not match.
4652 	 */
4653 	cbucket = &cache->uc_freebucket;
4654 #ifdef NUMA
4655 	if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4656 		if (PCPU_GET(domain) != itemdomain) {
4657 			cbucket = &cache->uc_crossbucket;
4658 			if (cbucket->ucb_cnt != 0)
4659 				counter_u64_add(zone->uz_xdomain,
4660 				    cbucket->ucb_cnt);
4661 		}
4662 	}
4663 #endif
4664 	bucket = cache_bucket_unload(cbucket);
4665 	KASSERT(bucket == NULL || bucket->ub_cnt == bucket->ub_entries,
4666 	    ("cache_free: Entered with non-full free bucket."));
4667 
4668 	/* We are no longer associated with this CPU. */
4669 	critical_exit();
4670 
4671 	/*
4672 	 * Don't let SMR zones operate without a free bucket.  Force
4673 	 * a synchronize and re-use this one.  We will only degrade
4674 	 * to a synchronize every bucket_size items rather than every
4675 	 * item if we fail to allocate a bucket.
4676 	 */
4677 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0) {
4678 		if (bucket != NULL)
4679 			bucket->ub_seq = smr_advance(zone->uz_smr);
4680 		newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4681 		if (newbucket == NULL && bucket != NULL) {
4682 			bucket_drain(zone, bucket);
4683 			newbucket = bucket;
4684 			bucket = NULL;
4685 		}
4686 	} else if (!bucketdisable)
4687 		newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4688 
4689 	if (bucket != NULL)
4690 		zone_free_bucket(zone, bucket, udata, itemdomain, true);
4691 
4692 	critical_enter();
4693 	if ((bucket = newbucket) == NULL)
4694 		return (false);
4695 	cache = &zone->uz_cpu[curcpu];
4696 #ifdef NUMA
4697 	/*
4698 	 * Check to see if we should be populating the cross bucket.  If it
4699 	 * is already populated we will fall through and attempt to populate
4700 	 * the free bucket.
4701 	 */
4702 	if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4703 		if (PCPU_GET(domain) != itemdomain &&
4704 		    cache->uc_crossbucket.ucb_bucket == NULL) {
4705 			cache_bucket_load_cross(cache, bucket);
4706 			return (true);
4707 		}
4708 	}
4709 #endif
4710 	/*
4711 	 * We may have lost the race to fill the bucket or switched CPUs.
4712 	 */
4713 	if (cache->uc_freebucket.ucb_bucket != NULL) {
4714 		critical_exit();
4715 		bucket_free(zone, bucket, udata);
4716 		critical_enter();
4717 	} else
4718 		cache_bucket_load_free(cache, bucket);
4719 
4720 	return (true);
4721 }
4722 
4723 static void
4724 slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item)
4725 {
4726 	uma_keg_t keg;
4727 	uma_domain_t dom;
4728 	int freei;
4729 
4730 	keg = zone->uz_keg;
4731 	KEG_LOCK_ASSERT(keg, slab->us_domain);
4732 
4733 	/* Do we need to remove from any lists? */
4734 	dom = &keg->uk_domain[slab->us_domain];
4735 	if (slab->us_freecount + 1 == keg->uk_ipers) {
4736 		LIST_REMOVE(slab, us_link);
4737 		LIST_INSERT_HEAD(&dom->ud_free_slab, slab, us_link);
4738 		dom->ud_free_slabs++;
4739 	} else if (slab->us_freecount == 0) {
4740 		LIST_REMOVE(slab, us_link);
4741 		LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
4742 	}
4743 
4744 	/* Slab management. */
4745 	freei = slab_item_index(slab, keg, item);
4746 	BIT_SET(keg->uk_ipers, freei, &slab->us_free);
4747 	slab->us_freecount++;
4748 
4749 	/* Keg statistics. */
4750 	dom->ud_free_items++;
4751 }
4752 
4753 static void
4754 zone_release(void *arg, void **bucket, int cnt)
4755 {
4756 	struct mtx *lock;
4757 	uma_zone_t zone;
4758 	uma_slab_t slab;
4759 	uma_keg_t keg;
4760 	uint8_t *mem;
4761 	void *item;
4762 	int i;
4763 
4764 	zone = arg;
4765 	keg = zone->uz_keg;
4766 	lock = NULL;
4767 	if (__predict_false((zone->uz_flags & UMA_ZFLAG_HASH) != 0))
4768 		lock = KEG_LOCK(keg, 0);
4769 	for (i = 0; i < cnt; i++) {
4770 		item = bucket[i];
4771 		if (__predict_true((zone->uz_flags & UMA_ZFLAG_VTOSLAB) != 0)) {
4772 			slab = vtoslab((vm_offset_t)item);
4773 		} else {
4774 			mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
4775 			if ((zone->uz_flags & UMA_ZFLAG_HASH) != 0)
4776 				slab = hash_sfind(&keg->uk_hash, mem);
4777 			else
4778 				slab = (uma_slab_t)(mem + keg->uk_pgoff);
4779 		}
4780 		if (lock != KEG_LOCKPTR(keg, slab->us_domain)) {
4781 			if (lock != NULL)
4782 				mtx_unlock(lock);
4783 			lock = KEG_LOCK(keg, slab->us_domain);
4784 		}
4785 		slab_free_item(zone, slab, item);
4786 	}
4787 	if (lock != NULL)
4788 		mtx_unlock(lock);
4789 }
4790 
4791 /*
4792  * Frees a single item to any zone.
