xref: /freebsd/sys/vm/uma_core.c (revision d8a0fe102c0cfdfcd5b818f850eff09d8536c9bc)
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/queue.h>
66 #include <sys/malloc.h>
67 #include <sys/ktr.h>
68 #include <sys/lock.h>
69 #include <sys/sysctl.h>
70 #include <sys/mutex.h>
71 #include <sys/proc.h>
72 #include <sys/random.h>
73 #include <sys/rwlock.h>
74 #include <sys/sbuf.h>
75 #include <sys/sched.h>
76 #include <sys/smp.h>
77 #include <sys/taskqueue.h>
78 #include <sys/vmmeter.h>
79 
80 #include <vm/vm.h>
81 #include <vm/vm_object.h>
82 #include <vm/vm_page.h>
83 #include <vm/vm_pageout.h>
84 #include <vm/vm_param.h>
85 #include <vm/vm_map.h>
86 #include <vm/vm_kern.h>
87 #include <vm/vm_extern.h>
88 #include <vm/uma.h>
89 #include <vm/uma_int.h>
90 #include <vm/uma_dbg.h>
91 
92 #include <ddb/ddb.h>
93 
94 #ifdef DEBUG_MEMGUARD
95 #include <vm/memguard.h>
96 #endif
97 
98 /*
99  * This is the zone and keg from which all zones are spawned.  The idea is that
100  * even the zone & keg heads are allocated from the allocator, so we use the
101  * bss section to bootstrap us.
102  */
103 static struct uma_keg masterkeg;
104 static struct uma_zone masterzone_k;
105 static struct uma_zone masterzone_z;
106 static uma_zone_t kegs = &masterzone_k;
107 static uma_zone_t zones = &masterzone_z;
108 
109 /* This is the zone from which all of uma_slab_t's are allocated. */
110 static uma_zone_t slabzone;
111 
112 /*
113  * The initial hash tables come out of this zone so they can be allocated
114  * prior to malloc coming up.
115  */
116 static uma_zone_t hashzone;
117 
118 /* The boot-time adjusted value for cache line alignment. */
119 int uma_align_cache = 64 - 1;
120 
121 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets");
122 
123 /*
124  * Are we allowed to allocate buckets?
125  */
126 static int bucketdisable = 1;
127 
128 /* Linked list of all kegs in the system */
129 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs);
130 
131 /* Linked list of all cache-only zones in the system */
132 static LIST_HEAD(,uma_zone) uma_cachezones =
133     LIST_HEAD_INITIALIZER(uma_cachezones);
134 
135 /* This RW lock protects the keg list */
136 static struct rwlock_padalign __exclusive_cache_line uma_rwlock;
137 
138 /*
139  * Pointer and counter to pool of pages, that is preallocated at
140  * startup to bootstrap UMA.  Early zones continue to use the pool
141  * until it is depleted, so allocations may happen after boot, thus
142  * we need a mutex to protect it.
143  */
144 static char *bootmem;
145 static int boot_pages;
146 static struct mtx uma_boot_pages_mtx;
147 
148 static struct sx uma_drain_lock;
149 
150 /* kmem soft limit. */
151 static unsigned long uma_kmem_limit;
152 static volatile unsigned long uma_kmem_total;
153 
154 /* Is the VM done starting up? */
155 static int booted = 0;
156 #define	UMA_STARTUP	1
157 #define	UMA_STARTUP2	2
158 
159 /*
160  * This is the handle used to schedule events that need to happen
161  * outside of the allocation fast path.
162  */
163 static struct callout uma_callout;
164 #define	UMA_TIMEOUT	20		/* Seconds for callout interval. */
165 
166 /*
167  * This structure is passed as the zone ctor arg so that I don't have to create
168  * a special allocation function just for zones.
169  */
170 struct uma_zctor_args {
171 	const char *name;
172 	size_t size;
173 	uma_ctor ctor;
174 	uma_dtor dtor;
175 	uma_init uminit;
176 	uma_fini fini;
177 	uma_import import;
178 	uma_release release;
179 	void *arg;
180 	uma_keg_t keg;
181 	int align;
182 	uint32_t flags;
183 };
184 
185 struct uma_kctor_args {
186 	uma_zone_t zone;
187 	size_t size;
188 	uma_init uminit;
189 	uma_fini fini;
190 	int align;
191 	uint32_t flags;
192 };
193 
194 struct uma_bucket_zone {
195 	uma_zone_t	ubz_zone;
196 	char		*ubz_name;
197 	int		ubz_entries;	/* Number of items it can hold. */
198 	int		ubz_maxsize;	/* Maximum allocation size per-item. */
199 };
200 
201 /*
202  * Compute the actual number of bucket entries to pack them in power
203  * of two sizes for more efficient space utilization.
204  */
205 #define	BUCKET_SIZE(n)						\
206     (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *))
207 
208 #define	BUCKET_MAX	BUCKET_SIZE(256)
209 
210 struct uma_bucket_zone bucket_zones[] = {
211 	{ NULL, "4 Bucket", BUCKET_SIZE(4), 4096 },
212 	{ NULL, "6 Bucket", BUCKET_SIZE(6), 3072 },
213 	{ NULL, "8 Bucket", BUCKET_SIZE(8), 2048 },
214 	{ NULL, "12 Bucket", BUCKET_SIZE(12), 1536 },
215 	{ NULL, "16 Bucket", BUCKET_SIZE(16), 1024 },
216 	{ NULL, "32 Bucket", BUCKET_SIZE(32), 512 },
217 	{ NULL, "64 Bucket", BUCKET_SIZE(64), 256 },
218 	{ NULL, "128 Bucket", BUCKET_SIZE(128), 128 },
219 	{ NULL, "256 Bucket", BUCKET_SIZE(256), 64 },
220 	{ NULL, NULL, 0}
221 };
222 
223 /*
224  * Flags and enumerations to be passed to internal functions.
225  */
226 enum zfreeskip { SKIP_NONE = 0, SKIP_DTOR, SKIP_FINI };
227 
228 /* Prototypes.. */
229 
230 static void *noobj_alloc(uma_zone_t, vm_size_t, uint8_t *, int);
231 static void *page_alloc(uma_zone_t, vm_size_t, uint8_t *, int);
232 static void *startup_alloc(uma_zone_t, vm_size_t, 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);
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);
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 zone, void *, int flags);
260 static uma_slab_t zone_fetch_slab(uma_zone_t zone, uma_keg_t last, int flags);
261 static uma_slab_t zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int flags);
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 zone, void **bucket, int max, int flags);
267 static void zone_release(uma_zone_t zone, void **bucket, int cnt);
268 static void uma_zero_item(void *item, uma_zone_t zone);
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);
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 		    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 
739 	if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
740 		return;
741 
742 	b1 = b2 = NULL;
743 	ZONE_LOCK(zone);
744 	critical_enter();
745 	cache = &zone->uz_cpu[curcpu];
746 	if (cache->uc_allocbucket) {
747 		if (cache->uc_allocbucket->ub_cnt != 0)
748 			LIST_INSERT_HEAD(&zone->uz_buckets,
749 			    cache->uc_allocbucket, ub_link);
750 		else
751 			b1 = cache->uc_allocbucket;
752 		cache->uc_allocbucket = NULL;
753 	}
754 	if (cache->uc_freebucket) {
755 		if (cache->uc_freebucket->ub_cnt != 0)
756 			LIST_INSERT_HEAD(&zone->uz_buckets,
757 			    cache->uc_freebucket, ub_link);
758 		else
759 			b2 = cache->uc_freebucket;
760 		cache->uc_freebucket = NULL;
761 	}
762 	critical_exit();
763 	ZONE_UNLOCK(zone);
764 	if (b1)
765 		bucket_free(zone, b1, NULL);
766 	if (b2)
767 		bucket_free(zone, b2, NULL);
768 }
769 
770 /*
771  * Safely drain per-CPU caches of a zone(s) to alloc bucket.
772  * This is an expensive call because it needs to bind to all CPUs
773  * one by one and enter a critical section on each of them in order
774  * to safely access their cache buckets.
775  * Zone lock must not be held on call this function.
776  */
777 static void
778 cache_drain_safe(uma_zone_t zone)
779 {
780 	int cpu;
781 
782 	/*
783 	 * Polite bucket sizes shrinking was not enouth, shrink aggressively.
784 	 */
785 	if (zone)
786 		cache_shrink(zone);
787 	else
788 		zone_foreach(cache_shrink);
789 
790 	CPU_FOREACH(cpu) {
791 		thread_lock(curthread);
792 		sched_bind(curthread, cpu);
793 		thread_unlock(curthread);
794 
795 		if (zone)
796 			cache_drain_safe_cpu(zone);
797 		else
798 			zone_foreach(cache_drain_safe_cpu);
799 	}
800 	thread_lock(curthread);
801 	sched_unbind(curthread);
802 	thread_unlock(curthread);
803 }
804 
805 /*
806  * Drain the cached buckets from a zone.  Expects a locked zone on entry.
807  */
808 static void
809 bucket_cache_drain(uma_zone_t zone)
810 {
811 	uma_bucket_t bucket;
812 
813 	/*
814 	 * Drain the bucket queues and free the buckets, we just keep two per
815 	 * cpu (alloc/free).
816 	 */
817 	while ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) {
818 		LIST_REMOVE(bucket, ub_link);
819 		ZONE_UNLOCK(zone);
820 		bucket_drain(zone, bucket);
821 		bucket_free(zone, bucket, NULL);
822 		ZONE_LOCK(zone);
823 	}
824 
825 	/*
826 	 * Shrink further bucket sizes.  Price of single zone lock collision
827 	 * is probably lower then price of global cache drain.
828 	 */
829 	if (zone->uz_count > zone->uz_count_min)
830 		zone->uz_count--;
831 }
832 
833 static void
834 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start)
835 {
836 	uint8_t *mem;
837 	int i;
838 	uint8_t flags;
839 
840 	CTR4(KTR_UMA, "keg_free_slab keg %s(%p) slab %p, returning %d bytes",
841 	    keg->uk_name, keg, slab, PAGE_SIZE * keg->uk_ppera);
842 
843 	mem = slab->us_data;
844 	flags = slab->us_flags;
845 	i = start;
846 	if (keg->uk_fini != NULL) {
847 		for (i--; i > -1; i--)
848 			keg->uk_fini(slab->us_data + (keg->uk_rsize * i),
849 			    keg->uk_size);
850 	}
851 	if (keg->uk_flags & UMA_ZONE_OFFPAGE)
852 		zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE);
853 	keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags);
854 	uma_total_dec(PAGE_SIZE * keg->uk_ppera);
855 }
856 
857 /*
858  * Frees pages from a keg back to the system.  This is done on demand from
859  * the pageout daemon.
860  *
861  * Returns nothing.
862  */
863 static void
864 keg_drain(uma_keg_t keg)
865 {
866 	struct slabhead freeslabs = { 0 };
867 	uma_slab_t slab, tmp;
868 
869 	/*
870 	 * We don't want to take pages from statically allocated kegs at this
871 	 * time
872 	 */
873 	if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL)
874 		return;
875 
876 	CTR3(KTR_UMA, "keg_drain %s(%p) free items: %u",
877 	    keg->uk_name, keg, keg->uk_free);
878 	KEG_LOCK(keg);
879 	if (keg->uk_free == 0)
880 		goto finished;
881 
882 	LIST_FOREACH_SAFE(slab, &keg->uk_free_slab, us_link, tmp) {
883 		/* We have nowhere to free these to. */
884 		if (slab->us_flags & UMA_SLAB_BOOT)
885 			continue;
886 
887 		LIST_REMOVE(slab, us_link);
888 		keg->uk_pages -= keg->uk_ppera;
889 		keg->uk_free -= keg->uk_ipers;
890 
891 		if (keg->uk_flags & UMA_ZONE_HASH)
892 			UMA_HASH_REMOVE(&keg->uk_hash, slab, slab->us_data);
893 
894 		SLIST_INSERT_HEAD(&freeslabs, slab, us_hlink);
895 	}
896 finished:
897 	KEG_UNLOCK(keg);
898 
899 	while ((slab = SLIST_FIRST(&freeslabs)) != NULL) {
900 		SLIST_REMOVE(&freeslabs, slab, uma_slab, us_hlink);
901 		keg_free_slab(keg, slab, keg->uk_ipers);
902 	}
903 }
904 
905 static void
906 zone_drain_wait(uma_zone_t zone, int waitok)
907 {
908 
909 	/*
910 	 * Set draining to interlock with zone_dtor() so we can release our
911 	 * locks as we go.  Only dtor() should do a WAITOK call since it
912 	 * is the only call that knows the structure will still be available
913 	 * when it wakes up.
914 	 */
915 	ZONE_LOCK(zone);
916 	while (zone->uz_flags & UMA_ZFLAG_DRAINING) {
917 		if (waitok == M_NOWAIT)
918 			goto out;
919 		msleep(zone, zone->uz_lockptr, PVM, "zonedrain", 1);
920 	}
921 	zone->uz_flags |= UMA_ZFLAG_DRAINING;
922 	bucket_cache_drain(zone);
923 	ZONE_UNLOCK(zone);
924 	/*
925 	 * The DRAINING flag protects us from being freed while
926 	 * we're running.  Normally the uma_rwlock would protect us but we
927 	 * must be able to release and acquire the right lock for each keg.
928 	 */
929 	zone_foreach_keg(zone, &keg_drain);
930 	ZONE_LOCK(zone);
931 	zone->uz_flags &= ~UMA_ZFLAG_DRAINING;
932 	wakeup(zone);
933 out:
934 	ZONE_UNLOCK(zone);
935 }
936 
937 void
938 zone_drain(uma_zone_t zone)
939 {
940 
941 	zone_drain_wait(zone, M_NOWAIT);
942 }
943 
944 /*
945  * Allocate a new slab for a keg.  This does not insert the slab onto a list.
946  *
947  * Arguments:
948  *	wait  Shall we wait?
949  *
950  * Returns:
951  *	The slab that was allocated or NULL if there is no memory and the
952  *	caller specified M_NOWAIT.
