xref: /freebsd/sys/vm/uma_core.c (revision d184218c18d067f8fd47203f54ab02a7b2ed9b11)
1 /*-
2  * Copyright (c) 2002-2005, 2009 Jeffrey Roberson <jeff@FreeBSD.org>
3  * Copyright (c) 2004, 2005 Bosko Milekic <bmilekic@FreeBSD.org>
4  * Copyright (c) 2004-2006 Robert N. M. Watson
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice unmodified, this list of conditions, and the following
12  *    disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
18  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
19  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
20  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
21  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
22  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
23  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
24  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
25  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
26  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
27  */
28 
29 /*
30  * uma_core.c  Implementation of the Universal Memory allocator
31  *
32  * This allocator is intended to replace the multitude of similar object caches
33  * in the standard FreeBSD kernel.  The intent is to be flexible as well as
34  * effecient.  A primary design goal is to return unused memory to the rest of
35  * the system.  This will make the system as a whole more flexible due to the
36  * ability to move memory to subsystems which most need it instead of leaving
37  * pools of reserved memory unused.
38  *
39  * The basic ideas stem from similar slab/zone based allocators whose algorithms
40  * are well known.
41  *
42  */
43 
44 /*
45  * TODO:
46  *	- Improve memory usage for large allocations
47  *	- Investigate cache size adjustments
48  */
49 
50 #include <sys/cdefs.h>
51 __FBSDID("$FreeBSD$");
52 
53 /* I should really use ktr.. */
54 /*
55 #define UMA_DEBUG 1
56 #define UMA_DEBUG_ALLOC 1
57 #define UMA_DEBUG_ALLOC_1 1
58 */
59 
60 #include "opt_ddb.h"
61 #include "opt_param.h"
62 #include "opt_vm.h"
63 
64 #include <sys/param.h>
65 #include <sys/systm.h>
66 #include <sys/kernel.h>
67 #include <sys/types.h>
68 #include <sys/queue.h>
69 #include <sys/malloc.h>
70 #include <sys/ktr.h>
71 #include <sys/lock.h>
72 #include <sys/sysctl.h>
73 #include <sys/mutex.h>
74 #include <sys/proc.h>
75 #include <sys/rwlock.h>
76 #include <sys/sbuf.h>
77 #include <sys/smp.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 static uma_zone_t slabrefzone;	/* With refcounters (for UMA_ZONE_REFCNT) */
112 
113 /*
114  * The initial hash tables come out of this zone so they can be allocated
115  * prior to malloc coming up.
116  */
117 static uma_zone_t hashzone;
118 
119 /* The boot-time adjusted value for cache line alignment. */
120 int uma_align_cache = 64 - 1;
121 
122 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets");
123 
124 /*
125  * Are we allowed to allocate buckets?
126  */
127 static int bucketdisable = 1;
128 
129 /* Linked list of all kegs in the system */
130 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs);
131 
132 /* This mutex protects the keg list */
133 static struct mtx uma_mtx;
134 
135 /* Linked list of boot time pages */
136 static LIST_HEAD(,uma_slab) uma_boot_pages =
137     LIST_HEAD_INITIALIZER(uma_boot_pages);
138 
139 /* This mutex protects the boot time pages list */
140 static struct mtx uma_boot_pages_mtx;
141 
142 /* Is the VM done starting up? */
143 static int booted = 0;
144 #define	UMA_STARTUP	1
145 #define	UMA_STARTUP2	2
146 
147 /* Maximum number of allowed items-per-slab if the slab header is OFFPAGE */
148 static u_int uma_max_ipers;
149 static u_int uma_max_ipers_ref;
150 
151 /*
152  * This is the handle used to schedule events that need to happen
153  * outside of the allocation fast path.
154  */
155 static struct callout uma_callout;
156 #define	UMA_TIMEOUT	20		/* Seconds for callout interval. */
157 
158 /*
159  * This structure is passed as the zone ctor arg so that I don't have to create
160  * a special allocation function just for zones.
161  */
162 struct uma_zctor_args {
163 	const char *name;
164 	size_t size;
165 	uma_ctor ctor;
166 	uma_dtor dtor;
167 	uma_init uminit;
168 	uma_fini fini;
169 	uma_keg_t keg;
170 	int align;
171 	uint32_t flags;
172 };
173 
174 struct uma_kctor_args {
175 	uma_zone_t zone;
176 	size_t size;
177 	uma_init uminit;
178 	uma_fini fini;
179 	int align;
180 	uint32_t flags;
181 };
182 
183 struct uma_bucket_zone {
184 	uma_zone_t	ubz_zone;
185 	char		*ubz_name;
186 	int		ubz_entries;
187 };
188 
189 #define	BUCKET_MAX	128
190 
191 struct uma_bucket_zone bucket_zones[] = {
192 	{ NULL, "16 Bucket", 16 },
193 	{ NULL, "32 Bucket", 32 },
194 	{ NULL, "64 Bucket", 64 },
195 	{ NULL, "128 Bucket", 128 },
196 	{ NULL, NULL, 0}
197 };
198 
199 #define	BUCKET_SHIFT	4
200 #define	BUCKET_ZONES	((BUCKET_MAX >> BUCKET_SHIFT) + 1)
201 
202 /*
203  * bucket_size[] maps requested bucket sizes to zones that allocate a bucket
204  * of approximately the right size.
205  */
206 static uint8_t bucket_size[BUCKET_ZONES];
207 
208 /*
209  * Flags and enumerations to be passed to internal functions.
210  */
211 enum zfreeskip { SKIP_NONE, SKIP_DTOR, SKIP_FINI };
212 
213 #define	ZFREE_STATFAIL	0x00000001	/* Update zone failure statistic. */
214 #define	ZFREE_STATFREE	0x00000002	/* Update zone free statistic. */
215 
216 /* Prototypes.. */
217 
218 static void *noobj_alloc(uma_zone_t, int, uint8_t *, int);
219 static void *page_alloc(uma_zone_t, int, uint8_t *, int);
220 static void *startup_alloc(uma_zone_t, int, uint8_t *, int);
221 static void page_free(void *, int, uint8_t);
222 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int);
223 static void cache_drain(uma_zone_t);
224 static void bucket_drain(uma_zone_t, uma_bucket_t);
225 static void bucket_cache_drain(uma_zone_t zone);
226 static int keg_ctor(void *, int, void *, int);
227 static void keg_dtor(void *, int, void *);
228 static int zone_ctor(void *, int, void *, int);
229 static void zone_dtor(void *, int, void *);
230 static int zero_init(void *, int, int);
231 static void keg_small_init(uma_keg_t keg);
232 static void keg_large_init(uma_keg_t keg);
233 static void zone_foreach(void (*zfunc)(uma_zone_t));
234 static void zone_timeout(uma_zone_t zone);
235 static int hash_alloc(struct uma_hash *);
236 static int hash_expand(struct uma_hash *, struct uma_hash *);
237 static void hash_free(struct uma_hash *hash);
238 static void uma_timeout(void *);
239 static void uma_startup3(void);
240 static void *zone_alloc_item(uma_zone_t, void *, int);
241 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip,
242     int);
243 static void bucket_enable(void);
244 static void bucket_init(void);
245 static uma_bucket_t bucket_alloc(int, int);
246 static void bucket_free(uma_bucket_t);
247 static void bucket_zone_drain(void);
248 static int zone_alloc_bucket(uma_zone_t zone, int flags);
249 static uma_slab_t zone_fetch_slab(uma_zone_t zone, uma_keg_t last, int flags);
250 static uma_slab_t zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int flags);
251 static void *slab_alloc_item(uma_zone_t zone, uma_slab_t slab);
252 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit,
253     uma_fini fini, int align, uint32_t flags);
254 static inline void zone_relock(uma_zone_t zone, uma_keg_t keg);
255 static inline void keg_relock(uma_keg_t keg, uma_zone_t zone);
256 
257 void uma_print_zone(uma_zone_t);
258 void uma_print_stats(void);
259 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS);
260 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS);
261 
262 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
263 
264 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLTYPE_INT,
265     0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones");
266 
267 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLTYPE_STRUCT,
268     0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats");
269 
270 static int zone_warnings = 1;
271 TUNABLE_INT("vm.zone_warnings", &zone_warnings);
272 SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RW, &zone_warnings, 0,
273     "Warn when UMA zones becomes full");
274 
275 /*
276  * This routine checks to see whether or not it's safe to enable buckets.
277  */
278 
279 static void
280 bucket_enable(void)
281 {
282 	bucketdisable = vm_page_count_min();
283 }
284 
285 /*
286  * Initialize bucket_zones, the array of zones of buckets of various sizes.
287  *
288  * For each zone, calculate the memory required for each bucket, consisting
289  * of the header and an array of pointers.  Initialize bucket_size[] to point
290  * the range of appropriate bucket sizes at the zone.
291  */
292 static void
293 bucket_init(void)
294 {
295 	struct uma_bucket_zone *ubz;
296 	int i;
297 	int j;
298 
299 	for (i = 0, j = 0; bucket_zones[j].ubz_entries != 0; j++) {
300 		int size;
301 
302 		ubz = &bucket_zones[j];
303 		size = roundup(sizeof(struct uma_bucket), sizeof(void *));
304 		size += sizeof(void *) * ubz->ubz_entries;
305 		ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
306 		    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
307 		    UMA_ZFLAG_INTERNAL | UMA_ZFLAG_BUCKET);
308 		for (; i <= ubz->ubz_entries; i += (1 << BUCKET_SHIFT))
309 			bucket_size[i >> BUCKET_SHIFT] = j;
310 	}
311 }
312 
313 /*
314  * Given a desired number of entries for a bucket, return the zone from which
315  * to allocate the bucket.
316  */
317 static struct uma_bucket_zone *
318 bucket_zone_lookup(int entries)
319 {
320 	int idx;
321 
322 	idx = howmany(entries, 1 << BUCKET_SHIFT);
323 	return (&bucket_zones[bucket_size[idx]]);
324 }
325 
326 static uma_bucket_t
327 bucket_alloc(int entries, int bflags)
328 {
329 	struct uma_bucket_zone *ubz;
330 	uma_bucket_t bucket;
331 
332 	/*
333 	 * This is to stop us from allocating per cpu buckets while we're
334 	 * running out of vm.boot_pages.  Otherwise, we would exhaust the
335 	 * boot pages.  This also prevents us from allocating buckets in
336 	 * low memory situations.
337 	 */
338 	if (bucketdisable)
339 		return (NULL);
340 
341 	ubz = bucket_zone_lookup(entries);
342 	bucket = zone_alloc_item(ubz->ubz_zone, NULL, bflags);
343 	if (bucket) {
344 #ifdef INVARIANTS
345 		bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
346 #endif
347 		bucket->ub_cnt = 0;
348 		bucket->ub_entries = ubz->ubz_entries;
349 	}
350 
351 	return (bucket);
352 }
353 
354 static void
355 bucket_free(uma_bucket_t bucket)
356 {
357 	struct uma_bucket_zone *ubz;
358 
359 	ubz = bucket_zone_lookup(bucket->ub_entries);
360 	zone_free_item(ubz->ubz_zone, bucket, NULL, SKIP_NONE,
361 	    ZFREE_STATFREE);
362 }
363 
364 static void
365 bucket_zone_drain(void)
366 {
367 	struct uma_bucket_zone *ubz;
368 
369 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
370 		zone_drain(ubz->ubz_zone);
371 }
372 
373 static void
374 zone_log_warning(uma_zone_t zone)
375 {
376 	static const struct timeval warninterval = { 300, 0 };
377 
378 	if (!zone_warnings || zone->uz_warning == NULL)
379 		return;
380 
381 	if (ratecheck(&zone->uz_ratecheck, &warninterval))
382 		printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning);
383 }
384 
385 static inline uma_keg_t
386 zone_first_keg(uma_zone_t zone)
387 {
388 
389 	return (LIST_FIRST(&zone->uz_kegs)->kl_keg);
390 }
391 
392 static void
393 zone_foreach_keg(uma_zone_t zone, void (*kegfn)(uma_keg_t))
394 {
395 	uma_klink_t klink;
396 
397 	LIST_FOREACH(klink, &zone->uz_kegs, kl_link)
398 		kegfn(klink->kl_keg);
399 }
400 
401 /*
402  * Routine called by timeout which is used to fire off some time interval
403  * based calculations.  (stats, hash size, etc.)
404  *
405  * Arguments:
406  *	arg   Unused
407  *
408  * Returns:
409  *	Nothing
410  */
411 static void
412 uma_timeout(void *unused)
413 {
414 	bucket_enable();
415 	zone_foreach(zone_timeout);
416 
417 	/* Reschedule this event */
418 	callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
419 }
420 
421 /*
422  * Routine to perform timeout driven calculations.  This expands the
423  * hashes and does per cpu statistics aggregation.
424  *
425  *  Returns nothing.
426  */
427 static void
428 keg_timeout(uma_keg_t keg)
429 {
430 
431 	KEG_LOCK(keg);
432 	/*
433 	 * Expand the keg hash table.
434 	 *
435 	 * This is done if the number of slabs is larger than the hash size.
436 	 * What I'm trying to do here is completely reduce collisions.  This
437 	 * may be a little aggressive.  Should I allow for two collisions max?
438 	 */
439 	if (keg->uk_flags & UMA_ZONE_HASH &&
440 	    keg->uk_pages / keg->uk_ppera >= keg->uk_hash.uh_hashsize) {
441 		struct uma_hash newhash;
442 		struct uma_hash oldhash;
443 		int ret;
444 
445 		/*
446 		 * This is so involved because allocating and freeing
447 		 * while the keg lock is held will lead to deadlock.
448 		 * I have to do everything in stages and check for
449 		 * races.
450 		 */
451 		newhash = keg->uk_hash;
452 		KEG_UNLOCK(keg);
453 		ret = hash_alloc(&newhash);
454 		KEG_LOCK(keg);
455 		if (ret) {
456 			if (hash_expand(&keg->uk_hash, &newhash)) {
457 				oldhash = keg->uk_hash;
458 				keg->uk_hash = newhash;
459 			} else
460 				oldhash = newhash;
461 
462 			KEG_UNLOCK(keg);
463 			hash_free(&oldhash);
464 			KEG_LOCK(keg);
465 		}
466 	}
467 	KEG_UNLOCK(keg);
468 }
469 
470 static void
471 zone_timeout(uma_zone_t zone)
472 {
473 
474 	zone_foreach_keg(zone, &keg_timeout);
475 }
476 
477 /*
478  * Allocate and zero fill the next sized hash table from the appropriate
479  * backing store.
480  *
481  * Arguments:
482  *	hash  A new hash structure with the old hash size in uh_hashsize
483  *
484  * Returns:
485  *	1 on sucess and 0 on failure.
486  */
487 static int
488 hash_alloc(struct uma_hash *hash)
489 {
490 	int oldsize;
491 	int alloc;
492 
493 	oldsize = hash->uh_hashsize;
494 
495 	/* We're just going to go to a power of two greater */
496 	if (oldsize)  {
497 		hash->uh_hashsize = oldsize * 2;
498 		alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
499 		hash->uh_slab_hash = (struct slabhead *)malloc(alloc,
500 		    M_UMAHASH, M_NOWAIT);
501 	} else {
502 		alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
503 		hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
504 		    M_WAITOK);
505 		hash->uh_hashsize = UMA_HASH_SIZE_INIT;
506 	}
507 	if (hash->uh_slab_hash) {
508 		bzero(hash->uh_slab_hash, alloc);
509 		hash->uh_hashmask = hash->uh_hashsize - 1;
510 		return (1);
511 	}
512 
513 	return (0);
514 }
515 
516 /*
517  * Expands the hash table for HASH zones.  This is done from zone_timeout
518  * to reduce collisions.  This must not be done in the regular allocation
519  * path, otherwise, we can recurse on the vm while allocating pages.
