xref: /freebsd/sys/vm/uma_core.c (revision 884a2a699669ec61e2366e3e358342dbc94be24a)
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 
63 #include <sys/param.h>
64 #include <sys/systm.h>
65 #include <sys/kernel.h>
66 #include <sys/types.h>
67 #include <sys/queue.h>
68 #include <sys/malloc.h>
69 #include <sys/ktr.h>
70 #include <sys/lock.h>
71 #include <sys/sysctl.h>
72 #include <sys/mutex.h>
73 #include <sys/proc.h>
74 #include <sys/sbuf.h>
75 #include <sys/smp.h>
76 #include <sys/vmmeter.h>
77 
78 #include <vm/vm.h>
79 #include <vm/vm_object.h>
80 #include <vm/vm_page.h>
81 #include <vm/vm_param.h>
82 #include <vm/vm_map.h>
83 #include <vm/vm_kern.h>
84 #include <vm/vm_extern.h>
85 #include <vm/uma.h>
86 #include <vm/uma_int.h>
87 #include <vm/uma_dbg.h>
88 
89 #include <ddb/ddb.h>
90 
91 /*
92  * This is the zone and keg from which all zones are spawned.  The idea is that
93  * even the zone & keg heads are allocated from the allocator, so we use the
94  * bss section to bootstrap us.
95  */
96 static struct uma_keg masterkeg;
97 static struct uma_zone masterzone_k;
98 static struct uma_zone masterzone_z;
99 static uma_zone_t kegs = &masterzone_k;
100 static uma_zone_t zones = &masterzone_z;
101 
102 /* This is the zone from which all of uma_slab_t's are allocated. */
103 static uma_zone_t slabzone;
104 static uma_zone_t slabrefzone;	/* With refcounters (for UMA_ZONE_REFCNT) */
105 
106 /*
107  * The initial hash tables come out of this zone so they can be allocated
108  * prior to malloc coming up.
109  */
110 static uma_zone_t hashzone;
111 
112 /* The boot-time adjusted value for cache line alignment. */
113 int uma_align_cache = 64 - 1;
114 
115 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets");
116 
117 /*
118  * Are we allowed to allocate buckets?
119  */
120 static int bucketdisable = 1;
121 
122 /* Linked list of all kegs in the system */
123 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs);
124 
125 /* This mutex protects the keg list */
126 static struct mtx uma_mtx;
127 
128 /* Linked list of boot time pages */
129 static LIST_HEAD(,uma_slab) uma_boot_pages =
130     LIST_HEAD_INITIALIZER(uma_boot_pages);
131 
132 /* This mutex protects the boot time pages list */
133 static struct mtx uma_boot_pages_mtx;
134 
135 /* Is the VM done starting up? */
136 static int booted = 0;
137 #define	UMA_STARTUP	1
138 #define	UMA_STARTUP2	2
139 
140 /* Maximum number of allowed items-per-slab if the slab header is OFFPAGE */
141 static u_int uma_max_ipers;
142 static u_int uma_max_ipers_ref;
143 
144 /*
145  * This is the handle used to schedule events that need to happen
146  * outside of the allocation fast path.
147  */
148 static struct callout uma_callout;
149 #define	UMA_TIMEOUT	20		/* Seconds for callout interval. */
150 
151 /*
152  * This structure is passed as the zone ctor arg so that I don't have to create
153  * a special allocation function just for zones.
154  */
155 struct uma_zctor_args {
156 	char *name;
157 	size_t size;
158 	uma_ctor ctor;
159 	uma_dtor dtor;
160 	uma_init uminit;
161 	uma_fini fini;
162 	uma_keg_t keg;
163 	int align;
164 	u_int32_t flags;
165 };
166 
167 struct uma_kctor_args {
168 	uma_zone_t zone;
169 	size_t size;
170 	uma_init uminit;
171 	uma_fini fini;
172 	int align;
173 	u_int32_t flags;
174 };
175 
176 struct uma_bucket_zone {
177 	uma_zone_t	ubz_zone;
178 	char		*ubz_name;
179 	int		ubz_entries;
180 };
181 
182 #define	BUCKET_MAX	128
183 
184 struct uma_bucket_zone bucket_zones[] = {
185 	{ NULL, "16 Bucket", 16 },
186 	{ NULL, "32 Bucket", 32 },
187 	{ NULL, "64 Bucket", 64 },
188 	{ NULL, "128 Bucket", 128 },
189 	{ NULL, NULL, 0}
190 };
191 
192 #define	BUCKET_SHIFT	4
193 #define	BUCKET_ZONES	((BUCKET_MAX >> BUCKET_SHIFT) + 1)
194 
195 /*
196  * bucket_size[] maps requested bucket sizes to zones that allocate a bucket
197  * of approximately the right size.