953  */
954 static uma_slab_t
955 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int wait)
956 {
957 	uma_alloc allocf;
958 	uma_slab_t slab;
959 	unsigned long size;
960 	uint8_t *mem;
961 	uint8_t flags;
962 	int i;
963 
964 	mtx_assert(&keg->uk_lock, MA_OWNED);
965 	slab = NULL;
966 	mem = NULL;
967 
968 	allocf = keg->uk_allocf;
969 	KEG_UNLOCK(keg);
970 	size = keg->uk_ppera * PAGE_SIZE;
971 
972 	if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
973 		slab = zone_alloc_item(keg->uk_slabzone, NULL, wait);
974 		if (slab == NULL)
975 			goto out;
976 	}
977 
978 	/*
979 	 * This reproduces the old vm_zone behavior of zero filling pages the
980 	 * first time they are added to a zone.
981 	 *
982 	 * Malloced items are zeroed in uma_zalloc.
983 	 */
984 
985 	if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
986 		wait |= M_ZERO;
987 	else
988 		wait &= ~M_ZERO;
989 
990 	if (keg->uk_flags & UMA_ZONE_NODUMP)
991 		wait |= M_NODUMP;
992 
993 	/* zone is passed for legacy reasons. */
994 	mem = allocf(zone, size, &flags, wait);
995 	if (mem == NULL) {
996 		if (keg->uk_flags & UMA_ZONE_OFFPAGE)
997 			zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE);
998 		slab = NULL;
999 		goto out;
1000 	}
1001 	uma_total_inc(size);
1002 
1003 	/* Point the slab into the allocated memory */
1004 	if (!(keg->uk_flags & UMA_ZONE_OFFPAGE))
1005 		slab = (uma_slab_t )(mem + keg->uk_pgoff);
1006 
1007 	if (keg->uk_flags & UMA_ZONE_VTOSLAB)
1008 		for (i = 0; i < keg->uk_ppera; i++)
1009 			vsetslab((vm_offset_t)mem + (i * PAGE_SIZE), slab);
1010 
1011 	slab->us_keg = keg;
1012 	slab->us_data = mem;
1013 	slab->us_freecount = keg->uk_ipers;
1014 	slab->us_flags = flags;
1015 	BIT_FILL(SLAB_SETSIZE, &slab->us_free);
1016 #ifdef INVARIANTS
1017 	BIT_ZERO(SLAB_SETSIZE, &slab->us_debugfree);
1018 #endif
1019 
1020 	if (keg->uk_init != NULL) {
1021 		for (i = 0; i < keg->uk_ipers; i++)
1022 			if (keg->uk_init(slab->us_data + (keg->uk_rsize * i),
1023 			    keg->uk_size, wait) != 0)
1024 				break;
1025 		if (i != keg->uk_ipers) {
1026 			keg_free_slab(keg, slab, i);
1027 			slab = NULL;
1028 			goto out;
1029 		}
1030 	}
1031 out:
1032 	KEG_LOCK(keg);
1033 
1034 	CTR3(KTR_UMA, "keg_alloc_slab: allocated slab %p for %s(%p)",
1035 	    slab, keg->uk_name, keg);
1036 
1037 	if (slab != NULL) {
1038 		if (keg->uk_flags & UMA_ZONE_HASH)
1039 			UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
1040 
1041 		keg->uk_pages += keg->uk_ppera;
1042 		keg->uk_free += keg->uk_ipers;
1043 	}
1044 
1045 	return (slab);
1046 }
1047 
1048 /*
1049  * This function is intended to be used early on in place of page_alloc() so
1050  * that we may use the boot time page cache to satisfy allocations before
1051  * the VM is ready.
1052  */
1053 static void *
1054 startup_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *pflag, int wait)
1055 {
1056 	uma_keg_t keg;
1057 	void *mem;
1058 	int pages;
1059 
1060 	keg = zone_first_keg(zone);
1061 	pages = howmany(bytes, PAGE_SIZE);
1062 	KASSERT(pages > 0, ("startup_alloc can't reserve 0 pages\n"));
1063 
1064 	/*
1065 	 * Check our small startup cache to see if it has pages remaining.
1066 	 */
1067 	mtx_lock(&uma_boot_pages_mtx);
1068 	if (pages <= boot_pages) {
1069 		mem = bootmem;
1070 		boot_pages -= pages;
1071 		bootmem += pages * PAGE_SIZE;
1072 		mtx_unlock(&uma_boot_pages_mtx);
1073 		*pflag = UMA_SLAB_BOOT;
1074 		return (mem);
1075 	}
1076 	mtx_unlock(&uma_boot_pages_mtx);
1077 	if (booted < UMA_STARTUP2)
1078 		panic("UMA: Increase vm.boot_pages");
1079 	/*
1080 	 * Now that we've booted reset these users to their real allocator.
1081 	 */
1082 #ifdef UMA_MD_SMALL_ALLOC
1083 	keg->uk_allocf = (keg->uk_ppera > 1) ? page_alloc : uma_small_alloc;
1084 #else
1085 	keg->uk_allocf = page_alloc;
1086 #endif
1087 	return keg->uk_allocf(zone, bytes, pflag, wait);
1088 }
1089 
1090 /*
1091  * Allocates a number of pages from the system
1092  *
1093  * Arguments:
1094  *	bytes  The number of bytes requested
1095  *	wait  Shall we wait?
1096  *
1097  * Returns:
1098  *	A pointer to the alloced memory or possibly
1099  *	NULL if M_NOWAIT is set.
1100  */
1101 static void *
1102 page_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *pflag, int wait)
1103 {
1104 	void *p;	/* Returned page */
1105 
1106 	*pflag = UMA_SLAB_KERNEL;
1107 	p = (void *) kmem_malloc(kernel_arena, bytes, wait);
1108 
1109 	return (p);
1110 }
1111 
1112 /*
1113  * Allocates a number of pages from within an object
1114  *
1115  * Arguments:
1116  *	bytes  The number of bytes requested
1117  *	wait   Shall we wait?
1118  *
1119  * Returns:
1120  *	A pointer to the alloced memory or possibly
1121  *	NULL if M_NOWAIT is set.
1122  */
1123 static void *
1124 noobj_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *flags, int wait)
1125 {
1126 	TAILQ_HEAD(, vm_page) alloctail;
1127 	u_long npages;
1128 	vm_offset_t retkva, zkva;
1129 	vm_page_t p, p_next;
1130 	uma_keg_t keg;
1131 
1132 	TAILQ_INIT(&alloctail);
1133 	keg = zone_first_keg(zone);
1134 
1135 	npages = howmany(bytes, PAGE_SIZE);
1136 	while (npages > 0) {
1137 		p = vm_page_alloc(NULL, 0, VM_ALLOC_INTERRUPT |
1138 		    VM_ALLOC_WIRED | VM_ALLOC_NOOBJ |
1139 		    ((wait & M_WAITOK) != 0 ? VM_ALLOC_WAITOK :
1140 		    VM_ALLOC_NOWAIT));
1141 		if (p != NULL) {
1142 			/*
1143 			 * Since the page does not belong to an object, its
1144 			 * listq is unused.
1145 			 */
1146 			TAILQ_INSERT_TAIL(&alloctail, p, listq);
1147 			npages--;
1148 			continue;
1149 		}
1150 		/*
1151 		 * Page allocation failed, free intermediate pages and
1152 		 * exit.
1153 		 */
1154 		TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1155 			vm_page_unwire(p, PQ_NONE);
1156 			vm_page_free(p);
1157 		}
1158 		return (NULL);
1159 	}
1160 	*flags = UMA_SLAB_PRIV;
1161 	zkva = keg->uk_kva +
1162 	    atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
1163 	retkva = zkva;
1164 	TAILQ_FOREACH(p, &alloctail, listq) {
1165 		pmap_qenter(zkva, &p, 1);
1166 		zkva += PAGE_SIZE;
1167 	}
1168 
1169 	return ((void *)retkva);
1170 }
1171 
1172 /*
1173  * Frees a number of pages to the system
1174  *
1175  * Arguments:
1176  *	mem   A pointer to the memory to be freed
1177  *	size  The size of the memory being freed
1178  *	flags The original p->us_flags field
1179  *
1180  * Returns:
1181  *	Nothing
1182  */
1183 static void
1184 page_free(void *mem, vm_size_t size, uint8_t flags)
1185 {
1186 	struct vmem *vmem;
1187 
1188 	if (flags & UMA_SLAB_KERNEL)
1189 		vmem = kernel_arena;
1190 	else
1191 		panic("UMA: page_free used with invalid flags %x", flags);
1192 
1193 	kmem_free(vmem, (vm_offset_t)mem, size);
1194 }
1195 
1196 /*
1197  * Zero fill initializer
1198  *
1199  * Arguments/Returns follow uma_init specifications
1200  */
1201 static int
1202 zero_init(void *mem, int size, int flags)
1203 {
1204 	bzero(mem, size);
1205 	return (0);
1206 }
1207 
1208 /*
1209  * Finish creating a small uma keg.  This calculates ipers, and the keg size.
1210  *
1211  * Arguments
1212  *	keg  The zone we should initialize
1213  *
1214  * Returns
1215  *	Nothing
1216  */
1217 static void
1218 keg_small_init(uma_keg_t keg)
1219 {
1220 	u_int rsize;
1221 	u_int memused;
1222 	u_int wastedspace;
1223 	u_int shsize;
1224 	u_int slabsize;
1225 
1226 	if (keg->uk_flags & UMA_ZONE_PCPU) {
1227 		u_int ncpus = (mp_maxid + 1) ? (mp_maxid + 1) : MAXCPU;
1228 
1229 		slabsize = sizeof(struct pcpu);
1230 		keg->uk_ppera = howmany(ncpus * sizeof(struct pcpu),
1231 		    PAGE_SIZE);
1232 	} else {
1233 		slabsize = UMA_SLAB_SIZE;
1234 		keg->uk_ppera = 1;
1235 	}
1236 
1237 	/*
1238 	 * Calculate the size of each allocation (rsize) according to
1239 	 * alignment.  If the requested size is smaller than we have
1240 	 * allocation bits for we round it up.
1241 	 */
1242 	rsize = keg->uk_size;
1243 	if (rsize < slabsize / SLAB_SETSIZE)
1244 		rsize = slabsize / SLAB_SETSIZE;
1245 	if (rsize & keg->uk_align)
1246 		rsize = (rsize & ~keg->uk_align) + (keg->uk_align + 1);
1247 	keg->uk_rsize = rsize;
1248 
1249 	KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
1250 	    keg->uk_rsize < sizeof(struct pcpu),
1251 	    ("%s: size %u too large", __func__, keg->uk_rsize));
1252 
1253 	if (keg->uk_flags & UMA_ZONE_OFFPAGE)
1254 		shsize = 0;
1255 	else
1256 		shsize = sizeof(struct uma_slab);
1257 
1258 	keg->uk_ipers = (slabsize - shsize) / rsize;
1259 	KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE,
1260 	    ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers));
1261 
1262 	memused = keg->uk_ipers * rsize + shsize;
1263 	wastedspace = slabsize - memused;
1264 
1265 	/*
1266 	 * We can't do OFFPAGE if we're internal or if we've been
1267 	 * asked to not go to the VM for buckets.  If we do this we
1268 	 * may end up going to the VM  for slabs which we do not
1269 	 * want to do if we're UMA_ZFLAG_CACHEONLY as a result
1270 	 * of UMA_ZONE_VM, which clearly forbids it.
1271 	 */
1272 	if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) ||
1273 	    (keg->uk_flags & UMA_ZFLAG_CACHEONLY))
1274 		return;
1275 
1276 	/*
1277 	 * See if using an OFFPAGE slab will limit our waste.  Only do
1278 	 * this if it permits more items per-slab.
1279 	 *
1280 	 * XXX We could try growing slabsize to limit max waste as well.
1281 	 * Historically this was not done because the VM could not
1282 	 * efficiently handle contiguous allocations.
1283 	 */
1284 	if ((wastedspace >= slabsize / UMA_MAX_WASTE) &&
1285 	    (keg->uk_ipers < (slabsize / keg->uk_rsize))) {
1286 		keg->uk_ipers = slabsize / keg->uk_rsize;
1287 		KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE,
1288 		    ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers));
1289 		CTR6(KTR_UMA, "UMA decided we need offpage slab headers for "
1290 		    "keg: %s(%p), calculated wastedspace = %d, "
1291 		    "maximum wasted space allowed = %d, "
1292 		    "calculated ipers = %d, "
1293 		    "new wasted space = %d\n", keg->uk_name, keg, wastedspace,
1294 		    slabsize / UMA_MAX_WASTE, keg->uk_ipers,
1295 		    slabsize - keg->uk_ipers * keg->uk_rsize);
1296 		keg->uk_flags |= UMA_ZONE_OFFPAGE;
1297 	}
1298 
1299 	if ((keg->uk_flags & UMA_ZONE_OFFPAGE) &&
1300 	    (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1301 		keg->uk_flags |= UMA_ZONE_HASH;
1302 }
1303 
1304 /*
1305  * Finish creating a large (> UMA_SLAB_SIZE) uma kegs.  Just give in and do
1306  * OFFPAGE for now.  When I can allow for more dynamic slab sizes this will be
1307  * more complicated.
1308  *
1309  * Arguments
1310  *	keg  The keg we should initialize
1311  *
1312  * Returns
1313  *	Nothing
1314  */
1315 static void
1316 keg_large_init(uma_keg_t keg)
1317 {
1318 	u_int shsize;
1319 
1320 	KASSERT(keg != NULL, ("Keg is null in keg_large_init"));
1321 	KASSERT((keg->uk_flags & UMA_ZFLAG_CACHEONLY) == 0,
1322 	    ("keg_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY keg"));
1323 	KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0,
1324 	    ("%s: Cannot large-init a UMA_ZONE_PCPU keg", __func__));
1325 
1326 	keg->uk_ppera = howmany(keg->uk_size, PAGE_SIZE);
1327 	keg->uk_ipers = 1;
1328 	keg->uk_rsize = keg->uk_size;
1329 
1330 	/* Check whether we have enough space to not do OFFPAGE. */
1331 	if ((keg->uk_flags & UMA_ZONE_OFFPAGE) == 0) {
1332 		shsize = sizeof(struct uma_slab);
1333 		if (shsize & UMA_ALIGN_PTR)
1334 			shsize = (shsize & ~UMA_ALIGN_PTR) +
1335 			    (UMA_ALIGN_PTR + 1);
1336 
1337 		if (PAGE_SIZE * keg->uk_ppera - keg->uk_rsize < shsize) {
1338 			/*
1339 			 * We can't do OFFPAGE if we're internal, in which case
1340 			 * we need an extra page per allocation to contain the
1341 			 * slab header.