520  *
521  * Arguments:
522  *	oldhash  The hash you want to expand
523  *	newhash  The hash structure for the new table
524  *
525  * Returns:
526  *	Nothing
527  *
528  * Discussion:
529  */
530 static int
531 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
532 {
533 	uma_slab_t slab;
534 	int hval;
535 	int i;
536 
537 	if (!newhash->uh_slab_hash)
538 		return (0);
539 
540 	if (oldhash->uh_hashsize >= newhash->uh_hashsize)
541 		return (0);
542 
543 	/*
544 	 * I need to investigate hash algorithms for resizing without a
545 	 * full rehash.
546 	 */
547 
548 	for (i = 0; i < oldhash->uh_hashsize; i++)
549 		while (!SLIST_EMPTY(&oldhash->uh_slab_hash[i])) {
550 			slab = SLIST_FIRST(&oldhash->uh_slab_hash[i]);
551 			SLIST_REMOVE_HEAD(&oldhash->uh_slab_hash[i], us_hlink);
552 			hval = UMA_HASH(newhash, slab->us_data);
553 			SLIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
554 			    slab, us_hlink);
555 		}
556 
557 	return (1);
558 }
559 
560 /*
561  * Free the hash bucket to the appropriate backing store.
562  *
563  * Arguments:
564  *	slab_hash  The hash bucket we're freeing
565  *	hashsize   The number of entries in that hash bucket
566  *
567  * Returns:
568  *	Nothing
569  */
570 static void
571 hash_free(struct uma_hash *hash)
572 {
573 	if (hash->uh_slab_hash == NULL)
574 		return;
575 	if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
576 		zone_free_item(hashzone,
577 		    hash->uh_slab_hash, NULL, SKIP_NONE, ZFREE_STATFREE);
578 	else
579 		free(hash->uh_slab_hash, M_UMAHASH);
580 }
581 
582 /*
583  * Frees all outstanding items in a bucket
584  *
585  * Arguments:
586  *	zone   The zone to free to, must be unlocked.
587  *	bucket The free/alloc bucket with items, cpu queue must be locked.
588  *
589  * Returns:
590  *	Nothing
591  */
592 
593 static void
594 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
595 {
596 	void *item;
597 
598 	if (bucket == NULL)
599 		return;
600 
601 	while (bucket->ub_cnt > 0)  {
602 		bucket->ub_cnt--;
603 		item = bucket->ub_bucket[bucket->ub_cnt];
604 #ifdef INVARIANTS
605 		bucket->ub_bucket[bucket->ub_cnt] = NULL;
606 		KASSERT(item != NULL,
607 		    ("bucket_drain: botched ptr, item is NULL"));
608 #endif
609 		zone_free_item(zone, item, NULL, SKIP_DTOR, 0);
610 	}
611 }
612 
613 /*
614  * Drains the per cpu caches for a zone.
615  *
616  * NOTE: This may only be called while the zone is being turn down, and not
617  * during normal operation.  This is necessary in order that we do not have
618  * to migrate CPUs to drain the per-CPU caches.
619  *
620  * Arguments:
621  *	zone     The zone to drain, must be unlocked.
622  *
623  * Returns:
624  *	Nothing
625  */
626 static void
627 cache_drain(uma_zone_t zone)
628 {
629 	uma_cache_t cache;
630 	int cpu;
631 
632 	/*
633 	 * XXX: It is safe to not lock the per-CPU caches, because we're
634 	 * tearing down the zone anyway.  I.e., there will be no further use
635 	 * of the caches at this point.
636 	 *
637 	 * XXX: It would good to be able to assert that the zone is being
638 	 * torn down to prevent improper use of cache_drain().
639 	 *
640 	 * XXX: We lock the zone before passing into bucket_cache_drain() as
641 	 * it is used elsewhere.  Should the tear-down path be made special
642 	 * there in some form?
643 	 */
644 	CPU_FOREACH(cpu) {
645 		cache = &zone->uz_cpu[cpu];
646 		bucket_drain(zone, cache->uc_allocbucket);
647 		bucket_drain(zone, cache->uc_freebucket);
648 		if (cache->uc_allocbucket != NULL)
649 			bucket_free(cache->uc_allocbucket);
650 		if (cache->uc_freebucket != NULL)
651 			bucket_free(cache->uc_freebucket);
652 		cache->uc_allocbucket = cache->uc_freebucket = NULL;
653 	}
654 	ZONE_LOCK(zone);
655 	bucket_cache_drain(zone);
656 	ZONE_UNLOCK(zone);
657 }
658 
659 /*
660  * Drain the cached buckets from a zone.  Expects a locked zone on entry.
661  */
662 static void
663 bucket_cache_drain(uma_zone_t zone)
664 {
665 	uma_bucket_t bucket;
666 
667 	/*
668 	 * Drain the bucket queues and free the buckets, we just keep two per
669 	 * cpu (alloc/free).
670 	 */
671 	while ((bucket = LIST_FIRST(&zone->uz_full_bucket)) != NULL) {
672 		LIST_REMOVE(bucket, ub_link);
673 		ZONE_UNLOCK(zone);
674 		bucket_drain(zone, bucket);
675 		bucket_free(bucket);
676 		ZONE_LOCK(zone);
677 	}
678 
679 	/* Now we do the free queue.. */
680 	while ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) {
681 		LIST_REMOVE(bucket, ub_link);
682 		bucket_free(bucket);
683 	}
684 }
685 
686 /*
687  * Frees pages from a keg back to the system.  This is done on demand from
688  * the pageout daemon.
689  *
690  * Returns nothing.
691  */
692 static void
693 keg_drain(uma_keg_t keg)
694 {
695 	struct slabhead freeslabs = { 0 };
696 	uma_slab_t slab;
697 	uma_slab_t n;
698 	uint8_t flags;
699 	uint8_t *mem;
700 	int i;
701 
702 	/*
703 	 * We don't want to take pages from statically allocated kegs at this
704 	 * time
705 	 */
706 	if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL)
707 		return;
708 
709 #ifdef UMA_DEBUG
710 	printf("%s free items: %u\n", keg->uk_name, keg->uk_free);
711 #endif
712 	KEG_LOCK(keg);
713 	if (keg->uk_free == 0)
714 		goto finished;
715 
716 	slab = LIST_FIRST(&keg->uk_free_slab);
717 	while (slab) {
718 		n = LIST_NEXT(slab, us_link);
719 
720 		/* We have no where to free these to */
721 		if (slab->us_flags & UMA_SLAB_BOOT) {
722 			slab = n;
723 			continue;
724 		}
725 
726 		LIST_REMOVE(slab, us_link);
727 		keg->uk_pages -= keg->uk_ppera;
728 		keg->uk_free -= keg->uk_ipers;
729 
730 		if (keg->uk_flags & UMA_ZONE_HASH)
731 			UMA_HASH_REMOVE(&keg->uk_hash, slab, slab->us_data);
732 
733 		SLIST_INSERT_HEAD(&freeslabs, slab, us_hlink);
734 
735 		slab = n;
736 	}
737 finished:
738 	KEG_UNLOCK(keg);
739 
740 	while ((slab = SLIST_FIRST(&freeslabs)) != NULL) {
741 		SLIST_REMOVE(&freeslabs, slab, uma_slab, us_hlink);
742 		if (keg->uk_fini)
743 			for (i = 0; i < keg->uk_ipers; i++)
744 				keg->uk_fini(
745 				    slab->us_data + (keg->uk_rsize * i),
746 				    keg->uk_size);
747 		flags = slab->us_flags;
748 		mem = slab->us_data;
749 
750 		if (keg->uk_flags & UMA_ZONE_VTOSLAB) {
751 			vm_object_t obj;
752 
753 			if (flags & UMA_SLAB_KMEM)
754 				obj = kmem_object;
755 			else if (flags & UMA_SLAB_KERNEL)
756 				obj = kernel_object;
757 			else
758 				obj = NULL;
759 			for (i = 0; i < keg->uk_ppera; i++)
760 				vsetobj((vm_offset_t)mem + (i * PAGE_SIZE),
761 				    obj);
762 		}
763 		if (keg->uk_flags & UMA_ZONE_OFFPAGE)
764 			zone_free_item(keg->uk_slabzone, slab, NULL,
765 			    SKIP_NONE, ZFREE_STATFREE);
766 #ifdef UMA_DEBUG
767 		printf("%s: Returning %d bytes.\n",
768 		    keg->uk_name, PAGE_SIZE * keg->uk_ppera);
769 #endif
770 		keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags);
771 	}
772 }
773 
774 static void
775 zone_drain_wait(uma_zone_t zone, int waitok)
776 {
777 
778 	/*
779 	 * Set draining to interlock with zone_dtor() so we can release our
780 	 * locks as we go.  Only dtor() should do a WAITOK call since it
781 	 * is the only call that knows the structure will still be available
782 	 * when it wakes up.
783 	 */
784 	ZONE_LOCK(zone);
785 	while (zone->uz_flags & UMA_ZFLAG_DRAINING) {
786 		if (waitok == M_NOWAIT)
787 			goto out;
788 		mtx_unlock(&uma_mtx);
789 		msleep(zone, zone->uz_lock, PVM, "zonedrain", 1);
790 		mtx_lock(&uma_mtx);
791 	}
792 	zone->uz_flags |= UMA_ZFLAG_DRAINING;
793 	bucket_cache_drain(zone);
794 	ZONE_UNLOCK(zone);
795 	/*
796 	 * The DRAINING flag protects us from being freed while
797 	 * we're running.  Normally the uma_mtx would protect us but we
798 	 * must be able to release and acquire the right lock for each keg.
799 	 */
800 	zone_foreach_keg(zone, &keg_drain);
801 	ZONE_LOCK(zone);
802 	zone->uz_flags &= ~UMA_ZFLAG_DRAINING;
803 	wakeup(zone);
804 out:
805 	ZONE_UNLOCK(zone);
806 }
807 
808 void
809 zone_drain(uma_zone_t zone)
810 {
811 
812 	zone_drain_wait(zone, M_NOWAIT);
813 }
814 
815 /*
816  * Allocate a new slab for a keg.  This does not insert the slab onto a list.
817  *
818  * Arguments:
819  *	wait  Shall we wait?
820  *
821  * Returns:
822  *	The slab that was allocated or NULL if there is no memory and the
823  *	caller specified M_NOWAIT.
824  */
825 static uma_slab_t
826 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int wait)
827 {
828 	uma_slabrefcnt_t slabref;
829 	uma_alloc allocf;
830 	uma_slab_t slab;
831 	uint8_t *mem;
832 	uint8_t flags;
833 	int i;
834 
835 	mtx_assert(&keg->uk_lock, MA_OWNED);
836 	slab = NULL;
837 
838 #ifdef UMA_DEBUG
839 	printf("slab_zalloc:  Allocating a new slab for %s\n", keg->uk_name);
840 #endif
841 	allocf = keg->uk_allocf;
842 	KEG_UNLOCK(keg);
843 
844 	if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
845 		slab = zone_alloc_item(keg->uk_slabzone, NULL, wait);
846 		if (slab == NULL) {
847 			KEG_LOCK(keg);
848 			return NULL;
849 		}
850 	}
851 
852 	/*
853 	 * This reproduces the old vm_zone behavior of zero filling pages the
854 	 * first time they are added to a zone.
855 	 *
856 	 * Malloced items are zeroed in uma_zalloc.
857 	 */
858 
859 	if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
860 		wait |= M_ZERO;
861 	else
862 		wait &= ~M_ZERO;
863 
864 	if (keg->uk_flags & UMA_ZONE_NODUMP)
865 		wait |= M_NODUMP;
866 
867 	/* zone is passed for legacy reasons. */
868 	mem = allocf(zone, keg->uk_ppera * PAGE_SIZE, &flags, wait);
869 	if (mem == NULL) {
870 		if (keg->uk_flags & UMA_ZONE_OFFPAGE)
871 			zone_free_item(keg->uk_slabzone, slab, NULL,
872 			    SKIP_NONE, ZFREE_STATFREE);
873 		KEG_LOCK(keg);
874 		return (NULL);
875 	}
876 
877 	/* Point the slab into the allocated memory */
878 	if (!(keg->uk_flags & UMA_ZONE_OFFPAGE))
879 		slab = (uma_slab_t )(mem + keg->uk_pgoff);
880 
881 	if (keg->uk_flags & UMA_ZONE_VTOSLAB)
882 		for (i = 0; i < keg->uk_ppera; i++)
883 			vsetslab((vm_offset_t)mem + (i * PAGE_SIZE), slab);
884 
885 	slab->us_keg = keg;
886 	slab->us_data = mem;
887 	slab->us_freecount = keg->uk_ipers;
888 	slab->us_firstfree = 0;
889 	slab->us_flags = flags;
890 
891 	if (keg->uk_flags & UMA_ZONE_REFCNT) {
892 		slabref = (uma_slabrefcnt_t)slab;
893 		for (i = 0; i < keg->uk_ipers; i++) {
894 			slabref->us_freelist[i].us_refcnt = 0;
895 			slabref->us_freelist[i].us_item = i+1;
896 		}
897 	} else {
898 		for (i = 0; i < keg->uk_ipers; i++)
899 			slab->us_freelist[i].us_item = i+1;
900 	}
901 
902 	if (keg->uk_init != NULL) {
903 		for (i = 0; i < keg->uk_ipers; i++)
904 			if (keg->uk_init(slab->us_data + (keg->uk_rsize * i),
905 			    keg->uk_size, wait) != 0)
906 				break;
907 		if (i != keg->uk_ipers) {
908 			if (keg->uk_fini != NULL) {
909 				for (i--; i > -1; i--)
910 					keg->uk_fini(slab->us_data +
911 					    (keg->uk_rsize * i),
912 					    keg->uk_size);
913 			}
914 			if (keg->uk_flags & UMA_ZONE_VTOSLAB) {
915 				vm_object_t obj;
916 
917 				if (flags & UMA_SLAB_KMEM)
918 					obj = kmem_object;
919 				else if (flags & UMA_SLAB_KERNEL)
920 					obj = kernel_object;
921 				else
922 					obj = NULL;
923 				for (i = 0; i < keg->uk_ppera; i++)
924 					vsetobj((vm_offset_t)mem +
925 					    (i * PAGE_SIZE), obj);
926 			}
927 			if (keg->uk_flags & UMA_ZONE_OFFPAGE)
928 				zone_free_item(keg->uk_slabzone, slab,
929 				    NULL, SKIP_NONE, ZFREE_STATFREE);
930 			keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera,
931 			    flags);
932 			KEG_LOCK(keg);
933 			return (NULL);
934 		}
935 	}
936 	KEG_LOCK(keg);
937 
938 	if (keg->uk_flags & UMA_ZONE_HASH)
939 		UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
940 
941 	keg->uk_pages += keg->uk_ppera;
942 	keg->uk_free += keg->uk_ipers;
943 
944 	return (slab);
945 }
946 
947 /*
948  * This function is intended to be used early on in place of page_alloc() so
949  * that we may use the boot time page cache to satisfy allocations before
950  * the VM is ready.
951  */
952 static void *
953 startup_alloc(uma_zone_t zone, int bytes, uint8_t *pflag, int wait)
954 {
955 	uma_keg_t keg;
956 	uma_slab_t tmps;
957 	int pages, check_pages;
958 
959 	keg = zone_first_keg(zone);
960 	pages = howmany(bytes, PAGE_SIZE);
961 	check_pages = pages - 1;
962 	KASSERT(pages > 0, ("startup_alloc can't reserve 0 pages\n"));
963 
964 	/*
965 	 * Check our small startup cache to see if it has pages remaining.
966 	 */
967 	mtx_lock(&uma_boot_pages_mtx);
968 
969 	/* First check if we have enough room. */
970 	tmps = LIST_FIRST(&uma_boot_pages);
971 	while (tmps != NULL && check_pages-- > 0)
972 		tmps = LIST_NEXT(tmps, us_link);
973 	if (tmps != NULL) {
974 		/*
975 		 * It's ok to lose tmps references.  The last one will
976 		 * have tmps->us_data pointing to the start address of
977 		 * "pages" contiguous pages of memory.
