xref: /freebsd/sys/vm/uma_core.c (revision d8ffc21c5ca6f7d4f2d9a65dc6308699af0b6a01)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2002-2019 Jeffrey Roberson <jeff@FreeBSD.org>
5  * Copyright (c) 2004, 2005 Bosko Milekic <bmilekic@FreeBSD.org>
6  * Copyright (c) 2004-2006 Robert N. M. Watson
7  * All rights reserved.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice unmodified, this list of conditions, and the following
14  *    disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
20  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
21  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
22  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
23  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
24  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
25  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
26  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
27  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
28  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
29  */
30 
31 /*
32  * uma_core.c  Implementation of the Universal Memory allocator
33  *
34  * This allocator is intended to replace the multitude of similar object caches
35  * in the standard FreeBSD kernel.  The intent is to be flexible as well as
36  * efficient.  A primary design goal is to return unused memory to the rest of
37  * the system.  This will make the system as a whole more flexible due to the
38  * ability to move memory to subsystems which most need it instead of leaving
39  * pools of reserved memory unused.
40  *
41  * The basic ideas stem from similar slab/zone based allocators whose algorithms
42  * are well known.
43  *
44  */
45 
46 /*
47  * TODO:
48  *	- Improve memory usage for large allocations
49  *	- Investigate cache size adjustments
50  */
51 
52 #include <sys/cdefs.h>
53 __FBSDID("$FreeBSD$");
54 
55 #include "opt_ddb.h"
56 #include "opt_param.h"
57 #include "opt_vm.h"
58 
59 #include <sys/param.h>
60 #include <sys/systm.h>
61 #include <sys/bitset.h>
62 #include <sys/domainset.h>
63 #include <sys/eventhandler.h>
64 #include <sys/kernel.h>
65 #include <sys/types.h>
66 #include <sys/limits.h>
67 #include <sys/queue.h>
68 #include <sys/malloc.h>
69 #include <sys/ktr.h>
70 #include <sys/lock.h>
71 #include <sys/sysctl.h>
72 #include <sys/mutex.h>
73 #include <sys/proc.h>
74 #include <sys/random.h>
75 #include <sys/rwlock.h>
76 #include <sys/sbuf.h>
77 #include <sys/sched.h>
78 #include <sys/sleepqueue.h>
79 #include <sys/smp.h>
80 #include <sys/taskqueue.h>
81 #include <sys/vmmeter.h>
82 
83 #include <vm/vm.h>
84 #include <vm/vm_domainset.h>
85 #include <vm/vm_object.h>
86 #include <vm/vm_page.h>
87 #include <vm/vm_pageout.h>
88 #include <vm/vm_param.h>
89 #include <vm/vm_phys.h>
90 #include <vm/vm_pagequeue.h>
91 #include <vm/vm_map.h>
92 #include <vm/vm_kern.h>
93 #include <vm/vm_extern.h>
94 #include <vm/uma.h>
95 #include <vm/uma_int.h>
96 #include <vm/uma_dbg.h>
97 
98 #include <ddb/ddb.h>
99 
100 #ifdef DEBUG_MEMGUARD
101 #include <vm/memguard.h>
102 #endif
103 
104 #include <machine/md_var.h>
105 
106 /*
107  * This is the zone and keg from which all zones are spawned.
108  */
109 static uma_zone_t kegs;
110 static uma_zone_t zones;
111 
112 /*
113  * These are the two zones from which all offpage uma_slab_ts are allocated.
114  *
115  * One zone is for slab headers that can represent a larger number of items,
116  * making the slabs themselves more efficient, and the other zone is for
117  * headers that are smaller and represent fewer items, making the headers more
118  * efficient.
119  */
120 #define	SLABZONE_SIZE(setsize)					\
121     (sizeof(struct uma_hash_slab) + BITSET_SIZE(setsize) * SLAB_BITSETS)
122 #define	SLABZONE0_SETSIZE	(PAGE_SIZE / 16)
123 #define	SLABZONE1_SETSIZE	SLAB_MAX_SETSIZE
124 #define	SLABZONE0_SIZE	SLABZONE_SIZE(SLABZONE0_SETSIZE)
125 #define	SLABZONE1_SIZE	SLABZONE_SIZE(SLABZONE1_SETSIZE)
126 static uma_zone_t slabzones[2];
127 
128 /*
129  * The initial hash tables come out of this zone so they can be allocated
130  * prior to malloc coming up.
131  */
132 static uma_zone_t hashzone;
133 
134 /* The boot-time adjusted value for cache line alignment. */
135 int uma_align_cache = 64 - 1;
136 
137 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets");
138 static MALLOC_DEFINE(M_UMA, "UMA", "UMA Misc");
139 
140 /*
141  * Are we allowed to allocate buckets?
142  */
143 static int bucketdisable = 1;
144 
145 /* Linked list of all kegs in the system */
146 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs);
147 
148 /* Linked list of all cache-only zones in the system */
149 static LIST_HEAD(,uma_zone) uma_cachezones =
150     LIST_HEAD_INITIALIZER(uma_cachezones);
151 
152 /* This RW lock protects the keg list */
153 static struct rwlock_padalign __exclusive_cache_line uma_rwlock;
154 
155 /*
156  * First available virual address for boot time allocations.
157  */
158 static vm_offset_t bootstart;
159 static vm_offset_t bootmem;
160 
161 static struct sx uma_reclaim_lock;
162 
163 /*
164  * kmem soft limit, initialized by uma_set_limit().  Ensure that early
165  * allocations don't trigger a wakeup of the reclaim thread.
166  */
167 unsigned long uma_kmem_limit = LONG_MAX;
168 SYSCTL_ULONG(_vm, OID_AUTO, uma_kmem_limit, CTLFLAG_RD, &uma_kmem_limit, 0,
169     "UMA kernel memory soft limit");
170 unsigned long uma_kmem_total;
171 SYSCTL_ULONG(_vm, OID_AUTO, uma_kmem_total, CTLFLAG_RD, &uma_kmem_total, 0,
172     "UMA kernel memory usage");
173 
174 /* Is the VM done starting up? */
175 static enum {
176 	BOOT_COLD,
177 	BOOT_KVA,
178 	BOOT_RUNNING,
179 	BOOT_SHUTDOWN,
180 } booted = BOOT_COLD;
181 
182 /*
183  * This is the handle used to schedule events that need to happen
184  * outside of the allocation fast path.
185  */
186 static struct callout uma_callout;
187 #define	UMA_TIMEOUT	20		/* Seconds for callout interval. */
188 
189 /*
190  * This structure is passed as the zone ctor arg so that I don't have to create
191  * a special allocation function just for zones.
192  */
193 struct uma_zctor_args {
194 	const char *name;
195 	size_t size;
196 	uma_ctor ctor;
197 	uma_dtor dtor;
198 	uma_init uminit;
199 	uma_fini fini;
200 	uma_import import;
201 	uma_release release;
202 	void *arg;
203 	uma_keg_t keg;
204 	int align;
205 	uint32_t flags;
206 };
207 
208 struct uma_kctor_args {
209 	uma_zone_t zone;
210 	size_t size;
211 	uma_init uminit;
212 	uma_fini fini;
213 	int align;
214 	uint32_t flags;
215 };
216 
217 struct uma_bucket_zone {
218 	uma_zone_t	ubz_zone;
219 	char		*ubz_name;
220 	int		ubz_entries;	/* Number of items it can hold. */
221 	int		ubz_maxsize;	/* Maximum allocation size per-item. */
222 };
223 
224 /*
225  * Compute the actual number of bucket entries to pack them in power
226  * of two sizes for more efficient space utilization.
227  */
228 #define	BUCKET_SIZE(n)						\
229     (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *))
230 
231 #define	BUCKET_MAX	BUCKET_SIZE(256)
232 #define	BUCKET_MIN	BUCKET_SIZE(4)
233 
234 struct uma_bucket_zone bucket_zones[] = {
235 	{ NULL, "4 Bucket", BUCKET_SIZE(4), 4096 },
236 	{ NULL, "6 Bucket", BUCKET_SIZE(6), 3072 },
237 	{ NULL, "8 Bucket", BUCKET_SIZE(8), 2048 },
238 	{ NULL, "12 Bucket", BUCKET_SIZE(12), 1536 },
239 	{ NULL, "16 Bucket", BUCKET_SIZE(16), 1024 },
240 	{ NULL, "32 Bucket", BUCKET_SIZE(32), 512 },
241 	{ NULL, "64 Bucket", BUCKET_SIZE(64), 256 },
242 	{ NULL, "128 Bucket", BUCKET_SIZE(128), 128 },
243 	{ NULL, "256 Bucket", BUCKET_SIZE(256), 64 },
244 	{ NULL, NULL, 0}
245 };
246 
247 /*
248  * Flags and enumerations to be passed to internal functions.
249  */
250 enum zfreeskip {
251 	SKIP_NONE =	0,
252 	SKIP_CNT =	0x00000001,
253 	SKIP_DTOR =	0x00010000,
254 	SKIP_FINI =	0x00020000,
255 };
256 
257 /* Prototypes.. */
258 
259 void	uma_startup1(vm_offset_t);
260 void	uma_startup2(void);
261 
262 static void *noobj_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
263 static void *page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
264 static void *pcpu_page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
265 static void *startup_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
266 static void page_free(void *, vm_size_t, uint8_t);
267 static void pcpu_page_free(void *, vm_size_t, uint8_t);
268 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int, int, int);
269 static void cache_drain(uma_zone_t);
270 static void bucket_drain(uma_zone_t, uma_bucket_t);
271 static void bucket_cache_reclaim(uma_zone_t zone, bool);
272 static int keg_ctor(void *, int, void *, int);
273 static void keg_dtor(void *, int, void *);
274 static int zone_ctor(void *, int, void *, int);
275 static void zone_dtor(void *, int, void *);
276 static int zero_init(void *, int, int);
277 static void zone_foreach(void (*zfunc)(uma_zone_t, void *), void *);
278 static void zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *), void *);
279 static void zone_timeout(uma_zone_t zone, void *);
280 static int hash_alloc(struct uma_hash *, u_int);
281 static int hash_expand(struct uma_hash *, struct uma_hash *);
282 static void hash_free(struct uma_hash *hash);
283 static void uma_timeout(void *);
284 static void uma_startup3(void);
285 static void uma_shutdown(void);
286 static void *zone_alloc_item(uma_zone_t, void *, int, int);
287 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip);
288 static int zone_alloc_limit(uma_zone_t zone, int count, int flags);
289 static void zone_free_limit(uma_zone_t zone, int count);
290 static void bucket_enable(void);
291 static void bucket_init(void);
292 static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int);
293 static void bucket_free(uma_zone_t zone, uma_bucket_t, void *);
294 static void bucket_zone_drain(void);
295 static uma_bucket_t zone_alloc_bucket(uma_zone_t, void *, int, int);
296 static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab);
297 static void slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item);
298 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit,
299     uma_fini fini, int align, uint32_t flags);
300 static int zone_import(void *, void **, int, int, int);
301 static void zone_release(void *, void **, int);
302 static bool cache_alloc(uma_zone_t, uma_cache_t, void *, int);
303 static bool cache_free(uma_zone_t, uma_cache_t, void *, void *, int);
304 
305 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS);
306 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS);
307 static int sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS);
308 static int sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS);
309 static int sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS);
310 static int sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS);
311 static int sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS);
312 
313 static uint64_t uma_zone_get_allocs(uma_zone_t zone);
314 
315 #ifdef INVARIANTS
316 static uint64_t uma_keg_get_allocs(uma_keg_t zone);
317 static inline struct noslabbits *slab_dbg_bits(uma_slab_t slab, uma_keg_t keg);
318 
319 static bool uma_dbg_kskip(uma_keg_t keg, void *mem);
320 static bool uma_dbg_zskip(uma_zone_t zone, void *mem);
321 static void uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item);
322 static void uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item);
323 
324 static SYSCTL_NODE(_vm, OID_AUTO, debug, CTLFLAG_RD, 0,
325     "Memory allocation debugging");
326 
327 static u_int dbg_divisor = 1;
328 SYSCTL_UINT(_vm_debug, OID_AUTO, divisor,
329     CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &dbg_divisor, 0,
330     "Debug & thrash every this item in memory allocator");
331 
332 static counter_u64_t uma_dbg_cnt = EARLY_COUNTER;
333 static counter_u64_t uma_skip_cnt = EARLY_COUNTER;
334 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, trashed, CTLFLAG_RD,
335     &uma_dbg_cnt, "memory items debugged");
336 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, skipped, CTLFLAG_RD,
337     &uma_skip_cnt, "memory items skipped, not debugged");
338 #endif
339 
340 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
341 
342 SYSCTL_NODE(_vm, OID_AUTO, uma, CTLFLAG_RW, 0, "Universal Memory Allocator");
343 
344 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLFLAG_MPSAFE|CTLTYPE_INT,
345     0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones");
346 
347 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLFLAG_MPSAFE|CTLTYPE_STRUCT,
348     0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats");
349 
350 static int zone_warnings = 1;
351 SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RWTUN, &zone_warnings, 0,
352     "Warn when UMA zones becomes full");
353 
354 /*
355  * Select the slab zone for an offpage slab with the given maximum item count.
356  */
357 static inline uma_zone_t
358 slabzone(int ipers)
359 {
360 
361 	return (slabzones[ipers > SLABZONE0_SETSIZE]);
362 }
363 
364 /*
365  * This routine checks to see whether or not it's safe to enable buckets.
366  */
367 static void
368 bucket_enable(void)
369 {
370 
371 	KASSERT(booted >= BOOT_KVA, ("Bucket enable before init"));
372 	bucketdisable = vm_page_count_min();
373 }
374 
375 /*
376  * Initialize bucket_zones, the array of zones of buckets of various sizes.
377  *
378  * For each zone, calculate the memory required for each bucket, consisting
379  * of the header and an array of pointers.
380  */
381 static void
382 bucket_init(void)
383 {
384 	struct uma_bucket_zone *ubz;
385 	int size;
386 
387 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) {
388 		size = roundup(sizeof(struct uma_bucket), sizeof(void *));
389 		size += sizeof(void *) * ubz->ubz_entries;
390 		ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
391 		    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
392 		    UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET |
393 		    UMA_ZONE_FIRSTTOUCH);
394 	}
395 }
396 
397 /*
398  * Given a desired number of entries for a bucket, return the zone from which
399  * to allocate the bucket.
400  */
401 static struct uma_bucket_zone *
402 bucket_zone_lookup(int entries)
403 {
404 	struct uma_bucket_zone *ubz;
405 
406 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
407 		if (ubz->ubz_entries >= entries)
408 			return (ubz);
409 	ubz--;
410 	return (ubz);
411 }
412 
413 static struct uma_bucket_zone *
414 bucket_zone_max(uma_zone_t zone, int nitems)
415 {
416 	struct uma_bucket_zone *ubz;
417 	int bpcpu;
418 
419 	bpcpu = 2;
420 	if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
421 		/* Count the cross-domain bucket. */
422 		bpcpu++;
423 
424 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
425 		if (ubz->ubz_entries * bpcpu * mp_ncpus > nitems)
426 			break;
427 	if (ubz == &bucket_zones[0])
428 		ubz = NULL;
429 	else
430 		ubz--;
431 	return (ubz);
432 }
433 
434 static int
435 bucket_select(int size)
436 {
437 	struct uma_bucket_zone *ubz;
438 
439 	ubz = &bucket_zones[0];
440 	if (size > ubz->ubz_maxsize)
441 		return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1);
442 
443 	for (; ubz->ubz_entries != 0; ubz++)
444 		if (ubz->ubz_maxsize < size)
445 			break;
446 	ubz--;
447 	return (ubz->ubz_entries);
448 }
449 
450 static uma_bucket_t
451 bucket_alloc(uma_zone_t zone, void *udata, int flags)
452 {
453 	struct uma_bucket_zone *ubz;
454 	uma_bucket_t bucket;
455 
456 	/*
457 	 * Don't allocate buckets in low memory situations.
458 	 */
459 	if (bucketdisable)
460 		return (NULL);
461 
462 	/*
463 	 * To limit bucket recursion we store the original zone flags
464 	 * in a cookie passed via zalloc_arg/zfree_arg.  This allows the
465 	 * NOVM flag to persist even through deep recursions.  We also
466 	 * store ZFLAG_BUCKET once we have recursed attempting to allocate
467 	 * a bucket for a bucket zone so we do not allow infinite bucket
468 	 * recursion.  This cookie will even persist to frees of unused
469 	 * buckets via the allocation path or bucket allocations in the
470 	 * free path.
471 	 */
472 	if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
473 		udata = (void *)(uintptr_t)zone->uz_flags;
474 	else {
475 		if ((uintptr_t)udata & UMA_ZFLAG_BUCKET)
476 			return (NULL);
477 		udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET);
478 	}
479 	if ((uintptr_t)udata & UMA_ZFLAG_CACHEONLY)
480 		flags |= M_NOVM;
481 	ubz = bucket_zone_lookup(zone->uz_bucket_size);
482 	if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0)
483 		ubz++;
484 	bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags);
485 	if (bucket) {
486 #ifdef INVARIANTS
487 		bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
488 #endif
489 		bucket->ub_cnt = 0;
490 		bucket->ub_entries = ubz->ubz_entries;
491 	}
492 
493 	return (bucket);
494 }
495 
496 static void
497 bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata)
498 {
499 	struct uma_bucket_zone *ubz;
500 
501 	KASSERT(bucket->ub_cnt == 0,
502 	    ("bucket_free: Freeing a non free bucket."));
503 	if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
504 		udata = (void *)(uintptr_t)zone->uz_flags;
505 	ubz = bucket_zone_lookup(bucket->ub_entries);
506 	uma_zfree_arg(ubz->ubz_zone, bucket, udata);
507 }
508 
509 static void
510 bucket_zone_drain(void)
511 {
512 	struct uma_bucket_zone *ubz;
513 
514 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
515 		uma_zone_reclaim(ubz->ubz_zone, UMA_RECLAIM_DRAIN);
516 }
517 
518 /*
519  * Attempt to satisfy an allocation by retrieving a full bucket from one of the
520  * zone's caches.
521  */
522 static uma_bucket_t
523 zone_fetch_bucket(uma_zone_t zone, uma_zone_domain_t zdom)
524 {
525 	uma_bucket_t bucket;
526 
527 	ZONE_LOCK_ASSERT(zone);
528 
529 	if ((bucket = TAILQ_FIRST(&zdom->uzd_buckets)) != NULL) {
530 		MPASS(zdom->uzd_nitems >= bucket->ub_cnt);
531 		TAILQ_REMOVE(&zdom->uzd_buckets, bucket, ub_link);
532 		zdom->uzd_nitems -= bucket->ub_cnt;
533 		if (zdom->uzd_imin > zdom->uzd_nitems)
534 			zdom->uzd_imin = zdom->uzd_nitems;
535 		zone->uz_bkt_count -= bucket->ub_cnt;
536 	}
537 	return (bucket);
538 }
539 
540 /*
541  * Insert a full bucket into the specified cache.  The "ws" parameter indicates
542  * whether the bucket's contents should be counted as part of the zone's working
543  * set.
544  */
545 static void
546 zone_put_bucket(uma_zone_t zone, uma_zone_domain_t zdom, uma_bucket_t bucket,
547     const bool ws)
548 {
549 
550 	ZONE_LOCK_ASSERT(zone);
551 	KASSERT(!ws || zone->uz_bkt_count < zone->uz_bkt_max,
552 	    ("%s: zone %p overflow", __func__, zone));
553 
554 	if (ws)
555 		TAILQ_INSERT_HEAD(&zdom->uzd_buckets, bucket, ub_link);
556 	else
557 		TAILQ_INSERT_TAIL(&zdom->uzd_buckets, bucket, ub_link);
558 	zdom->uzd_nitems += bucket->ub_cnt;
559 	if (ws && zdom->uzd_imax < zdom->uzd_nitems)
560 		zdom->uzd_imax = zdom->uzd_nitems;
561 	zone->uz_bkt_count += bucket->ub_cnt;
562 }
563 
564 /* Pops an item out of a per-cpu cache bucket. */
565 static inline void *
566 cache_bucket_pop(uma_cache_t cache, uma_cache_bucket_t bucket)
567 {
568 	void *item;
569 
570 	CRITICAL_ASSERT(curthread);
571 
572 	bucket->ucb_cnt--;
573 	item = bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt];
574 #ifdef INVARIANTS
575 	bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = NULL;
576 	KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled."));
577 #endif
578 	cache->uc_allocs++;
579 
580 	return (item);
581 }
582 
583 /* Pushes an item into a per-cpu cache bucket. */
584 static inline void
585 cache_bucket_push(uma_cache_t cache, uma_cache_bucket_t bucket, void *item)
586 {
587 
588 	CRITICAL_ASSERT(curthread);
589 	KASSERT(bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] == NULL,
590 	    ("uma_zfree: Freeing to non free bucket index."));
591 
592 	bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = item;
593 	bucket->ucb_cnt++;
594 	cache->uc_frees++;
595 }
596 
597 /*
598  * Unload a UMA bucket from a per-cpu cache.
599  */
600 static inline uma_bucket_t
601 cache_bucket_unload(uma_cache_bucket_t bucket)
602 {
603 	uma_bucket_t b;
604 
605 	b = bucket->ucb_bucket;
606 	if (b != NULL) {
607 		MPASS(b->ub_entries == bucket->ucb_entries);
608 		b->ub_cnt = bucket->ucb_cnt;
609 		bucket->ucb_bucket = NULL;
610 		bucket->ucb_entries = bucket->ucb_cnt = 0;
611 	}
612 
613 	return (b);
614 }
615 
616 static inline uma_bucket_t
617 cache_bucket_unload_alloc(uma_cache_t cache)
618 {
619 
620 	return (cache_bucket_unload(&cache->uc_allocbucket));
621 }
622 
623 static inline uma_bucket_t
624 cache_bucket_unload_free(uma_cache_t cache)
625 {
626 
627 	return (cache_bucket_unload(&cache->uc_freebucket));
628 }
629 
630 static inline uma_bucket_t
631 cache_bucket_unload_cross(uma_cache_t cache)
632 {
633 
634 	return (cache_bucket_unload(&cache->uc_crossbucket));
635 }
636 
637 /*
638  * Load a bucket into a per-cpu cache bucket.
639  */
640 static inline void
641 cache_bucket_load(uma_cache_bucket_t bucket, uma_bucket_t b)
642 {
643 
644 	CRITICAL_ASSERT(curthread);
645 	MPASS(bucket->ucb_bucket == NULL);
646 
647 	bucket->ucb_bucket = b;
648 	bucket->ucb_cnt = b->ub_cnt;
649 	bucket->ucb_entries = b->ub_entries;
650 }
651 
652 static inline void
653 cache_bucket_load_alloc(uma_cache_t cache, uma_bucket_t b)
654 {
655 
656 	cache_bucket_load(&cache->uc_allocbucket, b);
657 }
658 
659 static inline void
660 cache_bucket_load_free(uma_cache_t cache, uma_bucket_t b)
661 {
662 
663 	cache_bucket_load(&cache->uc_freebucket, b);
664 }
665 
666 #ifdef NUMA
667 static inline void
668 cache_bucket_load_cross(uma_cache_t cache, uma_bucket_t b)
669 {
670 
671 	cache_bucket_load(&cache->uc_crossbucket, b);
672 }
673 #endif
674 
675 /*
676  * Copy and preserve ucb_spare.