1342 			 */
1343 			if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) == 0)
1344 				keg->uk_flags |= UMA_ZONE_OFFPAGE;
1345 			else
1346 				keg->uk_ppera++;
1347 		}
1348 	}
1349 
1350 	if ((keg->uk_flags & UMA_ZONE_OFFPAGE) &&
1351 	    (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1352 		keg->uk_flags |= UMA_ZONE_HASH;
1353 }
1354 
1355 static void
1356 keg_cachespread_init(uma_keg_t keg)
1357 {
1358 	int alignsize;
1359 	int trailer;
1360 	int pages;
1361 	int rsize;
1362 
1363 	KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0,
1364 	    ("%s: Cannot cachespread-init a UMA_ZONE_PCPU keg", __func__));
1365 
1366 	alignsize = keg->uk_align + 1;
1367 	rsize = keg->uk_size;
1368 	/*
1369 	 * We want one item to start on every align boundary in a page.  To
1370 	 * do this we will span pages.  We will also extend the item by the
1371 	 * size of align if it is an even multiple of align.  Otherwise, it
1372 	 * would fall on the same boundary every time.
1373 	 */
1374 	if (rsize & keg->uk_align)
1375 		rsize = (rsize & ~keg->uk_align) + alignsize;
1376 	if ((rsize & alignsize) == 0)
1377 		rsize += alignsize;
1378 	trailer = rsize - keg->uk_size;
1379 	pages = (rsize * (PAGE_SIZE / alignsize)) / PAGE_SIZE;
1380 	pages = MIN(pages, (128 * 1024) / PAGE_SIZE);
1381 	keg->uk_rsize = rsize;
1382 	keg->uk_ppera = pages;
1383 	keg->uk_ipers = ((pages * PAGE_SIZE) + trailer) / rsize;
1384 	keg->uk_flags |= UMA_ZONE_OFFPAGE | UMA_ZONE_VTOSLAB;
1385 	KASSERT(keg->uk_ipers <= SLAB_SETSIZE,
1386 	    ("%s: keg->uk_ipers too high(%d) increase max_ipers", __func__,
1387 	    keg->uk_ipers));
1388 }
1389 
1390 /*
1391  * Keg header ctor.  This initializes all fields, locks, etc.  And inserts
1392  * the keg onto the global keg list.
1393  *
1394  * Arguments/Returns follow uma_ctor specifications
1395  *	udata  Actually uma_kctor_args
1396  */
1397 static int
1398 keg_ctor(void *mem, int size, void *udata, int flags)
1399 {
1400 	struct uma_kctor_args *arg = udata;
1401 	uma_keg_t keg = mem;
1402 	uma_zone_t zone;
1403 
1404 	bzero(keg, size);
1405 	keg->uk_size = arg->size;
1406 	keg->uk_init = arg->uminit;
1407 	keg->uk_fini = arg->fini;
1408 	keg->uk_align = arg->align;
1409 	keg->uk_free = 0;
1410 	keg->uk_reserve = 0;
1411 	keg->uk_pages = 0;
1412 	keg->uk_flags = arg->flags;
1413 	keg->uk_slabzone = NULL;
1414 
1415 	/*
1416 	 * The master zone is passed to us at keg-creation time.
1417 	 */
1418 	zone = arg->zone;
1419 	keg->uk_name = zone->uz_name;
1420 
1421 	if (arg->flags & UMA_ZONE_VM)
1422 		keg->uk_flags |= UMA_ZFLAG_CACHEONLY;
1423 
1424 	if (arg->flags & UMA_ZONE_ZINIT)
1425 		keg->uk_init = zero_init;
1426 
1427 	if (arg->flags & UMA_ZONE_MALLOC)
1428 		keg->uk_flags |= UMA_ZONE_VTOSLAB;
1429 
1430 	if (arg->flags & UMA_ZONE_PCPU)
1431 #ifdef SMP
1432 		keg->uk_flags |= UMA_ZONE_OFFPAGE;
1433 #else
1434 		keg->uk_flags &= ~UMA_ZONE_PCPU;
1435 #endif
1436 
1437 	if (keg->uk_flags & UMA_ZONE_CACHESPREAD) {
1438 		keg_cachespread_init(keg);
1439 	} else {
1440 		if (keg->uk_size > (UMA_SLAB_SIZE - sizeof(struct uma_slab)))
1441 			keg_large_init(keg);
1442 		else
1443 			keg_small_init(keg);
1444 	}
1445 
1446 	if (keg->uk_flags & UMA_ZONE_OFFPAGE)
1447 		keg->uk_slabzone = slabzone;
1448 
1449 	/*
1450 	 * If we haven't booted yet we need allocations to go through the
1451 	 * startup cache until the vm is ready.
1452 	 */
1453 	if (booted < UMA_STARTUP2)
1454 		keg->uk_allocf = startup_alloc;
1455 #ifdef UMA_MD_SMALL_ALLOC
1456 	else if (keg->uk_ppera == 1)
1457 		keg->uk_allocf = uma_small_alloc;
1458 #endif
1459 	else
1460 		keg->uk_allocf = page_alloc;
1461 #ifdef UMA_MD_SMALL_ALLOC
1462 	if (keg->uk_ppera == 1)
1463 		keg->uk_freef = uma_small_free;
1464 	else
1465 #endif
1466 		keg->uk_freef = page_free;
1467 
1468 	/*
1469 	 * Initialize keg's lock
1470 	 */
1471 	KEG_LOCK_INIT(keg, (arg->flags & UMA_ZONE_MTXCLASS));
1472 
1473 	/*
1474 	 * If we're putting the slab header in the actual page we need to
1475 	 * figure out where in each page it goes.  This calculates a right
1476 	 * justified offset into the memory on an ALIGN_PTR boundary.
1477 	 */
1478 	if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) {
1479 		u_int totsize;
1480 
1481 		/* Size of the slab struct and free list */
1482 		totsize = sizeof(struct uma_slab);
1483 
1484 		if (totsize & UMA_ALIGN_PTR)
1485 			totsize = (totsize & ~UMA_ALIGN_PTR) +
1486 			    (UMA_ALIGN_PTR + 1);
1487 		keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - totsize;
1488 
1489 		/*
1490 		 * The only way the following is possible is if with our
1491 		 * UMA_ALIGN_PTR adjustments we are now bigger than
1492 		 * UMA_SLAB_SIZE.  I haven't checked whether this is
1493 		 * mathematically possible for all cases, so we make
1494 		 * sure here anyway.
1495 		 */
1496 		totsize = keg->uk_pgoff + sizeof(struct uma_slab);
1497 		if (totsize > PAGE_SIZE * keg->uk_ppera) {
1498 			printf("zone %s ipers %d rsize %d size %d\n",
1499 			    zone->uz_name, keg->uk_ipers, keg->uk_rsize,
1500 			    keg->uk_size);
1501 			panic("UMA slab won't fit.");
1502 		}
1503 	}
1504 
1505 	if (keg->uk_flags & UMA_ZONE_HASH)
1506 		hash_alloc(&keg->uk_hash);
1507 
1508 	CTR5(KTR_UMA, "keg_ctor %p zone %s(%p) out %d free %d\n",
1509 	    keg, zone->uz_name, zone,
1510 	    (keg->uk_pages / keg->uk_ppera) * keg->uk_ipers - keg->uk_free,
1511 	    keg->uk_free);
1512 
1513 	LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
1514 
1515 	rw_wlock(&uma_rwlock);
1516 	LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
1517 	rw_wunlock(&uma_rwlock);
1518 	return (0);
1519 }
1520 
1521 /*
1522  * Zone header ctor.  This initializes all fields, locks, etc.
1523  *
1524  * Arguments/Returns follow uma_ctor specifications
1525  *	udata  Actually uma_zctor_args
1526  */
1527 static int
1528 zone_ctor(void *mem, int size, void *udata, int flags)
1529 {
1530 	struct uma_zctor_args *arg = udata;
1531 	uma_zone_t zone = mem;
1532 	uma_zone_t z;
1533 	uma_keg_t keg;
1534 
1535 	bzero(zone, size);
1536 	zone->uz_name = arg->name;
1537 	zone->uz_ctor = arg->ctor;
1538 	zone->uz_dtor = arg->dtor;
1539 	zone->uz_slab = zone_fetch_slab;
1540 	zone->uz_init = NULL;
1541 	zone->uz_fini = NULL;
1542 	zone->uz_allocs = 0;
1543 	zone->uz_frees = 0;
1544 	zone->uz_fails = 0;
1545 	zone->uz_sleeps = 0;
1546 	zone->uz_count = 0;
1547 	zone->uz_count_min = 0;
1548 	zone->uz_flags = 0;
1549 	zone->uz_warning = NULL;
1550 	timevalclear(&zone->uz_ratecheck);
1551 	keg = arg->keg;
1552 
1553 	ZONE_LOCK_INIT(zone, (arg->flags & UMA_ZONE_MTXCLASS));
1554 
1555 	/*
1556 	 * This is a pure cache zone, no kegs.
1557 	 */
1558 	if (arg->import) {
1559 		if (arg->flags & UMA_ZONE_VM)
1560 			arg->flags |= UMA_ZFLAG_CACHEONLY;
1561 		zone->uz_flags = arg->flags;
1562 		zone->uz_size = arg->size;
1563 		zone->uz_import = arg->import;
1564 		zone->uz_release = arg->release;
1565 		zone->uz_arg = arg->arg;
1566 		zone->uz_lockptr = &zone->uz_lock;
1567 		rw_wlock(&uma_rwlock);
1568 		LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link);
1569 		rw_wunlock(&uma_rwlock);
1570 		goto out;
1571 	}
1572 
1573 	/*
1574 	 * Use the regular zone/keg/slab allocator.
1575 	 */
1576 	zone->uz_import = (uma_import)zone_import;
1577 	zone->uz_release = (uma_release)zone_release;
1578 	zone->uz_arg = zone;
1579 
1580 	if (arg->flags & UMA_ZONE_SECONDARY) {
1581 		KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
1582 		zone->uz_init = arg->uminit;
1583 		zone->uz_fini = arg->fini;
1584 		zone->uz_lockptr = &keg->uk_lock;
1585 		zone->uz_flags |= UMA_ZONE_SECONDARY;
1586 		rw_wlock(&uma_rwlock);
1587 		ZONE_LOCK(zone);
1588 		LIST_FOREACH(z, &keg->uk_zones, uz_link) {
1589 			if (LIST_NEXT(z, uz_link) == NULL) {
1590 				LIST_INSERT_AFTER(z, zone, uz_link);
1591 				break;
1592 			}
1593 		}
1594 		ZONE_UNLOCK(zone);
1595 		rw_wunlock(&uma_rwlock);
1596 	} else if (keg == NULL) {
1597 		if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
1598 		    arg->align, arg->flags)) == NULL)
1599 			return (ENOMEM);
1600 	} else {
1601 		struct uma_kctor_args karg;
1602 		int error;
1603 
1604 		/* We should only be here from uma_startup() */
1605 		karg.size = arg->size;
1606 		karg.uminit = arg->uminit;
1607 		karg.fini = arg->fini;
1608 		karg.align = arg->align;
1609 		karg.flags = arg->flags;
1610 		karg.zone = zone;
1611 		error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
1612 		    flags);
1613 		if (error)
1614 			return (error);
1615 	}
1616 
1617 	/*
1618 	 * Link in the first keg.
1619 	 */
1620 	zone->uz_klink.kl_keg = keg;
1621 	LIST_INSERT_HEAD(&zone->uz_kegs, &zone->uz_klink, kl_link);
1622 	zone->uz_lockptr = &keg->uk_lock;
1623 	zone->uz_size = keg->uk_size;
1624 	zone->uz_flags |= (keg->uk_flags &
1625 	    (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
1626 
1627 	/*
1628 	 * Some internal zones don't have room allocated for the per cpu
1629 	 * caches.  If we're internal, bail out here.
1630 	 */
1631 	if (keg->uk_flags & UMA_ZFLAG_INTERNAL) {
1632 		KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
1633 		    ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
1634 		return (0);
1635 	}
1636 
1637 out:
1638 	if ((arg->flags & UMA_ZONE_MAXBUCKET) == 0)
1639 		zone->uz_count = bucket_select(zone->uz_size);
1640 	else
1641 		zone->uz_count = BUCKET_MAX;
1642 	zone->uz_count_min = zone->uz_count;
1643 
1644 	return (0);
1645 }
1646 
1647 /*
1648  * Keg header dtor.  This frees all data, destroys locks, frees the hash
1649  * table and removes the keg from the global list.
1650  *
1651  * Arguments/Returns follow uma_dtor specifications
1652  *	udata  unused
1653  */
1654 static void
1655 keg_dtor(void *arg, int size, void *udata)
1656 {
1657 	uma_keg_t keg;
1658 
1659 	keg = (uma_keg_t)arg;
1660 	KEG_LOCK(keg);
1661 	if (keg->uk_free != 0) {
1662 		printf("Freed UMA keg (%s) was not empty (%d items). "
1663 		    " Lost %d pages of memory.\n",
1664 		    keg->uk_name ? keg->uk_name : "",
1665 		    keg->uk_free, keg->uk_pages);
1666 	}
1667 	KEG_UNLOCK(keg);
1668 
1669 	hash_free(&keg->uk_hash);
1670 
1671 	KEG_LOCK_FINI(keg);
1672 }
1673 
1674 /*
1675  * Zone header dtor.
1676  *
1677  * Arguments/Returns follow uma_dtor specifications
1678  *	udata  unused
1679  */
1680 static void
1681 zone_dtor(void *arg, int size, void *udata)
1682 {
1683 	uma_klink_t klink;
1684 	uma_zone_t zone;
1685 	uma_keg_t keg;
1686 
1687 	zone = (uma_zone_t)arg;
1688 	keg = zone_first_keg(zone);
1689 
1690 	if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
1691 		cache_drain(zone);
1692 
1693 	rw_wlock(&uma_rwlock);
1694 	LIST_REMOVE(zone, uz_link);
1695 	rw_wunlock(&uma_rwlock);
1696 	/*
1697 	 * XXX there are some races here where
1698 	 * the zone can be drained but zone lock
1699 	 * released and then refilled before we
1700 	 * remove it... we dont care for now
1701 	 */
1702 	zone_drain_wait(zone, M_WAITOK);
1703 	/*
1704 	 * Unlink all of our kegs.
