xref: /freebsd/sys/vm/uma_core.c (revision 81966bce06dac45f42bda62b14dba0756ef28505)
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/sbuf.h>
76 #include <sys/smp.h>
77 #include <sys/vmmeter.h>
78 
79 #include <vm/vm.h>
80 #include <vm/vm_object.h>
81 #include <vm/vm_page.h>
82 #include <vm/vm_param.h>
83 #include <vm/vm_map.h>
84 #include <vm/vm_kern.h>
85 #include <vm/vm_extern.h>
86 #include <vm/uma.h>
87 #include <vm/uma_int.h>
88 #include <vm/uma_dbg.h>
89 
90 #include <ddb/ddb.h>
91 
92 #ifdef DEBUG_MEMGUARD
93 #include <vm/memguard.h>
94 #endif
95 
96 /*
97  * This is the zone and keg from which all zones are spawned.  The idea is that
98  * even the zone & keg heads are allocated from the allocator, so we use the
99  * bss section to bootstrap us.
100  */
101 static struct uma_keg masterkeg;
102 static struct uma_zone masterzone_k;
103 static struct uma_zone masterzone_z;
104 static uma_zone_t kegs = &masterzone_k;
105 static uma_zone_t zones = &masterzone_z;
106 
107 /* This is the zone from which all of uma_slab_t's are allocated. */
108 static uma_zone_t slabzone;
109 static uma_zone_t slabrefzone;	/* With refcounters (for UMA_ZONE_REFCNT) */
110 
111 /*
112  * The initial hash tables come out of this zone so they can be allocated
113  * prior to malloc coming up.
114  */
115 static uma_zone_t hashzone;
116 
117 /* The boot-time adjusted value for cache line alignment. */
118 int uma_align_cache = 64 - 1;
119 
120 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets");
121 
122 /*
123  * Are we allowed to allocate buckets?
124  */
125 static int bucketdisable = 1;
126 
127 /* Linked list of all kegs in the system */
128 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs);
129 
130 /* This mutex protects the keg list */
131 static struct mtx uma_mtx;
132 
133 /* Linked list of boot time pages */
134 static LIST_HEAD(,uma_slab) uma_boot_pages =
135     LIST_HEAD_INITIALIZER(uma_boot_pages);
136 
137 /* This mutex protects the boot time pages list */
138 static struct mtx uma_boot_pages_mtx;
139 
140 /* Is the VM done starting up? */
141 static int booted = 0;
142 #define	UMA_STARTUP	1
143 #define	UMA_STARTUP2	2
144 
145 /* Maximum number of allowed items-per-slab if the slab header is OFFPAGE */
146 static u_int uma_max_ipers;
147 static u_int uma_max_ipers_ref;
148 
149 /*
150  * This is the handle used to schedule events that need to happen
151  * outside of the allocation fast path.
152  */
153 static struct callout uma_callout;
154 #define	UMA_TIMEOUT	20		/* Seconds for callout interval. */
155 
156 /*
157  * This structure is passed as the zone ctor arg so that I don't have to create
158  * a special allocation function just for zones.
159  */
160 struct uma_zctor_args {
161 	char *name;
162 	size_t size;
163 	uma_ctor ctor;
164 	uma_dtor dtor;
165 	uma_init uminit;
166 	uma_fini fini;
167 	uma_keg_t keg;
168 	int align;
169 	u_int32_t flags;
170 };
171 
172 struct uma_kctor_args {
173 	uma_zone_t zone;
174 	size_t size;
175 	uma_init uminit;
176 	uma_fini fini;
177 	int align;
178 	u_int32_t flags;
179 };
180 
181 struct uma_bucket_zone {
182 	uma_zone_t	ubz_zone;
183 	char		*ubz_name;
184 	int		ubz_entries;
185 };
186 
187 #define	BUCKET_MAX	128
188 
189 struct uma_bucket_zone bucket_zones[] = {
190 	{ NULL, "16 Bucket", 16 },
191 	{ NULL, "32 Bucket", 32 },
192 	{ NULL, "64 Bucket", 64 },
193 	{ NULL, "128 Bucket", 128 },
194 	{ NULL, NULL, 0}
195 };
196 
197 #define	BUCKET_SHIFT	4
198 #define	BUCKET_ZONES	((BUCKET_MAX >> BUCKET_SHIFT) + 1)
199 
200 /*
201  * bucket_size[] maps requested bucket sizes to zones that allocate a bucket
202  * of approximately the right size.
203  */
204 static uint8_t bucket_size[BUCKET_ZONES];
205 
206 /*
207  * Flags and enumerations to be passed to internal functions.
208  */
209 enum zfreeskip { SKIP_NONE, SKIP_DTOR, SKIP_FINI };
210 
211 #define	ZFREE_STATFAIL	0x00000001	/* Update zone failure statistic. */
212 #define	ZFREE_STATFREE	0x00000002	/* Update zone free statistic. */
213 
214 /* Prototypes.. */
215 
216 static void *obj_alloc(uma_zone_t, int, u_int8_t *, int);
217 static void *page_alloc(uma_zone_t, int, u_int8_t *, int);
218 static void *startup_alloc(uma_zone_t, int, u_int8_t *, int);
219 static void page_free(void *, int, u_int8_t);
220 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int);
221 static void cache_drain(uma_zone_t);
222 static void bucket_drain(uma_zone_t, uma_bucket_t);
223 static void bucket_cache_drain(uma_zone_t zone);
224 static int keg_ctor(void *, int, void *, int);
225 static void keg_dtor(void *, int, void *);
226 static int zone_ctor(void *, int, void *, int);
227 static void zone_dtor(void *, int, void *);
228 static int zero_init(void *, int, int);
229 static void keg_small_init(uma_keg_t keg);
230 static void keg_large_init(uma_keg_t keg);
231 static void zone_foreach(void (*zfunc)(uma_zone_t));
232 static void zone_timeout(uma_zone_t zone);
233 static int hash_alloc(struct uma_hash *);
234 static int hash_expand(struct uma_hash *, struct uma_hash *);
235 static void hash_free(struct uma_hash *hash);
236 static void uma_timeout(void *);
237 static void uma_startup3(void);
238 static void *zone_alloc_item(uma_zone_t, void *, int);
239 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip,
240     int);
241 static void bucket_enable(void);
242 static void bucket_init(void);
243 static uma_bucket_t bucket_alloc(int, int);
244 static void bucket_free(uma_bucket_t);
245 static void bucket_zone_drain(void);
246 static int zone_alloc_bucket(uma_zone_t zone, int flags);
247 static uma_slab_t zone_fetch_slab(uma_zone_t zone, uma_keg_t last, int flags);
248 static uma_slab_t zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int flags);
249 static void *slab_alloc_item(uma_zone_t zone, uma_slab_t slab);
250 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit,
251     uma_fini fini, int align, u_int32_t flags);
252 static inline void zone_relock(uma_zone_t zone, uma_keg_t keg);
253 static inline void keg_relock(uma_keg_t keg, uma_zone_t zone);
254 
255 void uma_print_zone(uma_zone_t);
256 void uma_print_stats(void);
257 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS);
258 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS);
259 
260 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
261 
262 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLTYPE_INT,
263     0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones");
264 
265 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLTYPE_STRUCT,
266     0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats");
267 
268 /*
269  * This routine checks to see whether or not it's safe to enable buckets.
270  */
271 
272 static void
273 bucket_enable(void)
274 {
275 	if (cnt.v_free_count < cnt.v_free_min)
276 		bucketdisable = 1;
277 	else
278 		bucketdisable = 0;
279 }
280 
281 /*
282  * Initialize bucket_zones, the array of zones of buckets of various sizes.
283  *
284  * For each zone, calculate the memory required for each bucket, consisting
285  * of the header and an array of pointers.  Initialize bucket_size[] to point
286  * the range of appropriate bucket sizes at the zone.
287  */
288 static void
289 bucket_init(void)
290 {
291 	struct uma_bucket_zone *ubz;
292 	int i;
293 	int j;
294 
295 	for (i = 0, j = 0; bucket_zones[j].ubz_entries != 0; j++) {
296 		int size;
297 
298 		ubz = &bucket_zones[j];
299 		size = roundup(sizeof(struct uma_bucket), sizeof(void *));
300 		size += sizeof(void *) * ubz->ubz_entries;
301 		ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
302 		    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
303 		    UMA_ZFLAG_INTERNAL | UMA_ZFLAG_BUCKET);
304 		for (; i <= ubz->ubz_entries; i += (1 << BUCKET_SHIFT))
305 			bucket_size[i >> BUCKET_SHIFT] = j;
306 	}
307 }
308 
309 /*
310  * Given a desired number of entries for a bucket, return the zone from which
311  * to allocate the bucket.
312  */
313 static struct uma_bucket_zone *
314 bucket_zone_lookup(int entries)
315 {
316 	int idx;
317 
318 	idx = howmany(entries, 1 << BUCKET_SHIFT);
319 	return (&bucket_zones[bucket_size[idx]]);
320 }
321 
322 static uma_bucket_t
323 bucket_alloc(int entries, int bflags)
324 {
325 	struct uma_bucket_zone *ubz;
326 	uma_bucket_t bucket;
327 
328 	/*
329 	 * This is to stop us from allocating per cpu buckets while we're
330 	 * running out of vm.boot_pages.  Otherwise, we would exhaust the
331 	 * boot pages.  This also prevents us from allocating buckets in
332 	 * low memory situations.
333 	 */
334 	if (bucketdisable)
335 		return (NULL);
336 
337 	ubz = bucket_zone_lookup(entries);
338 	bucket = zone_alloc_item(ubz->ubz_zone, NULL, bflags);
339 	if (bucket) {
340 #ifdef INVARIANTS
341 		bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
342 #endif
343 		bucket->ub_cnt = 0;
344 		bucket->ub_entries = ubz->ubz_entries;
345 	}
346 
347 	return (bucket);
348 }
349 
350 static void
351 bucket_free(uma_bucket_t bucket)
352 {
353 	struct uma_bucket_zone *ubz;
354 
355 	ubz = bucket_zone_lookup(bucket->ub_entries);
356 	zone_free_item(ubz->ubz_zone, bucket, NULL, SKIP_NONE,
357 	    ZFREE_STATFREE);
358 }
359 
360 static void
361 bucket_zone_drain(void)
362 {
363 	struct uma_bucket_zone *ubz;
364 
365 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
366 		zone_drain(ubz->ubz_zone);
367 }
368 
369 static inline uma_keg_t
370 zone_first_keg(uma_zone_t zone)
371 {
372 
373 	return (LIST_FIRST(&zone->uz_kegs)->kl_keg);
374 }
375 
376 static void
377 zone_foreach_keg(uma_zone_t zone, void (*kegfn)(uma_keg_t))
378 {
379 	uma_klink_t klink;
380 
381 	LIST_FOREACH(klink, &zone->uz_kegs, kl_link)
382 		kegfn(klink->kl_keg);
383 }
384 
385 /*
386  * Routine called by timeout which is used to fire off some time interval
387  * based calculations.  (stats, hash size, etc.)
388  *
389  * Arguments:
390  *	arg   Unused
391  *
392  * Returns:
393  *	Nothing
394  */
395 static void
396 uma_timeout(void *unused)
397 {
398 	bucket_enable();
399 	zone_foreach(zone_timeout);
400 
401 	/* Reschedule this event */
402 	callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
403 }
404 
405 /*
406  * Routine to perform timeout driven calculations.  This expands the
407  * hashes and does per cpu statistics aggregation.
408  *
409  *  Returns nothing.
410  */
411 static void
412 keg_timeout(uma_keg_t keg)
413 {
414 
415 	KEG_LOCK(keg);
416 	/*
417 	 * Expand the keg hash table.
418 	 *
419 	 * This is done if the number of slabs is larger than the hash size.
420 	 * What I'm trying to do here is completely reduce collisions.  This
421 	 * may be a little aggressive.  Should I allow for two collisions max?
422 	 */
423 	if (keg->uk_flags & UMA_ZONE_HASH &&
424 	    keg->uk_pages / keg->uk_ppera >= keg->uk_hash.uh_hashsize) {
425 		struct uma_hash newhash;
426 		struct uma_hash oldhash;
427 		int ret;
428 
429 		/*
430 		 * This is so involved because allocating and freeing
431 		 * while the keg lock is held will lead to deadlock.
432 		 * I have to do everything in stages and check for
433 		 * races.
434 		 */
435 		newhash = keg->uk_hash;
436 		KEG_UNLOCK(keg);
437 		ret = hash_alloc(&newhash);
438 		KEG_LOCK(keg);
439 		if (ret) {
440 			if (hash_expand(&keg->uk_hash, &newhash)) {
441 				oldhash = keg->uk_hash;
442 				keg->uk_hash = newhash;
443 			} else
444 				oldhash = newhash;
445 
446 			KEG_UNLOCK(keg);
447 			hash_free(&oldhash);
448 			KEG_LOCK(keg);
449 		}
450 	}
451 	KEG_UNLOCK(keg);
452 }
453 
454 static void
455 zone_timeout(uma_zone_t zone)
456 {
457 
458 	zone_foreach_keg(zone, &keg_timeout);
459 }
460 
461 /*
462  * Allocate and zero fill the next sized hash table from the appropriate
463  * backing store.
464  *
465  * Arguments:
466  *	hash  A new hash structure with the old hash size in uh_hashsize
467  *
468  * Returns:
469  *	1 on sucess and 0 on failure.
470  */
471 static int
472 hash_alloc(struct uma_hash *hash)
473 {
474 	int oldsize;
475 	int alloc;
476 
477 	oldsize = hash->uh_hashsize;
478 
479 	/* We're just going to go to a power of two greater */
480 	if (oldsize)  {
481 		hash->uh_hashsize = oldsize * 2;
482 		alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
483 		hash->uh_slab_hash = (struct slabhead *)malloc(alloc,
484 		    M_UMAHASH, M_NOWAIT);
485 	} else {
486 		alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
487 		hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
488 		    M_WAITOK);
489 		hash->uh_hashsize = UMA_HASH_SIZE_INIT;
490 	}
491 	if (hash->uh_slab_hash) {
492 		bzero(hash->uh_slab_hash, alloc);
493 		hash->uh_hashmask = hash->uh_hashsize - 1;
494 		return (1);
495 	}
496 
497 	return (0);
498 }
499 
500 /*
501  * Expands the hash table for HASH zones.  This is done from zone_timeout
502  * to reduce collisions.  This must not be done in the regular allocation
503  * path, otherwise, we can recurse on the vm while allocating pages.
504  *
505  * Arguments:
506  *	oldhash  The hash you want to expand
507  *	newhash  The hash structure for the new table
508  *
509  * Returns:
510  *	Nothing
511  *
512  * Discussion:
513  */
514 static int
515 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
516 {
517 	uma_slab_t slab;
518 	int hval;
519 	int i;
520 
521 	if (!newhash->uh_slab_hash)
522 		return (0);
523 
524 	if (oldhash->uh_hashsize >= newhash->uh_hashsize)
525 		return (0);
526 
527 	/*
528 	 * I need to investigate hash algorithms for resizing without a
529 	 * full rehash.
530 	 */
531 
532 	for (i = 0; i < oldhash->uh_hashsize; i++)
533 		while (!SLIST_EMPTY(&oldhash->uh_slab_hash[i])) {
534 			slab = SLIST_FIRST(&oldhash->uh_slab_hash[i]);
535 			SLIST_REMOVE_HEAD(&oldhash->uh_slab_hash[i], us_hlink);
536 			hval = UMA_HASH(newhash, slab->us_data);
537 			SLIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
538 			    slab, us_hlink);
539 		}
540 
541 	return (1);
542 }
543 
544 /*
545  * Free the hash bucket to the appropriate backing store.