4793  *
4794  * Arguments:
4795  *	zone   The zone to free to
4796  *	item   The item we're freeing
4797  *	udata  User supplied data for the dtor
4798  *	skip   Skip dtors and finis
4799  */
4800 static __noinline void
4801 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
4802 {
4803 
4804 	/*
4805 	 * If a free is sent directly to an SMR zone we have to
4806 	 * synchronize immediately because the item can instantly
4807 	 * be reallocated. This should only happen in degenerate
4808 	 * cases when no memory is available for per-cpu caches.
4809 	 */
4810 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0 && skip == SKIP_NONE)
4811 		smr_synchronize(zone->uz_smr);
4812 
4813 	item_dtor(zone, item, zone->uz_size, udata, skip);
4814 
4815 	if (skip < SKIP_FINI && zone->uz_fini) {
4816 		kasan_mark_item_valid(zone, item);
4817 		zone->uz_fini(item, zone->uz_size);
4818 		kasan_mark_item_invalid(zone, item);
4819 	}
4820 
4821 	zone->uz_release(zone->uz_arg, &item, 1);
4822 
4823 	if (skip & SKIP_CNT)
4824 		return;
4825 
4826 	counter_u64_add(zone->uz_frees, 1);
4827 
4828 	if (zone->uz_max_items > 0)
4829 		zone_free_limit(zone, 1);
4830 }
4831 
4832 /* See uma.h */
4833 int
4834 uma_zone_set_max(uma_zone_t zone, int nitems)
4835 {
4836 
4837 	/*
4838 	 * If the limit is small, we may need to constrain the maximum per-CPU
4839 	 * cache size, or disable caching entirely.
4840 	 */
4841 	uma_zone_set_maxcache(zone, nitems);
4842 
4843 	/*
4844 	 * XXX This can misbehave if the zone has any allocations with
4845 	 * no limit and a limit is imposed.  There is currently no
4846 	 * way to clear a limit.
4847 	 */
4848 	ZONE_LOCK(zone);
4849 	zone->uz_max_items = nitems;
4850 	zone->uz_flags |= UMA_ZFLAG_LIMIT;
4851 	zone_update_caches(zone);
4852 	/* We may need to wake waiters. */
4853 	wakeup(&zone->uz_max_items);
4854 	ZONE_UNLOCK(zone);
4855 
4856 	return (nitems);
4857 }
4858 
4859 /* See uma.h */
4860 void
4861 uma_zone_set_maxcache(uma_zone_t zone, int nitems)
4862 {
4863 	int bpcpu, bpdom, bsize, nb;
4864 
4865 	ZONE_LOCK(zone);
4866 
4867 	/*
4868 	 * Compute a lower bound on the number of items that may be cached in
4869 	 * the zone.  Each CPU gets at least two buckets, and for cross-domain
4870 	 * frees we use an additional bucket per CPU and per domain.  Select the
4871 	 * largest bucket size that does not exceed half of the requested limit,
4872 	 * with the left over space given to the full bucket cache.
4873 	 */
4874 	bpdom = 0;
4875 	bpcpu = 2;
4876 #ifdef NUMA
4877 	if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 && vm_ndomains > 1) {
4878 		bpcpu++;
4879 		bpdom++;
4880 	}
4881 #endif
4882 	nb = bpcpu * mp_ncpus + bpdom * vm_ndomains;
4883 	bsize = nitems / nb / 2;
4884 	if (bsize > BUCKET_MAX)
4885 		bsize = BUCKET_MAX;
4886 	else if (bsize == 0 && nitems / nb > 0)
4887 		bsize = 1;
4888 	zone->uz_bucket_size_max = zone->uz_bucket_size = bsize;
4889 	if (zone->uz_bucket_size_min > zone->uz_bucket_size_max)
4890 		zone->uz_bucket_size_min = zone->uz_bucket_size_max;
4891 	zone->uz_bucket_max = nitems - nb * bsize;
4892 	ZONE_UNLOCK(zone);
4893 }
4894 
4895 /* See uma.h */
4896 int
4897 uma_zone_get_max(uma_zone_t zone)
4898 {
4899 	int nitems;
4900 
4901 	nitems = atomic_load_64(&zone->uz_max_items);
4902 
4903 	return (nitems);
4904 }
4905 
4906 /* See uma.h */
4907 void
4908 uma_zone_set_warning(uma_zone_t zone, const char *warning)
4909 {
4910 
4911 	ZONE_ASSERT_COLD(zone);
4912 	zone->uz_warning = warning;
4913 }
4914 
4915 /* See uma.h */
4916 void
4917 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction)
4918 {
4919 
4920 	ZONE_ASSERT_COLD(zone);
4921 	TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone);
4922 }
4923 
4924 /* See uma.h */
4925 int
4926 uma_zone_get_cur(uma_zone_t zone)
4927 {
4928 	int64_t nitems;
4929 	u_int i;
4930 
4931 	nitems = 0;
4932 	if (zone->uz_allocs != EARLY_COUNTER && zone->uz_frees != EARLY_COUNTER)
4933 		nitems = counter_u64_fetch(zone->uz_allocs) -
4934 		    counter_u64_fetch(zone->uz_frees);
4935 	CPU_FOREACH(i)
4936 		nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs) -
4937 		    atomic_load_64(&zone->uz_cpu[i].uc_frees);
4938 
4939 	return (nitems < 0 ? 0 : nitems);
4940 }
4941 
4942 static uint64_t
4943 uma_zone_get_allocs(uma_zone_t zone)
4944 {
4945 	uint64_t nitems;
4946 	u_int i;
4947 
4948 	nitems = 0;
4949 	if (zone->uz_allocs != EARLY_COUNTER)
4950 		nitems = counter_u64_fetch(zone->uz_allocs);
4951 	CPU_FOREACH(i)