1705 	 */
1706 	while ((klink = LIST_FIRST(&zone->uz_kegs)) != NULL) {
1707 		klink->kl_keg = NULL;
1708 		LIST_REMOVE(klink, kl_link);
1709 		if (klink == &zone->uz_klink)
1710 			continue;
1711 		free(klink, M_TEMP);
1712 	}
1713 	/*
1714 	 * We only destroy kegs from non secondary zones.
1715 	 */
1716 	if (keg != NULL && (zone->uz_flags & UMA_ZONE_SECONDARY) == 0)  {
1717 		rw_wlock(&uma_rwlock);
1718 		LIST_REMOVE(keg, uk_link);
1719 		rw_wunlock(&uma_rwlock);
1720 		zone_free_item(kegs, keg, NULL, SKIP_NONE);
1721 	}
1722 	ZONE_LOCK_FINI(zone);
1723 }
1724 
1725 /*
1726  * Traverses every zone in the system and calls a callback
1727  *
1728  * Arguments:
1729  *	zfunc  A pointer to a function which accepts a zone
1730  *		as an argument.
1731  *
1732  * Returns:
1733  *	Nothing
1734  */
1735 static void
1736 zone_foreach(void (*zfunc)(uma_zone_t))
1737 {
1738 	uma_keg_t keg;
1739 	uma_zone_t zone;
1740 
1741 	rw_rlock(&uma_rwlock);
1742 	LIST_FOREACH(keg, &uma_kegs, uk_link) {
1743 		LIST_FOREACH(zone, &keg->uk_zones, uz_link)
1744 			zfunc(zone);
1745 	}
1746 	rw_runlock(&uma_rwlock);
1747 }
1748 
1749 /* Public functions */
1750 /* See uma.h */
1751 void
1752 uma_startup(void *mem, int npages)
1753 {
1754 	struct uma_zctor_args args;
1755 
1756 	rw_init(&uma_rwlock, "UMA lock");
1757 
1758 	/* "manually" create the initial zone */
1759 	memset(&args, 0, sizeof(args));
1760 	args.name = "UMA Kegs";
1761 	args.size = sizeof(struct uma_keg);
1762 	args.ctor = keg_ctor;
1763 	args.dtor = keg_dtor;
1764 	args.uminit = zero_init;
1765 	args.fini = NULL;
1766 	args.keg = &masterkeg;
1767 	args.align = 32 - 1;
1768 	args.flags = UMA_ZFLAG_INTERNAL;
1769 	/* The initial zone has no Per cpu queues so it's smaller */
1770 	zone_ctor(kegs, sizeof(struct uma_zone), &args, M_WAITOK);
1771 
1772 	mtx_init(&uma_boot_pages_mtx, "UMA boot pages", NULL, MTX_DEF);
1773 	bootmem = mem;
1774 	boot_pages = npages;
1775 
1776 	args.name = "UMA Zones";
1777 	args.size = sizeof(struct uma_zone) +
1778 	    (sizeof(struct uma_cache) * (mp_maxid + 1));
1779 	args.ctor = zone_ctor;
1780 	args.dtor = zone_dtor;
1781 	args.uminit = zero_init;
1782 	args.fini = NULL;
1783 	args.keg = NULL;
1784 	args.align = 32 - 1;
1785 	args.flags = UMA_ZFLAG_INTERNAL;
1786 	/* The initial zone has no Per cpu queues so it's smaller */
1787 	zone_ctor(zones, sizeof(struct uma_zone), &args, M_WAITOK);
1788 
1789 	/* Now make a zone for slab headers */
1790 	slabzone = uma_zcreate("UMA Slabs",
1791 				sizeof(struct uma_slab),
1792 				NULL, NULL, NULL, NULL,
1793 				UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1794 
1795 	hashzone = uma_zcreate("UMA Hash",
1796 	    sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
1797 	    NULL, NULL, NULL, NULL,
1798 	    UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1799 
1800 	bucket_init();
1801 
1802 	booted = UMA_STARTUP;
1803 }
1804 
1805 /* see uma.h */
1806 void
1807 uma_startup2(void)
1808 {
1809 	booted = UMA_STARTUP2;
1810 	bucket_enable();
1811 	sx_init(&uma_drain_lock, "umadrain");
1812 }
1813 
1814 /*
1815  * Initialize our callout handle
1816  *
1817  */
1818 
1819 static void
1820 uma_startup3(void)
1821 {
1822 
1823 	callout_init(&uma_callout, 1);
1824 	callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
1825 }
1826 
1827 static uma_keg_t
1828 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
1829 		int align, uint32_t flags)
1830 {
1831 	struct uma_kctor_args args;
1832 
1833 	args.size = size;
1834 	args.uminit = uminit;
1835 	args.fini = fini;
1836 	args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
1837 	args.flags = flags;
1838 	args.zone = zone;
1839 	return (zone_alloc_item(kegs, &args, M_WAITOK));
1840 }
1841 
1842 /* See uma.h */
1843 void
1844 uma_set_align(int align)
1845 {
1846 
1847 	if (align != UMA_ALIGN_CACHE)
1848 		uma_align_cache = align;
1849 }
1850 
1851 /* See uma.h */
1852 uma_zone_t
1853 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
1854 		uma_init uminit, uma_fini fini, int align, uint32_t flags)
1855 
1856 {
1857 	struct uma_zctor_args args;
1858 	uma_zone_t res;
1859 	bool locked;
1860 
1861 	KASSERT(powerof2(align + 1), ("invalid zone alignment %d for \"%s\"",
1862 	    align, name));
1863 
1864 	/* This stuff is essential for the zone ctor */
1865 	memset(&args, 0, sizeof(args));
1866 	args.name = name;
1867 	args.size = size;
1868 	args.ctor = ctor;
1869 	args.dtor = dtor;
1870 	args.uminit = uminit;
1871 	args.fini = fini;
1872 #ifdef  INVARIANTS
1873 	/*
1874 	 * If a zone is being created with an empty constructor and
1875 	 * destructor, pass UMA constructor/destructor which checks for
1876 	 * memory use after free.
1877 	 */
1878 	if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOFREE))) &&
1879 	    ctor == NULL && dtor == NULL && uminit == NULL && fini == NULL) {
1880 		args.ctor = trash_ctor;
1881 		args.dtor = trash_dtor;
1882 		args.uminit = trash_init;
1883 		args.fini = trash_fini;
1884 	}
1885 #endif
1886 	args.align = align;
1887 	args.flags = flags;
1888 	args.keg = NULL;
1889 
1890 	if (booted < UMA_STARTUP2) {
1891 		locked = false;
1892 	} else {
1893 		sx_slock(&uma_drain_lock);
1894 		locked = true;
1895 	}
1896 	res = zone_alloc_item(zones, &args, M_WAITOK);
1897 	if (locked)
1898 		sx_sunlock(&uma_drain_lock);
1899 	return (res);
1900 }
1901 
1902 /* See uma.h */
1903 uma_zone_t
1904 uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor,
1905 		    uma_init zinit, uma_fini zfini, uma_zone_t master)
1906 {
1907 	struct uma_zctor_args args;
1908 	uma_keg_t keg;
1909 	uma_zone_t res;
1910 	bool locked;
1911 
1912 	keg = zone_first_keg(master);
1913 	memset(&args, 0, sizeof(args));
1914 	args.name = name;
1915 	args.size = keg->uk_size;
1916 	args.ctor = ctor;
1917 	args.dtor = dtor;
1918 	args.uminit = zinit;
1919 	args.fini = zfini;
1920 	args.align = keg->uk_align;
1921 	args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
1922 	args.keg = keg;
1923 
1924 	if (booted < UMA_STARTUP2) {
1925 		locked = false;
1926 	} else {
1927 		sx_slock(&uma_drain_lock);
1928 		locked = true;
1929 	}
1930 	/* XXX Attaches only one keg of potentially many. */
1931 	res = zone_alloc_item(zones, &args, M_WAITOK);
1932 	if (locked)
1933 		sx_sunlock(&uma_drain_lock);
1934 	return (res);
1935 }
1936 
1937 /* See uma.h */
1938 uma_zone_t
1939 uma_zcache_create(char *name, int size, uma_ctor ctor, uma_dtor dtor,
1940 		    uma_init zinit, uma_fini zfini, uma_import zimport,
1941 		    uma_release zrelease, void *arg, int flags)
1942 {
1943 	struct uma_zctor_args args;
1944 
1945 	memset(&args, 0, sizeof(args));
1946 	args.name = name;
1947 	args.size = size;
1948 	args.ctor = ctor;
1949 	args.dtor = dtor;
1950 	args.uminit = zinit;
1951 	args.fini = zfini;
1952 	args.import = zimport;
1953 	args.release = zrelease;
1954 	args.arg = arg;
1955 	args.align = 0;
1956 	args.flags = flags;
1957 
1958 	return (zone_alloc_item(zones, &args, M_WAITOK));
1959 }
1960 
1961 static void
1962 zone_lock_pair(uma_zone_t a, uma_zone_t b)
1963 {
1964 	if (a < b) {
1965 		ZONE_LOCK(a);
1966 		mtx_lock_flags(b->uz_lockptr, MTX_DUPOK);
1967 	} else {
1968 		ZONE_LOCK(b);
1969 		mtx_lock_flags(a->uz_lockptr, MTX_DUPOK);
1970 	}
1971 }
1972 
1973 static void
1974 zone_unlock_pair(uma_zone_t a, uma_zone_t b)
1975 {
1976 
1977 	ZONE_UNLOCK(a);
1978 	ZONE_UNLOCK(b);
1979 }
1980 
1981 int
1982 uma_zsecond_add(uma_zone_t zone, uma_zone_t master)
1983 {
1984 	uma_klink_t klink;
1985 	uma_klink_t kl;
1986 	int error;
1987 
1988 	error = 0;
1989 	klink = malloc(sizeof(*klink), M_TEMP, M_WAITOK | M_ZERO);
1990 
1991 	zone_lock_pair(zone, master);
1992 	/*
1993 	 * zone must use vtoslab() to resolve objects and must already be
1994 	 * a secondary.
1995 	 */
1996 	if ((zone->uz_flags & (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY))
1997 	    != (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) {
1998 		error = EINVAL;
1999 		goto out;
2000 	}
2001 	/*
2002 	 * The new master must also use vtoslab().
2003 	 */
2004 	if ((zone->uz_flags & UMA_ZONE_VTOSLAB) != UMA_ZONE_VTOSLAB) {
2005 		error = EINVAL;
2006 		goto out;
2007 	}
2008 
2009 	/*
2010 	 * The underlying object must be the same size.  rsize
2011 	 * may be different.
2012 	 */
2013 	if (master->uz_size != zone->uz_size) {
2014 		error = E2BIG;
2015 		goto out;
2016 	}
2017 	/*
2018 	 * Put it at the end of the list.
2019 	 */
2020 	klink->kl_keg = zone_first_keg(master);
2021 	LIST_FOREACH(kl, &zone->uz_kegs, kl_link) {
2022 		if (LIST_NEXT(kl, kl_link) == NULL) {
2023 			LIST_INSERT_AFTER(kl, klink, kl_link);
2024 			break;
2025 		}
2026 	}
2027 	klink = NULL;
2028 	zone->uz_flags |= UMA_ZFLAG_MULTI;
2029 	zone->uz_slab = zone_fetch_slab_multi;
2030 
2031 out:
2032 	zone_unlock_pair(zone, master);
2033 	if (klink != NULL)
2034 		free(klink, M_TEMP);
2035 
2036 	return (error);
2037 }
2038 
2039 
2040 /* See uma.h */
2041 void
2042 uma_zdestroy(uma_zone_t zone)
2043 {
2044 
2045 	sx_slock(&uma_drain_lock);
2046 	zone_free_item(zones, zone, NULL, SKIP_NONE);
2047 	sx_sunlock(&uma_drain_lock);
2048 }
2049 
2050 void
2051 uma_zwait(uma_zone_t zone)
2052 {
2053 	void *item;
2054 
2055 	item = uma_zalloc_arg(zone, NULL, M_WAITOK);
2056 	uma_zfree(zone, item);
2057 }
2058 
2059 /* See uma.h */
2060 void *
2061 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
2062 {
2063 	void *item;
2064 	uma_cache_t cache;
2065 	uma_bucket_t bucket;
2066 	int lockfail;
2067 	int cpu;
2068 
2069 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
2070 	random_harvest_fast_uma(&zone, sizeof(zone), 1, RANDOM_UMA);
2071 
2072 	/* This is the fast path allocation */
2073 	CTR4(KTR_UMA, "uma_zalloc_arg thread %x zone %s(%p) flags %d",
2074 	    curthread, zone->uz_name, zone, flags);
2075 
2076 	if (flags & M_WAITOK) {
2077 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
2078 		    "uma_zalloc_arg: zone \"%s\"", zone->uz_name);
2079 	}
2080 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
2081 	    ("uma_zalloc_arg: called with spinlock or critical section held"));
2082 
2083 #ifdef DEBUG_MEMGUARD
2084 	if (memguard_cmp_zone(zone)) {
2085 		item = memguard_alloc(zone->uz_size, flags);
2086 		if (item != NULL) {
2087 			if (zone->uz_init != NULL &&
2088 			    zone->uz_init(item, zone->uz_size, flags) != 0)
2089 				return (NULL);
2090 			if (zone->uz_ctor != NULL &&
2091 			    zone->uz_ctor(item, zone->uz_size, udata,
2092 			    flags) != 0) {
2093 			    	zone->uz_fini(item, zone->uz_size);
2094 				return (NULL);
2095 			}
2096 			return (item);
2097 		}
2098 		/* This is unfortunate but should not be fatal. */
2099 	}
2100 #endif
2101 	/*
2102 	 * If possible, allocate from the per-CPU cache.  There are two
2103 	 * requirements for safe access to the per-CPU cache: (1) the thread
2104 	 * accessing the cache must not be preempted or yield during access,
2105 	 * and (2) the thread must not migrate CPUs without switching which
2106 	 * cache it accesses.  We rely on a critical section to prevent
2107 	 * preemption and migration.  We release the critical section in
2108 	 * order to acquire the zone mutex if we are unable to allocate from
2109 	 * the current cache; when we re-acquire the critical section, we
2110 	 * must detect and handle migration if it has occurred.