677  */
678 static inline void
679 cache_bucket_copy(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
680 {
681 
682 	b1->ucb_bucket = b2->ucb_bucket;
683 	b1->ucb_entries = b2->ucb_entries;
684 	b1->ucb_cnt = b2->ucb_cnt;
685 }
686 
687 /*
688  * Swap two cache buckets.
689  */
690 static inline void
691 cache_bucket_swap(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
692 {
693 	struct uma_cache_bucket b3;
694 
695 	CRITICAL_ASSERT(curthread);
696 
697 	cache_bucket_copy(&b3, b1);
698 	cache_bucket_copy(b1, b2);
699 	cache_bucket_copy(b2, &b3);
700 }
701 
702 static void
703 zone_log_warning(uma_zone_t zone)
704 {
705 	static const struct timeval warninterval = { 300, 0 };
706 
707 	if (!zone_warnings || zone->uz_warning == NULL)
708 		return;
709 
710 	if (ratecheck(&zone->uz_ratecheck, &warninterval))
711 		printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning);
712 }
713 
714 static inline void
715 zone_maxaction(uma_zone_t zone)
716 {
717 
718 	if (zone->uz_maxaction.ta_func != NULL)
719 		taskqueue_enqueue(taskqueue_thread, &zone->uz_maxaction);
720 }
721 
722 /*
723  * Routine called by timeout which is used to fire off some time interval
724  * based calculations.  (stats, hash size, etc.)
725  *
726  * Arguments:
727  *	arg   Unused
728  *
729  * Returns:
730  *	Nothing
731  */
732 static void
733 uma_timeout(void *unused)
734 {
735 	bucket_enable();
736 	zone_foreach(zone_timeout, NULL);
737 
738 	/* Reschedule this event */
739 	callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
740 }
741 
742 /*
743  * Update the working set size estimate for the zone's bucket cache.
744  * The constants chosen here are somewhat arbitrary.  With an update period of
745  * 20s (UMA_TIMEOUT), this estimate is dominated by zone activity over the
746  * last 100s.
747  */
748 static void
749 zone_domain_update_wss(uma_zone_domain_t zdom)
750 {
751 	long wss;
752 
753 	MPASS(zdom->uzd_imax >= zdom->uzd_imin);
754 	wss = zdom->uzd_imax - zdom->uzd_imin;
755 	zdom->uzd_imax = zdom->uzd_imin = zdom->uzd_nitems;
756 	zdom->uzd_wss = (4 * wss + zdom->uzd_wss) / 5;
757 }
758 
759 /*
760  * Routine to perform timeout driven calculations.  This expands the
761  * hashes and does per cpu statistics aggregation.
762  *
763  *  Returns nothing.
764  */
765 static void
766 zone_timeout(uma_zone_t zone, void *unused)
767 {
768 	uma_keg_t keg;
769 	u_int slabs, pages;
770 
771 	if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
772 		goto update_wss;
773 
774 	keg = zone->uz_keg;
775 
776 	/*
777 	 * Hash zones are non-numa by definition so the first domain
778 	 * is the only one present.
779 	 */
780 	KEG_LOCK(keg, 0);
781 	pages = keg->uk_domain[0].ud_pages;
782 
783 	/*
784 	 * Expand the keg hash table.
785 	 *
786 	 * This is done if the number of slabs is larger than the hash size.
787 	 * What I'm trying to do here is completely reduce collisions.  This
788 	 * may be a little aggressive.  Should I allow for two collisions max?
789 	 */
790 	if ((slabs = pages / keg->uk_ppera) > keg->uk_hash.uh_hashsize) {
791 		struct uma_hash newhash;
792 		struct uma_hash oldhash;
793 		int ret;
794 
795 		/*
796 		 * This is so involved because allocating and freeing
797 		 * while the keg lock is held will lead to deadlock.
798 		 * I have to do everything in stages and check for
799 		 * races.
800 		 */
801 		KEG_UNLOCK(keg, 0);
802 		ret = hash_alloc(&newhash, 1 << fls(slabs));
803 		KEG_LOCK(keg, 0);
804 		if (ret) {
805 			if (hash_expand(&keg->uk_hash, &newhash)) {
806 				oldhash = keg->uk_hash;
807 				keg->uk_hash = newhash;
808 			} else
809 				oldhash = newhash;
810 
811 			KEG_UNLOCK(keg, 0);
812 			hash_free(&oldhash);
813 			goto update_wss;
814 		}
815 	}
816 	KEG_UNLOCK(keg, 0);
817 
818 update_wss:
819 	ZONE_LOCK(zone);
820 	for (int i = 0; i < vm_ndomains; i++)
821 		zone_domain_update_wss(&zone->uz_domain[i]);
822 	ZONE_UNLOCK(zone);
823 }
824 
825 /*
826  * Allocate and zero fill the next sized hash table from the appropriate
827  * backing store.
828  *
829  * Arguments:
830  *	hash  A new hash structure with the old hash size in uh_hashsize
831  *
832  * Returns:
833  *	1 on success and 0 on failure.
834  */
835 static int
836 hash_alloc(struct uma_hash *hash, u_int size)
837 {
838 	size_t alloc;
839 
840 	KASSERT(powerof2(size), ("hash size must be power of 2"));
841 	if (size > UMA_HASH_SIZE_INIT)  {
842 		hash->uh_hashsize = size;
843 		alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
844 		hash->uh_slab_hash = malloc(alloc, M_UMAHASH, M_NOWAIT);
845 	} else {
846 		alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
847 		hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
848 		    UMA_ANYDOMAIN, M_WAITOK);
849 		hash->uh_hashsize = UMA_HASH_SIZE_INIT;
850 	}
851 	if (hash->uh_slab_hash) {
852 		bzero(hash->uh_slab_hash, alloc);
853 		hash->uh_hashmask = hash->uh_hashsize - 1;
854 		return (1);
855 	}
856 
857 	return (0);
858 }
859 
860 /*
861  * Expands the hash table for HASH zones.  This is done from zone_timeout
862  * to reduce collisions.  This must not be done in the regular allocation
863  * path, otherwise, we can recurse on the vm while allocating pages.
864  *
865  * Arguments:
866  *	oldhash  The hash you want to expand
867  *	newhash  The hash structure for the new table
868  *
869  * Returns:
870  *	Nothing
871  *
872  * Discussion:
873  */
874 static int
875 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
876 {
877 	uma_hash_slab_t slab;
878 	u_int hval;
879 	u_int idx;
880 
881 	if (!newhash->uh_slab_hash)
882 		return (0);
883 
884 	if (oldhash->uh_hashsize >= newhash->uh_hashsize)
885 		return (0);
886 
887 	/*
888 	 * I need to investigate hash algorithms for resizing without a
889 	 * full rehash.
890 	 */
891 
892 	for (idx = 0; idx < oldhash->uh_hashsize; idx++)
893 		while (!LIST_EMPTY(&oldhash->uh_slab_hash[idx])) {
894 			slab = LIST_FIRST(&oldhash->uh_slab_hash[idx]);
895 			LIST_REMOVE(slab, uhs_hlink);
896 			hval = UMA_HASH(newhash, slab->uhs_data);
897 			LIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
898 			    slab, uhs_hlink);
899 		}
900 
901 	return (1);
902 }
903 
904 /*
905  * Free the hash bucket to the appropriate backing store.
906  *
907  * Arguments:
908  *	slab_hash  The hash bucket we're freeing
909  *	hashsize   The number of entries in that hash bucket
910  *
911  * Returns:
912  *	Nothing
913  */
914 static void
915 hash_free(struct uma_hash *hash)
916 {
917 	if (hash->uh_slab_hash == NULL)
918 		return;
919 	if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
920 		zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE);
921 	else
922 		free(hash->uh_slab_hash, M_UMAHASH);
923 }
924 
925 /*
926  * Frees all outstanding items in a bucket
927  *
928  * Arguments:
929  *	zone   The zone to free to, must be unlocked.
930  *	bucket The free/alloc bucket with items.
931  *
932  * Returns:
933  *	Nothing
934  */
935 
936 static void
937 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
938 {
939 	int i;
940 
941 	if (bucket == NULL || bucket->ub_cnt == 0)
942 		return;
943 
944 	if (zone->uz_fini)
945 		for (i = 0; i < bucket->ub_cnt; i++)
946 			zone->uz_fini(bucket->ub_bucket[i], zone->uz_size);
947 	zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt);
948 	if (zone->uz_max_items > 0)
949 		zone_free_limit(zone, bucket->ub_cnt);
950 	bucket->ub_cnt = 0;
951 }
952 
953 /*
954  * Drains the per cpu caches for a zone.
955  *
956  * NOTE: This may only be called while the zone is being torn down, and not
957  * during normal operation.  This is necessary in order that we do not have
958  * to migrate CPUs to drain the per-CPU caches.
959  *
960  * Arguments:
961  *	zone     The zone to drain, must be unlocked.
962  *
963  * Returns:
964  *	Nothing
965  */
966 static void
967 cache_drain(uma_zone_t zone)
968 {
969 	uma_cache_t cache;
970 	uma_bucket_t bucket;
971 	int cpu;
972 
973 	/*
974 	 * XXX: It is safe to not lock the per-CPU caches, because we're
975 	 * tearing down the zone anyway.  I.e., there will be no further use
976 	 * of the caches at this point.
977 	 *
978 	 * XXX: It would good to be able to assert that the zone is being
979 	 * torn down to prevent improper use of cache_drain().
980 	 */
981 	CPU_FOREACH(cpu) {
982 		cache = &zone->uz_cpu[cpu];
983 		bucket = cache_bucket_unload_alloc(cache);
984 		if (bucket != NULL) {
985 			bucket_drain(zone, bucket);
986 			bucket_free(zone, bucket, NULL);
987 		}
988 		bucket = cache_bucket_unload_free(cache);
989 		if (bucket != NULL) {
990 			bucket_drain(zone, bucket);
991 			bucket_free(zone, bucket, NULL);
992 		}
993 		bucket = cache_bucket_unload_cross(cache);
994 		if (bucket != NULL) {
995 			bucket_drain(zone, bucket);
996 			bucket_free(zone, bucket, NULL);
997 		}
998 	}
999 	bucket_cache_reclaim(zone, true);
1000 }
1001 
1002 static void
1003 cache_shrink(uma_zone_t zone, void *unused)
1004 {
1005 
1006 	if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1007 		return;
1008 
1009 	ZONE_LOCK(zone);
1010 	zone->uz_bucket_size =
1011 	    (zone->uz_bucket_size_min + zone->uz_bucket_size) / 2;
1012 	ZONE_UNLOCK(zone);
1013 }
1014 
1015 static void
1016 cache_drain_safe_cpu(uma_zone_t zone, void *unused)
1017 {
1018 	uma_cache_t cache;
1019 	uma_bucket_t b1, b2, b3;
1020 	int domain;
1021 
1022 	if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1023 		return;
1024 
1025 	b1 = b2 = b3 = NULL;
1026 	ZONE_LOCK(zone);
1027 	critical_enter();
1028 	if (zone->uz_flags & UMA_ZONE_FIRSTTOUCH)
1029 		domain = PCPU_GET(domain);
1030 	else
1031 		domain = 0;
1032 	cache = &zone->uz_cpu[curcpu];
1033 	b1 = cache_bucket_unload_alloc(cache);
1034 	if (b1 != NULL && b1->ub_cnt != 0) {
1035 		zone_put_bucket(zone, &zone->uz_domain[domain], b1, false);
1036 		b1 = NULL;
1037 	}
1038 	b2 = cache_bucket_unload_free(cache);
1039 	if (b2 != NULL && b2->ub_cnt != 0) {
1040 		zone_put_bucket(zone, &zone->uz_domain[domain], b2, false);
1041 		b2 = NULL;
1042 	}
1043 	b3 = cache_bucket_unload_cross(cache);
1044 	critical_exit();
1045 	ZONE_UNLOCK(zone);
1046 	if (b1)
1047 		bucket_free(zone, b1, NULL);
1048 	if (b2)
1049 		bucket_free(zone, b2, NULL);
1050 	if (b3) {
1051 		bucket_drain(zone, b3);
1052 		bucket_free(zone, b3, NULL);
1053 	}
1054 }
1055 
1056 /*
1057  * Safely drain per-CPU caches of a zone(s) to alloc bucket.
1058  * This is an expensive call because it needs to bind to all CPUs
1059  * one by one and enter a critical section on each of them in order
1060  * to safely access their cache buckets.
1061  * Zone lock must not be held on call this function.
1062  */
1063 static void
1064 pcpu_cache_drain_safe(uma_zone_t zone)
1065 {
1066 	int cpu;
1067 
1068 	/*
1069 	 * Polite bucket sizes shrinking was not enough, shrink aggressively.
1070 	 */
1071 	if (zone)
1072 		cache_shrink(zone, NULL);
1073 	else
1074 		zone_foreach(cache_shrink, NULL);
1075 
1076 	CPU_FOREACH(cpu) {
1077 		thread_lock(curthread);
1078 		sched_bind(curthread, cpu);
1079 		thread_unlock(curthread);
1080 
1081 		if (zone)
1082 			cache_drain_safe_cpu(zone, NULL);
1083 		else
1084 			zone_foreach(cache_drain_safe_cpu, NULL);
1085 	}
1086 	thread_lock(curthread);
1087 	sched_unbind(curthread);
1088 	thread_unlock(curthread);
1089 }
1090 
1091 /*
1092  * Reclaim cached buckets from a zone.  All buckets are reclaimed if the caller
1093  * requested a drain, otherwise the per-domain caches are trimmed to either
1094  * estimated working set size.
1095  */
1096 static void
1097 bucket_cache_reclaim(uma_zone_t zone, bool drain)
1098 {
1099 	uma_zone_domain_t zdom;
1100 	uma_bucket_t bucket;
1101 	long target, tofree;
1102 	int i;
1103 
1104 	for (i = 0; i < vm_ndomains; i++) {
1105 		/*
1106 		 * The cross bucket is partially filled and not part of
1107 		 * the item count.  Reclaim it individually here.
1108 		 */
1109 		zdom = &zone->uz_domain[i];
1110 		ZONE_CROSS_LOCK(zone);
1111 		bucket = zdom->uzd_cross;
1112 		zdom->uzd_cross = NULL;
1113 		ZONE_CROSS_UNLOCK(zone);
1114 		if (bucket != NULL) {
1115 			bucket_drain(zone, bucket);
1116 			bucket_free(zone, bucket, NULL);
1117 		}
1118 
1119 		/*
1120 		 * Shrink the zone bucket size to ensure that the per-CPU caches
1121 		 * don't grow too large.
1122 		 */
1123 		ZONE_LOCK(zone);
1124 		if (i == 0 && zone->uz_bucket_size > zone->uz_bucket_size_min)
1125 			zone->uz_bucket_size--;
1126 
1127 		/*
1128 		 * If we were asked to drain the zone, we are done only once
1129 		 * this bucket cache is empty.  Otherwise, we reclaim items in
1130 		 * excess of the zone's estimated working set size.  If the
1131 		 * difference nitems - imin is larger than the WSS estimate,
1132 		 * then the estimate will grow at the end of this interval and
1133 		 * we ignore the historical average.
1134 		 */
1135 		target = drain ? 0 : lmax(zdom->uzd_wss, zdom->uzd_nitems -
1136 		    zdom->uzd_imin);
1137 		while (zdom->uzd_nitems > target) {
1138 			bucket = TAILQ_LAST(&zdom->uzd_buckets, uma_bucketlist);
1139 			if (bucket == NULL)
1140 				break;
1141 			tofree = bucket->ub_cnt;
1142 			TAILQ_REMOVE(&zdom->uzd_buckets, bucket, ub_link);
1143 			zdom->uzd_nitems -= tofree;
1144 
1145 			/*
1146 			 * Shift the bounds of the current WSS interval to avoid
1147 			 * perturbing the estimate.
1148 			 */
1149 			zdom->uzd_imax -= lmin(zdom->uzd_imax, tofree);
1150 			zdom->uzd_imin -= lmin(zdom->uzd_imin, tofree);
1151 
1152 			ZONE_UNLOCK(zone);
1153 			bucket_drain(zone, bucket);
1154 			bucket_free(zone, bucket, NULL);
1155 			ZONE_LOCK(zone);
1156 		}
1157 		ZONE_UNLOCK(zone);
1158 	}
1159 }
1160 
1161 static void
1162 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start)
1163 {
1164 	uint8_t *mem;
1165 	int i;
1166 	uint8_t flags;
1167 
1168 	CTR4(KTR_UMA, "keg_free_slab keg %s(%p) slab %p, returning %d bytes",
1169 	    keg->uk_name, keg, slab, PAGE_SIZE * keg->uk_ppera);
1170 
1171 	mem = slab_data(slab, keg);
1172 	flags = slab->us_flags;
1173 	i = start;
1174 	if (keg->uk_fini != NULL) {
1175 		for (i--; i > -1; i--)
1176 #ifdef INVARIANTS
1177 		/*
1178 		 * trash_fini implies that dtor was trash_dtor. trash_fini
1179 		 * would check that memory hasn't been modified since free,
1180 		 * which executed trash_dtor.
1181 		 * That's why we need to run uma_dbg_kskip() check here,
1182 		 * albeit we don't make skip check for other init/fini
1183 		 * invocations.
1184 		 */
1185 		if (!uma_dbg_kskip(keg, slab_item(slab, keg, i)) ||
1186 		    keg->uk_fini != trash_fini)
1187 #endif
1188 			keg->uk_fini(slab_item(slab, keg, i), keg->uk_size);
1189 	}
1190 	if (keg->uk_flags & UMA_ZFLAG_OFFPAGE)
1191 		zone_free_item(slabzone(keg->uk_ipers), slab_tohashslab(slab),
1192 		    NULL, SKIP_NONE);
1193 	keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags);
1194 	uma_total_dec(PAGE_SIZE * keg->uk_ppera);
1195 }
1196 
1197 /*
1198  * Frees pages from a keg back to the system.  This is done on demand from
1199  * the pageout daemon.
1200  *
1201  * Returns nothing.
1202  */
1203 static void
1204 keg_drain(uma_keg_t keg)
1205 {
1206 	struct slabhead freeslabs = { 0 };
1207 	uma_domain_t dom;
1208 	uma_slab_t slab, tmp;
1209 	int i, n;
1210 
1211 	/*
1212 	 * We don't want to take pages from statically allocated kegs at this
1213 	 * time
1214 	 */
1215 	if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL)
1216 		return;
1217 
1218 	for (i = 0; i < vm_ndomains; i++) {
1219 		CTR4(KTR_UMA, "keg_drain %s(%p) domain %d free items: %u",
1220 		    keg->uk_name, keg, i, dom->ud_free);
1221 		n = 0;
1222 		dom = &keg->uk_domain[i];
1223 		KEG_LOCK(keg, i);
1224 		LIST_FOREACH_SAFE(slab, &dom->ud_free_slab, us_link, tmp) {
1225 			if (keg->uk_flags & UMA_ZFLAG_HASH)
1226 				UMA_HASH_REMOVE(&keg->uk_hash, slab);
1227 			n++;
1228 			LIST_REMOVE(slab, us_link);
1229 			LIST_INSERT_HEAD(&freeslabs, slab, us_link);
1230 		}
1231 		dom->ud_pages -= n * keg->uk_ppera;
1232 		dom->ud_free -= n * keg->uk_ipers;
1233 		KEG_UNLOCK(keg, i);
1234 	}
1235 
1236 	while ((slab = LIST_FIRST(&freeslabs)) != NULL) {
1237 		LIST_REMOVE(slab, us_link);
1238 		keg_free_slab(keg, slab, keg->uk_ipers);
1239 	}
1240 }
1241 
1242 static void
1243 zone_reclaim(uma_zone_t zone, int waitok, bool drain)
1244 {
1245 
1246 	/*
1247 	 * Set draining to interlock with zone_dtor() so we can release our
1248 	 * locks as we go.  Only dtor() should do a WAITOK call since it
1249 	 * is the only call that knows the structure will still be available
1250 	 * when it wakes up.
1251 	 */
1252 	ZONE_LOCK(zone);
1253 	while (zone->uz_flags & UMA_ZFLAG_RECLAIMING) {
1254 		if (waitok == M_NOWAIT)
1255 			goto out;
1256 		msleep(zone, &zone->uz_lock, PVM, "zonedrain", 1);
1257 	}
1258 	zone->uz_flags |= UMA_ZFLAG_RECLAIMING;
1259 	ZONE_UNLOCK(zone);
1260 	bucket_cache_reclaim(zone, drain);
1261 
1262 	/*
1263 	 * The DRAINING flag protects us from being freed while
1264 	 * we're running.  Normally the uma_rwlock would protect us but we
1265 	 * must be able to release and acquire the right lock for each keg.
1266 	 */
1267 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0)
1268 		keg_drain(zone->uz_keg);
1269 	ZONE_LOCK(zone);
1270 	zone->uz_flags &= ~UMA_ZFLAG_RECLAIMING;
1271 	wakeup(zone);
1272 out:
1273 	ZONE_UNLOCK(zone);
1274 }
1275 
1276 static void
1277 zone_drain(uma_zone_t zone, void *unused)
1278 {
1279 
1280 	zone_reclaim(zone, M_NOWAIT, true);
1281 }
1282 
1283 static void
1284 zone_trim(uma_zone_t zone, void *unused)
1285 {
1286 
1287 	zone_reclaim(zone, M_NOWAIT, false);
1288 }
1289 
1290 /*
1291  * Allocate a new slab for a keg and inserts it into the partial slab list.
1292  * The keg should be unlocked on entry.  If the allocation succeeds it will
1293  * be locked on return.
1294  *
1295  * Arguments:
1296  *	flags   Wait flags for the item initialization routine
1297  *	aflags  Wait flags for the slab allocation
1298  *
1299  * Returns:
1300  *	The slab that was allocated or NULL if there is no memory and the
1301  *	caller specified M_NOWAIT.
1302  */
1303 static uma_slab_t
1304 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int domain, int flags,
1305     int aflags)
1306 {
1307 	uma_domain_t dom;
1308 	uma_alloc allocf;
1309 	uma_slab_t slab;
1310 	unsigned long size;
1311 	uint8_t *mem;
1312 	uint8_t sflags;
1313 	int i;
1314 
1315 	KASSERT(domain >= 0 && domain < vm_ndomains,
1316 	    ("keg_alloc_slab: domain %d out of range", domain));
1317 
1318 	allocf = keg->uk_allocf;
1319 	slab = NULL;
1320 	mem = NULL;
1321 	if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) {
1322 		uma_hash_slab_t hslab;
1323 		hslab = zone_alloc_item(slabzone(keg->uk_ipers), NULL,
1324 		    domain, aflags);
1325 		if (hslab == NULL)
1326 			goto fail;
1327 		slab = &hslab->uhs_slab;
1328 	}
1329 
1330 	/*
1331 	 * This reproduces the old vm_zone behavior of zero filling pages the
1332 	 * first time they are added to a zone.
1333 	 *
1334 	 * Malloced items are zeroed in uma_zalloc.