4952 		nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs);
4953 
4954 	return (nitems);
4955 }
4956 
4957 static uint64_t
4958 uma_zone_get_frees(uma_zone_t zone)
4959 {
4960 	uint64_t nitems;
4961 	u_int i;
4962 
4963 	nitems = 0;
4964 	if (zone->uz_frees != EARLY_COUNTER)
4965 		nitems = counter_u64_fetch(zone->uz_frees);
4966 	CPU_FOREACH(i)
4967 		nitems += atomic_load_64(&zone->uz_cpu[i].uc_frees);
4968 
4969 	return (nitems);
4970 }
4971 
4972 #ifdef INVARIANTS
4973 /* Used only for KEG_ASSERT_COLD(). */
4974 static uint64_t
4975 uma_keg_get_allocs(uma_keg_t keg)
4976 {
4977 	uma_zone_t z;
4978 	uint64_t nitems;
4979 
4980 	nitems = 0;
4981 	LIST_FOREACH(z, &keg->uk_zones, uz_link)
4982 		nitems += uma_zone_get_allocs(z);
4983 
4984 	return (nitems);
4985 }
4986 #endif
4987 
4988 /* See uma.h */
4989 void
4990 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
4991 {
4992 	uma_keg_t keg;
4993 
4994 	KEG_GET(zone, keg);
4995 	KEG_ASSERT_COLD(keg);
4996 	keg->uk_init = uminit;
4997 }
4998 
4999 /* See uma.h */
5000 void
5001 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
5002 {
5003 	uma_keg_t keg;
5004 
5005 	KEG_GET(zone, keg);
5006 	KEG_ASSERT_COLD(keg);
5007 	keg->uk_fini = fini;
5008 }
5009 
5010 /* See uma.h */
5011 void
5012 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
5013 {
5014 
5015 	ZONE_ASSERT_COLD(zone);
5016 	zone->uz_init = zinit;
5017 }
5018 
5019 /* See uma.h */
5020 void
5021 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
5022 {
5023 
5024 	ZONE_ASSERT_COLD(zone);
5025 	zone->uz_fini = zfini;
5026 }
5027 
5028 /* See uma.h */
5029 void
5030 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
5031 {
5032 	uma_keg_t keg;
5033 
5034 	KEG_GET(zone, keg);
5035 	KEG_ASSERT_COLD(keg);
5036 	keg->uk_freef = freef;
5037 }
5038 
5039 /* See uma.h */
5040 void
5041 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
5042 {
5043 	uma_keg_t keg;
5044 
5045 	KEG_GET(zone, keg);
5046 	KEG_ASSERT_COLD(keg);
5047 	keg->uk_allocf = allocf;
5048 }
5049 
5050 /* See uma.h */
5051 void
5052 uma_zone_set_smr(uma_zone_t zone, smr_t smr)
5053 {
5054 
5055 	ZONE_ASSERT_COLD(zone);
5056 
5057 	KASSERT(smr != NULL, ("Got NULL smr"));
5058 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
5059 	    ("zone %p (%s) already uses SMR", zone, zone->uz_name));
5060 	zone->uz_flags |= UMA_ZONE_SMR;
5061 	zone->uz_smr = smr;
5062 	zone_update_caches(zone);
5063 }
5064 
5065 smr_t
5066 uma_zone_get_smr(uma_zone_t zone)
5067 {
5068 
5069 	return (zone->uz_smr);
5070 }
5071 
5072 /* See uma.h */
5073 void
5074 uma_zone_reserve(uma_zone_t zone, int items)
5075 {
5076 	uma_keg_t keg;
5077 
5078 	KEG_GET(zone, keg);
5079 	KEG_ASSERT_COLD(keg);
5080 	keg->uk_reserve = items;
5081 }
5082 
5083 /* See uma.h */
5084 int
5085 uma_zone_reserve_kva(uma_zone_t zone, int count)
5086 {
5087 	uma_keg_t keg;
5088 	vm_offset_t kva;
5089 	u_int pages;
5090 
5091 	KEG_GET(zone, keg);
5092 	KEG_ASSERT_COLD(keg);
5093 	ZONE_ASSERT_COLD(zone);
5094 
5095 	pages = howmany(count, keg->uk_ipers) * keg->uk_ppera;
5096 
5097 #ifdef UMA_MD_SMALL_ALLOC
5098 	if (keg->uk_ppera > 1) {
5099 #else
5100 	if (1) {
5101 #endif
5102 		kva = kva_alloc((vm_size_t)pages * PAGE_SIZE);
5103 		if (kva == 0)
5104 			return (0);
5105 	} else
5106 		kva = 0;
5107 
5108 	MPASS(keg->uk_kva == 0);
5109 	keg->uk_kva = kva;
5110 	keg->uk_offset = 0;
5111 	zone->uz_max_items = pages * keg->uk_ipers;
5112 #ifdef UMA_MD_SMALL_ALLOC
5113 	keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
5114 #else
5115 	keg->uk_allocf = noobj_alloc;
5116 #endif
5117 	keg->uk_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
5118 	zone->uz_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
5119 	zone_update_caches(zone);
5120 
5121 	return (1);
5122 }
5123 
5124 /* See uma.h */
5125 void
5126 uma_prealloc(uma_zone_t zone, int items)
5127 {
5128 	struct vm_domainset_iter di;
5129 	uma_domain_t dom;
5130 	uma_slab_t slab;
5131 	uma_keg_t keg;
5132 	int aflags, domain, slabs;
5133 
5134 	KEG_GET(zone, keg);
5135 	slabs = howmany(items, keg->uk_ipers);
5136 	while (slabs-- > 0) {
5137 		aflags = M_NOWAIT;
5138 		vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
5139 		    &aflags);
5140 		for (;;) {
5141 			slab = keg_alloc_slab(keg, zone, domain, M_WAITOK,
5142 			    aflags);
5143 			if (slab != NULL) {
5144 				dom = &keg->uk_domain[slab->us_domain];
5145 				/*
5146 				 * keg_alloc_slab() always returns a slab on the
5147 				 * partial list.