2111 	 */
2112 	critical_enter();
2113 	cpu = curcpu;
2114 	cache = &zone->uz_cpu[cpu];
2115 
2116 zalloc_start:
2117 	bucket = cache->uc_allocbucket;
2118 	if (bucket != NULL && bucket->ub_cnt > 0) {
2119 		bucket->ub_cnt--;
2120 		item = bucket->ub_bucket[bucket->ub_cnt];
2121 #ifdef INVARIANTS
2122 		bucket->ub_bucket[bucket->ub_cnt] = NULL;
2123 #endif
2124 		KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled."));
2125 		cache->uc_allocs++;
2126 		critical_exit();
2127 		if (zone->uz_ctor != NULL &&
2128 		    zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2129 			atomic_add_long(&zone->uz_fails, 1);
2130 			zone_free_item(zone, item, udata, SKIP_DTOR);
2131 			return (NULL);
2132 		}
2133 #ifdef INVARIANTS
2134 		uma_dbg_alloc(zone, NULL, item);
2135 #endif
2136 		if (flags & M_ZERO)
2137 			uma_zero_item(item, zone);
2138 		return (item);
2139 	}
2140 
2141 	/*
2142 	 * We have run out of items in our alloc bucket.
2143 	 * See if we can switch with our free bucket.
2144 	 */
2145 	bucket = cache->uc_freebucket;
2146 	if (bucket != NULL && bucket->ub_cnt > 0) {
2147 		CTR2(KTR_UMA,
2148 		    "uma_zalloc: zone %s(%p) swapping empty with alloc",
2149 		    zone->uz_name, zone);
2150 		cache->uc_freebucket = cache->uc_allocbucket;
2151 		cache->uc_allocbucket = bucket;
2152 		goto zalloc_start;
2153 	}
2154 
2155 	/*
2156 	 * Discard any empty allocation bucket while we hold no locks.
2157 	 */
2158 	bucket = cache->uc_allocbucket;
2159 	cache->uc_allocbucket = NULL;
2160 	critical_exit();
2161 	if (bucket != NULL)
2162 		bucket_free(zone, bucket, udata);
2163 
2164 	/* Short-circuit for zones without buckets and low memory. */
2165 	if (zone->uz_count == 0 || bucketdisable)
2166 		goto zalloc_item;
2167 
2168 	/*
2169 	 * Attempt to retrieve the item from the per-CPU cache has failed, so
2170 	 * we must go back to the zone.  This requires the zone lock, so we
2171 	 * must drop the critical section, then re-acquire it when we go back
2172 	 * to the cache.  Since the critical section is released, we may be
2173 	 * preempted or migrate.  As such, make sure not to maintain any
2174 	 * thread-local state specific to the cache from prior to releasing
2175 	 * the critical section.
2176 	 */
2177 	lockfail = 0;
2178 	if (ZONE_TRYLOCK(zone) == 0) {
2179 		/* Record contention to size the buckets. */
2180 		ZONE_LOCK(zone);
2181 		lockfail = 1;
2182 	}
2183 	critical_enter();
2184 	cpu = curcpu;
2185 	cache = &zone->uz_cpu[cpu];
2186 
2187 	/*
2188 	 * Since we have locked the zone we may as well send back our stats.
2189 	 */
2190 	atomic_add_long(&zone->uz_allocs, cache->uc_allocs);
2191 	atomic_add_long(&zone->uz_frees, cache->uc_frees);
2192 	cache->uc_allocs = 0;
2193 	cache->uc_frees = 0;
2194 
2195 	/* See if we lost the race to fill the cache. */
2196 	if (cache->uc_allocbucket != NULL) {
2197 		ZONE_UNLOCK(zone);
2198 		goto zalloc_start;
2199 	}
2200 
2201 	/*
2202 	 * Check the zone's cache of buckets.
2203 	 */
2204 	if ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) {
2205 		KASSERT(bucket->ub_cnt != 0,
2206 		    ("uma_zalloc_arg: Returning an empty bucket."));
2207 
2208 		LIST_REMOVE(bucket, ub_link);
2209 		cache->uc_allocbucket = bucket;
2210 		ZONE_UNLOCK(zone);
2211 		goto zalloc_start;
2212 	}
2213 	/* We are no longer associated with this CPU. */
2214 	critical_exit();
2215 
2216 	/*
2217 	 * We bump the uz count when the cache size is insufficient to
2218 	 * handle the working set.
2219 	 */
2220 	if (lockfail && zone->uz_count < BUCKET_MAX)
2221 		zone->uz_count++;
2222 	ZONE_UNLOCK(zone);
2223 
2224 	/*
2225 	 * Now lets just fill a bucket and put it on the free list.  If that
2226 	 * works we'll restart the allocation from the beginning and it
2227 	 * will use the just filled bucket.
2228 	 */
2229 	bucket = zone_alloc_bucket(zone, udata, flags);
2230 	CTR3(KTR_UMA, "uma_zalloc: zone %s(%p) bucket zone returned %p",
2231 	    zone->uz_name, zone, bucket);
2232 	if (bucket != NULL) {
2233 		ZONE_LOCK(zone);
2234 		critical_enter();
2235 		cpu = curcpu;
2236 		cache = &zone->uz_cpu[cpu];
2237 		/*
2238 		 * See if we lost the race or were migrated.  Cache the
2239 		 * initialized bucket to make this less likely or claim
2240 		 * the memory directly.
2241 		 */
2242 		if (cache->uc_allocbucket == NULL)
2243 			cache->uc_allocbucket = bucket;
2244 		else
2245 			LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link);
2246 		ZONE_UNLOCK(zone);
2247 		goto zalloc_start;
2248 	}
2249 
2250 	/*
2251 	 * We may not be able to get a bucket so return an actual item.
2252 	 */
2253 zalloc_item:
2254 	item = zone_alloc_item(zone, udata, flags);
2255 
2256 	return (item);
2257 }
2258 
2259 static uma_slab_t
2260 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int flags)
2261 {
2262 	uma_slab_t slab;
2263 	int reserve;
2264 
2265 	mtx_assert(&keg->uk_lock, MA_OWNED);
2266 	slab = NULL;
2267 	reserve = 0;
2268 	if ((flags & M_USE_RESERVE) == 0)
2269 		reserve = keg->uk_reserve;
2270 
2271 	for (;;) {
2272 		/*
2273 		 * Find a slab with some space.  Prefer slabs that are partially
2274 		 * used over those that are totally full.  This helps to reduce
2275 		 * fragmentation.
2276 		 */
2277 		if (keg->uk_free > reserve) {
2278 			if (!LIST_EMPTY(&keg->uk_part_slab)) {
2279 				slab = LIST_FIRST(&keg->uk_part_slab);
2280 			} else {
2281 				slab = LIST_FIRST(&keg->uk_free_slab);
2282 				LIST_REMOVE(slab, us_link);
2283 				LIST_INSERT_HEAD(&keg->uk_part_slab, slab,
2284 				    us_link);
2285 			}
2286 			MPASS(slab->us_keg == keg);
2287 			return (slab);
2288 		}
2289 
2290 		/*
2291 		 * M_NOVM means don't ask at all!
2292 		 */
2293 		if (flags & M_NOVM)
2294 			break;
2295 
2296 		if (keg->uk_maxpages && keg->uk_pages >= keg->uk_maxpages) {
2297 			keg->uk_flags |= UMA_ZFLAG_FULL;
2298 			/*
2299 			 * If this is not a multi-zone, set the FULL bit.
2300 			 * Otherwise slab_multi() takes care of it.
2301 			 */
2302 			if ((zone->uz_flags & UMA_ZFLAG_MULTI) == 0) {
2303 				zone->uz_flags |= UMA_ZFLAG_FULL;
2304 				zone_log_warning(zone);
2305 				zone_maxaction(zone);
2306 			}
2307 			if (flags & M_NOWAIT)
2308 				break;
2309 			zone->uz_sleeps++;
2310 			msleep(keg, &keg->uk_lock, PVM, "keglimit", 0);
2311 			continue;
2312 		}
2313 		slab = keg_alloc_slab(keg, zone, flags);
2314 		/*
2315 		 * If we got a slab here it's safe to mark it partially used
2316 		 * and return.  We assume that the caller is going to remove
2317 		 * at least one item.
2318 		 */
2319 		if (slab) {
2320 			MPASS(slab->us_keg == keg);
2321 			LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2322 			return (slab);
2323 		}
2324 		/*
2325 		 * We might not have been able to get a slab but another cpu
2326 		 * could have while we were unlocked.  Check again before we
2327 		 * fail.
2328 		 */
2329 		flags |= M_NOVM;
2330 	}
2331 	return (slab);
2332 }
2333 
2334 static uma_slab_t
2335 zone_fetch_slab(uma_zone_t zone, uma_keg_t keg, int flags)
2336 {
2337 	uma_slab_t slab;
2338 
2339 	if (keg == NULL) {
2340 		keg = zone_first_keg(zone);
2341 		KEG_LOCK(keg);
2342 	}
2343 
2344 	for (;;) {
2345 		slab = keg_fetch_slab(keg, zone, flags);
2346 		if (slab)
2347 			return (slab);
2348 		if (flags & (M_NOWAIT | M_NOVM))
2349 			break;
2350 	}
2351 	KEG_UNLOCK(keg);
2352 	return (NULL);
2353 }
2354 
2355 /*
2356  * uma_zone_fetch_slab_multi:  Fetches a slab from one available keg.  Returns
2357  * with the keg locked.  On NULL no lock is held.
2358  *
2359  * The last pointer is used to seed the search.  It is not required.
2360  */
2361 static uma_slab_t
2362 zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int rflags)
2363 {
2364 	uma_klink_t klink;
2365 	uma_slab_t slab;
2366 	uma_keg_t keg;
2367 	int flags;
2368 	int empty;
2369 	int full;
2370 
2371 	/*
2372 	 * Don't wait on the first pass.  This will skip limit tests
2373 	 * as well.  We don't want to block if we can find a provider
2374 	 * without blocking.
2375 	 */
2376 	flags = (rflags & ~M_WAITOK) | M_NOWAIT;
2377 	/*
2378 	 * Use the last slab allocated as a hint for where to start
2379 	 * the search.
2380 	 */
2381 	if (last != NULL) {
2382 		slab = keg_fetch_slab(last, zone, flags);
2383 		if (slab)
2384 			return (slab);
2385 		KEG_UNLOCK(last);
2386 	}
2387 	/*
2388 	 * Loop until we have a slab incase of transient failures
2389 	 * while M_WAITOK is specified.  I'm not sure this is 100%
2390 	 * required but we've done it for so long now.
2391 	 */
2392 	for (;;) {
2393 		empty = 0;
2394 		full = 0;
2395 		/*
2396 		 * Search the available kegs for slabs.  Be careful to hold the
2397 		 * correct lock while calling into the keg layer.
2398 		 */
2399 		LIST_FOREACH(klink, &zone->uz_kegs, kl_link) {
2400 			keg = klink->kl_keg;
2401 			KEG_LOCK(keg);
2402 			if ((keg->uk_flags & UMA_ZFLAG_FULL) == 0) {
2403 				slab = keg_fetch_slab(keg, zone, flags);
2404 				if (slab)
2405 					return (slab);
2406 			}
2407 			if (keg->uk_flags & UMA_ZFLAG_FULL)
2408 				full++;
2409 			else
2410 				empty++;
2411 			KEG_UNLOCK(keg);
2412 		}
2413 		if (rflags & (M_NOWAIT | M_NOVM))
2414 			break;
2415 		flags = rflags;
2416 		/*
2417 		 * All kegs are full.  XXX We can't atomically check all kegs
2418 		 * and sleep so just sleep for a short period and retry.
2419 		 */
2420 		if (full && !empty) {
2421 			ZONE_LOCK(zone);
2422 			zone->uz_flags |= UMA_ZFLAG_FULL;
2423 			zone->uz_sleeps++;
2424 			zone_log_warning(zone);
2425 			zone_maxaction(zone);
2426 			msleep(zone, zone->uz_lockptr, PVM,
2427 			    "zonelimit", hz/100);
2428 			zone->uz_flags &= ~UMA_ZFLAG_FULL;
2429 			ZONE_UNLOCK(zone);
2430 			continue;
2431 		}
2432 	}
2433 	return (NULL);
2434 }
2435 
2436 static void *
2437 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
2438 {
2439 	void *item;
2440 	uint8_t freei;
2441 
2442 	MPASS(keg == slab->us_keg);
2443 	mtx_assert(&keg->uk_lock, MA_OWNED);
2444 
2445 	freei = BIT_FFS(SLAB_SETSIZE, &slab->us_free) - 1;
2446 	BIT_CLR(SLAB_SETSIZE, freei, &slab->us_free);
2447 	item = slab->us_data + (keg->uk_rsize * freei);
2448 	slab->us_freecount--;
2449 	keg->uk_free--;
2450 
2451 	/* Move this slab to the full list */
2452 	if (slab->us_freecount == 0) {
2453 		LIST_REMOVE(slab, us_link);
2454 		LIST_INSERT_HEAD(&keg->uk_full_slab, slab, us_link);
2455 	}
2456 
2457 	return (item);
2458 }
2459 
2460 static int
2461 zone_import(uma_zone_t zone, void **bucket, int max, int flags)
2462 {
2463 	uma_slab_t slab;
2464 	uma_keg_t keg;
2465 	int i;
2466 
2467 	slab = NULL;
2468 	keg = NULL;
2469 	/* Try to keep the buckets totally full */
2470 	for (i = 0; i < max; ) {
2471 		if ((slab = zone->uz_slab(zone, keg, flags)) == NULL)
2472 			break;
2473 		keg = slab->us_keg;
2474 		while (slab->us_freecount && i < max) {
2475 			bucket[i++] = slab_alloc_item(keg, slab);
2476 			if (keg->uk_free <= keg->uk_reserve)
2477 				break;
2478 		}
2479 		/* Don't grab more than one slab at a time. */
2480 		flags &= ~M_WAITOK;
2481 		flags |= M_NOWAIT;
2482 	}
2483 	if (slab != NULL)
2484 		KEG_UNLOCK(keg);
2485 
2486 	return i;
2487 }
2488 
2489 static uma_bucket_t
2490 zone_alloc_bucket(uma_zone_t zone, void *udata, int flags)
2491 {
2492 	uma_bucket_t bucket;
2493 	int max;
2494 
2495 	/* Don't wait for buckets, preserve caller's NOVM setting. */
2496 	bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
2497 	if (bucket == NULL)
2498 		return (NULL);
2499 
2500 	max = MIN(bucket->ub_entries, zone->uz_count);
2501 	bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
2502 	    max, flags);
2503 
2504 	/*
2505 	 * Initialize the memory if necessary.
