xref: /freebsd/sys/vm/uma_core.c (revision d0ba1baed3f6e4936a0c1b89c25f6c59168ef6de)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2002-2005, 2009, 2013 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/eventhandler.h>
63 #include <sys/kernel.h>
64 #include <sys/types.h>
65 #include <sys/limits.h>
66 #include <sys/queue.h>
67 #include <sys/malloc.h>
68 #include <sys/ktr.h>
69 #include <sys/lock.h>
70 #include <sys/sysctl.h>
71 #include <sys/mutex.h>
72 #include <sys/proc.h>
73 #include <sys/random.h>
74 #include <sys/rwlock.h>
75 #include <sys/sbuf.h>
76 #include <sys/sched.h>
77 #include <sys/smp.h>
78 #include <sys/taskqueue.h>
79 #include <sys/vmmeter.h>
80 
81 #include <vm/vm.h>
82 #include <vm/vm_object.h>
83 #include <vm/vm_page.h>
84 #include <vm/vm_pageout.h>
85 #include <vm/vm_param.h>
86 #include <vm/vm_phys.h>
87 #include <vm/vm_map.h>
88 #include <vm/vm_kern.h>
89 #include <vm/vm_extern.h>
90 #include <vm/uma.h>
91 #include <vm/uma_int.h>
92 #include <vm/uma_dbg.h>
93 
94 #include <ddb/ddb.h>
95 
96 #ifdef DEBUG_MEMGUARD
97 #include <vm/memguard.h>
98 #endif
99 
100 /*
101  * This is the zone and keg from which all zones are spawned.
102  */
103 static uma_zone_t kegs;
104 static uma_zone_t zones;
105 
106 /* This is the zone from which all offpage uma_slab_ts are allocated. */
107 static uma_zone_t slabzone;
108 
109 /*
110  * The initial hash tables come out of this zone so they can be allocated
111  * prior to malloc coming up.
112  */
113 static uma_zone_t hashzone;
114 
115 /* The boot-time adjusted value for cache line alignment. */
116 int uma_align_cache = 64 - 1;
117 
118 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets");
119 
120 /*
121  * Are we allowed to allocate buckets?
122  */
123 static int bucketdisable = 1;
124 
125 /* Linked list of all kegs in the system */
126 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs);
127 
128 /* Linked list of all cache-only zones in the system */
129 static LIST_HEAD(,uma_zone) uma_cachezones =
130     LIST_HEAD_INITIALIZER(uma_cachezones);
131 
132 /* This RW lock protects the keg list */
133 static struct rwlock_padalign __exclusive_cache_line uma_rwlock;
134 
135 /*
136  * Pointer and counter to pool of pages, that is preallocated at
137  * startup to bootstrap UMA.
138  */
139 static char *bootmem;
140 static int boot_pages;
141 
142 static struct sx uma_drain_lock;
143 
144 /* kmem soft limit. */
145 static unsigned long uma_kmem_limit = LONG_MAX;
146 static volatile unsigned long uma_kmem_total;
147 
148 /* Is the VM done starting up? */
149 static enum { BOOT_COLD = 0, BOOT_STRAPPED, BOOT_PAGEALLOC, BOOT_BUCKETS,
150     BOOT_RUNNING } booted = BOOT_COLD;
151 
152 /*
153  * This is the handle used to schedule events that need to happen
154  * outside of the allocation fast path.
155  */
156 static struct callout uma_callout;
157 #define	UMA_TIMEOUT	20		/* Seconds for callout interval. */
158 
159 /*
160  * This structure is passed as the zone ctor arg so that I don't have to create
161  * a special allocation function just for zones.
162  */
163 struct uma_zctor_args {
164 	const char *name;
165 	size_t size;
166 	uma_ctor ctor;
167 	uma_dtor dtor;
168 	uma_init uminit;
169 	uma_fini fini;
170 	uma_import import;
171 	uma_release release;
172 	void *arg;
173 	uma_keg_t keg;
174 	int align;
175 	uint32_t flags;
176 };
177 
178 struct uma_kctor_args {
179 	uma_zone_t zone;
180 	size_t size;
181 	uma_init uminit;
182 	uma_fini fini;
183 	int align;
184 	uint32_t flags;
185 };
186 
187 struct uma_bucket_zone {
188 	uma_zone_t	ubz_zone;
189 	char		*ubz_name;
190 	int		ubz_entries;	/* Number of items it can hold. */
191 	int		ubz_maxsize;	/* Maximum allocation size per-item. */
192 };
193 
194 /*
195  * Compute the actual number of bucket entries to pack them in power
196  * of two sizes for more efficient space utilization.
197  */
198 #define	BUCKET_SIZE(n)						\
199     (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *))
200 
201 #define	BUCKET_MAX	BUCKET_SIZE(256)
202 
203 struct uma_bucket_zone bucket_zones[] = {
204 	{ NULL, "4 Bucket", BUCKET_SIZE(4), 4096 },
205 	{ NULL, "6 Bucket", BUCKET_SIZE(6), 3072 },
206 	{ NULL, "8 Bucket", BUCKET_SIZE(8), 2048 },
207 	{ NULL, "12 Bucket", BUCKET_SIZE(12), 1536 },
208 	{ NULL, "16 Bucket", BUCKET_SIZE(16), 1024 },
209 	{ NULL, "32 Bucket", BUCKET_SIZE(32), 512 },
210 	{ NULL, "64 Bucket", BUCKET_SIZE(64), 256 },
211 	{ NULL, "128 Bucket", BUCKET_SIZE(128), 128 },
212 	{ NULL, "256 Bucket", BUCKET_SIZE(256), 64 },
213 	{ NULL, NULL, 0}
214 };
215 
216 /*
217  * Flags and enumerations to be passed to internal functions.
218  */
219 enum zfreeskip { SKIP_NONE = 0, SKIP_DTOR, SKIP_FINI };
220 
221 #define	UMA_ANYDOMAIN	-1	/* Special value for domain search. */
222 
223 /* Prototypes.. */
224 
225 int	uma_startup_count(int);
226 void	uma_startup(void *, int);
227 void	uma_startup1(void);
228 void	uma_startup2(void);
229 
230 static void *noobj_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
231 static void *page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
232 static void *startup_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
233 static void page_free(void *, vm_size_t, uint8_t);
234 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int, int);
235 static void cache_drain(uma_zone_t);
236 static void bucket_drain(uma_zone_t, uma_bucket_t);
237 static void bucket_cache_drain(uma_zone_t zone);
238 static int keg_ctor(void *, int, void *, int);
239 static void keg_dtor(void *, int, void *);
240 static int zone_ctor(void *, int, void *, int);
241 static void zone_dtor(void *, int, void *);
242 static int zero_init(void *, int, int);
243 static void keg_small_init(uma_keg_t keg);
244 static void keg_large_init(uma_keg_t keg);
245 static void zone_foreach(void (*zfunc)(uma_zone_t));
246 static void zone_timeout(uma_zone_t zone);
247 static int hash_alloc(struct uma_hash *);
248 static int hash_expand(struct uma_hash *, struct uma_hash *);
249 static void hash_free(struct uma_hash *hash);
250 static void uma_timeout(void *);
251 static void uma_startup3(void);
252 static void *zone_alloc_item(uma_zone_t, void *, int, int);
253 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip);
254 static void bucket_enable(void);
255 static void bucket_init(void);
256 static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int);
257 static void bucket_free(uma_zone_t zone, uma_bucket_t, void *);
258 static void bucket_zone_drain(void);
259 static uma_bucket_t zone_alloc_bucket(uma_zone_t, void *, int, int);
260 static uma_slab_t zone_fetch_slab(uma_zone_t, uma_keg_t, int, int);
261 static uma_slab_t zone_fetch_slab_multi(uma_zone_t, uma_keg_t, int, int);
262 static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab);
263 static void slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item);
264 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit,
265     uma_fini fini, int align, uint32_t flags);
266 static int zone_import(uma_zone_t, void **, int, int, int);
267 static void zone_release(uma_zone_t, void **, int);
268 static void uma_zero_item(void *, uma_zone_t);
269 
270 void uma_print_zone(uma_zone_t);
271 void uma_print_stats(void);
272 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS);
273 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS);
274 
275 #ifdef INVARIANTS
276 static void uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item);
277 static void uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item);
278 #endif
279 
280 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
281 
282 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLTYPE_INT,
283     0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones");
284 
285 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLTYPE_STRUCT,
286     0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats");
287 
288 static int zone_warnings = 1;
289 SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RWTUN, &zone_warnings, 0,
290     "Warn when UMA zones becomes full");
291 
292 /* Adjust bytes under management by UMA. */
293 static inline void
294 uma_total_dec(unsigned long size)
295 {
296 
297 	atomic_subtract_long(&uma_kmem_total, size);
298 }
299 
300 static inline void
301 uma_total_inc(unsigned long size)
302 {
303 
304 	if (atomic_fetchadd_long(&uma_kmem_total, size) > uma_kmem_limit)
305 		uma_reclaim_wakeup();
306 }
307 
308 /*
309  * This routine checks to see whether or not it's safe to enable buckets.
310  */
311 static void
312 bucket_enable(void)
313 {
314 	bucketdisable = vm_page_count_min();
315 }
316 
317 /*
318  * Initialize bucket_zones, the array of zones of buckets of various sizes.
319  *
320  * For each zone, calculate the memory required for each bucket, consisting
321  * of the header and an array of pointers.
322  */
323 static void
324 bucket_init(void)
325 {
326 	struct uma_bucket_zone *ubz;
327 	int size;
328 
329 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) {
330 		size = roundup(sizeof(struct uma_bucket), sizeof(void *));
331 		size += sizeof(void *) * ubz->ubz_entries;
332 		ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
333 		    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
334 		    UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET | UMA_ZONE_NUMA);
335 	}
336 }
337 
338 /*
339  * Given a desired number of entries for a bucket, return the zone from which
340  * to allocate the bucket.
341  */
342 static struct uma_bucket_zone *
343 bucket_zone_lookup(int entries)
344 {
345 	struct uma_bucket_zone *ubz;
346 
347 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
348 		if (ubz->ubz_entries >= entries)
349 			return (ubz);
350 	ubz--;
351 	return (ubz);
352 }
353 
354 static int
355 bucket_select(int size)
356 {
357 	struct uma_bucket_zone *ubz;
358 
359 	ubz = &bucket_zones[0];
360 	if (size > ubz->ubz_maxsize)
361 		return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1);
362 
363 	for (; ubz->ubz_entries != 0; ubz++)
364 		if (ubz->ubz_maxsize < size)
365 			break;
366 	ubz--;
367 	return (ubz->ubz_entries);
368 }
369 
370 static uma_bucket_t
371 bucket_alloc(uma_zone_t zone, void *udata, int flags)
372 {
373 	struct uma_bucket_zone *ubz;
374 	uma_bucket_t bucket;
375 
376 	/*
377 	 * This is to stop us from allocating per cpu buckets while we're
378 	 * running out of vm.boot_pages.  Otherwise, we would exhaust the
379 	 * boot pages.  This also prevents us from allocating buckets in
380 	 * low memory situations.
381 	 */
382 	if (bucketdisable)
383 		return (NULL);
384 	/*
385 	 * To limit bucket recursion we store the original zone flags
386 	 * in a cookie passed via zalloc_arg/zfree_arg.  This allows the
387 	 * NOVM flag to persist even through deep recursions.  We also
388 	 * store ZFLAG_BUCKET once we have recursed attempting to allocate
389 	 * a bucket for a bucket zone so we do not allow infinite bucket
390 	 * recursion.  This cookie will even persist to frees of unused
391 	 * buckets via the allocation path or bucket allocations in the
392 	 * free path.
393 	 */
394 	if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
395 		udata = (void *)(uintptr_t)zone->uz_flags;
396 	else {
397 		if ((uintptr_t)udata & UMA_ZFLAG_BUCKET)
398 			return (NULL);
399 		udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET);
400 	}
401 	if ((uintptr_t)udata & UMA_ZFLAG_CACHEONLY)
402 		flags |= M_NOVM;
403 	ubz = bucket_zone_lookup(zone->uz_count);
404 	if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0)
405 		ubz++;
406 	bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags);
407 	if (bucket) {
408 #ifdef INVARIANTS
409 		bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
410 #endif
411 		bucket->ub_cnt = 0;
412 		bucket->ub_entries = ubz->ubz_entries;
413 	}
414 
415 	return (bucket);
416 }
417 
418 static void
419 bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata)
420 {
421 	struct uma_bucket_zone *ubz;
422 
423 	KASSERT(bucket->ub_cnt == 0,
424 	    ("bucket_free: Freeing a non free bucket."));
425 	if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
426 		udata = (void *)(uintptr_t)zone->uz_flags;
427 	ubz = bucket_zone_lookup(bucket->ub_entries);
428 	uma_zfree_arg(ubz->ubz_zone, bucket, udata);
429 }
430 
431 static void
432 bucket_zone_drain(void)
433 {
434 	struct uma_bucket_zone *ubz;
435 
436 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
437 		zone_drain(ubz->ubz_zone);
438 }
439 
440 static void
441 zone_log_warning(uma_zone_t zone)
442 {
443 	static const struct timeval warninterval = { 300, 0 };
444 
445 	if (!zone_warnings || zone->uz_warning == NULL)
446 		return;
447 
448 	if (ratecheck(&zone->uz_ratecheck, &warninterval))
449 		printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning);
450 }
451 
452 static inline void
453 zone_maxaction(uma_zone_t zone)
454 {
455 
456 	if (zone->uz_maxaction.ta_func != NULL)
457 		taskqueue_enqueue(taskqueue_thread, &zone->uz_maxaction);
458 }
459 
460 static void
461 zone_foreach_keg(uma_zone_t zone, void (*kegfn)(uma_keg_t))
462 {
463 	uma_klink_t klink;
464 
465 	LIST_FOREACH(klink, &zone->uz_kegs, kl_link)
466 		kegfn(klink->kl_keg);
467 }
468 
469 /*
470  * Routine called by timeout which is used to fire off some time interval
471  * based calculations.  (stats, hash size, etc.)
472  *
473  * Arguments:
474  *	arg   Unused
475  *
476  * Returns:
477  *	Nothing
478  */
479 static void
480 uma_timeout(void *unused)
481 {
482 	bucket_enable();
483 	zone_foreach(zone_timeout);
484 
485 	/* Reschedule this event */
486 	callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
487 }
488 
489 /*
490  * Routine to perform timeout driven calculations.  This expands the
491  * hashes and does per cpu statistics aggregation.
492  *
493  *  Returns nothing.
494  */
495 static void
496 keg_timeout(uma_keg_t keg)
497 {
498 
499 	KEG_LOCK(keg);
500 	/*
501 	 * Expand the keg hash table.
502 	 *
503 	 * This is done if the number of slabs is larger than the hash size.
504 	 * What I'm trying to do here is completely reduce collisions.  This
505 	 * may be a little aggressive.  Should I allow for two collisions max?
506 	 */
507 	if (keg->uk_flags & UMA_ZONE_HASH &&
508 	    keg->uk_pages / keg->uk_ppera >= keg->uk_hash.uh_hashsize) {
509 		struct uma_hash newhash;
510 		struct uma_hash oldhash;
511 		int ret;
512 
513 		/*
514 		 * This is so involved because allocating and freeing
515 		 * while the keg lock is held will lead to deadlock.
516 		 * I have to do everything in stages and check for
517 		 * races.
518 		 */
519 		newhash = keg->uk_hash;
520 		KEG_UNLOCK(keg);
521 		ret = hash_alloc(&newhash);
522 		KEG_LOCK(keg);
523 		if (ret) {
524 			if (hash_expand(&keg->uk_hash, &newhash)) {
525 				oldhash = keg->uk_hash;
526 				keg->uk_hash = newhash;
527 			} else
528 				oldhash = newhash;
529 
530 			KEG_UNLOCK(keg);
531 			hash_free(&oldhash);
532 			return;
533 		}
534 	}
535 	KEG_UNLOCK(keg);
536 }
537 
538 static void
539 zone_timeout(uma_zone_t zone)
540 {
541 
542 	zone_foreach_keg(zone, &keg_timeout);
543 }
544 
545 /*
546  * Allocate and zero fill the next sized hash table from the appropriate
547  * backing store.
548  *
549  * Arguments:
550  *	hash  A new hash structure with the old hash size in uh_hashsize
551  *
552  * Returns:
553  *	1 on success and 0 on failure.
554  */
555 static int
556 hash_alloc(struct uma_hash *hash)
557 {
558 	int oldsize;
559 	int alloc;
560 
561 	oldsize = hash->uh_hashsize;
562 
563 	/* We're just going to go to a power of two greater */
564 	if (oldsize)  {
565 		hash->uh_hashsize = oldsize * 2;
566 		alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
567 		hash->uh_slab_hash = (struct slabhead *)malloc(alloc,
568 		    M_UMAHASH, M_NOWAIT);
569 	} else {
570 		alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
571 		hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
572 		    UMA_ANYDOMAIN, M_WAITOK);
573 		hash->uh_hashsize = UMA_HASH_SIZE_INIT;
574 	}
575 	if (hash->uh_slab_hash) {
576 		bzero(hash->uh_slab_hash, alloc);
577 		hash->uh_hashmask = hash->uh_hashsize - 1;
578 		return (1);
579 	}
580 
581 	return (0);
582 }
583 
584 /*
585  * Expands the hash table for HASH zones.  This is done from zone_timeout
586  * to reduce collisions.  This must not be done in the regular allocation
587  * path, otherwise, we can recurse on the vm while allocating pages.
588  *
589  * Arguments:
590  *	oldhash  The hash you want to expand
591  *	newhash  The hash structure for the new table
592  *
593  * Returns:
594  *	Nothing
595  *
596  * Discussion:
597  */
598 static int
599 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
600 {
601 	uma_slab_t slab;
602 	int hval;
603 	int i;
604 
605 	if (!newhash->uh_slab_hash)
606 		return (0);
607 
608 	if (oldhash->uh_hashsize >= newhash->uh_hashsize)
609 		return (0);
610 
611 	/*
612 	 * I need to investigate hash algorithms for resizing without a
613 	 * full rehash.