978 		 */
979 		while (pages-- > 0) {
980 			tmps = LIST_FIRST(&uma_boot_pages);
981 			LIST_REMOVE(tmps, us_link);
982 		}
983 		mtx_unlock(&uma_boot_pages_mtx);
984 		*pflag = tmps->us_flags;
985 		return (tmps->us_data);
986 	}
987 	mtx_unlock(&uma_boot_pages_mtx);
988 	if (booted < UMA_STARTUP2)
989 		panic("UMA: Increase vm.boot_pages");
990 	/*
991 	 * Now that we've booted reset these users to their real allocator.
992 	 */
993 #ifdef UMA_MD_SMALL_ALLOC
994 	keg->uk_allocf = (keg->uk_ppera > 1) ? page_alloc : uma_small_alloc;
995 #else
996 	keg->uk_allocf = page_alloc;
997 #endif
998 	return keg->uk_allocf(zone, bytes, pflag, wait);
999 }
1000 
1001 /*
1002  * Allocates a number of pages from the system
1003  *
1004  * Arguments:
1005  *	bytes  The number of bytes requested
1006  *	wait  Shall we wait?
1007  *
1008  * Returns:
1009  *	A pointer to the alloced memory or possibly
1010  *	NULL if M_NOWAIT is set.
1011  */
1012 static void *
1013 page_alloc(uma_zone_t zone, int bytes, uint8_t *pflag, int wait)
1014 {
1015 	void *p;	/* Returned page */
1016 
1017 	*pflag = UMA_SLAB_KMEM;
1018 	p = (void *) kmem_malloc(kmem_map, bytes, wait);
1019 
1020 	return (p);
1021 }
1022 
1023 /*
1024  * Allocates a number of pages from within an object
1025  *
1026  * Arguments:
1027  *	bytes  The number of bytes requested
1028  *	wait   Shall we wait?
1029  *
1030  * Returns:
1031  *	A pointer to the alloced memory or possibly
1032  *	NULL if M_NOWAIT is set.
1033  */
1034 static void *
1035 noobj_alloc(uma_zone_t zone, int bytes, uint8_t *flags, int wait)
1036 {
1037 	TAILQ_HEAD(, vm_page) alloctail;
1038 	u_long npages;
1039 	vm_offset_t retkva, zkva;
1040 	vm_page_t p, p_next;
1041 	uma_keg_t keg;
1042 
1043 	TAILQ_INIT(&alloctail);
1044 	keg = zone_first_keg(zone);
1045 
1046 	npages = howmany(bytes, PAGE_SIZE);
1047 	while (npages > 0) {
1048 		p = vm_page_alloc(NULL, 0, VM_ALLOC_INTERRUPT |
1049 		    VM_ALLOC_WIRED | VM_ALLOC_NOOBJ);
1050 		if (p != NULL) {
1051 			/*
1052 			 * Since the page does not belong to an object, its
1053 			 * listq is unused.
1054 			 */
1055 			TAILQ_INSERT_TAIL(&alloctail, p, listq);
1056 			npages--;
1057 			continue;
1058 		}
1059 		if (wait & M_WAITOK) {
1060 			VM_WAIT;
1061 			continue;
1062 		}
1063 
1064 		/*
1065 		 * Page allocation failed, free intermediate pages and
1066 		 * exit.
1067 		 */
1068 		TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1069 			vm_page_unwire(p, 0);
1070 			vm_page_free(p);
1071 		}
1072 		return (NULL);
1073 	}
1074 	*flags = UMA_SLAB_PRIV;
1075 	zkva = keg->uk_kva +
1076 	    atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
1077 	retkva = zkva;
1078 	TAILQ_FOREACH(p, &alloctail, listq) {
1079 		pmap_qenter(zkva, &p, 1);
1080 		zkva += PAGE_SIZE;
1081 	}
1082 
1083 	return ((void *)retkva);
1084 }
1085 
1086 /*
1087  * Frees a number of pages to the system
1088  *
1089  * Arguments:
1090  *	mem   A pointer to the memory to be freed
1091  *	size  The size of the memory being freed
1092  *	flags The original p->us_flags field
1093  *
1094  * Returns:
1095  *	Nothing
1096  */
1097 static void
1098 page_free(void *mem, int size, uint8_t flags)
1099 {
1100 	vm_map_t map;
1101 
1102 	if (flags & UMA_SLAB_KMEM)
1103 		map = kmem_map;
1104 	else if (flags & UMA_SLAB_KERNEL)
1105 		map = kernel_map;
1106 	else
1107 		panic("UMA: page_free used with invalid flags %d", flags);
1108 
1109 	kmem_free(map, (vm_offset_t)mem, size);
1110 }
1111 
1112 /*
1113  * Zero fill initializer
1114  *
1115  * Arguments/Returns follow uma_init specifications
1116  */
1117 static int
1118 zero_init(void *mem, int size, int flags)
1119 {
1120 	bzero(mem, size);
1121 	return (0);
1122 }
1123 
1124 /*
1125  * Finish creating a small uma keg.  This calculates ipers, and the keg size.
1126  *
1127  * Arguments
1128  *	keg  The zone we should initialize
1129  *
1130  * Returns
1131  *	Nothing
1132  */
1133 static void
1134 keg_small_init(uma_keg_t keg)
1135 {
1136 	u_int rsize;
1137 	u_int memused;
1138 	u_int wastedspace;
1139 	u_int shsize;
1140 
1141 	if (keg->uk_flags & UMA_ZONE_PCPU) {
1142 		keg->uk_slabsize = sizeof(struct pcpu);
1143 		keg->uk_ppera = howmany(mp_ncpus * sizeof(struct pcpu),
1144 		    PAGE_SIZE);
1145 	} else {
1146 		keg->uk_slabsize = UMA_SLAB_SIZE;
1147 		keg->uk_ppera = 1;
1148 	}
1149 
1150 	rsize = keg->uk_size;
1151 
1152 	if (rsize & keg->uk_align)
1153 		rsize = (rsize & ~keg->uk_align) + (keg->uk_align + 1);
1154 	if (rsize < keg->uk_slabsize / 256)
1155 		rsize = keg->uk_slabsize / 256;
1156 
1157 	keg->uk_rsize = rsize;
1158 
1159 	KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
1160 	    keg->uk_rsize < sizeof(struct pcpu),
1161 	    ("%s: size %u too large", __func__, keg->uk_rsize));
1162 
1163 	if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
1164 		shsize = 0;
1165 	} else if (keg->uk_flags & UMA_ZONE_REFCNT) {
1166 		rsize += UMA_FRITMREF_SZ;	/* linkage & refcnt */
1167 		shsize = sizeof(struct uma_slab_refcnt);
1168 	} else {
1169 		rsize += UMA_FRITM_SZ;	/* Account for linkage */
1170 		shsize = sizeof(struct uma_slab);
1171 	}
1172 
1173 	keg->uk_ipers = (keg->uk_slabsize - shsize) / rsize;
1174 	KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= 256,
1175 	    ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers));
1176 
1177 	memused = keg->uk_ipers * rsize + shsize;
1178 	wastedspace = keg->uk_slabsize - memused;
1179 
1180 	/*
1181 	 * We can't do OFFPAGE if we're internal or if we've been
1182 	 * asked to not go to the VM for buckets.  If we do this we
1183 	 * may end up going to the VM (kmem_map) for slabs which we
1184 	 * do not want to do if we're UMA_ZFLAG_CACHEONLY as a
1185 	 * result of UMA_ZONE_VM, which clearly forbids it.
1186 	 */
1187 	if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) ||
1188 	    (keg->uk_flags & UMA_ZFLAG_CACHEONLY))
1189 		return;
1190 
1191 	if ((wastedspace >= keg->uk_slabsize / UMA_MAX_WASTE) &&
1192 	    (keg->uk_ipers < (keg->uk_slabsize / keg->uk_rsize))) {
1193 		keg->uk_ipers = keg->uk_slabsize / keg->uk_rsize;
1194 		KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= 256,
1195 		    ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers));
1196 #ifdef UMA_DEBUG
1197 		printf("UMA decided we need offpage slab headers for "
1198 		    "keg: %s, calculated wastedspace = %d, "
1199 		    "maximum wasted space allowed = %d, "
1200 		    "calculated ipers = %d, "
1201 		    "new wasted space = %d\n", keg->uk_name, wastedspace,
1202 		    keg->uk_slabsize / UMA_MAX_WASTE, keg->uk_ipers,
1203 		    keg->uk_slabsize - keg->uk_ipers * keg->uk_rsize);
1204 #endif
1205 		keg->uk_flags |= UMA_ZONE_OFFPAGE;
1206 	}
1207 
1208 	if ((keg->uk_flags & UMA_ZONE_OFFPAGE) &&
1209 	    (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1210 		keg->uk_flags |= UMA_ZONE_HASH;
1211 }
1212 
1213 /*
1214  * Finish creating a large (> UMA_SLAB_SIZE) uma kegs.  Just give in and do
1215  * OFFPAGE for now.  When I can allow for more dynamic slab sizes this will be
1216  * more complicated.
1217  *
1218  * Arguments
1219  *	keg  The keg we should initialize
1220  *
1221  * Returns
1222  *	Nothing
1223  */
1224 static void
1225 keg_large_init(uma_keg_t keg)
1226 {
1227 
1228 	KASSERT(keg != NULL, ("Keg is null in keg_large_init"));
1229 	KASSERT((keg->uk_flags & UMA_ZFLAG_CACHEONLY) == 0,
1230 	    ("keg_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY keg"));
1231 	KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0,
1232 	    ("%s: Cannot large-init a UMA_ZONE_PCPU keg", __func__));
1233 
1234 	keg->uk_ppera = howmany(keg->uk_size, PAGE_SIZE);
1235 	keg->uk_slabsize = keg->uk_ppera * PAGE_SIZE;
1236 	keg->uk_ipers = 1;
1237 	keg->uk_rsize = keg->uk_size;
1238 
1239 	/* We can't do OFFPAGE if we're internal, bail out here. */
1240 	if (keg->uk_flags & UMA_ZFLAG_INTERNAL)
1241 		return;
1242 
1243 	keg->uk_flags |= UMA_ZONE_OFFPAGE;
1244 	if ((keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1245 		keg->uk_flags |= UMA_ZONE_HASH;
1246 }
1247 
1248 static void
1249 keg_cachespread_init(uma_keg_t keg)
1250 {
1251 	int alignsize;
1252 	int trailer;
1253 	int pages;
1254 	int rsize;
1255 
1256 	KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0,
1257 	    ("%s: Cannot cachespread-init a UMA_ZONE_PCPU keg", __func__));
1258 
1259 	alignsize = keg->uk_align + 1;
1260 	rsize = keg->uk_size;
1261 	/*
1262 	 * We want one item to start on every align boundary in a page.  To
1263 	 * do this we will span pages.  We will also extend the item by the
1264 	 * size of align if it is an even multiple of align.  Otherwise, it
1265 	 * would fall on the same boundary every time.
1266 	 */
1267 	if (rsize & keg->uk_align)
1268 		rsize = (rsize & ~keg->uk_align) + alignsize;
1269 	if ((rsize & alignsize) == 0)
1270 		rsize += alignsize;
1271 	trailer = rsize - keg->uk_size;
1272 	pages = (rsize * (PAGE_SIZE / alignsize)) / PAGE_SIZE;
1273 	pages = MIN(pages, (128 * 1024) / PAGE_SIZE);
1274 	keg->uk_rsize = rsize;
1275 	keg->uk_ppera = pages;
1276 	keg->uk_slabsize = UMA_SLAB_SIZE;
1277 	keg->uk_ipers = ((pages * PAGE_SIZE) + trailer) / rsize;
1278 	keg->uk_flags |= UMA_ZONE_OFFPAGE | UMA_ZONE_VTOSLAB;
1279 	KASSERT(keg->uk_ipers <= uma_max_ipers,
1280 	    ("%s: keg->uk_ipers too high(%d) increase max_ipers", __func__,
1281 	    keg->uk_ipers));
1282 }
1283 
1284 /*
1285  * Keg header ctor.  This initializes all fields, locks, etc.  And inserts
1286  * the keg onto the global keg list.
1287  *
1288  * Arguments/Returns follow uma_ctor specifications
1289  *	udata  Actually uma_kctor_args
1290  */
1291 static int
1292 keg_ctor(void *mem, int size, void *udata, int flags)
1293 {
1294 	struct uma_kctor_args *arg = udata;
1295 	uma_keg_t keg = mem;
1296 	uma_zone_t zone;
1297 
1298 	bzero(keg, size);
1299 	keg->uk_size = arg->size;
1300 	keg->uk_init = arg->uminit;
1301 	keg->uk_fini = arg->fini;
1302 	keg->uk_align = arg->align;
1303 	keg->uk_free = 0;
1304 	keg->uk_pages = 0;
1305 	keg->uk_flags = arg->flags;
1306 	keg->uk_allocf = page_alloc;
1307 	keg->uk_freef = page_free;
1308 	keg->uk_recurse = 0;
1309 	keg->uk_slabzone = NULL;
1310 
1311 	/*
1312 	 * The master zone is passed to us at keg-creation time.
1313 	 */
1314 	zone = arg->zone;
1315 	keg->uk_name = zone->uz_name;
1316 
1317 	if (arg->flags & UMA_ZONE_VM)
1318 		keg->uk_flags |= UMA_ZFLAG_CACHEONLY;
1319 
1320 	if (arg->flags & UMA_ZONE_ZINIT)
1321 		keg->uk_init = zero_init;
1322 
1323 	if (arg->flags & UMA_ZONE_REFCNT || arg->flags & UMA_ZONE_MALLOC)
1324 		keg->uk_flags |= UMA_ZONE_VTOSLAB;
1325 
1326 	if (arg->flags & UMA_ZONE_PCPU)
1327 #ifdef SMP
1328 		keg->uk_flags |= UMA_ZONE_OFFPAGE;
1329 #else
1330 		keg->uk_flags &= ~UMA_ZONE_PCPU;
1331 #endif
1332 
1333 	/*
1334 	 * The +UMA_FRITM_SZ added to uk_size is to account for the
1335 	 * linkage that is added to the size in keg_small_init().  If
1336 	 * we don't account for this here then we may end up in
1337 	 * keg_small_init() with a calculated 'ipers' of 0.
1338 	 */
1339 	if (keg->uk_flags & UMA_ZONE_REFCNT) {
1340 		if (keg->uk_flags & UMA_ZONE_CACHESPREAD)
1341 			keg_cachespread_init(keg);
1342 		else if ((keg->uk_size+UMA_FRITMREF_SZ) >
1343 		    (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt)))
1344 			keg_large_init(keg);
1345 		else
1346 			keg_small_init(keg);
1347 	} else {
1348 		if (keg->uk_flags & UMA_ZONE_CACHESPREAD)
1349 			keg_cachespread_init(keg);
1350 		else if ((keg->uk_size+UMA_FRITM_SZ) >
1351 		    (UMA_SLAB_SIZE - sizeof(struct uma_slab)))
1352 			keg_large_init(keg);
1353 		else
1354 			keg_small_init(keg);
1355 	}
1356 
1357 	if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
1358 		if (keg->uk_flags & UMA_ZONE_REFCNT)
1359 			keg->uk_slabzone = slabrefzone;
1360 		else
1361 			keg->uk_slabzone = slabzone;
1362 	}
1363 
1364 	/*
1365 	 * If we haven't booted yet we need allocations to go through the
1366 	 * startup cache until the vm is ready.
1367 	 */
1368 	if (keg->uk_ppera == 1) {
1369 #ifdef UMA_MD_SMALL_ALLOC
1370 		keg->uk_allocf = uma_small_alloc;
1371 		keg->uk_freef = uma_small_free;
1372 
1373 		if (booted < UMA_STARTUP)
1374 			keg->uk_allocf = startup_alloc;
1375 #else
1376 		if (booted < UMA_STARTUP2)
1377 			keg->uk_allocf = startup_alloc;
1378 #endif
1379 	} else if (booted < UMA_STARTUP2 &&
1380 	    (keg->uk_flags & UMA_ZFLAG_INTERNAL))
1381 		keg->uk_allocf = startup_alloc;
1382 
1383 	/*
1384 	 * Initialize keg's lock (shared among zones).