5148 				 */
5149 				LIST_REMOVE(slab, us_link);
5150 				LIST_INSERT_HEAD(&dom->ud_free_slab, slab,
5151 				    us_link);
5152 				dom->ud_free_slabs++;
5153 				KEG_UNLOCK(keg, slab->us_domain);
5154 				break;
5155 			}
5156 			if (vm_domainset_iter_policy(&di, &domain) != 0)
5157 				vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask, 0);
5158 		}
5159 	}
5160 }
5161 
5162 /*
5163  * Returns a snapshot of memory consumption in bytes.
5164  */
5165 size_t
5166 uma_zone_memory(uma_zone_t zone)
5167 {
5168 	size_t sz;
5169 	int i;
5170 
5171 	sz = 0;
5172 	if (zone->uz_flags & UMA_ZFLAG_CACHE) {
5173 		for (i = 0; i < vm_ndomains; i++)
5174 			sz += ZDOM_GET(zone, i)->uzd_nitems;
5175 		return (sz * zone->uz_size);
5176 	}
5177 	for (i = 0; i < vm_ndomains; i++)
5178 		sz += zone->uz_keg->uk_domain[i].ud_pages;
5179 
5180 	return (sz * PAGE_SIZE);
5181 }
5182 
5183 /* See uma.h */
5184 void
5185 uma_reclaim(int req)
5186 {
5187 	uma_reclaim_domain(req, UMA_ANYDOMAIN);
5188 }
5189 
5190 void
5191 uma_reclaim_domain(int req, int domain)
5192 {
5193 	void *arg;
5194 
5195 	bucket_enable();
5196 
5197 	arg = (void *)(uintptr_t)domain;
5198 	sx_slock(&uma_reclaim_lock);
5199 	switch (req) {
5200 	case UMA_RECLAIM_TRIM:
5201 		zone_foreach(zone_trim, arg);
5202 		break;
5203 	case UMA_RECLAIM_DRAIN:
5204 		zone_foreach(zone_drain, arg);
5205 		break;
5206 	case UMA_RECLAIM_DRAIN_CPU:
5207 		zone_foreach(zone_drain, arg);
5208 		pcpu_cache_drain_safe(NULL);
5209 		zone_foreach(zone_drain, arg);
5210 		break;
5211 	default:
5212 		panic("unhandled reclamation request %d", req);
5213 	}
5214 
5215 	/*
5216 	 * Some slabs may have been freed but this zone will be visited early
5217 	 * we visit again so that we can free pages that are empty once other
5218 	 * zones are drained.  We have to do the same for buckets.
5219 	 */
5220 	zone_drain(slabzones[0], arg);
5221 	zone_drain(slabzones[1], arg);
5222 	bucket_zone_drain(domain);
5223 	sx_sunlock(&uma_reclaim_lock);
5224 }
5225 
5226 static volatile int uma_reclaim_needed;
5227 
5228 void
5229 uma_reclaim_wakeup(void)
5230 {
5231 
5232 	if (atomic_fetchadd_int(&uma_reclaim_needed, 1) == 0)
5233 		wakeup(uma_reclaim);
5234 }
5235 
5236 void
5237 uma_reclaim_worker(void *arg __unused)
5238 {
5239 
5240 	for (;;) {
5241 		sx_xlock(&uma_reclaim_lock);
5242 		while (atomic_load_int(&uma_reclaim_needed) == 0)
5243 			sx_sleep(uma_reclaim, &uma_reclaim_lock, PVM, "umarcl",
5244 			    hz);
5245 		sx_xunlock(&uma_reclaim_lock);
5246 		EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM);
5247 		uma_reclaim(UMA_RECLAIM_DRAIN_CPU);
5248 		atomic_store_int(&uma_reclaim_needed, 0);
5249 		/* Don't fire more than once per-second. */
5250 		pause("umarclslp", hz);
5251 	}
5252 }
5253 
5254 /* See uma.h */
5255 void
5256 uma_zone_reclaim(uma_zone_t zone, int req)
5257 {
5258 	uma_zone_reclaim_domain(zone, req, UMA_ANYDOMAIN);
5259 }
5260 
5261 void
5262 uma_zone_reclaim_domain(uma_zone_t zone, int req, int domain)
5263 {
5264 	void *arg;
5265 
5266 	arg = (void *)(uintptr_t)domain;
5267 	switch (req) {
5268 	case UMA_RECLAIM_TRIM:
5269 		zone_trim(zone, arg);
5270 		break;
5271 	case UMA_RECLAIM_DRAIN:
5272 		zone_drain(zone, arg);
5273 		break;
5274 	case UMA_RECLAIM_DRAIN_CPU:
5275 		pcpu_cache_drain_safe(zone);
5276 		zone_drain(zone, arg);
5277 		break;
5278 	default:
5279 		panic("unhandled reclamation request %d", req);
5280 	}
5281 }
5282 
5283 /* See uma.h */
5284 int
5285 uma_zone_exhausted(uma_zone_t zone)
5286 {
5287 
5288 	return (atomic_load_32(&zone->uz_sleepers) > 0);
5289 }
5290 
5291 unsigned long
5292 uma_limit(void)
5293 {
5294 
5295 	return (uma_kmem_limit);
5296 }
5297 
5298 void
5299 uma_set_limit(unsigned long limit)
5300 {
5301 
5302 	uma_kmem_limit = limit;
5303 }
5304 
5305 unsigned long
5306 uma_size(void)
5307 {
5308 
5309 	return (atomic_load_long(&uma_kmem_total));
5310 }
5311 
5312 long
5313 uma_avail(void)
5314 {
5315 
5316 	return (uma_kmem_limit - uma_size());
5317 }
5318 
5319 #ifdef DDB
5320 /*
5321  * Generate statistics across both the zone and its per-cpu cache's.  Return
5322  * desired statistics if the pointer is non-NULL for that statistic.