546  *
547  * Arguments:
548  *	slab_hash  The hash bucket we're freeing
549  *	hashsize   The number of entries in that hash bucket
550  *
551  * Returns:
552  *	Nothing
553  */
554 static void
555 hash_free(struct uma_hash *hash)
556 {
557 	if (hash->uh_slab_hash == NULL)
558 		return;
559 	if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
560 		zone_free_item(hashzone,
561 		    hash->uh_slab_hash, NULL, SKIP_NONE, ZFREE_STATFREE);
562 	else
563 		free(hash->uh_slab_hash, M_UMAHASH);
564 }
565 
566 /*
567  * Frees all outstanding items in a bucket
568  *
569  * Arguments:
570  *	zone   The zone to free to, must be unlocked.
571  *	bucket The free/alloc bucket with items, cpu queue must be locked.
572  *
573  * Returns:
574  *	Nothing
575  */
576 
577 static void
578 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
579 {
580 	void *item;
581 
582 	if (bucket == NULL)
583 		return;
584 
585 	while (bucket->ub_cnt > 0)  {
586 		bucket->ub_cnt--;
587 		item = bucket->ub_bucket[bucket->ub_cnt];
588 #ifdef INVARIANTS
589 		bucket->ub_bucket[bucket->ub_cnt] = NULL;
590 		KASSERT(item != NULL,
591 		    ("bucket_drain: botched ptr, item is NULL"));
592 #endif
593 		zone_free_item(zone, item, NULL, SKIP_DTOR, 0);
594 	}
595 }
596 
597 /*
598  * Drains the per cpu caches for a zone.
599  *
600  * NOTE: This may only be called while the zone is being turn down, and not
601  * during normal operation.  This is necessary in order that we do not have
602  * to migrate CPUs to drain the per-CPU caches.
603  *
604  * Arguments:
605  *	zone     The zone to drain, must be unlocked.
606  *
607  * Returns:
608  *	Nothing
609  */
610 static void
611 cache_drain(uma_zone_t zone)
612 {
613 	uma_cache_t cache;
614 	int cpu;
615 
616 	/*
617 	 * XXX: It is safe to not lock the per-CPU caches, because we're
618 	 * tearing down the zone anyway.  I.e., there will be no further use
619 	 * of the caches at this point.
620 	 *
621 	 * XXX: It would good to be able to assert that the zone is being
622 	 * torn down to prevent improper use of cache_drain().
623 	 *
624 	 * XXX: We lock the zone before passing into bucket_cache_drain() as
625 	 * it is used elsewhere.  Should the tear-down path be made special
626 	 * there in some form?
627 	 */
628 	CPU_FOREACH(cpu) {
629 		cache = &zone->uz_cpu[cpu];
630 		bucket_drain(zone, cache->uc_allocbucket);
631 		bucket_drain(zone, cache->uc_freebucket);
632 		if (cache->uc_allocbucket != NULL)
633 			bucket_free(cache->uc_allocbucket);
634 		if (cache->uc_freebucket != NULL)
635 			bucket_free(cache->uc_freebucket);
636 		cache->uc_allocbucket = cache->uc_freebucket = NULL;
637 	}
638 	ZONE_LOCK(zone);
639 	bucket_cache_drain(zone);
640 	ZONE_UNLOCK(zone);
641 }
642 
643 /*
644  * Drain the cached buckets from a zone.  Expects a locked zone on entry.
645  */
646 static void
647 bucket_cache_drain(uma_zone_t zone)
648 {
649 	uma_bucket_t bucket;
650 
651 	/*
652 	 * Drain the bucket queues and free the buckets, we just keep two per
653 	 * cpu (alloc/free).
654 	 */
655 	while ((bucket = LIST_FIRST(&zone->uz_full_bucket)) != NULL) {
656 		LIST_REMOVE(bucket, ub_link);
657 		ZONE_UNLOCK(zone);
658 		bucket_drain(zone, bucket);
659 		bucket_free(bucket);
660 		ZONE_LOCK(zone);
661 	}
662 
663 	/* Now we do the free queue.. */
664 	while ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) {
665 		LIST_REMOVE(bucket, ub_link);
666 		bucket_free(bucket);
667 	}
668 }
669 
670 /*
671  * Frees pages from a keg back to the system.  This is done on demand from
672  * the pageout daemon.
673  *
674  * Returns nothing.
675  */
676 static void
677 keg_drain(uma_keg_t keg)
678 {
679 	struct slabhead freeslabs = { 0 };
680 	uma_slab_t slab;
681 	uma_slab_t n;
682 	u_int8_t flags;
683 	u_int8_t *mem;
684 	int i;
685 
686 	/*
687 	 * We don't want to take pages from statically allocated kegs at this
688 	 * time
689 	 */
690 	if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL)
691 		return;
692 
693 #ifdef UMA_DEBUG
694 	printf("%s free items: %u\n", keg->uk_name, keg->uk_free);
695 #endif
696 	KEG_LOCK(keg);
697 	if (keg->uk_free == 0)
698 		goto finished;
699 
700 	slab = LIST_FIRST(&keg->uk_free_slab);
701 	while (slab) {
702 		n = LIST_NEXT(slab, us_link);
703 
704 		/* We have no where to free these to */
705 		if (slab->us_flags & UMA_SLAB_BOOT) {
706 			slab = n;
707 			continue;
708 		}
709 
710 		LIST_REMOVE(slab, us_link);
711 		keg->uk_pages -= keg->uk_ppera;
712 		keg->uk_free -= keg->uk_ipers;
713 
714 		if (keg->uk_flags & UMA_ZONE_HASH)
715 			UMA_HASH_REMOVE(&keg->uk_hash, slab, slab->us_data);
716 
717 		SLIST_INSERT_HEAD(&freeslabs, slab, us_hlink);
718 
719 		slab = n;
720 	}
721 finished:
722 	KEG_UNLOCK(keg);
723 
724 	while ((slab = SLIST_FIRST(&freeslabs)) != NULL) {
725 		SLIST_REMOVE(&freeslabs, slab, uma_slab, us_hlink);
726 		if (keg->uk_fini)
727 			for (i = 0; i < keg->uk_ipers; i++)
728 				keg->uk_fini(
729 				    slab->us_data + (keg->uk_rsize * i),
730 				    keg->uk_size);
731 		flags = slab->us_flags;
732 		mem = slab->us_data;
733 
734 		if (keg->uk_flags & UMA_ZONE_VTOSLAB) {
735 			vm_object_t obj;
736 
737 			if (flags & UMA_SLAB_KMEM)
738 				obj = kmem_object;
739 			else if (flags & UMA_SLAB_KERNEL)
740 				obj = kernel_object;
741 			else
742 				obj = NULL;
743 			for (i = 0; i < keg->uk_ppera; i++)
744 				vsetobj((vm_offset_t)mem + (i * PAGE_SIZE),
745 				    obj);
746 		}
747 		if (keg->uk_flags & UMA_ZONE_OFFPAGE)
748 			zone_free_item(keg->uk_slabzone, slab, NULL,
749 			    SKIP_NONE, ZFREE_STATFREE);
750 #ifdef UMA_DEBUG
751 		printf("%s: Returning %d bytes.\n",
752 		    keg->uk_name, UMA_SLAB_SIZE * keg->uk_ppera);
753 #endif
754 		keg->uk_freef(mem, UMA_SLAB_SIZE * keg->uk_ppera, flags);
755 	}
756 }
757 
758 static void
759 zone_drain_wait(uma_zone_t zone, int waitok)
760 {
761 
762 	/*
763 	 * Set draining to interlock with zone_dtor() so we can release our
764 	 * locks as we go.  Only dtor() should do a WAITOK call since it
765 	 * is the only call that knows the structure will still be available
766 	 * when it wakes up.
767 	 */
768 	ZONE_LOCK(zone);
769 	while (zone->uz_flags & UMA_ZFLAG_DRAINING) {
770 		if (waitok == M_NOWAIT)
771 			goto out;
772 		mtx_unlock(&uma_mtx);
773 		msleep(zone, zone->uz_lock, PVM, "zonedrain", 1);
774 		mtx_lock(&uma_mtx);
775 	}
776 	zone->uz_flags |= UMA_ZFLAG_DRAINING;
777 	bucket_cache_drain(zone);
778 	ZONE_UNLOCK(zone);
779 	/*
780 	 * The DRAINING flag protects us from being freed while
781 	 * we're running.  Normally the uma_mtx would protect us but we
782 	 * must be able to release and acquire the right lock for each keg.
783 	 */
784 	zone_foreach_keg(zone, &keg_drain);
785 	ZONE_LOCK(zone);
786 	zone->uz_flags &= ~UMA_ZFLAG_DRAINING;
787 	wakeup(zone);
788 out:
789 	ZONE_UNLOCK(zone);
790 }
791 
792 void
793 zone_drain(uma_zone_t zone)
794 {
795 
796 	zone_drain_wait(zone, M_NOWAIT);
797 }
798 
799 /*
800  * Allocate a new slab for a keg.  This does not insert the slab onto a list.
801  *
802  * Arguments:
803  *	wait  Shall we wait?
804  *
805  * Returns:
806  *	The slab that was allocated or NULL if there is no memory and the
807  *	caller specified M_NOWAIT.
808  */
809 static uma_slab_t
810 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int wait)
811 {
812 	uma_slabrefcnt_t slabref;
813 	uma_alloc allocf;
814 	uma_slab_t slab;
815 	u_int8_t *mem;
816 	u_int8_t flags;
817 	int i;
818 
819 	mtx_assert(&keg->uk_lock, MA_OWNED);
820 	slab = NULL;
821 
822 #ifdef UMA_DEBUG
823 	printf("slab_zalloc:  Allocating a new slab for %s\n", keg->uk_name);
824 #endif
825 	allocf = keg->uk_allocf;
826 	KEG_UNLOCK(keg);
827 
828 	if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
829 		slab = zone_alloc_item(keg->uk_slabzone, NULL, wait);
830 		if (slab == NULL) {
831 			KEG_LOCK(keg);
832 			return NULL;
833 		}
834 	}
835 
836 	/*
837 	 * This reproduces the old vm_zone behavior of zero filling pages the
838 	 * first time they are added to a zone.
839 	 *
840 	 * Malloced items are zeroed in uma_zalloc.
841 	 */
842 
843 	if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
844 		wait |= M_ZERO;
845 	else
846 		wait &= ~M_ZERO;
847 
848 	/* zone is passed for legacy reasons. */
849 	mem = allocf(zone, keg->uk_ppera * UMA_SLAB_SIZE, &flags, wait);
850 	if (mem == NULL) {
851 		if (keg->uk_flags & UMA_ZONE_OFFPAGE)
852 			zone_free_item(keg->uk_slabzone, slab, NULL,
853 			    SKIP_NONE, ZFREE_STATFREE);
854 		KEG_LOCK(keg);
855 		return (NULL);
856 	}
857 
858 	/* Point the slab into the allocated memory */
859 	if (!(keg->uk_flags & UMA_ZONE_OFFPAGE))
860 		slab = (uma_slab_t )(mem + keg->uk_pgoff);
861 
862 	if (keg->uk_flags & UMA_ZONE_VTOSLAB)
863 		for (i = 0; i < keg->uk_ppera; i++)
864 			vsetslab((vm_offset_t)mem + (i * PAGE_SIZE), slab);
865 
866 	slab->us_keg = keg;
867 	slab->us_data = mem;
868 	slab->us_freecount = keg->uk_ipers;
869 	slab->us_firstfree = 0;
870 	slab->us_flags = flags;
871 
872 	if (keg->uk_flags & UMA_ZONE_REFCNT) {
873 		slabref = (uma_slabrefcnt_t)slab;
874 		for (i = 0; i < keg->uk_ipers; i++) {
875 			slabref->us_freelist[i].us_refcnt = 0;
876 			slabref->us_freelist[i].us_item = i+1;
877 		}
878 	} else {
879 		for (i = 0; i < keg->uk_ipers; i++)
880 			slab->us_freelist[i].us_item = i+1;
881 	}
882 
883 	if (keg->uk_init != NULL) {
884 		for (i = 0; i < keg->uk_ipers; i++)
885 			if (keg->uk_init(slab->us_data + (keg->uk_rsize * i),
886 			    keg->uk_size, wait) != 0)
887 				break;
888 		if (i != keg->uk_ipers) {
889 			if (keg->uk_fini != NULL) {
890 				for (i--; i > -1; i--)
891 					keg->uk_fini(slab->us_data +
892 					    (keg->uk_rsize * i),
893 					    keg->uk_size);
894 			}
895 			if (keg->uk_flags & UMA_ZONE_VTOSLAB) {
896 				vm_object_t obj;
897 
898 				if (flags & UMA_SLAB_KMEM)
899 					obj = kmem_object;
900 				else if (flags & UMA_SLAB_KERNEL)
901 					obj = kernel_object;
902 				else
903 					obj = NULL;
904 				for (i = 0; i < keg->uk_ppera; i++)
905 					vsetobj((vm_offset_t)mem +
906 					    (i * PAGE_SIZE), obj);
907 			}
908 			if (keg->uk_flags & UMA_ZONE_OFFPAGE)
909 				zone_free_item(keg->uk_slabzone, slab,
910 				    NULL, SKIP_NONE, ZFREE_STATFREE);
911 			keg->uk_freef(mem, UMA_SLAB_SIZE * keg->uk_ppera,
912 			    flags);
913 			KEG_LOCK(keg);
914 			return (NULL);
915 		}
916 	}
917 	KEG_LOCK(keg);
918 
919 	if (keg->uk_flags & UMA_ZONE_HASH)
920 		UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
921 
922 	keg->uk_pages += keg->uk_ppera;
923 	keg->uk_free += keg->uk_ipers;
924 
925 	return (slab);
926 }
927 
928 /*
929  * This function is intended to be used early on in place of page_alloc() so
930  * that we may use the boot time page cache to satisfy allocations before
931  * the VM is ready.
932  */
933 static void *
934 startup_alloc(uma_zone_t zone, int bytes, u_int8_t *pflag, int wait)
935 {
936 	uma_keg_t keg;
937 	uma_slab_t tmps;
938 	int pages, check_pages;
939 
940 	keg = zone_first_keg(zone);
941 	pages = howmany(bytes, PAGE_SIZE);
942 	check_pages = pages - 1;
943 	KASSERT(pages > 0, ("startup_alloc can't reserve 0 pages\n"));
944 
945 	/*
946 	 * Check our small startup cache to see if it has pages remaining.
947 	 */
948 	mtx_lock(&uma_boot_pages_mtx);
949 
950 	/* First check if we have enough room. */
951 	tmps = LIST_FIRST(&uma_boot_pages);
952 	while (tmps != NULL && check_pages-- > 0)
953 		tmps = LIST_NEXT(tmps, us_link);
954 	if (tmps != NULL) {
955 		/*
956 		 * It's ok to lose tmps references.  The last one will
957 		 * have tmps->us_data pointing to the start address of
958 		 * "pages" contiguous pages of memory.
959 		 */
960 		while (pages-- > 0) {
961 			tmps = LIST_FIRST(&uma_boot_pages);
962 			LIST_REMOVE(tmps, us_link);
963 		}
964 		mtx_unlock(&uma_boot_pages_mtx);
965 		*pflag = tmps->us_flags;
966 		return (tmps->us_data);
967 	}
968 	mtx_unlock(&uma_boot_pages_mtx);
969 	if (booted < UMA_STARTUP2)
970 		panic("UMA: Increase vm.boot_pages");
971 	/*
972 	 * Now that we've booted reset these users to their real allocator.