614 	 */
615 
616 	for (i = 0; i < oldhash->uh_hashsize; i++)
617 		while (!SLIST_EMPTY(&oldhash->uh_slab_hash[i])) {
618 			slab = SLIST_FIRST(&oldhash->uh_slab_hash[i]);
619 			SLIST_REMOVE_HEAD(&oldhash->uh_slab_hash[i], us_hlink);
620 			hval = UMA_HASH(newhash, slab->us_data);
621 			SLIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
622 			    slab, us_hlink);
623 		}
624 
625 	return (1);
626 }
627 
628 /*
629  * Free the hash bucket to the appropriate backing store.
630  *
631  * Arguments:
632  *	slab_hash  The hash bucket we're freeing
633  *	hashsize   The number of entries in that hash bucket
634  *
635  * Returns:
636  *	Nothing
637  */
638 static void
639 hash_free(struct uma_hash *hash)
640 {
641 	if (hash->uh_slab_hash == NULL)
642 		return;
643 	if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
644 		zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE);
645 	else
646 		free(hash->uh_slab_hash, M_UMAHASH);
647 }
648 
649 /*
650  * Frees all outstanding items in a bucket
651  *
652  * Arguments:
653  *	zone   The zone to free to, must be unlocked.
654  *	bucket The free/alloc bucket with items, cpu queue must be locked.
655  *
656  * Returns:
657  *	Nothing
658  */
659 
660 static void
661 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
662 {
663 	int i;
664 
665 	if (bucket == NULL)
666 		return;
667 
668 	if (zone->uz_fini)
669 		for (i = 0; i < bucket->ub_cnt; i++)
670 			zone->uz_fini(bucket->ub_bucket[i], zone->uz_size);
671 	zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt);
672 	bucket->ub_cnt = 0;
673 }
674 
675 /*
676  * Drains the per cpu caches for a zone.
677  *
678  * NOTE: This may only be called while the zone is being turn down, and not
679  * during normal operation.  This is necessary in order that we do not have
680  * to migrate CPUs to drain the per-CPU caches.
681  *
682  * Arguments:
683  *	zone     The zone to drain, must be unlocked.
684  *
685  * Returns:
686  *	Nothing
687  */
688 static void
689 cache_drain(uma_zone_t zone)
690 {
691 	uma_cache_t cache;
692 	int cpu;
693 
694 	/*
695 	 * XXX: It is safe to not lock the per-CPU caches, because we're
696 	 * tearing down the zone anyway.  I.e., there will be no further use
697 	 * of the caches at this point.
698 	 *
699 	 * XXX: It would good to be able to assert that the zone is being
700 	 * torn down to prevent improper use of cache_drain().
701 	 *
702 	 * XXX: We lock the zone before passing into bucket_cache_drain() as
703 	 * it is used elsewhere.  Should the tear-down path be made special
704 	 * there in some form?
705 	 */
706 	CPU_FOREACH(cpu) {
707 		cache = &zone->uz_cpu[cpu];
708 		bucket_drain(zone, cache->uc_allocbucket);
709 		bucket_drain(zone, cache->uc_freebucket);
710 		if (cache->uc_allocbucket != NULL)
711 			bucket_free(zone, cache->uc_allocbucket, NULL);
712 		if (cache->uc_freebucket != NULL)
713 			bucket_free(zone, cache->uc_freebucket, NULL);
714 		cache->uc_allocbucket = cache->uc_freebucket = NULL;
715 	}
716 	ZONE_LOCK(zone);
717 	bucket_cache_drain(zone);
718 	ZONE_UNLOCK(zone);
719 }
720 
721 static void
722 cache_shrink(uma_zone_t zone)
723 {
724 
725 	if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
726 		return;
727 
728 	ZONE_LOCK(zone);
729 	zone->uz_count = (zone->uz_count_min + zone->uz_count) / 2;
730 	ZONE_UNLOCK(zone);
731 }
732 
733 static void
734 cache_drain_safe_cpu(uma_zone_t zone)
735 {
736 	uma_cache_t cache;
737 	uma_bucket_t b1, b2;
738 	int domain;
739 
740 	if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
741 		return;
742 
743 	b1 = b2 = NULL;
744 	ZONE_LOCK(zone);
745 	critical_enter();
746 	if (zone->uz_flags & UMA_ZONE_NUMA)
747 		domain = PCPU_GET(domain);
748 	else
749 		domain = 0;
750 	cache = &zone->uz_cpu[curcpu];
751 	if (cache->uc_allocbucket) {
752 		if (cache->uc_allocbucket->ub_cnt != 0)
753 			LIST_INSERT_HEAD(&zone->uz_domain[domain].uzd_buckets,
754 			    cache->uc_allocbucket, ub_link);
755 		else
756 			b1 = cache->uc_allocbucket;
757 		cache->uc_allocbucket = NULL;
758 	}
759 	if (cache->uc_freebucket) {
760 		if (cache->uc_freebucket->ub_cnt != 0)
761 			LIST_INSERT_HEAD(&zone->uz_domain[domain].uzd_buckets,
762 			    cache->uc_freebucket, ub_link);
763 		else
764 			b2 = cache->uc_freebucket;
765 		cache->uc_freebucket = NULL;
766 	}
767 	critical_exit();
768 	ZONE_UNLOCK(zone);
769 	if (b1)
770 		bucket_free(zone, b1, NULL);
771 	if (b2)
772 		bucket_free(zone, b2, NULL);
773 }
774 
775 /*
776  * Safely drain per-CPU caches of a zone(s) to alloc bucket.
777  * This is an expensive call because it needs to bind to all CPUs
778  * one by one and enter a critical section on each of them in order
779  * to safely access their cache buckets.
780  * Zone lock must not be held on call this function.
781  */
782 static void
783 cache_drain_safe(uma_zone_t zone)
784 {
785 	int cpu;
786 
787 	/*
788 	 * Polite bucket sizes shrinking was not enouth, shrink aggressively.
789 	 */
790 	if (zone)
791 		cache_shrink(zone);
792 	else
793 		zone_foreach(cache_shrink);
794 
795 	CPU_FOREACH(cpu) {
796 		thread_lock(curthread);
797 		sched_bind(curthread, cpu);
798 		thread_unlock(curthread);
799 
800 		if (zone)
801 			cache_drain_safe_cpu(zone);
802 		else
803 			zone_foreach(cache_drain_safe_cpu);
804 	}
805 	thread_lock(curthread);
806 	sched_unbind(curthread);
807 	thread_unlock(curthread);
808 }
809 
810 /*
811  * Drain the cached buckets from a zone.  Expects a locked zone on entry.
812  */
813 static void
814 bucket_cache_drain(uma_zone_t zone)
815 {
816 	uma_zone_domain_t zdom;
817 	uma_bucket_t bucket;
818 	int i;
819 
820 	/*
821 	 * Drain the bucket queues and free the buckets.
822 	 */
823 	for (i = 0; i < vm_ndomains; i++) {
824 		zdom = &zone->uz_domain[i];
825 		while ((bucket = LIST_FIRST(&zdom->uzd_buckets)) != NULL) {
826 			LIST_REMOVE(bucket, ub_link);
827 			ZONE_UNLOCK(zone);
828 			bucket_drain(zone, bucket);
829 			bucket_free(zone, bucket, NULL);
830 			ZONE_LOCK(zone);
831 		}
832 	}
833 
834 	/*
835 	 * Shrink further bucket sizes.  Price of single zone lock collision
836 	 * is probably lower then price of global cache drain.
837 	 */
838 	if (zone->uz_count > zone->uz_count_min)
839 		zone->uz_count--;
840 }
841 
842 static void
843 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start)
844 {
845 	uint8_t *mem;
846 	int i;
847 	uint8_t flags;
848 
849 	CTR4(KTR_UMA, "keg_free_slab keg %s(%p) slab %p, returning %d bytes",
850 	    keg->uk_name, keg, slab, PAGE_SIZE * keg->uk_ppera);
851 
852 	mem = slab->us_data;
853 	flags = slab->us_flags;
854 	i = start;
855 	if (keg->uk_fini != NULL) {
856 		for (i--; i > -1; i--)
857 			keg->uk_fini(slab->us_data + (keg->uk_rsize * i),
858 			    keg->uk_size);
859 	}
860 	if (keg->uk_flags & UMA_ZONE_OFFPAGE)
861 		zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE);
862 	keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags);
863 	uma_total_dec(PAGE_SIZE * keg->uk_ppera);
864 }
865 
866 /*
867  * Frees pages from a keg back to the system.  This is done on demand from
868  * the pageout daemon.
869  *
870  * Returns nothing.
871  */
872 static void
873 keg_drain(uma_keg_t keg)
874 {
875 	struct slabhead freeslabs = { 0 };
876 	uma_domain_t dom;
877 	uma_slab_t slab, tmp;
878 	int i;
879 
880 	/*
881 	 * We don't want to take pages from statically allocated kegs at this
882 	 * time
883 	 */
884 	if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL)
885 		return;
886 
887 	CTR3(KTR_UMA, "keg_drain %s(%p) free items: %u",
888 	    keg->uk_name, keg, keg->uk_free);
889 	KEG_LOCK(keg);
890 	if (keg->uk_free == 0)
891 		goto finished;
892 
893 	for (i = 0; i < vm_ndomains; i++) {
894 		dom = &keg->uk_domain[i];
895 		LIST_FOREACH_SAFE(slab, &dom->ud_free_slab, us_link, tmp) {
896 			/* We have nowhere to free these to. */
897 			if (slab->us_flags & UMA_SLAB_BOOT)
898 				continue;
899 
900 			LIST_REMOVE(slab, us_link);
901 			keg->uk_pages -= keg->uk_ppera;
902 			keg->uk_free -= keg->uk_ipers;
903 
904 			if (keg->uk_flags & UMA_ZONE_HASH)
905 				UMA_HASH_REMOVE(&keg->uk_hash, slab,
906 				    slab->us_data);
907 
908 			SLIST_INSERT_HEAD(&freeslabs, slab, us_hlink);
909 		}
910 	}
911 
912 finished:
913 	KEG_UNLOCK(keg);
914 
915 	while ((slab = SLIST_FIRST(&freeslabs)) != NULL) {
916 		SLIST_REMOVE(&freeslabs, slab, uma_slab, us_hlink);
917 		keg_free_slab(keg, slab, keg->uk_ipers);
918 	}
919 }
920 
921 static void
922 zone_drain_wait(uma_zone_t zone, int waitok)
923 {
924 
925 	/*
926 	 * Set draining to interlock with zone_dtor() so we can release our
927 	 * locks as we go.  Only dtor() should do a WAITOK call since it
928 	 * is the only call that knows the structure will still be available
929 	 * when it wakes up.
930 	 */
931 	ZONE_LOCK(zone);
932 	while (zone->uz_flags & UMA_ZFLAG_DRAINING) {
933 		if (waitok == M_NOWAIT)
934 			goto out;
935 		msleep(zone, zone->uz_lockptr, PVM, "zonedrain", 1);
936 	}
937 	zone->uz_flags |= UMA_ZFLAG_DRAINING;
938 	bucket_cache_drain(zone);
939 	ZONE_UNLOCK(zone);
940 	/*
941 	 * The DRAINING flag protects us from being freed while
942 	 * we're running.  Normally the uma_rwlock would protect us but we
943 	 * must be able to release and acquire the right lock for each keg.
944 	 */
945 	zone_foreach_keg(zone, &keg_drain);
946 	ZONE_LOCK(zone);
947 	zone->uz_flags &= ~UMA_ZFLAG_DRAINING;
948 	wakeup(zone);
949 out:
950 	ZONE_UNLOCK(zone);
951 }
952 
953 void
954 zone_drain(uma_zone_t zone)
955 {
956 
957 	zone_drain_wait(zone, M_NOWAIT);
958 }
959 
960 /*
961  * Allocate a new slab for a keg.  This does not insert the slab onto a list.
962  *
963  * Arguments:
964  *	wait  Shall we wait?
965  *
966  * Returns:
967  *	The slab that was allocated or NULL if there is no memory and the
968  *	caller specified M_NOWAIT.
969  */
970 static uma_slab_t
971 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int domain, int wait)
972 {
973 	uma_alloc allocf;
974 	uma_slab_t slab;
975 	unsigned long size;
976 	uint8_t *mem;
977 	uint8_t flags;
978 	int i;
979 
980 	KASSERT(domain >= 0 && domain < vm_ndomains,
981 	    ("keg_alloc_slab: domain %d out of range", domain));
982 	mtx_assert(&keg->uk_lock, MA_OWNED);
983 	slab = NULL;
984 	mem = NULL;
985 
986 	allocf = keg->uk_allocf;
987 	KEG_UNLOCK(keg);
988 	size = keg->uk_ppera * PAGE_SIZE;
989 
990 	if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
991 		slab = zone_alloc_item(keg->uk_slabzone, NULL, domain, wait);
992 		if (slab == NULL)
993 			goto out;
994 	}
995 
996 	/*
997 	 * This reproduces the old vm_zone behavior of zero filling pages the
998 	 * first time they are added to a zone.
999 	 *
1000 	 * Malloced items are zeroed in uma_zalloc.
1001 	 */
1002 
1003 	if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
1004 		wait |= M_ZERO;
1005 	else
1006 		wait &= ~M_ZERO;
1007 
1008 	if (keg->uk_flags & UMA_ZONE_NODUMP)
1009 		wait |= M_NODUMP;
1010 
1011 	/* zone is passed for legacy reasons. */
1012 	mem = allocf(zone, size, domain, &flags, wait);
1013 	if (mem == NULL) {
1014 		if (keg->uk_flags & UMA_ZONE_OFFPAGE)
1015 			zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE);
1016 		slab = NULL;
1017 		goto out;
1018 	}
1019 	uma_total_inc(size);
1020 
1021 	/* Point the slab into the allocated memory */
1022 	if (!(keg->uk_flags & UMA_ZONE_OFFPAGE))
1023 		slab = (uma_slab_t )(mem + keg->uk_pgoff);
1024 
1025 	if (keg->uk_flags & UMA_ZONE_VTOSLAB)
1026 		for (i = 0; i < keg->uk_ppera; i++)
1027 			vsetslab((vm_offset_t)mem + (i * PAGE_SIZE), slab);
1028 
1029 	slab->us_keg = keg;
1030 	slab->us_data = mem;
1031 	slab->us_freecount = keg->uk_ipers;
1032 	slab->us_flags = flags;
1033 	slab->us_domain = domain;
1034 	BIT_FILL(SLAB_SETSIZE, &slab->us_free);
1035 #ifdef INVARIANTS
1036 	BIT_ZERO(SLAB_SETSIZE, &slab->us_debugfree);
1037 #endif
1038 
1039 	if (keg->uk_init != NULL) {
1040 		for (i = 0; i < keg->uk_ipers; i++)
1041 			if (keg->uk_init(slab->us_data + (keg->uk_rsize * i),
1042 			    keg->uk_size, wait) != 0)
1043 				break;
1044 		if (i != keg->uk_ipers) {
1045 			keg_free_slab(keg, slab, i);
1046 			slab = NULL;
1047 			goto out;
1048 		}
1049 	}
1050 out:
1051 	KEG_LOCK(keg);
1052 
1053 	CTR3(KTR_UMA, "keg_alloc_slab: allocated slab %p for %s(%p)",
1054 	    slab, keg->uk_name, keg);
1055 
1056 	if (slab != NULL) {
1057 		if (keg->uk_flags & UMA_ZONE_HASH)
1058 			UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
1059 
1060 		keg->uk_pages += keg->uk_ppera;
1061 		keg->uk_free += keg->uk_ipers;
1062 	}
1063 
1064 	return (slab);
1065 }
1066 
1067 /*
1068  * This function is intended to be used early on in place of page_alloc() so
1069  * that we may use the boot time page cache to satisfy allocations before
1070  * the VM is ready.
1071  */
1072 static void *
1073 startup_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1074     int wait)
1075 {
1076 	uma_keg_t keg;
1077 	void *mem;
1078 	int pages;
1079 
1080 	keg = zone_first_keg(zone);
1081 
1082 	/*
1083 	 * If we are in BOOT_BUCKETS or higher, than switch to real
1084 	 * allocator.  Zones with page sized slabs switch at BOOT_PAGEALLOC.
1085 	 */
1086 	switch (booted) {
1087 		case BOOT_COLD:
1088 		case BOOT_STRAPPED:
1089 			break;
1090 		case BOOT_PAGEALLOC:
1091 			if (keg->uk_ppera > 1)
1092 				break;
1093 		case BOOT_BUCKETS:
1094 		case BOOT_RUNNING:
1095 #ifdef UMA_MD_SMALL_ALLOC
1096 			keg->uk_allocf = (keg->uk_ppera > 1) ?
1097 			    page_alloc : uma_small_alloc;
1098 #else
1099 			keg->uk_allocf = page_alloc;
1100 #endif
1101 			return keg->uk_allocf(zone, bytes, domain, pflag, wait);
1102 	}
1103 
1104 	/*
1105 	 * Check our small startup cache to see if it has pages remaining.
1106 	 */
1107 	pages = howmany(bytes, PAGE_SIZE);
1108 	KASSERT(pages > 0, ("%s can't reserve 0 pages", __func__));
1109 	if (pages > boot_pages)
1110 		panic("UMA zone \"%s\": Increase vm.boot_pages", zone->uz_name);
1111 #ifdef DIAGNOSTIC
1112 	printf("%s from \"%s\", %d boot pages left\n", __func__, zone->uz_name,
1113 	    boot_pages);
1114 #endif
1115 	mem = bootmem;
1116 	boot_pages -= pages;
1117 	bootmem += pages * PAGE_SIZE;
1118 	*pflag = UMA_SLAB_BOOT;
1119 
1120 	return (mem);
1121 }
1122 
1123 /*
1124  * Allocates a number of pages from the system
1125  *
1126  * Arguments:
1127  *	bytes  The number of bytes requested
1128  *	wait  Shall we wait?