1385 	 */
1386 	if (arg->flags & UMA_ZONE_MTXCLASS)
1387 		KEG_LOCK_INIT(keg, 1);
1388 	else
1389 		KEG_LOCK_INIT(keg, 0);
1390 
1391 	/*
1392 	 * If we're putting the slab header in the actual page we need to
1393 	 * figure out where in each page it goes.  This calculates a right
1394 	 * justified offset into the memory on an ALIGN_PTR boundary.
1395 	 */
1396 	if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) {
1397 		u_int totsize;
1398 
1399 		/* Size of the slab struct and free list */
1400 		if (keg->uk_flags & UMA_ZONE_REFCNT)
1401 			totsize = sizeof(struct uma_slab_refcnt) +
1402 			    keg->uk_ipers * UMA_FRITMREF_SZ;
1403 		else
1404 			totsize = sizeof(struct uma_slab) +
1405 			    keg->uk_ipers * UMA_FRITM_SZ;
1406 
1407 		if (totsize & UMA_ALIGN_PTR)
1408 			totsize = (totsize & ~UMA_ALIGN_PTR) +
1409 			    (UMA_ALIGN_PTR + 1);
1410 		keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - totsize;
1411 
1412 		if (keg->uk_flags & UMA_ZONE_REFCNT)
1413 			totsize = keg->uk_pgoff + sizeof(struct uma_slab_refcnt)
1414 			    + keg->uk_ipers * UMA_FRITMREF_SZ;
1415 		else
1416 			totsize = keg->uk_pgoff + sizeof(struct uma_slab)
1417 			    + keg->uk_ipers * UMA_FRITM_SZ;
1418 
1419 		/*
1420 		 * The only way the following is possible is if with our
1421 		 * UMA_ALIGN_PTR adjustments we are now bigger than
1422 		 * UMA_SLAB_SIZE.  I haven't checked whether this is
1423 		 * mathematically possible for all cases, so we make
1424 		 * sure here anyway.
1425 		 */
1426 		if (totsize > PAGE_SIZE * keg->uk_ppera) {
1427 			printf("zone %s ipers %d rsize %d size %d\n",
1428 			    zone->uz_name, keg->uk_ipers, keg->uk_rsize,
1429 			    keg->uk_size);
1430 			panic("UMA slab won't fit.");
1431 		}
1432 	}
1433 
1434 	if (keg->uk_flags & UMA_ZONE_HASH)
1435 		hash_alloc(&keg->uk_hash);
1436 
1437 #ifdef UMA_DEBUG
1438 	printf("UMA: %s(%p) size %d(%d) flags %#x ipers %d ppera %d out %d free %d\n",
1439 	    zone->uz_name, zone, keg->uk_size, keg->uk_rsize, keg->uk_flags,
1440 	    keg->uk_ipers, keg->uk_ppera,
1441 	    (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free);
1442 #endif
1443 
1444 	LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
1445 
1446 	mtx_lock(&uma_mtx);
1447 	LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
1448 	mtx_unlock(&uma_mtx);
1449 	return (0);
1450 }
1451 
1452 /*
1453  * Zone header ctor.  This initializes all fields, locks, etc.
1454  *
1455  * Arguments/Returns follow uma_ctor specifications
1456  *	udata  Actually uma_zctor_args
1457  */
1458 static int
1459 zone_ctor(void *mem, int size, void *udata, int flags)
1460 {
1461 	struct uma_zctor_args *arg = udata;
1462 	uma_zone_t zone = mem;
1463 	uma_zone_t z;
1464 	uma_keg_t keg;
1465 
1466 	bzero(zone, size);
1467 	zone->uz_name = arg->name;
1468 	zone->uz_ctor = arg->ctor;
1469 	zone->uz_dtor = arg->dtor;
1470 	zone->uz_slab = zone_fetch_slab;
1471 	zone->uz_init = NULL;
1472 	zone->uz_fini = NULL;
1473 	zone->uz_allocs = 0;
1474 	zone->uz_frees = 0;
1475 	zone->uz_fails = 0;
1476 	zone->uz_sleeps = 0;
1477 	zone->uz_fills = zone->uz_count = 0;
1478 	zone->uz_flags = 0;
1479 	zone->uz_warning = NULL;
1480 	timevalclear(&zone->uz_ratecheck);
1481 	keg = arg->keg;
1482 
1483 	if (arg->flags & UMA_ZONE_SECONDARY) {
1484 		KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
1485 		zone->uz_init = arg->uminit;
1486 		zone->uz_fini = arg->fini;
1487 		zone->uz_lock = &keg->uk_lock;
1488 		zone->uz_flags |= UMA_ZONE_SECONDARY;
1489 		mtx_lock(&uma_mtx);
1490 		ZONE_LOCK(zone);
1491 		LIST_FOREACH(z, &keg->uk_zones, uz_link) {
1492 			if (LIST_NEXT(z, uz_link) == NULL) {
1493 				LIST_INSERT_AFTER(z, zone, uz_link);
1494 				break;
1495 			}
1496 		}
1497 		ZONE_UNLOCK(zone);
1498 		mtx_unlock(&uma_mtx);
1499 	} else if (keg == NULL) {
1500 		if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
1501 		    arg->align, arg->flags)) == NULL)
1502 			return (ENOMEM);
1503 	} else {
1504 		struct uma_kctor_args karg;
1505 		int error;
1506 
1507 		/* We should only be here from uma_startup() */
1508 		karg.size = arg->size;
1509 		karg.uminit = arg->uminit;
1510 		karg.fini = arg->fini;
1511 		karg.align = arg->align;
1512 		karg.flags = arg->flags;
1513 		karg.zone = zone;
1514 		error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
1515 		    flags);
1516 		if (error)
1517 			return (error);
1518 	}
1519 	/*
1520 	 * Link in the first keg.
1521 	 */
1522 	zone->uz_klink.kl_keg = keg;
1523 	LIST_INSERT_HEAD(&zone->uz_kegs, &zone->uz_klink, kl_link);
1524 	zone->uz_lock = &keg->uk_lock;
1525 	zone->uz_size = keg->uk_size;
1526 	zone->uz_flags |= (keg->uk_flags &
1527 	    (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
1528 
1529 	/*
1530 	 * Some internal zones don't have room allocated for the per cpu
1531 	 * caches.  If we're internal, bail out here.
1532 	 */
1533 	if (keg->uk_flags & UMA_ZFLAG_INTERNAL) {
1534 		KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
1535 		    ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
1536 		return (0);
1537 	}
1538 
1539 	if (keg->uk_flags & UMA_ZONE_MAXBUCKET)
1540 		zone->uz_count = BUCKET_MAX;
1541 	else if (keg->uk_ipers <= BUCKET_MAX)
1542 		zone->uz_count = keg->uk_ipers;
1543 	else
1544 		zone->uz_count = BUCKET_MAX;
1545 	return (0);
1546 }
1547 
1548 /*
1549  * Keg header dtor.  This frees all data, destroys locks, frees the hash
1550  * table and removes the keg from the global list.
1551  *
1552  * Arguments/Returns follow uma_dtor specifications
1553  *	udata  unused
1554  */
1555 static void
1556 keg_dtor(void *arg, int size, void *udata)
1557 {
1558 	uma_keg_t keg;
1559 
1560 	keg = (uma_keg_t)arg;
1561 	KEG_LOCK(keg);
1562 	if (keg->uk_free != 0) {
1563 		printf("Freed UMA keg was not empty (%d items). "
1564 		    " Lost %d pages of memory.\n",
1565 		    keg->uk_free, keg->uk_pages);
1566 	}
1567 	KEG_UNLOCK(keg);
1568 
1569 	hash_free(&keg->uk_hash);
1570 
1571 	KEG_LOCK_FINI(keg);
1572 }
1573 
1574 /*
1575  * Zone header dtor.
1576  *
1577  * Arguments/Returns follow uma_dtor specifications
1578  *	udata  unused
1579  */
1580 static void
1581 zone_dtor(void *arg, int size, void *udata)
1582 {
1583 	uma_klink_t klink;
1584 	uma_zone_t zone;
1585 	uma_keg_t keg;
1586 
1587 	zone = (uma_zone_t)arg;
1588 	keg = zone_first_keg(zone);
1589 
1590 	if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
1591 		cache_drain(zone);
1592 
1593 	mtx_lock(&uma_mtx);
1594 	LIST_REMOVE(zone, uz_link);
1595 	mtx_unlock(&uma_mtx);
1596 	/*
1597 	 * XXX there are some races here where
1598 	 * the zone can be drained but zone lock
1599 	 * released and then refilled before we
1600 	 * remove it... we dont care for now
1601 	 */
1602 	zone_drain_wait(zone, M_WAITOK);
1603 	/*
1604 	 * Unlink all of our kegs.
1605 	 */
1606 	while ((klink = LIST_FIRST(&zone->uz_kegs)) != NULL) {
1607 		klink->kl_keg = NULL;
1608 		LIST_REMOVE(klink, kl_link);
1609 		if (klink == &zone->uz_klink)
1610 			continue;
1611 		free(klink, M_TEMP);
1612 	}
1613 	/*
1614 	 * We only destroy kegs from non secondary zones.
1615 	 */
1616 	if ((zone->uz_flags & UMA_ZONE_SECONDARY) == 0)  {
1617 		mtx_lock(&uma_mtx);
1618 		LIST_REMOVE(keg, uk_link);
1619 		mtx_unlock(&uma_mtx);
1620 		zone_free_item(kegs, keg, NULL, SKIP_NONE,
1621 		    ZFREE_STATFREE);
1622 	}
1623 }
1624 
1625 /*
1626  * Traverses every zone in the system and calls a callback
1627  *
1628  * Arguments:
1629  *	zfunc  A pointer to a function which accepts a zone
1630  *		as an argument.
1631  *
1632  * Returns:
1633  *	Nothing
1634  */
1635 static void
1636 zone_foreach(void (*zfunc)(uma_zone_t))
1637 {
1638 	uma_keg_t keg;
1639 	uma_zone_t zone;
1640 
1641 	mtx_lock(&uma_mtx);
1642 	LIST_FOREACH(keg, &uma_kegs, uk_link) {
1643 		LIST_FOREACH(zone, &keg->uk_zones, uz_link)
1644 			zfunc(zone);
1645 	}
1646 	mtx_unlock(&uma_mtx);
1647 }
1648 
1649 /* Public functions */
1650 /* See uma.h */
1651 void
1652 uma_startup(void *bootmem, int boot_pages)
1653 {
1654 	struct uma_zctor_args args;
1655 	uma_slab_t slab;
1656 	u_int slabsize;
1657 	u_int objsize, totsize, wsize;
1658 	int i;
1659 
1660 #ifdef UMA_DEBUG
1661 	printf("Creating uma keg headers zone and keg.\n");
1662 #endif
1663 	mtx_init(&uma_mtx, "UMA lock", NULL, MTX_DEF);
1664 
1665 	/*
1666 	 * Figure out the maximum number of items-per-slab we'll have if
1667 	 * we're using the OFFPAGE slab header to track free items, given
1668 	 * all possible object sizes and the maximum desired wastage
1669 	 * (UMA_MAX_WASTE).
1670 	 *
1671 	 * We iterate until we find an object size for
1672 	 * which the calculated wastage in keg_small_init() will be
1673 	 * enough to warrant OFFPAGE.  Since wastedspace versus objsize
1674 	 * is an overall increasing see-saw function, we find the smallest
1675 	 * objsize such that the wastage is always acceptable for objects
1676 	 * with that objsize or smaller.  Since a smaller objsize always
1677 	 * generates a larger possible uma_max_ipers, we use this computed
1678 	 * objsize to calculate the largest ipers possible.  Since the
1679 	 * ipers calculated for OFFPAGE slab headers is always larger than
1680 	 * the ipers initially calculated in keg_small_init(), we use
1681 	 * the former's equation (UMA_SLAB_SIZE / keg->uk_rsize) to
1682 	 * obtain the maximum ipers possible for offpage slab headers.
1683 	 *
1684 	 * It should be noted that ipers versus objsize is an inversly
1685 	 * proportional function which drops off rather quickly so as
1686 	 * long as our UMA_MAX_WASTE is such that the objsize we calculate
1687 	 * falls into the portion of the inverse relation AFTER the steep
1688 	 * falloff, then uma_max_ipers shouldn't be too high (~10 on i386).
1689 	 *
1690 	 * Note that we have 8-bits (1 byte) to use as a freelist index
1691 	 * inside the actual slab header itself and this is enough to
1692 	 * accomodate us.  In the worst case, a UMA_SMALLEST_UNIT sized
1693 	 * object with offpage slab header would have ipers =
1694 	 * UMA_SLAB_SIZE / UMA_SMALLEST_UNIT (currently = 256), which is
1695 	 * 1 greater than what our byte-integer freelist index can
1696 	 * accomodate, but we know that this situation never occurs as
1697 	 * for UMA_SMALLEST_UNIT-sized objects, we will never calculate
1698 	 * that we need to go to offpage slab headers.  Or, if we do,
1699 	 * then we trap that condition below and panic in the INVARIANTS case.
1700 	 */
1701 	wsize = UMA_SLAB_SIZE - sizeof(struct uma_slab) -
1702 	    (UMA_SLAB_SIZE / UMA_MAX_WASTE);
1703 	totsize = wsize;
1704 	objsize = UMA_SMALLEST_UNIT;
1705 	while (totsize >= wsize) {
1706 		totsize = (UMA_SLAB_SIZE - sizeof(struct uma_slab)) /
1707 		    (objsize + UMA_FRITM_SZ);
1708 		totsize *= (UMA_FRITM_SZ + objsize);
1709 		objsize++;
1710 	}
1711 	if (objsize > UMA_SMALLEST_UNIT)
1712 		objsize--;
1713 	uma_max_ipers = MAX(UMA_SLAB_SIZE / objsize, 64);
1714 
1715 	wsize = UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt) -
1716 	    (UMA_SLAB_SIZE / UMA_MAX_WASTE);
1717 	totsize = wsize;
1718 	objsize = UMA_SMALLEST_UNIT;
1719 	while (totsize >= wsize) {
1720 		totsize = (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt)) /
1721 		    (objsize + UMA_FRITMREF_SZ);
1722 		totsize *= (UMA_FRITMREF_SZ + objsize);
1723 		objsize++;
1724 	}
1725 	if (objsize > UMA_SMALLEST_UNIT)
1726 		objsize--;
1727 	uma_max_ipers_ref = MAX(UMA_SLAB_SIZE / objsize, 64);
1728 
1729 	KASSERT((uma_max_ipers_ref <= 256) && (uma_max_ipers <= 256),
1730 	    ("uma_startup: calculated uma_max_ipers values too large!"));
1731 
1732 #ifdef UMA_DEBUG
1733 	printf("Calculated uma_max_ipers (for OFFPAGE) is %d\n", uma_max_ipers);
1734 	printf("Calculated uma_max_ipers_ref (for OFFPAGE) is %d\n",
1735 	    uma_max_ipers_ref);
1736 #endif
1737 
1738 	/* "manually" create the initial zone */
1739 	args.name = "UMA Kegs";
1740 	args.size = sizeof(struct uma_keg);
1741 	args.ctor = keg_ctor;
1742 	args.dtor = keg_dtor;
1743 	args.uminit = zero_init;
1744 	args.fini = NULL;
1745 	args.keg = &masterkeg;
1746 	args.align = 32 - 1;
1747 	args.flags = UMA_ZFLAG_INTERNAL;
1748 	/* The initial zone has no Per cpu queues so it's smaller */
1749 	zone_ctor(kegs, sizeof(struct uma_zone), &args, M_WAITOK);
1750 
1751 #ifdef UMA_DEBUG
1752 	printf("Filling boot free list.\n");
1753 #endif
1754 	for (i = 0; i < boot_pages; i++) {
1755 		slab = (uma_slab_t)((uint8_t *)bootmem + (i * UMA_SLAB_SIZE));
1756 		slab->us_data = (uint8_t *)slab;
1757 		slab->us_flags = UMA_SLAB_BOOT;
1758 		LIST_INSERT_HEAD(&uma_boot_pages, slab, us_link);
1759 	}
1760 	mtx_init(&uma_boot_pages_mtx, "UMA boot pages", NULL, MTX_DEF);
1761 
1762 #ifdef UMA_DEBUG
1763 	printf("Creating uma zone headers zone and keg.\n");
1764 #endif
1765 	args.name = "UMA Zones";
1766 	args.size = sizeof(struct uma_zone) +
1767 	    (sizeof(struct uma_cache) * (mp_maxid + 1));
1768 	args.ctor = zone_ctor;
1769 	args.dtor = zone_dtor;
1770 	args.uminit = zero_init;
1771 	args.fini = NULL;
1772 	args.keg = NULL;
1773 	args.align = 32 - 1;
1774 	args.flags = UMA_ZFLAG_INTERNAL;
1775 	/* The initial zone has no Per cpu queues so it's smaller */
1776 	zone_ctor(zones, sizeof(struct uma_zone), &args, M_WAITOK);
1777 
1778 #ifdef UMA_DEBUG
1779 	printf("Initializing pcpu cache locks.\n");
1780 #endif
1781 #ifdef UMA_DEBUG
1782 	printf("Creating slab and hash zones.\n");
1783 #endif
1784 
1785 	/*
1786 	 * This is the max number of free list items we'll have with
1787 	 * offpage slabs.