1335 	 */
1336 
1337 	if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
1338 		aflags |= M_ZERO;
1339 	else
1340 		aflags &= ~M_ZERO;
1341 
1342 	if (keg->uk_flags & UMA_ZONE_NODUMP)
1343 		aflags |= M_NODUMP;
1344 
1345 	/* zone is passed for legacy reasons. */
1346 	size = keg->uk_ppera * PAGE_SIZE;
1347 	mem = allocf(zone, size, domain, &sflags, aflags);
1348 	if (mem == NULL) {
1349 		if (keg->uk_flags & UMA_ZFLAG_OFFPAGE)
1350 			zone_free_item(slabzone(keg->uk_ipers),
1351 			    slab_tohashslab(slab), NULL, SKIP_NONE);
1352 		goto fail;
1353 	}
1354 	uma_total_inc(size);
1355 
1356 	/* For HASH zones all pages go to the same uma_domain. */
1357 	if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
1358 		domain = 0;
1359 
1360 	/* Point the slab into the allocated memory */
1361 	if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE))
1362 		slab = (uma_slab_t )(mem + keg->uk_pgoff);
1363 	else
1364 		slab_tohashslab(slab)->uhs_data = mem;
1365 
1366 	if (keg->uk_flags & UMA_ZFLAG_VTOSLAB)
1367 		for (i = 0; i < keg->uk_ppera; i++)
1368 			vsetzoneslab((vm_offset_t)mem + (i * PAGE_SIZE),
1369 			    zone, slab);
1370 
1371 	slab->us_freecount = keg->uk_ipers;
1372 	slab->us_flags = sflags;
1373 	slab->us_domain = domain;
1374 
1375 	BIT_FILL(keg->uk_ipers, &slab->us_free);
1376 #ifdef INVARIANTS
1377 	BIT_ZERO(keg->uk_ipers, slab_dbg_bits(slab, keg));
1378 #endif
1379 
1380 	if (keg->uk_init != NULL) {
1381 		for (i = 0; i < keg->uk_ipers; i++)
1382 			if (keg->uk_init(slab_item(slab, keg, i),
1383 			    keg->uk_size, flags) != 0)
1384 				break;
1385 		if (i != keg->uk_ipers) {
1386 			keg_free_slab(keg, slab, i);
1387 			goto fail;
1388 		}
1389 	}
1390 	KEG_LOCK(keg, domain);
1391 
1392 	CTR3(KTR_UMA, "keg_alloc_slab: allocated slab %p for %s(%p)",
1393 	    slab, keg->uk_name, keg);
1394 
1395 	if (keg->uk_flags & UMA_ZFLAG_HASH)
1396 		UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
1397 
1398 	/*
1399 	 * If we got a slab here it's safe to mark it partially used
1400 	 * and return.  We assume that the caller is going to remove
1401 	 * at least one item.
1402 	 */
1403 	dom = &keg->uk_domain[domain];
1404 	LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
1405 	dom->ud_pages += keg->uk_ppera;
1406 	dom->ud_free += keg->uk_ipers;
1407 
1408 	return (slab);
1409 
1410 fail:
1411 	return (NULL);
1412 }
1413 
1414 /*
1415  * This function is intended to be used early on in place of page_alloc() so
1416  * that we may use the boot time page cache to satisfy allocations before
1417  * the VM is ready.
1418  */
1419 static void *
1420 startup_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1421     int wait)
1422 {
1423 	vm_paddr_t pa;
1424 	vm_page_t m;
1425 	void *mem;
1426 	int pages;
1427 	int i;
1428 
1429 	pages = howmany(bytes, PAGE_SIZE);
1430 	KASSERT(pages > 0, ("%s can't reserve 0 pages", __func__));
1431 
1432 	*pflag = UMA_SLAB_BOOT;
1433 	m = vm_page_alloc_contig_domain(NULL, 0, domain,
1434 	    malloc2vm_flags(wait) | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED, pages,
1435 	    (vm_paddr_t)0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT);
1436 	if (m == NULL)
1437 		return (NULL);
1438 
1439 	pa = VM_PAGE_TO_PHYS(m);
1440 	for (i = 0; i < pages; i++, pa += PAGE_SIZE) {
1441 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
1442     defined(__riscv) || defined(__powerpc64__)
1443 		if ((wait & M_NODUMP) == 0)
1444 			dump_add_page(pa);
1445 #endif
1446 	}
1447 	/* Allocate KVA and indirectly advance bootmem. */
1448 	mem = (void *)pmap_map(&bootmem, m->phys_addr,
1449 	    m->phys_addr + (pages * PAGE_SIZE), VM_PROT_READ | VM_PROT_WRITE);
1450         if ((wait & M_ZERO) != 0)
1451                 bzero(mem, pages * PAGE_SIZE);
1452 
1453         return (mem);
1454 }
1455 
1456 static void
1457 startup_free(void *mem, vm_size_t bytes)
1458 {
1459 	vm_offset_t va;
1460 	vm_page_t m;
1461 
1462 	va = (vm_offset_t)mem;
1463 	m = PHYS_TO_VM_PAGE(pmap_kextract(va));
1464 	pmap_remove(kernel_pmap, va, va + bytes);
1465 	for (; bytes != 0; bytes -= PAGE_SIZE, m++) {
1466 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
1467     defined(__riscv) || defined(__powerpc64__)
1468 		dump_drop_page(VM_PAGE_TO_PHYS(m));
1469 #endif
1470 		vm_page_unwire_noq(m);
1471 		vm_page_free(m);
1472 	}
1473 }
1474 
1475 /*
1476  * Allocates a number of pages from the system
1477  *
1478  * Arguments:
1479  *	bytes  The number of bytes requested
1480  *	wait  Shall we wait?
1481  *
1482  * Returns:
1483  *	A pointer to the alloced memory or possibly
1484  *	NULL if M_NOWAIT is set.
1485  */
1486 static void *
1487 page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1488     int wait)
1489 {
1490 	void *p;	/* Returned page */
1491 
1492 	*pflag = UMA_SLAB_KERNEL;
1493 	p = (void *)kmem_malloc_domainset(DOMAINSET_FIXED(domain), bytes, wait);
1494 
1495 	return (p);
1496 }
1497 
1498 static void *
1499 pcpu_page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1500     int wait)
1501 {
1502 	struct pglist alloctail;
1503 	vm_offset_t addr, zkva;
1504 	int cpu, flags;
1505 	vm_page_t p, p_next;
1506 #ifdef NUMA
1507 	struct pcpu *pc;
1508 #endif
1509 
1510 	MPASS(bytes == (mp_maxid + 1) * PAGE_SIZE);
1511 
1512 	TAILQ_INIT(&alloctail);
1513 	flags = VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | VM_ALLOC_NOOBJ |
1514 	    malloc2vm_flags(wait);
1515 	*pflag = UMA_SLAB_KERNEL;
1516 	for (cpu = 0; cpu <= mp_maxid; cpu++) {
1517 		if (CPU_ABSENT(cpu)) {
1518 			p = vm_page_alloc(NULL, 0, flags);
1519 		} else {
1520 #ifndef NUMA
1521 			p = vm_page_alloc(NULL, 0, flags);
1522 #else
1523 			pc = pcpu_find(cpu);
1524 			p = vm_page_alloc_domain(NULL, 0, pc->pc_domain, flags);
1525 			if (__predict_false(p == NULL))
1526 				p = vm_page_alloc(NULL, 0, flags);
1527 #endif
1528 		}
1529 		if (__predict_false(p == NULL))
1530 			goto fail;
1531 		TAILQ_INSERT_TAIL(&alloctail, p, listq);
1532 	}
1533 	if ((addr = kva_alloc(bytes)) == 0)
1534 		goto fail;
1535 	zkva = addr;
1536 	TAILQ_FOREACH(p, &alloctail, listq) {
1537 		pmap_qenter(zkva, &p, 1);
1538 		zkva += PAGE_SIZE;
1539 	}
1540 	return ((void*)addr);
1541 fail:
1542 	TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1543 		vm_page_unwire_noq(p);
1544 		vm_page_free(p);
1545 	}
1546 	return (NULL);
1547 }
1548 
1549 /*
1550  * Allocates a number of pages from within an object
1551  *
1552  * Arguments:
1553  *	bytes  The number of bytes requested
1554  *	wait   Shall we wait?
1555  *
1556  * Returns:
1557  *	A pointer to the alloced memory or possibly
1558  *	NULL if M_NOWAIT is set.
1559  */
1560 static void *
1561 noobj_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags,
1562     int wait)
1563 {
1564 	TAILQ_HEAD(, vm_page) alloctail;
1565 	u_long npages;
1566 	vm_offset_t retkva, zkva;
1567 	vm_page_t p, p_next;
1568 	uma_keg_t keg;
1569 
1570 	TAILQ_INIT(&alloctail);
1571 	keg = zone->uz_keg;
1572 
1573 	npages = howmany(bytes, PAGE_SIZE);
1574 	while (npages > 0) {
1575 		p = vm_page_alloc_domain(NULL, 0, domain, VM_ALLOC_INTERRUPT |
1576 		    VM_ALLOC_WIRED | VM_ALLOC_NOOBJ |
1577 		    ((wait & M_WAITOK) != 0 ? VM_ALLOC_WAITOK :
1578 		    VM_ALLOC_NOWAIT));
1579 		if (p != NULL) {
1580 			/*
1581 			 * Since the page does not belong to an object, its
1582 			 * listq is unused.
1583 			 */
1584 			TAILQ_INSERT_TAIL(&alloctail, p, listq);
1585 			npages--;
1586 			continue;
1587 		}
1588 		/*
1589 		 * Page allocation failed, free intermediate pages and
1590 		 * exit.
1591 		 */
1592 		TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1593 			vm_page_unwire_noq(p);
1594 			vm_page_free(p);
1595 		}
1596 		return (NULL);
1597 	}
1598 	*flags = UMA_SLAB_PRIV;
1599 	zkva = keg->uk_kva +
1600 	    atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
1601 	retkva = zkva;
1602 	TAILQ_FOREACH(p, &alloctail, listq) {
1603 		pmap_qenter(zkva, &p, 1);
1604 		zkva += PAGE_SIZE;
1605 	}
1606 
1607 	return ((void *)retkva);
1608 }
1609 
1610 /*
1611  * Frees a number of pages to the system
1612  *
1613  * Arguments:
1614  *	mem   A pointer to the memory to be freed
1615  *	size  The size of the memory being freed
1616  *	flags The original p->us_flags field
1617  *
1618  * Returns:
1619  *	Nothing
1620  */
1621 static void
1622 page_free(void *mem, vm_size_t size, uint8_t flags)
1623 {
1624 
1625 	if ((flags & UMA_SLAB_BOOT) != 0) {
1626 		startup_free(mem, size);
1627 		return;
1628 	}
1629 
1630 	if ((flags & UMA_SLAB_KERNEL) == 0)
1631 		panic("UMA: page_free used with invalid flags %x", flags);
1632 
1633 	kmem_free((vm_offset_t)mem, size);
1634 }
1635 
1636 /*
1637  * Frees pcpu zone allocations
1638  *
1639  * Arguments:
1640  *	mem   A pointer to the memory to be freed
1641  *	size  The size of the memory being freed
1642  *	flags The original p->us_flags field
1643  *
1644  * Returns:
1645  *	Nothing
1646  */
1647 static void
1648 pcpu_page_free(void *mem, vm_size_t size, uint8_t flags)
1649 {
1650 	vm_offset_t sva, curva;
1651 	vm_paddr_t paddr;
1652 	vm_page_t m;
1653 
1654 	MPASS(size == (mp_maxid+1)*PAGE_SIZE);
1655 	sva = (vm_offset_t)mem;
1656 	for (curva = sva; curva < sva + size; curva += PAGE_SIZE) {
1657 		paddr = pmap_kextract(curva);
1658 		m = PHYS_TO_VM_PAGE(paddr);
1659 		vm_page_unwire_noq(m);
1660 		vm_page_free(m);
1661 	}
1662 	pmap_qremove(sva, size >> PAGE_SHIFT);
1663 	kva_free(sva, size);
1664 }
1665 
1666 
1667 /*
1668  * Zero fill initializer
1669  *
1670  * Arguments/Returns follow uma_init specifications
1671  */
1672 static int
1673 zero_init(void *mem, int size, int flags)
1674 {
1675 	bzero(mem, size);
1676 	return (0);
1677 }
1678 
1679 #ifdef INVARIANTS
1680 struct noslabbits *
1681 slab_dbg_bits(uma_slab_t slab, uma_keg_t keg)
1682 {
1683 
1684 	return ((void *)((char *)&slab->us_free + BITSET_SIZE(keg->uk_ipers)));
1685 }
1686 #endif
1687 
1688 /*
1689  * Actual size of embedded struct slab (!OFFPAGE).
1690  */
1691 size_t
1692 slab_sizeof(int nitems)
1693 {
1694 	size_t s;
1695 
1696 	s = sizeof(struct uma_slab) + BITSET_SIZE(nitems) * SLAB_BITSETS;
1697 	return (roundup(s, UMA_ALIGN_PTR + 1));
1698 }
1699 
1700 /*
1701  * Size of memory for embedded slabs (!OFFPAGE).
1702  */
1703 size_t
1704 slab_space(int nitems)
1705 {
1706 	return (UMA_SLAB_SIZE - slab_sizeof(nitems));
1707 }
1708 
1709 #define	UMA_FIXPT_SHIFT	31
1710 #define	UMA_FRAC_FIXPT(n, d)						\
1711 	((uint32_t)(((uint64_t)(n) << UMA_FIXPT_SHIFT) / (d)))
1712 #define	UMA_FIXPT_PCT(f)						\
1713 	((u_int)(((uint64_t)100 * (f)) >> UMA_FIXPT_SHIFT))
1714 #define	UMA_PCT_FIXPT(pct)	UMA_FRAC_FIXPT((pct), 100)
1715 #define	UMA_MIN_EFF	UMA_PCT_FIXPT(100 - UMA_MAX_WASTE)
1716 
1717 /*
1718  * Compute the number of items that will fit in a slab.  If hdr is true, the
1719  * item count may be limited to provide space in the slab for an inline slab
1720  * header.  Otherwise, all slab space will be provided for item storage.
1721  */
1722 static u_int
1723 slab_ipers_hdr(u_int size, u_int rsize, u_int slabsize, bool hdr)
1724 {
1725 	u_int ipers;
1726 	u_int padpi;
1727 
1728 	/* The padding between items is not needed after the last item. */
1729 	padpi = rsize - size;
1730 
1731 	if (hdr) {
1732 		/*
1733 		 * Start with the maximum item count and remove items until
1734 		 * the slab header first alongside the allocatable memory.
1735 		 */
1736 		for (ipers = MIN(SLAB_MAX_SETSIZE,
1737 		    (slabsize + padpi - slab_sizeof(1)) / rsize);
1738 		    ipers > 0 &&
1739 		    ipers * rsize - padpi + slab_sizeof(ipers) > slabsize;
1740 		    ipers--)
1741 			continue;
1742 	} else {
1743 		ipers = MIN((slabsize + padpi) / rsize, SLAB_MAX_SETSIZE);
1744 	}
1745 
1746 	return (ipers);
1747 }
1748 
1749 /*
1750  * Compute the number of items that will fit in a slab for a startup zone.
1751  */
1752 int
1753 slab_ipers(size_t size, int align)
1754 {
1755 	int rsize;
1756 
1757 	rsize = roundup(size, align + 1); /* Assume no CACHESPREAD */
1758 	return (slab_ipers_hdr(size, rsize, UMA_SLAB_SIZE, true));
1759 }
1760 
1761 /*
1762  * Determine the format of a uma keg.  This determines where the slab header
1763  * will be placed (inline or offpage) and calculates ipers, rsize, and ppera.
1764  *
1765  * Arguments
1766  *	keg  The zone we should initialize
1767  *
1768  * Returns
1769  *	Nothing
1770  */
1771 static void
1772 keg_layout(uma_keg_t keg)
1773 {
1774 	u_int alignsize;
1775 	u_int eff;
1776 	u_int eff_offpage;
1777 	u_int format;
1778 	u_int ipers;
1779 	u_int ipers_offpage;
1780 	u_int pages;
1781 	u_int rsize;
1782 	u_int slabsize;
1783 
1784 	KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
1785 	    (keg->uk_size <= UMA_PCPU_ALLOC_SIZE &&
1786 	     (keg->uk_flags & UMA_ZONE_CACHESPREAD) == 0),
1787 	    ("%s: cannot configure for PCPU: keg=%s, size=%u, flags=0x%b",
1788 	     __func__, keg->uk_name, keg->uk_size, keg->uk_flags,
1789 	     PRINT_UMA_ZFLAGS));
1790 	KASSERT((keg->uk_flags &
1791 	    (UMA_ZFLAG_INTERNAL | UMA_ZFLAG_CACHEONLY)) == 0 ||
1792 	    (keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0,
1793 	    ("%s: incompatible flags 0x%b", __func__, keg->uk_flags,
1794 	     PRINT_UMA_ZFLAGS));
1795 
1796 	alignsize = keg->uk_align + 1;
1797 	format = 0;
1798 	ipers = 0;
1799 
1800 	/*
1801 	 * Calculate the size of each allocation (rsize) according to
1802 	 * alignment.  If the requested size is smaller than we have
1803 	 * allocation bits for we round it up.
1804 	 */
1805 	rsize = MAX(keg->uk_size, UMA_SMALLEST_UNIT);
1806 	rsize = roundup2(rsize, alignsize);
1807 
1808 	if ((keg->uk_flags & UMA_ZONE_PCPU) != 0) {
1809 		slabsize = UMA_PCPU_ALLOC_SIZE;
1810 		pages = mp_maxid + 1;
1811 	} else if ((keg->uk_flags & UMA_ZONE_CACHESPREAD) != 0) {
1812 		/*
1813 		 * We want one item to start on every align boundary in a page.
1814 		 * To do this we will span pages.  We will also extend the item
1815 		 * by the size of align if it is an even multiple of align.
1816 		 * Otherwise, it would fall on the same boundary every time.
1817 		 */
1818 		if ((rsize & alignsize) == 0)
1819 			rsize += alignsize;
1820 		slabsize = rsize * (PAGE_SIZE / alignsize);
1821 		slabsize = MIN(slabsize, rsize * SLAB_MAX_SETSIZE);
1822 		slabsize = MIN(slabsize, UMA_CACHESPREAD_MAX_SIZE);
1823 		pages = howmany(slabsize, PAGE_SIZE);
1824 		slabsize = ptoa(pages);
1825 	} else {
1826 		/*
1827 		 * Choose a slab size of as many pages as it takes to represent
1828 		 * a single item.  We will then try to fit as many additional
1829 		 * items into the slab as possible.  At some point, we may want
1830 		 * to increase the slab size for awkward item sizes in order to
1831 		 * increase efficiency.
1832 		 */
1833 		pages = howmany(keg->uk_size, PAGE_SIZE);
1834 		slabsize = ptoa(pages);
1835 	}
1836 
1837 	/* Evaluate an inline slab layout. */
1838 	if ((keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0)
1839 		ipers = slab_ipers_hdr(keg->uk_size, rsize, slabsize, true);
1840 
1841 	/* TODO: vm_page-embedded slab. */
1842 
1843 	/*
1844 	 * We can't do OFFPAGE if we're internal or if we've been
1845 	 * asked to not go to the VM for buckets.  If we do this we
1846 	 * may end up going to the VM  for slabs which we do not
1847 	 * want to do if we're UMA_ZFLAG_CACHEONLY as a result
1848 	 * of UMA_ZONE_VM, which clearly forbids it.
1849 	 */
1850 	if ((keg->uk_flags &
1851 	    (UMA_ZFLAG_INTERNAL | UMA_ZFLAG_CACHEONLY)) != 0) {
1852 		if (ipers == 0) {
1853 			/* We need an extra page for the slab header. */
1854 			pages++;
1855 			slabsize = ptoa(pages);
1856 			ipers = slab_ipers_hdr(keg->uk_size, rsize, slabsize,
1857 			    true);
1858 		}
1859 		goto out;
1860 	}
1861 
1862 	/*
1863 	 * See if using an OFFPAGE slab will improve our efficiency.
1864 	 * Only do this if we are below our efficiency threshold.
1865 	 *
1866 	 * XXX We could try growing slabsize to limit max waste as well.
1867 	 * Historically this was not done because the VM could not
1868 	 * efficiently handle contiguous allocations.
1869 	 */
1870 	eff = UMA_FRAC_FIXPT(ipers * rsize, slabsize);
1871 	ipers_offpage = slab_ipers_hdr(keg->uk_size, rsize, slabsize, false);
1872 	eff_offpage = UMA_FRAC_FIXPT(ipers_offpage * rsize,
1873 	    slabsize + slabzone(ipers_offpage)->uz_keg->uk_rsize);
1874 	if (ipers == 0 || (eff < UMA_MIN_EFF && eff < eff_offpage)) {
1875 		CTR5(KTR_UMA, "UMA decided we need offpage slab headers for "
1876 		    "keg: %s(%p), minimum efficiency allowed = %u%%, "
1877 		    "old efficiency = %u%%, offpage efficiency = %u%%",
1878 		    keg->uk_name, keg, UMA_FIXPT_PCT(UMA_MIN_EFF),
1879 		    UMA_FIXPT_PCT(eff), UMA_FIXPT_PCT(eff_offpage));
1880 		format = UMA_ZFLAG_OFFPAGE;
1881 		ipers = ipers_offpage;
1882 	}
1883 
1884 out:
1885 	/*
1886 	 * How do we find the slab header if it is offpage or if not all item
1887 	 * start addresses are in the same page?  We could solve the latter
1888 	 * case with vaddr alignment, but we don't.
1889 	 */
1890 	if ((format & UMA_ZFLAG_OFFPAGE) != 0 ||
1891 	    (ipers - 1) * rsize >= PAGE_SIZE) {
1892 		if ((keg->uk_flags & UMA_ZONE_NOTPAGE) != 0)
1893 			format |= UMA_ZFLAG_HASH;
1894 		else
1895 			format |= UMA_ZFLAG_VTOSLAB;
1896 	}
1897 	keg->uk_ipers = ipers;
1898 	keg->uk_rsize = rsize;
1899 	keg->uk_flags |= format;
1900 	keg->uk_ppera = pages;
1901 	CTR6(KTR_UMA, "%s: keg=%s, flags=%#x, rsize=%u, ipers=%u, ppera=%u",
1902 	    __func__, keg->uk_name, keg->uk_flags, rsize, ipers, pages);
1903 	KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_MAX_SETSIZE,
1904 	    ("%s: keg=%s, flags=0x%b, rsize=%u, ipers=%u, ppera=%u", __func__,
1905 	     keg->uk_name, keg->uk_flags, PRINT_UMA_ZFLAGS, rsize, ipers,
1906 	     pages));
1907 }
1908 
1909 /*
1910  * Keg header ctor.  This initializes all fields, locks, etc.  And inserts
1911  * the keg onto the global keg list.
1912  *
1913  * Arguments/Returns follow uma_ctor specifications
1914  *	udata  Actually uma_kctor_args
1915  */
1916 static int
1917 keg_ctor(void *mem, int size, void *udata, int flags)
1918 {
1919 	struct uma_kctor_args *arg = udata;
1920 	uma_keg_t keg = mem;
1921 	uma_zone_t zone;
1922 	int i;
1923 
1924 	bzero(keg, size);
1925 	keg->uk_size = arg->size;
1926 	keg->uk_init = arg->uminit;
1927 	keg->uk_fini = arg->fini;
1928 	keg->uk_align = arg->align;
1929 	keg->uk_reserve = 0;
1930 	keg->uk_flags = arg->flags;
1931 
1932 	/*
1933 	 * We use a global round-robin policy by default.  Zones with
1934 	 * UMA_ZONE_FIRSTTOUCH set will use first-touch instead, in which
1935 	 * case the iterator is never run.