2506 	 */
2507 	if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
2508 		int i;
2509 
2510 		for (i = 0; i < bucket->ub_cnt; i++)
2511 			if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size,
2512 			    flags) != 0)
2513 				break;
2514 		/*
2515 		 * If we couldn't initialize the whole bucket, put the
2516 		 * rest back onto the freelist.
2517 		 */
2518 		if (i != bucket->ub_cnt) {
2519 			zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
2520 			    bucket->ub_cnt - i);
2521 #ifdef INVARIANTS
2522 			bzero(&bucket->ub_bucket[i],
2523 			    sizeof(void *) * (bucket->ub_cnt - i));
2524 #endif
2525 			bucket->ub_cnt = i;
2526 		}
2527 	}
2528 
2529 	if (bucket->ub_cnt == 0) {
2530 		bucket_free(zone, bucket, udata);
2531 		atomic_add_long(&zone->uz_fails, 1);
2532 		return (NULL);
2533 	}
2534 
2535 	return (bucket);
2536 }
2537 
2538 /*
2539  * Allocates a single item from a zone.
2540  *
2541  * Arguments
2542  *	zone   The zone to alloc for.
2543  *	udata  The data to be passed to the constructor.
2544  *	flags  M_WAITOK, M_NOWAIT, M_ZERO.
2545  *
2546  * Returns
2547  *	NULL if there is no memory and M_NOWAIT is set
2548  *	An item if successful
2549  */
2550 
2551 static void *
2552 zone_alloc_item(uma_zone_t zone, void *udata, int flags)
2553 {
2554 	void *item;
2555 
2556 	item = NULL;
2557 
2558 	if (zone->uz_import(zone->uz_arg, &item, 1, flags) != 1)
2559 		goto fail;
2560 	atomic_add_long(&zone->uz_allocs, 1);
2561 
2562 	/*
2563 	 * We have to call both the zone's init (not the keg's init)
2564 	 * and the zone's ctor.  This is because the item is going from
2565 	 * a keg slab directly to the user, and the user is expecting it
2566 	 * to be both zone-init'd as well as zone-ctor'd.
2567 	 */
2568 	if (zone->uz_init != NULL) {
2569 		if (zone->uz_init(item, zone->uz_size, flags) != 0) {
2570 			zone_free_item(zone, item, udata, SKIP_FINI);
2571 			goto fail;
2572 		}
2573 	}
2574 	if (zone->uz_ctor != NULL) {
2575 		if (zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2576 			zone_free_item(zone, item, udata, SKIP_DTOR);
2577 			goto fail;
2578 		}
2579 	}
2580 #ifdef INVARIANTS
2581 	uma_dbg_alloc(zone, NULL, item);
2582 #endif
2583 	if (flags & M_ZERO)
2584 		uma_zero_item(item, zone);
2585 
2586 	CTR3(KTR_UMA, "zone_alloc_item item %p from %s(%p)", item,
2587 	    zone->uz_name, zone);
2588 
2589 	return (item);
2590 
2591 fail:
2592 	CTR2(KTR_UMA, "zone_alloc_item failed from %s(%p)",
2593 	    zone->uz_name, zone);
2594 	atomic_add_long(&zone->uz_fails, 1);
2595 	return (NULL);
2596 }
2597 
2598 /* See uma.h */
2599 void
2600 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
2601 {
2602 	uma_cache_t cache;
2603 	uma_bucket_t bucket;
2604 	int lockfail;
2605 	int cpu;
2606 
2607 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
2608 	random_harvest_fast_uma(&zone, sizeof(zone), 1, RANDOM_UMA);
2609 
2610 	CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread,
2611 	    zone->uz_name);
2612 
2613 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
2614 	    ("uma_zfree_arg: called with spinlock or critical section held"));
2615 
2616         /* uma_zfree(..., NULL) does nothing, to match free(9). */
2617         if (item == NULL)
2618                 return;
2619 #ifdef DEBUG_MEMGUARD
2620 	if (is_memguard_addr(item)) {
2621 		if (zone->uz_dtor != NULL)
2622 			zone->uz_dtor(item, zone->uz_size, udata);
2623 		if (zone->uz_fini != NULL)
2624 			zone->uz_fini(item, zone->uz_size);
2625 		memguard_free(item);
2626 		return;
2627 	}
2628 #endif
2629 #ifdef INVARIANTS
2630 	if (zone->uz_flags & UMA_ZONE_MALLOC)
2631 		uma_dbg_free(zone, udata, item);
2632 	else
2633 		uma_dbg_free(zone, NULL, item);
2634 #endif
2635 	if (zone->uz_dtor != NULL)
2636 		zone->uz_dtor(item, zone->uz_size, udata);
2637 
2638 	/*
2639 	 * The race here is acceptable.  If we miss it we'll just have to wait
2640 	 * a little longer for the limits to be reset.
2641 	 */
2642 	if (zone->uz_flags & UMA_ZFLAG_FULL)
2643 		goto zfree_item;
2644 
2645 	/*
2646 	 * If possible, free to the per-CPU cache.  There are two
2647 	 * requirements for safe access to the per-CPU cache: (1) the thread
2648 	 * accessing the cache must not be preempted or yield during access,
2649 	 * and (2) the thread must not migrate CPUs without switching which
2650 	 * cache it accesses.  We rely on a critical section to prevent
2651 	 * preemption and migration.  We release the critical section in
2652 	 * order to acquire the zone mutex if we are unable to free to the
2653 	 * current cache; when we re-acquire the critical section, we must
2654 	 * detect and handle migration if it has occurred.
2655 	 */
2656 zfree_restart:
2657 	critical_enter();
2658 	cpu = curcpu;
2659 	cache = &zone->uz_cpu[cpu];
2660 
2661 zfree_start:
2662 	/*
2663 	 * Try to free into the allocbucket first to give LIFO ordering
2664 	 * for cache-hot datastructures.  Spill over into the freebucket
2665 	 * if necessary.  Alloc will swap them if one runs dry.
2666 	 */
2667 	bucket = cache->uc_allocbucket;
2668 	if (bucket == NULL || bucket->ub_cnt >= bucket->ub_entries)
2669 		bucket = cache->uc_freebucket;
2670 	if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) {
2671 		KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL,
2672 		    ("uma_zfree: Freeing to non free bucket index."));
2673 		bucket->ub_bucket[bucket->ub_cnt] = item;
2674 		bucket->ub_cnt++;
2675 		cache->uc_frees++;
2676 		critical_exit();
2677 		return;
2678 	}
2679 
2680 	/*
2681 	 * We must go back the zone, which requires acquiring the zone lock,
2682 	 * which in turn means we must release and re-acquire the critical
2683 	 * section.  Since the critical section is released, we may be
2684 	 * preempted or migrate.  As such, make sure not to maintain any
2685 	 * thread-local state specific to the cache from prior to releasing
2686 	 * the critical section.
2687 	 */
2688 	critical_exit();
2689 	if (zone->uz_count == 0 || bucketdisable)
2690 		goto zfree_item;
2691 
2692 	lockfail = 0;
2693 	if (ZONE_TRYLOCK(zone) == 0) {
2694 		/* Record contention to size the buckets. */
2695 		ZONE_LOCK(zone);
2696 		lockfail = 1;
2697 	}
2698 	critical_enter();
2699 	cpu = curcpu;
2700 	cache = &zone->uz_cpu[cpu];
2701 
2702 	/*
2703 	 * Since we have locked the zone we may as well send back our stats.
2704 	 */
2705 	atomic_add_long(&zone->uz_allocs, cache->uc_allocs);
2706 	atomic_add_long(&zone->uz_frees, cache->uc_frees);
2707 	cache->uc_allocs = 0;
2708 	cache->uc_frees = 0;
2709 
2710 	bucket = cache->uc_freebucket;
2711 	if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) {
2712 		ZONE_UNLOCK(zone);
2713 		goto zfree_start;
2714 	}
2715 	cache->uc_freebucket = NULL;
2716 	/* We are no longer associated with this CPU. */
2717 	critical_exit();
2718 
2719 	/* Can we throw this on the zone full list? */
2720 	if (bucket != NULL) {
2721 		CTR3(KTR_UMA,
2722 		    "uma_zfree: zone %s(%p) putting bucket %p on free list",
2723 		    zone->uz_name, zone, bucket);
2724 		/* ub_cnt is pointing to the last free item */
2725 		KASSERT(bucket->ub_cnt != 0,
2726 		    ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n"));
2727 		LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link);
2728 	}
2729 
2730 	/*
2731 	 * We bump the uz count when the cache size is insufficient to
2732 	 * handle the working set.
2733 	 */
2734 	if (lockfail && zone->uz_count < BUCKET_MAX)
2735 		zone->uz_count++;
2736 	ZONE_UNLOCK(zone);
2737 
2738 	bucket = bucket_alloc(zone, udata, M_NOWAIT);
2739 	CTR3(KTR_UMA, "uma_zfree: zone %s(%p) allocated bucket %p",
2740 	    zone->uz_name, zone, bucket);
2741 	if (bucket) {
2742 		critical_enter();
2743 		cpu = curcpu;
2744 		cache = &zone->uz_cpu[cpu];
2745 		if (cache->uc_freebucket == NULL) {
2746 			cache->uc_freebucket = bucket;
2747 			goto zfree_start;
2748 		}
2749 		/*
2750 		 * We lost the race, start over.  We have to drop our
2751 		 * critical section to free the bucket.
2752 		 */
2753 		critical_exit();
2754 		bucket_free(zone, bucket, udata);
2755 		goto zfree_restart;
2756 	}
2757 
2758 	/*
2759 	 * If nothing else caught this, we'll just do an internal free.
2760 	 */
2761 zfree_item:
2762 	zone_free_item(zone, item, udata, SKIP_DTOR);
2763 
2764 	return;
2765 }
2766 
2767 static void
2768 slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item)
2769 {
2770 	uint8_t freei;
2771 
2772 	mtx_assert(&keg->uk_lock, MA_OWNED);
2773 	MPASS(keg == slab->us_keg);
2774 
2775 	/* Do we need to remove from any lists? */
2776 	if (slab->us_freecount+1 == keg->uk_ipers) {
2777 		LIST_REMOVE(slab, us_link);
2778 		LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
2779 	} else if (slab->us_freecount == 0) {
2780 		LIST_REMOVE(slab, us_link);
2781 		LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2782 	}
2783 
2784 	/* Slab management. */
2785 	freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
2786 	BIT_SET(SLAB_SETSIZE, freei, &slab->us_free);
2787 	slab->us_freecount++;
2788 
2789 	/* Keg statistics. */
2790 	keg->uk_free++;
2791 }
2792 
2793 static void
2794 zone_release(uma_zone_t zone, void **bucket, int cnt)
2795 {
2796 	void *item;
2797 	uma_slab_t slab;
2798 	uma_keg_t keg;
2799 	uint8_t *mem;
2800 	int clearfull;
2801 	int i;
2802 
2803 	clearfull = 0;
2804 	keg = zone_first_keg(zone);
2805 	KEG_LOCK(keg);
2806 	for (i = 0; i < cnt; i++) {
2807 		item = bucket[i];
2808 		if (!(zone->uz_flags & UMA_ZONE_VTOSLAB)) {
2809 			mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
2810 			if (zone->uz_flags & UMA_ZONE_HASH) {
2811 				slab = hash_sfind(&keg->uk_hash, mem);
2812 			} else {
2813 				mem += keg->uk_pgoff;
2814 				slab = (uma_slab_t)mem;
2815 			}
2816 		} else {
2817 			slab = vtoslab((vm_offset_t)item);
2818 			if (slab->us_keg != keg) {
2819 				KEG_UNLOCK(keg);
2820 				keg = slab->us_keg;
2821 				KEG_LOCK(keg);
2822 			}
2823 		}
2824 		slab_free_item(keg, slab, item);
2825 		if (keg->uk_flags & UMA_ZFLAG_FULL) {
2826 			if (keg->uk_pages < keg->uk_maxpages) {
2827 				keg->uk_flags &= ~UMA_ZFLAG_FULL;
2828 				clearfull = 1;
2829 			}
2830 
2831 			/*
2832 			 * We can handle one more allocation. Since we're
2833 			 * clearing ZFLAG_FULL, wake up all procs blocked
2834 			 * on pages. This should be uncommon, so keeping this
2835 			 * simple for now (rather than adding count of blocked
2836 			 * threads etc).
2837 			 */
2838 			wakeup(keg);
2839 		}
2840 	}
2841 	KEG_UNLOCK(keg);
2842 	if (clearfull) {
2843 		ZONE_LOCK(zone);
2844 		zone->uz_flags &= ~UMA_ZFLAG_FULL;
2845 		wakeup(zone);
2846 		ZONE_UNLOCK(zone);
2847 	}
2848 
2849 }
2850 
2851 /*
2852  * Frees a single item to any zone.