1129  *
1130  * Returns:
1131  *	A pointer to the alloced memory or possibly
1132  *	NULL if M_NOWAIT is set.
1133  */
1134 static void *
1135 page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1136     int wait)
1137 {
1138 	void *p;	/* Returned page */
1139 
1140 	*pflag = UMA_SLAB_KERNEL;
1141 	p = (void *) kmem_malloc_domain(domain, bytes, wait);
1142 
1143 	return (p);
1144 }
1145 
1146 /*
1147  * Allocates a number of pages from within an object
1148  *
1149  * Arguments:
1150  *	bytes  The number of bytes requested
1151  *	wait   Shall we wait?
1152  *
1153  * Returns:
1154  *	A pointer to the alloced memory or possibly
1155  *	NULL if M_NOWAIT is set.
1156  */
1157 static void *
1158 noobj_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags,
1159     int wait)
1160 {
1161 	TAILQ_HEAD(, vm_page) alloctail;
1162 	u_long npages;
1163 	vm_offset_t retkva, zkva;
1164 	vm_page_t p, p_next;
1165 	uma_keg_t keg;
1166 
1167 	TAILQ_INIT(&alloctail);
1168 	keg = zone_first_keg(zone);
1169 
1170 	npages = howmany(bytes, PAGE_SIZE);
1171 	while (npages > 0) {
1172 		p = vm_page_alloc_domain(NULL, 0, domain, VM_ALLOC_INTERRUPT |
1173 		    VM_ALLOC_WIRED | VM_ALLOC_NOOBJ |
1174 		    ((wait & M_WAITOK) != 0 ? VM_ALLOC_WAITOK :
1175 		    VM_ALLOC_NOWAIT));
1176 		if (p != NULL) {
1177 			/*
1178 			 * Since the page does not belong to an object, its
1179 			 * listq is unused.
1180 			 */
1181 			TAILQ_INSERT_TAIL(&alloctail, p, listq);
1182 			npages--;
1183 			continue;
1184 		}
1185 		/*
1186 		 * Page allocation failed, free intermediate pages and
1187 		 * exit.
1188 		 */
1189 		TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1190 			vm_page_unwire(p, PQ_NONE);
1191 			vm_page_free(p);
1192 		}
1193 		return (NULL);
1194 	}
1195 	*flags = UMA_SLAB_PRIV;
1196 	zkva = keg->uk_kva +
1197 	    atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
1198 	retkva = zkva;
1199 	TAILQ_FOREACH(p, &alloctail, listq) {
1200 		pmap_qenter(zkva, &p, 1);
1201 		zkva += PAGE_SIZE;
1202 	}
1203 
1204 	return ((void *)retkva);
1205 }
1206 
1207 /*
1208  * Frees a number of pages to the system
1209  *
1210  * Arguments:
1211  *	mem   A pointer to the memory to be freed
1212  *	size  The size of the memory being freed
1213  *	flags The original p->us_flags field
1214  *
1215  * Returns:
1216  *	Nothing
1217  */
1218 static void
1219 page_free(void *mem, vm_size_t size, uint8_t flags)
1220 {
1221 	struct vmem *vmem;
1222 
1223 	if (flags & UMA_SLAB_KERNEL)
1224 		vmem = kernel_arena;
1225 	else
1226 		panic("UMA: page_free used with invalid flags %x", flags);
1227 
1228 	kmem_free(vmem, (vm_offset_t)mem, size);
1229 }
1230 
1231 /*
1232  * Zero fill initializer
1233  *
1234  * Arguments/Returns follow uma_init specifications
1235  */
1236 static int
1237 zero_init(void *mem, int size, int flags)
1238 {
1239 	bzero(mem, size);
1240 	return (0);
1241 }
1242 
1243 /*
1244  * Finish creating a small uma keg.  This calculates ipers, and the keg size.
1245  *
1246  * Arguments
1247  *	keg  The zone we should initialize
1248  *
1249  * Returns
1250  *	Nothing
1251  */
1252 static void
1253 keg_small_init(uma_keg_t keg)
1254 {
1255 	u_int rsize;
1256 	u_int memused;
1257 	u_int wastedspace;
1258 	u_int shsize;
1259 	u_int slabsize;
1260 
1261 	if (keg->uk_flags & UMA_ZONE_PCPU) {
1262 		u_int ncpus = (mp_maxid + 1) ? (mp_maxid + 1) : MAXCPU;
1263 
1264 		slabsize = sizeof(struct pcpu);
1265 		keg->uk_ppera = howmany(ncpus * sizeof(struct pcpu),
1266 		    PAGE_SIZE);
1267 	} else {
1268 		slabsize = UMA_SLAB_SIZE;
1269 		keg->uk_ppera = 1;
1270 	}
1271 
1272 	/*
1273 	 * Calculate the size of each allocation (rsize) according to
1274 	 * alignment.  If the requested size is smaller than we have
1275 	 * allocation bits for we round it up.
1276 	 */
1277 	rsize = keg->uk_size;
1278 	if (rsize < slabsize / SLAB_SETSIZE)
1279 		rsize = slabsize / SLAB_SETSIZE;
1280 	if (rsize & keg->uk_align)
1281 		rsize = (rsize & ~keg->uk_align) + (keg->uk_align + 1);
1282 	keg->uk_rsize = rsize;
1283 
1284 	KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
1285 	    keg->uk_rsize < sizeof(struct pcpu),
1286 	    ("%s: size %u too large", __func__, keg->uk_rsize));
1287 
1288 	if (keg->uk_flags & UMA_ZONE_OFFPAGE)
1289 		shsize = 0;
1290 	else
1291 		shsize = sizeof(struct uma_slab);
1292 
1293 	if (rsize <= slabsize - shsize)
1294 		keg->uk_ipers = (slabsize - shsize) / rsize;
1295 	else {
1296 		/* Handle special case when we have 1 item per slab, so
1297 		 * alignment requirement can be relaxed. */
1298 		KASSERT(keg->uk_size <= slabsize - shsize,
1299 		    ("%s: size %u greater than slab", __func__, keg->uk_size));
1300 		keg->uk_ipers = 1;
1301 	}
1302 	KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE,
1303 	    ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers));
1304 
1305 	memused = keg->uk_ipers * rsize + shsize;
1306 	wastedspace = slabsize - memused;
1307 
1308 	/*
1309 	 * We can't do OFFPAGE if we're internal or if we've been
1310 	 * asked to not go to the VM for buckets.  If we do this we
1311 	 * may end up going to the VM  for slabs which we do not
1312 	 * want to do if we're UMA_ZFLAG_CACHEONLY as a result
1313 	 * of UMA_ZONE_VM, which clearly forbids it.
1314 	 */
1315 	if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) ||
1316 	    (keg->uk_flags & UMA_ZFLAG_CACHEONLY))
1317 		return;
1318 
1319 	/*
1320 	 * See if using an OFFPAGE slab will limit our waste.  Only do
1321 	 * this if it permits more items per-slab.
1322 	 *
1323 	 * XXX We could try growing slabsize to limit max waste as well.
1324 	 * Historically this was not done because the VM could not
1325 	 * efficiently handle contiguous allocations.
1326 	 */
1327 	if ((wastedspace >= slabsize / UMA_MAX_WASTE) &&
1328 	    (keg->uk_ipers < (slabsize / keg->uk_rsize))) {
1329 		keg->uk_ipers = slabsize / keg->uk_rsize;
1330 		KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE,
1331 		    ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers));
1332 		CTR6(KTR_UMA, "UMA decided we need offpage slab headers for "
1333 		    "keg: %s(%p), calculated wastedspace = %d, "
1334 		    "maximum wasted space allowed = %d, "
1335 		    "calculated ipers = %d, "
1336 		    "new wasted space = %d\n", keg->uk_name, keg, wastedspace,
1337 		    slabsize / UMA_MAX_WASTE, keg->uk_ipers,
1338 		    slabsize - keg->uk_ipers * keg->uk_rsize);
1339 		keg->uk_flags |= UMA_ZONE_OFFPAGE;
1340 	}
1341 
1342 	if ((keg->uk_flags & UMA_ZONE_OFFPAGE) &&
1343 	    (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1344 		keg->uk_flags |= UMA_ZONE_HASH;
1345 }
1346 
1347 /*
1348  * Finish creating a large (> UMA_SLAB_SIZE) uma kegs.  Just give in and do
1349  * OFFPAGE for now.  When I can allow for more dynamic slab sizes this will be
1350  * more complicated.
1351  *
1352  * Arguments
1353  *	keg  The keg we should initialize
1354  *
1355  * Returns
1356  *	Nothing
1357  */
1358 static void
1359 keg_large_init(uma_keg_t keg)
1360 {
1361 	u_int shsize;
1362 
1363 	KASSERT(keg != NULL, ("Keg is null in keg_large_init"));
1364 	KASSERT((keg->uk_flags & UMA_ZFLAG_CACHEONLY) == 0,
1365 	    ("keg_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY keg"));
1366 	KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0,
1367 	    ("%s: Cannot large-init a UMA_ZONE_PCPU keg", __func__));
1368 
1369 	keg->uk_ppera = howmany(keg->uk_size, PAGE_SIZE);
1370 	keg->uk_ipers = 1;
1371 	keg->uk_rsize = keg->uk_size;
1372 
1373 	/* Check whether we have enough space to not do OFFPAGE. */
1374 	if ((keg->uk_flags & UMA_ZONE_OFFPAGE) == 0) {
1375 		shsize = sizeof(struct uma_slab);
1376 		if (shsize & UMA_ALIGN_PTR)
1377 			shsize = (shsize & ~UMA_ALIGN_PTR) +
1378 			    (UMA_ALIGN_PTR + 1);
1379 
1380 		if (PAGE_SIZE * keg->uk_ppera - keg->uk_rsize < shsize) {
1381 			/*
1382 			 * We can't do OFFPAGE if we're internal, in which case
1383 			 * we need an extra page per allocation to contain the
1384 			 * slab header.
1385 			 */
1386 			if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) == 0)
1387 				keg->uk_flags |= UMA_ZONE_OFFPAGE;
1388 			else
1389 				keg->uk_ppera++;
1390 		}
1391 	}
1392 
1393 	if ((keg->uk_flags & UMA_ZONE_OFFPAGE) &&
1394 	    (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1395 		keg->uk_flags |= UMA_ZONE_HASH;
1396 }
1397 
1398 static void
1399 keg_cachespread_init(uma_keg_t keg)
1400 {
1401 	int alignsize;
1402 	int trailer;
1403 	int pages;
1404 	int rsize;
1405 
1406 	KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0,
1407 	    ("%s: Cannot cachespread-init a UMA_ZONE_PCPU keg", __func__));
1408 
1409 	alignsize = keg->uk_align + 1;
1410 	rsize = keg->uk_size;
1411 	/*
1412 	 * We want one item to start on every align boundary in a page.  To
1413 	 * do this we will span pages.  We will also extend the item by the
1414 	 * size of align if it is an even multiple of align.  Otherwise, it
1415 	 * would fall on the same boundary every time.
1416 	 */
1417 	if (rsize & keg->uk_align)
1418 		rsize = (rsize & ~keg->uk_align) + alignsize;
1419 	if ((rsize & alignsize) == 0)
1420 		rsize += alignsize;
1421 	trailer = rsize - keg->uk_size;
1422 	pages = (rsize * (PAGE_SIZE / alignsize)) / PAGE_SIZE;
1423 	pages = MIN(pages, (128 * 1024) / PAGE_SIZE);
1424 	keg->uk_rsize = rsize;
1425 	keg->uk_ppera = pages;
1426 	keg->uk_ipers = ((pages * PAGE_SIZE) + trailer) / rsize;
1427 	keg->uk_flags |= UMA_ZONE_OFFPAGE | UMA_ZONE_VTOSLAB;
1428 	KASSERT(keg->uk_ipers <= SLAB_SETSIZE,
1429 	    ("%s: keg->uk_ipers too high(%d) increase max_ipers", __func__,
1430 	    keg->uk_ipers));
1431 }
1432 
1433 /*
1434  * Keg header ctor.  This initializes all fields, locks, etc.  And inserts
1435  * the keg onto the global keg list.
1436  *
1437  * Arguments/Returns follow uma_ctor specifications
1438  *	udata  Actually uma_kctor_args
1439  */
1440 static int
1441 keg_ctor(void *mem, int size, void *udata, int flags)
1442 {
1443 	struct uma_kctor_args *arg = udata;
1444 	uma_keg_t keg = mem;
1445 	uma_zone_t zone;
1446 
1447 	bzero(keg, size);
1448 	keg->uk_size = arg->size;
1449 	keg->uk_init = arg->uminit;
1450 	keg->uk_fini = arg->fini;
1451 	keg->uk_align = arg->align;
1452 	keg->uk_cursor = 0;
1453 	keg->uk_free = 0;
1454 	keg->uk_reserve = 0;
1455 	keg->uk_pages = 0;
1456 	keg->uk_flags = arg->flags;
1457 	keg->uk_slabzone = NULL;
1458 
1459 	/*
1460 	 * The master zone is passed to us at keg-creation time.
1461 	 */
1462 	zone = arg->zone;
1463 	keg->uk_name = zone->uz_name;
1464 
1465 	if (arg->flags & UMA_ZONE_VM)
1466 		keg->uk_flags |= UMA_ZFLAG_CACHEONLY;
1467 
1468 	if (arg->flags & UMA_ZONE_ZINIT)
1469 		keg->uk_init = zero_init;
1470 
1471 	if (arg->flags & UMA_ZONE_MALLOC)
1472 		keg->uk_flags |= UMA_ZONE_VTOSLAB;
1473 
1474 	if (arg->flags & UMA_ZONE_PCPU)
1475 #ifdef SMP
1476 		keg->uk_flags |= UMA_ZONE_OFFPAGE;
1477 #else
1478 		keg->uk_flags &= ~UMA_ZONE_PCPU;
1479 #endif
1480 
1481 	if (keg->uk_flags & UMA_ZONE_CACHESPREAD) {
1482 		keg_cachespread_init(keg);
1483 	} else {
1484 		if (keg->uk_size > UMA_SLAB_SPACE)
1485 			keg_large_init(keg);
1486 		else
1487 			keg_small_init(keg);
1488 	}
1489 
1490 	if (keg->uk_flags & UMA_ZONE_OFFPAGE)
1491 		keg->uk_slabzone = slabzone;
1492 
1493 	/*
1494 	 * If we haven't booted yet we need allocations to go through the
1495 	 * startup cache until the vm is ready.
1496 	 */
1497 	if (booted < BOOT_PAGEALLOC)
1498 		keg->uk_allocf = startup_alloc;
1499 #ifdef UMA_MD_SMALL_ALLOC
1500 	else if (keg->uk_ppera == 1)
1501 		keg->uk_allocf = uma_small_alloc;
1502 #endif
1503 	else
1504 		keg->uk_allocf = page_alloc;
1505 #ifdef UMA_MD_SMALL_ALLOC
1506 	if (keg->uk_ppera == 1)
1507 		keg->uk_freef = uma_small_free;
1508 	else
1509 #endif
1510 		keg->uk_freef = page_free;
1511 
1512 	/*
1513 	 * Initialize keg's lock
1514 	 */
1515 	KEG_LOCK_INIT(keg, (arg->flags & UMA_ZONE_MTXCLASS));
1516 
1517 	/*
1518 	 * If we're putting the slab header in the actual page we need to
1519 	 * figure out where in each page it goes.  This calculates a right
1520 	 * justified offset into the memory on an ALIGN_PTR boundary.
1521 	 */
1522 	if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) {
1523 		u_int totsize;
1524 
1525 		/* Size of the slab struct and free list */
1526 		totsize = sizeof(struct uma_slab);
1527 
1528 		if (totsize & UMA_ALIGN_PTR)
1529 			totsize = (totsize & ~UMA_ALIGN_PTR) +
1530 			    (UMA_ALIGN_PTR + 1);
1531 		keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - totsize;
1532 
1533 		/*
1534 		 * The only way the following is possible is if with our
1535 		 * UMA_ALIGN_PTR adjustments we are now bigger than
1536 		 * UMA_SLAB_SIZE.  I haven't checked whether this is
1537 		 * mathematically possible for all cases, so we make
1538 		 * sure here anyway.
1539 		 */
1540 		totsize = keg->uk_pgoff + sizeof(struct uma_slab);
1541 		if (totsize > PAGE_SIZE * keg->uk_ppera) {
1542 			printf("zone %s ipers %d rsize %d size %d\n",
1543 			    zone->uz_name, keg->uk_ipers, keg->uk_rsize,
1544 			    keg->uk_size);
1545 			panic("UMA slab won't fit.");
1546 		}
1547 	}
1548 
1549 	if (keg->uk_flags & UMA_ZONE_HASH)
1550 		hash_alloc(&keg->uk_hash);
1551 
1552 	CTR5(KTR_UMA, "keg_ctor %p zone %s(%p) out %d free %d\n",
1553 	    keg, zone->uz_name, zone,
1554 	    (keg->uk_pages / keg->uk_ppera) * keg->uk_ipers - keg->uk_free,
1555 	    keg->uk_free);
1556 
1557 	LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
1558 
1559 	rw_wlock(&uma_rwlock);
1560 	LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
1561 	rw_wunlock(&uma_rwlock);
1562 	return (0);
1563 }
1564 
1565 /*
1566  * Zone header ctor.  This initializes all fields, locks, etc.