1936 	 */
1937 	keg->uk_dr.dr_policy = DOMAINSET_RR();
1938 	keg->uk_dr.dr_iter = 0;
1939 
1940 	/*
1941 	 * The master zone is passed to us at keg-creation time.
1942 	 */
1943 	zone = arg->zone;
1944 	keg->uk_name = zone->uz_name;
1945 
1946 	if (arg->flags & UMA_ZONE_VM)
1947 		keg->uk_flags |= UMA_ZFLAG_CACHEONLY;
1948 
1949 	if (arg->flags & UMA_ZONE_ZINIT)
1950 		keg->uk_init = zero_init;
1951 
1952 	if (arg->flags & UMA_ZONE_MALLOC)
1953 		keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
1954 
1955 #ifndef SMP
1956 	keg->uk_flags &= ~UMA_ZONE_PCPU;
1957 #endif
1958 
1959 	keg_layout(keg);
1960 
1961 	/*
1962 	 * Use a first-touch NUMA policy for all kegs that pmap_extract()
1963 	 * will work on with the exception of critical VM structures
1964 	 * necessary for paging.
1965 	 *
1966 	 * Zones may override the default by specifying either.
1967 	 */
1968 #ifdef NUMA
1969 	if ((keg->uk_flags &
1970 	    (UMA_ZFLAG_HASH | UMA_ZONE_VM | UMA_ZONE_ROUNDROBIN)) == 0)
1971 		keg->uk_flags |= UMA_ZONE_FIRSTTOUCH;
1972 	else if ((keg->uk_flags & UMA_ZONE_FIRSTTOUCH) == 0)
1973 		keg->uk_flags |= UMA_ZONE_ROUNDROBIN;
1974 #endif
1975 
1976 	/*
1977 	 * If we haven't booted yet we need allocations to go through the
1978 	 * startup cache until the vm is ready.
1979 	 */
1980 #ifdef UMA_MD_SMALL_ALLOC
1981 	if (keg->uk_ppera == 1)
1982 		keg->uk_allocf = uma_small_alloc;
1983 	else
1984 #endif
1985 	if (booted < BOOT_KVA)
1986 		keg->uk_allocf = startup_alloc;
1987 	else if (keg->uk_flags & UMA_ZONE_PCPU)
1988 		keg->uk_allocf = pcpu_page_alloc;
1989 	else
1990 		keg->uk_allocf = page_alloc;
1991 #ifdef UMA_MD_SMALL_ALLOC
1992 	if (keg->uk_ppera == 1)
1993 		keg->uk_freef = uma_small_free;
1994 	else
1995 #endif
1996 	if (keg->uk_flags & UMA_ZONE_PCPU)
1997 		keg->uk_freef = pcpu_page_free;
1998 	else
1999 		keg->uk_freef = page_free;
2000 
2001 	/*
2002 	 * Initialize keg's locks.
2003 	 */
2004 	for (i = 0; i < vm_ndomains; i++)
2005 		KEG_LOCK_INIT(keg, i, (arg->flags & UMA_ZONE_MTXCLASS));
2006 
2007 	/*
2008 	 * If we're putting the slab header in the actual page we need to
2009 	 * figure out where in each page it goes.  See slab_sizeof
2010 	 * definition.
2011 	 */
2012 	if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE)) {
2013 		size_t shsize;
2014 
2015 		shsize = slab_sizeof(keg->uk_ipers);
2016 		keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - shsize;
2017 		/*
2018 		 * The only way the following is possible is if with our
2019 		 * UMA_ALIGN_PTR adjustments we are now bigger than
2020 		 * UMA_SLAB_SIZE.  I haven't checked whether this is
2021 		 * mathematically possible for all cases, so we make
2022 		 * sure here anyway.
2023 		 */
2024 		KASSERT(keg->uk_pgoff + shsize <= PAGE_SIZE * keg->uk_ppera,
2025 		    ("zone %s ipers %d rsize %d size %d slab won't fit",
2026 		    zone->uz_name, keg->uk_ipers, keg->uk_rsize, keg->uk_size));
2027 	}
2028 
2029 	if (keg->uk_flags & UMA_ZFLAG_HASH)
2030 		hash_alloc(&keg->uk_hash, 0);
2031 
2032 	CTR3(KTR_UMA, "keg_ctor %p zone %s(%p)", keg, zone->uz_name, zone);
2033 
2034 	LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
2035 
2036 	rw_wlock(&uma_rwlock);
2037 	LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
2038 	rw_wunlock(&uma_rwlock);
2039 	return (0);
2040 }
2041 
2042 static void
2043 zone_kva_available(uma_zone_t zone, void *unused)
2044 {
2045 	uma_keg_t keg;
2046 
2047 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
2048 		return;
2049 	KEG_GET(zone, keg);
2050 	if (keg->uk_allocf == startup_alloc)
2051 		keg->uk_allocf = page_alloc;
2052 }
2053 
2054 static void
2055 zone_alloc_counters(uma_zone_t zone, void *unused)
2056 {
2057 
2058 	zone->uz_allocs = counter_u64_alloc(M_WAITOK);
2059 	zone->uz_frees = counter_u64_alloc(M_WAITOK);
2060 	zone->uz_fails = counter_u64_alloc(M_WAITOK);
2061 }
2062 
2063 static void
2064 zone_alloc_sysctl(uma_zone_t zone, void *unused)
2065 {
2066 	uma_zone_domain_t zdom;
2067 	uma_domain_t dom;
2068 	uma_keg_t keg;
2069 	struct sysctl_oid *oid, *domainoid;
2070 	int domains, i, cnt;
2071 	static const char *nokeg = "cache zone";
2072 	char *c;
2073 
2074 	/*
2075 	 * Make a sysctl safe copy of the zone name by removing
2076 	 * any special characters and handling dups by appending
2077 	 * an index.
2078 	 */
2079 	if (zone->uz_namecnt != 0) {
2080 		/* Count the number of decimal digits and '_' separator. */
2081 		for (i = 1, cnt = zone->uz_namecnt; cnt != 0; i++)
2082 			cnt /= 10;
2083 		zone->uz_ctlname = malloc(strlen(zone->uz_name) + i + 1,
2084 		    M_UMA, M_WAITOK);
2085 		sprintf(zone->uz_ctlname, "%s_%d", zone->uz_name,
2086 		    zone->uz_namecnt);
2087 	} else
2088 		zone->uz_ctlname = strdup(zone->uz_name, M_UMA);
2089 	for (c = zone->uz_ctlname; *c != '\0'; c++)
2090 		if (strchr("./\\ -", *c) != NULL)
2091 			*c = '_';
2092 
2093 	/*
2094 	 * Basic parameters at the root.
2095 	 */
2096 	zone->uz_oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_vm_uma),
2097 	    OID_AUTO, zone->uz_ctlname, CTLFLAG_RD, NULL, "");
2098 	oid = zone->uz_oid;
2099 	SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2100 	    "size", CTLFLAG_RD, &zone->uz_size, 0, "Allocation size");
2101 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2102 	    "flags", CTLFLAG_RD | CTLTYPE_STRING | CTLFLAG_MPSAFE,
2103 	    zone, 0, sysctl_handle_uma_zone_flags, "A",
2104 	    "Allocator configuration flags");
2105 	SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2106 	    "bucket_size", CTLFLAG_RD, &zone->uz_bucket_size, 0,
2107 	    "Desired per-cpu cache size");
2108 	SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2109 	    "bucket_size_max", CTLFLAG_RD, &zone->uz_bucket_size_max, 0,
2110 	    "Maximum allowed per-cpu cache size");
2111 
2112 	/*
2113 	 * keg if present.
2114 	 */
2115 	if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
2116 		domains = vm_ndomains;
2117 	else
2118 		domains = 1;
2119 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2120 	    "keg", CTLFLAG_RD, NULL, "");
2121 	keg = zone->uz_keg;
2122 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0) {
2123 		SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2124 		    "name", CTLFLAG_RD, keg->uk_name, "Keg name");
2125 		SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2126 		    "rsize", CTLFLAG_RD, &keg->uk_rsize, 0,
2127 		    "Real object size with alignment");
2128 		SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2129 		    "ppera", CTLFLAG_RD, &keg->uk_ppera, 0,
2130 		    "pages per-slab allocation");
2131 		SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2132 		    "ipers", CTLFLAG_RD, &keg->uk_ipers, 0,
2133 		    "items available per-slab");
2134 		SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2135 		    "align", CTLFLAG_RD, &keg->uk_align, 0,
2136 		    "item alignment mask");
2137 		SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2138 		    "efficiency", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2139 		    keg, 0, sysctl_handle_uma_slab_efficiency, "I",
2140 		    "Slab utilization (100 - internal fragmentation %)");
2141 		domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(oid),
2142 		    OID_AUTO, "domain", CTLFLAG_RD, NULL, "");
2143 		for (i = 0; i < domains; i++) {
2144 			dom = &keg->uk_domain[i];
2145 			oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2146 			    OID_AUTO, VM_DOMAIN(i)->vmd_name, CTLFLAG_RD,
2147 			    NULL, "");
2148 			SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2149 			    "pages", CTLFLAG_RD, &dom->ud_pages, 0,
2150 			    "Total pages currently allocated from VM");
2151 			SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2152 			    "free", CTLFLAG_RD, &dom->ud_free, 0,
2153 			    "items free in the slab layer");
2154 		}
2155 	} else
2156 		SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2157 		    "name", CTLFLAG_RD, nokeg, "Keg name");
2158 
2159 	/*
2160 	 * Information about zone limits.
2161 	 */
2162 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2163 	    "limit", CTLFLAG_RD, NULL, "");
2164 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2165 	    "items", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2166 	    zone, 0, sysctl_handle_uma_zone_items, "QU",
2167 	    "current number of allocated items if limit is set");
2168 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2169 	    "max_items", CTLFLAG_RD, &zone->uz_max_items, 0,
2170 	    "Maximum number of cached items");
2171 	SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2172 	    "sleepers", CTLFLAG_RD, &zone->uz_sleepers, 0,
2173 	    "Number of threads sleeping at limit");
2174 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2175 	    "sleeps", CTLFLAG_RD, &zone->uz_sleeps, 0,
2176 	    "Total zone limit sleeps");
2177 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2178 	    "bucket_max", CTLFLAG_RD, &zone->uz_bkt_max, 0,
2179 	    "Maximum number of items in the bucket cache");
2180 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2181 	    "bucket_cnt", CTLFLAG_RD, &zone->uz_bkt_count, 0,
2182 	    "Number of items in the bucket cache");
2183 
2184 	/*
2185 	 * Per-domain zone information.
2186 	 */
2187 	domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid),
2188 	    OID_AUTO, "domain", CTLFLAG_RD, NULL, "");
2189 	if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2190 		domains = 1;
2191 	for (i = 0; i < domains; i++) {
2192 		zdom = &zone->uz_domain[i];
2193 		oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2194 		    OID_AUTO, VM_DOMAIN(i)->vmd_name, CTLFLAG_RD, NULL, "");
2195 		SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2196 		    "nitems", CTLFLAG_RD, &zdom->uzd_nitems,
2197 		    "number of items in this domain");
2198 		SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2199 		    "imax", CTLFLAG_RD, &zdom->uzd_imax,
2200 		    "maximum item count in this period");
2201 		SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2202 		    "imin", CTLFLAG_RD, &zdom->uzd_imin,
2203 		    "minimum item count in this period");
2204 		SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2205 		    "wss", CTLFLAG_RD, &zdom->uzd_wss,
2206 		    "Working set size");
2207 	}
2208 
2209 	/*
2210 	 * General statistics.
2211 	 */
2212 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2213 	    "stats", CTLFLAG_RD, NULL, "");
2214 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2215 	    "current", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2216 	    zone, 1, sysctl_handle_uma_zone_cur, "I",
2217 	    "Current number of allocated items");
2218 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2219 	    "allocs", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2220 	    zone, 0, sysctl_handle_uma_zone_allocs, "QU",
2221 	    "Total allocation calls");
2222 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2223 	    "frees", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2224 	    zone, 0, sysctl_handle_uma_zone_frees, "QU",
2225 	    "Total free calls");
2226 	SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2227 	    "fails", CTLFLAG_RD, &zone->uz_fails,
2228 	    "Number of allocation failures");
2229 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2230 	    "xdomain", CTLFLAG_RD, &zone->uz_xdomain, 0,
2231 	    "Free calls from the wrong domain");
2232 }
2233 
2234 struct uma_zone_count {
2235 	const char	*name;
2236 	int		count;
2237 };
2238 
2239 static void
2240 zone_count(uma_zone_t zone, void *arg)
2241 {
2242 	struct uma_zone_count *cnt;
2243 
2244 	cnt = arg;
2245 	/*
2246 	 * Some zones are rapidly created with identical names and
2247 	 * destroyed out of order.  This can lead to gaps in the count.
2248 	 * Use one greater than the maximum observed for this name.
2249 	 */
2250 	if (strcmp(zone->uz_name, cnt->name) == 0)
2251 		cnt->count = MAX(cnt->count,
2252 		    zone->uz_namecnt + 1);
2253 }
2254 
2255 static void
2256 zone_update_caches(uma_zone_t zone)
2257 {
2258 	int i;
2259 
2260 	for (i = 0; i <= mp_maxid; i++) {
2261 		cache_set_uz_size(&zone->uz_cpu[i], zone->uz_size);
2262 		cache_set_uz_flags(&zone->uz_cpu[i], zone->uz_flags);
2263 	}
2264 }
2265 
2266 /*
2267  * Zone header ctor.  This initializes all fields, locks, etc.
2268  *
2269  * Arguments/Returns follow uma_ctor specifications
2270  *	udata  Actually uma_zctor_args
2271  */
2272 static int
2273 zone_ctor(void *mem, int size, void *udata, int flags)
2274 {
2275 	struct uma_zone_count cnt;
2276 	struct uma_zctor_args *arg = udata;
2277 	uma_zone_t zone = mem;
2278 	uma_zone_t z;
2279 	uma_keg_t keg;
2280 	int i;
2281 
2282 	bzero(zone, size);
2283 	zone->uz_name = arg->name;
2284 	zone->uz_ctor = arg->ctor;
2285 	zone->uz_dtor = arg->dtor;
2286 	zone->uz_init = NULL;
2287 	zone->uz_fini = NULL;
2288 	zone->uz_sleeps = 0;
2289 	zone->uz_xdomain = 0;
2290 	zone->uz_bucket_size = 0;
2291 	zone->uz_bucket_size_min = 0;
2292 	zone->uz_bucket_size_max = BUCKET_MAX;
2293 	zone->uz_flags = 0;
2294 	zone->uz_warning = NULL;
2295 	/* The domain structures follow the cpu structures. */
2296 	zone->uz_domain = (struct uma_zone_domain *)&zone->uz_cpu[mp_ncpus];
2297 	zone->uz_bkt_max = ULONG_MAX;
2298 	timevalclear(&zone->uz_ratecheck);
2299 
2300 	/* Count the number of duplicate names. */
2301 	cnt.name = arg->name;
2302 	cnt.count = 0;
2303 	zone_foreach(zone_count, &cnt);
2304 	zone->uz_namecnt = cnt.count;
2305 	ZONE_LOCK_INIT(zone, (arg->flags & UMA_ZONE_MTXCLASS));
2306 	ZONE_CROSS_LOCK_INIT(zone);
2307 
2308 	for (i = 0; i < vm_ndomains; i++)
2309 		TAILQ_INIT(&zone->uz_domain[i].uzd_buckets);
2310 
2311 #ifdef INVARIANTS
2312 	if (arg->uminit == trash_init && arg->fini == trash_fini)
2313 		zone->uz_flags |= UMA_ZFLAG_TRASH | UMA_ZFLAG_CTORDTOR;
2314 #endif
2315 
2316 	/*
2317 	 * This is a pure cache zone, no kegs.
2318 	 */
2319 	if (arg->import) {
2320 		KASSERT((arg->flags & UMA_ZFLAG_CACHE) != 0,
2321 		    ("zone_ctor: Import specified for non-cache zone."));
2322 		if (arg->flags & UMA_ZONE_VM)
2323 			arg->flags |= UMA_ZFLAG_CACHEONLY;
2324 		zone->uz_flags = arg->flags;
2325 		zone->uz_size = arg->size;
2326 		zone->uz_import = arg->import;
2327 		zone->uz_release = arg->release;
2328 		zone->uz_arg = arg->arg;
2329 		rw_wlock(&uma_rwlock);
2330 		LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link);
2331 		rw_wunlock(&uma_rwlock);
2332 		goto out;
2333 	}
2334 
2335 	/*
2336 	 * Use the regular zone/keg/slab allocator.
2337 	 */
2338 	zone->uz_import = zone_import;
2339 	zone->uz_release = zone_release;
2340 	zone->uz_arg = zone;
2341 	keg = arg->keg;
2342 
2343 	if (arg->flags & UMA_ZONE_SECONDARY) {
2344 		KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
2345 		    ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
2346 		KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
2347 		zone->uz_init = arg->uminit;
2348 		zone->uz_fini = arg->fini;
2349 		zone->uz_flags |= UMA_ZONE_SECONDARY;
2350 		rw_wlock(&uma_rwlock);
2351 		ZONE_LOCK(zone);
2352 		LIST_FOREACH(z, &keg->uk_zones, uz_link) {
2353 			if (LIST_NEXT(z, uz_link) == NULL) {
2354 				LIST_INSERT_AFTER(z, zone, uz_link);
2355 				break;
2356 			}
2357 		}
2358 		ZONE_UNLOCK(zone);
2359 		rw_wunlock(&uma_rwlock);
2360 	} else if (keg == NULL) {
2361 		if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
2362 		    arg->align, arg->flags)) == NULL)
2363 			return (ENOMEM);
2364 	} else {
2365 		struct uma_kctor_args karg;
2366 		int error;
2367 
2368 		/* We should only be here from uma_startup() */
2369 		karg.size = arg->size;
2370 		karg.uminit = arg->uminit;
2371 		karg.fini = arg->fini;
2372 		karg.align = arg->align;
2373 		karg.flags = arg->flags;
2374 		karg.zone = zone;
2375 		error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
2376 		    flags);
2377 		if (error)
2378 			return (error);
2379 	}
2380 
2381 	/* Inherit properties from the keg. */
2382 	zone->uz_keg = keg;
2383 	zone->uz_size = keg->uk_size;
2384 	zone->uz_flags |= (keg->uk_flags &
2385 	    (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
2386 
2387 out:
2388 	if (__predict_true(booted >= BOOT_RUNNING)) {
2389 		zone_alloc_counters(zone, NULL);
2390 		zone_alloc_sysctl(zone, NULL);
2391 	} else {
2392 		zone->uz_allocs = EARLY_COUNTER;
2393 		zone->uz_frees = EARLY_COUNTER;
2394 		zone->uz_fails = EARLY_COUNTER;
2395 	}
2396 
2397 	KASSERT((arg->flags & (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET)) !=
2398 	    (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET),
2399 	    ("Invalid zone flag combination"));
2400 	if (arg->flags & UMA_ZFLAG_INTERNAL)
2401 		zone->uz_bucket_size_max = zone->uz_bucket_size = 0;
2402 	if ((arg->flags & UMA_ZONE_MAXBUCKET) != 0)
2403 		zone->uz_bucket_size = BUCKET_MAX;
2404 	else if ((arg->flags & UMA_ZONE_MINBUCKET) != 0)
2405 		zone->uz_bucket_size_max = zone->uz_bucket_size = BUCKET_MIN;
2406 	else if ((arg->flags & UMA_ZONE_NOBUCKET) != 0)
2407 		zone->uz_bucket_size = 0;
2408 	else
2409 		zone->uz_bucket_size = bucket_select(zone->uz_size);
2410 	zone->uz_bucket_size_min = zone->uz_bucket_size;
2411 	if (zone->uz_dtor != NULL || zone->uz_ctor != NULL)
2412 		zone->uz_flags |= UMA_ZFLAG_CTORDTOR;
2413 	zone_update_caches(zone);
2414 
2415 	return (0);
2416 }
2417 
2418 /*
2419  * Keg header dtor.  This frees all data, destroys locks, frees the hash
2420  * table and removes the keg from the global list.
2421  *
2422  * Arguments/Returns follow uma_dtor specifications
2423  *	udata  unused
2424  */
2425 static void
2426 keg_dtor(void *arg, int size, void *udata)
2427 {
2428 	uma_keg_t keg;
2429 	uint32_t free, pages;
2430 	int i;
2431 
2432 	keg = (uma_keg_t)arg;
2433 	free = pages = 0;
2434 	for (i = 0; i < vm_ndomains; i++) {
2435 		free += keg->uk_domain[i].ud_free;
2436 		pages += keg->uk_domain[i].ud_pages;
2437 		KEG_LOCK_FINI(keg, i);
2438 	}
2439 	if (free != 0)
2440 		printf("Freed UMA keg (%s) was not empty (%u items). "
2441 		    " Lost %u pages of memory.\n",
2442 		    keg->uk_name ? keg->uk_name : "",
2443 		    free, pages);
2444 
2445 	hash_free(&keg->uk_hash);
2446 }
2447 
2448 /*
2449  * Zone header dtor.
2450  *
2451  * Arguments/Returns follow uma_dtor specifications
2452  *	udata  unused
2453  */
2454 static void
2455 zone_dtor(void *arg, int size, void *udata)
2456 {
2457 	uma_zone_t zone;
2458 	uma_keg_t keg;
2459 
2460 	zone = (uma_zone_t)arg;
2461 
2462 	sysctl_remove_oid(zone->uz_oid, 1, 1);
2463 
2464 	if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
2465 		cache_drain(zone);
2466 
2467 	rw_wlock(&uma_rwlock);
2468 	LIST_REMOVE(zone, uz_link);
2469 	rw_wunlock(&uma_rwlock);
2470 	/*
2471 	 * XXX there are some races here where
2472 	 * the zone can be drained but zone lock
2473 	 * released and then refilled before we
2474 	 * remove it... we dont care for now
2475 	 */
2476 	zone_reclaim(zone, M_WAITOK, true);
2477 	/*
2478 	 * We only destroy kegs from non secondary/non cache zones.
2479 	 */
2480 	if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) {
2481 		keg = zone->uz_keg;
2482 		rw_wlock(&uma_rwlock);
2483 		LIST_REMOVE(keg, uk_link);
2484 		rw_wunlock(&uma_rwlock);
2485 		zone_free_item(kegs, keg, NULL, SKIP_NONE);
2486 	}
2487 	counter_u64_free(zone->uz_allocs);
2488 	counter_u64_free(zone->uz_frees);
2489 	counter_u64_free(zone->uz_fails);
2490 	free(zone->uz_ctlname, M_UMA);
2491 	ZONE_LOCK_FINI(zone);
2492 	ZONE_CROSS_LOCK_FINI(zone);
2493 }
2494 
2495 static void
2496 zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *arg), void *arg)
2497 {
2498 	uma_keg_t keg;
2499 	uma_zone_t zone;
2500 
2501 	LIST_FOREACH(keg, &uma_kegs, uk_link) {
2502 		LIST_FOREACH(zone, &keg->uk_zones, uz_link)
2503 			zfunc(zone, arg);
2504 	}
2505 	LIST_FOREACH(zone, &uma_cachezones, uz_link)
2506 		zfunc(zone, arg);
2507 }
2508 
2509 /*
2510  * Traverses every zone in the system and calls a callback
2511  *
2512  * Arguments:
2513  *	zfunc  A pointer to a function which accepts a zone
2514  *		as an argument.