5323  *
5324  * Note: does not update the zone statistics, as it can't safely clear the
5325  * per-CPU cache statistic.
5326  *
5327  */
5328 static void
5329 uma_zone_sumstat(uma_zone_t z, long *cachefreep, uint64_t *allocsp,
5330     uint64_t *freesp, uint64_t *sleepsp, uint64_t *xdomainp)
5331 {
5332 	uma_cache_t cache;
5333 	uint64_t allocs, frees, sleeps, xdomain;
5334 	int cachefree, cpu;
5335 
5336 	allocs = frees = sleeps = xdomain = 0;
5337 	cachefree = 0;
5338 	CPU_FOREACH(cpu) {
5339 		cache = &z->uz_cpu[cpu];
5340 		cachefree += cache->uc_allocbucket.ucb_cnt;
5341 		cachefree += cache->uc_freebucket.ucb_cnt;
5342 		xdomain += cache->uc_crossbucket.ucb_cnt;
5343 		cachefree += cache->uc_crossbucket.ucb_cnt;
5344 		allocs += cache->uc_allocs;
5345 		frees += cache->uc_frees;
5346 	}
5347 	allocs += counter_u64_fetch(z->uz_allocs);
5348 	frees += counter_u64_fetch(z->uz_frees);
5349 	xdomain += counter_u64_fetch(z->uz_xdomain);
5350 	sleeps += z->uz_sleeps;
5351 	if (cachefreep != NULL)
5352 		*cachefreep = cachefree;
5353 	if (allocsp != NULL)
5354 		*allocsp = allocs;
5355 	if (freesp != NULL)
5356 		*freesp = frees;
5357 	if (sleepsp != NULL)
5358 		*sleepsp = sleeps;
5359 	if (xdomainp != NULL)
5360 		*xdomainp = xdomain;
5361 }
5362 #endif /* DDB */
5363 
5364 static int
5365 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
5366 {
5367 	uma_keg_t kz;
5368 	uma_zone_t z;
5369 	int count;
5370 
5371 	count = 0;
5372 	rw_rlock(&uma_rwlock);
5373 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
5374 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
5375 			count++;
5376 	}
5377 	LIST_FOREACH(z, &uma_cachezones, uz_link)
5378 		count++;
5379 
5380 	rw_runlock(&uma_rwlock);
5381 	return (sysctl_handle_int(oidp, &count, 0, req));
5382 }
5383 
5384 static void
5385 uma_vm_zone_stats(struct uma_type_header *uth, uma_zone_t z, struct sbuf *sbuf,
5386     struct uma_percpu_stat *ups, bool internal)
5387 {
5388 	uma_zone_domain_t zdom;
5389 	uma_cache_t cache;
5390 	int i;
5391 
5392 	for (i = 0; i < vm_ndomains; i++) {
5393 		zdom = ZDOM_GET(z, i);
5394 		uth->uth_zone_free += zdom->uzd_nitems;
5395 	}
5396 	uth->uth_allocs = counter_u64_fetch(z->uz_allocs);
5397 	uth->uth_frees = counter_u64_fetch(z->uz_frees);
5398 	uth->uth_fails = counter_u64_fetch(z->uz_fails);
5399 	uth->uth_xdomain = counter_u64_fetch(z->uz_xdomain);
5400 	uth->uth_sleeps = z->uz_sleeps;
5401 
5402 	for (i = 0; i < mp_maxid + 1; i++) {
5403 		bzero(&ups[i], sizeof(*ups));
5404 		if (internal || CPU_ABSENT(i))
5405 			continue;
5406 		cache = &z->uz_cpu[i];
5407 		ups[i].ups_cache_free += cache->uc_allocbucket.ucb_cnt;
5408 		ups[i].ups_cache_free += cache->uc_freebucket.ucb_cnt;
5409 		ups[i].ups_cache_free += cache->uc_crossbucket.ucb_cnt;
5410 		ups[i].ups_allocs = cache->uc_allocs;
5411 		ups[i].ups_frees = cache->uc_frees;
5412 	}
5413 }
5414 
5415 static int
5416 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
5417 {
5418 	struct uma_stream_header ush;
5419 	struct uma_type_header uth;
5420 	struct uma_percpu_stat *ups;
5421 	struct sbuf sbuf;
5422 	uma_keg_t kz;
5423 	uma_zone_t z;
5424 	uint64_t items;
5425 	uint32_t kfree, pages;
5426 	int count, error, i;
5427 
5428 	error = sysctl_wire_old_buffer(req, 0);
5429 	if (error != 0)
5430 		return (error);
5431 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
5432 	sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL);
5433 	ups = malloc((mp_maxid + 1) * sizeof(*ups), M_TEMP, M_WAITOK);
5434 
5435 	count = 0;
5436 	rw_rlock(&uma_rwlock);
5437 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
5438 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
5439 			count++;
5440 	}
5441 
5442 	LIST_FOREACH(z, &uma_cachezones, uz_link)
5443 		count++;
5444 
5445 	/*
5446 	 * Insert stream header.