1567  *
1568  * Arguments/Returns follow uma_ctor specifications
1569  *	udata  Actually uma_zctor_args
1570  */
1571 static int
1572 zone_ctor(void *mem, int size, void *udata, int flags)
1573 {
1574 	struct uma_zctor_args *arg = udata;
1575 	uma_zone_t zone = mem;
1576 	uma_zone_t z;
1577 	uma_keg_t keg;
1578 
1579 	bzero(zone, size);
1580 	zone->uz_name = arg->name;
1581 	zone->uz_ctor = arg->ctor;
1582 	zone->uz_dtor = arg->dtor;
1583 	zone->uz_slab = zone_fetch_slab;
1584 	zone->uz_init = NULL;
1585 	zone->uz_fini = NULL;
1586 	zone->uz_allocs = 0;
1587 	zone->uz_frees = 0;
1588 	zone->uz_fails = 0;
1589 	zone->uz_sleeps = 0;
1590 	zone->uz_count = 0;
1591 	zone->uz_count_min = 0;
1592 	zone->uz_flags = 0;
1593 	zone->uz_warning = NULL;
1594 	/* The domain structures follow the cpu structures. */
1595 	zone->uz_domain = (struct uma_zone_domain *)&zone->uz_cpu[mp_ncpus];
1596 	timevalclear(&zone->uz_ratecheck);
1597 	keg = arg->keg;
1598 
1599 	ZONE_LOCK_INIT(zone, (arg->flags & UMA_ZONE_MTXCLASS));
1600 
1601 	/*
1602 	 * This is a pure cache zone, no kegs.
1603 	 */
1604 	if (arg->import) {
1605 		if (arg->flags & UMA_ZONE_VM)
1606 			arg->flags |= UMA_ZFLAG_CACHEONLY;
1607 		zone->uz_flags = arg->flags;
1608 		zone->uz_size = arg->size;
1609 		zone->uz_import = arg->import;
1610 		zone->uz_release = arg->release;
1611 		zone->uz_arg = arg->arg;
1612 		zone->uz_lockptr = &zone->uz_lock;
1613 		rw_wlock(&uma_rwlock);
1614 		LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link);
1615 		rw_wunlock(&uma_rwlock);
1616 		goto out;
1617 	}
1618 
1619 	/*
1620 	 * Use the regular zone/keg/slab allocator.
1621 	 */
1622 	zone->uz_import = (uma_import)zone_import;
1623 	zone->uz_release = (uma_release)zone_release;
1624 	zone->uz_arg = zone;
1625 
1626 	if (arg->flags & UMA_ZONE_SECONDARY) {
1627 		KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
1628 		zone->uz_init = arg->uminit;
1629 		zone->uz_fini = arg->fini;
1630 		zone->uz_lockptr = &keg->uk_lock;
1631 		zone->uz_flags |= UMA_ZONE_SECONDARY;
1632 		rw_wlock(&uma_rwlock);
1633 		ZONE_LOCK(zone);
1634 		LIST_FOREACH(z, &keg->uk_zones, uz_link) {
1635 			if (LIST_NEXT(z, uz_link) == NULL) {
1636 				LIST_INSERT_AFTER(z, zone, uz_link);
1637 				break;
1638 			}
1639 		}
1640 		ZONE_UNLOCK(zone);
1641 		rw_wunlock(&uma_rwlock);
1642 	} else if (keg == NULL) {
1643 		if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
1644 		    arg->align, arg->flags)) == NULL)
1645 			return (ENOMEM);
1646 	} else {
1647 		struct uma_kctor_args karg;
1648 		int error;
1649 
1650 		/* We should only be here from uma_startup() */
1651 		karg.size = arg->size;
1652 		karg.uminit = arg->uminit;
1653 		karg.fini = arg->fini;
1654 		karg.align = arg->align;
1655 		karg.flags = arg->flags;
1656 		karg.zone = zone;
1657 		error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
1658 		    flags);
1659 		if (error)
1660 			return (error);
1661 	}
1662 
1663 	/*
1664 	 * Link in the first keg.
1665 	 */
1666 	zone->uz_klink.kl_keg = keg;
1667 	LIST_INSERT_HEAD(&zone->uz_kegs, &zone->uz_klink, kl_link);
1668 	zone->uz_lockptr = &keg->uk_lock;
1669 	zone->uz_size = keg->uk_size;
1670 	zone->uz_flags |= (keg->uk_flags &
1671 	    (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
1672 
1673 	/*
1674 	 * Some internal zones don't have room allocated for the per cpu
1675 	 * caches.  If we're internal, bail out here.
1676 	 */
1677 	if (keg->uk_flags & UMA_ZFLAG_INTERNAL) {
1678 		KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
1679 		    ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
1680 		return (0);
1681 	}
1682 
1683 out:
1684 	KASSERT((arg->flags & (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET)) !=
1685 	    (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET),
1686 	    ("Invalid zone flag combination"));
1687 	if ((arg->flags & UMA_ZONE_MAXBUCKET) != 0)
1688 		zone->uz_count = BUCKET_MAX;
1689 	else if ((arg->flags & UMA_ZONE_NOBUCKET) != 0)
1690 		zone->uz_count = 0;
1691 	else
1692 		zone->uz_count = bucket_select(zone->uz_size);
1693 	zone->uz_count_min = zone->uz_count;
1694 
1695 	return (0);
1696 }
1697 
1698 /*
1699  * Keg header dtor.  This frees all data, destroys locks, frees the hash
1700  * table and removes the keg from the global list.
1701  *
1702  * Arguments/Returns follow uma_dtor specifications
1703  *	udata  unused
1704  */
1705 static void
1706 keg_dtor(void *arg, int size, void *udata)
1707 {
1708 	uma_keg_t keg;
1709 
1710 	keg = (uma_keg_t)arg;
1711 	KEG_LOCK(keg);
1712 	if (keg->uk_free != 0) {
1713 		printf("Freed UMA keg (%s) was not empty (%d items). "
1714 		    " Lost %d pages of memory.\n",
1715 		    keg->uk_name ? keg->uk_name : "",
1716 		    keg->uk_free, keg->uk_pages);
1717 	}
1718 	KEG_UNLOCK(keg);
1719 
1720 	hash_free(&keg->uk_hash);
1721 
1722 	KEG_LOCK_FINI(keg);
1723 }
1724 
1725 /*
1726  * Zone header dtor.
1727  *
1728  * Arguments/Returns follow uma_dtor specifications
1729  *	udata  unused
1730  */
1731 static void
1732 zone_dtor(void *arg, int size, void *udata)
1733 {
1734 	uma_klink_t klink;
1735 	uma_zone_t zone;
1736 	uma_keg_t keg;
1737 
1738 	zone = (uma_zone_t)arg;
1739 	keg = zone_first_keg(zone);
1740 
1741 	if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
1742 		cache_drain(zone);
1743 
1744 	rw_wlock(&uma_rwlock);
1745 	LIST_REMOVE(zone, uz_link);
1746 	rw_wunlock(&uma_rwlock);
1747 	/*
1748 	 * XXX there are some races here where
1749 	 * the zone can be drained but zone lock
1750 	 * released and then refilled before we
1751 	 * remove it... we dont care for now
1752 	 */
1753 	zone_drain_wait(zone, M_WAITOK);
1754 	/*
1755 	 * Unlink all of our kegs.
1756 	 */
1757 	while ((klink = LIST_FIRST(&zone->uz_kegs)) != NULL) {
1758 		klink->kl_keg = NULL;
1759 		LIST_REMOVE(klink, kl_link);
1760 		if (klink == &zone->uz_klink)
1761 			continue;
1762 		free(klink, M_TEMP);
1763 	}
1764 	/*
1765 	 * We only destroy kegs from non secondary zones.
1766 	 */
1767 	if (keg != NULL && (zone->uz_flags & UMA_ZONE_SECONDARY) == 0)  {
1768 		rw_wlock(&uma_rwlock);
1769 		LIST_REMOVE(keg, uk_link);
1770 		rw_wunlock(&uma_rwlock);
1771 		zone_free_item(kegs, keg, NULL, SKIP_NONE);
1772 	}
1773 	ZONE_LOCK_FINI(zone);
1774 }
1775 
1776 /*
1777  * Traverses every zone in the system and calls a callback
1778  *
1779  * Arguments:
1780  *	zfunc  A pointer to a function which accepts a zone
1781  *		as an argument.
1782  *
1783  * Returns:
1784  *	Nothing
1785  */
1786 static void
1787 zone_foreach(void (*zfunc)(uma_zone_t))
1788 {
1789 	uma_keg_t keg;
1790 	uma_zone_t zone;
1791 
1792 	rw_rlock(&uma_rwlock);
1793 	LIST_FOREACH(keg, &uma_kegs, uk_link) {
1794 		LIST_FOREACH(zone, &keg->uk_zones, uz_link)
1795 			zfunc(zone);
1796 	}
1797 	rw_runlock(&uma_rwlock);
1798 }
1799 
1800 /*
1801  * Count how many pages do we need to bootstrap.  VM supplies
1802  * its need in early zones in the argument, we add up our zones,
1803  * which consist of: UMA Slabs, UMA Hash and 9 Bucket zones. The
1804  * zone of zones and zone of kegs are accounted separately.
1805  */
1806 #define	UMA_BOOT_ZONES	11
1807 /* Zone of zones and zone of kegs have arbitrary alignment. */
1808 #define	UMA_BOOT_ALIGN	32
1809 static int zsize, ksize;
1810 int
1811 uma_startup_count(int vm_zones)
1812 {
1813 	int zones, pages;
1814 
1815 	ksize = sizeof(struct uma_keg) +
1816 	    (sizeof(struct uma_domain) * vm_ndomains);
1817 	zsize = sizeof(struct uma_zone) +
1818 	    (sizeof(struct uma_cache) * (mp_maxid + 1)) +
1819 	    (sizeof(struct uma_zone_domain) * vm_ndomains);
1820 
1821 	/*
1822 	 * Memory for the zone of kegs and its keg,
1823 	 * and for zone of zones.
1824 	 */
1825 	pages = howmany(roundup(zsize, CACHE_LINE_SIZE) * 2 +
1826 	    roundup(ksize, CACHE_LINE_SIZE), PAGE_SIZE);
1827 
1828 #ifdef	UMA_MD_SMALL_ALLOC
1829 	zones = UMA_BOOT_ZONES;
1830 #else
1831 	zones = UMA_BOOT_ZONES + vm_zones;
1832 	vm_zones = 0;
1833 #endif
1834 
1835 	/* Memory for the rest of startup zones, UMA and VM, ... */
1836 	if (zsize > UMA_SLAB_SPACE)
1837 		pages += (zones + vm_zones) *
1838 		    howmany(roundup2(zsize, UMA_BOOT_ALIGN), UMA_SLAB_SIZE);
1839 	else if (roundup2(zsize, UMA_BOOT_ALIGN) > UMA_SLAB_SPACE)
1840 		pages += zones;
1841 	else
1842 		pages += howmany(zones,
1843 		    UMA_SLAB_SPACE / roundup2(zsize, UMA_BOOT_ALIGN));
1844 
1845 	/* ... and their kegs. Note that zone of zones allocates a keg! */
1846 	pages += howmany(zones + 1,
1847 	    UMA_SLAB_SPACE / roundup2(ksize, UMA_BOOT_ALIGN));
1848 
1849 	/*
1850 	 * Most of startup zones are not going to be offpages, that's
1851 	 * why we use UMA_SLAB_SPACE instead of UMA_SLAB_SIZE in all
1852 	 * calculations.  Some large bucket zones will be offpage, and
1853 	 * thus will allocate hashes.  We take conservative approach
1854 	 * and assume that all zones may allocate hash.  This may give
1855 	 * us some positive inaccuracy, usually an extra single page.
1856 	 */
1857 	pages += howmany(zones, UMA_SLAB_SPACE /
1858 	    (sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT));
1859 
1860 	return (pages);
1861 }
1862 
1863 void
1864 uma_startup(void *mem, int npages)
1865 {
1866 	struct uma_zctor_args args;
1867 	uma_keg_t masterkeg;
1868 	uintptr_t m;
1869 
1870 #ifdef DIAGNOSTIC
1871 	printf("Entering %s with %d boot pages configured\n", __func__, npages);
1872 #endif
1873 
1874 	rw_init(&uma_rwlock, "UMA lock");
1875 
1876 	/* Use bootpages memory for the zone of zones and zone of kegs. */
1877 	m = (uintptr_t)mem;
1878 	zones = (uma_zone_t)m;
1879 	m += roundup(zsize, CACHE_LINE_SIZE);
1880 	kegs = (uma_zone_t)m;
1881 	m += roundup(zsize, CACHE_LINE_SIZE);
1882 	masterkeg = (uma_keg_t)m;
1883 	m += roundup(ksize, CACHE_LINE_SIZE);
1884 	m = roundup(m, PAGE_SIZE);
1885 	npages -= (m - (uintptr_t)mem) / PAGE_SIZE;
1886 	mem = (void *)m;
1887 
1888 	/* "manually" create the initial zone */
1889 	memset(&args, 0, sizeof(args));
1890 	args.name = "UMA Kegs";
1891 	args.size = ksize;
1892 	args.ctor = keg_ctor;
1893 	args.dtor = keg_dtor;
1894 	args.uminit = zero_init;
1895 	args.fini = NULL;
1896 	args.keg = masterkeg;
1897 	args.align = UMA_BOOT_ALIGN - 1;
1898 	args.flags = UMA_ZFLAG_INTERNAL;
1899 	zone_ctor(kegs, zsize, &args, M_WAITOK);
1900 
1901 	bootmem = mem;
1902 	boot_pages = npages;
1903 
1904 	args.name = "UMA Zones";
1905 	args.size = zsize;
1906 	args.ctor = zone_ctor;
1907 	args.dtor = zone_dtor;
1908 	args.uminit = zero_init;
1909 	args.fini = NULL;
1910 	args.keg = NULL;
1911 	args.align = UMA_BOOT_ALIGN - 1;
1912 	args.flags = UMA_ZFLAG_INTERNAL;
1913 	zone_ctor(zones, zsize, &args, M_WAITOK);
1914 
1915 	/* Now make a zone for slab headers */
1916 	slabzone = uma_zcreate("UMA Slabs",
1917 				sizeof(struct uma_slab),
1918 				NULL, NULL, NULL, NULL,
1919 				UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1920 
1921 	hashzone = uma_zcreate("UMA Hash",
1922 	    sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
1923 	    NULL, NULL, NULL, NULL,
1924 	    UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1925 
1926 	bucket_init();
1927 
1928 	booted = BOOT_STRAPPED;
1929 }
1930 
1931 void
1932 uma_startup1(void)
1933 {
1934 
1935 #ifdef DIAGNOSTIC
1936 	printf("Entering %s with %d boot pages left\n", __func__, boot_pages);
1937 #endif
1938 	booted = BOOT_PAGEALLOC;
1939 }
1940 
1941 void
1942 uma_startup2(void)
1943 {
1944 
1945 #ifdef DIAGNOSTIC
1946 	printf("Entering %s with %d boot pages left\n", __func__, boot_pages);
1947 #endif
1948 	booted = BOOT_BUCKETS;
1949 	sx_init(&uma_drain_lock, "umadrain");
1950 	bucket_enable();
1951 }
1952 
1953 /*
1954  * Initialize our callout handle
1955  *
1956  */
1957 static void
1958 uma_startup3(void)
1959 {
1960 
1961 	booted = BOOT_RUNNING;
1962 	callout_init(&uma_callout, 1);
1963 	callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
1964 }
1965 
1966 static uma_keg_t
1967 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
1968 		int align, uint32_t flags)
1969 {
1970 	struct uma_kctor_args args;
1971 
1972 	args.size = size;
1973 	args.uminit = uminit;
1974 	args.fini = fini;
1975 	args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
1976 	args.flags = flags;
1977 	args.zone = zone;
1978 	return (zone_alloc_item(kegs, &args, UMA_ANYDOMAIN, M_WAITOK));
1979 }
1980 
1981 /* Public functions */
1982 /* See uma.h */
1983 void
1984 uma_set_align(int align)
1985 {
1986 
1987 	if (align != UMA_ALIGN_CACHE)
1988 		uma_align_cache = align;
1989 }
1990 
1991 /* See uma.h */
1992 uma_zone_t
1993 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
1994 		uma_init uminit, uma_fini fini, int align, uint32_t flags)
1995 
1996 {
1997 	struct uma_zctor_args args;
1998 	uma_zone_t res;
1999 	bool locked;
2000 
2001 	KASSERT(powerof2(align + 1), ("invalid zone alignment %d for \"%s\"",
2002 	    align, name));
2003 
2004 	/* This stuff is essential for the zone ctor */
2005 	memset(&args, 0, sizeof(args));
2006 	args.name = name;
2007 	args.size = size;
2008 	args.ctor = ctor;
2009 	args.dtor = dtor;
2010 	args.uminit = uminit;
2011 	args.fini = fini;
2012 #ifdef  INVARIANTS
2013 	/*
2014 	 * If a zone is being created with an empty constructor and
2015 	 * destructor, pass UMA constructor/destructor which checks for
2016 	 * memory use after free.