1788 	 */
1789 	slabsize = uma_max_ipers * UMA_FRITM_SZ;
1790 	slabsize += sizeof(struct uma_slab);
1791 
1792 	/* Now make a zone for slab headers */
1793 	slabzone = uma_zcreate("UMA Slabs",
1794 				slabsize,
1795 				NULL, NULL, NULL, NULL,
1796 				UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1797 
1798 	/*
1799 	 * We also create a zone for the bigger slabs with reference
1800 	 * counts in them, to accomodate UMA_ZONE_REFCNT zones.
1801 	 */
1802 	slabsize = uma_max_ipers_ref * UMA_FRITMREF_SZ;
1803 	slabsize += sizeof(struct uma_slab_refcnt);
1804 	slabrefzone = uma_zcreate("UMA RCntSlabs",
1805 				  slabsize,
1806 				  NULL, NULL, NULL, NULL,
1807 				  UMA_ALIGN_PTR,
1808 				  UMA_ZFLAG_INTERNAL);
1809 
1810 	hashzone = uma_zcreate("UMA Hash",
1811 	    sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
1812 	    NULL, NULL, NULL, NULL,
1813 	    UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1814 
1815 	bucket_init();
1816 
1817 	booted = UMA_STARTUP;
1818 
1819 #ifdef UMA_DEBUG
1820 	printf("UMA startup complete.\n");
1821 #endif
1822 }
1823 
1824 /* see uma.h */
1825 void
1826 uma_startup2(void)
1827 {
1828 	booted = UMA_STARTUP2;
1829 	bucket_enable();
1830 #ifdef UMA_DEBUG
1831 	printf("UMA startup2 complete.\n");
1832 #endif
1833 }
1834 
1835 /*
1836  * Initialize our callout handle
1837  *
1838  */
1839 
1840 static void
1841 uma_startup3(void)
1842 {
1843 #ifdef UMA_DEBUG
1844 	printf("Starting callout.\n");
1845 #endif
1846 	callout_init(&uma_callout, CALLOUT_MPSAFE);
1847 	callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
1848 #ifdef UMA_DEBUG
1849 	printf("UMA startup3 complete.\n");
1850 #endif
1851 }
1852 
1853 static uma_keg_t
1854 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
1855 		int align, uint32_t flags)
1856 {
1857 	struct uma_kctor_args args;
1858 
1859 	args.size = size;
1860 	args.uminit = uminit;
1861 	args.fini = fini;
1862 	args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
1863 	args.flags = flags;
1864 	args.zone = zone;
1865 	return (zone_alloc_item(kegs, &args, M_WAITOK));
1866 }
1867 
1868 /* See uma.h */
1869 void
1870 uma_set_align(int align)
1871 {
1872 
1873 	if (align != UMA_ALIGN_CACHE)
1874 		uma_align_cache = align;
1875 }
1876 
1877 /* See uma.h */
1878 uma_zone_t
1879 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
1880 		uma_init uminit, uma_fini fini, int align, uint32_t flags)
1881 
1882 {
1883 	struct uma_zctor_args args;
1884 
1885 	/* This stuff is essential for the zone ctor */
1886 	args.name = name;
1887 	args.size = size;
1888 	args.ctor = ctor;
1889 	args.dtor = dtor;
1890 	args.uminit = uminit;
1891 	args.fini = fini;
1892 	args.align = align;
1893 	args.flags = flags;
1894 	args.keg = NULL;
1895 
1896 	return (zone_alloc_item(zones, &args, M_WAITOK));
1897 }
1898 
1899 /* See uma.h */
1900 uma_zone_t
1901 uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor,
1902 		    uma_init zinit, uma_fini zfini, uma_zone_t master)
1903 {
1904 	struct uma_zctor_args args;
1905 	uma_keg_t keg;
1906 
1907 	keg = zone_first_keg(master);
1908 	args.name = name;
1909 	args.size = keg->uk_size;
1910 	args.ctor = ctor;
1911 	args.dtor = dtor;
1912 	args.uminit = zinit;
1913 	args.fini = zfini;
1914 	args.align = keg->uk_align;
1915 	args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
1916 	args.keg = keg;
1917 
1918 	/* XXX Attaches only one keg of potentially many. */
1919 	return (zone_alloc_item(zones, &args, M_WAITOK));
1920 }
1921 
1922 static void
1923 zone_lock_pair(uma_zone_t a, uma_zone_t b)
1924 {
1925 	if (a < b) {
1926 		ZONE_LOCK(a);
1927 		mtx_lock_flags(b->uz_lock, MTX_DUPOK);
1928 	} else {
1929 		ZONE_LOCK(b);
1930 		mtx_lock_flags(a->uz_lock, MTX_DUPOK);
1931 	}
1932 }
1933 
1934 static void
1935 zone_unlock_pair(uma_zone_t a, uma_zone_t b)
1936 {
1937 
1938 	ZONE_UNLOCK(a);
1939 	ZONE_UNLOCK(b);
1940 }
1941 
1942 int
1943 uma_zsecond_add(uma_zone_t zone, uma_zone_t master)
1944 {
1945 	uma_klink_t klink;
1946 	uma_klink_t kl;
1947 	int error;
1948 
1949 	error = 0;
1950 	klink = malloc(sizeof(*klink), M_TEMP, M_WAITOK | M_ZERO);
1951 
1952 	zone_lock_pair(zone, master);
1953 	/*
1954 	 * zone must use vtoslab() to resolve objects and must already be
1955 	 * a secondary.
1956 	 */
1957 	if ((zone->uz_flags & (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY))
1958 	    != (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) {
1959 		error = EINVAL;
1960 		goto out;
1961 	}
1962 	/*
1963 	 * The new master must also use vtoslab().
1964 	 */
1965 	if ((zone->uz_flags & UMA_ZONE_VTOSLAB) != UMA_ZONE_VTOSLAB) {
1966 		error = EINVAL;
1967 		goto out;
1968 	}
1969 	/*
1970 	 * Both must either be refcnt, or not be refcnt.
1971 	 */
1972 	if ((zone->uz_flags & UMA_ZONE_REFCNT) !=
1973 	    (master->uz_flags & UMA_ZONE_REFCNT)) {
1974 		error = EINVAL;
1975 		goto out;
1976 	}
1977 	/*
1978 	 * The underlying object must be the same size.  rsize
1979 	 * may be different.
1980 	 */
1981 	if (master->uz_size != zone->uz_size) {
1982 		error = E2BIG;
1983 		goto out;
1984 	}
1985 	/*
1986 	 * Put it at the end of the list.
1987 	 */
1988 	klink->kl_keg = zone_first_keg(master);
1989 	LIST_FOREACH(kl, &zone->uz_kegs, kl_link) {
1990 		if (LIST_NEXT(kl, kl_link) == NULL) {
1991 			LIST_INSERT_AFTER(kl, klink, kl_link);
1992 			break;
1993 		}
1994 	}
1995 	klink = NULL;
1996 	zone->uz_flags |= UMA_ZFLAG_MULTI;
1997 	zone->uz_slab = zone_fetch_slab_multi;
1998 
1999 out:
2000 	zone_unlock_pair(zone, master);
2001 	if (klink != NULL)
2002 		free(klink, M_TEMP);
2003 
2004 	return (error);
2005 }
2006 
2007 
2008 /* See uma.h */
2009 void
2010 uma_zdestroy(uma_zone_t zone)
2011 {
2012 
2013 	zone_free_item(zones, zone, NULL, SKIP_NONE, ZFREE_STATFREE);
2014 }
2015 
2016 /* See uma.h */
2017 void *
2018 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
2019 {
2020 	void *item;
2021 	uma_cache_t cache;
2022 	uma_bucket_t bucket;
2023 	int cpu;
2024 
2025 	/* This is the fast path allocation */
2026 #ifdef UMA_DEBUG_ALLOC_1
2027 	printf("Allocating one item from %s(%p)\n", zone->uz_name, zone);
2028 #endif
2029 	CTR3(KTR_UMA, "uma_zalloc_arg thread %x zone %s flags %d", curthread,
2030 	    zone->uz_name, flags);
2031 
2032 	if (flags & M_WAITOK) {
2033 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
2034 		    "uma_zalloc_arg: zone \"%s\"", zone->uz_name);
2035 	}
2036 #ifdef DEBUG_MEMGUARD
2037 	if (memguard_cmp_zone(zone)) {
2038 		item = memguard_alloc(zone->uz_size, flags);
2039 		if (item != NULL) {
2040 			/*
2041 			 * Avoid conflict with the use-after-free
2042 			 * protecting infrastructure from INVARIANTS.
2043 			 */
2044 			if (zone->uz_init != NULL &&
2045 			    zone->uz_init != mtrash_init &&
2046 			    zone->uz_init(item, zone->uz_size, flags) != 0)
2047 				return (NULL);
2048 			if (zone->uz_ctor != NULL &&
2049 			    zone->uz_ctor != mtrash_ctor &&
2050 			    zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2051 			    	zone->uz_fini(item, zone->uz_size);
2052 				return (NULL);
2053 			}
2054 			return (item);
2055 		}
2056 		/* This is unfortunate but should not be fatal. */
2057 	}
2058 #endif
2059 	/*
2060 	 * If possible, allocate from the per-CPU cache.  There are two
2061 	 * requirements for safe access to the per-CPU cache: (1) the thread
2062 	 * accessing the cache must not be preempted or yield during access,
2063 	 * and (2) the thread must not migrate CPUs without switching which
2064 	 * cache it accesses.  We rely on a critical section to prevent
2065 	 * preemption and migration.  We release the critical section in
2066 	 * order to acquire the zone mutex if we are unable to allocate from
2067 	 * the current cache; when we re-acquire the critical section, we
2068 	 * must detect and handle migration if it has occurred.
2069 	 */
2070 zalloc_restart:
2071 	critical_enter();
2072 	cpu = curcpu;
2073 	cache = &zone->uz_cpu[cpu];
2074 
2075 zalloc_start:
2076 	bucket = cache->uc_allocbucket;
2077 
2078 	if (bucket) {
2079 		if (bucket->ub_cnt > 0) {
2080 			bucket->ub_cnt--;
2081 			item = bucket->ub_bucket[bucket->ub_cnt];
2082 #ifdef INVARIANTS
2083 			bucket->ub_bucket[bucket->ub_cnt] = NULL;
2084 #endif
2085 			KASSERT(item != NULL,
2086 			    ("uma_zalloc: Bucket pointer mangled."));
2087 			cache->uc_allocs++;
2088 			critical_exit();
2089 #ifdef INVARIANTS
2090 			ZONE_LOCK(zone);
2091 			uma_dbg_alloc(zone, NULL, item);
2092 			ZONE_UNLOCK(zone);
2093 #endif
2094 			if (zone->uz_ctor != NULL) {
2095 				if (zone->uz_ctor(item, zone->uz_size,
2096 				    udata, flags) != 0) {
2097 					zone_free_item(zone, item, udata,
2098 					    SKIP_DTOR, ZFREE_STATFAIL |
2099 					    ZFREE_STATFREE);
2100 					return (NULL);
2101 				}
2102 			}
2103 			if (flags & M_ZERO)
2104 				bzero(item, zone->uz_size);
2105 			return (item);
2106 		} else if (cache->uc_freebucket) {
2107 			/*
2108 			 * We have run out of items in our allocbucket.
2109 			 * See if we can switch with our free bucket.
2110 			 */
2111 			if (cache->uc_freebucket->ub_cnt > 0) {
2112 #ifdef UMA_DEBUG_ALLOC
2113 				printf("uma_zalloc: Swapping empty with"
2114 				    " alloc.\n");
2115 #endif
2116 				bucket = cache->uc_freebucket;
2117 				cache->uc_freebucket = cache->uc_allocbucket;
2118 				cache->uc_allocbucket = bucket;
2119 
2120 				goto zalloc_start;
2121 			}
2122 		}
2123 	}
2124 	/*
2125 	 * Attempt to retrieve the item from the per-CPU cache has failed, so
2126 	 * we must go back to the zone.  This requires the zone lock, so we
2127 	 * must drop the critical section, then re-acquire it when we go back
2128 	 * to the cache.  Since the critical section is released, we may be
2129 	 * preempted or migrate.  As such, make sure not to maintain any
2130 	 * thread-local state specific to the cache from prior to releasing
2131 	 * the critical section.
2132 	 */
2133 	critical_exit();
2134 	ZONE_LOCK(zone);
2135 	critical_enter();
2136 	cpu = curcpu;
2137 	cache = &zone->uz_cpu[cpu];
2138 	bucket = cache->uc_allocbucket;
2139 	if (bucket != NULL) {
2140 		if (bucket->ub_cnt > 0) {
2141 			ZONE_UNLOCK(zone);
2142 			goto zalloc_start;
2143 		}
2144 		bucket = cache->uc_freebucket;
2145 		if (bucket != NULL && bucket->ub_cnt > 0) {
2146 			ZONE_UNLOCK(zone);
2147 			goto zalloc_start;
2148 		}
2149 	}
2150 
2151 	/* Since we have locked the zone we may as well send back our stats */
2152 	zone->uz_allocs += cache->uc_allocs;
2153 	cache->uc_allocs = 0;
2154 	zone->uz_frees += cache->uc_frees;
2155 	cache->uc_frees = 0;
2156 
2157 	/* Our old one is now a free bucket */
2158 	if (cache->uc_allocbucket) {
2159 		KASSERT(cache->uc_allocbucket->ub_cnt == 0,
2160 		    ("uma_zalloc_arg: Freeing a non free bucket."));
2161 		LIST_INSERT_HEAD(&zone->uz_free_bucket,
2162 		    cache->uc_allocbucket, ub_link);
2163 		cache->uc_allocbucket = NULL;
2164 	}
2165 
2166 	/* Check the free list for a new alloc bucket */
2167 	if ((bucket = LIST_FIRST(&zone->uz_full_bucket)) != NULL) {
2168 		KASSERT(bucket->ub_cnt != 0,
2169 		    ("uma_zalloc_arg: Returning an empty bucket."));
2170 
2171 		LIST_REMOVE(bucket, ub_link);
2172 		cache->uc_allocbucket = bucket;
2173 		ZONE_UNLOCK(zone);
2174 		goto zalloc_start;
2175 	}
2176 	/* We are no longer associated with this CPU. */
2177 	critical_exit();
2178 
2179 	/* Bump up our uz_count so we get here less */
2180 	if (zone->uz_count < BUCKET_MAX)
2181 		zone->uz_count++;
2182 
2183 	/*
2184 	 * Now lets just fill a bucket and put it on the free list.  If that
2185 	 * works we'll restart the allocation from the begining.