2515  *
2516  * Returns:
2517  *	Nothing
2518  */
2519 static void
2520 zone_foreach(void (*zfunc)(uma_zone_t, void *arg), void *arg)
2521 {
2522 
2523 	rw_rlock(&uma_rwlock);
2524 	zone_foreach_unlocked(zfunc, arg);
2525 	rw_runlock(&uma_rwlock);
2526 }
2527 
2528 /*
2529  * Initialize the kernel memory allocator.  This is done after pages can be
2530  * allocated but before general KVA is available.
2531  */
2532 void
2533 uma_startup1(vm_offset_t virtual_avail)
2534 {
2535 	struct uma_zctor_args args;
2536 	size_t ksize, zsize, size;
2537 	uma_keg_t masterkeg;
2538 	uintptr_t m;
2539 	uint8_t pflag;
2540 
2541 	bootstart = bootmem = virtual_avail;
2542 
2543 	rw_init(&uma_rwlock, "UMA lock");
2544 	sx_init(&uma_reclaim_lock, "umareclaim");
2545 
2546 	ksize = sizeof(struct uma_keg) +
2547 	    (sizeof(struct uma_domain) * vm_ndomains);
2548 	ksize = roundup(ksize, UMA_SUPER_ALIGN);
2549 	zsize = sizeof(struct uma_zone) +
2550 	    (sizeof(struct uma_cache) * (mp_maxid + 1)) +
2551 	    (sizeof(struct uma_zone_domain) * vm_ndomains);
2552 	zsize = roundup(zsize, UMA_SUPER_ALIGN);
2553 
2554 	/* Allocate the zone of zones, zone of kegs, and zone of zones keg. */
2555 	size = (zsize * 2) + ksize;
2556 	m = (uintptr_t)startup_alloc(NULL, size, 0, &pflag, M_NOWAIT | M_ZERO);
2557 	zones = (uma_zone_t)m;
2558 	m += zsize;
2559 	kegs = (uma_zone_t)m;
2560 	m += zsize;
2561 	masterkeg = (uma_keg_t)m;
2562 
2563 	/* "manually" create the initial zone */
2564 	memset(&args, 0, sizeof(args));
2565 	args.name = "UMA Kegs";
2566 	args.size = ksize;
2567 	args.ctor = keg_ctor;
2568 	args.dtor = keg_dtor;
2569 	args.uminit = zero_init;
2570 	args.fini = NULL;
2571 	args.keg = masterkeg;
2572 	args.align = UMA_SUPER_ALIGN - 1;
2573 	args.flags = UMA_ZFLAG_INTERNAL;
2574 	zone_ctor(kegs, zsize, &args, M_WAITOK);
2575 
2576 	args.name = "UMA Zones";
2577 	args.size = zsize;
2578 	args.ctor = zone_ctor;
2579 	args.dtor = zone_dtor;
2580 	args.uminit = zero_init;
2581 	args.fini = NULL;
2582 	args.keg = NULL;
2583 	args.align = UMA_SUPER_ALIGN - 1;
2584 	args.flags = UMA_ZFLAG_INTERNAL;
2585 	zone_ctor(zones, zsize, &args, M_WAITOK);
2586 
2587 	/* Now make zones for slab headers */
2588 	slabzones[0] = uma_zcreate("UMA Slabs 0", SLABZONE0_SIZE,
2589 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2590 	slabzones[1] = uma_zcreate("UMA Slabs 1", SLABZONE1_SIZE,
2591 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2592 
2593 	hashzone = uma_zcreate("UMA Hash",
2594 	    sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
2595 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2596 
2597 	bucket_init();
2598 }
2599 
2600 #ifndef UMA_MD_SMALL_ALLOC
2601 extern void vm_radix_reserve_kva(void);
2602 #endif
2603 
2604 /*
2605  * Advertise the availability of normal kva allocations and switch to
2606  * the default back-end allocator.  Marks the KVA we consumed on startup
2607  * as used in the map.
2608  */
2609 void
2610 uma_startup2(void)
2611 {
2612 
2613 	if (!PMAP_HAS_DMAP) {
2614 		vm_map_lock(kernel_map);
2615 		(void)vm_map_insert(kernel_map, NULL, 0, bootstart, bootmem,
2616 		    VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT);
2617 		vm_map_unlock(kernel_map);
2618 	}
2619 
2620 #ifndef UMA_MD_SMALL_ALLOC
2621 	/* Set up radix zone to use noobj_alloc. */
2622 	vm_radix_reserve_kva();
2623 #endif
2624 
2625 	booted = BOOT_KVA;
2626 	zone_foreach_unlocked(zone_kva_available, NULL);
2627 	bucket_enable();
2628 }
2629 
2630 /*
2631  * Finish our initialization steps.
2632  */
2633 static void
2634 uma_startup3(void)
2635 {
2636 
2637 #ifdef INVARIANTS
2638 	TUNABLE_INT_FETCH("vm.debug.divisor", &dbg_divisor);
2639 	uma_dbg_cnt = counter_u64_alloc(M_WAITOK);
2640 	uma_skip_cnt = counter_u64_alloc(M_WAITOK);
2641 #endif
2642 	zone_foreach_unlocked(zone_alloc_counters, NULL);
2643 	zone_foreach_unlocked(zone_alloc_sysctl, NULL);
2644 	callout_init(&uma_callout, 1);
2645 	callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
2646 	booted = BOOT_RUNNING;
2647 
2648 	EVENTHANDLER_REGISTER(shutdown_post_sync, uma_shutdown, NULL,
2649 	    EVENTHANDLER_PRI_FIRST);
2650 }
2651 
2652 static void
2653 uma_shutdown(void)
2654 {
2655 
2656 	booted = BOOT_SHUTDOWN;
2657 }
2658 
2659 static uma_keg_t
2660 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
2661 		int align, uint32_t flags)
2662 {
2663 	struct uma_kctor_args args;
2664 
2665 	args.size = size;
2666 	args.uminit = uminit;
2667 	args.fini = fini;
2668 	args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
2669 	args.flags = flags;
2670 	args.zone = zone;
2671 	return (zone_alloc_item(kegs, &args, UMA_ANYDOMAIN, M_WAITOK));
2672 }
2673 
2674 /* Public functions */
2675 /* See uma.h */
2676 void
2677 uma_set_align(int align)
2678 {
2679 
2680 	if (align != UMA_ALIGN_CACHE)
2681 		uma_align_cache = align;
2682 }
2683 
2684 /* See uma.h */
2685 uma_zone_t
2686 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
2687 		uma_init uminit, uma_fini fini, int align, uint32_t flags)
2688 
2689 {
2690 	struct uma_zctor_args args;
2691 	uma_zone_t res;
2692 
2693 	KASSERT(powerof2(align + 1), ("invalid zone alignment %d for \"%s\"",
2694 	    align, name));
2695 
2696 	/* This stuff is essential for the zone ctor */
2697 	memset(&args, 0, sizeof(args));
2698 	args.name = name;
2699 	args.size = size;
2700 	args.ctor = ctor;
2701 	args.dtor = dtor;
2702 	args.uminit = uminit;
2703 	args.fini = fini;
2704 #ifdef  INVARIANTS
2705 	/*
2706 	 * Inject procedures which check for memory use after free if we are
2707 	 * allowed to scramble the memory while it is not allocated.  This
2708 	 * requires that: UMA is actually able to access the memory, no init
2709 	 * or fini procedures, no dependency on the initial value of the
2710 	 * memory, and no (legitimate) use of the memory after free.  Note,
2711 	 * the ctor and dtor do not need to be empty.
2712 	 */
2713 	if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOTOUCH |
2714 	    UMA_ZONE_NOFREE))) && uminit == NULL && fini == NULL) {
2715 		args.uminit = trash_init;
2716 		args.fini = trash_fini;
2717 	}
2718 #endif
2719 	args.align = align;
2720 	args.flags = flags;
2721 	args.keg = NULL;
2722 
2723 	sx_slock(&uma_reclaim_lock);
2724 	res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
2725 	sx_sunlock(&uma_reclaim_lock);
2726 
2727 	return (res);
2728 }
2729 
2730 /* See uma.h */
2731 uma_zone_t
2732 uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor,
2733 		    uma_init zinit, uma_fini zfini, uma_zone_t master)
2734 {
2735 	struct uma_zctor_args args;
2736 	uma_keg_t keg;
2737 	uma_zone_t res;
2738 
2739 	keg = master->uz_keg;
2740 	memset(&args, 0, sizeof(args));
2741 	args.name = name;
2742 	args.size = keg->uk_size;
2743 	args.ctor = ctor;
2744 	args.dtor = dtor;
2745 	args.uminit = zinit;
2746 	args.fini = zfini;
2747 	args.align = keg->uk_align;
2748 	args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
2749 	args.keg = keg;
2750 
2751 	sx_slock(&uma_reclaim_lock);
2752 	res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
2753 	sx_sunlock(&uma_reclaim_lock);
2754 
2755 	return (res);
2756 }
2757 
2758 /* See uma.h */
2759 uma_zone_t
2760 uma_zcache_create(char *name, int size, uma_ctor ctor, uma_dtor dtor,
2761 		    uma_init zinit, uma_fini zfini, uma_import zimport,
2762 		    uma_release zrelease, void *arg, int flags)
2763 {
2764 	struct uma_zctor_args args;
2765 
2766 	memset(&args, 0, sizeof(args));
2767 	args.name = name;
2768 	args.size = size;
2769 	args.ctor = ctor;
2770 	args.dtor = dtor;
2771 	args.uminit = zinit;
2772 	args.fini = zfini;
2773 	args.import = zimport;
2774 	args.release = zrelease;
2775 	args.arg = arg;
2776 	args.align = 0;
2777 	args.flags = flags | UMA_ZFLAG_CACHE;
2778 
2779 	return (zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK));
2780 }
2781 
2782 /* See uma.h */
2783 void
2784 uma_zdestroy(uma_zone_t zone)
2785 {
2786 
2787 	/*
2788 	 * Large slabs are expensive to reclaim, so don't bother doing
2789 	 * unnecessary work if we're shutting down.
2790 	 */
2791 	if (booted == BOOT_SHUTDOWN &&
2792 	    zone->uz_fini == NULL && zone->uz_release == zone_release)
2793 		return;
2794 	sx_slock(&uma_reclaim_lock);
2795 	zone_free_item(zones, zone, NULL, SKIP_NONE);
2796 	sx_sunlock(&uma_reclaim_lock);
2797 }
2798 
2799 void
2800 uma_zwait(uma_zone_t zone)
2801 {
2802 	void *item;
2803 
2804 	item = uma_zalloc_arg(zone, NULL, M_WAITOK);
2805 	uma_zfree(zone, item);
2806 }
2807 
2808 void *
2809 uma_zalloc_pcpu_arg(uma_zone_t zone, void *udata, int flags)
2810 {
2811 	void *item;
2812 #ifdef SMP
2813 	int i;
2814 
2815 	MPASS(zone->uz_flags & UMA_ZONE_PCPU);
2816 #endif
2817 	item = uma_zalloc_arg(zone, udata, flags & ~M_ZERO);
2818 	if (item != NULL && (flags & M_ZERO)) {
2819 #ifdef SMP
2820 		for (i = 0; i <= mp_maxid; i++)
2821 			bzero(zpcpu_get_cpu(item, i), zone->uz_size);
2822 #else
2823 		bzero(item, zone->uz_size);
2824 #endif
2825 	}
2826 	return (item);
2827 }
2828 
2829 /*
2830  * A stub while both regular and pcpu cases are identical.
2831  */
2832 void
2833 uma_zfree_pcpu_arg(uma_zone_t zone, void *item, void *udata)
2834 {
2835 
2836 #ifdef SMP
2837 	MPASS(zone->uz_flags & UMA_ZONE_PCPU);
2838 #endif
2839 	uma_zfree_arg(zone, item, udata);
2840 }
2841 
2842 #ifdef INVARIANTS
2843 #define	UMA_ALWAYS_CTORDTOR	1
2844 #else
2845 #define	UMA_ALWAYS_CTORDTOR	0
2846 #endif
2847 
2848 static void *
2849 item_ctor(uma_zone_t zone, int size, void *udata, int flags, void *item)
2850 {
2851 #ifdef INVARIANTS
2852 	bool skipdbg;
2853 
2854 	skipdbg = uma_dbg_zskip(zone, item);
2855 	if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 &&
2856 	    zone->uz_ctor != trash_ctor)
2857 		trash_ctor(item, size, udata, flags);
2858 #endif
2859 	if (__predict_false(zone->uz_ctor != NULL) &&
2860 	    zone->uz_ctor(item, size, udata, flags) != 0) {
2861 		counter_u64_add(zone->uz_fails, 1);
2862 		zone_free_item(zone, item, udata, SKIP_DTOR | SKIP_CNT);
2863 		return (NULL);
2864 	}
2865 #ifdef INVARIANTS
2866 	if (!skipdbg)
2867 		uma_dbg_alloc(zone, NULL, item);
2868 #endif
2869 	if (flags & M_ZERO)
2870 		bzero(item, size);
2871 
2872 	return (item);
2873 }
2874 
2875 static inline void
2876 item_dtor(uma_zone_t zone, void *item, int size, void *udata,
2877     enum zfreeskip skip)
2878 {
2879 #ifdef INVARIANTS
2880 	bool skipdbg;
2881 
2882 	skipdbg = uma_dbg_zskip(zone, item);
2883 	if (skip == SKIP_NONE && !skipdbg) {
2884 		if ((zone->uz_flags & UMA_ZONE_MALLOC) != 0)
2885 			uma_dbg_free(zone, udata, item);
2886 		else
2887 			uma_dbg_free(zone, NULL, item);
2888 	}
2889 #endif
2890 	if (__predict_true(skip < SKIP_DTOR)) {
2891 		if (zone->uz_dtor != NULL)
2892 			zone->uz_dtor(item, size, udata);
2893 #ifdef INVARIANTS
2894 		if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 &&
2895 		    zone->uz_dtor != trash_dtor)
2896 			trash_dtor(item, size, udata);
2897 #endif
2898 	}
2899 }
2900 
2901 /* See uma.h */
2902 void *
2903 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
2904 {
2905 	uma_cache_bucket_t bucket;
2906 	uma_cache_t cache;
2907 	void *item;
2908 	int domain, size, uz_flags;
2909 
2910 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
2911 	random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
2912 
2913 	/* This is the fast path allocation */
2914 	CTR3(KTR_UMA, "uma_zalloc_arg zone %s(%p) flags %d", zone->uz_name,
2915 	    zone, flags);
2916 
2917 #ifdef WITNESS
2918 	if (flags & M_WAITOK) {
2919 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
2920 		    "uma_zalloc_arg: zone \"%s\"", zone->uz_name);
2921 	}
2922 #endif
2923 
2924 #ifdef INVARIANTS
2925 	KASSERT((flags & M_EXEC) == 0, ("uma_zalloc_arg: called with M_EXEC"));
2926 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
2927 	    ("uma_zalloc_arg: called with spinlock or critical section held"));
2928 	if (zone->uz_flags & UMA_ZONE_PCPU)
2929 		KASSERT((flags & M_ZERO) == 0, ("allocating from a pcpu zone "
2930 		    "with M_ZERO passed"));
2931 #endif
2932 
2933 #ifdef DEBUG_MEMGUARD
2934 	if (memguard_cmp_zone(zone)) {
2935 		item = memguard_alloc(zone->uz_size, flags);
2936 		if (item != NULL) {
2937 			if (zone->uz_init != NULL &&
2938 			    zone->uz_init(item, zone->uz_size, flags) != 0)
2939 				return (NULL);
2940 			if (zone->uz_ctor != NULL &&
2941 			    zone->uz_ctor(item, zone->uz_size, udata,
2942 			    flags) != 0) {
2943 				counter_u64_add(zone->uz_fails, 1);
2944 			    	zone->uz_fini(item, zone->uz_size);
2945 				return (NULL);
2946 			}
2947 			return (item);
2948 		}
2949 		/* This is unfortunate but should not be fatal. */
2950 	}
2951 #endif
2952 	/*
2953 	 * If possible, allocate from the per-CPU cache.  There are two
2954 	 * requirements for safe access to the per-CPU cache: (1) the thread
2955 	 * accessing the cache must not be preempted or yield during access,
2956 	 * and (2) the thread must not migrate CPUs without switching which
2957 	 * cache it accesses.  We rely on a critical section to prevent
2958 	 * preemption and migration.  We release the critical section in
2959 	 * order to acquire the zone mutex if we are unable to allocate from
2960 	 * the current cache; when we re-acquire the critical section, we
2961 	 * must detect and handle migration if it has occurred.
2962 	 */
2963 	critical_enter();
2964 	do {
2965 		cache = &zone->uz_cpu[curcpu];
2966 		bucket = &cache->uc_allocbucket;
2967 		size = cache_uz_size(cache);
2968 		uz_flags = cache_uz_flags(cache);
2969 		if (__predict_true(bucket->ucb_cnt != 0)) {
2970 			item = cache_bucket_pop(cache, bucket);
2971 			critical_exit();
2972 			if (__predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0 ||
2973 			    UMA_ALWAYS_CTORDTOR))
2974 				return (item_ctor(zone, size, udata, flags, item));
2975 			if (flags & M_ZERO)
2976 				bzero(item, size);
2977 			return (item);
2978 		}
2979 	} while (cache_alloc(zone, cache, udata, flags));
2980 	critical_exit();
2981 
2982 	/*
2983 	 * We can not get a bucket so try to return a single item.
2984 	 */
2985 	if (uz_flags & UMA_ZONE_FIRSTTOUCH)
2986 		domain = PCPU_GET(domain);
2987 	else
2988 		domain = UMA_ANYDOMAIN;
2989 	return (zone_alloc_item(zone, udata, domain, flags));
2990 }
2991 
2992 /*
2993  * Replenish an alloc bucket and possibly restore an old one.  Called in
2994  * a critical section.  Returns in a critical section.
2995  *
2996  * A false return value indicates an allocation failure.
2997  * A true return value indicates success and the caller should retry.
2998  */
2999 static __noinline bool
3000 cache_alloc(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3001 {
3002 	uma_zone_domain_t zdom;
3003 	uma_bucket_t bucket;
3004 	int domain;
3005 	bool lockfail;
3006 
3007 	CRITICAL_ASSERT(curthread);
3008 
3009 	/*
3010 	 * If we have run out of items in our alloc bucket see
3011 	 * if we can switch with the free bucket.
3012 	 */
3013 	if (cache->uc_freebucket.ucb_cnt != 0) {
3014 		cache_bucket_swap(&cache->uc_freebucket, &cache->uc_allocbucket);
3015 		return (true);
3016 	}
3017 
3018 	/*
3019 	 * Discard any empty allocation bucket while we hold no locks.
3020 	 */
3021 	bucket = cache_bucket_unload_alloc(cache);
3022 	critical_exit();
3023 	if (bucket != NULL)
3024 		bucket_free(zone, bucket, udata);
3025 
3026 	/* Short-circuit for zones without buckets and low memory. */
3027 	if (zone->uz_bucket_size == 0 || bucketdisable) {
3028 		critical_enter();
3029 		return (false);
3030 	}
3031 
3032 	/*
3033 	 * Attempt to retrieve the item from the per-CPU cache has failed, so
3034 	 * we must go back to the zone.  This requires the zone lock, so we
3035 	 * must drop the critical section, then re-acquire it when we go back
3036 	 * to the cache.  Since the critical section is released, we may be
3037 	 * preempted or migrate.  As such, make sure not to maintain any
3038 	 * thread-local state specific to the cache from prior to releasing
3039 	 * the critical section.
3040 	 */
3041 	lockfail = 0;
3042 	if (ZONE_TRYLOCK(zone) == 0) {
3043 		/* Record contention to size the buckets. */
3044 		ZONE_LOCK(zone);
3045 		lockfail = 1;
3046 	}
3047 
3048 	/* See if we lost the race to fill the cache. */
3049 	critical_enter();
3050 	cache = &zone->uz_cpu[curcpu];
3051 	if (cache->uc_allocbucket.ucb_bucket != NULL) {
3052 		ZONE_UNLOCK(zone);
3053 		return (true);
3054 	}
3055 
3056 	/*
3057 	 * Check the zone's cache of buckets.
3058 	 */
3059 	if (zone->uz_flags & UMA_ZONE_FIRSTTOUCH) {
3060 		domain = PCPU_GET(domain);
3061 		zdom = &zone->uz_domain[domain];
3062 	} else {
3063 		domain = UMA_ANYDOMAIN;
3064 		zdom = &zone->uz_domain[0];
3065 	}
3066 
3067 	if ((bucket = zone_fetch_bucket(zone, zdom)) != NULL) {
3068 		ZONE_UNLOCK(zone);
3069 		KASSERT(bucket->ub_cnt != 0,
3070 		    ("uma_zalloc_arg: Returning an empty bucket."));
3071 		cache_bucket_load_alloc(cache, bucket);
3072 		return (true);
3073 	}
3074 	/* We are no longer associated with this CPU. */
3075 	critical_exit();
3076 
3077 	/*
3078 	 * We bump the uz count when the cache size is insufficient to
3079 	 * handle the working set.
3080 	 */
3081 	if (lockfail && zone->uz_bucket_size < zone->uz_bucket_size_max)
3082 		zone->uz_bucket_size++;
3083 	ZONE_UNLOCK(zone);
3084 
3085 	/*
3086 	 * Fill a bucket and attempt to use it as the alloc bucket.
3087 	 */
3088 	bucket = zone_alloc_bucket(zone, udata, domain, flags);
3089 	CTR3(KTR_UMA, "uma_zalloc: zone %s(%p) bucket zone returned %p",
3090 	    zone->uz_name, zone, bucket);
3091 	if (bucket == NULL) {
3092 		critical_enter();
3093 		return (false);
3094 	}
3095 
3096 	/*
3097 	 * See if we lost the race or were migrated.  Cache the
3098 	 * initialized bucket to make this less likely or claim
3099 	 * the memory directly.