973 	 */
974 #ifdef UMA_MD_SMALL_ALLOC
975 	keg->uk_allocf = (keg->uk_ppera > 1) ? page_alloc : uma_small_alloc;
976 #else
977 	keg->uk_allocf = page_alloc;
978 #endif
979 	return keg->uk_allocf(zone, bytes, pflag, wait);
980 }
981 
982 /*
983  * Allocates a number of pages from the system
984  *
985  * Arguments:
986  *	bytes  The number of bytes requested
987  *	wait  Shall we wait?
988  *
989  * Returns:
990  *	A pointer to the alloced memory or possibly
991  *	NULL if M_NOWAIT is set.
992  */
993 static void *
994 page_alloc(uma_zone_t zone, int bytes, u_int8_t *pflag, int wait)
995 {
996 	void *p;	/* Returned page */
997 
998 	*pflag = UMA_SLAB_KMEM;
999 	p = (void *) kmem_malloc(kmem_map, bytes, wait);
1000 
1001 	return (p);
1002 }
1003 
1004 /*
1005  * Allocates a number of pages from within an object
1006  *
1007  * Arguments:
1008  *	bytes  The number of bytes requested
1009  *	wait   Shall we wait?
1010  *
1011  * Returns:
1012  *	A pointer to the alloced memory or possibly
1013  *	NULL if M_NOWAIT is set.
1014  */
1015 static void *
1016 obj_alloc(uma_zone_t zone, int bytes, u_int8_t *flags, int wait)
1017 {
1018 	vm_object_t object;
1019 	vm_offset_t retkva, zkva;
1020 	vm_page_t p;
1021 	int pages, startpages;
1022 	uma_keg_t keg;
1023 
1024 	keg = zone_first_keg(zone);
1025 	object = keg->uk_obj;
1026 	retkva = 0;
1027 
1028 	/*
1029 	 * This looks a little weird since we're getting one page at a time.
1030 	 */
1031 	VM_OBJECT_LOCK(object);
1032 	p = TAILQ_LAST(&object->memq, pglist);
1033 	pages = p != NULL ? p->pindex + 1 : 0;
1034 	startpages = pages;
1035 	zkva = keg->uk_kva + pages * PAGE_SIZE;
1036 	for (; bytes > 0; bytes -= PAGE_SIZE) {
1037 		p = vm_page_alloc(object, pages,
1038 		    VM_ALLOC_INTERRUPT | VM_ALLOC_WIRED);
1039 		if (p == NULL) {
1040 			if (pages != startpages)
1041 				pmap_qremove(retkva, pages - startpages);
1042 			while (pages != startpages) {
1043 				pages--;
1044 				p = TAILQ_LAST(&object->memq, pglist);
1045 				vm_page_unwire(p, 0);
1046 				vm_page_free(p);
1047 			}
1048 			retkva = 0;
1049 			goto done;
1050 		}
1051 		pmap_qenter(zkva, &p, 1);
1052 		if (retkva == 0)
1053 			retkva = zkva;
1054 		zkva += PAGE_SIZE;
1055 		pages += 1;
1056 	}
1057 done:
1058 	VM_OBJECT_UNLOCK(object);
1059 	*flags = UMA_SLAB_PRIV;
1060 
1061 	return ((void *)retkva);
1062 }
1063 
1064 /*
1065  * Frees a number of pages to the system
1066  *
1067  * Arguments:
1068  *	mem   A pointer to the memory to be freed
1069  *	size  The size of the memory being freed
1070  *	flags The original p->us_flags field
1071  *
1072  * Returns:
1073  *	Nothing
1074  */
1075 static void
1076 page_free(void *mem, int size, u_int8_t flags)
1077 {
1078 	vm_map_t map;
1079 
1080 	if (flags & UMA_SLAB_KMEM)
1081 		map = kmem_map;
1082 	else if (flags & UMA_SLAB_KERNEL)
1083 		map = kernel_map;
1084 	else
1085 		panic("UMA: page_free used with invalid flags %d", flags);
1086 
1087 	kmem_free(map, (vm_offset_t)mem, size);
1088 }
1089 
1090 /*
1091  * Zero fill initializer
1092  *
1093  * Arguments/Returns follow uma_init specifications
1094  */
1095 static int
1096 zero_init(void *mem, int size, int flags)
1097 {
1098 	bzero(mem, size);
1099 	return (0);
1100 }
1101 
1102 /*
1103  * Finish creating a small uma keg.  This calculates ipers, and the keg size.
1104  *
1105  * Arguments
1106  *	keg  The zone we should initialize
1107  *
1108  * Returns
1109  *	Nothing
1110  */
1111 static void
1112 keg_small_init(uma_keg_t keg)
1113 {
1114 	u_int rsize;
1115 	u_int memused;
1116 	u_int wastedspace;
1117 	u_int shsize;
1118 
1119 	KASSERT(keg != NULL, ("Keg is null in keg_small_init"));
1120 	rsize = keg->uk_size;
1121 
1122 	if (rsize < UMA_SMALLEST_UNIT)
1123 		rsize = UMA_SMALLEST_UNIT;
1124 	if (rsize & keg->uk_align)
1125 		rsize = (rsize & ~keg->uk_align) + (keg->uk_align + 1);
1126 
1127 	keg->uk_rsize = rsize;
1128 	keg->uk_ppera = 1;
1129 
1130 	if (keg->uk_flags & UMA_ZONE_REFCNT) {
1131 		rsize += UMA_FRITMREF_SZ;	/* linkage & refcnt */
1132 		shsize = sizeof(struct uma_slab_refcnt);
1133 	} else {
1134 		rsize += UMA_FRITM_SZ;	/* Account for linkage */
1135 		shsize = sizeof(struct uma_slab);
1136 	}
1137 
1138 	keg->uk_ipers = (UMA_SLAB_SIZE - shsize) / rsize;
1139 	KASSERT(keg->uk_ipers != 0, ("keg_small_init: ipers is 0"));
1140 	memused = keg->uk_ipers * rsize + shsize;
1141 	wastedspace = UMA_SLAB_SIZE - memused;
1142 
1143 	/*
1144 	 * We can't do OFFPAGE if we're internal or if we've been
1145 	 * asked to not go to the VM for buckets.  If we do this we
1146 	 * may end up going to the VM (kmem_map) for slabs which we
1147 	 * do not want to do if we're UMA_ZFLAG_CACHEONLY as a
1148 	 * result of UMA_ZONE_VM, which clearly forbids it.
1149 	 */
1150 	if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) ||
1151 	    (keg->uk_flags & UMA_ZFLAG_CACHEONLY))
1152 		return;
1153 
1154 	if ((wastedspace >= UMA_MAX_WASTE) &&
1155 	    (keg->uk_ipers < (UMA_SLAB_SIZE / keg->uk_rsize))) {
1156 		keg->uk_ipers = UMA_SLAB_SIZE / keg->uk_rsize;
1157 		KASSERT(keg->uk_ipers <= 255,
1158 		    ("keg_small_init: keg->uk_ipers too high!"));
1159 #ifdef UMA_DEBUG
1160 		printf("UMA decided we need offpage slab headers for "
1161 		    "keg: %s, calculated wastedspace = %d, "
1162 		    "maximum wasted space allowed = %d, "
1163 		    "calculated ipers = %d, "
1164 		    "new wasted space = %d\n", keg->uk_name, wastedspace,
1165 		    UMA_MAX_WASTE, keg->uk_ipers,
1166 		    UMA_SLAB_SIZE - keg->uk_ipers * keg->uk_rsize);
1167 #endif
1168 		keg->uk_flags |= UMA_ZONE_OFFPAGE;
1169 		if ((keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1170 			keg->uk_flags |= UMA_ZONE_HASH;
1171 	}
1172 }
1173 
1174 /*
1175  * Finish creating a large (> UMA_SLAB_SIZE) uma kegs.  Just give in and do
1176  * OFFPAGE for now.  When I can allow for more dynamic slab sizes this will be
1177  * more complicated.
1178  *
1179  * Arguments
1180  *	keg  The keg we should initialize
1181  *
1182  * Returns
1183  *	Nothing
1184  */
1185 static void
1186 keg_large_init(uma_keg_t keg)
1187 {
1188 	int pages;
1189 
1190 	KASSERT(keg != NULL, ("Keg is null in keg_large_init"));
1191 	KASSERT((keg->uk_flags & UMA_ZFLAG_CACHEONLY) == 0,
1192 	    ("keg_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY keg"));
1193 
1194 	pages = keg->uk_size / UMA_SLAB_SIZE;
1195 
1196 	/* Account for remainder */
1197 	if ((pages * UMA_SLAB_SIZE) < keg->uk_size)
1198 		pages++;
1199 
1200 	keg->uk_ppera = pages;
1201 	keg->uk_ipers = 1;
1202 	keg->uk_rsize = keg->uk_size;
1203 
1204 	/* We can't do OFFPAGE if we're internal, bail out here. */
1205 	if (keg->uk_flags & UMA_ZFLAG_INTERNAL)
1206 		return;
1207 
1208 	keg->uk_flags |= UMA_ZONE_OFFPAGE;
1209 	if ((keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1210 		keg->uk_flags |= UMA_ZONE_HASH;
1211 }
1212 
1213 static void
1214 keg_cachespread_init(uma_keg_t keg)
1215 {
1216 	int alignsize;
1217 	int trailer;
1218 	int pages;
1219 	int rsize;
1220 
1221 	alignsize = keg->uk_align + 1;
1222 	rsize = keg->uk_size;
1223 	/*
1224 	 * We want one item to start on every align boundary in a page.  To
1225 	 * do this we will span pages.  We will also extend the item by the
1226 	 * size of align if it is an even multiple of align.  Otherwise, it
1227 	 * would fall on the same boundary every time.
1228 	 */
1229 	if (rsize & keg->uk_align)
1230 		rsize = (rsize & ~keg->uk_align) + alignsize;
1231 	if ((rsize & alignsize) == 0)
1232 		rsize += alignsize;
1233 	trailer = rsize - keg->uk_size;
1234 	pages = (rsize * (PAGE_SIZE / alignsize)) / PAGE_SIZE;
1235 	pages = MIN(pages, (128 * 1024) / PAGE_SIZE);
1236 	keg->uk_rsize = rsize;
1237 	keg->uk_ppera = pages;
1238 	keg->uk_ipers = ((pages * PAGE_SIZE) + trailer) / rsize;
1239 	keg->uk_flags |= UMA_ZONE_OFFPAGE | UMA_ZONE_VTOSLAB;
1240 	KASSERT(keg->uk_ipers <= uma_max_ipers,
1241 	    ("keg_small_init: keg->uk_ipers too high(%d) increase max_ipers",
1242 	    keg->uk_ipers));
1243 }
1244 
1245 /*
1246  * Keg header ctor.  This initializes all fields, locks, etc.  And inserts
1247  * the keg onto the global keg list.
1248  *
1249  * Arguments/Returns follow uma_ctor specifications
1250  *	udata  Actually uma_kctor_args
1251  */
1252 static int
1253 keg_ctor(void *mem, int size, void *udata, int flags)
1254 {
1255 	struct uma_kctor_args *arg = udata;
1256 	uma_keg_t keg = mem;
1257 	uma_zone_t zone;
1258 
1259 	bzero(keg, size);
1260 	keg->uk_size = arg->size;
1261 	keg->uk_init = arg->uminit;
1262 	keg->uk_fini = arg->fini;
1263 	keg->uk_align = arg->align;
1264 	keg->uk_free = 0;
1265 	keg->uk_pages = 0;
1266 	keg->uk_flags = arg->flags;
1267 	keg->uk_allocf = page_alloc;
1268 	keg->uk_freef = page_free;
1269 	keg->uk_recurse = 0;
1270 	keg->uk_slabzone = NULL;
1271 
1272 	/*
1273 	 * The master zone is passed to us at keg-creation time.
1274 	 */
1275 	zone = arg->zone;
1276 	keg->uk_name = zone->uz_name;
1277 
1278 	if (arg->flags & UMA_ZONE_VM)
1279 		keg->uk_flags |= UMA_ZFLAG_CACHEONLY;
1280 
1281 	if (arg->flags & UMA_ZONE_ZINIT)
1282 		keg->uk_init = zero_init;
1283 
1284 	if (arg->flags & UMA_ZONE_REFCNT || arg->flags & UMA_ZONE_MALLOC)
1285 		keg->uk_flags |= UMA_ZONE_VTOSLAB;
1286 
1287 	/*
1288 	 * The +UMA_FRITM_SZ added to uk_size is to account for the
1289 	 * linkage that is added to the size in keg_small_init().  If
1290 	 * we don't account for this here then we may end up in
1291 	 * keg_small_init() with a calculated 'ipers' of 0.
1292 	 */
1293 	if (keg->uk_flags & UMA_ZONE_REFCNT) {
1294 		if (keg->uk_flags & UMA_ZONE_CACHESPREAD)
1295 			keg_cachespread_init(keg);
1296 		else if ((keg->uk_size+UMA_FRITMREF_SZ) >
1297 		    (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt)))
1298 			keg_large_init(keg);
1299 		else
1300 			keg_small_init(keg);
1301 	} else {
1302 		if (keg->uk_flags & UMA_ZONE_CACHESPREAD)
1303 			keg_cachespread_init(keg);
1304 		else if ((keg->uk_size+UMA_FRITM_SZ) >
1305 		    (UMA_SLAB_SIZE - sizeof(struct uma_slab)))
1306 			keg_large_init(keg);
1307 		else
1308 			keg_small_init(keg);
1309 	}
1310 
1311 	if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
1312 		if (keg->uk_flags & UMA_ZONE_REFCNT)
1313 			keg->uk_slabzone = slabrefzone;
1314 		else
1315 			keg->uk_slabzone = slabzone;
1316 	}
1317 
1318 	/*
1319 	 * If we haven't booted yet we need allocations to go through the
1320 	 * startup cache until the vm is ready.
1321 	 */
1322 	if (keg->uk_ppera == 1) {
1323 #ifdef UMA_MD_SMALL_ALLOC
1324 		keg->uk_allocf = uma_small_alloc;
1325 		keg->uk_freef = uma_small_free;
1326 
1327 		if (booted < UMA_STARTUP)
1328 			keg->uk_allocf = startup_alloc;
1329 #else
1330 		if (booted < UMA_STARTUP2)
1331 			keg->uk_allocf = startup_alloc;
1332 #endif
1333 	} else if (booted < UMA_STARTUP2 &&
1334 	    (keg->uk_flags & UMA_ZFLAG_INTERNAL))
1335 		keg->uk_allocf = startup_alloc;
1336 
1337 	/*
1338 	 * Initialize keg's lock (shared among zones).
1339 	 */
1340 	if (arg->flags & UMA_ZONE_MTXCLASS)
1341 		KEG_LOCK_INIT(keg, 1);
1342 	else
1343 		KEG_LOCK_INIT(keg, 0);
1344 
1345 	/*
1346 	 * If we're putting the slab header in the actual page we need to
1347 	 * figure out where in each page it goes.  This calculates a right
1348 	 * justified offset into the memory on an ALIGN_PTR boundary.