2186 	 */
2187 	if (zone_alloc_bucket(zone, flags)) {
2188 		ZONE_UNLOCK(zone);
2189 		goto zalloc_restart;
2190 	}
2191 	ZONE_UNLOCK(zone);
2192 	/*
2193 	 * We may not be able to get a bucket so return an actual item.
2194 	 */
2195 #ifdef UMA_DEBUG
2196 	printf("uma_zalloc_arg: Bucketzone returned NULL\n");
2197 #endif
2198 
2199 	item = zone_alloc_item(zone, udata, flags);
2200 	return (item);
2201 }
2202 
2203 static uma_slab_t
2204 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int flags)
2205 {
2206 	uma_slab_t slab;
2207 
2208 	mtx_assert(&keg->uk_lock, MA_OWNED);
2209 	slab = NULL;
2210 
2211 	for (;;) {
2212 		/*
2213 		 * Find a slab with some space.  Prefer slabs that are partially
2214 		 * used over those that are totally full.  This helps to reduce
2215 		 * fragmentation.
2216 		 */
2217 		if (keg->uk_free != 0) {
2218 			if (!LIST_EMPTY(&keg->uk_part_slab)) {
2219 				slab = LIST_FIRST(&keg->uk_part_slab);
2220 			} else {
2221 				slab = LIST_FIRST(&keg->uk_free_slab);
2222 				LIST_REMOVE(slab, us_link);
2223 				LIST_INSERT_HEAD(&keg->uk_part_slab, slab,
2224 				    us_link);
2225 			}
2226 			MPASS(slab->us_keg == keg);
2227 			return (slab);
2228 		}
2229 
2230 		/*
2231 		 * M_NOVM means don't ask at all!
2232 		 */
2233 		if (flags & M_NOVM)
2234 			break;
2235 
2236 		if (keg->uk_maxpages && keg->uk_pages >= keg->uk_maxpages) {
2237 			keg->uk_flags |= UMA_ZFLAG_FULL;
2238 			/*
2239 			 * If this is not a multi-zone, set the FULL bit.
2240 			 * Otherwise slab_multi() takes care of it.
2241 			 */
2242 			if ((zone->uz_flags & UMA_ZFLAG_MULTI) == 0) {
2243 				zone->uz_flags |= UMA_ZFLAG_FULL;
2244 				zone_log_warning(zone);
2245 			}
2246 			if (flags & M_NOWAIT)
2247 				break;
2248 			zone->uz_sleeps++;
2249 			msleep(keg, &keg->uk_lock, PVM, "keglimit", 0);
2250 			continue;
2251 		}
2252 		keg->uk_recurse++;
2253 		slab = keg_alloc_slab(keg, zone, flags);
2254 		keg->uk_recurse--;
2255 		/*
2256 		 * If we got a slab here it's safe to mark it partially used
2257 		 * and return.  We assume that the caller is going to remove
2258 		 * at least one item.
2259 		 */
2260 		if (slab) {
2261 			MPASS(slab->us_keg == keg);
2262 			LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2263 			return (slab);
2264 		}
2265 		/*
2266 		 * We might not have been able to get a slab but another cpu
2267 		 * could have while we were unlocked.  Check again before we
2268 		 * fail.
2269 		 */
2270 		flags |= M_NOVM;
2271 	}
2272 	return (slab);
2273 }
2274 
2275 static inline void
2276 zone_relock(uma_zone_t zone, uma_keg_t keg)
2277 {
2278 	if (zone->uz_lock != &keg->uk_lock) {
2279 		KEG_UNLOCK(keg);
2280 		ZONE_LOCK(zone);
2281 	}
2282 }
2283 
2284 static inline void
2285 keg_relock(uma_keg_t keg, uma_zone_t zone)
2286 {
2287 	if (zone->uz_lock != &keg->uk_lock) {
2288 		ZONE_UNLOCK(zone);
2289 		KEG_LOCK(keg);
2290 	}
2291 }
2292 
2293 static uma_slab_t
2294 zone_fetch_slab(uma_zone_t zone, uma_keg_t keg, int flags)
2295 {
2296 	uma_slab_t slab;
2297 
2298 	if (keg == NULL)
2299 		keg = zone_first_keg(zone);
2300 	/*
2301 	 * This is to prevent us from recursively trying to allocate
2302 	 * buckets.  The problem is that if an allocation forces us to
2303 	 * grab a new bucket we will call page_alloc, which will go off
2304 	 * and cause the vm to allocate vm_map_entries.  If we need new
2305 	 * buckets there too we will recurse in kmem_alloc and bad
2306 	 * things happen.  So instead we return a NULL bucket, and make
2307 	 * the code that allocates buckets smart enough to deal with it
2308 	 */
2309 	if (keg->uk_flags & UMA_ZFLAG_BUCKET && keg->uk_recurse != 0)
2310 		return (NULL);
2311 
2312 	for (;;) {
2313 		slab = keg_fetch_slab(keg, zone, flags);
2314 		if (slab)
2315 			return (slab);
2316 		if (flags & (M_NOWAIT | M_NOVM))
2317 			break;
2318 	}
2319 	return (NULL);
2320 }
2321 
2322 /*
2323  * uma_zone_fetch_slab_multi:  Fetches a slab from one available keg.  Returns
2324  * with the keg locked.  Caller must call zone_relock() afterwards if the
2325  * zone lock is required.  On NULL the zone lock is held.
2326  *
2327  * The last pointer is used to seed the search.  It is not required.
2328  */
2329 static uma_slab_t
2330 zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int rflags)
2331 {
2332 	uma_klink_t klink;
2333 	uma_slab_t slab;
2334 	uma_keg_t keg;
2335 	int flags;
2336 	int empty;
2337 	int full;
2338 
2339 	/*
2340 	 * Don't wait on the first pass.  This will skip limit tests
2341 	 * as well.  We don't want to block if we can find a provider
2342 	 * without blocking.
2343 	 */
2344 	flags = (rflags & ~M_WAITOK) | M_NOWAIT;
2345 	/*
2346 	 * Use the last slab allocated as a hint for where to start
2347 	 * the search.
2348 	 */
2349 	if (last) {
2350 		slab = keg_fetch_slab(last, zone, flags);
2351 		if (slab)
2352 			return (slab);
2353 		zone_relock(zone, last);
2354 		last = NULL;
2355 	}
2356 	/*
2357 	 * Loop until we have a slab incase of transient failures
2358 	 * while M_WAITOK is specified.  I'm not sure this is 100%
2359 	 * required but we've done it for so long now.
2360 	 */
2361 	for (;;) {
2362 		empty = 0;
2363 		full = 0;
2364 		/*
2365 		 * Search the available kegs for slabs.  Be careful to hold the
2366 		 * correct lock while calling into the keg layer.
2367 		 */
2368 		LIST_FOREACH(klink, &zone->uz_kegs, kl_link) {
2369 			keg = klink->kl_keg;
2370 			keg_relock(keg, zone);
2371 			if ((keg->uk_flags & UMA_ZFLAG_FULL) == 0) {
2372 				slab = keg_fetch_slab(keg, zone, flags);
2373 				if (slab)
2374 					return (slab);
2375 			}
2376 			if (keg->uk_flags & UMA_ZFLAG_FULL)
2377 				full++;
2378 			else
2379 				empty++;
2380 			zone_relock(zone, keg);
2381 		}
2382 		if (rflags & (M_NOWAIT | M_NOVM))
2383 			break;
2384 		flags = rflags;
2385 		/*
2386 		 * All kegs are full.  XXX We can't atomically check all kegs
2387 		 * and sleep so just sleep for a short period and retry.
2388 		 */
2389 		if (full && !empty) {
2390 			zone->uz_flags |= UMA_ZFLAG_FULL;
2391 			zone->uz_sleeps++;
2392 			zone_log_warning(zone);
2393 			msleep(zone, zone->uz_lock, PVM, "zonelimit", hz/100);
2394 			zone->uz_flags &= ~UMA_ZFLAG_FULL;
2395 			continue;
2396 		}
2397 	}
2398 	return (NULL);
2399 }
2400 
2401 static void *
2402 slab_alloc_item(uma_zone_t zone, uma_slab_t slab)
2403 {
2404 	uma_keg_t keg;
2405 	uma_slabrefcnt_t slabref;
2406 	void *item;
2407 	uint8_t freei;
2408 
2409 	keg = slab->us_keg;
2410 	mtx_assert(&keg->uk_lock, MA_OWNED);
2411 
2412 	freei = slab->us_firstfree;
2413 	if (keg->uk_flags & UMA_ZONE_REFCNT) {
2414 		slabref = (uma_slabrefcnt_t)slab;
2415 		slab->us_firstfree = slabref->us_freelist[freei].us_item;
2416 	} else {
2417 		slab->us_firstfree = slab->us_freelist[freei].us_item;
2418 	}
2419 	item = slab->us_data + (keg->uk_rsize * freei);
2420 
2421 	slab->us_freecount--;
2422 	keg->uk_free--;
2423 #ifdef INVARIANTS
2424 	uma_dbg_alloc(zone, slab, item);
2425 #endif
2426 	/* Move this slab to the full list */
2427 	if (slab->us_freecount == 0) {
2428 		LIST_REMOVE(slab, us_link);
2429 		LIST_INSERT_HEAD(&keg->uk_full_slab, slab, us_link);
2430 	}
2431 
2432 	return (item);
2433 }
2434 
2435 static int
2436 zone_alloc_bucket(uma_zone_t zone, int flags)
2437 {
2438 	uma_bucket_t bucket;
2439 	uma_slab_t slab;
2440 	uma_keg_t keg;
2441 	int16_t saved;
2442 	int max, origflags = flags;
2443 
2444 	/*
2445 	 * Try this zone's free list first so we don't allocate extra buckets.
2446 	 */
2447 	if ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) {
2448 		KASSERT(bucket->ub_cnt == 0,
2449 		    ("zone_alloc_bucket: Bucket on free list is not empty."));
2450 		LIST_REMOVE(bucket, ub_link);
2451 	} else {
2452 		int bflags;
2453 
2454 		bflags = (flags & ~M_ZERO);
2455 		if (zone->uz_flags & UMA_ZFLAG_CACHEONLY)
2456 			bflags |= M_NOVM;
2457 
2458 		ZONE_UNLOCK(zone);
2459 		bucket = bucket_alloc(zone->uz_count, bflags);
2460 		ZONE_LOCK(zone);
2461 	}
2462 
2463 	if (bucket == NULL) {
2464 		return (0);
2465 	}
2466 
2467 #ifdef SMP
2468 	/*
2469 	 * This code is here to limit the number of simultaneous bucket fills
2470 	 * for any given zone to the number of per cpu caches in this zone. This
2471 	 * is done so that we don't allocate more memory than we really need.
2472 	 */
2473 	if (zone->uz_fills >= mp_ncpus)
2474 		goto done;
2475 
2476 #endif
2477 	zone->uz_fills++;
2478 
2479 	max = MIN(bucket->ub_entries, zone->uz_count);
2480 	/* Try to keep the buckets totally full */
2481 	saved = bucket->ub_cnt;
2482 	slab = NULL;
2483 	keg = NULL;
2484 	while (bucket->ub_cnt < max &&
2485 	    (slab = zone->uz_slab(zone, keg, flags)) != NULL) {
2486 		keg = slab->us_keg;
2487 		while (slab->us_freecount && bucket->ub_cnt < max) {
2488 			bucket->ub_bucket[bucket->ub_cnt++] =
2489 			    slab_alloc_item(zone, slab);
2490 		}
2491 
2492 		/* Don't block on the next fill */
2493 		flags |= M_NOWAIT;
2494 	}
2495 	if (slab)
2496 		zone_relock(zone, keg);
2497 
2498 	/*
2499 	 * We unlock here because we need to call the zone's init.
2500 	 * It should be safe to unlock because the slab dealt with
2501 	 * above is already on the appropriate list within the keg
2502 	 * and the bucket we filled is not yet on any list, so we
2503 	 * own it.
2504 	 */
2505 	if (zone->uz_init != NULL) {
2506 		int i;
2507 
2508 		ZONE_UNLOCK(zone);
2509 		for (i = saved; i < bucket->ub_cnt; i++)
2510 			if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size,
2511 			    origflags) != 0)
2512 				break;
2513 		/*
2514 		 * If we couldn't initialize the whole bucket, put the
2515 		 * rest back onto the freelist.
2516 		 */
2517 		if (i != bucket->ub_cnt) {
2518 			int j;
2519 
2520 			for (j = i; j < bucket->ub_cnt; j++) {
2521 				zone_free_item(zone, bucket->ub_bucket[j],
2522 				    NULL, SKIP_FINI, 0);
2523 #ifdef INVARIANTS
2524 				bucket->ub_bucket[j] = NULL;
2525 #endif
2526 			}
2527 			bucket->ub_cnt = i;
2528 		}
2529 		ZONE_LOCK(zone);
2530 	}
2531 
2532 	zone->uz_fills--;
2533 	if (bucket->ub_cnt != 0) {
2534 		LIST_INSERT_HEAD(&zone->uz_full_bucket,
2535 		    bucket, ub_link);
2536 		return (1);
2537 	}
2538 #ifdef SMP
2539 done:
2540 #endif
2541 	bucket_free(bucket);
2542 
2543 	return (0);
2544 }
2545 /*
2546  * Allocates an item for an internal zone
2547  *
2548  * Arguments
2549  *	zone   The zone to alloc for.
2550  *	udata  The data to be passed to the constructor.
2551  *	flags  M_WAITOK, M_NOWAIT, M_ZERO.
2552  *
2553  * Returns
2554  *	NULL if there is no memory and M_NOWAIT is set
2555  *	An item if successful
2556  */
2557 
2558 static void *
2559 zone_alloc_item(uma_zone_t zone, void *udata, int flags)
2560 {
2561 	uma_slab_t slab;
2562 	void *item;
2563 
2564 	item = NULL;
2565 
2566 #ifdef UMA_DEBUG_ALLOC
2567 	printf("INTERNAL: Allocating one item from %s(%p)\n", zone->uz_name, zone);
2568 #endif
2569 	ZONE_LOCK(zone);
2570 
2571 	slab = zone->uz_slab(zone, NULL, flags);
2572 	if (slab == NULL) {
2573 		zone->uz_fails++;
2574 		ZONE_UNLOCK(zone);
2575 		return (NULL);
2576 	}
2577 
2578 	item = slab_alloc_item(zone, slab);
2579 
2580 	zone_relock(zone, slab->us_keg);
2581 	zone->uz_allocs++;
2582 	ZONE_UNLOCK(zone);
2583 
2584 	/*
2585 	 * We have to call both the zone's init (not the keg's init)
2586 	 * and the zone's ctor.  This is because the item is going from
2587 	 * a keg slab directly to the user, and the user is expecting it
2588 	 * to be both zone-init'd as well as zone-ctor'd.