198  */
199 static uint8_t bucket_size[BUCKET_ZONES];
200 
201 /*
202  * Flags and enumerations to be passed to internal functions.
203  */
204 enum zfreeskip { SKIP_NONE, SKIP_DTOR, SKIP_FINI };
205 
206 #define	ZFREE_STATFAIL	0x00000001	/* Update zone failure statistic. */
207 #define	ZFREE_STATFREE	0x00000002	/* Update zone free statistic. */
208 
209 /* Prototypes.. */
210 
211 static void *obj_alloc(uma_zone_t, int, u_int8_t *, int);
212 static void *page_alloc(uma_zone_t, int, u_int8_t *, int);
213 static void *startup_alloc(uma_zone_t, int, u_int8_t *, int);
214 static void page_free(void *, int, u_int8_t);
215 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int);
216 static void cache_drain(uma_zone_t);
217 static void bucket_drain(uma_zone_t, uma_bucket_t);
218 static void bucket_cache_drain(uma_zone_t zone);
219 static int keg_ctor(void *, int, void *, int);
220 static void keg_dtor(void *, int, void *);
221 static int zone_ctor(void *, int, void *, int);
222 static void zone_dtor(void *, int, void *);
223 static int zero_init(void *, int, int);
224 static void keg_small_init(uma_keg_t keg);
225 static void keg_large_init(uma_keg_t keg);
226 static void zone_foreach(void (*zfunc)(uma_zone_t));
227 static void zone_timeout(uma_zone_t zone);
228 static int hash_alloc(struct uma_hash *);
229 static int hash_expand(struct uma_hash *, struct uma_hash *);
230 static void hash_free(struct uma_hash *hash);
231 static void uma_timeout(void *);
232 static void uma_startup3(void);
233 static void *zone_alloc_item(uma_zone_t, void *, int);
234 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip,
235     int);
236 static void bucket_enable(void);
237 static void bucket_init(void);
238 static uma_bucket_t bucket_alloc(int, int);
239 static void bucket_free(uma_bucket_t);
240 static void bucket_zone_drain(void);
241 static int zone_alloc_bucket(uma_zone_t zone, int flags);
242 static uma_slab_t zone_fetch_slab(uma_zone_t zone, uma_keg_t last, int flags);
243 static uma_slab_t zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int flags);
244 static void *slab_alloc_item(uma_zone_t zone, uma_slab_t slab);
245 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit,
246     uma_fini fini, int align, u_int32_t flags);
247 static inline void zone_relock(uma_zone_t zone, uma_keg_t keg);
248 static inline void keg_relock(uma_keg_t keg, uma_zone_t zone);
249 
250 void uma_print_zone(uma_zone_t);
251 void uma_print_stats(void);
252 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS);
253 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS);
254 
255 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
256 
257 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLTYPE_INT,
258     0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones");
259 
260 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLTYPE_STRUCT,
261     0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats");
262 
263 /*
264  * This routine checks to see whether or not it's safe to enable buckets.
265  */
266 
267 static void
268 bucket_enable(void)
269 {
270 	if (cnt.v_free_count < cnt.v_free_min)
271 		bucketdisable = 1;
272 	else
273 		bucketdisable = 0;
274 }
275 
276 /*
277  * Initialize bucket_zones, the array of zones of buckets of various sizes.
278  *
279  * For each zone, calculate the memory required for each bucket, consisting
280  * of the header and an array of pointers.  Initialize bucket_size[] to point
281  * the range of appropriate bucket sizes at the zone.
282  */
283 static void
284 bucket_init(void)
285 {
286 	struct uma_bucket_zone *ubz;
287 	int i;
288 	int j;
289 
290 	for (i = 0, j = 0; bucket_zones[j].ubz_entries != 0; j++) {
291 		int size;
292 
293 		ubz = &bucket_zones[j];
294 		size = roundup(sizeof(struct uma_bucket), sizeof(void *));
295 		size += sizeof(void *) * ubz->ubz_entries;
296 		ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
297 		    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
298 		    UMA_ZFLAG_INTERNAL | UMA_ZFLAG_BUCKET);
299 		for (; i <= ubz->ubz_entries; i += (1 << BUCKET_SHIFT))
300 			bucket_size[i >> BUCKET_SHIFT] = j;
301 	}
302 }
303 
304 /*
305  * Given a desired number of entries for a bucket, return the zone from which
306  * to allocate the bucket.