2017 	 */
2018 	if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOFREE))) &&
2019 	    ctor == NULL && dtor == NULL && uminit == NULL && fini == NULL) {
2020 		args.ctor = trash_ctor;
2021 		args.dtor = trash_dtor;
2022 		args.uminit = trash_init;
2023 		args.fini = trash_fini;
2024 	}
2025 #endif
2026 	args.align = align;
2027 	args.flags = flags;
2028 	args.keg = NULL;
2029 
2030 	if (booted < BOOT_BUCKETS) {
2031 		locked = false;
2032 	} else {
2033 		sx_slock(&uma_drain_lock);
2034 		locked = true;
2035 	}
2036 	res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
2037 	if (locked)
2038 		sx_sunlock(&uma_drain_lock);
2039 	return (res);
2040 }
2041 
2042 /* See uma.h */
2043 uma_zone_t
2044 uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor,
2045 		    uma_init zinit, uma_fini zfini, uma_zone_t master)
2046 {
2047 	struct uma_zctor_args args;
2048 	uma_keg_t keg;
2049 	uma_zone_t res;
2050 	bool locked;
2051 
2052 	keg = zone_first_keg(master);
2053 	memset(&args, 0, sizeof(args));
2054 	args.name = name;
2055 	args.size = keg->uk_size;
2056 	args.ctor = ctor;
2057 	args.dtor = dtor;
2058 	args.uminit = zinit;
2059 	args.fini = zfini;
2060 	args.align = keg->uk_align;
2061 	args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
2062 	args.keg = keg;
2063 
2064 	if (booted < BOOT_BUCKETS) {
2065 		locked = false;
2066 	} else {
2067 		sx_slock(&uma_drain_lock);
2068 		locked = true;
2069 	}
2070 	/* XXX Attaches only one keg of potentially many. */
2071 	res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
2072 	if (locked)
2073 		sx_sunlock(&uma_drain_lock);
2074 	return (res);
2075 }
2076 
2077 /* See uma.h */
2078 uma_zone_t
2079 uma_zcache_create(char *name, int size, uma_ctor ctor, uma_dtor dtor,
2080 		    uma_init zinit, uma_fini zfini, uma_import zimport,
2081 		    uma_release zrelease, void *arg, int flags)
2082 {
2083 	struct uma_zctor_args args;
2084 
2085 	memset(&args, 0, sizeof(args));
2086 	args.name = name;
2087 	args.size = size;
2088 	args.ctor = ctor;
2089 	args.dtor = dtor;
2090 	args.uminit = zinit;
2091 	args.fini = zfini;
2092 	args.import = zimport;
2093 	args.release = zrelease;
2094 	args.arg = arg;
2095 	args.align = 0;
2096 	args.flags = flags;
2097 
2098 	return (zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK));
2099 }
2100 
2101 static void
2102 zone_lock_pair(uma_zone_t a, uma_zone_t b)
2103 {
2104 	if (a < b) {
2105 		ZONE_LOCK(a);
2106 		mtx_lock_flags(b->uz_lockptr, MTX_DUPOK);
2107 	} else {
2108 		ZONE_LOCK(b);
2109 		mtx_lock_flags(a->uz_lockptr, MTX_DUPOK);
2110 	}
2111 }
2112 
2113 static void
2114 zone_unlock_pair(uma_zone_t a, uma_zone_t b)
2115 {
2116 
2117 	ZONE_UNLOCK(a);
2118 	ZONE_UNLOCK(b);
2119 }
2120 
2121 int
2122 uma_zsecond_add(uma_zone_t zone, uma_zone_t master)
2123 {
2124 	uma_klink_t klink;
2125 	uma_klink_t kl;
2126 	int error;
2127 
2128 	error = 0;
2129 	klink = malloc(sizeof(*klink), M_TEMP, M_WAITOK | M_ZERO);
2130 
2131 	zone_lock_pair(zone, master);
2132 	/*
2133 	 * zone must use vtoslab() to resolve objects and must already be
2134 	 * a secondary.
2135 	 */
2136 	if ((zone->uz_flags & (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY))
2137 	    != (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) {
2138 		error = EINVAL;
2139 		goto out;
2140 	}
2141 	/*
2142 	 * The new master must also use vtoslab().
2143 	 */
2144 	if ((zone->uz_flags & UMA_ZONE_VTOSLAB) != UMA_ZONE_VTOSLAB) {
2145 		error = EINVAL;
2146 		goto out;
2147 	}
2148 
2149 	/*
2150 	 * The underlying object must be the same size.  rsize
2151 	 * may be different.
2152 	 */
2153 	if (master->uz_size != zone->uz_size) {
2154 		error = E2BIG;
2155 		goto out;
2156 	}
2157 	/*
2158 	 * Put it at the end of the list.
2159 	 */
2160 	klink->kl_keg = zone_first_keg(master);
2161 	LIST_FOREACH(kl, &zone->uz_kegs, kl_link) {
2162 		if (LIST_NEXT(kl, kl_link) == NULL) {
2163 			LIST_INSERT_AFTER(kl, klink, kl_link);
2164 			break;
2165 		}
2166 	}
2167 	klink = NULL;
2168 	zone->uz_flags |= UMA_ZFLAG_MULTI;
2169 	zone->uz_slab = zone_fetch_slab_multi;
2170 
2171 out:
2172 	zone_unlock_pair(zone, master);
2173 	if (klink != NULL)
2174 		free(klink, M_TEMP);
2175 
2176 	return (error);
2177 }
2178 
2179 
2180 /* See uma.h */
2181 void
2182 uma_zdestroy(uma_zone_t zone)
2183 {
2184 
2185 	sx_slock(&uma_drain_lock);
2186 	zone_free_item(zones, zone, NULL, SKIP_NONE);
2187 	sx_sunlock(&uma_drain_lock);
2188 }
2189 
2190 void
2191 uma_zwait(uma_zone_t zone)
2192 {
2193 	void *item;
2194 
2195 	item = uma_zalloc_arg(zone, NULL, M_WAITOK);
2196 	uma_zfree(zone, item);
2197 }
2198 
2199 /* See uma.h */
2200 void *
2201 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
2202 {
2203 	uma_zone_domain_t zdom;
2204 	uma_bucket_t bucket;
2205 	uma_cache_t cache;
2206 	void *item;
2207 	int cpu, domain, lockfail;
2208 
2209 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
2210 	random_harvest_fast_uma(&zone, sizeof(zone), 1, RANDOM_UMA);
2211 
2212 	/* This is the fast path allocation */
2213 	CTR4(KTR_UMA, "uma_zalloc_arg thread %x zone %s(%p) flags %d",
2214 	    curthread, zone->uz_name, zone, flags);
2215 
2216 	if (flags & M_WAITOK) {
2217 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
2218 		    "uma_zalloc_arg: zone \"%s\"", zone->uz_name);
2219 	}
2220 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
2221 	    ("uma_zalloc_arg: called with spinlock or critical section held"));
2222 
2223 #ifdef DEBUG_MEMGUARD
2224 	if (memguard_cmp_zone(zone)) {
2225 		item = memguard_alloc(zone->uz_size, flags);
2226 		if (item != NULL) {
2227 			if (zone->uz_init != NULL &&
2228 			    zone->uz_init(item, zone->uz_size, flags) != 0)
2229 				return (NULL);
2230 			if (zone->uz_ctor != NULL &&
2231 			    zone->uz_ctor(item, zone->uz_size, udata,
2232 			    flags) != 0) {
2233 			    	zone->uz_fini(item, zone->uz_size);
2234 				return (NULL);
2235 			}
2236 			return (item);
2237 		}
2238 		/* This is unfortunate but should not be fatal. */
2239 	}
2240 #endif
2241 	/*
2242 	 * If possible, allocate from the per-CPU cache.  There are two
2243 	 * requirements for safe access to the per-CPU cache: (1) the thread
2244 	 * accessing the cache must not be preempted or yield during access,
2245 	 * and (2) the thread must not migrate CPUs without switching which
2246 	 * cache it accesses.  We rely on a critical section to prevent
2247 	 * preemption and migration.  We release the critical section in
2248 	 * order to acquire the zone mutex if we are unable to allocate from
2249 	 * the current cache; when we re-acquire the critical section, we
2250 	 * must detect and handle migration if it has occurred.
2251 	 */
2252 	critical_enter();
2253 	cpu = curcpu;
2254 	cache = &zone->uz_cpu[cpu];
2255 
2256 zalloc_start:
2257 	bucket = cache->uc_allocbucket;
2258 	if (bucket != NULL && bucket->ub_cnt > 0) {
2259 		bucket->ub_cnt--;
2260 		item = bucket->ub_bucket[bucket->ub_cnt];
2261 #ifdef INVARIANTS
2262 		bucket->ub_bucket[bucket->ub_cnt] = NULL;
2263 #endif
2264 		KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled."));
2265 		cache->uc_allocs++;
2266 		critical_exit();
2267 		if (zone->uz_ctor != NULL &&
2268 		    zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2269 			atomic_add_long(&zone->uz_fails, 1);
2270 			zone_free_item(zone, item, udata, SKIP_DTOR);
2271 			return (NULL);
2272 		}
2273 #ifdef INVARIANTS
2274 		uma_dbg_alloc(zone, NULL, item);
2275 #endif
2276 		if (flags & M_ZERO)
2277 			uma_zero_item(item, zone);
2278 		return (item);
2279 	}
2280 
2281 	/*
2282 	 * We have run out of items in our alloc bucket.
2283 	 * See if we can switch with our free bucket.
2284 	 */
2285 	bucket = cache->uc_freebucket;
2286 	if (bucket != NULL && bucket->ub_cnt > 0) {
2287 		CTR2(KTR_UMA,
2288 		    "uma_zalloc: zone %s(%p) swapping empty with alloc",
2289 		    zone->uz_name, zone);
2290 		cache->uc_freebucket = cache->uc_allocbucket;
2291 		cache->uc_allocbucket = bucket;
2292 		goto zalloc_start;
2293 	}
2294 
2295 	/*
2296 	 * Discard any empty allocation bucket while we hold no locks.
2297 	 */
2298 	bucket = cache->uc_allocbucket;
2299 	cache->uc_allocbucket = NULL;
2300 	critical_exit();
2301 	if (bucket != NULL)
2302 		bucket_free(zone, bucket, udata);
2303 
2304 	if (zone->uz_flags & UMA_ZONE_NUMA)
2305 		domain = PCPU_GET(domain);
2306 	else
2307 		domain = UMA_ANYDOMAIN;
2308 
2309 	/* Short-circuit for zones without buckets and low memory. */
2310 	if (zone->uz_count == 0 || bucketdisable)
2311 		goto zalloc_item;
2312 
2313 	/*
2314 	 * Attempt to retrieve the item from the per-CPU cache has failed, so
2315 	 * we must go back to the zone.  This requires the zone lock, so we
2316 	 * must drop the critical section, then re-acquire it when we go back
2317 	 * to the cache.  Since the critical section is released, we may be
2318 	 * preempted or migrate.  As such, make sure not to maintain any
2319 	 * thread-local state specific to the cache from prior to releasing
2320 	 * the critical section.
2321 	 */
2322 	lockfail = 0;
2323 	if (ZONE_TRYLOCK(zone) == 0) {
2324 		/* Record contention to size the buckets. */
2325 		ZONE_LOCK(zone);
2326 		lockfail = 1;
2327 	}
2328 	critical_enter();
2329 	cpu = curcpu;
2330 	cache = &zone->uz_cpu[cpu];
2331 
2332 	/* See if we lost the race to fill the cache. */
2333 	if (cache->uc_allocbucket != NULL) {
2334 		ZONE_UNLOCK(zone);
2335 		goto zalloc_start;
2336 	}
2337 
2338 	/*
2339 	 * Check the zone's cache of buckets.
2340 	 */
2341 	if (domain == UMA_ANYDOMAIN)
2342 		zdom = &zone->uz_domain[0];
2343 	else
2344 		zdom = &zone->uz_domain[domain];
2345 	if ((bucket = LIST_FIRST(&zdom->uzd_buckets)) != NULL) {
2346 		KASSERT(bucket->ub_cnt != 0,
2347 		    ("uma_zalloc_arg: Returning an empty bucket."));
2348 
2349 		LIST_REMOVE(bucket, ub_link);
2350 		cache->uc_allocbucket = bucket;
2351 		ZONE_UNLOCK(zone);
2352 		goto zalloc_start;
2353 	}
2354 	/* We are no longer associated with this CPU. */
2355 	critical_exit();
2356 
2357 	/*
2358 	 * We bump the uz count when the cache size is insufficient to
2359 	 * handle the working set.
2360 	 */
2361 	if (lockfail && zone->uz_count < BUCKET_MAX)
2362 		zone->uz_count++;
2363 	ZONE_UNLOCK(zone);
2364 
2365 	/*
2366 	 * Now lets just fill a bucket and put it on the free list.  If that
2367 	 * works we'll restart the allocation from the beginning and it
2368 	 * will use the just filled bucket.
2369 	 */
2370 	bucket = zone_alloc_bucket(zone, udata, domain, flags);
2371 	CTR3(KTR_UMA, "uma_zalloc: zone %s(%p) bucket zone returned %p",
2372 	    zone->uz_name, zone, bucket);
2373 	if (bucket != NULL) {
2374 		ZONE_LOCK(zone);
2375 		critical_enter();
2376 		cpu = curcpu;
2377 		cache = &zone->uz_cpu[cpu];
2378 		/*
2379 		 * See if we lost the race or were migrated.  Cache the
2380 		 * initialized bucket to make this less likely or claim
2381 		 * the memory directly.
2382 		 */
2383 		if (cache->uc_allocbucket != NULL ||
2384 		    (zone->uz_flags & UMA_ZONE_NUMA &&
2385 		    domain != PCPU_GET(domain)))
2386 			LIST_INSERT_HEAD(&zdom->uzd_buckets, bucket, ub_link);
2387 		else
2388 			cache->uc_allocbucket = bucket;
2389 		ZONE_UNLOCK(zone);
2390 		goto zalloc_start;
2391 	}
2392 
2393 	/*
2394 	 * We may not be able to get a bucket so return an actual item.
2395 	 */
2396 zalloc_item:
2397 	item = zone_alloc_item(zone, udata, domain, flags);
2398 
2399 	return (item);
2400 }
2401 
2402 void *
2403 uma_zalloc_domain(uma_zone_t zone, void *udata, int domain, int flags)
2404 {
2405 
2406 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
2407 	random_harvest_fast_uma(&zone, sizeof(zone), 1, RANDOM_UMA);
2408 
2409 	/* This is the fast path allocation */
2410 	CTR5(KTR_UMA,
2411 	    "uma_zalloc_domain thread %x zone %s(%p) domain %d flags %d",
2412 	    curthread, zone->uz_name, zone, domain, flags);
2413 
2414 	if (flags & M_WAITOK) {
2415 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
2416 		    "uma_zalloc_domain: zone \"%s\"", zone->uz_name);
2417 	}
2418 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
2419 	    ("uma_zalloc_domain: called with spinlock or critical section held"));
2420 
2421 	return (zone_alloc_item(zone, udata, domain, flags));
2422 }
2423 
2424 /*
2425  * Find a slab with some space.  Prefer slabs that are partially used over those
2426  * that are totally full.  This helps to reduce fragmentation.
2427  *
2428  * If 'rr' is 1, search all domains starting from 'domain'.  Otherwise check
2429  * only 'domain'.
2430  */
2431 static uma_slab_t
2432 keg_first_slab(uma_keg_t keg, int domain, int rr)
2433 {
2434 	uma_domain_t dom;
2435 	uma_slab_t slab;
2436 	int start;
2437 
2438 	KASSERT(domain >= 0 && domain < vm_ndomains,
2439 	    ("keg_first_slab: domain %d out of range", domain));
2440 
2441 	slab = NULL;
2442 	start = domain;
2443 	do {
2444 		dom = &keg->uk_domain[domain];
2445 		if (!LIST_EMPTY(&dom->ud_part_slab))
2446 			return (LIST_FIRST(&dom->ud_part_slab));
2447 		if (!LIST_EMPTY(&dom->ud_free_slab)) {
2448 			slab = LIST_FIRST(&dom->ud_free_slab);
2449 			LIST_REMOVE(slab, us_link);
2450 			LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
2451 			return (slab);
2452 		}
2453 		if (rr)
2454 			domain = (domain + 1) % vm_ndomains;
2455 	} while (domain != start);
2456 
2457 	return (NULL);
2458 }
2459 
2460 static uma_slab_t
2461 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int rdomain, int flags)
2462 {
2463 	uma_domain_t dom;
2464 	uma_slab_t slab;
2465 	int allocflags, domain, reserve, rr, start;
2466 
2467 	mtx_assert(&keg->uk_lock, MA_OWNED);
2468 	slab = NULL;
2469 	reserve = 0;
2470 	allocflags = flags;
2471 	if ((flags & M_USE_RESERVE) == 0)
2472 		reserve = keg->uk_reserve;
2473 
2474 	/*
2475 	 * Round-robin for non first-touch zones when there is more than one
2476 	 * domain.
2477 	 */
2478 	if (vm_ndomains == 1)
2479 		rdomain = 0;
2480 	rr = rdomain == UMA_ANYDOMAIN;
2481 	if (rr) {
2482 		keg->uk_cursor = (keg->uk_cursor + 1) % vm_ndomains;
2483 		domain = start = keg->uk_cursor;
2484 		/* Only block on the second pass. */
2485 		if ((flags & (M_WAITOK | M_NOVM)) == M_WAITOK)
2486 			allocflags = (allocflags & ~M_WAITOK) | M_NOWAIT;
2487 	} else
2488 		domain = start = rdomain;
2489 
2490 again:
2491 	do {
2492 		if (keg->uk_free > reserve &&
2493 		    (slab = keg_first_slab(keg, domain, rr)) != NULL) {
2494 			MPASS(slab->us_keg == keg);
2495 			return (slab);
2496 		}
2497 
2498 		/*
2499 		 * M_NOVM means don't ask at all!
2500 		 */
2501 		if (flags & M_NOVM)
2502 			break;
2503 
2504 		if (keg->uk_maxpages && keg->uk_pages >= keg->uk_maxpages) {
2505 			keg->uk_flags |= UMA_ZFLAG_FULL;
2506 			/*
2507 			 * If this is not a multi-zone, set the FULL bit.
2508 			 * Otherwise slab_multi() takes care of it.
2509 			 */
2510 			if ((zone->uz_flags & UMA_ZFLAG_MULTI) == 0) {
2511 				zone->uz_flags |= UMA_ZFLAG_FULL;
2512 				zone_log_warning(zone);
2513 				zone_maxaction(zone);
2514 			}
2515 			if (flags & M_NOWAIT)
2516 				return (NULL);
2517 			zone->uz_sleeps++;
2518 			msleep(keg, &keg->uk_lock, PVM, "keglimit", 0);
2519 			continue;
2520 		}
2521 		slab = keg_alloc_slab(keg, zone, domain, allocflags);
2522 		/*
2523 		 * If we got a slab here it's safe to mark it partially used
2524 		 * and return.  We assume that the caller is going to remove
2525 		 * at least one item.