2589 	 */
2590 	if (zone->uz_init != NULL) {
2591 		if (zone->uz_init(item, zone->uz_size, flags) != 0) {
2592 			zone_free_item(zone, item, udata, SKIP_FINI,
2593 			    ZFREE_STATFAIL | ZFREE_STATFREE);
2594 			return (NULL);
2595 		}
2596 	}
2597 	if (zone->uz_ctor != NULL) {
2598 		if (zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2599 			zone_free_item(zone, item, udata, SKIP_DTOR,
2600 			    ZFREE_STATFAIL | ZFREE_STATFREE);
2601 			return (NULL);
2602 		}
2603 	}
2604 	if (flags & M_ZERO)
2605 		bzero(item, zone->uz_size);
2606 
2607 	return (item);
2608 }
2609 
2610 /* See uma.h */
2611 void
2612 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
2613 {
2614 	uma_cache_t cache;
2615 	uma_bucket_t bucket;
2616 	int bflags;
2617 	int cpu;
2618 
2619 #ifdef UMA_DEBUG_ALLOC_1
2620 	printf("Freeing item %p to %s(%p)\n", item, zone->uz_name, zone);
2621 #endif
2622 	CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread,
2623 	    zone->uz_name);
2624 
2625         /* uma_zfree(..., NULL) does nothing, to match free(9). */
2626         if (item == NULL)
2627                 return;
2628 #ifdef DEBUG_MEMGUARD
2629 	if (is_memguard_addr(item)) {
2630 		if (zone->uz_dtor != NULL && zone->uz_dtor != mtrash_dtor)
2631 			zone->uz_dtor(item, zone->uz_size, udata);
2632 		if (zone->uz_fini != NULL && zone->uz_fini != mtrash_fini)
2633 			zone->uz_fini(item, zone->uz_size);
2634 		memguard_free(item);
2635 		return;
2636 	}
2637 #endif
2638 	if (zone->uz_dtor)
2639 		zone->uz_dtor(item, zone->uz_size, udata);
2640 
2641 #ifdef INVARIANTS
2642 	ZONE_LOCK(zone);
2643 	if (zone->uz_flags & UMA_ZONE_MALLOC)
2644 		uma_dbg_free(zone, udata, item);
2645 	else
2646 		uma_dbg_free(zone, NULL, item);
2647 	ZONE_UNLOCK(zone);
2648 #endif
2649 	/*
2650 	 * The race here is acceptable.  If we miss it we'll just have to wait
2651 	 * a little longer for the limits to be reset.
2652 	 */
2653 	if (zone->uz_flags & UMA_ZFLAG_FULL)
2654 		goto zfree_internal;
2655 
2656 	/*
2657 	 * If possible, free to the per-CPU cache.  There are two
2658 	 * requirements for safe access to the per-CPU cache: (1) the thread
2659 	 * accessing the cache must not be preempted or yield during access,
2660 	 * and (2) the thread must not migrate CPUs without switching which
2661 	 * cache it accesses.  We rely on a critical section to prevent
2662 	 * preemption and migration.  We release the critical section in
2663 	 * order to acquire the zone mutex if we are unable to free to the
2664 	 * current cache; when we re-acquire the critical section, we must
2665 	 * detect and handle migration if it has occurred.
2666 	 */
2667 zfree_restart:
2668 	critical_enter();
2669 	cpu = curcpu;
2670 	cache = &zone->uz_cpu[cpu];
2671 
2672 zfree_start:
2673 	bucket = cache->uc_freebucket;
2674 
2675 	if (bucket) {
2676 		/*
2677 		 * Do we have room in our bucket? It is OK for this uz count
2678 		 * check to be slightly out of sync.
2679 		 */
2680 
2681 		if (bucket->ub_cnt < bucket->ub_entries) {
2682 			KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL,
2683 			    ("uma_zfree: Freeing to non free bucket index."));
2684 			bucket->ub_bucket[bucket->ub_cnt] = item;
2685 			bucket->ub_cnt++;
2686 			cache->uc_frees++;
2687 			critical_exit();
2688 			return;
2689 		} else if (cache->uc_allocbucket) {
2690 #ifdef UMA_DEBUG_ALLOC
2691 			printf("uma_zfree: Swapping buckets.\n");
2692 #endif
2693 			/*
2694 			 * We have run out of space in our freebucket.
2695 			 * See if we can switch with our alloc bucket.
2696 			 */
2697 			if (cache->uc_allocbucket->ub_cnt <
2698 			    cache->uc_freebucket->ub_cnt) {
2699 				bucket = cache->uc_freebucket;
2700 				cache->uc_freebucket = cache->uc_allocbucket;
2701 				cache->uc_allocbucket = bucket;
2702 				goto zfree_start;
2703 			}
2704 		}
2705 	}
2706 	/*
2707 	 * We can get here for two reasons:
2708 	 *
2709 	 * 1) The buckets are NULL
2710 	 * 2) The alloc and free buckets are both somewhat full.
2711 	 *
2712 	 * We must go back the zone, which requires acquiring the zone lock,
2713 	 * which in turn means we must release and re-acquire the critical
2714 	 * section.  Since the critical section is released, we may be
2715 	 * preempted or migrate.  As such, make sure not to maintain any
2716 	 * thread-local state specific to the cache from prior to releasing
2717 	 * the critical section.
2718 	 */
2719 	critical_exit();
2720 	ZONE_LOCK(zone);
2721 	critical_enter();
2722 	cpu = curcpu;
2723 	cache = &zone->uz_cpu[cpu];
2724 	if (cache->uc_freebucket != NULL) {
2725 		if (cache->uc_freebucket->ub_cnt <
2726 		    cache->uc_freebucket->ub_entries) {
2727 			ZONE_UNLOCK(zone);
2728 			goto zfree_start;
2729 		}
2730 		if (cache->uc_allocbucket != NULL &&
2731 		    (cache->uc_allocbucket->ub_cnt <
2732 		    cache->uc_freebucket->ub_cnt)) {
2733 			ZONE_UNLOCK(zone);
2734 			goto zfree_start;
2735 		}
2736 	}
2737 
2738 	/* Since we have locked the zone we may as well send back our stats */
2739 	zone->uz_allocs += cache->uc_allocs;
2740 	cache->uc_allocs = 0;
2741 	zone->uz_frees += cache->uc_frees;
2742 	cache->uc_frees = 0;
2743 
2744 	bucket = cache->uc_freebucket;
2745 	cache->uc_freebucket = NULL;
2746 
2747 	/* Can we throw this on the zone full list? */
2748 	if (bucket != NULL) {
2749 #ifdef UMA_DEBUG_ALLOC
2750 		printf("uma_zfree: Putting old bucket on the free list.\n");
2751 #endif
2752 		/* ub_cnt is pointing to the last free item */
2753 		KASSERT(bucket->ub_cnt != 0,
2754 		    ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n"));
2755 		LIST_INSERT_HEAD(&zone->uz_full_bucket,
2756 		    bucket, ub_link);
2757 	}
2758 	if ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) {
2759 		LIST_REMOVE(bucket, ub_link);
2760 		ZONE_UNLOCK(zone);
2761 		cache->uc_freebucket = bucket;
2762 		goto zfree_start;
2763 	}
2764 	/* We are no longer associated with this CPU. */
2765 	critical_exit();
2766 
2767 	/* And the zone.. */
2768 	ZONE_UNLOCK(zone);
2769 
2770 #ifdef UMA_DEBUG_ALLOC
2771 	printf("uma_zfree: Allocating new free bucket.\n");
2772 #endif
2773 	bflags = M_NOWAIT;
2774 
2775 	if (zone->uz_flags & UMA_ZFLAG_CACHEONLY)
2776 		bflags |= M_NOVM;
2777 	bucket = bucket_alloc(zone->uz_count, bflags);
2778 	if (bucket) {
2779 		ZONE_LOCK(zone);
2780 		LIST_INSERT_HEAD(&zone->uz_free_bucket,
2781 		    bucket, ub_link);
2782 		ZONE_UNLOCK(zone);
2783 		goto zfree_restart;
2784 	}
2785 
2786 	/*
2787 	 * If nothing else caught this, we'll just do an internal free.
2788 	 */
2789 zfree_internal:
2790 	zone_free_item(zone, item, udata, SKIP_DTOR, ZFREE_STATFREE);
2791 
2792 	return;
2793 }
2794 
2795 /*
2796  * Frees an item to an INTERNAL zone or allocates a free bucket
2797  *
2798  * Arguments:
2799  *	zone   The zone to free to
2800  *	item   The item we're freeing
2801  *	udata  User supplied data for the dtor
2802  *	skip   Skip dtors and finis
2803  */
2804 static void
2805 zone_free_item(uma_zone_t zone, void *item, void *udata,
2806     enum zfreeskip skip, int flags)
2807 {
2808 	uma_slab_t slab;
2809 	uma_slabrefcnt_t slabref;
2810 	uma_keg_t keg;
2811 	uint8_t *mem;
2812 	uint8_t freei;
2813 	int clearfull;
2814 
2815 	if (skip < SKIP_DTOR && zone->uz_dtor)
2816 		zone->uz_dtor(item, zone->uz_size, udata);
2817 
2818 	if (skip < SKIP_FINI && zone->uz_fini)
2819 		zone->uz_fini(item, zone->uz_size);
2820 
2821 	ZONE_LOCK(zone);
2822 
2823 	if (flags & ZFREE_STATFAIL)
2824 		zone->uz_fails++;
2825 	if (flags & ZFREE_STATFREE)
2826 		zone->uz_frees++;
2827 
2828 	if (!(zone->uz_flags & UMA_ZONE_VTOSLAB)) {
2829 		mem = (uint8_t *)((unsigned long)item & (~UMA_SLAB_MASK));
2830 		keg = zone_first_keg(zone); /* Must only be one. */
2831 		if (zone->uz_flags & UMA_ZONE_HASH) {
2832 			slab = hash_sfind(&keg->uk_hash, mem);
2833 		} else {
2834 			mem += keg->uk_pgoff;
2835 			slab = (uma_slab_t)mem;
2836 		}
2837 	} else {
2838 		/* This prevents redundant lookups via free(). */
2839 		if ((zone->uz_flags & UMA_ZONE_MALLOC) && udata != NULL)
2840 			slab = (uma_slab_t)udata;
2841 		else
2842 			slab = vtoslab((vm_offset_t)item);
2843 		keg = slab->us_keg;
2844 		keg_relock(keg, zone);
2845 	}
2846 	MPASS(keg == slab->us_keg);
2847 
2848 	/* Do we need to remove from any lists? */
2849 	if (slab->us_freecount+1 == keg->uk_ipers) {
2850 		LIST_REMOVE(slab, us_link);
2851 		LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
2852 	} else if (slab->us_freecount == 0) {
2853 		LIST_REMOVE(slab, us_link);
2854 		LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2855 	}
2856 
2857 	/* Slab management stuff */
2858 	freei = ((unsigned long)item - (unsigned long)slab->us_data)
2859 		/ keg->uk_rsize;
2860 
2861 #ifdef INVARIANTS
2862 	if (!skip)
2863 		uma_dbg_free(zone, slab, item);
2864 #endif
2865 
2866 	if (keg->uk_flags & UMA_ZONE_REFCNT) {
2867 		slabref = (uma_slabrefcnt_t)slab;
2868 		slabref->us_freelist[freei].us_item = slab->us_firstfree;
2869 	} else {
2870 		slab->us_freelist[freei].us_item = slab->us_firstfree;
2871 	}
2872 	slab->us_firstfree = freei;
2873 	slab->us_freecount++;
2874 
2875 	/* Zone statistics */
2876 	keg->uk_free++;
2877 
2878 	clearfull = 0;
2879 	if (keg->uk_flags & UMA_ZFLAG_FULL) {
2880 		if (keg->uk_pages < keg->uk_maxpages) {
2881 			keg->uk_flags &= ~UMA_ZFLAG_FULL;
2882 			clearfull = 1;
2883 		}
2884 
2885 		/*
2886 		 * We can handle one more allocation. Since we're clearing ZFLAG_FULL,
2887 		 * wake up all procs blocked on pages. This should be uncommon, so
2888 		 * keeping this simple for now (rather than adding count of blocked
2889 		 * threads etc).
2890 		 */
2891 		wakeup(keg);
2892 	}
2893 	if (clearfull) {
2894 		zone_relock(zone, keg);
2895 		zone->uz_flags &= ~UMA_ZFLAG_FULL;
2896 		wakeup(zone);
2897 		ZONE_UNLOCK(zone);
2898 	} else
2899 		KEG_UNLOCK(keg);
2900 }
2901 
2902 /* See uma.h */
2903 int
2904 uma_zone_set_max(uma_zone_t zone, int nitems)
2905 {
2906 	uma_keg_t keg;
2907 
2908 	ZONE_LOCK(zone);
2909 	keg = zone_first_keg(zone);
2910 	keg->uk_maxpages = (nitems / keg->uk_ipers) * keg->uk_ppera;
2911 	if (keg->uk_maxpages * keg->uk_ipers < nitems)
2912 		keg->uk_maxpages += keg->uk_ppera;
2913 	nitems = keg->uk_maxpages * keg->uk_ipers;
2914 	ZONE_UNLOCK(zone);
2915 
2916 	return (nitems);
2917 }
2918 
2919 /* See uma.h */
2920 int
2921 uma_zone_get_max(uma_zone_t zone)
2922 {
2923 	int nitems;
2924 	uma_keg_t keg;
2925 
2926 	ZONE_LOCK(zone);
2927 	keg = zone_first_keg(zone);
2928 	nitems = keg->uk_maxpages * keg->uk_ipers;
2929 	ZONE_UNLOCK(zone);
2930 
2931 	return (nitems);
2932 }
2933 
2934 /* See uma.h */
2935 void
2936 uma_zone_set_warning(uma_zone_t zone, const char *warning)
2937 {
2938 
2939 	ZONE_LOCK(zone);
2940 	zone->uz_warning = warning;
2941 	ZONE_UNLOCK(zone);
2942 }
2943 
2944 /* See uma.h */
2945 int
2946 uma_zone_get_cur(uma_zone_t zone)
2947 {
2948 	int64_t nitems;
2949 	u_int i;
2950 
2951 	ZONE_LOCK(zone);
2952 	nitems = zone->uz_allocs - zone->uz_frees;
2953 	CPU_FOREACH(i) {
2954 		/*
2955 		 * See the comment in sysctl_vm_zone_stats() regarding the
2956 		 * safety of accessing the per-cpu caches. With the zone lock
2957 		 * held, it is safe, but can potentially result in stale data.