1349 	 */
1350 	if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) {
1351 		u_int totsize;
1352 
1353 		/* Size of the slab struct and free list */
1354 		if (keg->uk_flags & UMA_ZONE_REFCNT)
1355 			totsize = sizeof(struct uma_slab_refcnt) +
1356 			    keg->uk_ipers * UMA_FRITMREF_SZ;
1357 		else
1358 			totsize = sizeof(struct uma_slab) +
1359 			    keg->uk_ipers * UMA_FRITM_SZ;
1360 
1361 		if (totsize & UMA_ALIGN_PTR)
1362 			totsize = (totsize & ~UMA_ALIGN_PTR) +
1363 			    (UMA_ALIGN_PTR + 1);
1364 		keg->uk_pgoff = (UMA_SLAB_SIZE * keg->uk_ppera) - totsize;
1365 
1366 		if (keg->uk_flags & UMA_ZONE_REFCNT)
1367 			totsize = keg->uk_pgoff + sizeof(struct uma_slab_refcnt)
1368 			    + keg->uk_ipers * UMA_FRITMREF_SZ;
1369 		else
1370 			totsize = keg->uk_pgoff + sizeof(struct uma_slab)
1371 			    + keg->uk_ipers * UMA_FRITM_SZ;
1372 
1373 		/*
1374 		 * The only way the following is possible is if with our
1375 		 * UMA_ALIGN_PTR adjustments we are now bigger than
1376 		 * UMA_SLAB_SIZE.  I haven't checked whether this is
1377 		 * mathematically possible for all cases, so we make
1378 		 * sure here anyway.
1379 		 */
1380 		if (totsize > UMA_SLAB_SIZE * keg->uk_ppera) {
1381 			printf("zone %s ipers %d rsize %d size %d\n",
1382 			    zone->uz_name, keg->uk_ipers, keg->uk_rsize,
1383 			    keg->uk_size);
1384 			panic("UMA slab won't fit.");
1385 		}
1386 	}
1387 
1388 	if (keg->uk_flags & UMA_ZONE_HASH)
1389 		hash_alloc(&keg->uk_hash);
1390 
1391 #ifdef UMA_DEBUG
1392 	printf("UMA: %s(%p) size %d(%d) flags %d ipers %d ppera %d out %d free %d\n",
1393 	    zone->uz_name, zone, keg->uk_size, keg->uk_rsize, keg->uk_flags,
1394 	    keg->uk_ipers, keg->uk_ppera,
1395 	    (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free);
1396 #endif
1397 
1398 	LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
1399 
1400 	mtx_lock(&uma_mtx);
1401 	LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
1402 	mtx_unlock(&uma_mtx);
1403 	return (0);
1404 }
1405 
1406 /*
1407  * Zone header ctor.  This initializes all fields, locks, etc.
1408  *
1409  * Arguments/Returns follow uma_ctor specifications
1410  *	udata  Actually uma_zctor_args
1411  */
1412 static int
1413 zone_ctor(void *mem, int size, void *udata, int flags)
1414 {
1415 	struct uma_zctor_args *arg = udata;
1416 	uma_zone_t zone = mem;
1417 	uma_zone_t z;
1418 	uma_keg_t keg;
1419 
1420 	bzero(zone, size);
1421 	zone->uz_name = arg->name;
1422 	zone->uz_ctor = arg->ctor;
1423 	zone->uz_dtor = arg->dtor;
1424 	zone->uz_slab = zone_fetch_slab;
1425 	zone->uz_init = NULL;
1426 	zone->uz_fini = NULL;
1427 	zone->uz_allocs = 0;
1428 	zone->uz_frees = 0;
1429 	zone->uz_fails = 0;
1430 	zone->uz_sleeps = 0;
1431 	zone->uz_fills = zone->uz_count = 0;
1432 	zone->uz_flags = 0;
1433 	keg = arg->keg;
1434 
1435 	if (arg->flags & UMA_ZONE_SECONDARY) {
1436 		KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
1437 		zone->uz_init = arg->uminit;
1438 		zone->uz_fini = arg->fini;
1439 		zone->uz_lock = &keg->uk_lock;
1440 		zone->uz_flags |= UMA_ZONE_SECONDARY;
1441 		mtx_lock(&uma_mtx);
1442 		ZONE_LOCK(zone);
1443 		LIST_FOREACH(z, &keg->uk_zones, uz_link) {
1444 			if (LIST_NEXT(z, uz_link) == NULL) {
1445 				LIST_INSERT_AFTER(z, zone, uz_link);
1446 				break;
1447 			}
1448 		}
1449 		ZONE_UNLOCK(zone);
1450 		mtx_unlock(&uma_mtx);
1451 	} else if (keg == NULL) {
1452 		if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
1453 		    arg->align, arg->flags)) == NULL)
1454 			return (ENOMEM);
1455 	} else {
1456 		struct uma_kctor_args karg;
1457 		int error;
1458 
1459 		/* We should only be here from uma_startup() */
1460 		karg.size = arg->size;
1461 		karg.uminit = arg->uminit;
1462 		karg.fini = arg->fini;
1463 		karg.align = arg->align;
1464 		karg.flags = arg->flags;
1465 		karg.zone = zone;
1466 		error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
1467 		    flags);
1468 		if (error)
1469 			return (error);
1470 	}
1471 	/*
1472 	 * Link in the first keg.
1473 	 */
1474 	zone->uz_klink.kl_keg = keg;
1475 	LIST_INSERT_HEAD(&zone->uz_kegs, &zone->uz_klink, kl_link);
1476 	zone->uz_lock = &keg->uk_lock;
1477 	zone->uz_size = keg->uk_size;
1478 	zone->uz_flags |= (keg->uk_flags &
1479 	    (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
1480 
1481 	/*
1482 	 * Some internal zones don't have room allocated for the per cpu
1483 	 * caches.  If we're internal, bail out here.
1484 	 */
1485 	if (keg->uk_flags & UMA_ZFLAG_INTERNAL) {
1486 		KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
1487 		    ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
1488 		return (0);
1489 	}
1490 
1491 	if (keg->uk_flags & UMA_ZONE_MAXBUCKET)
1492 		zone->uz_count = BUCKET_MAX;
1493 	else if (keg->uk_ipers <= BUCKET_MAX)
1494 		zone->uz_count = keg->uk_ipers;
1495 	else
1496 		zone->uz_count = BUCKET_MAX;
1497 	return (0);
1498 }
1499 
1500 /*
1501  * Keg header dtor.  This frees all data, destroys locks, frees the hash
1502  * table and removes the keg from the global list.
1503  *
1504  * Arguments/Returns follow uma_dtor specifications
1505  *	udata  unused
1506  */
1507 static void
1508 keg_dtor(void *arg, int size, void *udata)
1509 {
1510 	uma_keg_t keg;
1511 
1512 	keg = (uma_keg_t)arg;
1513 	KEG_LOCK(keg);
1514 	if (keg->uk_free != 0) {
1515 		printf("Freed UMA keg was not empty (%d items). "
1516 		    " Lost %d pages of memory.\n",
1517 		    keg->uk_free, keg->uk_pages);
1518 	}
1519 	KEG_UNLOCK(keg);
1520 
1521 	hash_free(&keg->uk_hash);
1522 
1523 	KEG_LOCK_FINI(keg);
1524 }
1525 
1526 /*
1527  * Zone header dtor.
1528  *
1529  * Arguments/Returns follow uma_dtor specifications
1530  *	udata  unused
1531  */
1532 static void
1533 zone_dtor(void *arg, int size, void *udata)
1534 {
1535 	uma_klink_t klink;
1536 	uma_zone_t zone;
1537 	uma_keg_t keg;
1538 
1539 	zone = (uma_zone_t)arg;
1540 	keg = zone_first_keg(zone);
1541 
1542 	if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
1543 		cache_drain(zone);
1544 
1545 	mtx_lock(&uma_mtx);
1546 	LIST_REMOVE(zone, uz_link);
1547 	mtx_unlock(&uma_mtx);
1548 	/*
1549 	 * XXX there are some races here where
1550 	 * the zone can be drained but zone lock
1551 	 * released and then refilled before we
1552 	 * remove it... we dont care for now
1553 	 */
1554 	zone_drain_wait(zone, M_WAITOK);
1555 	/*
1556 	 * Unlink all of our kegs.
1557 	 */
1558 	while ((klink = LIST_FIRST(&zone->uz_kegs)) != NULL) {
1559 		klink->kl_keg = NULL;
1560 		LIST_REMOVE(klink, kl_link);
1561 		if (klink == &zone->uz_klink)
1562 			continue;
1563 		free(klink, M_TEMP);
1564 	}
1565 	/*
1566 	 * We only destroy kegs from non secondary zones.
1567 	 */
1568 	if ((zone->uz_flags & UMA_ZONE_SECONDARY) == 0)  {
1569 		mtx_lock(&uma_mtx);
1570 		LIST_REMOVE(keg, uk_link);
1571 		mtx_unlock(&uma_mtx);
1572 		zone_free_item(kegs, keg, NULL, SKIP_NONE,
1573 		    ZFREE_STATFREE);
1574 	}
1575 }
1576 
1577 /*
1578  * Traverses every zone in the system and calls a callback
1579  *
1580  * Arguments:
1581  *	zfunc  A pointer to a function which accepts a zone
1582  *		as an argument.
1583  *
1584  * Returns:
1585  *	Nothing
1586  */
1587 static void
1588 zone_foreach(void (*zfunc)(uma_zone_t))
1589 {
1590 	uma_keg_t keg;
1591 	uma_zone_t zone;
1592 
1593 	mtx_lock(&uma_mtx);
1594 	LIST_FOREACH(keg, &uma_kegs, uk_link) {
1595 		LIST_FOREACH(zone, &keg->uk_zones, uz_link)
1596 			zfunc(zone);
1597 	}
1598 	mtx_unlock(&uma_mtx);
1599 }
1600 
1601 /* Public functions */
1602 /* See uma.h */
1603 void
1604 uma_startup(void *bootmem, int boot_pages)
1605 {
1606 	struct uma_zctor_args args;
1607 	uma_slab_t slab;
1608 	u_int slabsize;
1609 	u_int objsize, totsize, wsize;
1610 	int i;
1611 
1612 #ifdef UMA_DEBUG
1613 	printf("Creating uma keg headers zone and keg.\n");
1614 #endif
1615 	mtx_init(&uma_mtx, "UMA lock", NULL, MTX_DEF);
1616 
1617 	/*
1618 	 * Figure out the maximum number of items-per-slab we'll have if
1619 	 * we're using the OFFPAGE slab header to track free items, given
1620 	 * all possible object sizes and the maximum desired wastage
1621 	 * (UMA_MAX_WASTE).
1622 	 *
1623 	 * We iterate until we find an object size for
1624 	 * which the calculated wastage in keg_small_init() will be
1625 	 * enough to warrant OFFPAGE.  Since wastedspace versus objsize
1626 	 * is an overall increasing see-saw function, we find the smallest
1627 	 * objsize such that the wastage is always acceptable for objects
1628 	 * with that objsize or smaller.  Since a smaller objsize always
1629 	 * generates a larger possible uma_max_ipers, we use this computed
1630 	 * objsize to calculate the largest ipers possible.  Since the
1631 	 * ipers calculated for OFFPAGE slab headers is always larger than
1632 	 * the ipers initially calculated in keg_small_init(), we use
1633 	 * the former's equation (UMA_SLAB_SIZE / keg->uk_rsize) to
1634 	 * obtain the maximum ipers possible for offpage slab headers.
1635 	 *
1636 	 * It should be noted that ipers versus objsize is an inversly
1637 	 * proportional function which drops off rather quickly so as
1638 	 * long as our UMA_MAX_WASTE is such that the objsize we calculate
1639 	 * falls into the portion of the inverse relation AFTER the steep
1640 	 * falloff, then uma_max_ipers shouldn't be too high (~10 on i386).
1641 	 *
1642 	 * Note that we have 8-bits (1 byte) to use as a freelist index
1643 	 * inside the actual slab header itself and this is enough to
1644 	 * accomodate us.  In the worst case, a UMA_SMALLEST_UNIT sized
1645 	 * object with offpage slab header would have ipers =
1646 	 * UMA_SLAB_SIZE / UMA_SMALLEST_UNIT (currently = 256), which is
1647 	 * 1 greater than what our byte-integer freelist index can
1648 	 * accomodate, but we know that this situation never occurs as
1649 	 * for UMA_SMALLEST_UNIT-sized objects, we will never calculate
1650 	 * that we need to go to offpage slab headers.  Or, if we do,
1651 	 * then we trap that condition below and panic in the INVARIANTS case.
1652 	 */
1653 	wsize = UMA_SLAB_SIZE - sizeof(struct uma_slab) - UMA_MAX_WASTE;
1654 	totsize = wsize;
1655 	objsize = UMA_SMALLEST_UNIT;
1656 	while (totsize >= wsize) {
1657 		totsize = (UMA_SLAB_SIZE - sizeof(struct uma_slab)) /
1658 		    (objsize + UMA_FRITM_SZ);
1659 		totsize *= (UMA_FRITM_SZ + objsize);
1660 		objsize++;
1661 	}
1662 	if (objsize > UMA_SMALLEST_UNIT)
1663 		objsize--;
1664 	uma_max_ipers = MAX(UMA_SLAB_SIZE / objsize, 64);
1665 
1666 	wsize = UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt) - UMA_MAX_WASTE;
1667 	totsize = wsize;
1668 	objsize = UMA_SMALLEST_UNIT;
1669 	while (totsize >= wsize) {
1670 		totsize = (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt)) /
1671 		    (objsize + UMA_FRITMREF_SZ);
1672 		totsize *= (UMA_FRITMREF_SZ + objsize);
1673 		objsize++;
1674 	}
1675 	if (objsize > UMA_SMALLEST_UNIT)
1676 		objsize--;
1677 	uma_max_ipers_ref = MAX(UMA_SLAB_SIZE / objsize, 64);
1678 
1679 	KASSERT((uma_max_ipers_ref <= 255) && (uma_max_ipers <= 255),
1680 	    ("uma_startup: calculated uma_max_ipers values too large!"));
1681 
1682 #ifdef UMA_DEBUG
1683 	printf("Calculated uma_max_ipers (for OFFPAGE) is %d\n", uma_max_ipers);
1684 	printf("Calculated uma_max_ipers_slab (for OFFPAGE) is %d\n",
1685 	    uma_max_ipers_ref);
1686 #endif
1687 
1688 	/* "manually" create the initial zone */
1689 	args.name = "UMA Kegs";
1690 	args.size = sizeof(struct uma_keg);
1691 	args.ctor = keg_ctor;
1692 	args.dtor = keg_dtor;
1693 	args.uminit = zero_init;
1694 	args.fini = NULL;
1695 	args.keg = &masterkeg;
1696 	args.align = 32 - 1;
1697 	args.flags = UMA_ZFLAG_INTERNAL;
1698 	/* The initial zone has no Per cpu queues so it's smaller */
1699 	zone_ctor(kegs, sizeof(struct uma_zone), &args, M_WAITOK);
1700 
1701 #ifdef UMA_DEBUG
1702 	printf("Filling boot free list.\n");
1703 #endif
1704 	for (i = 0; i < boot_pages; i++) {
1705 		slab = (uma_slab_t)((u_int8_t *)bootmem + (i * UMA_SLAB_SIZE));
1706 		slab->us_data = (u_int8_t *)slab;
1707 		slab->us_flags = UMA_SLAB_BOOT;
1708 		LIST_INSERT_HEAD(&uma_boot_pages, slab, us_link);
1709 	}
1710 	mtx_init(&uma_boot_pages_mtx, "UMA boot pages", NULL, MTX_DEF);
1711 
1712 #ifdef UMA_DEBUG
1713 	printf("Creating uma zone headers zone and keg.\n");
1714 #endif
1715 	args.name = "UMA Zones";
1716 	args.size = sizeof(struct uma_zone) +
1717 	    (sizeof(struct uma_cache) * (mp_maxid + 1));
1718 	args.ctor = zone_ctor;
1719 	args.dtor = zone_dtor;
1720 	args.uminit = zero_init;
1721 	args.fini = NULL;
1722 	args.keg = NULL;
1723 	args.align = 32 - 1;
1724 	args.flags = UMA_ZFLAG_INTERNAL;
1725 	/* The initial zone has no Per cpu queues so it's smaller */
1726 	zone_ctor(zones, sizeof(struct uma_zone), &args, M_WAITOK);
1727 
1728 #ifdef UMA_DEBUG
1729 	printf("Initializing pcpu cache locks.\n");
1730 #endif
1731 #ifdef UMA_DEBUG
1732 	printf("Creating slab and hash zones.\n");
1733 #endif
1734 
1735 	/*
1736 	 * This is the max number of free list items we'll have with
1737 	 * offpage slabs.