2526 		 */
2527 		if (slab) {
2528 			MPASS(slab->us_keg == keg);
2529 			dom = &keg->uk_domain[slab->us_domain];
2530 			LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
2531 			return (slab);
2532 		}
2533 		if (rr) {
2534 			keg->uk_cursor = (keg->uk_cursor + 1) % vm_ndomains;
2535 			domain = keg->uk_cursor;
2536 		}
2537 	} while (domain != start);
2538 
2539 	/* Retry domain scan with blocking. */
2540 	if (allocflags != flags) {
2541 		allocflags = flags;
2542 		goto again;
2543 	}
2544 
2545 	/*
2546 	 * We might not have been able to get a slab but another cpu
2547 	 * could have while we were unlocked.  Check again before we
2548 	 * fail.
2549 	 */
2550 	if (keg->uk_free > reserve &&
2551 	    (slab = keg_first_slab(keg, domain, rr)) != NULL) {
2552 		MPASS(slab->us_keg == keg);
2553 		return (slab);
2554 	}
2555 	return (NULL);
2556 }
2557 
2558 static uma_slab_t
2559 zone_fetch_slab(uma_zone_t zone, uma_keg_t keg, int domain, int flags)
2560 {
2561 	uma_slab_t slab;
2562 
2563 	if (keg == NULL) {
2564 		keg = zone_first_keg(zone);
2565 		KEG_LOCK(keg);
2566 	}
2567 
2568 	for (;;) {
2569 		slab = keg_fetch_slab(keg, zone, domain, flags);
2570 		if (slab)
2571 			return (slab);
2572 		if (flags & (M_NOWAIT | M_NOVM))
2573 			break;
2574 	}
2575 	KEG_UNLOCK(keg);
2576 	return (NULL);
2577 }
2578 
2579 /*
2580  * uma_zone_fetch_slab_multi:  Fetches a slab from one available keg.  Returns
2581  * with the keg locked.  On NULL no lock is held.
2582  *
2583  * The last pointer is used to seed the search.  It is not required.
2584  */
2585 static uma_slab_t
2586 zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int domain, int rflags)
2587 {
2588 	uma_klink_t klink;
2589 	uma_slab_t slab;
2590 	uma_keg_t keg;
2591 	int flags;
2592 	int empty;
2593 	int full;
2594 
2595 	/*
2596 	 * Don't wait on the first pass.  This will skip limit tests
2597 	 * as well.  We don't want to block if we can find a provider
2598 	 * without blocking.
2599 	 */
2600 	flags = (rflags & ~M_WAITOK) | M_NOWAIT;
2601 	/*
2602 	 * Use the last slab allocated as a hint for where to start
2603 	 * the search.
2604 	 */
2605 	if (last != NULL) {
2606 		slab = keg_fetch_slab(last, zone, domain, flags);
2607 		if (slab)
2608 			return (slab);
2609 		KEG_UNLOCK(last);
2610 	}
2611 	/*
2612 	 * Loop until we have a slab incase of transient failures
2613 	 * while M_WAITOK is specified.  I'm not sure this is 100%
2614 	 * required but we've done it for so long now.
2615 	 */
2616 	for (;;) {
2617 		empty = 0;
2618 		full = 0;
2619 		/*
2620 		 * Search the available kegs for slabs.  Be careful to hold the
2621 		 * correct lock while calling into the keg layer.
2622 		 */
2623 		LIST_FOREACH(klink, &zone->uz_kegs, kl_link) {
2624 			keg = klink->kl_keg;
2625 			KEG_LOCK(keg);
2626 			if ((keg->uk_flags & UMA_ZFLAG_FULL) == 0) {
2627 				slab = keg_fetch_slab(keg, zone, domain, flags);
2628 				if (slab)
2629 					return (slab);
2630 			}
2631 			if (keg->uk_flags & UMA_ZFLAG_FULL)
2632 				full++;
2633 			else
2634 				empty++;
2635 			KEG_UNLOCK(keg);
2636 		}
2637 		if (rflags & (M_NOWAIT | M_NOVM))
2638 			break;
2639 		flags = rflags;
2640 		/*
2641 		 * All kegs are full.  XXX We can't atomically check all kegs
2642 		 * and sleep so just sleep for a short period and retry.
2643 		 */
2644 		if (full && !empty) {
2645 			ZONE_LOCK(zone);
2646 			zone->uz_flags |= UMA_ZFLAG_FULL;
2647 			zone->uz_sleeps++;
2648 			zone_log_warning(zone);
2649 			zone_maxaction(zone);
2650 			msleep(zone, zone->uz_lockptr, PVM,
2651 			    "zonelimit", hz/100);
2652 			zone->uz_flags &= ~UMA_ZFLAG_FULL;
2653 			ZONE_UNLOCK(zone);
2654 			continue;
2655 		}
2656 	}
2657 	return (NULL);
2658 }
2659 
2660 static void *
2661 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
2662 {
2663 	uma_domain_t dom;
2664 	void *item;
2665 	uint8_t freei;
2666 
2667 	MPASS(keg == slab->us_keg);
2668 	mtx_assert(&keg->uk_lock, MA_OWNED);
2669 
2670 	freei = BIT_FFS(SLAB_SETSIZE, &slab->us_free) - 1;
2671 	BIT_CLR(SLAB_SETSIZE, freei, &slab->us_free);
2672 	item = slab->us_data + (keg->uk_rsize * freei);
2673 	slab->us_freecount--;
2674 	keg->uk_free--;
2675 
2676 	/* Move this slab to the full list */
2677 	if (slab->us_freecount == 0) {
2678 		LIST_REMOVE(slab, us_link);
2679 		dom = &keg->uk_domain[slab->us_domain];
2680 		LIST_INSERT_HEAD(&dom->ud_full_slab, slab, us_link);
2681 	}
2682 
2683 	return (item);
2684 }
2685 
2686 static int
2687 zone_import(uma_zone_t zone, void **bucket, int max, int domain, int flags)
2688 {
2689 	uma_slab_t slab;
2690 	uma_keg_t keg;
2691 	int stripe;
2692 	int i;
2693 
2694 	slab = NULL;
2695 	keg = NULL;
2696 	/* Try to keep the buckets totally full */
2697 	for (i = 0; i < max; ) {
2698 		if ((slab = zone->uz_slab(zone, keg, domain, flags)) == NULL)
2699 			break;
2700 		keg = slab->us_keg;
2701 		stripe = howmany(max, vm_ndomains);
2702 		while (slab->us_freecount && i < max) {
2703 			bucket[i++] = slab_alloc_item(keg, slab);
2704 			if (keg->uk_free <= keg->uk_reserve)
2705 				break;
2706 #ifdef NUMA
2707 			/*
2708 			 * If the zone is striped we pick a new slab for every
2709 			 * N allocations.  Eliminating this conditional will
2710 			 * instead pick a new domain for each bucket rather
2711 			 * than stripe within each bucket.  The current option
2712 			 * produces more fragmentation and requires more cpu
2713 			 * time but yields better distribution.
2714 			 */
2715 			if ((zone->uz_flags & UMA_ZONE_NUMA) == 0 &&
2716 			    vm_ndomains > 1 && --stripe == 0)
2717 				break;
2718 #endif
2719 		}
2720 		/* Don't block if we allocated any successfully. */
2721 		flags &= ~M_WAITOK;
2722 		flags |= M_NOWAIT;
2723 	}
2724 	if (slab != NULL)
2725 		KEG_UNLOCK(keg);
2726 
2727 	return i;
2728 }
2729 
2730 static uma_bucket_t
2731 zone_alloc_bucket(uma_zone_t zone, void *udata, int domain, int flags)
2732 {
2733 	uma_bucket_t bucket;
2734 	int max;
2735 
2736 	/* Don't wait for buckets, preserve caller's NOVM setting. */
2737 	bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
2738 	if (bucket == NULL)
2739 		return (NULL);
2740 
2741 	max = MIN(bucket->ub_entries, zone->uz_count);
2742 	bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
2743 	    max, domain, flags);
2744 
2745 	/*
2746 	 * Initialize the memory if necessary.
2747 	 */
2748 	if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
2749 		int i;
2750 
2751 		for (i = 0; i < bucket->ub_cnt; i++)
2752 			if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size,
2753 			    flags) != 0)
2754 				break;
2755 		/*
2756 		 * If we couldn't initialize the whole bucket, put the
2757 		 * rest back onto the freelist.
2758 		 */
2759 		if (i != bucket->ub_cnt) {
2760 			zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
2761 			    bucket->ub_cnt - i);
2762 #ifdef INVARIANTS
2763 			bzero(&bucket->ub_bucket[i],
2764 			    sizeof(void *) * (bucket->ub_cnt - i));
2765 #endif
2766 			bucket->ub_cnt = i;
2767 		}
2768 	}
2769 
2770 	if (bucket->ub_cnt == 0) {
2771 		bucket_free(zone, bucket, udata);
2772 		atomic_add_long(&zone->uz_fails, 1);
2773 		return (NULL);
2774 	}
2775 
2776 	return (bucket);
2777 }
2778 
2779 /*
2780  * Allocates a single item from a zone.
2781  *
2782  * Arguments
2783  *	zone   The zone to alloc for.
2784  *	udata  The data to be passed to the constructor.
2785  *	domain The domain to allocate from or UMA_ANYDOMAIN.
2786  *	flags  M_WAITOK, M_NOWAIT, M_ZERO.
2787  *
2788  * Returns
2789  *	NULL if there is no memory and M_NOWAIT is set
2790  *	An item if successful
2791  */
2792 
2793 static void *
2794 zone_alloc_item(uma_zone_t zone, void *udata, int domain, int flags)
2795 {
2796 	void *item;
2797 
2798 	item = NULL;
2799 
2800 	if (zone->uz_import(zone->uz_arg, &item, 1, domain, flags) != 1)
2801 		goto fail;
2802 	atomic_add_long(&zone->uz_allocs, 1);
2803 
2804 	/*
2805 	 * We have to call both the zone's init (not the keg's init)
2806 	 * and the zone's ctor.  This is because the item is going from
2807 	 * a keg slab directly to the user, and the user is expecting it
2808 	 * to be both zone-init'd as well as zone-ctor'd.
2809 	 */
2810 	if (zone->uz_init != NULL) {
2811 		if (zone->uz_init(item, zone->uz_size, flags) != 0) {
2812 			zone_free_item(zone, item, udata, SKIP_FINI);
2813 			goto fail;
2814 		}
2815 	}
2816 	if (zone->uz_ctor != NULL) {
2817 		if (zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2818 			zone_free_item(zone, item, udata, SKIP_DTOR);
2819 			goto fail;
2820 		}
2821 	}
2822 #ifdef INVARIANTS
2823 	uma_dbg_alloc(zone, NULL, item);
2824 #endif
2825 	if (flags & M_ZERO)
2826 		uma_zero_item(item, zone);
2827 
2828 	CTR3(KTR_UMA, "zone_alloc_item item %p from %s(%p)", item,
2829 	    zone->uz_name, zone);
2830 
2831 	return (item);
2832 
2833 fail:
2834 	CTR2(KTR_UMA, "zone_alloc_item failed from %s(%p)",
2835 	    zone->uz_name, zone);
2836 	atomic_add_long(&zone->uz_fails, 1);
2837 	return (NULL);
2838 }
2839 
2840 /* See uma.h */
2841 void
2842 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
2843 {
2844 	uma_cache_t cache;
2845 	uma_bucket_t bucket;
2846 	uma_zone_domain_t zdom;
2847 	int cpu, domain, lockfail;
2848 
2849 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
2850 	random_harvest_fast_uma(&zone, sizeof(zone), 1, RANDOM_UMA);
2851 
2852 	CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread,
2853 	    zone->uz_name);
2854 
2855 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
2856 	    ("uma_zfree_arg: called with spinlock or critical section held"));
2857 
2858         /* uma_zfree(..., NULL) does nothing, to match free(9). */
2859         if (item == NULL)
2860                 return;
2861 #ifdef DEBUG_MEMGUARD
2862 	if (is_memguard_addr(item)) {
2863 		if (zone->uz_dtor != NULL)
2864 			zone->uz_dtor(item, zone->uz_size, udata);
2865 		if (zone->uz_fini != NULL)
2866 			zone->uz_fini(item, zone->uz_size);
2867 		memguard_free(item);
2868 		return;
2869 	}
2870 #endif
2871 #ifdef INVARIANTS
2872 	if (zone->uz_flags & UMA_ZONE_MALLOC)
2873 		uma_dbg_free(zone, udata, item);
2874 	else
2875 		uma_dbg_free(zone, NULL, item);
2876 #endif
2877 	if (zone->uz_dtor != NULL)
2878 		zone->uz_dtor(item, zone->uz_size, udata);
2879 
2880 	/*
2881 	 * The race here is acceptable.  If we miss it we'll just have to wait
2882 	 * a little longer for the limits to be reset.
2883 	 */
2884 	if (zone->uz_flags & UMA_ZFLAG_FULL)
2885 		goto zfree_item;
2886 
2887 	/*
2888 	 * If possible, free to the per-CPU cache.  There are two
2889 	 * requirements for safe access to the per-CPU cache: (1) the thread
2890 	 * accessing the cache must not be preempted or yield during access,
2891 	 * and (2) the thread must not migrate CPUs without switching which
2892 	 * cache it accesses.  We rely on a critical section to prevent
2893 	 * preemption and migration.  We release the critical section in
2894 	 * order to acquire the zone mutex if we are unable to free to the
2895 	 * current cache; when we re-acquire the critical section, we must
2896 	 * detect and handle migration if it has occurred.
2897 	 */
2898 zfree_restart:
2899 	critical_enter();
2900 	cpu = curcpu;
2901 	cache = &zone->uz_cpu[cpu];
2902 
2903 zfree_start:
2904 	/*
2905 	 * Try to free into the allocbucket first to give LIFO ordering
2906 	 * for cache-hot datastructures.  Spill over into the freebucket
2907 	 * if necessary.  Alloc will swap them if one runs dry.
2908 	 */
2909 	bucket = cache->uc_allocbucket;
2910 	if (bucket == NULL || bucket->ub_cnt >= bucket->ub_entries)
2911 		bucket = cache->uc_freebucket;
2912 	if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) {
2913 		KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL,
2914 		    ("uma_zfree: Freeing to non free bucket index."));
2915 		bucket->ub_bucket[bucket->ub_cnt] = item;
2916 		bucket->ub_cnt++;
2917 		cache->uc_frees++;
2918 		critical_exit();
2919 		return;
2920 	}
2921 
2922 	/*
2923 	 * We must go back the zone, which requires acquiring the zone lock,
2924 	 * which in turn means we must release and re-acquire the critical
2925 	 * section.  Since the critical section is released, we may be
2926 	 * preempted or migrate.  As such, make sure not to maintain any
2927 	 * thread-local state specific to the cache from prior to releasing
2928 	 * the critical section.
2929 	 */
2930 	critical_exit();
2931 	if (zone->uz_count == 0 || bucketdisable)
2932 		goto zfree_item;
2933 
2934 	lockfail = 0;
2935 	if (ZONE_TRYLOCK(zone) == 0) {
2936 		/* Record contention to size the buckets. */
2937 		ZONE_LOCK(zone);
2938 		lockfail = 1;
2939 	}
2940 	critical_enter();
2941 	cpu = curcpu;
2942 	cache = &zone->uz_cpu[cpu];
2943 
2944 	/*
2945 	 * Since we have locked the zone we may as well send back our stats.
2946 	 */
2947 	atomic_add_long(&zone->uz_allocs, cache->uc_allocs);
2948 	atomic_add_long(&zone->uz_frees, cache->uc_frees);
2949 	cache->uc_allocs = 0;
2950 	cache->uc_frees = 0;
2951 
2952 	bucket = cache->uc_freebucket;
2953 	if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) {
2954 		ZONE_UNLOCK(zone);
2955 		goto zfree_start;
2956 	}
2957 	cache->uc_freebucket = NULL;
2958 	/* We are no longer associated with this CPU. */
2959 	critical_exit();
2960 
2961 	if ((zone->uz_flags & UMA_ZONE_NUMA) != 0)
2962 		domain = PCPU_GET(domain);
2963 	else
2964 		domain = 0;
2965 	zdom = &zone->uz_domain[0];
2966 
2967 	/* Can we throw this on the zone full list? */
2968 	if (bucket != NULL) {
2969 		CTR3(KTR_UMA,
2970 		    "uma_zfree: zone %s(%p) putting bucket %p on free list",
2971 		    zone->uz_name, zone, bucket);
2972 		/* ub_cnt is pointing to the last free item */
2973 		KASSERT(bucket->ub_cnt != 0,
2974 		    ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n"));
2975 		if ((zone->uz_flags & UMA_ZONE_NOBUCKETCACHE) != 0) {
2976 			ZONE_UNLOCK(zone);
2977 			bucket_drain(zone, bucket);
2978 			bucket_free(zone, bucket, udata);
2979 			goto zfree_restart;
2980 		} else
2981 			LIST_INSERT_HEAD(&zdom->uzd_buckets, bucket, ub_link);
2982 	}
2983 
2984 	/*
2985 	 * We bump the uz count when the cache size is insufficient to
2986 	 * handle the working set.
2987 	 */
2988 	if (lockfail && zone->uz_count < BUCKET_MAX)
2989 		zone->uz_count++;
2990 	ZONE_UNLOCK(zone);
2991 
2992 	bucket = bucket_alloc(zone, udata, M_NOWAIT);
2993 	CTR3(KTR_UMA, "uma_zfree: zone %s(%p) allocated bucket %p",
2994 	    zone->uz_name, zone, bucket);
2995 	if (bucket) {
2996 		critical_enter();
2997 		cpu = curcpu;
2998 		cache = &zone->uz_cpu[cpu];
2999 		if (cache->uc_freebucket == NULL &&
3000 		    ((zone->uz_flags & UMA_ZONE_NUMA) == 0 ||
3001 		    domain == PCPU_GET(domain))) {
3002 			cache->uc_freebucket = bucket;
3003 			goto zfree_start;
3004 		}
3005 		/*
3006 		 * We lost the race, start over.  We have to drop our
3007 		 * critical section to free the bucket.