3100 	 */
3101 	ZONE_LOCK(zone);
3102 	critical_enter();
3103 	cache = &zone->uz_cpu[curcpu];
3104 	if (cache->uc_allocbucket.ucb_bucket == NULL &&
3105 	    ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) == 0 ||
3106 	    domain == PCPU_GET(domain))) {
3107 		cache_bucket_load_alloc(cache, bucket);
3108 		zdom->uzd_imax += bucket->ub_cnt;
3109 	} else if (zone->uz_bkt_count >= zone->uz_bkt_max) {
3110 		critical_exit();
3111 		ZONE_UNLOCK(zone);
3112 		bucket_drain(zone, bucket);
3113 		bucket_free(zone, bucket, udata);
3114 		critical_enter();
3115 		return (true);
3116 	} else
3117 		zone_put_bucket(zone, zdom, bucket, false);
3118 	ZONE_UNLOCK(zone);
3119 	return (true);
3120 }
3121 
3122 void *
3123 uma_zalloc_domain(uma_zone_t zone, void *udata, int domain, int flags)
3124 {
3125 
3126 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3127 	random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3128 
3129 	/* This is the fast path allocation */
3130 	CTR4(KTR_UMA, "uma_zalloc_domain zone %s(%p) domain %d flags %d",
3131 	    zone->uz_name, zone, domain, flags);
3132 
3133 	if (flags & M_WAITOK) {
3134 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
3135 		    "uma_zalloc_domain: zone \"%s\"", zone->uz_name);
3136 	}
3137 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3138 	    ("uma_zalloc_domain: called with spinlock or critical section held"));
3139 
3140 	return (zone_alloc_item(zone, udata, domain, flags));
3141 }
3142 
3143 /*
3144  * Find a slab with some space.  Prefer slabs that are partially used over those
3145  * that are totally full.  This helps to reduce fragmentation.
3146  *
3147  * If 'rr' is 1, search all domains starting from 'domain'.  Otherwise check
3148  * only 'domain'.
3149  */
3150 static uma_slab_t
3151 keg_first_slab(uma_keg_t keg, int domain, bool rr)
3152 {
3153 	uma_domain_t dom;
3154 	uma_slab_t slab;
3155 	int start;
3156 
3157 	KASSERT(domain >= 0 && domain < vm_ndomains,
3158 	    ("keg_first_slab: domain %d out of range", domain));
3159 	KEG_LOCK_ASSERT(keg, domain);
3160 
3161 	slab = NULL;
3162 	start = domain;
3163 	do {
3164 		dom = &keg->uk_domain[domain];
3165 		if (!LIST_EMPTY(&dom->ud_part_slab))
3166 			return (LIST_FIRST(&dom->ud_part_slab));
3167 		if (!LIST_EMPTY(&dom->ud_free_slab)) {
3168 			slab = LIST_FIRST(&dom->ud_free_slab);
3169 			LIST_REMOVE(slab, us_link);
3170 			LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
3171 			return (slab);
3172 		}
3173 		if (rr)
3174 			domain = (domain + 1) % vm_ndomains;
3175 	} while (domain != start);
3176 
3177 	return (NULL);
3178 }
3179 
3180 /*
3181  * Fetch an existing slab from a free or partial list.  Returns with the
3182  * keg domain lock held if a slab was found or unlocked if not.
3183  */
3184 static uma_slab_t
3185 keg_fetch_free_slab(uma_keg_t keg, int domain, bool rr, int flags)
3186 {
3187 	uma_slab_t slab;
3188 	uint32_t reserve;
3189 
3190 	/* HASH has a single free list. */
3191 	if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
3192 		domain = 0;
3193 
3194 	KEG_LOCK(keg, domain);
3195 	reserve = (flags & M_USE_RESERVE) != 0 ? 0 : keg->uk_reserve;
3196 	if (keg->uk_domain[domain].ud_free <= reserve ||
3197 	    (slab = keg_first_slab(keg, domain, rr)) == NULL) {
3198 		KEG_UNLOCK(keg, domain);
3199 		return (NULL);
3200 	}
3201 	return (slab);
3202 }
3203 
3204 static uma_slab_t
3205 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int rdomain, const int flags)
3206 {
3207 	struct vm_domainset_iter di;
3208 	uma_slab_t slab;
3209 	int aflags, domain;
3210 	bool rr;
3211 
3212 restart:
3213 	/*
3214 	 * Use the keg's policy if upper layers haven't already specified a
3215 	 * domain (as happens with first-touch zones).
3216 	 *
3217 	 * To avoid races we run the iterator with the keg lock held, but that
3218 	 * means that we cannot allow the vm_domainset layer to sleep.  Thus,
3219 	 * clear M_WAITOK and handle low memory conditions locally.
3220 	 */
3221 	rr = rdomain == UMA_ANYDOMAIN;
3222 	if (rr) {
3223 		aflags = (flags & ~M_WAITOK) | M_NOWAIT;
3224 		vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
3225 		    &aflags);
3226 	} else {
3227 		aflags = flags;
3228 		domain = rdomain;
3229 	}
3230 
3231 	for (;;) {
3232 		slab = keg_fetch_free_slab(keg, domain, rr, flags);
3233 		if (slab != NULL)
3234 			return (slab);
3235 
3236 		/*
3237 		 * M_NOVM means don't ask at all!
3238 		 */
3239 		if (flags & M_NOVM)
3240 			break;
3241 
3242 		slab = keg_alloc_slab(keg, zone, domain, flags, aflags);
3243 		if (slab != NULL)
3244 			return (slab);
3245 		if (!rr && (flags & M_WAITOK) == 0)
3246 			break;
3247 		if (rr && vm_domainset_iter_policy(&di, &domain) != 0) {
3248 			if ((flags & M_WAITOK) != 0) {
3249 				vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask);
3250 				goto restart;
3251 			}
3252 			break;
3253 		}
3254 	}
3255 
3256 	/*
3257 	 * We might not have been able to get a slab but another cpu
3258 	 * could have while we were unlocked.  Check again before we
3259 	 * fail.
3260 	 */
3261 	if ((slab = keg_fetch_free_slab(keg, domain, rr, flags)) != NULL)
3262 		return (slab);
3263 
3264 	return (NULL);
3265 }
3266 
3267 static void *
3268 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
3269 {
3270 	uma_domain_t dom;
3271 	void *item;
3272 	int freei;
3273 
3274 	KEG_LOCK_ASSERT(keg, slab->us_domain);
3275 
3276 	dom = &keg->uk_domain[slab->us_domain];
3277 	freei = BIT_FFS(keg->uk_ipers, &slab->us_free) - 1;
3278 	BIT_CLR(keg->uk_ipers, freei, &slab->us_free);
3279 	item = slab_item(slab, keg, freei);
3280 	slab->us_freecount--;
3281 	dom->ud_free--;
3282 
3283 	/* Move this slab to the full list */
3284 	if (slab->us_freecount == 0) {
3285 		LIST_REMOVE(slab, us_link);
3286 		LIST_INSERT_HEAD(&dom->ud_full_slab, slab, us_link);
3287 	}
3288 
3289 	return (item);
3290 }
3291 
3292 static int
3293 zone_import(void *arg, void **bucket, int max, int domain, int flags)
3294 {
3295 	uma_domain_t dom;
3296 	uma_zone_t zone;
3297 	uma_slab_t slab;
3298 	uma_keg_t keg;
3299 #ifdef NUMA
3300 	int stripe;
3301 #endif
3302 	int i;
3303 
3304 	zone = arg;
3305 	slab = NULL;
3306 	keg = zone->uz_keg;
3307 	/* Try to keep the buckets totally full */
3308 	for (i = 0; i < max; ) {
3309 		if ((slab = keg_fetch_slab(keg, zone, domain, flags)) == NULL)
3310 			break;
3311 #ifdef NUMA
3312 		stripe = howmany(max, vm_ndomains);
3313 #endif
3314 		dom = &keg->uk_domain[slab->us_domain];
3315 		while (slab->us_freecount && i < max) {
3316 			bucket[i++] = slab_alloc_item(keg, slab);
3317 			if (dom->ud_free <= keg->uk_reserve)
3318 				break;
3319 #ifdef NUMA
3320 			/*
3321 			 * If the zone is striped we pick a new slab for every
3322 			 * N allocations.  Eliminating this conditional will
3323 			 * instead pick a new domain for each bucket rather
3324 			 * than stripe within each bucket.  The current option
3325 			 * produces more fragmentation and requires more cpu
3326 			 * time but yields better distribution.
3327 			 */
3328 			if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0 &&
3329 			    vm_ndomains > 1 && --stripe == 0)
3330 				break;
3331 #endif
3332 		}
3333 		KEG_UNLOCK(keg, slab->us_domain);
3334 		/* Don't block if we allocated any successfully. */
3335 		flags &= ~M_WAITOK;
3336 		flags |= M_NOWAIT;
3337 	}
3338 
3339 	return i;
3340 }
3341 
3342 static int
3343 zone_alloc_limit_hard(uma_zone_t zone, int count, int flags)
3344 {
3345 	uint64_t old, new, total, max;
3346 
3347 	/*
3348 	 * The hard case.  We're going to sleep because there were existing
3349 	 * sleepers or because we ran out of items.  This routine enforces
3350 	 * fairness by keeping fifo order.
3351 	 *
3352 	 * First release our ill gotten gains and make some noise.
3353 	 */
3354 	for (;;) {
3355 		zone_free_limit(zone, count);
3356 		zone_log_warning(zone);
3357 		zone_maxaction(zone);
3358 		if (flags & M_NOWAIT)
3359 			return (0);
3360 
3361 		/*
3362 		 * We need to allocate an item or set ourself as a sleeper
3363 		 * while the sleepq lock is held to avoid wakeup races.  This
3364 		 * is essentially a home rolled semaphore.
3365 		 */
3366 		sleepq_lock(&zone->uz_max_items);
3367 		old = zone->uz_items;
3368 		do {
3369 			MPASS(UZ_ITEMS_SLEEPERS(old) < UZ_ITEMS_SLEEPERS_MAX);
3370 			/* Cache the max since we will evaluate twice. */
3371 			max = zone->uz_max_items;
3372 			if (UZ_ITEMS_SLEEPERS(old) != 0 ||
3373 			    UZ_ITEMS_COUNT(old) >= max)
3374 				new = old + UZ_ITEMS_SLEEPER;
3375 			else
3376 				new = old + MIN(count, max - old);
3377 		} while (atomic_fcmpset_64(&zone->uz_items, &old, new) == 0);
3378 
3379 		/* We may have successfully allocated under the sleepq lock. */
3380 		if (UZ_ITEMS_SLEEPERS(new) == 0) {
3381 			sleepq_release(&zone->uz_max_items);
3382 			return (new - old);
3383 		}
3384 
3385 		/*
3386 		 * This is in a different cacheline from uz_items so that we
3387 		 * don't constantly invalidate the fastpath cacheline when we
3388 		 * adjust item counts.  This could be limited to toggling on
3389 		 * transitions.
3390 		 */
3391 		atomic_add_32(&zone->uz_sleepers, 1);
3392 		atomic_add_64(&zone->uz_sleeps, 1);
3393 
3394 		/*
3395 		 * We have added ourselves as a sleeper.  The sleepq lock
3396 		 * protects us from wakeup races.  Sleep now and then retry.
3397 		 */
3398 		sleepq_add(&zone->uz_max_items, NULL, "zonelimit", 0, 0);
3399 		sleepq_wait(&zone->uz_max_items, PVM);
3400 
3401 		/*
3402 		 * After wakeup, remove ourselves as a sleeper and try
3403 		 * again.  We no longer have the sleepq lock for protection.
3404 		 *
3405 		 * Subract ourselves as a sleeper while attempting to add
3406 		 * our count.
3407 		 */
3408 		atomic_subtract_32(&zone->uz_sleepers, 1);
3409 		old = atomic_fetchadd_64(&zone->uz_items,
3410 		    -(UZ_ITEMS_SLEEPER - count));
3411 		/* We're no longer a sleeper. */
3412 		old -= UZ_ITEMS_SLEEPER;
3413 
3414 		/*
3415 		 * If we're still at the limit, restart.  Notably do not
3416 		 * block on other sleepers.  Cache the max value to protect
3417 		 * against changes via sysctl.
3418 		 */
3419 		total = UZ_ITEMS_COUNT(old);
3420 		max = zone->uz_max_items;
3421 		if (total >= max)
3422 			continue;
3423 		/* Truncate if necessary, otherwise wake other sleepers. */
3424 		if (total + count > max) {
3425 			zone_free_limit(zone, total + count - max);
3426 			count = max - total;
3427 		} else if (total + count < max && UZ_ITEMS_SLEEPERS(old) != 0)
3428 			wakeup_one(&zone->uz_max_items);
3429 
3430 		return (count);
3431 	}
3432 }
3433 
3434 /*
3435  * Allocate 'count' items from our max_items limit.  Returns the number
3436  * available.  If M_NOWAIT is not specified it will sleep until at least
3437  * one item can be allocated.
3438  */
3439 static int
3440 zone_alloc_limit(uma_zone_t zone, int count, int flags)
3441 {
3442 	uint64_t old;
3443 	uint64_t max;
3444 
3445 	max = zone->uz_max_items;
3446 	MPASS(max > 0);
3447 
3448 	/*
3449 	 * We expect normal allocations to succeed with a simple
3450 	 * fetchadd.
3451 	 */
3452 	old = atomic_fetchadd_64(&zone->uz_items, count);
3453 	if (__predict_true(old + count <= max))
3454 		return (count);
3455 
3456 	/*
3457 	 * If we had some items and no sleepers just return the
3458 	 * truncated value.  We have to release the excess space
3459 	 * though because that may wake sleepers who weren't woken
3460 	 * because we were temporarily over the limit.
3461 	 */
3462 	if (old < max) {
3463 		zone_free_limit(zone, (old + count) - max);
3464 		return (max - old);
3465 	}
3466 	return (zone_alloc_limit_hard(zone, count, flags));
3467 }
3468 
3469 /*
3470  * Free a number of items back to the limit.
3471  */
3472 static void
3473 zone_free_limit(uma_zone_t zone, int count)
3474 {
3475 	uint64_t old;
3476 
3477 	MPASS(count > 0);
3478 
3479 	/*
3480 	 * In the common case we either have no sleepers or
3481 	 * are still over the limit and can just return.
3482 	 */
3483 	old = atomic_fetchadd_64(&zone->uz_items, -count);
3484 	if (__predict_true(UZ_ITEMS_SLEEPERS(old) == 0 ||
3485 	   UZ_ITEMS_COUNT(old) - count >= zone->uz_max_items))
3486 		return;
3487 
3488 	/*
3489 	 * Moderate the rate of wakeups.  Sleepers will continue
3490 	 * to generate wakeups if necessary.
3491 	 */
3492 	wakeup_one(&zone->uz_max_items);
3493 }
3494 
3495 static uma_bucket_t
3496 zone_alloc_bucket(uma_zone_t zone, void *udata, int domain, int flags)
3497 {
3498 	uma_bucket_t bucket;
3499 	int maxbucket, cnt;
3500 
3501 	CTR3(KTR_UMA, "zone_alloc_bucket zone %s(%p) domain %d", zone->uz_name,
3502 	    zone, domain);
3503 
3504 	/* Avoid allocs targeting empty domains. */
3505 	if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
3506 		domain = UMA_ANYDOMAIN;
3507 
3508 	if (zone->uz_max_items > 0)
3509 		maxbucket = zone_alloc_limit(zone, zone->uz_bucket_size,
3510 		    M_NOWAIT);
3511 	else
3512 		maxbucket = zone->uz_bucket_size;
3513 	if (maxbucket == 0)
3514 		return (false);
3515 
3516 	/* Don't wait for buckets, preserve caller's NOVM setting. */
3517 	bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
3518 	if (bucket == NULL) {
3519 		cnt = 0;
3520 		goto out;
3521 	}
3522 
3523 	bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
3524 	    MIN(maxbucket, bucket->ub_entries), domain, flags);
3525 
3526 	/*
3527 	 * Initialize the memory if necessary.
3528 	 */
3529 	if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
3530 		int i;
3531 
3532 		for (i = 0; i < bucket->ub_cnt; i++)
3533 			if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size,
3534 			    flags) != 0)
3535 				break;
3536 		/*
3537 		 * If we couldn't initialize the whole bucket, put the
3538 		 * rest back onto the freelist.
3539 		 */
3540 		if (i != bucket->ub_cnt) {
3541 			zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
3542 			    bucket->ub_cnt - i);
3543 #ifdef INVARIANTS
3544 			bzero(&bucket->ub_bucket[i],
3545 			    sizeof(void *) * (bucket->ub_cnt - i));
3546 #endif
3547 			bucket->ub_cnt = i;
3548 		}
3549 	}
3550 
3551 	cnt = bucket->ub_cnt;
3552 	if (bucket->ub_cnt == 0) {
3553 		bucket_free(zone, bucket, udata);
3554 		counter_u64_add(zone->uz_fails, 1);
3555 		bucket = NULL;
3556 	}
3557 out:
3558 	if (zone->uz_max_items > 0 && cnt < maxbucket)
3559 		zone_free_limit(zone, maxbucket - cnt);
3560 
3561 	return (bucket);
3562 }
3563 
3564 /*
3565  * Allocates a single item from a zone.
3566  *
3567  * Arguments
3568  *	zone   The zone to alloc for.
3569  *	udata  The data to be passed to the constructor.
3570  *	domain The domain to allocate from or UMA_ANYDOMAIN.
3571  *	flags  M_WAITOK, M_NOWAIT, M_ZERO.
3572  *
3573  * Returns
3574  *	NULL if there is no memory and M_NOWAIT is set
3575  *	An item if successful
3576  */
3577 
3578 static void *
3579 zone_alloc_item(uma_zone_t zone, void *udata, int domain, int flags)
3580 {
3581 	void *item;
3582 
3583 	if (zone->uz_max_items > 0 && zone_alloc_limit(zone, 1, flags) == 0)
3584 		return (NULL);
3585 
3586 	/* Avoid allocs targeting empty domains. */
3587 	if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
3588 		domain = UMA_ANYDOMAIN;
3589 
3590 	if (zone->uz_import(zone->uz_arg, &item, 1, domain, flags) != 1)
3591 		goto fail_cnt;
3592 
3593 	/*
3594 	 * We have to call both the zone's init (not the keg's init)
3595 	 * and the zone's ctor.  This is because the item is going from
3596 	 * a keg slab directly to the user, and the user is expecting it
3597 	 * to be both zone-init'd as well as zone-ctor'd.
3598 	 */
3599 	if (zone->uz_init != NULL) {
3600 		if (zone->uz_init(item, zone->uz_size, flags) != 0) {
3601 			zone_free_item(zone, item, udata, SKIP_FINI | SKIP_CNT);
3602 			goto fail_cnt;
3603 		}
3604 	}
3605 	item = item_ctor(zone, zone->uz_size, udata, flags, item);
3606 	if (item == NULL)
3607 		goto fail;
3608 
3609 	counter_u64_add(zone->uz_allocs, 1);
3610 	CTR3(KTR_UMA, "zone_alloc_item item %p from %s(%p)", item,
3611 	    zone->uz_name, zone);
3612 
3613 	return (item);
3614 
3615 fail_cnt:
3616 	counter_u64_add(zone->uz_fails, 1);
3617 fail:
3618 	if (zone->uz_max_items > 0)
3619 		zone_free_limit(zone, 1);
3620 	CTR2(KTR_UMA, "zone_alloc_item failed from %s(%p)",
3621 	    zone->uz_name, zone);
3622 
3623 	return (NULL);
3624 }
3625 
3626 /* See uma.h */
3627 void
3628 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
3629 {
3630 	uma_cache_t cache;
3631 	uma_cache_bucket_t bucket;
3632 	int domain, itemdomain, uz_flags;
3633 
3634 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3635 	random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3636 
3637 	CTR2(KTR_UMA, "uma_zfree_arg zone %s(%p)", zone->uz_name, zone);
3638 
3639 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3640 	    ("uma_zfree_arg: called with spinlock or critical section held"));
3641 
3642         /* uma_zfree(..., NULL) does nothing, to match free(9). */
3643         if (item == NULL)
3644                 return;
3645 #ifdef DEBUG_MEMGUARD
3646 	if (is_memguard_addr(item)) {
3647 		if (zone->uz_dtor != NULL)
3648 			zone->uz_dtor(item, zone->uz_size, udata);
3649 		if (zone->uz_fini != NULL)
3650 			zone->uz_fini(item, zone->uz_size);
3651 		memguard_free(item);
3652 		return;
3653 	}
3654 #endif
3655 
3656 	/*
3657 	 * We are accessing the per-cpu cache without a critical section to
3658 	 * fetch size and flags.  This is acceptable, if we are preempted we
3659 	 * will simply read another cpu's line.
3660 	 */
3661 	cache = &zone->uz_cpu[curcpu];
3662 	uz_flags = cache_uz_flags(cache);
3663 	if (__predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0 ||
3664 	    UMA_ALWAYS_CTORDTOR))
3665 		item_dtor(zone, item, cache_uz_size(cache), udata, SKIP_NONE);
3666 
3667 	/*
3668 	 * The race here is acceptable.  If we miss it we'll just have to wait
3669 	 * a little longer for the limits to be reset.
3670 	 */
3671 	if (__predict_false(uz_flags & UMA_ZFLAG_LIMIT)) {
3672 		if (zone->uz_sleepers > 0)
3673 			goto zfree_item;
3674 	}
3675 
3676 	/*
3677 	 * If possible, free to the per-CPU cache.  There are two
3678 	 * requirements for safe access to the per-CPU cache: (1) the thread
3679 	 * accessing the cache must not be preempted or yield during access,
3680 	 * and (2) the thread must not migrate CPUs without switching which
3681 	 * cache it accesses.  We rely on a critical section to prevent
3682 	 * preemption and migration.  We release the critical section in
3683 	 * order to acquire the zone mutex if we are unable to free to the
3684 	 * current cache; when we re-acquire the critical section, we must
3685 	 * detect and handle migration if it has occurred.
3686 	 */
3687 	domain = itemdomain = 0;
3688 #ifdef NUMA
3689 	if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
3690 		itemdomain = _vm_phys_domain(pmap_kextract((vm_offset_t)item));
3691 #endif
3692 	critical_enter();
3693 	do {
3694 		cache = &zone->uz_cpu[curcpu];
3695 #ifdef NUMA
3696 		domain = PCPU_GET(domain);
3697 		if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
3698 		    domain != itemdomain) {
3699 			bucket = &cache->uc_crossbucket;
3700 		} else
3701 #endif
3702 		{
3703 			/*
3704 			 * Try to free into the allocbucket first to give LIFO
3705 			 * ordering for cache-hot datastructures.  Spill over
3706 			 * into the freebucket if necessary.  Alloc will swap
3707 			 * them if one runs dry.
3708 			 */
3709 			bucket = &cache->uc_allocbucket;
3710 			if (__predict_false(bucket->ucb_cnt >=
3711 			    bucket->ucb_entries))
3712 				bucket = &cache->uc_freebucket;
3713 		}
3714 		if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
3715 			cache_bucket_push(cache, bucket, item);
3716 			critical_exit();
3717 			return;
3718 		}
3719 	} while (cache_free(zone, cache, udata, item, itemdomain));
3720 	critical_exit();
3721 
3722 	/*
3723 	 * If nothing else caught this, we'll just do an internal free.
3724 	 */
3725 zfree_item:
3726 	zone_free_item(zone, item, udata, SKIP_DTOR);
3727 }
3728 
3729 #ifdef NUMA
3730 /*
3731  * sort crossdomain free buckets to domain correct buckets and cache
3732  * them.
3733  */
3734 static void
3735 zone_free_cross(uma_zone_t zone, uma_bucket_t bucket, void *udata)
3736 {
3737 	struct uma_bucketlist fullbuckets;
3738 	uma_zone_domain_t zdom;
3739 	uma_bucket_t b;
3740 	void *item;
3741 	int domain;
3742 
3743 	CTR3(KTR_UMA,
3744 	    "uma_zfree: zone %s(%p) draining cross bucket %p",
3745 	    zone->uz_name, zone, bucket);
3746 
3747 	TAILQ_INIT(&fullbuckets);
3748 
3749 	/*
3750 	 * To avoid having ndomain * ndomain buckets for sorting we have a
3751 	 * lock on the current crossfree bucket.  A full matrix with
3752 	 * per-domain locking could be used if necessary.