1738 	 */
1739 	slabsize = uma_max_ipers * UMA_FRITM_SZ;
1740 	slabsize += sizeof(struct uma_slab);
1741 
1742 	/* Now make a zone for slab headers */
1743 	slabzone = uma_zcreate("UMA Slabs",
1744 				slabsize,
1745 				NULL, NULL, NULL, NULL,
1746 				UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1747 
1748 	/*
1749 	 * We also create a zone for the bigger slabs with reference
1750 	 * counts in them, to accomodate UMA_ZONE_REFCNT zones.
1751 	 */
1752 	slabsize = uma_max_ipers_ref * UMA_FRITMREF_SZ;
1753 	slabsize += sizeof(struct uma_slab_refcnt);
1754 	slabrefzone = uma_zcreate("UMA RCntSlabs",
1755 				  slabsize,
1756 				  NULL, NULL, NULL, NULL,
1757 				  UMA_ALIGN_PTR,
1758 				  UMA_ZFLAG_INTERNAL);
1759 
1760 	hashzone = uma_zcreate("UMA Hash",
1761 	    sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
1762 	    NULL, NULL, NULL, NULL,
1763 	    UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1764 
1765 	bucket_init();
1766 
1767 	booted = UMA_STARTUP;
1768 
1769 #ifdef UMA_DEBUG
1770 	printf("UMA startup complete.\n");
1771 #endif
1772 }
1773 
1774 /* see uma.h */
1775 void
1776 uma_startup2(void)
1777 {
1778 	booted = UMA_STARTUP2;
1779 	bucket_enable();
1780 #ifdef UMA_DEBUG
1781 	printf("UMA startup2 complete.\n");
1782 #endif
1783 }
1784 
1785 /*
1786  * Initialize our callout handle
1787  *
1788  */
1789 
1790 static void
1791 uma_startup3(void)
1792 {
1793 #ifdef UMA_DEBUG
1794 	printf("Starting callout.\n");
1795 #endif
1796 	callout_init(&uma_callout, CALLOUT_MPSAFE);
1797 	callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
1798 #ifdef UMA_DEBUG
1799 	printf("UMA startup3 complete.\n");
1800 #endif
1801 }
1802 
1803 static uma_keg_t
1804 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
1805 		int align, u_int32_t flags)
1806 {
1807 	struct uma_kctor_args args;
1808 
1809 	args.size = size;
1810 	args.uminit = uminit;
1811 	args.fini = fini;
1812 	args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
1813 	args.flags = flags;
1814 	args.zone = zone;
1815 	return (zone_alloc_item(kegs, &args, M_WAITOK));
1816 }
1817 
1818 /* See uma.h */
1819 void
1820 uma_set_align(int align)
1821 {
1822 
1823 	if (align != UMA_ALIGN_CACHE)
1824 		uma_align_cache = align;
1825 }
1826 
1827 /* See uma.h */
1828 uma_zone_t
1829 uma_zcreate(char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
1830 		uma_init uminit, uma_fini fini, int align, u_int32_t flags)
1831 
1832 {
1833 	struct uma_zctor_args args;
1834 
1835 	/* This stuff is essential for the zone ctor */
1836 	args.name = name;
1837 	args.size = size;
1838 	args.ctor = ctor;
1839 	args.dtor = dtor;
1840 	args.uminit = uminit;
1841 	args.fini = fini;
1842 	args.align = align;
1843 	args.flags = flags;
1844 	args.keg = NULL;
1845 
1846 	return (zone_alloc_item(zones, &args, M_WAITOK));
1847 }
1848 
1849 /* See uma.h */
1850 uma_zone_t
1851 uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor,
1852 		    uma_init zinit, uma_fini zfini, uma_zone_t master)
1853 {
1854 	struct uma_zctor_args args;
1855 	uma_keg_t keg;
1856 
1857 	keg = zone_first_keg(master);
1858 	args.name = name;
1859 	args.size = keg->uk_size;
1860 	args.ctor = ctor;
1861 	args.dtor = dtor;
1862 	args.uminit = zinit;
1863 	args.fini = zfini;
1864 	args.align = keg->uk_align;
1865 	args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
1866 	args.keg = keg;
1867 
1868 	/* XXX Attaches only one keg of potentially many. */
1869 	return (zone_alloc_item(zones, &args, M_WAITOK));
1870 }
1871 
1872 static void
1873 zone_lock_pair(uma_zone_t a, uma_zone_t b)
1874 {
1875 	if (a < b) {
1876 		ZONE_LOCK(a);
1877 		mtx_lock_flags(b->uz_lock, MTX_DUPOK);
1878 	} else {
1879 		ZONE_LOCK(b);
1880 		mtx_lock_flags(a->uz_lock, MTX_DUPOK);
1881 	}
1882 }
1883 
1884 static void
1885 zone_unlock_pair(uma_zone_t a, uma_zone_t b)
1886 {
1887 
1888 	ZONE_UNLOCK(a);
1889 	ZONE_UNLOCK(b);
1890 }
1891 
1892 int
1893 uma_zsecond_add(uma_zone_t zone, uma_zone_t master)
1894 {
1895 	uma_klink_t klink;
1896 	uma_klink_t kl;
1897 	int error;
1898 
1899 	error = 0;
1900 	klink = malloc(sizeof(*klink), M_TEMP, M_WAITOK | M_ZERO);
1901 
1902 	zone_lock_pair(zone, master);
1903 	/*
1904 	 * zone must use vtoslab() to resolve objects and must already be
1905 	 * a secondary.
1906 	 */
1907 	if ((zone->uz_flags & (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY))
1908 	    != (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) {
1909 		error = EINVAL;
1910 		goto out;
1911 	}
1912 	/*
1913 	 * The new master must also use vtoslab().
1914 	 */
1915 	if ((zone->uz_flags & UMA_ZONE_VTOSLAB) != UMA_ZONE_VTOSLAB) {
1916 		error = EINVAL;
1917 		goto out;
1918 	}
1919 	/*
1920 	 * Both must either be refcnt, or not be refcnt.
1921 	 */
1922 	if ((zone->uz_flags & UMA_ZONE_REFCNT) !=
1923 	    (master->uz_flags & UMA_ZONE_REFCNT)) {
1924 		error = EINVAL;
1925 		goto out;
1926 	}
1927 	/*
1928 	 * The underlying object must be the same size.  rsize
1929 	 * may be different.
1930 	 */
1931 	if (master->uz_size != zone->uz_size) {
1932 		error = E2BIG;
1933 		goto out;
1934 	}
1935 	/*
1936 	 * Put it at the end of the list.
1937 	 */
1938 	klink->kl_keg = zone_first_keg(master);
1939 	LIST_FOREACH(kl, &zone->uz_kegs, kl_link) {
1940 		if (LIST_NEXT(kl, kl_link) == NULL) {
1941 			LIST_INSERT_AFTER(kl, klink, kl_link);
1942 			break;
1943 		}
1944 	}
1945 	klink = NULL;
1946 	zone->uz_flags |= UMA_ZFLAG_MULTI;
1947 	zone->uz_slab = zone_fetch_slab_multi;
1948 
1949 out:
1950 	zone_unlock_pair(zone, master);
1951 	if (klink != NULL)
1952 		free(klink, M_TEMP);
1953 
1954 	return (error);
1955 }
1956 
1957 
1958 /* See uma.h */
1959 void
1960 uma_zdestroy(uma_zone_t zone)
1961 {
1962 
1963 	zone_free_item(zones, zone, NULL, SKIP_NONE, ZFREE_STATFREE);
1964 }
1965 
1966 /* See uma.h */
1967 void *
1968 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
1969 {
1970 	void *item;
1971 	uma_cache_t cache;
1972 	uma_bucket_t bucket;
1973 	int cpu;
1974 
1975 	/* This is the fast path allocation */
1976 #ifdef UMA_DEBUG_ALLOC_1
1977 	printf("Allocating one item from %s(%p)\n", zone->uz_name, zone);
1978 #endif
1979 	CTR3(KTR_UMA, "uma_zalloc_arg thread %x zone %s flags %d", curthread,
1980 	    zone->uz_name, flags);
1981 
1982 	if (flags & M_WAITOK) {
1983 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
1984 		    "uma_zalloc_arg: zone \"%s\"", zone->uz_name);
1985 	}
1986 #ifdef DEBUG_MEMGUARD
1987 	if (memguard_cmp_zone(zone)) {
1988 		item = memguard_alloc(zone->uz_size, flags);
1989 		if (item != NULL) {
1990 			/*
1991 			 * Avoid conflict with the use-after-free
1992 			 * protecting infrastructure from INVARIANTS.
1993 			 */
1994 			if (zone->uz_init != NULL &&
1995 			    zone->uz_init != mtrash_init &&
1996 			    zone->uz_init(item, zone->uz_size, flags) != 0)
1997 				return (NULL);
1998 			if (zone->uz_ctor != NULL &&
1999 			    zone->uz_ctor != mtrash_ctor &&
2000 			    zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2001 			    	zone->uz_fini(item, zone->uz_size);
2002 				return (NULL);
2003 			}
2004 			return (item);
2005 		}
2006 		/* This is unfortunate but should not be fatal. */
2007 	}
2008 #endif
2009 	/*
2010 	 * If possible, allocate from the per-CPU cache.  There are two
2011 	 * requirements for safe access to the per-CPU cache: (1) the thread
2012 	 * accessing the cache must not be preempted or yield during access,
2013 	 * and (2) the thread must not migrate CPUs without switching which
2014 	 * cache it accesses.  We rely on a critical section to prevent
2015 	 * preemption and migration.  We release the critical section in
2016 	 * order to acquire the zone mutex if we are unable to allocate from
2017 	 * the current cache; when we re-acquire the critical section, we
2018 	 * must detect and handle migration if it has occurred.
2019 	 */
2020 zalloc_restart:
2021 	critical_enter();
2022 	cpu = curcpu;
2023 	cache = &zone->uz_cpu[cpu];
2024 
2025 zalloc_start:
2026 	bucket = cache->uc_allocbucket;
2027 
2028 	if (bucket) {
2029 		if (bucket->ub_cnt > 0) {
2030 			bucket->ub_cnt--;
2031 			item = bucket->ub_bucket[bucket->ub_cnt];
2032 #ifdef INVARIANTS
2033 			bucket->ub_bucket[bucket->ub_cnt] = NULL;
2034 #endif
2035 			KASSERT(item != NULL,
2036 			    ("uma_zalloc: Bucket pointer mangled."));
2037 			cache->uc_allocs++;
2038 			critical_exit();
2039 #ifdef INVARIANTS
2040 			ZONE_LOCK(zone);
2041 			uma_dbg_alloc(zone, NULL, item);
2042 			ZONE_UNLOCK(zone);
2043 #endif
2044 			if (zone->uz_ctor != NULL) {
2045 				if (zone->uz_ctor(item, zone->uz_size,
2046 				    udata, flags) != 0) {
2047 					zone_free_item(zone, item, udata,
2048 					    SKIP_DTOR, ZFREE_STATFAIL |
2049 					    ZFREE_STATFREE);
2050 					return (NULL);
2051 				}
2052 			}
2053 			if (flags & M_ZERO)
2054 				bzero(item, zone->uz_size);
2055 			return (item);
2056 		} else if (cache->uc_freebucket) {
2057 			/*
2058 			 * We have run out of items in our allocbucket.
2059 			 * See if we can switch with our free bucket.
2060 			 */
2061 			if (cache->uc_freebucket->ub_cnt > 0) {
2062 #ifdef UMA_DEBUG_ALLOC
2063 				printf("uma_zalloc: Swapping empty with"
2064 				    " alloc.\n");
2065 #endif
2066 				bucket = cache->uc_freebucket;
2067 				cache->uc_freebucket = cache->uc_allocbucket;
2068 				cache->uc_allocbucket = bucket;
2069 
2070 				goto zalloc_start;
2071 			}
2072 		}
2073 	}
2074 	/*
2075 	 * Attempt to retrieve the item from the per-CPU cache has failed, so
2076 	 * we must go back to the zone.  This requires the zone lock, so we
2077 	 * must drop the critical section, then re-acquire it when we go back
2078 	 * to the cache.  Since the critical section is released, we may be
2079 	 * preempted or migrate.  As such, make sure not to maintain any
2080 	 * thread-local state specific to the cache from prior to releasing
2081 	 * the critical section.
2082 	 */
2083 	critical_exit();
2084 	ZONE_LOCK(zone);
2085 	critical_enter();
2086 	cpu = curcpu;
2087 	cache = &zone->uz_cpu[cpu];
2088 	bucket = cache->uc_allocbucket;
2089 	if (bucket != NULL) {
2090 		if (bucket->ub_cnt > 0) {
2091 			ZONE_UNLOCK(zone);
2092 			goto zalloc_start;
2093 		}
2094 		bucket = cache->uc_freebucket;
2095 		if (bucket != NULL && bucket->ub_cnt > 0) {
2096 			ZONE_UNLOCK(zone);
2097 			goto zalloc_start;
2098 		}
2099 	}
2100 
2101 	/* Since we have locked the zone we may as well send back our stats */
2102 	zone->uz_allocs += cache->uc_allocs;
2103 	cache->uc_allocs = 0;
2104 	zone->uz_frees += cache->uc_frees;
2105 	cache->uc_frees = 0;
2106 
2107 	/* Our old one is now a free bucket */
2108 	if (cache->uc_allocbucket) {
2109 		KASSERT(cache->uc_allocbucket->ub_cnt == 0,
2110 		    ("uma_zalloc_arg: Freeing a non free bucket."));
2111 		LIST_INSERT_HEAD(&zone->uz_free_bucket,
2112 		    cache->uc_allocbucket, ub_link);
2113 		cache->uc_allocbucket = NULL;
2114 	}
2115 
2116 	/* Check the free list for a new alloc bucket */
2117 	if ((bucket = LIST_FIRST(&zone->uz_full_bucket)) != NULL) {
2118 		KASSERT(bucket->ub_cnt != 0,
2119 		    ("uma_zalloc_arg: Returning an empty bucket."));
2120 
2121 		LIST_REMOVE(bucket, ub_link);
2122 		cache->uc_allocbucket = bucket;
2123 		ZONE_UNLOCK(zone);
2124 		goto zalloc_start;
2125 	}
2126 	/* We are no longer associated with this CPU. */
2127 	critical_exit();
2128 
2129 	/* Bump up our uz_count so we get here less */
2130 	if (zone->uz_count < BUCKET_MAX)
2131 		zone->uz_count++;
2132 
2133 	/*
2134 	 * Now lets just fill a bucket and put it on the free list.  If that
2135 	 * works we'll restart the allocation from the begining.
2136 	 */
2137 	if (zone_alloc_bucket(zone, flags)) {
2138 		ZONE_UNLOCK(zone);
2139 		goto zalloc_restart;
2140 	}
2141 	ZONE_UNLOCK(zone);
2142 	/*
2143 	 * We may not be able to get a bucket so return an actual item.