5447 	 */
5448 	bzero(&ush, sizeof(ush));
5449 	ush.ush_version = UMA_STREAM_VERSION;
5450 	ush.ush_maxcpus = (mp_maxid + 1);
5451 	ush.ush_count = count;
5452 	(void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
5453 
5454 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
5455 		kfree = pages = 0;
5456 		for (i = 0; i < vm_ndomains; i++) {
5457 			kfree += kz->uk_domain[i].ud_free_items;
5458 			pages += kz->uk_domain[i].ud_pages;
5459 		}
5460 		LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5461 			bzero(&uth, sizeof(uth));
5462 			strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5463 			uth.uth_align = kz->uk_align;
5464 			uth.uth_size = kz->uk_size;
5465 			uth.uth_rsize = kz->uk_rsize;
5466 			if (z->uz_max_items > 0) {
5467 				items = UZ_ITEMS_COUNT(z->uz_items);
5468 				uth.uth_pages = (items / kz->uk_ipers) *
5469 					kz->uk_ppera;
5470 			} else
5471 				uth.uth_pages = pages;
5472 			uth.uth_maxpages = (z->uz_max_items / kz->uk_ipers) *
5473 			    kz->uk_ppera;
5474 			uth.uth_limit = z->uz_max_items;
5475 			uth.uth_keg_free = kfree;
5476 
5477 			/*
5478 			 * A zone is secondary is it is not the first entry
5479 			 * on the keg's zone list.
5480 			 */
5481 			if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
5482 			    (LIST_FIRST(&kz->uk_zones) != z))
5483 				uth.uth_zone_flags = UTH_ZONE_SECONDARY;
5484 			uma_vm_zone_stats(&uth, z, &sbuf, ups,
5485 			    kz->uk_flags & UMA_ZFLAG_INTERNAL);
5486 			(void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5487 			for (i = 0; i < mp_maxid + 1; i++)
5488 				(void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5489 		}
5490 	}
5491 	LIST_FOREACH(z, &uma_cachezones, uz_link) {
5492 		bzero(&uth, sizeof(uth));
5493 		strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5494 		uth.uth_size = z->uz_size;
5495 		uma_vm_zone_stats(&uth, z, &sbuf, ups, false);
5496 		(void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5497 		for (i = 0; i < mp_maxid + 1; i++)
5498 			(void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5499 	}
5500 
5501 	rw_runlock(&uma_rwlock);
5502 	error = sbuf_finish(&sbuf);
5503 	sbuf_delete(&sbuf);
5504 	free(ups, M_TEMP);
5505 	return (error);
5506 }
5507 
5508 int
5509 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS)
5510 {
5511 	uma_zone_t zone = *(uma_zone_t *)arg1;
5512 	int error, max;
5513 
5514 	max = uma_zone_get_max(zone);
5515 	error = sysctl_handle_int(oidp, &max, 0, req);
5516 	if (error || !req->newptr)
5517 		return (error);
5518 
5519 	uma_zone_set_max(zone, max);
5520 
5521 	return (0);
5522 }
5523 
5524 int
5525 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS)
5526 {
5527 	uma_zone_t zone;
5528 	int cur;
5529 
5530 	/*
5531 	 * Some callers want to add sysctls for global zones that
5532 	 * may not yet exist so they pass a pointer to a pointer.
5533 	 */
5534 	if (arg2 == 0)
5535 		zone = *(uma_zone_t *)arg1;
5536 	else
5537 		zone = arg1;
5538 	cur = uma_zone_get_cur(zone);
5539 	return (sysctl_handle_int(oidp, &cur, 0, req));
5540 }
5541 
5542 static int
5543 sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS)
5544 {
5545 	uma_zone_t zone = arg1;
5546 	uint64_t cur;
5547 
5548 	cur = uma_zone_get_allocs(zone);
5549 	return (sysctl_handle_64(oidp, &cur, 0, req));
5550 }
5551 
5552 static int
5553 sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS)
5554 {
5555 	uma_zone_t zone = arg1;
5556 	uint64_t cur;
5557 
5558 	cur = uma_zone_get_frees(zone);
5559 	return (sysctl_handle_64(oidp, &cur, 0, req));
5560 }
5561 
5562 static int
5563 sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS)
5564 {
5565 	struct sbuf sbuf;
5566 	uma_zone_t zone = arg1;
5567 	int error;
5568 
5569 	sbuf_new_for_sysctl(&sbuf, NULL, 0, req);
5570 	if (zone->uz_flags != 0)
5571 		sbuf_printf(&sbuf, "0x%b", zone->uz_flags, PRINT_UMA_ZFLAGS);
5572 	else
5573 		sbuf_printf(&sbuf, "0");
5574 	error = sbuf_finish(&sbuf);
5575 	sbuf_delete(&sbuf);
5576 
5577 	return (error);
5578 }
5579 
5580 static int
5581 sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS)
5582 {
5583 	uma_keg_t keg = arg1;
5584 	int avail, effpct, total;
5585 
5586 	total = keg->uk_ppera * PAGE_SIZE;
5587 	if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0)
5588 		total += slabzone(keg->uk_ipers)->uz_keg->uk_rsize;
5589 	/*
5590 	 * We consider the client's requested size and alignment here, not the
5591 	 * real size determination uk_rsize, because we also adjust the real
5592 	 * size for internal implementation reasons (max bitset size).