3753 	 */
3754 	ZONE_CROSS_LOCK(zone);
3755 	while (bucket->ub_cnt > 0) {
3756 		item = bucket->ub_bucket[bucket->ub_cnt - 1];
3757 		domain = _vm_phys_domain(pmap_kextract((vm_offset_t)item));
3758 		zdom = &zone->uz_domain[domain];
3759 		if (zdom->uzd_cross == NULL) {
3760 			zdom->uzd_cross = bucket_alloc(zone, udata, M_NOWAIT);
3761 			if (zdom->uzd_cross == NULL)
3762 				break;
3763 		}
3764 		zdom->uzd_cross->ub_bucket[zdom->uzd_cross->ub_cnt++] = item;
3765 		if (zdom->uzd_cross->ub_cnt == zdom->uzd_cross->ub_entries) {
3766 			TAILQ_INSERT_HEAD(&fullbuckets, zdom->uzd_cross,
3767 			    ub_link);
3768 			zdom->uzd_cross = NULL;
3769 		}
3770 		bucket->ub_cnt--;
3771 	}
3772 	ZONE_CROSS_UNLOCK(zone);
3773 	if (!TAILQ_EMPTY(&fullbuckets)) {
3774 		ZONE_LOCK(zone);
3775 		while ((b = TAILQ_FIRST(&fullbuckets)) != NULL) {
3776 			TAILQ_REMOVE(&fullbuckets, b, ub_link);
3777 			if (zone->uz_bkt_count >= zone->uz_bkt_max) {
3778 				ZONE_UNLOCK(zone);
3779 				bucket_drain(zone, b);
3780 				bucket_free(zone, b, udata);
3781 				ZONE_LOCK(zone);
3782 			} else {
3783 				domain = _vm_phys_domain(
3784 				    pmap_kextract(
3785 				    (vm_offset_t)b->ub_bucket[0]));
3786 				zdom = &zone->uz_domain[domain];
3787 				zone_put_bucket(zone, zdom, b, true);
3788 			}
3789 		}
3790 		ZONE_UNLOCK(zone);
3791 	}
3792 	if (bucket->ub_cnt != 0)
3793 		bucket_drain(zone, bucket);
3794 	bucket_free(zone, bucket, udata);
3795 }
3796 #endif
3797 
3798 static void
3799 zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata,
3800     int domain, int itemdomain)
3801 {
3802 	uma_zone_domain_t zdom;
3803 
3804 #ifdef NUMA
3805 	/*
3806 	 * Buckets coming from the wrong domain will be entirely for the
3807 	 * only other domain on two domain systems.  In this case we can
3808 	 * simply cache them.  Otherwise we need to sort them back to
3809 	 * correct domains.
3810 	 */
3811 	if (domain != itemdomain && vm_ndomains > 2) {
3812 		zone_free_cross(zone, bucket, udata);
3813 		return;
3814 	}
3815 #endif
3816 
3817 	/*
3818 	 * Attempt to save the bucket in the zone's domain bucket cache.
3819 	 *
3820 	 * We bump the uz count when the cache size is insufficient to
3821 	 * handle the working set.
3822 	 */
3823 	if (ZONE_TRYLOCK(zone) == 0) {
3824 		/* Record contention to size the buckets. */
3825 		ZONE_LOCK(zone);
3826 		if (zone->uz_bucket_size < zone->uz_bucket_size_max)
3827 			zone->uz_bucket_size++;
3828 	}
3829 
3830 	CTR3(KTR_UMA,
3831 	    "uma_zfree: zone %s(%p) putting bucket %p on free list",
3832 	    zone->uz_name, zone, bucket);
3833 	/* ub_cnt is pointing to the last free item */
3834 	KASSERT(bucket->ub_cnt == bucket->ub_entries,
3835 	    ("uma_zfree: Attempting to insert partial  bucket onto the full list.\n"));
3836 	if (zone->uz_bkt_count >= zone->uz_bkt_max) {
3837 		ZONE_UNLOCK(zone);
3838 		bucket_drain(zone, bucket);
3839 		bucket_free(zone, bucket, udata);
3840 	} else {
3841 		zdom = &zone->uz_domain[itemdomain];
3842 		zone_put_bucket(zone, zdom, bucket, true);
3843 		ZONE_UNLOCK(zone);
3844 	}
3845 }
3846 
3847 /*
3848  * Populate a free or cross bucket for the current cpu cache.  Free any
3849  * existing full bucket either to the zone cache or back to the slab layer.
3850  *
3851  * Enters and returns in a critical section.  false return indicates that
3852  * we can not satisfy this free in the cache layer.  true indicates that
3853  * the caller should retry.
3854  */
3855 static __noinline bool
3856 cache_free(uma_zone_t zone, uma_cache_t cache, void *udata, void *item,
3857     int itemdomain)
3858 {
3859 	uma_cache_bucket_t cbucket;
3860 	uma_bucket_t bucket;
3861 	int domain;
3862 
3863 	CRITICAL_ASSERT(curthread);
3864 
3865 	if (zone->uz_bucket_size == 0 || bucketdisable)
3866 		return false;
3867 
3868 	cache = &zone->uz_cpu[curcpu];
3869 
3870 	/*
3871 	 * FIRSTTOUCH domains need to free to the correct zdom.  When
3872 	 * enabled this is the zdom of the item.   The bucket is the
3873 	 * cross bucket if the current domain and itemdomain do not match.
3874 	 */
3875 	cbucket = &cache->uc_freebucket;
3876 #ifdef NUMA
3877 	if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0) {
3878 		domain = PCPU_GET(domain);
3879 		if (domain != itemdomain) {
3880 			cbucket = &cache->uc_crossbucket;
3881 			if (cbucket->ucb_cnt != 0)
3882 				atomic_add_64(&zone->uz_xdomain,
3883 				    cbucket->ucb_cnt);
3884 		}
3885 	} else
3886 #endif
3887 		itemdomain = domain = 0;
3888 	bucket = cache_bucket_unload(cbucket);
3889 
3890 	/* We are no longer associated with this CPU. */
3891 	critical_exit();
3892 
3893 	if (bucket != NULL)
3894 		zone_free_bucket(zone, bucket, udata, domain, itemdomain);
3895 
3896 	bucket = bucket_alloc(zone, udata, M_NOWAIT);
3897 	CTR3(KTR_UMA, "uma_zfree: zone %s(%p) allocated bucket %p",
3898 	    zone->uz_name, zone, bucket);
3899 	critical_enter();
3900 	if (bucket == NULL)
3901 		return (false);
3902 	cache = &zone->uz_cpu[curcpu];
3903 #ifdef NUMA
3904 	/*
3905 	 * Check to see if we should be populating the cross bucket.  If it
3906 	 * is already populated we will fall through and attempt to populate
3907 	 * the free bucket.
3908 	 */
3909 	if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0) {
3910 		domain = PCPU_GET(domain);
3911 		if (domain != itemdomain &&
3912 		    cache->uc_crossbucket.ucb_bucket == NULL) {
3913 			cache_bucket_load_cross(cache, bucket);
3914 			return (true);
3915 		}
3916 	}
3917 #endif
3918 	/*
3919 	 * We may have lost the race to fill the bucket or switched CPUs.
3920 	 */
3921 	if (cache->uc_freebucket.ucb_bucket != NULL) {
3922 		critical_exit();
3923 		bucket_free(zone, bucket, udata);
3924 		critical_enter();
3925 	} else
3926 		cache_bucket_load_free(cache, bucket);
3927 
3928 	return (true);
3929 }
3930 
3931 void
3932 uma_zfree_domain(uma_zone_t zone, void *item, void *udata)
3933 {
3934 
3935 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3936 	random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3937 
3938 	CTR2(KTR_UMA, "uma_zfree_domain zone %s(%p)", zone->uz_name, zone);
3939 
3940 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3941 	    ("uma_zfree_domain: called with spinlock or critical section held"));
3942 
3943         /* uma_zfree(..., NULL) does nothing, to match free(9). */
3944         if (item == NULL)
3945                 return;
3946 	zone_free_item(zone, item, udata, SKIP_NONE);
3947 }
3948 
3949 static void
3950 slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item)
3951 {
3952 	uma_keg_t keg;
3953 	uma_domain_t dom;
3954 	int freei;
3955 
3956 	keg = zone->uz_keg;
3957 	KEG_LOCK_ASSERT(keg, slab->us_domain);
3958 
3959 	/* Do we need to remove from any lists? */
3960 	dom = &keg->uk_domain[slab->us_domain];
3961 	if (slab->us_freecount+1 == keg->uk_ipers) {
3962 		LIST_REMOVE(slab, us_link);
3963 		LIST_INSERT_HEAD(&dom->ud_free_slab, slab, us_link);
3964 	} else if (slab->us_freecount == 0) {
3965 		LIST_REMOVE(slab, us_link);
3966 		LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
3967 	}
3968 
3969 	/* Slab management. */
3970 	freei = slab_item_index(slab, keg, item);
3971 	BIT_SET(keg->uk_ipers, freei, &slab->us_free);
3972 	slab->us_freecount++;
3973 
3974 	/* Keg statistics. */
3975 	dom->ud_free++;
3976 }
3977 
3978 static void
3979 zone_release(void *arg, void **bucket, int cnt)
3980 {
3981 	struct mtx *lock;
3982 	uma_zone_t zone;
3983 	uma_slab_t slab;
3984 	uma_keg_t keg;
3985 	uint8_t *mem;
3986 	void *item;
3987 	int i;
3988 
3989 	zone = arg;
3990 	keg = zone->uz_keg;
3991 	lock = NULL;
3992 	if (__predict_false((zone->uz_flags & UMA_ZFLAG_HASH) != 0))
3993 		lock = KEG_LOCK(keg, 0);
3994 	for (i = 0; i < cnt; i++) {
3995 		item = bucket[i];
3996 		if (__predict_true((zone->uz_flags & UMA_ZFLAG_VTOSLAB) != 0)) {
3997 			slab = vtoslab((vm_offset_t)item);
3998 		} else {
3999 			mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
4000 			if ((zone->uz_flags & UMA_ZFLAG_HASH) != 0)
4001 				slab = hash_sfind(&keg->uk_hash, mem);
4002 			else
4003 				slab = (uma_slab_t)(mem + keg->uk_pgoff);
4004 		}
4005 		if (lock != KEG_LOCKPTR(keg, slab->us_domain)) {
4006 			if (lock != NULL)
4007 				mtx_unlock(lock);
4008 			lock = KEG_LOCK(keg, slab->us_domain);
4009 		}
4010 		slab_free_item(zone, slab, item);
4011 	}
4012 	if (lock != NULL)
4013 		mtx_unlock(lock);
4014 }
4015 
4016 /*
4017  * Frees a single item to any zone.
4018  *
4019  * Arguments:
4020  *	zone   The zone to free to
4021  *	item   The item we're freeing
4022  *	udata  User supplied data for the dtor
4023  *	skip   Skip dtors and finis
4024  */
4025 static void
4026 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
4027 {
4028 
4029 	item_dtor(zone, item, zone->uz_size, udata, skip);
4030 
4031 	if (skip < SKIP_FINI && zone->uz_fini)
4032 		zone->uz_fini(item, zone->uz_size);
4033 
4034 	zone->uz_release(zone->uz_arg, &item, 1);
4035 
4036 	if (skip & SKIP_CNT)
4037 		return;
4038 
4039 	counter_u64_add(zone->uz_frees, 1);
4040 
4041 	if (zone->uz_max_items > 0)
4042 		zone_free_limit(zone, 1);
4043 }
4044 
4045 /* See uma.h */
4046 int
4047 uma_zone_set_max(uma_zone_t zone, int nitems)
4048 {
4049 	struct uma_bucket_zone *ubz;
4050 	int count;
4051 
4052 	/*
4053 	 * XXX This can misbehave if the zone has any allocations with
4054 	 * no limit and a limit is imposed.  There is currently no
4055 	 * way to clear a limit.
4056 	 */
4057 	ZONE_LOCK(zone);
4058 	ubz = bucket_zone_max(zone, nitems);
4059 	count = ubz != NULL ? ubz->ubz_entries : 0;
4060 	zone->uz_bucket_size_max = zone->uz_bucket_size = count;
4061 	if (zone->uz_bucket_size_min > zone->uz_bucket_size_max)
4062 		zone->uz_bucket_size_min = zone->uz_bucket_size_max;
4063 	zone->uz_max_items = nitems;
4064 	zone->uz_flags |= UMA_ZFLAG_LIMIT;
4065 	zone_update_caches(zone);
4066 	/* We may need to wake waiters. */
4067 	wakeup(&zone->uz_max_items);
4068 	ZONE_UNLOCK(zone);
4069 
4070 	return (nitems);
4071 }
4072 
4073 /* See uma.h */
4074 void
4075 uma_zone_set_maxcache(uma_zone_t zone, int nitems)
4076 {
4077 	struct uma_bucket_zone *ubz;
4078 	int bpcpu;
4079 
4080 	ZONE_LOCK(zone);
4081 	ubz = bucket_zone_max(zone, nitems);
4082 	if (ubz != NULL) {
4083 		bpcpu = 2;
4084 		if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4085 			/* Count the cross-domain bucket. */
4086 			bpcpu++;
4087 		nitems -= ubz->ubz_entries * bpcpu * mp_ncpus;
4088 		zone->uz_bucket_size_max = ubz->ubz_entries;
4089 	} else {
4090 		zone->uz_bucket_size_max = zone->uz_bucket_size = 0;
4091 	}
4092 	if (zone->uz_bucket_size_min > zone->uz_bucket_size_max)
4093 		zone->uz_bucket_size_min = zone->uz_bucket_size_max;
4094 	zone->uz_bkt_max = nitems;
4095 	ZONE_UNLOCK(zone);
4096 }
4097 
4098 /* See uma.h */
4099 int
4100 uma_zone_get_max(uma_zone_t zone)
4101 {
4102 	int nitems;
4103 
4104 	nitems = atomic_load_64(&zone->uz_max_items);
4105 
4106 	return (nitems);
4107 }
4108 
4109 /* See uma.h */
4110 void
4111 uma_zone_set_warning(uma_zone_t zone, const char *warning)
4112 {
4113 
4114 	ZONE_ASSERT_COLD(zone);
4115 	zone->uz_warning = warning;
4116 }
4117 
4118 /* See uma.h */
4119 void
4120 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction)
4121 {
4122 
4123 	ZONE_ASSERT_COLD(zone);
4124 	TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone);
4125 }
4126 
4127 /* See uma.h */
4128 int
4129 uma_zone_get_cur(uma_zone_t zone)
4130 {
4131 	int64_t nitems;
4132 	u_int i;
4133 
4134 	nitems = 0;
4135 	if (zone->uz_allocs != EARLY_COUNTER && zone->uz_frees != EARLY_COUNTER)
4136 		nitems = counter_u64_fetch(zone->uz_allocs) -
4137 		    counter_u64_fetch(zone->uz_frees);
4138 	CPU_FOREACH(i)
4139 		nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs) -
4140 		    atomic_load_64(&zone->uz_cpu[i].uc_frees);
4141 
4142 	return (nitems < 0 ? 0 : nitems);
4143 }
4144 
4145 static uint64_t
4146 uma_zone_get_allocs(uma_zone_t zone)
4147 {
4148 	uint64_t nitems;
4149 	u_int i;
4150 
4151 	nitems = 0;
4152 	if (zone->uz_allocs != EARLY_COUNTER)
4153 		nitems = counter_u64_fetch(zone->uz_allocs);
4154 	CPU_FOREACH(i)
4155 		nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs);
4156 
4157 	return (nitems);
4158 }
4159 
4160 static uint64_t
4161 uma_zone_get_frees(uma_zone_t zone)
4162 {
4163 	uint64_t nitems;
4164 	u_int i;
4165 
4166 	nitems = 0;
4167 	if (zone->uz_frees != EARLY_COUNTER)
4168 		nitems = counter_u64_fetch(zone->uz_frees);
4169 	CPU_FOREACH(i)
4170 		nitems += atomic_load_64(&zone->uz_cpu[i].uc_frees);
4171 
4172 	return (nitems);
4173 }
4174 
4175 #ifdef INVARIANTS
4176 /* Used only for KEG_ASSERT_COLD(). */
4177 static uint64_t
4178 uma_keg_get_allocs(uma_keg_t keg)
4179 {
4180 	uma_zone_t z;
4181 	uint64_t nitems;
4182 
4183 	nitems = 0;
4184 	LIST_FOREACH(z, &keg->uk_zones, uz_link)
4185 		nitems += uma_zone_get_allocs(z);
4186 
4187 	return (nitems);
4188 }
4189 #endif
4190 
4191 /* See uma.h */
4192 void
4193 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
4194 {
4195 	uma_keg_t keg;
4196 
4197 	KEG_GET(zone, keg);
4198 	KEG_ASSERT_COLD(keg);
4199 	keg->uk_init = uminit;
4200 }
4201 
4202 /* See uma.h */
4203 void
4204 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
4205 {
4206 	uma_keg_t keg;
4207 
4208 	KEG_GET(zone, keg);
4209 	KEG_ASSERT_COLD(keg);
4210 	keg->uk_fini = fini;
4211 }
4212 
4213 /* See uma.h */
4214 void
4215 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
4216 {
4217 
4218 	ZONE_ASSERT_COLD(zone);
4219 	zone->uz_init = zinit;
4220 }
4221 
4222 /* See uma.h */
4223 void
4224 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
4225 {
4226 
4227 	ZONE_ASSERT_COLD(zone);
4228 	zone->uz_fini = zfini;
4229 }
4230 
4231 /* See uma.h */
4232 void
4233 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
4234 {
4235 	uma_keg_t keg;
4236 
4237 	KEG_GET(zone, keg);
4238 	KEG_ASSERT_COLD(keg);
4239 	keg->uk_freef = freef;
4240 }
4241 
4242 /* See uma.h */
4243 void
4244 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
4245 {
4246 	uma_keg_t keg;
4247 
4248 	KEG_GET(zone, keg);
4249 	KEG_ASSERT_COLD(keg);
4250 	keg->uk_allocf = allocf;
4251 }
4252 
4253 /* See uma.h */
4254 void
4255 uma_zone_reserve(uma_zone_t zone, int items)
4256 {
4257 	uma_keg_t keg;
4258 
4259 	KEG_GET(zone, keg);
4260 	KEG_ASSERT_COLD(keg);
4261 	keg->uk_reserve = items;
4262 }
4263 
4264 /* See uma.h */
4265 int
4266 uma_zone_reserve_kva(uma_zone_t zone, int count)
4267 {
4268 	uma_keg_t keg;
4269 	vm_offset_t kva;
4270 	u_int pages;
4271 
4272 	KEG_GET(zone, keg);
4273 	KEG_ASSERT_COLD(keg);
4274 	ZONE_ASSERT_COLD(zone);
4275 
4276 	pages = howmany(count, keg->uk_ipers) * keg->uk_ppera;
4277 
4278 #ifdef UMA_MD_SMALL_ALLOC
4279 	if (keg->uk_ppera > 1) {
4280 #else
4281 	if (1) {
4282 #endif
4283 		kva = kva_alloc((vm_size_t)pages * PAGE_SIZE);
4284 		if (kva == 0)
4285 			return (0);
4286 	} else
4287 		kva = 0;
4288 
4289 	ZONE_LOCK(zone);
4290 	MPASS(keg->uk_kva == 0);
4291 	keg->uk_kva = kva;
4292 	keg->uk_offset = 0;
4293 	zone->uz_max_items = pages * keg->uk_ipers;
4294 #ifdef UMA_MD_SMALL_ALLOC
4295 	keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
4296 #else
4297 	keg->uk_allocf = noobj_alloc;
4298 #endif
4299 	keg->uk_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
4300 	zone->uz_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
4301 	zone_update_caches(zone);
4302 	ZONE_UNLOCK(zone);
4303 
4304 	return (1);
4305 }
4306 
4307 /* See uma.h */
4308 void
4309 uma_prealloc(uma_zone_t zone, int items)
4310 {
4311 	struct vm_domainset_iter di;
4312 	uma_domain_t dom;
4313 	uma_slab_t slab;
4314 	uma_keg_t keg;
4315 	int aflags, domain, slabs;
4316 
4317 	KEG_GET(zone, keg);
4318 	slabs = howmany(items, keg->uk_ipers);
4319 	while (slabs-- > 0) {
4320 		aflags = M_NOWAIT;
4321 		vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
4322 		    &aflags);
4323 		for (;;) {
4324 			slab = keg_alloc_slab(keg, zone, domain, M_WAITOK,
4325 			    aflags);
4326 			if (slab != NULL) {
4327 				dom = &keg->uk_domain[slab->us_domain];
4328 				LIST_REMOVE(slab, us_link);
4329 				LIST_INSERT_HEAD(&dom->ud_free_slab, slab,
4330 				    us_link);
4331 				KEG_UNLOCK(keg, slab->us_domain);
4332 				break;
4333 			}
4334 			if (vm_domainset_iter_policy(&di, &domain) != 0)
4335 				vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask);
4336 		}
4337 	}
4338 }
4339 
4340 /* See uma.h */
4341 void
4342 uma_reclaim(int req)
4343 {
4344 
4345 	CTR0(KTR_UMA, "UMA: vm asked us to release pages!");
4346 	sx_xlock(&uma_reclaim_lock);
4347 	bucket_enable();
4348 
4349 	switch (req) {
4350 	case UMA_RECLAIM_TRIM:
4351 		zone_foreach(zone_trim, NULL);
4352 		break;
4353 	case UMA_RECLAIM_DRAIN:
4354 	case UMA_RECLAIM_DRAIN_CPU:
4355 		zone_foreach(zone_drain, NULL);
4356 		if (req == UMA_RECLAIM_DRAIN_CPU) {
4357 			pcpu_cache_drain_safe(NULL);
4358 			zone_foreach(zone_drain, NULL);
4359 		}
4360 		break;
4361 	default:
4362 		panic("unhandled reclamation request %d", req);
4363 	}
4364 
4365 	/*
4366 	 * Some slabs may have been freed but this zone will be visited early
4367 	 * we visit again so that we can free pages that are empty once other
4368 	 * zones are drained.  We have to do the same for buckets.