2144 	 */
2145 #ifdef UMA_DEBUG
2146 	printf("uma_zalloc_arg: Bucketzone returned NULL\n");
2147 #endif
2148 
2149 	item = zone_alloc_item(zone, udata, flags);
2150 	return (item);
2151 }
2152 
2153 static uma_slab_t
2154 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int flags)
2155 {
2156 	uma_slab_t slab;
2157 
2158 	mtx_assert(&keg->uk_lock, MA_OWNED);
2159 	slab = NULL;
2160 
2161 	for (;;) {
2162 		/*
2163 		 * Find a slab with some space.  Prefer slabs that are partially
2164 		 * used over those that are totally full.  This helps to reduce
2165 		 * fragmentation.
2166 		 */
2167 		if (keg->uk_free != 0) {
2168 			if (!LIST_EMPTY(&keg->uk_part_slab)) {
2169 				slab = LIST_FIRST(&keg->uk_part_slab);
2170 			} else {
2171 				slab = LIST_FIRST(&keg->uk_free_slab);
2172 				LIST_REMOVE(slab, us_link);
2173 				LIST_INSERT_HEAD(&keg->uk_part_slab, slab,
2174 				    us_link);
2175 			}
2176 			MPASS(slab->us_keg == keg);
2177 			return (slab);
2178 		}
2179 
2180 		/*
2181 		 * M_NOVM means don't ask at all!
2182 		 */
2183 		if (flags & M_NOVM)
2184 			break;
2185 
2186 		if (keg->uk_maxpages && keg->uk_pages >= keg->uk_maxpages) {
2187 			keg->uk_flags |= UMA_ZFLAG_FULL;
2188 			/*
2189 			 * If this is not a multi-zone, set the FULL bit.
2190 			 * Otherwise slab_multi() takes care of it.
2191 			 */
2192 			if ((zone->uz_flags & UMA_ZFLAG_MULTI) == 0)
2193 				zone->uz_flags |= UMA_ZFLAG_FULL;
2194 			if (flags & M_NOWAIT)
2195 				break;
2196 			msleep(keg, &keg->uk_lock, PVM, "keglimit", 0);
2197 			continue;
2198 		}
2199 		keg->uk_recurse++;
2200 		slab = keg_alloc_slab(keg, zone, flags);
2201 		keg->uk_recurse--;
2202 		/*
2203 		 * If we got a slab here it's safe to mark it partially used
2204 		 * and return.  We assume that the caller is going to remove
2205 		 * at least one item.
2206 		 */
2207 		if (slab) {
2208 			MPASS(slab->us_keg == keg);
2209 			LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2210 			return (slab);
2211 		}
2212 		/*
2213 		 * We might not have been able to get a slab but another cpu
2214 		 * could have while we were unlocked.  Check again before we
2215 		 * fail.
2216 		 */
2217 		flags |= M_NOVM;
2218 	}
2219 	return (slab);
2220 }
2221 
2222 static inline void
2223 zone_relock(uma_zone_t zone, uma_keg_t keg)
2224 {
2225 	if (zone->uz_lock != &keg->uk_lock) {
2226 		KEG_UNLOCK(keg);
2227 		ZONE_LOCK(zone);
2228 	}
2229 }
2230 
2231 static inline void
2232 keg_relock(uma_keg_t keg, uma_zone_t zone)
2233 {
2234 	if (zone->uz_lock != &keg->uk_lock) {
2235 		ZONE_UNLOCK(zone);
2236 		KEG_LOCK(keg);
2237 	}
2238 }
2239 
2240 static uma_slab_t
2241 zone_fetch_slab(uma_zone_t zone, uma_keg_t keg, int flags)
2242 {
2243 	uma_slab_t slab;
2244 
2245 	if (keg == NULL)
2246 		keg = zone_first_keg(zone);
2247 	/*
2248 	 * This is to prevent us from recursively trying to allocate
2249 	 * buckets.  The problem is that if an allocation forces us to
2250 	 * grab a new bucket we will call page_alloc, which will go off
2251 	 * and cause the vm to allocate vm_map_entries.  If we need new
2252 	 * buckets there too we will recurse in kmem_alloc and bad
2253 	 * things happen.  So instead we return a NULL bucket, and make
2254 	 * the code that allocates buckets smart enough to deal with it
2255 	 */
2256 	if (keg->uk_flags & UMA_ZFLAG_BUCKET && keg->uk_recurse != 0)
2257 		return (NULL);
2258 
2259 	for (;;) {
2260 		slab = keg_fetch_slab(keg, zone, flags);
2261 		if (slab)
2262 			return (slab);
2263 		if (flags & (M_NOWAIT | M_NOVM))
2264 			break;
2265 	}
2266 	return (NULL);
2267 }
2268 
2269 /*
2270  * uma_zone_fetch_slab_multi:  Fetches a slab from one available keg.  Returns
2271  * with the keg locked.  Caller must call zone_relock() afterwards if the
2272  * zone lock is required.  On NULL the zone lock is held.
2273  *
2274  * The last pointer is used to seed the search.  It is not required.
2275  */
2276 static uma_slab_t
2277 zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int rflags)
2278 {
2279 	uma_klink_t klink;
2280 	uma_slab_t slab;
2281 	uma_keg_t keg;
2282 	int flags;
2283 	int empty;
2284 	int full;
2285 
2286 	/*
2287 	 * Don't wait on the first pass.  This will skip limit tests
2288 	 * as well.  We don't want to block if we can find a provider
2289 	 * without blocking.
2290 	 */
2291 	flags = (rflags & ~M_WAITOK) | M_NOWAIT;
2292 	/*
2293 	 * Use the last slab allocated as a hint for where to start
2294 	 * the search.
2295 	 */
2296 	if (last) {
2297 		slab = keg_fetch_slab(last, zone, flags);
2298 		if (slab)
2299 			return (slab);
2300 		zone_relock(zone, last);
2301 		last = NULL;
2302 	}
2303 	/*
2304 	 * Loop until we have a slab incase of transient failures
2305 	 * while M_WAITOK is specified.  I'm not sure this is 100%
2306 	 * required but we've done it for so long now.
2307 	 */
2308 	for (;;) {
2309 		empty = 0;
2310 		full = 0;
2311 		/*
2312 		 * Search the available kegs for slabs.  Be careful to hold the
2313 		 * correct lock while calling into the keg layer.
2314 		 */
2315 		LIST_FOREACH(klink, &zone->uz_kegs, kl_link) {
2316 			keg = klink->kl_keg;
2317 			keg_relock(keg, zone);
2318 			if ((keg->uk_flags & UMA_ZFLAG_FULL) == 0) {
2319 				slab = keg_fetch_slab(keg, zone, flags);
2320 				if (slab)
2321 					return (slab);
2322 			}
2323 			if (keg->uk_flags & UMA_ZFLAG_FULL)
2324 				full++;
2325 			else
2326 				empty++;
2327 			zone_relock(zone, keg);
2328 		}
2329 		if (rflags & (M_NOWAIT | M_NOVM))
2330 			break;
2331 		flags = rflags;
2332 		/*
2333 		 * All kegs are full.  XXX We can't atomically check all kegs
2334 		 * and sleep so just sleep for a short period and retry.
2335 		 */
2336 		if (full && !empty) {
2337 			zone->uz_flags |= UMA_ZFLAG_FULL;
2338 			zone->uz_sleeps++;
2339 			msleep(zone, zone->uz_lock, PVM, "zonelimit", hz/100);
2340 			zone->uz_flags &= ~UMA_ZFLAG_FULL;
2341 			continue;
2342 		}
2343 	}
2344 	return (NULL);
2345 }
2346 
2347 static void *
2348 slab_alloc_item(uma_zone_t zone, uma_slab_t slab)
2349 {
2350 	uma_keg_t keg;
2351 	uma_slabrefcnt_t slabref;
2352 	void *item;
2353 	u_int8_t freei;
2354 
2355 	keg = slab->us_keg;
2356 	mtx_assert(&keg->uk_lock, MA_OWNED);
2357 
2358 	freei = slab->us_firstfree;
2359 	if (keg->uk_flags & UMA_ZONE_REFCNT) {
2360 		slabref = (uma_slabrefcnt_t)slab;
2361 		slab->us_firstfree = slabref->us_freelist[freei].us_item;
2362 	} else {
2363 		slab->us_firstfree = slab->us_freelist[freei].us_item;
2364 	}
2365 	item = slab->us_data + (keg->uk_rsize * freei);
2366 
2367 	slab->us_freecount--;
2368 	keg->uk_free--;
2369 #ifdef INVARIANTS
2370 	uma_dbg_alloc(zone, slab, item);
2371 #endif
2372 	/* Move this slab to the full list */
2373 	if (slab->us_freecount == 0) {
2374 		LIST_REMOVE(slab, us_link);
2375 		LIST_INSERT_HEAD(&keg->uk_full_slab, slab, us_link);
2376 	}
2377 
2378 	return (item);
2379 }
2380 
2381 static int
2382 zone_alloc_bucket(uma_zone_t zone, int flags)
2383 {
2384 	uma_bucket_t bucket;
2385 	uma_slab_t slab;
2386 	uma_keg_t keg;
2387 	int16_t saved;
2388 	int max, origflags = flags;
2389 
2390 	/*
2391 	 * Try this zone's free list first so we don't allocate extra buckets.
2392 	 */
2393 	if ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) {
2394 		KASSERT(bucket->ub_cnt == 0,
2395 		    ("zone_alloc_bucket: Bucket on free list is not empty."));
2396 		LIST_REMOVE(bucket, ub_link);
2397 	} else {
2398 		int bflags;
2399 
2400 		bflags = (flags & ~M_ZERO);
2401 		if (zone->uz_flags & UMA_ZFLAG_CACHEONLY)
2402 			bflags |= M_NOVM;
2403 
2404 		ZONE_UNLOCK(zone);
2405 		bucket = bucket_alloc(zone->uz_count, bflags);
2406 		ZONE_LOCK(zone);
2407 	}
2408 
2409 	if (bucket == NULL) {
2410 		return (0);
2411 	}
2412 
2413 #ifdef SMP
2414 	/*
2415 	 * This code is here to limit the number of simultaneous bucket fills
2416 	 * for any given zone to the number of per cpu caches in this zone. This
2417 	 * is done so that we don't allocate more memory than we really need.
2418 	 */
2419 	if (zone->uz_fills >= mp_ncpus)
2420 		goto done;
2421 
2422 #endif
2423 	zone->uz_fills++;
2424 
2425 	max = MIN(bucket->ub_entries, zone->uz_count);
2426 	/* Try to keep the buckets totally full */
2427 	saved = bucket->ub_cnt;
2428 	slab = NULL;
2429 	keg = NULL;
2430 	while (bucket->ub_cnt < max &&
2431 	    (slab = zone->uz_slab(zone, keg, flags)) != NULL) {
2432 		keg = slab->us_keg;
2433 		while (slab->us_freecount && bucket->ub_cnt < max) {
2434 			bucket->ub_bucket[bucket->ub_cnt++] =
2435 			    slab_alloc_item(zone, slab);
2436 		}
2437 
2438 		/* Don't block on the next fill */
2439 		flags |= M_NOWAIT;
2440 	}
2441 	if (slab)
2442 		zone_relock(zone, keg);
2443 
2444 	/*
2445 	 * We unlock here because we need to call the zone's init.
2446 	 * It should be safe to unlock because the slab dealt with
2447 	 * above is already on the appropriate list within the keg
2448 	 * and the bucket we filled is not yet on any list, so we
2449 	 * own it.
2450 	 */
2451 	if (zone->uz_init != NULL) {
2452 		int i;
2453 
2454 		ZONE_UNLOCK(zone);
2455 		for (i = saved; i < bucket->ub_cnt; i++)
2456 			if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size,
2457 			    origflags) != 0)
2458 				break;
2459 		/*
2460 		 * If we couldn't initialize the whole bucket, put the
2461 		 * rest back onto the freelist.
2462 		 */
2463 		if (i != bucket->ub_cnt) {
2464 			int j;
2465 
2466 			for (j = i; j < bucket->ub_cnt; j++) {
2467 				zone_free_item(zone, bucket->ub_bucket[j],
2468 				    NULL, SKIP_FINI, 0);
2469 #ifdef INVARIANTS
2470 				bucket->ub_bucket[j] = NULL;
2471 #endif
2472 			}
2473 			bucket->ub_cnt = i;
2474 		}
2475 		ZONE_LOCK(zone);
2476 	}
2477 
2478 	zone->uz_fills--;
2479 	if (bucket->ub_cnt != 0) {
2480 		LIST_INSERT_HEAD(&zone->uz_full_bucket,
2481 		    bucket, ub_link);
2482 		return (1);
2483 	}
2484 #ifdef SMP
2485 done:
2486 #endif
2487 	bucket_free(bucket);
2488 
2489 	return (0);
2490 }
2491 /*
2492  * Allocates an item for an internal zone
2493  *
2494  * Arguments
2495  *	zone   The zone to alloc for.
2496  *	udata  The data to be passed to the constructor.
2497  *	flags  M_WAITOK, M_NOWAIT, M_ZERO.
2498  *
2499  * Returns
2500  *	NULL if there is no memory and M_NOWAIT is set
2501  *	An item if successful
2502  */
2503 
2504 static void *
2505 zone_alloc_item(uma_zone_t zone, void *udata, int flags)
2506 {
2507 	uma_slab_t slab;
2508 	void *item;
2509 
2510 	item = NULL;
2511 
2512 #ifdef UMA_DEBUG_ALLOC
2513 	printf("INTERNAL: Allocating one item from %s(%p)\n", zone->uz_name, zone);
2514 #endif
2515 	ZONE_LOCK(zone);
2516 
2517 	slab = zone->uz_slab(zone, NULL, flags);
2518 	if (slab == NULL) {
2519 		zone->uz_fails++;
2520 		ZONE_UNLOCK(zone);
2521 		return (NULL);
2522 	}
2523 
2524 	item = slab_alloc_item(zone, slab);
2525 
2526 	zone_relock(zone, slab->us_keg);
2527 	zone->uz_allocs++;
2528 	ZONE_UNLOCK(zone);
2529 
2530 	/*
2531 	 * We have to call both the zone's init (not the keg's init)
2532 	 * and the zone's ctor.  This is because the item is going from
2533 	 * a keg slab directly to the user, and the user is expecting it
2534 	 * to be both zone-init'd as well as zone-ctor'd.
2535 	 */
2536 	if (zone->uz_init != NULL) {
2537 		if (zone->uz_init(item, zone->uz_size, flags) != 0) {
2538 			zone_free_item(zone, item, udata, SKIP_FINI,
2539 			    ZFREE_STATFAIL | ZFREE_STATFREE);
2540 			return (NULL);
2541 		}
2542 	}
2543 	if (zone->uz_ctor != NULL) {
2544 		if (zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2545 			zone_free_item(zone, item, udata, SKIP_DTOR,
2546 			    ZFREE_STATFAIL | ZFREE_STATFREE);
2547 			return (NULL);
2548 		}
2549 	}
2550 	if (flags & M_ZERO)
2551 		bzero(item, zone->uz_size);
2552 
2553 	return (item);
2554 }
2555 
2556 /* See uma.h */
2557 void
2558 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
2559 {
2560 	uma_cache_t cache;
2561 	uma_bucket_t bucket;
2562 	int bflags;
2563 	int cpu;
2564 
2565 #ifdef UMA_DEBUG_ALLOC_1
2566 	printf("Freeing item %p to %s(%p)\n", item, zone->uz_name, zone);
2567 #endif
2568 	CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread,
2569 	    zone->uz_name);
2570 
2571         /* uma_zfree(..., NULL) does nothing, to match free(9). */
2572         if (item == NULL)
2573                 return;
2574 #ifdef DEBUG_MEMGUARD
2575 	if (is_memguard_addr(item)) {
2576 		if (zone->uz_dtor != NULL && zone->uz_dtor != mtrash_dtor)
2577 			zone->uz_dtor(item, zone->uz_size, udata);
2578 		if (zone->uz_fini != NULL && zone->uz_fini != mtrash_fini)
2579 			zone->uz_fini(item, zone->uz_size);
2580 		memguard_free(item);
2581 		return;
2582 	}
2583 #endif
2584 	if (zone->uz_dtor)
2585 		zone->uz_dtor(item, zone->uz_size, udata);
2586 
2587 #ifdef INVARIANTS
2588 	ZONE_LOCK(zone);
2589 	if (zone->uz_flags & UMA_ZONE_MALLOC)
2590 		uma_dbg_free(zone, udata, item);
2591 	else
2592 		uma_dbg_free(zone, NULL, item);
2593 	ZONE_UNLOCK(zone);
2594 #endif
2595 	/*
2596 	 * The race here is acceptable.  If we miss it we'll just have to wait
2597 	 * a little longer for the limits to be reset.