307  */
308 static struct uma_bucket_zone *
309 bucket_zone_lookup(int entries)
310 {
311 	int idx;
312 
313 	idx = howmany(entries, 1 << BUCKET_SHIFT);
314 	return (&bucket_zones[bucket_size[idx]]);
315 }
316 
317 static uma_bucket_t
318 bucket_alloc(int entries, int bflags)
319 {
320 	struct uma_bucket_zone *ubz;
321 	uma_bucket_t bucket;
322 
323 	/*
324 	 * This is to stop us from allocating per cpu buckets while we're
325 	 * running out of vm.boot_pages.  Otherwise, we would exhaust the
326 	 * boot pages.  This also prevents us from allocating buckets in
327 	 * low memory situations.
328 	 */
329 	if (bucketdisable)
330 		return (NULL);
331 
332 	ubz = bucket_zone_lookup(entries);
333 	bucket = zone_alloc_item(ubz->ubz_zone, NULL, bflags);
334 	if (bucket) {
335 #ifdef INVARIANTS
336 		bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
337 #endif
338 		bucket->ub_cnt = 0;
339 		bucket->ub_entries = ubz->ubz_entries;
340 	}
341 
342 	return (bucket);
343 }
344 
345 static void
346 bucket_free(uma_bucket_t bucket)
347 {
348 	struct uma_bucket_zone *ubz;
349 
350 	ubz = bucket_zone_lookup(bucket->ub_entries);
351 	zone_free_item(ubz->ubz_zone, bucket, NULL, SKIP_NONE,
352 	    ZFREE_STATFREE);
353 }
354 
355 static void
356 bucket_zone_drain(void)
357 {
358 	struct uma_bucket_zone *ubz;
359 
360 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
361 		zone_drain(ubz->ubz_zone);
362 }
363 
364 static inline uma_keg_t
365 zone_first_keg(uma_zone_t zone)
366 {
367 
368 	return (LIST_FIRST(&zone->uz_kegs)->kl_keg);
369 }
370 
371 static void
372 zone_foreach_keg(uma_zone_t zone, void (*kegfn)(uma_keg_t))
373 {
374 	uma_klink_t klink;
375 
376 	LIST_FOREACH(klink, &zone->uz_kegs, kl_link)
377 		kegfn(klink->kl_keg);
378 }
379 
380 /*
381  * Routine called by timeout which is used to fire off some time interval
382  * based calculations.  (stats, hash size, etc.)
383  *
384  * Arguments:
385  *	arg   Unused
386  *
387  * Returns:
388  *	Nothing
389  */
390 static void
391 uma_timeout(void *unused)
392 {
393 	bucket_enable();
394 	zone_foreach(zone_timeout);
395 
396 	/* Reschedule this event */
397 	callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
398 }
399 
400 /*
401  * Routine to perform timeout driven calculations.  This expands the
402  * hashes and does per cpu statistics aggregation.
403  *
404  *  Returns nothing.
405  */
406 static void
407 keg_timeout(uma_keg_t keg)
408 {
409 
410 	KEG_LOCK(keg);
411 	/*
412 	 * Expand the keg hash table.
413 	 *
414 	 * This is done if the number of slabs is larger than the hash size.
415 	 * What I'm trying to do here is completely reduce collisions.  This
416 	 * may be a little aggressive.  Should I allow for two collisions max?
417 	 */
418 	if (keg->uk_flags & UMA_ZONE_HASH &&
419 	    keg->uk_pages / keg->uk_ppera >= keg->uk_hash.uh_hashsize) {
420 		struct uma_hash newhash;
421 		struct uma_hash oldhash;
422 		int ret;
423 
424 		/*
425 		 * This is so involved because allocating and freeing
426 		 * while the keg lock is held will lead to deadlock.
427 		 * I have to do everything in stages and check for
428 		 * races.