2853  *
2854  * Arguments:
2855  *	zone   The zone to free to
2856  *	item   The item we're freeing
2857  *	udata  User supplied data for the dtor
2858  *	skip   Skip dtors and finis
2859  */
2860 static void
2861 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
2862 {
2863 
2864 #ifdef INVARIANTS
2865 	if (skip == SKIP_NONE) {
2866 		if (zone->uz_flags & UMA_ZONE_MALLOC)
2867 			uma_dbg_free(zone, udata, item);
2868 		else
2869 			uma_dbg_free(zone, NULL, item);
2870 	}
2871 #endif
2872 	if (skip < SKIP_DTOR && zone->uz_dtor)
2873 		zone->uz_dtor(item, zone->uz_size, udata);
2874 
2875 	if (skip < SKIP_FINI && zone->uz_fini)
2876 		zone->uz_fini(item, zone->uz_size);
2877 
2878 	atomic_add_long(&zone->uz_frees, 1);
2879 	zone->uz_release(zone->uz_arg, &item, 1);
2880 }
2881 
2882 /* See uma.h */
2883 int
2884 uma_zone_set_max(uma_zone_t zone, int nitems)
2885 {
2886 	uma_keg_t keg;
2887 
2888 	keg = zone_first_keg(zone);
2889 	if (keg == NULL)
2890 		return (0);
2891 	KEG_LOCK(keg);
2892 	keg->uk_maxpages = (nitems / keg->uk_ipers) * keg->uk_ppera;
2893 	if (keg->uk_maxpages * keg->uk_ipers < nitems)
2894 		keg->uk_maxpages += keg->uk_ppera;
2895 	nitems = (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers;
2896 	KEG_UNLOCK(keg);
2897 
2898 	return (nitems);
2899 }
2900 
2901 /* See uma.h */
2902 int
2903 uma_zone_get_max(uma_zone_t zone)
2904 {
2905 	int nitems;
2906 	uma_keg_t keg;
2907 
2908 	keg = zone_first_keg(zone);
2909 	if (keg == NULL)
2910 		return (0);
2911 	KEG_LOCK(keg);
2912 	nitems = (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers;
2913 	KEG_UNLOCK(keg);
2914 
2915 	return (nitems);
2916 }
2917 
2918 /* See uma.h */
2919 void
2920 uma_zone_set_warning(uma_zone_t zone, const char *warning)
2921 {
2922 
2923 	ZONE_LOCK(zone);
2924 	zone->uz_warning = warning;
2925 	ZONE_UNLOCK(zone);
2926 }
2927 
2928 /* See uma.h */
2929 void
2930 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction)
2931 {
2932 
2933 	ZONE_LOCK(zone);
2934 	TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone);
2935 	ZONE_UNLOCK(zone);
2936 }
2937 
2938 /* See uma.h */
2939 int
2940 uma_zone_get_cur(uma_zone_t zone)
2941 {
2942 	int64_t nitems;
2943 	u_int i;
2944 
2945 	ZONE_LOCK(zone);
2946 	nitems = zone->uz_allocs - zone->uz_frees;
2947 	CPU_FOREACH(i) {
2948 		/*
2949 		 * See the comment in sysctl_vm_zone_stats() regarding the
2950 		 * safety of accessing the per-cpu caches. With the zone lock
2951 		 * held, it is safe, but can potentially result in stale data.
2952 		 */
2953 		nitems += zone->uz_cpu[i].uc_allocs -
2954 		    zone->uz_cpu[i].uc_frees;
2955 	}
2956 	ZONE_UNLOCK(zone);
2957 
2958 	return (nitems < 0 ? 0 : nitems);
2959 }
2960 
2961 /* See uma.h */
2962 void
2963 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
2964 {
2965 	uma_keg_t keg;
2966 
2967 	keg = zone_first_keg(zone);
2968 	KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type"));
2969 	KEG_LOCK(keg);
2970 	KASSERT(keg->uk_pages == 0,
2971 	    ("uma_zone_set_init on non-empty keg"));
2972 	keg->uk_init = uminit;
2973 	KEG_UNLOCK(keg);
2974 }
2975 
2976 /* See uma.h */
2977 void
2978 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
2979 {
2980 	uma_keg_t keg;
2981 
2982 	keg = zone_first_keg(zone);
2983 	KASSERT(keg != NULL, ("uma_zone_set_fini: Invalid zone type"));
2984 	KEG_LOCK(keg);
2985 	KASSERT(keg->uk_pages == 0,
2986 	    ("uma_zone_set_fini on non-empty keg"));
2987 	keg->uk_fini = fini;
2988 	KEG_UNLOCK(keg);
2989 }
2990 
2991 /* See uma.h */
2992 void
2993 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
2994 {
2995 
2996 	ZONE_LOCK(zone);
2997 	KASSERT(zone_first_keg(zone)->uk_pages == 0,
2998 	    ("uma_zone_set_zinit on non-empty keg"));
2999 	zone->uz_init = zinit;
3000 	ZONE_UNLOCK(zone);
3001 }
3002 
3003 /* See uma.h */
3004 void
3005 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
3006 {
3007 
3008 	ZONE_LOCK(zone);
3009 	KASSERT(zone_first_keg(zone)->uk_pages == 0,
3010 	    ("uma_zone_set_zfini on non-empty keg"));
3011 	zone->uz_fini = zfini;
3012 	ZONE_UNLOCK(zone);
3013 }
3014 
3015 /* See uma.h */
3016 /* XXX uk_freef is not actually used with the zone locked */
3017 void
3018 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
3019 {
3020 	uma_keg_t keg;
3021 
3022 	keg = zone_first_keg(zone);
3023 	KASSERT(keg != NULL, ("uma_zone_set_freef: Invalid zone type"));
3024 	KEG_LOCK(keg);
3025 	keg->uk_freef = freef;
3026 	KEG_UNLOCK(keg);
3027 }
3028 
3029 /* See uma.h */
3030 /* XXX uk_allocf is not actually used with the zone locked */
3031 void
3032 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
3033 {
3034 	uma_keg_t keg;
3035 
3036 	keg = zone_first_keg(zone);
3037 	KEG_LOCK(keg);
3038 	keg->uk_allocf = allocf;
3039 	KEG_UNLOCK(keg);
3040 }
3041 
3042 /* See uma.h */
3043 void
3044 uma_zone_reserve(uma_zone_t zone, int items)
3045 {
3046 	uma_keg_t keg;
3047 
3048 	keg = zone_first_keg(zone);
3049 	if (keg == NULL)
3050 		return;
3051 	KEG_LOCK(keg);
3052 	keg->uk_reserve = items;
3053 	KEG_UNLOCK(keg);
3054 
3055 	return;
3056 }
3057 
3058 /* See uma.h */
3059 int
3060 uma_zone_reserve_kva(uma_zone_t zone, int count)
3061 {
3062 	uma_keg_t keg;
3063 	vm_offset_t kva;
3064 	u_int pages;
3065 
3066 	keg = zone_first_keg(zone);
3067 	if (keg == NULL)
3068 		return (0);
3069 	pages = count / keg->uk_ipers;
3070 
3071 	if (pages * keg->uk_ipers < count)
3072 		pages++;
3073 	pages *= keg->uk_ppera;
3074 
3075 #ifdef UMA_MD_SMALL_ALLOC
3076 	if (keg->uk_ppera > 1) {
3077 #else
3078 	if (1) {
3079 #endif
3080 		kva = kva_alloc((vm_size_t)pages * PAGE_SIZE);
3081 		if (kva == 0)
3082 			return (0);
3083 	} else
3084 		kva = 0;
3085 	KEG_LOCK(keg);
3086 	keg->uk_kva = kva;
3087 	keg->uk_offset = 0;
3088 	keg->uk_maxpages = pages;
3089 #ifdef UMA_MD_SMALL_ALLOC
3090 	keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
3091 #else
3092 	keg->uk_allocf = noobj_alloc;
3093 #endif
3094 	keg->uk_flags |= UMA_ZONE_NOFREE;
3095 	KEG_UNLOCK(keg);
3096 
3097 	return (1);
3098 }
3099 
3100 /* See uma.h */
3101 void
3102 uma_prealloc(uma_zone_t zone, int items)
3103 {
3104 	int slabs;
3105 	uma_slab_t slab;
3106 	uma_keg_t keg;
3107 
3108 	keg = zone_first_keg(zone);
3109 	if (keg == NULL)
3110 		return;
3111 	KEG_LOCK(keg);
3112 	slabs = items / keg->uk_ipers;
3113 	if (slabs * keg->uk_ipers < items)
3114 		slabs++;
3115 	while (slabs > 0) {
3116 		slab = keg_alloc_slab(keg, zone, M_WAITOK);
3117 		if (slab == NULL)
3118 			break;
3119 		MPASS(slab->us_keg == keg);
3120 		LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
3121 		slabs--;
3122 	}
3123 	KEG_UNLOCK(keg);
3124 }
3125 
3126 /* See uma.h */
3127 static void
3128 uma_reclaim_locked(bool kmem_danger)
3129 {
3130 
3131 	CTR0(KTR_UMA, "UMA: vm asked us to release pages!");
3132 	sx_assert(&uma_drain_lock, SA_XLOCKED);
3133 	bucket_enable();
3134 	zone_foreach(zone_drain);
3135 	if (vm_page_count_min() || kmem_danger) {
3136 		cache_drain_safe(NULL);
3137 		zone_foreach(zone_drain);
3138 	}
3139 	/*
3140 	 * Some slabs may have been freed but this zone will be visited early
3141 	 * we visit again so that we can free pages that are empty once other
3142 	 * zones are drained.  We have to do the same for buckets.
3143 	 */
3144 	zone_drain(slabzone);
3145 	bucket_zone_drain();
3146 }
3147 
3148 void
3149 uma_reclaim(void)
3150 {
3151 
3152 	sx_xlock(&uma_drain_lock);
3153 	uma_reclaim_locked(false);
3154 	sx_xunlock(&uma_drain_lock);
3155 }
3156 
3157 static volatile int uma_reclaim_needed;
3158 
3159 void
3160 uma_reclaim_wakeup(void)
3161 {
3162 
3163 	if (atomic_fetchadd_int(&uma_reclaim_needed, 1) == 0)
3164 		wakeup(uma_reclaim);
3165 }
3166 
3167 void
3168 uma_reclaim_worker(void *arg __unused)
3169 {
3170 
3171 	for (;;) {
3172 		sx_xlock(&uma_drain_lock);
3173 		while (atomic_load_int(&uma_reclaim_needed) == 0)
3174 			sx_sleep(uma_reclaim, &uma_drain_lock, PVM, "umarcl",
3175 			    hz);
3176 		sx_xunlock(&uma_drain_lock);
3177 		EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM);
3178 		sx_xlock(&uma_drain_lock);
3179 		uma_reclaim_locked(true);
3180 		atomic_store_int(&uma_reclaim_needed, 0);
3181 		sx_xunlock(&uma_drain_lock);
3182 		/* Don't fire more than once per-second. */
3183 		pause("umarclslp", hz);
3184 	}
3185 }
3186 
3187 /* See uma.h */
3188 int
3189 uma_zone_exhausted(uma_zone_t zone)
3190 {
3191 	int full;
3192 
3193 	ZONE_LOCK(zone);
3194 	full = (zone->uz_flags & UMA_ZFLAG_FULL);
3195 	ZONE_UNLOCK(zone);
3196 	return (full);
3197 }
3198 
3199 int
3200 uma_zone_exhausted_nolock(uma_zone_t zone)
3201 {
3202 	return (zone->uz_flags & UMA_ZFLAG_FULL);
3203 }
3204 
3205 void *
3206 uma_large_malloc(vm_size_t size, int wait)
3207 {
3208 	void *mem;
3209 	uma_slab_t slab;
3210 	uint8_t flags;
3211 
3212 	slab = zone_alloc_item(slabzone, NULL, wait);
3213 	if (slab == NULL)
3214 		return (NULL);
3215 	mem = page_alloc(NULL, size, &flags, wait);
3216 	if (mem) {
3217 		vsetslab((vm_offset_t)mem, slab);
3218 		slab->us_data = mem;
3219 		slab->us_flags = flags | UMA_SLAB_MALLOC;
3220 		slab->us_size = size;
3221 		uma_total_inc(size);
3222 	} else {
3223 		zone_free_item(slabzone, slab, NULL, SKIP_NONE);
3224 	}
3225 
3226 	return (mem);
3227 }
3228 
3229 void
3230 uma_large_free(uma_slab_t slab)
3231 {
3232 
3233 	page_free(slab->us_data, slab->us_size, slab->us_flags);
3234 	uma_total_dec(slab->us_size);
3235 	zone_free_item(slabzone, slab, NULL, SKIP_NONE);
3236 }
3237 
3238 static void
3239 uma_zero_item(void *item, uma_zone_t zone)
3240 {
3241 	int i;
3242 
3243 	if (zone->uz_flags & UMA_ZONE_PCPU) {
3244 		CPU_FOREACH(i)
3245 			bzero(zpcpu_get_cpu(item, i), zone->uz_size);
3246 	} else
3247 		bzero(item, zone->uz_size);
3248 }
3249 
3250 unsigned long
3251 uma_limit(void)
3252 {
3253 
3254 	return (uma_kmem_limit);
3255 }
3256 
3257 void
3258 uma_set_limit(unsigned long limit)
3259 {
3260 
3261 	uma_kmem_limit = limit;
3262 }
3263 
3264 unsigned long
3265 uma_size(void)
3266 {
3267 
3268 	return uma_kmem_total;
3269 }
3270 
3271 void
3272 uma_print_stats(void)
3273 {
3274 	zone_foreach(uma_print_zone);
3275 }
3276 
3277 static void
3278 slab_print(uma_slab_t slab)
3279 {
3280 	printf("slab: keg %p, data %p, freecount %d\n",
3281 		slab->us_keg, slab->us_data, slab->us_freecount);
3282 }
3283 
3284 static void
3285 cache_print(uma_cache_t cache)
3286 {
3287 	printf("alloc: %p(%d), free: %p(%d)\n",
3288 		cache->uc_allocbucket,
3289 		cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0,
3290 		cache->uc_freebucket,
3291 		cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0);
3292 }
3293 
3294 static void
3295 uma_print_keg(uma_keg_t keg)
3296 {
3297 	uma_slab_t slab;
3298 
3299 	printf("keg: %s(%p) size %d(%d) flags %#x ipers %d ppera %d "
3300 	    "out %d free %d limit %d\n",
3301 	    keg->uk_name, keg, keg->uk_size, keg->uk_rsize, keg->uk_flags,
3302 	    keg->uk_ipers, keg->uk_ppera,
3303 	    (keg->uk_pages / keg->uk_ppera) * keg->uk_ipers - keg->uk_free,
3304 	    keg->uk_free, (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers);
3305 	printf("Part slabs:\n");
3306 	LIST_FOREACH(slab, &keg->uk_part_slab, us_link)
3307 		slab_print(slab);
3308 	printf("Free slabs:\n");
3309 	LIST_FOREACH(slab, &keg->uk_free_slab, us_link)
3310 		slab_print(slab);
3311 	printf("Full slabs:\n");
3312 	LIST_FOREACH(slab, &keg->uk_full_slab, us_link)
3313 		slab_print(slab);
3314 }
3315 
3316 void
3317 uma_print_zone(uma_zone_t zone)
3318 {
3319 	uma_cache_t cache;
3320 	uma_klink_t kl;
3321 	int i;
3322 
3323 	printf("zone: %s(%p) size %d flags %#x\n",
3324 	    zone->uz_name, zone, zone->uz_size, zone->uz_flags);
3325 	LIST_FOREACH(kl, &zone->uz_kegs, kl_link)
3326 		uma_print_keg(kl->kl_keg);
3327 	CPU_FOREACH(i) {
3328 		cache = &zone->uz_cpu[i];
3329 		printf("CPU %d Cache:\n", i);
3330 		cache_print(cache);
3331 	}
3332 }
3333 
3334 #ifdef DDB
3335 /*
3336  * Generate statistics across both the zone and its per-cpu cache's.  Return
3337  * desired statistics if the pointer is non-NULL for that statistic.