3008 		 */
3009 		critical_exit();
3010 		bucket_free(zone, bucket, udata);
3011 		goto zfree_restart;
3012 	}
3013 
3014 	/*
3015 	 * If nothing else caught this, we'll just do an internal free.
3016 	 */
3017 zfree_item:
3018 	zone_free_item(zone, item, udata, SKIP_DTOR);
3019 
3020 	return;
3021 }
3022 
3023 void
3024 uma_zfree_domain(uma_zone_t zone, void *item, void *udata)
3025 {
3026 
3027 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3028 	random_harvest_fast_uma(&zone, sizeof(zone), 1, RANDOM_UMA);
3029 
3030 	CTR2(KTR_UMA, "uma_zfree_domain thread %x zone %s", curthread,
3031 	    zone->uz_name);
3032 
3033 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3034 	    ("uma_zfree_domain: called with spinlock or critical section held"));
3035 
3036         /* uma_zfree(..., NULL) does nothing, to match free(9). */
3037         if (item == NULL)
3038                 return;
3039 	zone_free_item(zone, item, udata, SKIP_NONE);
3040 }
3041 
3042 static void
3043 slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item)
3044 {
3045 	uma_domain_t dom;
3046 	uint8_t freei;
3047 
3048 	mtx_assert(&keg->uk_lock, MA_OWNED);
3049 	MPASS(keg == slab->us_keg);
3050 
3051 	dom = &keg->uk_domain[slab->us_domain];
3052 
3053 	/* Do we need to remove from any lists? */
3054 	if (slab->us_freecount+1 == keg->uk_ipers) {
3055 		LIST_REMOVE(slab, us_link);
3056 		LIST_INSERT_HEAD(&dom->ud_free_slab, slab, us_link);
3057 	} else if (slab->us_freecount == 0) {
3058 		LIST_REMOVE(slab, us_link);
3059 		LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
3060 	}
3061 
3062 	/* Slab management. */
3063 	freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
3064 	BIT_SET(SLAB_SETSIZE, freei, &slab->us_free);
3065 	slab->us_freecount++;
3066 
3067 	/* Keg statistics. */
3068 	keg->uk_free++;
3069 }
3070 
3071 static void
3072 zone_release(uma_zone_t zone, void **bucket, int cnt)
3073 {
3074 	void *item;
3075 	uma_slab_t slab;
3076 	uma_keg_t keg;
3077 	uint8_t *mem;
3078 	int clearfull;
3079 	int i;
3080 
3081 	clearfull = 0;
3082 	keg = zone_first_keg(zone);
3083 	KEG_LOCK(keg);
3084 	for (i = 0; i < cnt; i++) {
3085 		item = bucket[i];
3086 		if (!(zone->uz_flags & UMA_ZONE_VTOSLAB)) {
3087 			mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
3088 			if (zone->uz_flags & UMA_ZONE_HASH) {
3089 				slab = hash_sfind(&keg->uk_hash, mem);
3090 			} else {
3091 				mem += keg->uk_pgoff;
3092 				slab = (uma_slab_t)mem;
3093 			}
3094 		} else {
3095 			slab = vtoslab((vm_offset_t)item);
3096 			if (slab->us_keg != keg) {
3097 				KEG_UNLOCK(keg);
3098 				keg = slab->us_keg;
3099 				KEG_LOCK(keg);
3100 			}
3101 		}
3102 		slab_free_item(keg, slab, item);
3103 		if (keg->uk_flags & UMA_ZFLAG_FULL) {
3104 			if (keg->uk_pages < keg->uk_maxpages) {
3105 				keg->uk_flags &= ~UMA_ZFLAG_FULL;
3106 				clearfull = 1;
3107 			}
3108 
3109 			/*
3110 			 * We can handle one more allocation. Since we're
3111 			 * clearing ZFLAG_FULL, wake up all procs blocked
3112 			 * on pages. This should be uncommon, so keeping this
3113 			 * simple for now (rather than adding count of blocked
3114 			 * threads etc).
3115 			 */
3116 			wakeup(keg);
3117 		}
3118 	}
3119 	KEG_UNLOCK(keg);
3120 	if (clearfull) {
3121 		ZONE_LOCK(zone);
3122 		zone->uz_flags &= ~UMA_ZFLAG_FULL;
3123 		wakeup(zone);
3124 		ZONE_UNLOCK(zone);
3125 	}
3126 
3127 }
3128 
3129 /*
3130  * Frees a single item to any zone.
3131  *
3132  * Arguments:
3133  *	zone   The zone to free to
3134  *	item   The item we're freeing
3135  *	udata  User supplied data for the dtor
3136  *	skip   Skip dtors and finis
3137  */
3138 static void
3139 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
3140 {
3141 
3142 #ifdef INVARIANTS
3143 	if (skip == SKIP_NONE) {
3144 		if (zone->uz_flags & UMA_ZONE_MALLOC)
3145 			uma_dbg_free(zone, udata, item);
3146 		else
3147 			uma_dbg_free(zone, NULL, item);
3148 	}
3149 #endif
3150 	if (skip < SKIP_DTOR && zone->uz_dtor)
3151 		zone->uz_dtor(item, zone->uz_size, udata);
3152 
3153 	if (skip < SKIP_FINI && zone->uz_fini)
3154 		zone->uz_fini(item, zone->uz_size);
3155 
3156 	atomic_add_long(&zone->uz_frees, 1);
3157 	zone->uz_release(zone->uz_arg, &item, 1);
3158 }
3159 
3160 /* See uma.h */
3161 int
3162 uma_zone_set_max(uma_zone_t zone, int nitems)
3163 {
3164 	uma_keg_t keg;
3165 
3166 	keg = zone_first_keg(zone);
3167 	if (keg == NULL)
3168 		return (0);
3169 	KEG_LOCK(keg);
3170 	keg->uk_maxpages = (nitems / keg->uk_ipers) * keg->uk_ppera;
3171 	if (keg->uk_maxpages * keg->uk_ipers < nitems)
3172 		keg->uk_maxpages += keg->uk_ppera;
3173 	nitems = (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers;
3174 	KEG_UNLOCK(keg);
3175 
3176 	return (nitems);
3177 }
3178 
3179 /* See uma.h */
3180 int
3181 uma_zone_get_max(uma_zone_t zone)
3182 {
3183 	int nitems;
3184 	uma_keg_t keg;
3185 
3186 	keg = zone_first_keg(zone);
3187 	if (keg == NULL)
3188 		return (0);
3189 	KEG_LOCK(keg);
3190 	nitems = (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers;
3191 	KEG_UNLOCK(keg);
3192 
3193 	return (nitems);
3194 }
3195 
3196 /* See uma.h */
3197 void
3198 uma_zone_set_warning(uma_zone_t zone, const char *warning)
3199 {
3200 
3201 	ZONE_LOCK(zone);
3202 	zone->uz_warning = warning;
3203 	ZONE_UNLOCK(zone);
3204 }
3205 
3206 /* See uma.h */
3207 void
3208 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction)
3209 {
3210 
3211 	ZONE_LOCK(zone);
3212 	TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone);
3213 	ZONE_UNLOCK(zone);
3214 }
3215 
3216 /* See uma.h */
3217 int
3218 uma_zone_get_cur(uma_zone_t zone)
3219 {
3220 	int64_t nitems;
3221 	u_int i;
3222 
3223 	ZONE_LOCK(zone);
3224 	nitems = zone->uz_allocs - zone->uz_frees;
3225 	CPU_FOREACH(i) {
3226 		/*
3227 		 * See the comment in sysctl_vm_zone_stats() regarding the
3228 		 * safety of accessing the per-cpu caches. With the zone lock
3229 		 * held, it is safe, but can potentially result in stale data.
3230 		 */
3231 		nitems += zone->uz_cpu[i].uc_allocs -
3232 		    zone->uz_cpu[i].uc_frees;
3233 	}
3234 	ZONE_UNLOCK(zone);
3235 
3236 	return (nitems < 0 ? 0 : nitems);
3237 }
3238 
3239 /* See uma.h */
3240 void
3241 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
3242 {
3243 	uma_keg_t keg;
3244 
3245 	keg = zone_first_keg(zone);
3246 	KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type"));
3247 	KEG_LOCK(keg);
3248 	KASSERT(keg->uk_pages == 0,
3249 	    ("uma_zone_set_init on non-empty keg"));
3250 	keg->uk_init = uminit;
3251 	KEG_UNLOCK(keg);
3252 }
3253 
3254 /* See uma.h */
3255 void
3256 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
3257 {
3258 	uma_keg_t keg;
3259 
3260 	keg = zone_first_keg(zone);
3261 	KASSERT(keg != NULL, ("uma_zone_set_fini: Invalid zone type"));
3262 	KEG_LOCK(keg);
3263 	KASSERT(keg->uk_pages == 0,
3264 	    ("uma_zone_set_fini on non-empty keg"));
3265 	keg->uk_fini = fini;
3266 	KEG_UNLOCK(keg);
3267 }
3268 
3269 /* See uma.h */
3270 void
3271 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
3272 {
3273 
3274 	ZONE_LOCK(zone);
3275 	KASSERT(zone_first_keg(zone)->uk_pages == 0,
3276 	    ("uma_zone_set_zinit on non-empty keg"));
3277 	zone->uz_init = zinit;
3278 	ZONE_UNLOCK(zone);
3279 }
3280 
3281 /* See uma.h */
3282 void
3283 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
3284 {
3285 
3286 	ZONE_LOCK(zone);
3287 	KASSERT(zone_first_keg(zone)->uk_pages == 0,
3288 	    ("uma_zone_set_zfini on non-empty keg"));
3289 	zone->uz_fini = zfini;
3290 	ZONE_UNLOCK(zone);
3291 }
3292 
3293 /* See uma.h */
3294 /* XXX uk_freef is not actually used with the zone locked */
3295 void
3296 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
3297 {
3298 	uma_keg_t keg;
3299 
3300 	keg = zone_first_keg(zone);
3301 	KASSERT(keg != NULL, ("uma_zone_set_freef: Invalid zone type"));
3302 	KEG_LOCK(keg);
3303 	keg->uk_freef = freef;
3304 	KEG_UNLOCK(keg);
3305 }
3306 
3307 /* See uma.h */
3308 /* XXX uk_allocf is not actually used with the zone locked */
3309 void
3310 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
3311 {
3312 	uma_keg_t keg;
3313 
3314 	keg = zone_first_keg(zone);
3315 	KEG_LOCK(keg);
3316 	keg->uk_allocf = allocf;
3317 	KEG_UNLOCK(keg);
3318 }
3319 
3320 /* See uma.h */
3321 void
3322 uma_zone_reserve(uma_zone_t zone, int items)
3323 {
3324 	uma_keg_t keg;
3325 
3326 	keg = zone_first_keg(zone);
3327 	if (keg == NULL)
3328 		return;
3329 	KEG_LOCK(keg);
3330 	keg->uk_reserve = items;
3331 	KEG_UNLOCK(keg);
3332 
3333 	return;
3334 }
3335 
3336 /* See uma.h */
3337 int
3338 uma_zone_reserve_kva(uma_zone_t zone, int count)
3339 {
3340 	uma_keg_t keg;
3341 	vm_offset_t kva;
3342 	u_int pages;
3343 
3344 	keg = zone_first_keg(zone);
3345 	if (keg == NULL)
3346 		return (0);
3347 	pages = count / keg->uk_ipers;
3348 
3349 	if (pages * keg->uk_ipers < count)
3350 		pages++;
3351 	pages *= keg->uk_ppera;
3352 
3353 #ifdef UMA_MD_SMALL_ALLOC
3354 	if (keg->uk_ppera > 1) {
3355 #else
3356 	if (1) {
3357 #endif
3358 		kva = kva_alloc((vm_size_t)pages * PAGE_SIZE);
3359 		if (kva == 0)
3360 			return (0);
3361 	} else
3362 		kva = 0;
3363 	KEG_LOCK(keg);
3364 	keg->uk_kva = kva;
3365 	keg->uk_offset = 0;
3366 	keg->uk_maxpages = pages;
3367 #ifdef UMA_MD_SMALL_ALLOC
3368 	keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
3369 #else
3370 	keg->uk_allocf = noobj_alloc;
3371 #endif
3372 	keg->uk_flags |= UMA_ZONE_NOFREE;
3373 	KEG_UNLOCK(keg);
3374 
3375 	return (1);
3376 }
3377 
3378 /* See uma.h */
3379 void
3380 uma_prealloc(uma_zone_t zone, int items)
3381 {
3382 	uma_domain_t dom;
3383 	uma_slab_t slab;
3384 	uma_keg_t keg;
3385 	int domain, slabs;
3386 
3387 	keg = zone_first_keg(zone);
3388 	if (keg == NULL)
3389 		return;
3390 	KEG_LOCK(keg);
3391 	slabs = items / keg->uk_ipers;
3392 	domain = 0;
3393 	if (slabs * keg->uk_ipers < items)
3394 		slabs++;
3395 	while (slabs > 0) {
3396 		slab = keg_alloc_slab(keg, zone, domain, M_WAITOK);
3397 		if (slab == NULL)
3398 			break;
3399 		MPASS(slab->us_keg == keg);
3400 		dom = &keg->uk_domain[slab->us_domain];
3401 		LIST_INSERT_HEAD(&dom->ud_free_slab, slab, us_link);
3402 		slabs--;
3403 		domain = (domain + 1) % vm_ndomains;
3404 	}
3405 	KEG_UNLOCK(keg);
3406 }
3407 
3408 /* See uma.h */
3409 static void
3410 uma_reclaim_locked(bool kmem_danger)
3411 {
3412 
3413 	CTR0(KTR_UMA, "UMA: vm asked us to release pages!");
3414 	sx_assert(&uma_drain_lock, SA_XLOCKED);
3415 	bucket_enable();
3416 	zone_foreach(zone_drain);
3417 	if (vm_page_count_min() || kmem_danger) {
3418 		cache_drain_safe(NULL);
3419 		zone_foreach(zone_drain);
3420 	}
3421 	/*
3422 	 * Some slabs may have been freed but this zone will be visited early
3423 	 * we visit again so that we can free pages that are empty once other
3424 	 * zones are drained.  We have to do the same for buckets.