2598 	 */
2599 	if (zone->uz_flags & UMA_ZFLAG_FULL)
2600 		goto zfree_internal;
2601 
2602 	/*
2603 	 * If possible, free to the per-CPU cache.  There are two
2604 	 * requirements for safe access to the per-CPU cache: (1) the thread
2605 	 * accessing the cache must not be preempted or yield during access,
2606 	 * and (2) the thread must not migrate CPUs without switching which
2607 	 * cache it accesses.  We rely on a critical section to prevent
2608 	 * preemption and migration.  We release the critical section in
2609 	 * order to acquire the zone mutex if we are unable to free to the
2610 	 * current cache; when we re-acquire the critical section, we must
2611 	 * detect and handle migration if it has occurred.
2612 	 */
2613 zfree_restart:
2614 	critical_enter();
2615 	cpu = curcpu;
2616 	cache = &zone->uz_cpu[cpu];
2617 
2618 zfree_start:
2619 	bucket = cache->uc_freebucket;
2620 
2621 	if (bucket) {
2622 		/*
2623 		 * Do we have room in our bucket? It is OK for this uz count
2624 		 * check to be slightly out of sync.
2625 		 */
2626 
2627 		if (bucket->ub_cnt < bucket->ub_entries) {
2628 			KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL,
2629 			    ("uma_zfree: Freeing to non free bucket index."));
2630 			bucket->ub_bucket[bucket->ub_cnt] = item;
2631 			bucket->ub_cnt++;
2632 			cache->uc_frees++;
2633 			critical_exit();
2634 			return;
2635 		} else if (cache->uc_allocbucket) {
2636 #ifdef UMA_DEBUG_ALLOC
2637 			printf("uma_zfree: Swapping buckets.\n");
2638 #endif
2639 			/*
2640 			 * We have run out of space in our freebucket.
2641 			 * See if we can switch with our alloc bucket.
2642 			 */
2643 			if (cache->uc_allocbucket->ub_cnt <
2644 			    cache->uc_freebucket->ub_cnt) {
2645 				bucket = cache->uc_freebucket;
2646 				cache->uc_freebucket = cache->uc_allocbucket;
2647 				cache->uc_allocbucket = bucket;
2648 				goto zfree_start;
2649 			}
2650 		}
2651 	}
2652 	/*
2653 	 * We can get here for two reasons:
2654 	 *
2655 	 * 1) The buckets are NULL
2656 	 * 2) The alloc and free buckets are both somewhat full.
2657 	 *
2658 	 * We must go back the zone, which requires acquiring the zone lock,
2659 	 * which in turn means we must release and re-acquire the critical
2660 	 * section.  Since the critical section is released, we may be
2661 	 * preempted or migrate.  As such, make sure not to maintain any
2662 	 * thread-local state specific to the cache from prior to releasing
2663 	 * the critical section.
2664 	 */
2665 	critical_exit();
2666 	ZONE_LOCK(zone);
2667 	critical_enter();
2668 	cpu = curcpu;
2669 	cache = &zone->uz_cpu[cpu];
2670 	if (cache->uc_freebucket != NULL) {
2671 		if (cache->uc_freebucket->ub_cnt <
2672 		    cache->uc_freebucket->ub_entries) {
2673 			ZONE_UNLOCK(zone);
2674 			goto zfree_start;
2675 		}
2676 		if (cache->uc_allocbucket != NULL &&
2677 		    (cache->uc_allocbucket->ub_cnt <
2678 		    cache->uc_freebucket->ub_cnt)) {
2679 			ZONE_UNLOCK(zone);
2680 			goto zfree_start;
2681 		}
2682 	}
2683 
2684 	/* Since we have locked the zone we may as well send back our stats */
2685 	zone->uz_allocs += cache->uc_allocs;
2686 	cache->uc_allocs = 0;
2687 	zone->uz_frees += cache->uc_frees;
2688 	cache->uc_frees = 0;
2689 
2690 	bucket = cache->uc_freebucket;
2691 	cache->uc_freebucket = NULL;
2692 
2693 	/* Can we throw this on the zone full list? */
2694 	if (bucket != NULL) {
2695 #ifdef UMA_DEBUG_ALLOC
2696 		printf("uma_zfree: Putting old bucket on the free list.\n");
2697 #endif
2698 		/* ub_cnt is pointing to the last free item */
2699 		KASSERT(bucket->ub_cnt != 0,
2700 		    ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n"));
2701 		LIST_INSERT_HEAD(&zone->uz_full_bucket,
2702 		    bucket, ub_link);
2703 	}
2704 	if ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) {
2705 		LIST_REMOVE(bucket, ub_link);
2706 		ZONE_UNLOCK(zone);
2707 		cache->uc_freebucket = bucket;
2708 		goto zfree_start;
2709 	}
2710 	/* We are no longer associated with this CPU. */
2711 	critical_exit();
2712 
2713 	/* And the zone.. */
2714 	ZONE_UNLOCK(zone);
2715 
2716 #ifdef UMA_DEBUG_ALLOC
2717 	printf("uma_zfree: Allocating new free bucket.\n");
2718 #endif
2719 	bflags = M_NOWAIT;
2720 
2721 	if (zone->uz_flags & UMA_ZFLAG_CACHEONLY)
2722 		bflags |= M_NOVM;
2723 	bucket = bucket_alloc(zone->uz_count, bflags);
2724 	if (bucket) {
2725 		ZONE_LOCK(zone);
2726 		LIST_INSERT_HEAD(&zone->uz_free_bucket,
2727 		    bucket, ub_link);
2728 		ZONE_UNLOCK(zone);
2729 		goto zfree_restart;
2730 	}
2731 
2732 	/*
2733 	 * If nothing else caught this, we'll just do an internal free.
2734 	 */
2735 zfree_internal:
2736 	zone_free_item(zone, item, udata, SKIP_DTOR, ZFREE_STATFREE);
2737 
2738 	return;
2739 }
2740 
2741 /*
2742  * Frees an item to an INTERNAL zone or allocates a free bucket
2743  *
2744  * Arguments:
2745  *	zone   The zone to free to
2746  *	item   The item we're freeing
2747  *	udata  User supplied data for the dtor
2748  *	skip   Skip dtors and finis
2749  */
2750 static void
2751 zone_free_item(uma_zone_t zone, void *item, void *udata,
2752     enum zfreeskip skip, int flags)
2753 {
2754 	uma_slab_t slab;
2755 	uma_slabrefcnt_t slabref;
2756 	uma_keg_t keg;
2757 	u_int8_t *mem;
2758 	u_int8_t freei;
2759 	int clearfull;
2760 
2761 	if (skip < SKIP_DTOR && zone->uz_dtor)
2762 		zone->uz_dtor(item, zone->uz_size, udata);
2763 
2764 	if (skip < SKIP_FINI && zone->uz_fini)
2765 		zone->uz_fini(item, zone->uz_size);
2766 
2767 	ZONE_LOCK(zone);
2768 
2769 	if (flags & ZFREE_STATFAIL)
2770 		zone->uz_fails++;
2771 	if (flags & ZFREE_STATFREE)
2772 		zone->uz_frees++;
2773 
2774 	if (!(zone->uz_flags & UMA_ZONE_VTOSLAB)) {
2775 		mem = (u_int8_t *)((unsigned long)item & (~UMA_SLAB_MASK));
2776 		keg = zone_first_keg(zone); /* Must only be one. */
2777 		if (zone->uz_flags & UMA_ZONE_HASH) {
2778 			slab = hash_sfind(&keg->uk_hash, mem);
2779 		} else {
2780 			mem += keg->uk_pgoff;
2781 			slab = (uma_slab_t)mem;
2782 		}
2783 	} else {
2784 		/* This prevents redundant lookups via free(). */
2785 		if ((zone->uz_flags & UMA_ZONE_MALLOC) && udata != NULL)
2786 			slab = (uma_slab_t)udata;
2787 		else
2788 			slab = vtoslab((vm_offset_t)item);
2789 		keg = slab->us_keg;
2790 		keg_relock(keg, zone);
2791 	}
2792 	MPASS(keg == slab->us_keg);
2793 
2794 	/* Do we need to remove from any lists? */
2795 	if (slab->us_freecount+1 == keg->uk_ipers) {
2796 		LIST_REMOVE(slab, us_link);
2797 		LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
2798 	} else if (slab->us_freecount == 0) {
2799 		LIST_REMOVE(slab, us_link);
2800 		LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2801 	}
2802 
2803 	/* Slab management stuff */
2804 	freei = ((unsigned long)item - (unsigned long)slab->us_data)
2805 		/ keg->uk_rsize;
2806 
2807 #ifdef INVARIANTS
2808 	if (!skip)
2809 		uma_dbg_free(zone, slab, item);
2810 #endif
2811 
2812 	if (keg->uk_flags & UMA_ZONE_REFCNT) {
2813 		slabref = (uma_slabrefcnt_t)slab;
2814 		slabref->us_freelist[freei].us_item = slab->us_firstfree;
2815 	} else {
2816 		slab->us_freelist[freei].us_item = slab->us_firstfree;
2817 	}
2818 	slab->us_firstfree = freei;
2819 	slab->us_freecount++;
2820 
2821 	/* Zone statistics */
2822 	keg->uk_free++;
2823 
2824 	clearfull = 0;
2825 	if (keg->uk_flags & UMA_ZFLAG_FULL) {
2826 		if (keg->uk_pages < keg->uk_maxpages) {
2827 			keg->uk_flags &= ~UMA_ZFLAG_FULL;
2828 			clearfull = 1;
2829 		}
2830 
2831 		/*
2832 		 * We can handle one more allocation. Since we're clearing ZFLAG_FULL,
2833 		 * wake up all procs blocked on pages. This should be uncommon, so
2834 		 * keeping this simple for now (rather than adding count of blocked
2835 		 * threads etc).
2836 		 */
2837 		wakeup(keg);
2838 	}
2839 	if (clearfull) {
2840 		zone_relock(zone, keg);
2841 		zone->uz_flags &= ~UMA_ZFLAG_FULL;
2842 		wakeup(zone);
2843 		ZONE_UNLOCK(zone);
2844 	} else
2845 		KEG_UNLOCK(keg);
2846 }
2847 
2848 /* See uma.h */
2849 int
2850 uma_zone_set_max(uma_zone_t zone, int nitems)
2851 {
2852 	uma_keg_t keg;
2853 
2854 	ZONE_LOCK(zone);
2855 	keg = zone_first_keg(zone);
2856 	keg->uk_maxpages = (nitems / keg->uk_ipers) * keg->uk_ppera;
2857 	if (keg->uk_maxpages * keg->uk_ipers < nitems)
2858 		keg->uk_maxpages += keg->uk_ppera;
2859 	nitems = keg->uk_maxpages * keg->uk_ipers;
2860 	ZONE_UNLOCK(zone);
2861 
2862 	return (nitems);
2863 }
2864 
2865 /* See uma.h */
2866 int
2867 uma_zone_get_max(uma_zone_t zone)
2868 {
2869 	int nitems;
2870 	uma_keg_t keg;
2871 
2872 	ZONE_LOCK(zone);
2873 	keg = zone_first_keg(zone);
2874 	nitems = keg->uk_maxpages * keg->uk_ipers;
2875 	ZONE_UNLOCK(zone);
2876 
2877 	return (nitems);
2878 }
2879 
2880 /* See uma.h */
2881 int
2882 uma_zone_get_cur(uma_zone_t zone)
2883 {
2884 	int64_t nitems;
2885 	u_int i;
2886 
2887 	ZONE_LOCK(zone);
2888 	nitems = zone->uz_allocs - zone->uz_frees;
2889 	CPU_FOREACH(i) {
2890 		/*
2891 		 * See the comment in sysctl_vm_zone_stats() regarding the
2892 		 * safety of accessing the per-cpu caches. With the zone lock
2893 		 * held, it is safe, but can potentially result in stale data.