5593 	 */
5594 	avail = keg->uk_ipers * roundup2(keg->uk_size, keg->uk_align + 1);
5595 	if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
5596 		avail *= mp_maxid + 1;
5597 	effpct = 100 * avail / total;
5598 	return (sysctl_handle_int(oidp, &effpct, 0, req));
5599 }
5600 
5601 static int
5602 sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS)
5603 {
5604 	uma_zone_t zone = arg1;
5605 	uint64_t cur;
5606 
5607 	cur = UZ_ITEMS_COUNT(atomic_load_64(&zone->uz_items));
5608 	return (sysctl_handle_64(oidp, &cur, 0, req));
5609 }
5610 
5611 #ifdef INVARIANTS
5612 static uma_slab_t
5613 uma_dbg_getslab(uma_zone_t zone, void *item)
5614 {
5615 	uma_slab_t slab;
5616 	uma_keg_t keg;
5617 	uint8_t *mem;
5618 
5619 	/*
5620 	 * It is safe to return the slab here even though the
5621 	 * zone is unlocked because the item's allocation state
5622 	 * essentially holds a reference.
5623 	 */
5624 	mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
5625 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5626 		return (NULL);
5627 	if (zone->uz_flags & UMA_ZFLAG_VTOSLAB)
5628 		return (vtoslab((vm_offset_t)mem));
5629 	keg = zone->uz_keg;
5630 	if ((keg->uk_flags & UMA_ZFLAG_HASH) == 0)
5631 		return ((uma_slab_t)(mem + keg->uk_pgoff));
5632 	KEG_LOCK(keg, 0);
5633 	slab = hash_sfind(&keg->uk_hash, mem);
5634 	KEG_UNLOCK(keg, 0);
5635 
5636 	return (slab);
5637 }
5638 
5639 static bool
5640 uma_dbg_zskip(uma_zone_t zone, void *mem)
5641 {
5642 
5643 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5644 		return (true);
5645 
5646 	return (uma_dbg_kskip(zone->uz_keg, mem));
5647 }
5648 
5649 static bool
5650 uma_dbg_kskip(uma_keg_t keg, void *mem)
5651 {
5652 	uintptr_t idx;
5653 
5654 	if (dbg_divisor == 0)
5655 		return (true);
5656 
5657 	if (dbg_divisor == 1)
5658 		return (false);
5659 
5660 	idx = (uintptr_t)mem >> PAGE_SHIFT;
5661 	if (keg->uk_ipers > 1) {
5662 		idx *= keg->uk_ipers;
5663 		idx += ((uintptr_t)mem & PAGE_MASK) / keg->uk_rsize;
5664 	}
5665 
5666 	if ((idx / dbg_divisor) * dbg_divisor != idx) {
5667 		counter_u64_add(uma_skip_cnt, 1);
5668 		return (true);
5669 	}
5670 	counter_u64_add(uma_dbg_cnt, 1);
5671 
5672 	return (false);
5673 }
5674 
5675 /*
5676  * Set up the slab's freei data such that uma_dbg_free can function.
5677  *
5678  */
5679 static void
5680 uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item)
5681 {
5682 	uma_keg_t keg;
5683 	int freei;
5684 
5685 	if (slab == NULL) {
5686 		slab = uma_dbg_getslab(zone, item);
5687 		if (slab == NULL)
5688 			panic("uma: item %p did not belong to zone %s",
5689 			    item, zone->uz_name);
5690 	}
5691 	keg = zone->uz_keg;
5692 	freei = slab_item_index(slab, keg, item);
5693 
5694 	if (BIT_TEST_SET_ATOMIC(keg->uk_ipers, freei,
5695 	    slab_dbg_bits(slab, keg)))
5696 		panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)",
5697 		    item, zone, zone->uz_name, slab, freei);
5698 }
5699 
5700 /*
5701  * Verifies freed addresses.  Checks for alignment, valid slab membership
5702  * and duplicate frees.