3425 	 */
3426 	zone_drain(slabzone);
3427 	bucket_zone_drain();
3428 }
3429 
3430 void
3431 uma_reclaim(void)
3432 {
3433 
3434 	sx_xlock(&uma_drain_lock);
3435 	uma_reclaim_locked(false);
3436 	sx_xunlock(&uma_drain_lock);
3437 }
3438 
3439 static volatile int uma_reclaim_needed;
3440 
3441 void
3442 uma_reclaim_wakeup(void)
3443 {
3444 
3445 	if (atomic_fetchadd_int(&uma_reclaim_needed, 1) == 0)
3446 		wakeup(uma_reclaim);
3447 }
3448 
3449 void
3450 uma_reclaim_worker(void *arg __unused)
3451 {
3452 
3453 	for (;;) {
3454 		sx_xlock(&uma_drain_lock);
3455 		while (atomic_load_int(&uma_reclaim_needed) == 0)
3456 			sx_sleep(uma_reclaim, &uma_drain_lock, PVM, "umarcl",
3457 			    hz);
3458 		sx_xunlock(&uma_drain_lock);
3459 		EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM);
3460 		sx_xlock(&uma_drain_lock);
3461 		uma_reclaim_locked(true);
3462 		atomic_store_int(&uma_reclaim_needed, 0);
3463 		sx_xunlock(&uma_drain_lock);
3464 		/* Don't fire more than once per-second. */
3465 		pause("umarclslp", hz);
3466 	}
3467 }
3468 
3469 /* See uma.h */
3470 int
3471 uma_zone_exhausted(uma_zone_t zone)
3472 {
3473 	int full;
3474 
3475 	ZONE_LOCK(zone);
3476 	full = (zone->uz_flags & UMA_ZFLAG_FULL);
3477 	ZONE_UNLOCK(zone);
3478 	return (full);
3479 }
3480 
3481 int
3482 uma_zone_exhausted_nolock(uma_zone_t zone)
3483 {
3484 	return (zone->uz_flags & UMA_ZFLAG_FULL);
3485 }
3486 
3487 void *
3488 uma_large_malloc_domain(vm_size_t size, int domain, int wait)
3489 {
3490 	vm_offset_t addr;
3491 	uma_slab_t slab;
3492 
3493 	slab = zone_alloc_item(slabzone, NULL, domain, wait);
3494 	if (slab == NULL)
3495 		return (NULL);
3496 	if (domain == UMA_ANYDOMAIN)
3497 		addr = kmem_malloc(kernel_arena, size, wait);
3498 	else
3499 		addr = kmem_malloc_domain(domain, size, wait);
3500 	if (addr != 0) {
3501 		vsetslab(addr, slab);
3502 		slab->us_data = (void *)addr;
3503 		slab->us_flags = UMA_SLAB_KERNEL | UMA_SLAB_MALLOC;
3504 		slab->us_size = size;
3505 		slab->us_domain = vm_phys_domain(PHYS_TO_VM_PAGE(
3506 		    pmap_kextract(addr)));
3507 		uma_total_inc(size);
3508 	} else {
3509 		zone_free_item(slabzone, slab, NULL, SKIP_NONE);
3510 	}
3511 
3512 	return ((void *)addr);
3513 }
3514 
3515 void *
3516 uma_large_malloc(vm_size_t size, int wait)
3517 {
3518 
3519 	return uma_large_malloc_domain(size, UMA_ANYDOMAIN, wait);
3520 }
3521 
3522 void
3523 uma_large_free(uma_slab_t slab)
3524 {
3525 
3526 	KASSERT((slab->us_flags & UMA_SLAB_KERNEL) != 0,
3527 	    ("uma_large_free:  Memory not allocated with uma_large_malloc."));
3528 	kmem_free(kernel_arena, (vm_offset_t)slab->us_data, slab->us_size);
3529 	uma_total_dec(slab->us_size);
3530 	zone_free_item(slabzone, slab, NULL, SKIP_NONE);
3531 }
3532 
3533 static void
3534 uma_zero_item(void *item, uma_zone_t zone)
3535 {
3536 	int i;
3537 
3538 	if (zone->uz_flags & UMA_ZONE_PCPU) {
3539 		CPU_FOREACH(i)
3540 			bzero(zpcpu_get_cpu(item, i), zone->uz_size);
3541 	} else
3542 		bzero(item, zone->uz_size);
3543 }
3544 
3545 unsigned long
3546 uma_limit(void)
3547 {
3548 
3549 	return (uma_kmem_limit);
3550 }
3551 
3552 void
3553 uma_set_limit(unsigned long limit)
3554 {
3555 
3556 	uma_kmem_limit = limit;
3557 }
3558 
3559 unsigned long
3560 uma_size(void)
3561 {
3562 
3563 	return (uma_kmem_total);
3564 }
3565 
3566 long
3567 uma_avail(void)
3568 {
3569 
3570 	return (uma_kmem_limit - uma_kmem_total);
3571 }
3572 
3573 void
3574 uma_print_stats(void)
3575 {
3576 	zone_foreach(uma_print_zone);
3577 }
3578 
3579 static void
3580 slab_print(uma_slab_t slab)
3581 {
3582 	printf("slab: keg %p, data %p, freecount %d\n",
3583 		slab->us_keg, slab->us_data, slab->us_freecount);
3584 }
3585 
3586 static void
3587 cache_print(uma_cache_t cache)
3588 {
3589 	printf("alloc: %p(%d), free: %p(%d)\n",
3590 		cache->uc_allocbucket,
3591 		cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0,
3592 		cache->uc_freebucket,
3593 		cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0);
3594 }
3595 
3596 static void
3597 uma_print_keg(uma_keg_t keg)
3598 {
3599 	uma_domain_t dom;
3600 	uma_slab_t slab;
3601 	int i;
3602 
3603 	printf("keg: %s(%p) size %d(%d) flags %#x ipers %d ppera %d "
3604 	    "out %d free %d limit %d\n",
3605 	    keg->uk_name, keg, keg->uk_size, keg->uk_rsize, keg->uk_flags,
3606 	    keg->uk_ipers, keg->uk_ppera,
3607 	    (keg->uk_pages / keg->uk_ppera) * keg->uk_ipers - keg->uk_free,
3608 	    keg->uk_free, (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers);
3609 	for (i = 0; i < vm_ndomains; i++) {
3610 		dom = &keg->uk_domain[i];
3611 		printf("Part slabs:\n");
3612 		LIST_FOREACH(slab, &dom->ud_part_slab, us_link)
3613 			slab_print(slab);
3614 		printf("Free slabs:\n");
3615 		LIST_FOREACH(slab, &dom->ud_free_slab, us_link)
3616 			slab_print(slab);
3617 		printf("Full slabs:\n");
3618 		LIST_FOREACH(slab, &dom->ud_full_slab, us_link)
3619 			slab_print(slab);
3620 	}
3621 }
3622 
3623 void
3624 uma_print_zone(uma_zone_t zone)
3625 {
3626 	uma_cache_t cache;
3627 	uma_klink_t kl;
3628 	int i;
3629 
3630 	printf("zone: %s(%p) size %d flags %#x\n",
3631 	    zone->uz_name, zone, zone->uz_size, zone->uz_flags);
3632 	LIST_FOREACH(kl, &zone->uz_kegs, kl_link)
3633 		uma_print_keg(kl->kl_keg);
3634 	CPU_FOREACH(i) {
3635 		cache = &zone->uz_cpu[i];
3636 		printf("CPU %d Cache:\n", i);
3637 		cache_print(cache);
3638 	}
3639 }
3640 
3641 #ifdef DDB
3642 /*
3643  * Generate statistics across both the zone and its per-cpu cache's.  Return
3644  * desired statistics if the pointer is non-NULL for that statistic.
3645  *
3646  * Note: does not update the zone statistics, as it can't safely clear the
3647  * per-CPU cache statistic.
3648  *
3649  * XXXRW: Following the uc_allocbucket and uc_freebucket pointers here isn't
3650  * safe from off-CPU; we should modify the caches to track this information
3651  * directly so that we don't have to.
3652  */
3653 static void
3654 uma_zone_sumstat(uma_zone_t z, int *cachefreep, uint64_t *allocsp,
3655     uint64_t *freesp, uint64_t *sleepsp)
3656 {
3657 	uma_cache_t cache;
3658 	uint64_t allocs, frees, sleeps;
3659 	int cachefree, cpu;
3660 
3661 	allocs = frees = sleeps = 0;
3662 	cachefree = 0;
3663 	CPU_FOREACH(cpu) {
3664 		cache = &z->uz_cpu[cpu];
3665 		if (cache->uc_allocbucket != NULL)
3666 			cachefree += cache->uc_allocbucket->ub_cnt;
3667 		if (cache->uc_freebucket != NULL)
3668 			cachefree += cache->uc_freebucket->ub_cnt;
3669 		allocs += cache->uc_allocs;
3670 		frees += cache->uc_frees;
3671 	}
3672 	allocs += z->uz_allocs;
3673 	frees += z->uz_frees;
3674 	sleeps += z->uz_sleeps;
3675 	if (cachefreep != NULL)
3676 		*cachefreep = cachefree;
3677 	if (allocsp != NULL)
3678 		*allocsp = allocs;
3679 	if (freesp != NULL)
3680 		*freesp = frees;
3681 	if (sleepsp != NULL)
3682 		*sleepsp = sleeps;
3683 }
3684 #endif /* DDB */
3685 
3686 static int
3687 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
3688 {
3689 	uma_keg_t kz;
3690 	uma_zone_t z;
3691 	int count;
3692 
3693 	count = 0;
3694 	rw_rlock(&uma_rwlock);
3695 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3696 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
3697 			count++;
3698 	}
3699 	rw_runlock(&uma_rwlock);
3700 	return (sysctl_handle_int(oidp, &count, 0, req));
3701 }
3702 
3703 static int
3704 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
3705 {
3706 	struct uma_stream_header ush;
3707 	struct uma_type_header uth;
3708 	struct uma_percpu_stat *ups;
3709 	uma_bucket_t bucket;
3710 	uma_zone_domain_t zdom;
3711 	struct sbuf sbuf;
3712 	uma_cache_t cache;
3713 	uma_klink_t kl;
3714 	uma_keg_t kz;
3715 	uma_zone_t z;
3716 	uma_keg_t k;
3717 	int count, error, i;
3718 
3719 	error = sysctl_wire_old_buffer(req, 0);
3720 	if (error != 0)
3721 		return (error);
3722 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
3723 	sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL);
3724 	ups = malloc((mp_maxid + 1) * sizeof(*ups), M_TEMP, M_WAITOK);
3725 
3726 	count = 0;
3727 	rw_rlock(&uma_rwlock);
3728 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3729 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
3730 			count++;
3731 	}
3732 
3733 	/*
3734 	 * Insert stream header.
3735 	 */
3736 	bzero(&ush, sizeof(ush));
3737 	ush.ush_version = UMA_STREAM_VERSION;
3738 	ush.ush_maxcpus = (mp_maxid + 1);
3739 	ush.ush_count = count;
3740 	(void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
3741 
3742 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3743 		LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3744 			bzero(&uth, sizeof(uth));
3745 			ZONE_LOCK(z);
3746 			strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
3747 			uth.uth_align = kz->uk_align;
3748 			uth.uth_size = kz->uk_size;
3749 			uth.uth_rsize = kz->uk_rsize;
3750 			LIST_FOREACH(kl, &z->uz_kegs, kl_link) {
3751 				k = kl->kl_keg;
3752 				uth.uth_maxpages += k->uk_maxpages;
3753 				uth.uth_pages += k->uk_pages;
3754 				uth.uth_keg_free += k->uk_free;
3755 				uth.uth_limit = (k->uk_maxpages / k->uk_ppera)
3756 				    * k->uk_ipers;
3757 			}
3758 
3759 			/*
3760 			 * A zone is secondary is it is not the first entry
3761 			 * on the keg's zone list.
3762 			 */
3763 			if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
3764 			    (LIST_FIRST(&kz->uk_zones) != z))
3765 				uth.uth_zone_flags = UTH_ZONE_SECONDARY;
3766 
3767 			for (i = 0; i < vm_ndomains; i++) {
3768 				zdom = &z->uz_domain[i];
3769 				LIST_FOREACH(bucket, &zdom->uzd_buckets,
3770 				    ub_link)
3771 					uth.uth_zone_free += bucket->ub_cnt;
3772 			}
3773 			uth.uth_allocs = z->uz_allocs;
3774 			uth.uth_frees = z->uz_frees;
3775 			uth.uth_fails = z->uz_fails;
3776 			uth.uth_sleeps = z->uz_sleeps;
3777 			/*
3778 			 * While it is not normally safe to access the cache
3779 			 * bucket pointers while not on the CPU that owns the
3780 			 * cache, we only allow the pointers to be exchanged
3781 			 * without the zone lock held, not invalidated, so
3782 			 * accept the possible race associated with bucket
3783 			 * exchange during monitoring.
3784 			 */
3785 			for (i = 0; i < mp_maxid + 1; i++) {
3786 				bzero(&ups[i], sizeof(*ups));
3787 				if (kz->uk_flags & UMA_ZFLAG_INTERNAL ||
3788 				    CPU_ABSENT(i))
3789 					continue;
3790 				cache = &z->uz_cpu[i];
3791 				if (cache->uc_allocbucket != NULL)
3792 					ups[i].ups_cache_free +=
3793 					    cache->uc_allocbucket->ub_cnt;
3794 				if (cache->uc_freebucket != NULL)
3795 					ups[i].ups_cache_free +=
3796 					    cache->uc_freebucket->ub_cnt;
3797 				ups[i].ups_allocs = cache->uc_allocs;
3798 				ups[i].ups_frees = cache->uc_frees;
3799 			}
3800 			ZONE_UNLOCK(z);
3801 			(void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
3802 			for (i = 0; i < mp_maxid + 1; i++)
3803 				(void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
3804 		}
3805 	}
3806 	rw_runlock(&uma_rwlock);
3807 	error = sbuf_finish(&sbuf);
3808 	sbuf_delete(&sbuf);
3809 	free(ups, M_TEMP);
3810 	return (error);
3811 }
3812 
3813 int
3814 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS)
3815 {
3816 	uma_zone_t zone = *(uma_zone_t *)arg1;
3817 	int error, max;
3818 
3819 	max = uma_zone_get_max(zone);
3820 	error = sysctl_handle_int(oidp, &max, 0, req);
3821 	if (error || !req->newptr)
3822 		return (error);
3823 
3824 	uma_zone_set_max(zone, max);
3825 
3826 	return (0);
3827 }
3828 
3829 int
3830 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS)
3831 {
3832 	uma_zone_t zone = *(uma_zone_t *)arg1;
3833 	int cur;
3834 
3835 	cur = uma_zone_get_cur(zone);
3836 	return (sysctl_handle_int(oidp, &cur, 0, req));
3837 }
3838 
3839 #ifdef INVARIANTS
3840 static uma_slab_t
3841 uma_dbg_getslab(uma_zone_t zone, void *item)
3842 {
3843 	uma_slab_t slab;
3844 	uma_keg_t keg;
3845 	uint8_t *mem;
3846 
3847 	mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
3848 	if (zone->uz_flags & UMA_ZONE_VTOSLAB) {
3849 		slab = vtoslab((vm_offset_t)mem);
3850 	} else {
3851 		/*
3852 		 * It is safe to return the slab here even though the
3853 		 * zone is unlocked because the item's allocation state
3854 		 * essentially holds a reference.
3855 		 */
3856 		ZONE_LOCK(zone);
3857 		keg = LIST_FIRST(&zone->uz_kegs)->kl_keg;
3858 		if (keg->uk_flags & UMA_ZONE_HASH)
3859 			slab = hash_sfind(&keg->uk_hash, mem);
3860 		else
3861 			slab = (uma_slab_t)(mem + keg->uk_pgoff);
3862 		ZONE_UNLOCK(zone);
3863 	}
3864 
3865 	return (slab);
3866 }
3867 
3868 /*
3869  * Set up the slab's freei data such that uma_dbg_free can function.
3870  *
3871  */
3872 static void
3873 uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item)
3874 {
3875 	uma_keg_t keg;
3876 	int freei;
3877 
3878 	if (zone_first_keg(zone) == NULL)
3879 		return;
3880 	if (slab == NULL) {
3881 		slab = uma_dbg_getslab(zone, item);
3882 		if (slab == NULL)
3883 			panic("uma: item %p did not belong to zone %s\n",
3884 			    item, zone->uz_name);
3885 	}
3886 	keg = slab->us_keg;
3887 	freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
3888 
3889 	if (BIT_ISSET(SLAB_SETSIZE, freei, &slab->us_debugfree))
3890 		panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)\n",
3891 		    item, zone, zone->uz_name, slab, freei);
3892 	BIT_SET_ATOMIC(SLAB_SETSIZE, freei, &slab->us_debugfree);
3893 
3894 	return;
3895 }
3896 
3897 /*
3898  * Verifies freed addresses.  Checks for alignment, valid slab membership
3899  * and duplicate frees.
3900  *
3901  */
3902 static void
3903 uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item)
3904 {
3905 	uma_keg_t keg;
3906 	int freei;
3907 
3908 	if (zone_first_keg(zone) == NULL)
3909 		return;
3910 	if (slab == NULL) {
3911 		slab = uma_dbg_getslab(zone, item);
3912 		if (slab == NULL)
3913 			panic("uma: Freed item %p did not belong to zone %s\n",
3914 			    item, zone->uz_name);
3915 	}
3916 	keg = slab->us_keg;
3917 	freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
3918 
3919 	if (freei >= keg->uk_ipers)
3920 		panic("Invalid free of %p from zone %p(%s) slab %p(%d)\n",
3921 		    item, zone, zone->uz_name, slab, freei);
3922 
3923 	if (((freei * keg->uk_rsize) + slab->us_data) != item)
3924 		panic("Unaligned free of %p from zone %p(%s) slab %p(%d)\n",
3925 		    item, zone, zone->uz_name, slab, freei);
3926 
3927 	if (!BIT_ISSET(SLAB_SETSIZE, freei, &slab->us_debugfree))
3928 		panic("Duplicate free of %p from zone %p(%s) slab %p(%d)\n",
3929 		    item, zone, zone->uz_name, slab, freei);
3930 
3931 	BIT_CLR_ATOMIC(SLAB_SETSIZE, freei, &slab->us_debugfree);
3932 }
3933 #endif /* INVARIANTS */
3934 
3935 #ifdef DDB
3936 DB_SHOW_COMMAND(uma, db_show_uma)
3937 {
3938 	uma_bucket_t bucket;
3939 	uma_keg_t kz;
3940 	uma_zone_t z;
3941 	uma_zone_domain_t zdom;
3942 	uint64_t allocs, frees, sleeps;
3943 	int cachefree, i;
3944 
3945 	db_printf("%18s %8s %8s %8s %12s %8s %8s\n", "Zone", "Size", "Used",
3946 	    "Free", "Requests", "Sleeps", "Bucket");
3947 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3948 		LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3949 			if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
3950 				allocs = z->uz_allocs;
3951 				frees = z->uz_frees;
3952 				sleeps = z->uz_sleeps;
3953 				cachefree = 0;
3954 			} else
3955 				uma_zone_sumstat(z, &cachefree, &allocs,
3956 				    &frees, &sleeps);
3957 			if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
3958 			    (LIST_FIRST(&kz->uk_zones) != z)))
3959 				cachefree += kz->uk_free;
3960 			for (i = 0; i < vm_ndomains; i++) {
3961 				zdom = &z->uz_domain[i];
3962 				LIST_FOREACH(bucket, &zdom->uzd_buckets,
3963 				    ub_link)
3964 					cachefree += bucket->ub_cnt;
3965 			}
3966 			db_printf("%18s %8ju %8jd %8d %12ju %8ju %8u\n",
3967 			    z->uz_name, (uintmax_t)kz->uk_size,
3968 			    (intmax_t)(allocs - frees), cachefree,
3969 			    (uintmax_t)allocs, sleeps, z->uz_count);
3970 			if (db_pager_quit)
3971 				return;
3972 		}
3973 	}
3974 }
3975 
3976 DB_SHOW_COMMAND(umacache, db_show_umacache)
3977 {
3978 	uma_bucket_t bucket;
3979 	uma_zone_t z;
3980 	uma_zone_domain_t zdom;
3981 	uint64_t allocs, frees;
3982 	int cachefree, i;
3983 
3984 	db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
3985 	    "Requests", "Bucket");
3986 	LIST_FOREACH(z, &uma_cachezones, uz_link) {
3987 		uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL);
3988 		for (i = 0; i < vm_ndomains; i++) {
3989 			zdom = &z->uz_domain[i];
3990 			LIST_FOREACH(bucket, &zdom->uzd_buckets, ub_link)
3991 				cachefree += bucket->ub_cnt;
3992 		}
3993 		db_printf("%18s %8ju %8jd %8d %12ju %8u\n",
3994 		    z->uz_name, (uintmax_t)z->uz_size,
3995 		    (intmax_t)(allocs - frees), cachefree,
3996 		    (uintmax_t)allocs, z->uz_count);
3997 		if (db_pager_quit)
3998 			return;
3999 	}
4000 }
4001 #endif	/* DDB */
4002