429 		 */
430 		newhash = keg->uk_hash;
431 		KEG_UNLOCK(keg);
432 		ret = hash_alloc(&newhash);
433 		KEG_LOCK(keg);
434 		if (ret) {
435 			if (hash_expand(&keg->uk_hash, &newhash)) {
436 				oldhash = keg->uk_hash;
437 				keg->uk_hash = newhash;
438 			} else
439 				oldhash = newhash;
440 
441 			KEG_UNLOCK(keg);
442 			hash_free(&oldhash);
443 			KEG_LOCK(keg);
444 		}
445 	}
446 	KEG_UNLOCK(keg);
447 }
448 
449 static void
450 zone_timeout(uma_zone_t zone)
451 {
452 
453 	zone_foreach_keg(zone, &keg_timeout);
454 }
455 
456 /*
457  * Allocate and zero fill the next sized hash table from the appropriate
458  * backing store.
459  *
460  * Arguments:
461  *	hash  A new hash structure with the old hash size in uh_hashsize
462  *
463  * Returns:
464  *	1 on sucess and 0 on failure.
465  */
466 static int
467 hash_alloc(struct uma_hash *hash)
468 {
469 	int oldsize;
470 	int alloc;
471 
472 	oldsize = hash->uh_hashsize;
473 
474 	/* We're just going to go to a power of two greater */
475 	if (oldsize)  {
476 		hash->uh_hashsize = oldsize * 2;
477 		alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
478 		hash->uh_slab_hash = (struct slabhead *)malloc(alloc,
479 		    M_UMAHASH, M_NOWAIT);
480 	} else {
481 		alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
482 		hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
483 		    M_WAITOK);
484 		hash->uh_hashsize = UMA_HASH_SIZE_INIT;
485 	}
486 	if (hash->uh_slab_hash) {
487 		bzero(hash->uh_slab_hash, alloc);
488 		hash->uh_hashmask = hash->uh_hashsize - 1;
489 		return (1);
490 	}
491 
492 	return (0);
493 }
494 
495 /*
496  * Expands the hash table for HASH zones.  This is done from zone_timeout
497  * to reduce collisions.  This must not be done in the regular allocation
498  * path, otherwise, we can recurse on the vm while allocating pages.
499  *
500  * Arguments:
501  *	oldhash  The hash you want to expand
502  *	newhash  The hash structure for the new table
503  *
504  * Returns:
505  *	Nothing
506  *
507  * Discussion:
508  */
509 static int
510 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
511 {
512 	uma_slab_t slab;
513 	int hval;
514 	int i;
515 
516 	if (!newhash->uh_slab_hash)
517 		return (0);
518 
519 	if (oldhash->uh_hashsize >= newhash->uh_hashsize)
520 		return (0);
521 
522 	/*
523 	 * I need to investigate hash algorithms for resizing without a
524 	 * full rehash.
525 	 */
526 
527 	for (i = 0; i < oldhash->uh_hashsize; i++)
528 		while (!SLIST_EMPTY(&oldhash->uh_slab_hash[i])) {
529 			slab = SLIST_FIRST(&oldhash->uh_slab_hash[i]);
530 			SLIST_REMOVE_HEAD(&oldhash->uh_slab_hash[i], us_hlink);
531 			hval = UMA_HASH(newhash, slab->us_data);
532 			SLIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
533 			    slab, us_hlink);
534 		}
535 
536 	return (1);
537 }
538 
539 /*
540  * Free the hash bucket to the appropriate backing store.
541  *
542  * Arguments:
543  *	slab_hash  The hash bucket we're freeing
544  *	hashsize   The number of entries in that hash bucket
545  *
546  * Returns:
547  *	Nothing
548  */
549 static void
550 hash_free(struct uma_hash *hash)
551 {
552 	if (hash->uh_slab_hash == NULL)
553 		return;
554 	if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
555 		zone_free_item(hashzone,
556 		    hash->uh_slab_hash, NULL, SKIP_NONE, ZFREE_STATFREE);
557 	else
558 		free(hash->uh_slab_hash, M_UMAHASH);
559 }
560 
561 /*
562  * Frees all outstanding items in a bucket
563  *
564  * Arguments:
565  *	zone   The zone to free to, must be unlocked.
566  *	bucket The free/alloc bucket with items, cpu queue must be locked.
567  *
568  * Returns:
569  *	Nothing
570  */
571 
572 static void
573 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
574 {
575 	void *item;
576 
577 	if (bucket == NULL)
578 		return;
579 
580 	while (bucket->ub_cnt > 0)  {
581 		bucket->ub_cnt--;
582 		item = bucket->ub_bucket[bucket->ub_cnt];
583 #ifdef INVARIANTS
584 		bucket->ub_bucket[bucket->ub_cnt] = NULL;
585 		KASSERT(item != NULL,
586 		    ("bucket_drain: botched ptr, item is NULL"));
587 #endif
588 		zone_free_item(zone, item, NULL, SKIP_DTOR, 0);
589 	}
590 }
591 
592 /*
593  * Drains the per cpu caches for a zone.