2958 		 */
2959 		nitems += zone->uz_cpu[i].uc_allocs -
2960 		    zone->uz_cpu[i].uc_frees;
2961 	}
2962 	ZONE_UNLOCK(zone);
2963 
2964 	return (nitems < 0 ? 0 : nitems);
2965 }
2966 
2967 /* See uma.h */
2968 void
2969 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
2970 {
2971 	uma_keg_t keg;
2972 
2973 	ZONE_LOCK(zone);
2974 	keg = zone_first_keg(zone);
2975 	KASSERT(keg->uk_pages == 0,
2976 	    ("uma_zone_set_init on non-empty keg"));
2977 	keg->uk_init = uminit;
2978 	ZONE_UNLOCK(zone);
2979 }
2980 
2981 /* See uma.h */
2982 void
2983 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
2984 {
2985 	uma_keg_t keg;
2986 
2987 	ZONE_LOCK(zone);
2988 	keg = zone_first_keg(zone);
2989 	KASSERT(keg->uk_pages == 0,
2990 	    ("uma_zone_set_fini on non-empty keg"));
2991 	keg->uk_fini = fini;
2992 	ZONE_UNLOCK(zone);
2993 }
2994 
2995 /* See uma.h */
2996 void
2997 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
2998 {
2999 	ZONE_LOCK(zone);
3000 	KASSERT(zone_first_keg(zone)->uk_pages == 0,
3001 	    ("uma_zone_set_zinit on non-empty keg"));
3002 	zone->uz_init = zinit;
3003 	ZONE_UNLOCK(zone);
3004 }
3005 
3006 /* See uma.h */
3007 void
3008 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
3009 {
3010 	ZONE_LOCK(zone);
3011 	KASSERT(zone_first_keg(zone)->uk_pages == 0,
3012 	    ("uma_zone_set_zfini on non-empty keg"));
3013 	zone->uz_fini = zfini;
3014 	ZONE_UNLOCK(zone);
3015 }
3016 
3017 /* See uma.h */
3018 /* XXX uk_freef is not actually used with the zone locked */
3019 void
3020 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
3021 {
3022 
3023 	ZONE_LOCK(zone);
3024 	zone_first_keg(zone)->uk_freef = freef;
3025 	ZONE_UNLOCK(zone);
3026 }
3027 
3028 /* See uma.h */
3029 /* XXX uk_allocf is not actually used with the zone locked */
3030 void
3031 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
3032 {
3033 	uma_keg_t keg;
3034 
3035 	ZONE_LOCK(zone);
3036 	keg = zone_first_keg(zone);
3037 	keg->uk_flags |= UMA_ZFLAG_PRIVALLOC;
3038 	keg->uk_allocf = allocf;
3039 	ZONE_UNLOCK(zone);
3040 }
3041 
3042 /* See uma.h */
3043 int
3044 uma_zone_reserve_kva(uma_zone_t zone, int count)
3045 {
3046 	uma_keg_t keg;
3047 	vm_offset_t kva;
3048 	int pages;
3049 
3050 	keg = zone_first_keg(zone);
3051 	pages = count / keg->uk_ipers;
3052 
3053 	if (pages * keg->uk_ipers < count)
3054 		pages++;
3055 
3056 #ifdef UMA_MD_SMALL_ALLOC
3057 	if (keg->uk_ppera > 1) {
3058 #else
3059 	if (1) {
3060 #endif
3061 		kva = kmem_alloc_nofault(kernel_map, pages * UMA_SLAB_SIZE);
3062 		if (kva == 0)
3063 			return (0);
3064 	} else
3065 		kva = 0;
3066 	ZONE_LOCK(zone);
3067 	keg->uk_kva = kva;
3068 	keg->uk_offset = 0;
3069 	keg->uk_maxpages = pages;
3070 #ifdef UMA_MD_SMALL_ALLOC
3071 	keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
3072 #else
3073 	keg->uk_allocf = noobj_alloc;
3074 #endif
3075 	keg->uk_flags |= UMA_ZONE_NOFREE | UMA_ZFLAG_PRIVALLOC;
3076 	ZONE_UNLOCK(zone);
3077 	return (1);
3078 }
3079 
3080 /* See uma.h */
3081 void
3082 uma_prealloc(uma_zone_t zone, int items)
3083 {
3084 	int slabs;
3085 	uma_slab_t slab;
3086 	uma_keg_t keg;
3087 
3088 	keg = zone_first_keg(zone);
3089 	ZONE_LOCK(zone);
3090 	slabs = items / keg->uk_ipers;
3091 	if (slabs * keg->uk_ipers < items)
3092 		slabs++;
3093 	while (slabs > 0) {
3094 		slab = keg_alloc_slab(keg, zone, M_WAITOK);
3095 		if (slab == NULL)
3096 			break;
3097 		MPASS(slab->us_keg == keg);
3098 		LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
3099 		slabs--;
3100 	}
3101 	ZONE_UNLOCK(zone);
3102 }
3103 
3104 /* See uma.h */
3105 uint32_t *
3106 uma_find_refcnt(uma_zone_t zone, void *item)
3107 {
3108 	uma_slabrefcnt_t slabref;
3109 	uma_keg_t keg;
3110 	uint32_t *refcnt;
3111 	int idx;
3112 
3113 	slabref = (uma_slabrefcnt_t)vtoslab((vm_offset_t)item &
3114 	    (~UMA_SLAB_MASK));
3115 	keg = slabref->us_keg;
3116 	KASSERT(slabref != NULL && slabref->us_keg->uk_flags & UMA_ZONE_REFCNT,
3117 	    ("uma_find_refcnt(): zone possibly not UMA_ZONE_REFCNT"));
3118 	idx = ((unsigned long)item - (unsigned long)slabref->us_data)
3119 	    / keg->uk_rsize;
3120 	refcnt = &slabref->us_freelist[idx].us_refcnt;
3121 	return refcnt;
3122 }
3123 
3124 /* See uma.h */
3125 void
3126 uma_reclaim(void)
3127 {
3128 #ifdef UMA_DEBUG
3129 	printf("UMA: vm asked us to release pages!\n");
3130 #endif
3131 	bucket_enable();
3132 	zone_foreach(zone_drain);
3133 	/*
3134 	 * Some slabs may have been freed but this zone will be visited early
3135 	 * we visit again so that we can free pages that are empty once other
3136 	 * zones are drained.  We have to do the same for buckets.
3137 	 */
3138 	zone_drain(slabzone);
3139 	zone_drain(slabrefzone);
3140 	bucket_zone_drain();
3141 }
3142 
3143 /* See uma.h */
3144 int
3145 uma_zone_exhausted(uma_zone_t zone)
3146 {
3147 	int full;
3148 
3149 	ZONE_LOCK(zone);
3150 	full = (zone->uz_flags & UMA_ZFLAG_FULL);
3151 	ZONE_UNLOCK(zone);
3152 	return (full);
3153 }
3154 
3155 int
3156 uma_zone_exhausted_nolock(uma_zone_t zone)
3157 {
3158 	return (zone->uz_flags & UMA_ZFLAG_FULL);
3159 }
3160 
3161 void *
3162 uma_large_malloc(int size, int wait)
3163 {
3164 	void *mem;
3165 	uma_slab_t slab;
3166 	uint8_t flags;
3167 
3168 	slab = zone_alloc_item(slabzone, NULL, wait);
3169 	if (slab == NULL)
3170 		return (NULL);
3171 	mem = page_alloc(NULL, size, &flags, wait);
3172 	if (mem) {
3173 		vsetslab((vm_offset_t)mem, slab);
3174 		slab->us_data = mem;
3175 		slab->us_flags = flags | UMA_SLAB_MALLOC;
3176 		slab->us_size = size;
3177 	} else {
3178 		zone_free_item(slabzone, slab, NULL, SKIP_NONE,
3179 		    ZFREE_STATFAIL | ZFREE_STATFREE);
3180 	}
3181 
3182 	return (mem);
3183 }
3184 
3185 void
3186 uma_large_free(uma_slab_t slab)
3187 {
3188 	vsetobj((vm_offset_t)slab->us_data, kmem_object);
3189 	page_free(slab->us_data, slab->us_size, slab->us_flags);
3190 	zone_free_item(slabzone, slab, NULL, SKIP_NONE, ZFREE_STATFREE);
3191 }
3192 
3193 void
3194 uma_print_stats(void)
3195 {
3196 	zone_foreach(uma_print_zone);
3197 }
3198 
3199 static void
3200 slab_print(uma_slab_t slab)
3201 {
3202 	printf("slab: keg %p, data %p, freecount %d, firstfree %d\n",
3203 		slab->us_keg, slab->us_data, slab->us_freecount,
3204 		slab->us_firstfree);
3205 }
3206 
3207 static void
3208 cache_print(uma_cache_t cache)
3209 {
3210 	printf("alloc: %p(%d), free: %p(%d)\n",
3211 		cache->uc_allocbucket,
3212 		cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0,
3213 		cache->uc_freebucket,
3214 		cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0);
3215 }
3216 
3217 static void
3218 uma_print_keg(uma_keg_t keg)
3219 {
3220 	uma_slab_t slab;
3221 
3222 	printf("keg: %s(%p) size %d(%d) flags %#x ipers %d ppera %d "
3223 	    "out %d free %d limit %d\n",
3224 	    keg->uk_name, keg, keg->uk_size, keg->uk_rsize, keg->uk_flags,
3225 	    keg->uk_ipers, keg->uk_ppera,
3226 	    (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free,
3227 	    (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers);
3228 	printf("Part slabs:\n");
3229 	LIST_FOREACH(slab, &keg->uk_part_slab, us_link)
3230 		slab_print(slab);
3231 	printf("Free slabs:\n");
3232 	LIST_FOREACH(slab, &keg->uk_free_slab, us_link)
3233 		slab_print(slab);
3234 	printf("Full slabs:\n");
3235 	LIST_FOREACH(slab, &keg->uk_full_slab, us_link)
3236 		slab_print(slab);
3237 }
3238 
3239 void
3240 uma_print_zone(uma_zone_t zone)
3241 {
3242 	uma_cache_t cache;
3243 	uma_klink_t kl;
3244 	int i;
3245 
3246 	printf("zone: %s(%p) size %d flags %#x\n",
3247 	    zone->uz_name, zone, zone->uz_size, zone->uz_flags);
3248 	LIST_FOREACH(kl, &zone->uz_kegs, kl_link)
3249 		uma_print_keg(kl->kl_keg);
3250 	CPU_FOREACH(i) {
3251 		cache = &zone->uz_cpu[i];
3252 		printf("CPU %d Cache:\n", i);
3253 		cache_print(cache);
3254 	}
3255 }
3256 
3257 #ifdef DDB
3258 /*
3259  * Generate statistics across both the zone and its per-cpu cache's.  Return
3260  * desired statistics if the pointer is non-NULL for that statistic.
3261  *
3262  * Note: does not update the zone statistics, as it can't safely clear the
3263  * per-CPU cache statistic.
3264  *
3265  * XXXRW: Following the uc_allocbucket and uc_freebucket pointers here isn't
3266  * safe from off-CPU; we should modify the caches to track this information
3267  * directly so that we don't have to.
3268  */
3269 static void
3270 uma_zone_sumstat(uma_zone_t z, int *cachefreep, uint64_t *allocsp,
3271     uint64_t *freesp, uint64_t *sleepsp)
3272 {
3273 	uma_cache_t cache;
3274 	uint64_t allocs, frees, sleeps;
3275 	int cachefree, cpu;
3276 
3277 	allocs = frees = sleeps = 0;
3278 	cachefree = 0;
3279 	CPU_FOREACH(cpu) {
3280 		cache = &z->uz_cpu[cpu];
3281 		if (cache->uc_allocbucket != NULL)
3282 			cachefree += cache->uc_allocbucket->ub_cnt;
3283 		if (cache->uc_freebucket != NULL)
3284 			cachefree += cache->uc_freebucket->ub_cnt;
3285 		allocs += cache->uc_allocs;
3286 		frees += cache->uc_frees;
3287 	}
3288 	allocs += z->uz_allocs;
3289 	frees += z->uz_frees;
3290 	sleeps += z->uz_sleeps;
3291 	if (cachefreep != NULL)
3292 		*cachefreep = cachefree;
3293 	if (allocsp != NULL)
3294 		*allocsp = allocs;
3295 	if (freesp != NULL)
3296 		*freesp = frees;
3297 	if (sleepsp != NULL)
3298 		*sleepsp = sleeps;
3299 }
3300 #endif /* DDB */
3301 
3302 static int
3303 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
3304 {
3305 	uma_keg_t kz;
3306 	uma_zone_t z;
3307 	int count;
3308 
3309 	count = 0;
3310 	mtx_lock(&uma_mtx);
3311 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3312 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
3313 			count++;
3314 	}
3315 	mtx_unlock(&uma_mtx);
3316 	return (sysctl_handle_int(oidp, &count, 0, req));
3317 }
3318 
3319 static int
3320 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
3321 {
3322 	struct uma_stream_header ush;
3323 	struct uma_type_header uth;
3324 	struct uma_percpu_stat ups;
3325 	uma_bucket_t bucket;
3326 	struct sbuf sbuf;
3327 	uma_cache_t cache;
3328 	uma_klink_t kl;
3329 	uma_keg_t kz;
3330 	uma_zone_t z;
3331 	uma_keg_t k;
3332 	int count, error, i;
3333 
3334 	error = sysctl_wire_old_buffer(req, 0);
3335 	if (error != 0)
3336 		return (error);
3337 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
3338 
3339 	count = 0;
3340 	mtx_lock(&uma_mtx);
3341 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3342 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
3343 			count++;
3344 	}
3345 
3346 	/*
3347 	 * Insert stream header.
3348 	 */
3349 	bzero(&ush, sizeof(ush));
3350 	ush.ush_version = UMA_STREAM_VERSION;
3351 	ush.ush_maxcpus = (mp_maxid + 1);
3352 	ush.ush_count = count;
3353 	(void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
3354 
3355 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3356 		LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3357 			bzero(&uth, sizeof(uth));
3358 			ZONE_LOCK(z);
3359 			strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
3360 			uth.uth_align = kz->uk_align;
3361 			uth.uth_size = kz->uk_size;
3362 			uth.uth_rsize = kz->uk_rsize;
3363 			LIST_FOREACH(kl, &z->uz_kegs, kl_link) {
3364 				k = kl->kl_keg;
3365 				uth.uth_maxpages += k->uk_maxpages;
3366 				uth.uth_pages += k->uk_pages;
3367 				uth.uth_keg_free += k->uk_free;
3368 				uth.uth_limit = (k->uk_maxpages / k->uk_ppera)
3369 				    * k->uk_ipers;
3370 			}
3371 
3372 			/*
3373 			 * A zone is secondary is it is not the first entry
3374 			 * on the keg's zone list.
3375 			 */
3376 			if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
3377 			    (LIST_FIRST(&kz->uk_zones) != z))
3378 				uth.uth_zone_flags = UTH_ZONE_SECONDARY;
3379 
3380 			LIST_FOREACH(bucket, &z->uz_full_bucket, ub_link)
3381 				uth.uth_zone_free += bucket->ub_cnt;
3382 			uth.uth_allocs = z->uz_allocs;
3383 			uth.uth_frees = z->uz_frees;
3384 			uth.uth_fails = z->uz_fails;
3385 			uth.uth_sleeps = z->uz_sleeps;
3386 			(void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
3387 			/*
3388 			 * While it is not normally safe to access the cache
3389 			 * bucket pointers while not on the CPU that owns the
3390 			 * cache, we only allow the pointers to be exchanged
3391 			 * without the zone lock held, not invalidated, so
3392 			 * accept the possible race associated with bucket
3393 			 * exchange during monitoring.
3394 			 */
3395 			for (i = 0; i < (mp_maxid + 1); i++) {
3396 				bzero(&ups, sizeof(ups));
3397 				if (kz->uk_flags & UMA_ZFLAG_INTERNAL)
3398 					goto skip;
3399 				if (CPU_ABSENT(i))
3400 					goto skip;
3401 				cache = &z->uz_cpu[i];
3402 				if (cache->uc_allocbucket != NULL)
3403 					ups.ups_cache_free +=
3404 					    cache->uc_allocbucket->ub_cnt;
3405 				if (cache->uc_freebucket != NULL)
3406 					ups.ups_cache_free +=
3407 					    cache->uc_freebucket->ub_cnt;
3408 				ups.ups_allocs = cache->uc_allocs;
3409 				ups.ups_frees = cache->uc_frees;
3410 skip:
3411 				(void)sbuf_bcat(&sbuf, &ups, sizeof(ups));
3412 			}
3413 			ZONE_UNLOCK(z);
3414 		}
3415 	}
3416 	mtx_unlock(&uma_mtx);
3417 	error = sbuf_finish(&sbuf);
3418 	sbuf_delete(&sbuf);
3419 	return (error);
3420 }
3421 
3422 #ifdef DDB
3423 DB_SHOW_COMMAND(uma, db_show_uma)
3424 {
3425 	uint64_t allocs, frees, sleeps;
3426 	uma_bucket_t bucket;
3427 	uma_keg_t kz;
3428 	uma_zone_t z;
3429 	int cachefree;
3430 
3431 	db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
3432 	    "Requests", "Sleeps");
3433 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3434 		LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3435 			if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
3436 				allocs = z->uz_allocs;
3437 				frees = z->uz_frees;
3438 				sleeps = z->uz_sleeps;
3439 				cachefree = 0;
3440 			} else
3441 				uma_zone_sumstat(z, &cachefree, &allocs,
3442 				    &frees, &sleeps);
3443 			if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
3444 			    (LIST_FIRST(&kz->uk_zones) != z)))
3445 				cachefree += kz->uk_free;
3446 			LIST_FOREACH(bucket, &z->uz_full_bucket, ub_link)
3447 				cachefree += bucket->ub_cnt;
3448 			db_printf("%18s %8ju %8jd %8d %12ju %8ju\n", z->uz_name,
3449 			    (uintmax_t)kz->uk_size,
3450 			    (intmax_t)(allocs - frees), cachefree,
3451 			    (uintmax_t)allocs, sleeps);
3452 			if (db_pager_quit)
3453 				return;
3454 		}
3455 	}
3456 }
3457 #endif
3458