4369 	 */
4370 	zone_drain(slabzones[0], NULL);
4371 	zone_drain(slabzones[1], NULL);
4372 	bucket_zone_drain();
4373 	sx_xunlock(&uma_reclaim_lock);
4374 }
4375 
4376 static volatile int uma_reclaim_needed;
4377 
4378 void
4379 uma_reclaim_wakeup(void)
4380 {
4381 
4382 	if (atomic_fetchadd_int(&uma_reclaim_needed, 1) == 0)
4383 		wakeup(uma_reclaim);
4384 }
4385 
4386 void
4387 uma_reclaim_worker(void *arg __unused)
4388 {
4389 
4390 	for (;;) {
4391 		sx_xlock(&uma_reclaim_lock);
4392 		while (atomic_load_int(&uma_reclaim_needed) == 0)
4393 			sx_sleep(uma_reclaim, &uma_reclaim_lock, PVM, "umarcl",
4394 			    hz);
4395 		sx_xunlock(&uma_reclaim_lock);
4396 		EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM);
4397 		uma_reclaim(UMA_RECLAIM_DRAIN_CPU);
4398 		atomic_store_int(&uma_reclaim_needed, 0);
4399 		/* Don't fire more than once per-second. */
4400 		pause("umarclslp", hz);
4401 	}
4402 }
4403 
4404 /* See uma.h */
4405 void
4406 uma_zone_reclaim(uma_zone_t zone, int req)
4407 {
4408 
4409 	switch (req) {
4410 	case UMA_RECLAIM_TRIM:
4411 		zone_trim(zone, NULL);
4412 		break;
4413 	case UMA_RECLAIM_DRAIN:
4414 		zone_drain(zone, NULL);
4415 		break;
4416 	case UMA_RECLAIM_DRAIN_CPU:
4417 		pcpu_cache_drain_safe(zone);
4418 		zone_drain(zone, NULL);
4419 		break;
4420 	default:
4421 		panic("unhandled reclamation request %d", req);
4422 	}
4423 }
4424 
4425 /* See uma.h */
4426 int
4427 uma_zone_exhausted(uma_zone_t zone)
4428 {
4429 
4430 	return (atomic_load_32(&zone->uz_sleepers) > 0);
4431 }
4432 
4433 unsigned long
4434 uma_limit(void)
4435 {
4436 
4437 	return (uma_kmem_limit);
4438 }
4439 
4440 void
4441 uma_set_limit(unsigned long limit)
4442 {
4443 
4444 	uma_kmem_limit = limit;
4445 }
4446 
4447 unsigned long
4448 uma_size(void)
4449 {
4450 
4451 	return (atomic_load_long(&uma_kmem_total));
4452 }
4453 
4454 long
4455 uma_avail(void)
4456 {
4457 
4458 	return (uma_kmem_limit - uma_size());
4459 }
4460 
4461 #ifdef DDB
4462 /*
4463  * Generate statistics across both the zone and its per-cpu cache's.  Return
4464  * desired statistics if the pointer is non-NULL for that statistic.
4465  *
4466  * Note: does not update the zone statistics, as it can't safely clear the
4467  * per-CPU cache statistic.
4468  *
4469  */
4470 static void
4471 uma_zone_sumstat(uma_zone_t z, long *cachefreep, uint64_t *allocsp,
4472     uint64_t *freesp, uint64_t *sleepsp, uint64_t *xdomainp)
4473 {
4474 	uma_cache_t cache;
4475 	uint64_t allocs, frees, sleeps, xdomain;
4476 	int cachefree, cpu;
4477 
4478 	allocs = frees = sleeps = xdomain = 0;
4479 	cachefree = 0;
4480 	CPU_FOREACH(cpu) {
4481 		cache = &z->uz_cpu[cpu];
4482 		cachefree += cache->uc_allocbucket.ucb_cnt;
4483 		cachefree += cache->uc_freebucket.ucb_cnt;
4484 		xdomain += cache->uc_crossbucket.ucb_cnt;
4485 		cachefree += cache->uc_crossbucket.ucb_cnt;
4486 		allocs += cache->uc_allocs;
4487 		frees += cache->uc_frees;
4488 	}
4489 	allocs += counter_u64_fetch(z->uz_allocs);
4490 	frees += counter_u64_fetch(z->uz_frees);
4491 	sleeps += z->uz_sleeps;
4492 	xdomain += z->uz_xdomain;
4493 	if (cachefreep != NULL)
4494 		*cachefreep = cachefree;
4495 	if (allocsp != NULL)
4496 		*allocsp = allocs;
4497 	if (freesp != NULL)
4498 		*freesp = frees;
4499 	if (sleepsp != NULL)
4500 		*sleepsp = sleeps;
4501 	if (xdomainp != NULL)
4502 		*xdomainp = xdomain;
4503 }
4504 #endif /* DDB */
4505 
4506 static int
4507 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
4508 {
4509 	uma_keg_t kz;
4510 	uma_zone_t z;
4511 	int count;
4512 
4513 	count = 0;
4514 	rw_rlock(&uma_rwlock);
4515 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
4516 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
4517 			count++;
4518 	}
4519 	LIST_FOREACH(z, &uma_cachezones, uz_link)
4520 		count++;
4521 
4522 	rw_runlock(&uma_rwlock);
4523 	return (sysctl_handle_int(oidp, &count, 0, req));
4524 }
4525 
4526 static void
4527 uma_vm_zone_stats(struct uma_type_header *uth, uma_zone_t z, struct sbuf *sbuf,
4528     struct uma_percpu_stat *ups, bool internal)
4529 {
4530 	uma_zone_domain_t zdom;
4531 	uma_cache_t cache;
4532 	int i;
4533 
4534 
4535 	for (i = 0; i < vm_ndomains; i++) {
4536 		zdom = &z->uz_domain[i];
4537 		uth->uth_zone_free += zdom->uzd_nitems;
4538 	}
4539 	uth->uth_allocs = counter_u64_fetch(z->uz_allocs);
4540 	uth->uth_frees = counter_u64_fetch(z->uz_frees);
4541 	uth->uth_fails = counter_u64_fetch(z->uz_fails);
4542 	uth->uth_sleeps = z->uz_sleeps;
4543 	uth->uth_xdomain = z->uz_xdomain;
4544 
4545 	/*
4546 	 * While it is not normally safe to access the cache bucket pointers
4547 	 * while not on the CPU that owns the cache, we only allow the pointers
4548 	 * to be exchanged without the zone lock held, not invalidated, so
4549 	 * accept the possible race associated with bucket exchange during
4550 	 * monitoring.  Use atomic_load_ptr() to ensure that the bucket pointers
4551 	 * are loaded only once.
4552 	 */
4553 	for (i = 0; i < mp_maxid + 1; i++) {
4554 		bzero(&ups[i], sizeof(*ups));
4555 		if (internal || CPU_ABSENT(i))
4556 			continue;
4557 		cache = &z->uz_cpu[i];
4558 		ups[i].ups_cache_free += cache->uc_allocbucket.ucb_cnt;
4559 		ups[i].ups_cache_free += cache->uc_freebucket.ucb_cnt;
4560 		ups[i].ups_cache_free += cache->uc_crossbucket.ucb_cnt;
4561 		ups[i].ups_allocs = cache->uc_allocs;
4562 		ups[i].ups_frees = cache->uc_frees;
4563 	}
4564 }
4565 
4566 static int
4567 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
4568 {
4569 	struct uma_stream_header ush;
4570 	struct uma_type_header uth;
4571 	struct uma_percpu_stat *ups;
4572 	struct sbuf sbuf;
4573 	uma_keg_t kz;
4574 	uma_zone_t z;
4575 	uint64_t items;
4576 	uint32_t kfree, pages;
4577 	int count, error, i;
4578 
4579 	error = sysctl_wire_old_buffer(req, 0);
4580 	if (error != 0)
4581 		return (error);
4582 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
4583 	sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL);
4584 	ups = malloc((mp_maxid + 1) * sizeof(*ups), M_TEMP, M_WAITOK);
4585 
4586 	count = 0;
4587 	rw_rlock(&uma_rwlock);
4588 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
4589 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
4590 			count++;
4591 	}
4592 
4593 	LIST_FOREACH(z, &uma_cachezones, uz_link)
4594 		count++;
4595 
4596 	/*
4597 	 * Insert stream header.
4598 	 */
4599 	bzero(&ush, sizeof(ush));
4600 	ush.ush_version = UMA_STREAM_VERSION;
4601 	ush.ush_maxcpus = (mp_maxid + 1);
4602 	ush.ush_count = count;
4603 	(void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
4604 
4605 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
4606 		kfree = pages = 0;
4607 		for (i = 0; i < vm_ndomains; i++) {
4608 			kfree += kz->uk_domain[i].ud_free;
4609 			pages += kz->uk_domain[i].ud_pages;
4610 		}
4611 		LIST_FOREACH(z, &kz->uk_zones, uz_link) {
4612 			bzero(&uth, sizeof(uth));
4613 			ZONE_LOCK(z);
4614 			strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
4615 			uth.uth_align = kz->uk_align;
4616 			uth.uth_size = kz->uk_size;
4617 			uth.uth_rsize = kz->uk_rsize;
4618 			if (z->uz_max_items > 0) {
4619 				items = UZ_ITEMS_COUNT(z->uz_items);
4620 				uth.uth_pages = (items / kz->uk_ipers) *
4621 					kz->uk_ppera;
4622 			} else
4623 				uth.uth_pages = pages;
4624 			uth.uth_maxpages = (z->uz_max_items / kz->uk_ipers) *
4625 			    kz->uk_ppera;
4626 			uth.uth_limit = z->uz_max_items;
4627 			uth.uth_keg_free = kfree;
4628 
4629 			/*
4630 			 * A zone is secondary is it is not the first entry
4631 			 * on the keg's zone list.
4632 			 */
4633 			if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
4634 			    (LIST_FIRST(&kz->uk_zones) != z))
4635 				uth.uth_zone_flags = UTH_ZONE_SECONDARY;
4636 			uma_vm_zone_stats(&uth, z, &sbuf, ups,
4637 			    kz->uk_flags & UMA_ZFLAG_INTERNAL);
4638 			ZONE_UNLOCK(z);
4639 			(void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
4640 			for (i = 0; i < mp_maxid + 1; i++)
4641 				(void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
4642 		}
4643 	}
4644 	LIST_FOREACH(z, &uma_cachezones, uz_link) {
4645 		bzero(&uth, sizeof(uth));
4646 		ZONE_LOCK(z);
4647 		strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
4648 		uth.uth_size = z->uz_size;
4649 		uma_vm_zone_stats(&uth, z, &sbuf, ups, false);
4650 		ZONE_UNLOCK(z);
4651 		(void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
4652 		for (i = 0; i < mp_maxid + 1; i++)
4653 			(void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
4654 	}
4655 
4656 	rw_runlock(&uma_rwlock);
4657 	error = sbuf_finish(&sbuf);
4658 	sbuf_delete(&sbuf);
4659 	free(ups, M_TEMP);
4660 	return (error);
4661 }
4662 
4663 int
4664 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS)
4665 {
4666 	uma_zone_t zone = *(uma_zone_t *)arg1;
4667 	int error, max;
4668 
4669 	max = uma_zone_get_max(zone);
4670 	error = sysctl_handle_int(oidp, &max, 0, req);
4671 	if (error || !req->newptr)
4672 		return (error);
4673 
4674 	uma_zone_set_max(zone, max);
4675 
4676 	return (0);
4677 }
4678 
4679 int
4680 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS)
4681 {
4682 	uma_zone_t zone;
4683 	int cur;
4684 
4685 	/*
4686 	 * Some callers want to add sysctls for global zones that
4687 	 * may not yet exist so they pass a pointer to a pointer.
4688 	 */
4689 	if (arg2 == 0)
4690 		zone = *(uma_zone_t *)arg1;
4691 	else
4692 		zone = arg1;
4693 	cur = uma_zone_get_cur(zone);
4694 	return (sysctl_handle_int(oidp, &cur, 0, req));
4695 }
4696 
4697 static int
4698 sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS)
4699 {
4700 	uma_zone_t zone = arg1;
4701 	uint64_t cur;
4702 
4703 	cur = uma_zone_get_allocs(zone);
4704 	return (sysctl_handle_64(oidp, &cur, 0, req));
4705 }
4706 
4707 static int
4708 sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS)
4709 {
4710 	uma_zone_t zone = arg1;
4711 	uint64_t cur;
4712 
4713 	cur = uma_zone_get_frees(zone);
4714 	return (sysctl_handle_64(oidp, &cur, 0, req));
4715 }
4716 
4717 static int
4718 sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS)
4719 {
4720 	struct sbuf sbuf;
4721 	uma_zone_t zone = arg1;
4722 	int error;
4723 
4724 	sbuf_new_for_sysctl(&sbuf, NULL, 0, req);
4725 	if (zone->uz_flags != 0)
4726 		sbuf_printf(&sbuf, "0x%b", zone->uz_flags, PRINT_UMA_ZFLAGS);
4727 	else
4728 		sbuf_printf(&sbuf, "0");
4729 	error = sbuf_finish(&sbuf);
4730 	sbuf_delete(&sbuf);
4731 
4732 	return (error);
4733 }
4734 
4735 static int
4736 sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS)
4737 {
4738 	uma_keg_t keg = arg1;
4739 	int avail, effpct, total;
4740 
4741 	total = keg->uk_ppera * PAGE_SIZE;
4742 	if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0)
4743 		total += slabzone(keg->uk_ipers)->uz_keg->uk_rsize;
4744 	/*
4745 	 * We consider the client's requested size and alignment here, not the
4746 	 * real size determination uk_rsize, because we also adjust the real
4747 	 * size for internal implementation reasons (max bitset size).
4748 	 */
4749 	avail = keg->uk_ipers * roundup2(keg->uk_size, keg->uk_align + 1);
4750 	if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
4751 		avail *= mp_maxid + 1;
4752 	effpct = 100 * avail / total;
4753 	return (sysctl_handle_int(oidp, &effpct, 0, req));
4754 }
4755 
4756 static int
4757 sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS)
4758 {
4759 	uma_zone_t zone = arg1;
4760 	uint64_t cur;
4761 
4762 	cur = UZ_ITEMS_COUNT(atomic_load_64(&zone->uz_items));
4763 	return (sysctl_handle_64(oidp, &cur, 0, req));
4764 }
4765 
4766 #ifdef INVARIANTS
4767 static uma_slab_t
4768 uma_dbg_getslab(uma_zone_t zone, void *item)
4769 {
4770 	uma_slab_t slab;
4771 	uma_keg_t keg;
4772 	uint8_t *mem;
4773 
4774 	/*
4775 	 * It is safe to return the slab here even though the
4776 	 * zone is unlocked because the item's allocation state
4777 	 * essentially holds a reference.
4778 	 */
4779 	mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
4780 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
4781 		return (NULL);
4782 	if (zone->uz_flags & UMA_ZFLAG_VTOSLAB)
4783 		return (vtoslab((vm_offset_t)mem));
4784 	keg = zone->uz_keg;
4785 	if ((keg->uk_flags & UMA_ZFLAG_HASH) == 0)
4786 		return ((uma_slab_t)(mem + keg->uk_pgoff));
4787 	KEG_LOCK(keg, 0);
4788 	slab = hash_sfind(&keg->uk_hash, mem);
4789 	KEG_UNLOCK(keg, 0);
4790 
4791 	return (slab);
4792 }
4793 
4794 static bool
4795 uma_dbg_zskip(uma_zone_t zone, void *mem)
4796 {
4797 
4798 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
4799 		return (true);
4800 
4801 	return (uma_dbg_kskip(zone->uz_keg, mem));
4802 }
4803 
4804 static bool
4805 uma_dbg_kskip(uma_keg_t keg, void *mem)
4806 {
4807 	uintptr_t idx;
4808 
4809 	if (dbg_divisor == 0)
4810 		return (true);
4811 
4812 	if (dbg_divisor == 1)
4813 		return (false);
4814 
4815 	idx = (uintptr_t)mem >> PAGE_SHIFT;
4816 	if (keg->uk_ipers > 1) {
4817 		idx *= keg->uk_ipers;
4818 		idx += ((uintptr_t)mem & PAGE_MASK) / keg->uk_rsize;
4819 	}
4820 
4821 	if ((idx / dbg_divisor) * dbg_divisor != idx) {
4822 		counter_u64_add(uma_skip_cnt, 1);
4823 		return (true);
4824 	}
4825 	counter_u64_add(uma_dbg_cnt, 1);
4826 
4827 	return (false);
4828 }
4829 
4830 /*
4831  * Set up the slab's freei data such that uma_dbg_free can function.
4832  *
4833  */
4834 static void
4835 uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item)
4836 {
4837 	uma_keg_t keg;
4838 	int freei;
4839 
4840 	if (slab == NULL) {
4841 		slab = uma_dbg_getslab(zone, item);
4842 		if (slab == NULL)
4843 			panic("uma: item %p did not belong to zone %s\n",
4844 			    item, zone->uz_name);
4845 	}
4846 	keg = zone->uz_keg;
4847 	freei = slab_item_index(slab, keg, item);
4848 
4849 	if (BIT_ISSET(keg->uk_ipers, freei, slab_dbg_bits(slab, keg)))
4850 		panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)\n",
4851 		    item, zone, zone->uz_name, slab, freei);
4852 	BIT_SET_ATOMIC(keg->uk_ipers, freei, slab_dbg_bits(slab, keg));
4853 }
4854 
4855 /*
4856  * Verifies freed addresses.  Checks for alignment, valid slab membership
4857  * and duplicate frees.
4858  *
4859  */
4860 static void
4861 uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item)
4862 {
4863 	uma_keg_t keg;
4864 	int freei;
4865 
4866 	if (slab == NULL) {
4867 		slab = uma_dbg_getslab(zone, item);
4868 		if (slab == NULL)
4869 			panic("uma: Freed item %p did not belong to zone %s\n",
4870 			    item, zone->uz_name);
4871 	}
4872 	keg = zone->uz_keg;
4873 	freei = slab_item_index(slab, keg, item);
4874 
4875 	if (freei >= keg->uk_ipers)
4876 		panic("Invalid free of %p from zone %p(%s) slab %p(%d)\n",
4877 		    item, zone, zone->uz_name, slab, freei);
4878 
4879 	if (slab_item(slab, keg, freei) != item)
4880 		panic("Unaligned free of %p from zone %p(%s) slab %p(%d)\n",
4881 		    item, zone, zone->uz_name, slab, freei);
4882 
4883 	if (!BIT_ISSET(keg->uk_ipers, freei, slab_dbg_bits(slab, keg)))
4884 		panic("Duplicate free of %p from zone %p(%s) slab %p(%d)\n",
4885 		    item, zone, zone->uz_name, slab, freei);
4886 
4887 	BIT_CLR_ATOMIC(keg->uk_ipers, freei, slab_dbg_bits(slab, keg));
4888 }
4889 #endif /* INVARIANTS */
4890 
4891 #ifdef DDB
4892 static int64_t
4893 get_uma_stats(uma_keg_t kz, uma_zone_t z, uint64_t *allocs, uint64_t *used,
4894     uint64_t *sleeps, long *cachefree, uint64_t *xdomain)
4895 {
4896 	uint64_t frees;
4897 	int i;
4898 
4899 	if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
4900 		*allocs = counter_u64_fetch(z->uz_allocs);
4901 		frees = counter_u64_fetch(z->uz_frees);
4902 		*sleeps = z->uz_sleeps;
4903 		*cachefree = 0;
4904 		*xdomain = 0;
4905 	} else
4906 		uma_zone_sumstat(z, cachefree, allocs, &frees, sleeps,
4907 		    xdomain);
4908 	for (i = 0; i < vm_ndomains; i++) {
4909 		*cachefree += z->uz_domain[i].uzd_nitems;
4910 		if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
4911 		    (LIST_FIRST(&kz->uk_zones) != z)))
4912 			*cachefree += kz->uk_domain[i].ud_free;
4913 	}
4914 	*used = *allocs - frees;
4915 	return (((int64_t)*used + *cachefree) * kz->uk_size);
4916 }
4917 
4918 DB_SHOW_COMMAND(uma, db_show_uma)
4919 {
4920 	const char *fmt_hdr, *fmt_entry;
4921 	uma_keg_t kz;
4922 	uma_zone_t z;
4923 	uint64_t allocs, used, sleeps, xdomain;
4924 	long cachefree;
4925 	/* variables for sorting */
4926 	uma_keg_t cur_keg;
4927 	uma_zone_t cur_zone, last_zone;
4928 	int64_t cur_size, last_size, size;
4929 	int ties;
4930 
4931 	/* /i option produces machine-parseable CSV output */
4932 	if (modif[0] == 'i') {
4933 		fmt_hdr = "%s,%s,%s,%s,%s,%s,%s,%s,%s\n";
4934 		fmt_entry = "\"%s\",%ju,%jd,%ld,%ju,%ju,%u,%jd,%ju\n";
4935 	} else {
4936 		fmt_hdr = "%18s %6s %7s %7s %11s %7s %7s %10s %8s\n";
4937 		fmt_entry = "%18s %6ju %7jd %7ld %11ju %7ju %7u %10jd %8ju\n";
4938 	}
4939 
4940 	db_printf(fmt_hdr, "Zone", "Size", "Used", "Free", "Requests",
4941 	    "Sleeps", "Bucket", "Total Mem", "XFree");
4942 
4943 	/* Sort the zones with largest size first. */
4944 	last_zone = NULL;
4945 	last_size = INT64_MAX;
4946 	for (;;) {
4947 		cur_zone = NULL;
4948 		cur_size = -1;
4949 		ties = 0;
4950 		LIST_FOREACH(kz, &uma_kegs, uk_link) {
4951 			LIST_FOREACH(z, &kz->uk_zones, uz_link) {
4952 				/*
4953 				 * In the case of size ties, print out zones
4954 				 * in the order they are encountered.  That is,
4955 				 * when we encounter the most recently output
4956 				 * zone, we have already printed all preceding
4957 				 * ties, and we must print all following ties.
4958 				 */
4959 				if (z == last_zone) {
4960 					ties = 1;
4961 					continue;
4962 				}
4963 				size = get_uma_stats(kz, z, &allocs, &used,
4964 				    &sleeps, &cachefree, &xdomain);
4965 				if (size > cur_size && size < last_size + ties)
4966 				{
4967 					cur_size = size;
4968 					cur_zone = z;
4969 					cur_keg = kz;
4970 				}
4971 			}
4972 		}
4973 		if (cur_zone == NULL)
4974 			break;
4975 
4976 		size = get_uma_stats(cur_keg, cur_zone, &allocs, &used,
4977 		    &sleeps, &cachefree, &xdomain);
4978 		db_printf(fmt_entry, cur_zone->uz_name,
4979 		    (uintmax_t)cur_keg->uk_size, (intmax_t)used, cachefree,
4980 		    (uintmax_t)allocs, (uintmax_t)sleeps,
4981 		    (unsigned)cur_zone->uz_bucket_size, (intmax_t)size,
4982 		    xdomain);
4983 
4984 		if (db_pager_quit)
4985 			return;
4986 		last_zone = cur_zone;
4987 		last_size = cur_size;
4988 	}
4989 }
4990 
4991 DB_SHOW_COMMAND(umacache, db_show_umacache)
4992 {
4993 	uma_zone_t z;
4994 	uint64_t allocs, frees;
4995 	long cachefree;
4996 	int i;
4997 
4998 	db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
4999 	    "Requests", "Bucket");
5000 	LIST_FOREACH(z, &uma_cachezones, uz_link) {
5001 		uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL, NULL);
5002 		for (i = 0; i < vm_ndomains; i++)
5003 			cachefree += z->uz_domain[i].uzd_nitems;
5004 		db_printf("%18s %8ju %8jd %8ld %12ju %8u\n",
5005 		    z->uz_name, (uintmax_t)z->uz_size,
5006 		    (intmax_t)(allocs - frees), cachefree,
5007 		    (uintmax_t)allocs, z->uz_bucket_size);
5008 		if (db_pager_quit)
5009 			return;
5010 	}
5011 }
5012 #endif	/* DDB */
5013