2894 		 */
2895 		nitems += zone->uz_cpu[i].uc_allocs -
2896 		    zone->uz_cpu[i].uc_frees;
2897 	}
2898 	ZONE_UNLOCK(zone);
2899 
2900 	return (nitems < 0 ? 0 : nitems);
2901 }
2902 
2903 /* See uma.h */
2904 void
2905 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
2906 {
2907 	uma_keg_t keg;
2908 
2909 	ZONE_LOCK(zone);
2910 	keg = zone_first_keg(zone);
2911 	KASSERT(keg->uk_pages == 0,
2912 	    ("uma_zone_set_init on non-empty keg"));
2913 	keg->uk_init = uminit;
2914 	ZONE_UNLOCK(zone);
2915 }
2916 
2917 /* See uma.h */
2918 void
2919 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
2920 {
2921 	uma_keg_t keg;
2922 
2923 	ZONE_LOCK(zone);
2924 	keg = zone_first_keg(zone);
2925 	KASSERT(keg->uk_pages == 0,
2926 	    ("uma_zone_set_fini on non-empty keg"));
2927 	keg->uk_fini = fini;
2928 	ZONE_UNLOCK(zone);
2929 }
2930 
2931 /* See uma.h */
2932 void
2933 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
2934 {
2935 	ZONE_LOCK(zone);
2936 	KASSERT(zone_first_keg(zone)->uk_pages == 0,
2937 	    ("uma_zone_set_zinit on non-empty keg"));
2938 	zone->uz_init = zinit;
2939 	ZONE_UNLOCK(zone);
2940 }
2941 
2942 /* See uma.h */
2943 void
2944 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
2945 {
2946 	ZONE_LOCK(zone);
2947 	KASSERT(zone_first_keg(zone)->uk_pages == 0,
2948 	    ("uma_zone_set_zfini on non-empty keg"));
2949 	zone->uz_fini = zfini;
2950 	ZONE_UNLOCK(zone);
2951 }
2952 
2953 /* See uma.h */
2954 /* XXX uk_freef is not actually used with the zone locked */
2955 void
2956 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
2957 {
2958 
2959 	ZONE_LOCK(zone);
2960 	zone_first_keg(zone)->uk_freef = freef;
2961 	ZONE_UNLOCK(zone);
2962 }
2963 
2964 /* See uma.h */
2965 /* XXX uk_allocf is not actually used with the zone locked */
2966 void
2967 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
2968 {
2969 	uma_keg_t keg;
2970 
2971 	ZONE_LOCK(zone);
2972 	keg = zone_first_keg(zone);
2973 	keg->uk_flags |= UMA_ZFLAG_PRIVALLOC;
2974 	keg->uk_allocf = allocf;
2975 	ZONE_UNLOCK(zone);
2976 }
2977 
2978 /* See uma.h */
2979 int
2980 uma_zone_set_obj(uma_zone_t zone, struct vm_object *obj, int count)
2981 {
2982 	uma_keg_t keg;
2983 	vm_offset_t kva;
2984 	int pages;
2985 
2986 	keg = zone_first_keg(zone);
2987 	pages = count / keg->uk_ipers;
2988 
2989 	if (pages * keg->uk_ipers < count)
2990 		pages++;
2991 
2992 	kva = kmem_alloc_nofault(kernel_map, pages * UMA_SLAB_SIZE);
2993 
2994 	if (kva == 0)
2995 		return (0);
2996 	if (obj == NULL)
2997 		obj = vm_object_allocate(OBJT_PHYS, pages);
2998 	else {
2999 		VM_OBJECT_LOCK_INIT(obj, "uma object");
3000 		_vm_object_allocate(OBJT_PHYS, pages, obj);
3001 	}
3002 	ZONE_LOCK(zone);
3003 	keg->uk_kva = kva;
3004 	keg->uk_obj = obj;
3005 	keg->uk_maxpages = pages;
3006 	keg->uk_allocf = obj_alloc;
3007 	keg->uk_flags |= UMA_ZONE_NOFREE | UMA_ZFLAG_PRIVALLOC;
3008 	ZONE_UNLOCK(zone);
3009 	return (1);
3010 }
3011 
3012 /* See uma.h */
3013 void
3014 uma_prealloc(uma_zone_t zone, int items)
3015 {
3016 	int slabs;
3017 	uma_slab_t slab;
3018 	uma_keg_t keg;
3019 
3020 	keg = zone_first_keg(zone);
3021 	ZONE_LOCK(zone);
3022 	slabs = items / keg->uk_ipers;
3023 	if (slabs * keg->uk_ipers < items)
3024 		slabs++;
3025 	while (slabs > 0) {
3026 		slab = keg_alloc_slab(keg, zone, M_WAITOK);
3027 		if (slab == NULL)
3028 			break;
3029 		MPASS(slab->us_keg == keg);
3030 		LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
3031 		slabs--;
3032 	}
3033 	ZONE_UNLOCK(zone);
3034 }
3035 
3036 /* See uma.h */
3037 u_int32_t *
3038 uma_find_refcnt(uma_zone_t zone, void *item)
3039 {
3040 	uma_slabrefcnt_t slabref;
3041 	uma_keg_t keg;
3042 	u_int32_t *refcnt;
3043 	int idx;
3044 
3045 	slabref = (uma_slabrefcnt_t)vtoslab((vm_offset_t)item &
3046 	    (~UMA_SLAB_MASK));
3047 	keg = slabref->us_keg;
3048 	KASSERT(slabref != NULL && slabref->us_keg->uk_flags & UMA_ZONE_REFCNT,
3049 	    ("uma_find_refcnt(): zone possibly not UMA_ZONE_REFCNT"));
3050 	idx = ((unsigned long)item - (unsigned long)slabref->us_data)
3051 	    / keg->uk_rsize;
3052 	refcnt = &slabref->us_freelist[idx].us_refcnt;
3053 	return refcnt;
3054 }
3055 
3056 /* See uma.h */
3057 void
3058 uma_reclaim(void)
3059 {
3060 #ifdef UMA_DEBUG
3061 	printf("UMA: vm asked us to release pages!\n");
3062 #endif
3063 	bucket_enable();
3064 	zone_foreach(zone_drain);
3065 	/*
3066 	 * Some slabs may have been freed but this zone will be visited early
3067 	 * we visit again so that we can free pages that are empty once other
3068 	 * zones are drained.  We have to do the same for buckets.
3069 	 */
3070 	zone_drain(slabzone);
3071 	zone_drain(slabrefzone);
3072 	bucket_zone_drain();
3073 }
3074 
3075 /* See uma.h */
3076 int
3077 uma_zone_exhausted(uma_zone_t zone)
3078 {
3079 	int full;
3080 
3081 	ZONE_LOCK(zone);
3082 	full = (zone->uz_flags & UMA_ZFLAG_FULL);
3083 	ZONE_UNLOCK(zone);
3084 	return (full);
3085 }
3086 
3087 int
3088 uma_zone_exhausted_nolock(uma_zone_t zone)
3089 {
3090 	return (zone->uz_flags & UMA_ZFLAG_FULL);
3091 }
3092 
3093 void *
3094 uma_large_malloc(int size, int wait)
3095 {
3096 	void *mem;
3097 	uma_slab_t slab;
3098 	u_int8_t flags;
3099 
3100 	slab = zone_alloc_item(slabzone, NULL, wait);
3101 	if (slab == NULL)
3102 		return (NULL);
3103 	mem = page_alloc(NULL, size, &flags, wait);
3104 	if (mem) {
3105 		vsetslab((vm_offset_t)mem, slab);
3106 		slab->us_data = mem;
3107 		slab->us_flags = flags | UMA_SLAB_MALLOC;
3108 		slab->us_size = size;
3109 	} else {
3110 		zone_free_item(slabzone, slab, NULL, SKIP_NONE,
3111 		    ZFREE_STATFAIL | ZFREE_STATFREE);
3112 	}
3113 
3114 	return (mem);
3115 }
3116 
3117 void
3118 uma_large_free(uma_slab_t slab)
3119 {
3120 	vsetobj((vm_offset_t)slab->us_data, kmem_object);
3121 	page_free(slab->us_data, slab->us_size, slab->us_flags);
3122 	zone_free_item(slabzone, slab, NULL, SKIP_NONE, ZFREE_STATFREE);
3123 }
3124 
3125 void
3126 uma_print_stats(void)
3127 {
3128 	zone_foreach(uma_print_zone);
3129 }
3130 
3131 static void
3132 slab_print(uma_slab_t slab)
3133 {
3134 	printf("slab: keg %p, data %p, freecount %d, firstfree %d\n",
3135 		slab->us_keg, slab->us_data, slab->us_freecount,
3136 		slab->us_firstfree);
3137 }
3138 
3139 static void
3140 cache_print(uma_cache_t cache)
3141 {
3142 	printf("alloc: %p(%d), free: %p(%d)\n",
3143 		cache->uc_allocbucket,
3144 		cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0,
3145 		cache->uc_freebucket,
3146 		cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0);
3147 }
3148 
3149 static void
3150 uma_print_keg(uma_keg_t keg)
3151 {
3152 	uma_slab_t slab;
3153 
3154 	printf("keg: %s(%p) size %d(%d) flags %d ipers %d ppera %d "
3155 	    "out %d free %d limit %d\n",
3156 	    keg->uk_name, keg, keg->uk_size, keg->uk_rsize, keg->uk_flags,
3157 	    keg->uk_ipers, keg->uk_ppera,
3158 	    (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free,
3159 	    (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers);
3160 	printf("Part slabs:\n");
3161 	LIST_FOREACH(slab, &keg->uk_part_slab, us_link)
3162 		slab_print(slab);
3163 	printf("Free slabs:\n");
3164 	LIST_FOREACH(slab, &keg->uk_free_slab, us_link)
3165 		slab_print(slab);
3166 	printf("Full slabs:\n");
3167 	LIST_FOREACH(slab, &keg->uk_full_slab, us_link)
3168 		slab_print(slab);
3169 }
3170 
3171 void
3172 uma_print_zone(uma_zone_t zone)
3173 {
3174 	uma_cache_t cache;
3175 	uma_klink_t kl;
3176 	int i;
3177 
3178 	printf("zone: %s(%p) size %d flags %d\n",
3179 	    zone->uz_name, zone, zone->uz_size, zone->uz_flags);
3180 	LIST_FOREACH(kl, &zone->uz_kegs, kl_link)
3181 		uma_print_keg(kl->kl_keg);
3182 	CPU_FOREACH(i) {
3183 		cache = &zone->uz_cpu[i];
3184 		printf("CPU %d Cache:\n", i);
3185 		cache_print(cache);
3186 	}
3187 }
3188 
3189 #ifdef DDB
3190 /*
3191  * Generate statistics across both the zone and its per-cpu cache's.  Return
3192  * desired statistics if the pointer is non-NULL for that statistic.
3193  *
3194  * Note: does not update the zone statistics, as it can't safely clear the
3195  * per-CPU cache statistic.
3196  *
3197  * XXXRW: Following the uc_allocbucket and uc_freebucket pointers here isn't
3198  * safe from off-CPU; we should modify the caches to track this information
3199  * directly so that we don't have to.
3200  */
3201 static void
3202 uma_zone_sumstat(uma_zone_t z, int *cachefreep, u_int64_t *allocsp,
3203     u_int64_t *freesp, u_int64_t *sleepsp)
3204 {
3205 	uma_cache_t cache;
3206 	u_int64_t allocs, frees, sleeps;
3207 	int cachefree, cpu;
3208 
3209 	allocs = frees = sleeps = 0;
3210 	cachefree = 0;
3211 	CPU_FOREACH(cpu) {
3212 		cache = &z->uz_cpu[cpu];
3213 		if (cache->uc_allocbucket != NULL)
3214 			cachefree += cache->uc_allocbucket->ub_cnt;
3215 		if (cache->uc_freebucket != NULL)
3216 			cachefree += cache->uc_freebucket->ub_cnt;
3217 		allocs += cache->uc_allocs;
3218 		frees += cache->uc_frees;
3219 	}
3220 	allocs += z->uz_allocs;
3221 	frees += z->uz_frees;
3222 	sleeps += z->uz_sleeps;
3223 	if (cachefreep != NULL)
3224 		*cachefreep = cachefree;
3225 	if (allocsp != NULL)
3226 		*allocsp = allocs;
3227 	if (freesp != NULL)
3228 		*freesp = frees;
3229 	if (sleepsp != NULL)
3230 		*sleepsp = sleeps;
3231 }
3232 #endif /* DDB */
3233 
3234 static int
3235 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
3236 {
3237 	uma_keg_t kz;
3238 	uma_zone_t z;
3239 	int count;
3240 
3241 	count = 0;
3242 	mtx_lock(&uma_mtx);
3243 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3244 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
3245 			count++;
3246 	}
3247 	mtx_unlock(&uma_mtx);
3248 	return (sysctl_handle_int(oidp, &count, 0, req));
3249 }
3250 
3251 static int
3252 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
3253 {
3254 	struct uma_stream_header ush;
3255 	struct uma_type_header uth;
3256 	struct uma_percpu_stat ups;
3257 	uma_bucket_t bucket;
3258 	struct sbuf sbuf;
3259 	uma_cache_t cache;
3260 	uma_klink_t kl;
3261 	uma_keg_t kz;
3262 	uma_zone_t z;
3263 	uma_keg_t k;
3264 	int count, error, i;
3265 
3266 	error = sysctl_wire_old_buffer(req, 0);
3267 	if (error != 0)
3268 		return (error);
3269 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
3270 
3271 	count = 0;
3272 	mtx_lock(&uma_mtx);
3273 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3274 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
3275 			count++;
3276 	}
3277 
3278 	/*
3279 	 * Insert stream header.
3280 	 */
3281 	bzero(&ush, sizeof(ush));
3282 	ush.ush_version = UMA_STREAM_VERSION;
3283 	ush.ush_maxcpus = (mp_maxid + 1);
3284 	ush.ush_count = count;
3285 	(void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
3286 
3287 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3288 		LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3289 			bzero(&uth, sizeof(uth));
3290 			ZONE_LOCK(z);
3291 			strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
3292 			uth.uth_align = kz->uk_align;
3293 			uth.uth_size = kz->uk_size;
3294 			uth.uth_rsize = kz->uk_rsize;
3295 			LIST_FOREACH(kl, &z->uz_kegs, kl_link) {
3296 				k = kl->kl_keg;
3297 				uth.uth_maxpages += k->uk_maxpages;
3298 				uth.uth_pages += k->uk_pages;
3299 				uth.uth_keg_free += k->uk_free;
3300 				uth.uth_limit = (k->uk_maxpages / k->uk_ppera)
3301 				    * k->uk_ipers;
3302 			}
3303 
3304 			/*
3305 			 * A zone is secondary is it is not the first entry
3306 			 * on the keg's zone list.
3307 			 */
3308 			if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
3309 			    (LIST_FIRST(&kz->uk_zones) != z))
3310 				uth.uth_zone_flags = UTH_ZONE_SECONDARY;
3311 
3312 			LIST_FOREACH(bucket, &z->uz_full_bucket, ub_link)
3313 				uth.uth_zone_free += bucket->ub_cnt;
3314 			uth.uth_allocs = z->uz_allocs;
3315 			uth.uth_frees = z->uz_frees;
3316 			uth.uth_fails = z->uz_fails;
3317 			uth.uth_sleeps = z->uz_sleeps;
3318 			(void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
3319 			/*
3320 			 * While it is not normally safe to access the cache
3321 			 * bucket pointers while not on the CPU that owns the
3322 			 * cache, we only allow the pointers to be exchanged
3323 			 * without the zone lock held, not invalidated, so
3324 			 * accept the possible race associated with bucket
3325 			 * exchange during monitoring.
3326 			 */
3327 			for (i = 0; i < (mp_maxid + 1); i++) {
3328 				bzero(&ups, sizeof(ups));
3329 				if (kz->uk_flags & UMA_ZFLAG_INTERNAL)
3330 					goto skip;
3331 				if (CPU_ABSENT(i))
3332 					goto skip;
3333 				cache = &z->uz_cpu[i];
3334 				if (cache->uc_allocbucket != NULL)
3335 					ups.ups_cache_free +=
3336 					    cache->uc_allocbucket->ub_cnt;
3337 				if (cache->uc_freebucket != NULL)
3338 					ups.ups_cache_free +=
3339 					    cache->uc_freebucket->ub_cnt;
3340 				ups.ups_allocs = cache->uc_allocs;
3341 				ups.ups_frees = cache->uc_frees;
3342 skip:
3343 				(void)sbuf_bcat(&sbuf, &ups, sizeof(ups));
3344 			}
3345 			ZONE_UNLOCK(z);
3346 		}
3347 	}
3348 	mtx_unlock(&uma_mtx);
3349 	error = sbuf_finish(&sbuf);
3350 	sbuf_delete(&sbuf);
3351 	return (error);
3352 }
3353 
3354 #ifdef DDB
3355 DB_SHOW_COMMAND(uma, db_show_uma)
3356 {
3357 	u_int64_t allocs, frees, sleeps;
3358 	uma_bucket_t bucket;
3359 	uma_keg_t kz;
3360 	uma_zone_t z;
3361 	int cachefree;
3362 
3363 	db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
3364 	    "Requests", "Sleeps");
3365 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3366 		LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3367 			if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
3368 				allocs = z->uz_allocs;
3369 				frees = z->uz_frees;
3370 				sleeps = z->uz_sleeps;
3371 				cachefree = 0;
3372 			} else
3373 				uma_zone_sumstat(z, &cachefree, &allocs,
3374 				    &frees, &sleeps);
3375 			if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
3376 			    (LIST_FIRST(&kz->uk_zones) != z)))
3377 				cachefree += kz->uk_free;
3378 			LIST_FOREACH(bucket, &z->uz_full_bucket, ub_link)
3379 				cachefree += bucket->ub_cnt;
3380 			db_printf("%18s %8ju %8jd %8d %12ju %8ju\n", z->uz_name,
3381 			    (uintmax_t)kz->uk_size,
3382 			    (intmax_t)(allocs - frees), cachefree,
3383 			    (uintmax_t)allocs, sleeps);
3384 		}
3385 	}
3386 }
3387 #endif
3388