5703  *
5704  */
5705 static void
5706 uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item)
5707 {
5708 	uma_keg_t keg;
5709 	int freei;
5710 
5711 	if (slab == NULL) {
5712 		slab = uma_dbg_getslab(zone, item);
5713 		if (slab == NULL)
5714 			panic("uma: Freed item %p did not belong to zone %s",
5715 			    item, zone->uz_name);
5716 	}
5717 	keg = zone->uz_keg;
5718 	freei = slab_item_index(slab, keg, item);
5719 
5720 	if (freei >= keg->uk_ipers)
5721 		panic("Invalid free of %p from zone %p(%s) slab %p(%d)",
5722 		    item, zone, zone->uz_name, slab, freei);
5723 
5724 	if (slab_item(slab, keg, freei) != item)
5725 		panic("Unaligned free of %p from zone %p(%s) slab %p(%d)",
5726 		    item, zone, zone->uz_name, slab, freei);
5727 
5728 	if (!BIT_TEST_CLR_ATOMIC(keg->uk_ipers, freei,
5729 	    slab_dbg_bits(slab, keg)))
5730 		panic("Duplicate free of %p from zone %p(%s) slab %p(%d)",
5731 		    item, zone, zone->uz_name, slab, freei);
5732 }
5733 #endif /* INVARIANTS */
5734 
5735 #ifdef DDB
5736 static int64_t
5737 get_uma_stats(uma_keg_t kz, uma_zone_t z, uint64_t *allocs, uint64_t *used,
5738     uint64_t *sleeps, long *cachefree, uint64_t *xdomain)
5739 {
5740 	uint64_t frees;
5741 	int i;
5742 
5743 	if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
5744 		*allocs = counter_u64_fetch(z->uz_allocs);
5745 		frees = counter_u64_fetch(z->uz_frees);
5746 		*sleeps = z->uz_sleeps;
5747 		*cachefree = 0;
5748 		*xdomain = 0;
5749 	} else
5750 		uma_zone_sumstat(z, cachefree, allocs, &frees, sleeps,
5751 		    xdomain);
5752 	for (i = 0; i < vm_ndomains; i++) {
5753 		*cachefree += ZDOM_GET(z, i)->uzd_nitems;
5754 		if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
5755 		    (LIST_FIRST(&kz->uk_zones) != z)))
5756 			*cachefree += kz->uk_domain[i].ud_free_items;
5757 	}
5758 	*used = *allocs - frees;
5759 	return (((int64_t)*used + *cachefree) * kz->uk_size);
5760 }
5761 
5762 DB_SHOW_COMMAND(uma, db_show_uma)
5763 {
5764 	const char *fmt_hdr, *fmt_entry;
5765 	uma_keg_t kz;
5766 	uma_zone_t z;
5767 	uint64_t allocs, used, sleeps, xdomain;
5768 	long cachefree;
5769 	/* variables for sorting */
5770 	uma_keg_t cur_keg;
5771 	uma_zone_t cur_zone, last_zone;
5772 	int64_t cur_size, last_size, size;
5773 	int ties;
5774 
5775 	/* /i option produces machine-parseable CSV output */
5776 	if (modif[0] == 'i') {
5777 		fmt_hdr = "%s,%s,%s,%s,%s,%s,%s,%s,%s\n";
5778 		fmt_entry = "\"%s\",%ju,%jd,%ld,%ju,%ju,%u,%jd,%ju\n";
5779 	} else {
5780 		fmt_hdr = "%18s %6s %7s %7s %11s %7s %7s %10s %8s\n";
5781 		fmt_entry = "%18s %6ju %7jd %7ld %11ju %7ju %7u %10jd %8ju\n";
5782 	}
5783 
5784 	db_printf(fmt_hdr, "Zone", "Size", "Used", "Free", "Requests",
5785 	    "Sleeps", "Bucket", "Total Mem", "XFree");
5786 
5787 	/* Sort the zones with largest size first. */
5788 	last_zone = NULL;
5789 	last_size = INT64_MAX;
5790 	for (;;) {
5791 		cur_zone = NULL;
5792 		cur_size = -1;
5793 		ties = 0;
5794 		LIST_FOREACH(kz, &uma_kegs, uk_link) {
5795 			LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5796 				/*
5797 				 * In the case of size ties, print out zones
5798 				 * in the order they are encountered.  That is,
5799 				 * when we encounter the most recently output
5800 				 * zone, we have already printed all preceding
5801 				 * ties, and we must print all following ties.
5802 				 */
5803 				if (z == last_zone) {
5804 					ties = 1;
5805 					continue;
5806 				}
5807 				size = get_uma_stats(kz, z, &allocs, &used,
5808 				    &sleeps, &cachefree, &xdomain);
5809 				if (size > cur_size && size < last_size + ties)
5810 				{
5811 					cur_size = size;
5812 					cur_zone = z;
5813 					cur_keg = kz;
5814 				}
5815 			}
5816 		}
5817 		if (cur_zone == NULL)
5818 			break;
5819 
5820 		size = get_uma_stats(cur_keg, cur_zone, &allocs, &used,
5821 		    &sleeps, &cachefree, &xdomain);
5822 		db_printf(fmt_entry, cur_zone->uz_name,
5823 		    (uintmax_t)cur_keg->uk_size, (intmax_t)used, cachefree,
5824 		    (uintmax_t)allocs, (uintmax_t)sleeps,
5825 		    (unsigned)cur_zone->uz_bucket_size, (intmax_t)size,
5826 		    xdomain);
5827 
5828 		if (db_pager_quit)
5829 			return;
5830 		last_zone = cur_zone;
5831 		last_size = cur_size;
5832 	}
5833 }
5834 
5835 DB_SHOW_COMMAND(umacache, db_show_umacache)
5836 {
5837 	uma_zone_t z;
5838 	uint64_t allocs, frees;
5839 	long cachefree;
5840 	int i;
5841 
5842 	db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
5843 	    "Requests", "Bucket");
5844 	LIST_FOREACH(z, &uma_cachezones, uz_link) {
5845 		uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL, NULL);
5846 		for (i = 0; i < vm_ndomains; i++)
5847 			cachefree += ZDOM_GET(z, i)->uzd_nitems;
5848 		db_printf("%18s %8ju %8jd %8ld %12ju %8u\n",
5849 		    z->uz_name, (uintmax_t)z->uz_size,
5850 		    (intmax_t)(allocs - frees), cachefree,
5851 		    (uintmax_t)allocs, z->uz_bucket_size);
5852 		if (db_pager_quit)
5853 			return;
5854 	}
5855 }
5856 #endif	/* DDB */
5857