3338  *
3339  * Note: does not update the zone statistics, as it can't safely clear the
3340  * per-CPU cache statistic.
3341  *
3342  * XXXRW: Following the uc_allocbucket and uc_freebucket pointers here isn't
3343  * safe from off-CPU; we should modify the caches to track this information
3344  * directly so that we don't have to.
3345  */
3346 static void
3347 uma_zone_sumstat(uma_zone_t z, int *cachefreep, uint64_t *allocsp,
3348     uint64_t *freesp, uint64_t *sleepsp)
3349 {
3350 	uma_cache_t cache;
3351 	uint64_t allocs, frees, sleeps;
3352 	int cachefree, cpu;
3353 
3354 	allocs = frees = sleeps = 0;
3355 	cachefree = 0;
3356 	CPU_FOREACH(cpu) {
3357 		cache = &z->uz_cpu[cpu];
3358 		if (cache->uc_allocbucket != NULL)
3359 			cachefree += cache->uc_allocbucket->ub_cnt;
3360 		if (cache->uc_freebucket != NULL)
3361 			cachefree += cache->uc_freebucket->ub_cnt;
3362 		allocs += cache->uc_allocs;
3363 		frees += cache->uc_frees;
3364 	}
3365 	allocs += z->uz_allocs;
3366 	frees += z->uz_frees;
3367 	sleeps += z->uz_sleeps;
3368 	if (cachefreep != NULL)
3369 		*cachefreep = cachefree;
3370 	if (allocsp != NULL)
3371 		*allocsp = allocs;
3372 	if (freesp != NULL)
3373 		*freesp = frees;
3374 	if (sleepsp != NULL)
3375 		*sleepsp = sleeps;
3376 }
3377 #endif /* DDB */
3378 
3379 static int
3380 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
3381 {
3382 	uma_keg_t kz;
3383 	uma_zone_t z;
3384 	int count;
3385 
3386 	count = 0;
3387 	rw_rlock(&uma_rwlock);
3388 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3389 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
3390 			count++;
3391 	}
3392 	rw_runlock(&uma_rwlock);
3393 	return (sysctl_handle_int(oidp, &count, 0, req));
3394 }
3395 
3396 static int
3397 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
3398 {
3399 	struct uma_stream_header ush;
3400 	struct uma_type_header uth;
3401 	struct uma_percpu_stat ups;
3402 	uma_bucket_t bucket;
3403 	struct sbuf sbuf;
3404 	uma_cache_t cache;
3405 	uma_klink_t kl;
3406 	uma_keg_t kz;
3407 	uma_zone_t z;
3408 	uma_keg_t k;
3409 	int count, error, i;
3410 
3411 	error = sysctl_wire_old_buffer(req, 0);
3412 	if (error != 0)
3413 		return (error);
3414 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
3415 	sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL);
3416 
3417 	count = 0;
3418 	rw_rlock(&uma_rwlock);
3419 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3420 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
3421 			count++;
3422 	}
3423 
3424 	/*
3425 	 * Insert stream header.
3426 	 */
3427 	bzero(&ush, sizeof(ush));
3428 	ush.ush_version = UMA_STREAM_VERSION;
3429 	ush.ush_maxcpus = (mp_maxid + 1);
3430 	ush.ush_count = count;
3431 	(void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
3432 
3433 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3434 		LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3435 			bzero(&uth, sizeof(uth));
3436 			ZONE_LOCK(z);
3437 			strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
3438 			uth.uth_align = kz->uk_align;
3439 			uth.uth_size = kz->uk_size;
3440 			uth.uth_rsize = kz->uk_rsize;
3441 			LIST_FOREACH(kl, &z->uz_kegs, kl_link) {
3442 				k = kl->kl_keg;
3443 				uth.uth_maxpages += k->uk_maxpages;
3444 				uth.uth_pages += k->uk_pages;
3445 				uth.uth_keg_free += k->uk_free;
3446 				uth.uth_limit = (k->uk_maxpages / k->uk_ppera)
3447 				    * k->uk_ipers;
3448 			}
3449 
3450 			/*
3451 			 * A zone is secondary is it is not the first entry
3452 			 * on the keg's zone list.
3453 			 */
3454 			if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
3455 			    (LIST_FIRST(&kz->uk_zones) != z))
3456 				uth.uth_zone_flags = UTH_ZONE_SECONDARY;
3457 
3458 			LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3459 				uth.uth_zone_free += bucket->ub_cnt;
3460 			uth.uth_allocs = z->uz_allocs;
3461 			uth.uth_frees = z->uz_frees;
3462 			uth.uth_fails = z->uz_fails;
3463 			uth.uth_sleeps = z->uz_sleeps;
3464 			(void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
3465 			/*
3466 			 * While it is not normally safe to access the cache
3467 			 * bucket pointers while not on the CPU that owns the
3468 			 * cache, we only allow the pointers to be exchanged
3469 			 * without the zone lock held, not invalidated, so
3470 			 * accept the possible race associated with bucket
3471 			 * exchange during monitoring.
3472 			 */
3473 			for (i = 0; i < (mp_maxid + 1); i++) {
3474 				bzero(&ups, sizeof(ups));
3475 				if (kz->uk_flags & UMA_ZFLAG_INTERNAL)
3476 					goto skip;
3477 				if (CPU_ABSENT(i))
3478 					goto skip;
3479 				cache = &z->uz_cpu[i];
3480 				if (cache->uc_allocbucket != NULL)
3481 					ups.ups_cache_free +=
3482 					    cache->uc_allocbucket->ub_cnt;
3483 				if (cache->uc_freebucket != NULL)
3484 					ups.ups_cache_free +=
3485 					    cache->uc_freebucket->ub_cnt;
3486 				ups.ups_allocs = cache->uc_allocs;
3487 				ups.ups_frees = cache->uc_frees;
3488 skip:
3489 				(void)sbuf_bcat(&sbuf, &ups, sizeof(ups));
3490 			}
3491 			ZONE_UNLOCK(z);
3492 		}
3493 	}
3494 	rw_runlock(&uma_rwlock);
3495 	error = sbuf_finish(&sbuf);
3496 	sbuf_delete(&sbuf);
3497 	return (error);
3498 }
3499 
3500 int
3501 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS)
3502 {
3503 	uma_zone_t zone = *(uma_zone_t *)arg1;
3504 	int error, max;
3505 
3506 	max = uma_zone_get_max(zone);
3507 	error = sysctl_handle_int(oidp, &max, 0, req);
3508 	if (error || !req->newptr)
3509 		return (error);
3510 
3511 	uma_zone_set_max(zone, max);
3512 
3513 	return (0);
3514 }
3515 
3516 int
3517 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS)
3518 {
3519 	uma_zone_t zone = *(uma_zone_t *)arg1;
3520 	int cur;
3521 
3522 	cur = uma_zone_get_cur(zone);
3523 	return (sysctl_handle_int(oidp, &cur, 0, req));
3524 }
3525 
3526 #ifdef INVARIANTS
3527 static uma_slab_t
3528 uma_dbg_getslab(uma_zone_t zone, void *item)
3529 {
3530 	uma_slab_t slab;
3531 	uma_keg_t keg;
3532 	uint8_t *mem;
3533 
3534 	mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
3535 	if (zone->uz_flags & UMA_ZONE_VTOSLAB) {
3536 		slab = vtoslab((vm_offset_t)mem);
3537 	} else {
3538 		/*
3539 		 * It is safe to return the slab here even though the
3540 		 * zone is unlocked because the item's allocation state
3541 		 * essentially holds a reference.
3542 		 */
3543 		ZONE_LOCK(zone);
3544 		keg = LIST_FIRST(&zone->uz_kegs)->kl_keg;
3545 		if (keg->uk_flags & UMA_ZONE_HASH)
3546 			slab = hash_sfind(&keg->uk_hash, mem);
3547 		else
3548 			slab = (uma_slab_t)(mem + keg->uk_pgoff);
3549 		ZONE_UNLOCK(zone);
3550 	}
3551 
3552 	return (slab);
3553 }
3554 
3555 /*
3556  * Set up the slab's freei data such that uma_dbg_free can function.
3557  *
3558  */
3559 static void
3560 uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item)
3561 {
3562 	uma_keg_t keg;
3563 	int freei;
3564 
3565 	if (zone_first_keg(zone) == NULL)
3566 		return;
3567 	if (slab == NULL) {
3568 		slab = uma_dbg_getslab(zone, item);
3569 		if (slab == NULL)
3570 			panic("uma: item %p did not belong to zone %s\n",
3571 			    item, zone->uz_name);
3572 	}
3573 	keg = slab->us_keg;
3574 	freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
3575 
3576 	if (BIT_ISSET(SLAB_SETSIZE, freei, &slab->us_debugfree))
3577 		panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)\n",
3578 		    item, zone, zone->uz_name, slab, freei);
3579 	BIT_SET_ATOMIC(SLAB_SETSIZE, freei, &slab->us_debugfree);
3580 
3581 	return;
3582 }
3583 
3584 /*
3585  * Verifies freed addresses.  Checks for alignment, valid slab membership
3586  * and duplicate frees.
3587  *
3588  */
3589 static void
3590 uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item)
3591 {
3592 	uma_keg_t keg;
3593 	int freei;
3594 
3595 	if (zone_first_keg(zone) == NULL)
3596 		return;
3597 	if (slab == NULL) {
3598 		slab = uma_dbg_getslab(zone, item);
3599 		if (slab == NULL)
3600 			panic("uma: Freed item %p did not belong to zone %s\n",
3601 			    item, zone->uz_name);
3602 	}
3603 	keg = slab->us_keg;
3604 	freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
3605 
3606 	if (freei >= keg->uk_ipers)
3607 		panic("Invalid free of %p from zone %p(%s) slab %p(%d)\n",
3608 		    item, zone, zone->uz_name, slab, freei);
3609 
3610 	if (((freei * keg->uk_rsize) + slab->us_data) != item)
3611 		panic("Unaligned free of %p from zone %p(%s) slab %p(%d)\n",
3612 		    item, zone, zone->uz_name, slab, freei);
3613 
3614 	if (!BIT_ISSET(SLAB_SETSIZE, freei, &slab->us_debugfree))
3615 		panic("Duplicate free of %p from zone %p(%s) slab %p(%d)\n",
3616 		    item, zone, zone->uz_name, slab, freei);
3617 
3618 	BIT_CLR_ATOMIC(SLAB_SETSIZE, freei, &slab->us_debugfree);
3619 }
3620 #endif /* INVARIANTS */
3621 
3622 #ifdef DDB
3623 DB_SHOW_COMMAND(uma, db_show_uma)
3624 {
3625 	uint64_t allocs, frees, sleeps;
3626 	uma_bucket_t bucket;
3627 	uma_keg_t kz;
3628 	uma_zone_t z;
3629 	int cachefree;
3630 
3631 	db_printf("%18s %8s %8s %8s %12s %8s %8s\n", "Zone", "Size", "Used",
3632 	    "Free", "Requests", "Sleeps", "Bucket");
3633 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3634 		LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3635 			if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
3636 				allocs = z->uz_allocs;
3637 				frees = z->uz_frees;
3638 				sleeps = z->uz_sleeps;
3639 				cachefree = 0;
3640 			} else
3641 				uma_zone_sumstat(z, &cachefree, &allocs,
3642 				    &frees, &sleeps);
3643 			if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
3644 			    (LIST_FIRST(&kz->uk_zones) != z)))
3645 				cachefree += kz->uk_free;
3646 			LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3647 				cachefree += bucket->ub_cnt;
3648 			db_printf("%18s %8ju %8jd %8d %12ju %8ju %8u\n",
3649 			    z->uz_name, (uintmax_t)kz->uk_size,
3650 			    (intmax_t)(allocs - frees), cachefree,
3651 			    (uintmax_t)allocs, sleeps, z->uz_count);
3652 			if (db_pager_quit)
3653 				return;
3654 		}
3655 	}
3656 }
3657 
3658 DB_SHOW_COMMAND(umacache, db_show_umacache)
3659 {
3660 	uint64_t allocs, frees;
3661 	uma_bucket_t bucket;
3662 	uma_zone_t z;
3663 	int cachefree;
3664 
3665 	db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
3666 	    "Requests", "Bucket");
3667 	LIST_FOREACH(z, &uma_cachezones, uz_link) {
3668 		uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL);
3669 		LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3670 			cachefree += bucket->ub_cnt;
3671 		db_printf("%18s %8ju %8jd %8d %12ju %8u\n",
3672 		    z->uz_name, (uintmax_t)z->uz_size,
3673 		    (intmax_t)(allocs - frees), cachefree,
3674 		    (uintmax_t)allocs, z->uz_count);
3675 		if (db_pager_quit)
3676 			return;
3677 	}
3678 }
3679 #endif	/* DDB */
3680