594  *
595  * NOTE: This may only be called while the zone is being turn down, and not
596  * during normal operation.  This is necessary in order that we do not have
597  * to migrate CPUs to drain the per-CPU caches.
598  *
599  * Arguments:
600  *	zone     The zone to drain, must be unlocked.
601  *
602  * Returns:
603  *	Nothing
604  */
605 static void
606 cache_drain(uma_zone_t zone)
607 {
608 	uma_cache_t cache;
609 	int cpu;
610 
611 	/*
612 	 * XXX: It is safe to not lock the per-CPU caches, because we're
613 	 * tearing down the zone anyway.  I.e., there will be no further use
614 	 * of the caches at this point.
615 	 *
616 	 * XXX: It would good to be able to assert that the zone is being
617 	 * torn down to prevent improper use of cache_drain().
618 	 *
619 	 * XXX: We lock the zone before passing into bucket_cache_drain() as
620 	 * it is used elsewhere.  Should the tear-down path be made special
621 	 * there in some form?
622 	 */
623 	CPU_FOREACH(cpu) {
624 		cache = &zone->uz_cpu[cpu];
625 		bucket_drain(zone, cache->uc_allocbucket);
626 		bucket_drain(zone, cache->uc_freebucket);
627 		if (cache->uc_allocbucket != NULL)
628 			bucket_free(cache->uc_allocbucket);
629 		if (cache->uc_freebucket != NULL)
630 			bucket_free(cache->uc_freebucket);
631 		cache->uc_allocbucket = cache->uc_freebucket = NULL;
632 	}
633 	ZONE_LOCK(zone);
634 	bucket_cache_drain(zone);
635 	ZONE_UNLOCK(zone);
636 }
637 
638 /*
639  * Drain the cached buckets from a zone.  Expects a locked zone on entry.
640  */
641 static void
642 bucket_cache_drain(uma_zone_t zone)
643 {
644 	uma_bucket_t bucket;
645 
646 	/*
647 	 * Drain the bucket queues and free the buckets, we just keep two per
648 	 * cpu (alloc/free).
649 	 */
650 	while ((bucket = LIST_FIRST(&zone->uz_full_bucket)) != NULL) {
651 		LIST_REMOVE(bucket, ub_link);
652 		ZONE_UNLOCK(zone);
653 		bucket_drain(zone, bucket);
654 		bucket_free(bucket);
655 		ZONE_LOCK(zone);
656 	}
657 
658 	/* Now we do the free queue.. */
659 	while ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) {
660 		LIST_REMOVE(bucket, ub_link);
661 		bucket_free(bucket);
662 	}
663 }
664 
665 /*
666  * Frees pages from a keg back to the system.  This is done on demand from
667  * the pageout daemon.
668  *
669  * Returns nothing.
670  */
671 static void
672 keg_drain(uma_keg_t keg)
673 {
674 	struct slabhead freeslabs = { 0 };
675 	uma_slab_t slab;
676 	uma_slab_t n;
677 	u_int8_t flags;
678 	u_int8_t *mem;
679 	int i;
680 
681 	/*
682 	 * We don't want to take pages from statically allocated kegs at this
683 	 * time
684 	 */
685 	if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL)
686 		return;
687 
688 #ifdef UMA_DEBUG
689 	printf("%s free items: %u\n", keg->uk_name, keg->uk_free);
690 #endif
691 	KEG_LOCK(keg);
692 	if (keg->uk_free == 0)
693 		goto finished;
694 
695 	slab = LIST_FIRST(&keg->uk_free_slab);
696 	while (slab) {
697 		n = LIST_NEXT(slab, us_link);
698 
699 		/* We have no where to free these to */
700 		if (slab->us_flags & UMA_SLAB_BOOT) {
701 			slab = n;
702 			continue;
703 		}
704 
705 		LIST_REMOVE(slab, us_link);
706 		keg->uk_pages -= keg->uk_ppera;
707 		keg->uk_free -= keg->uk_ipers;
708 
709 		if (keg->uk_flags & UMA_ZONE_HASH)
710 			UMA_HASH_REMOVE(&keg->uk_hash, slab, slab->us_data);
711 
712 		SLIST_INSERT_HEAD(&freeslabs, slab, us_hlink);
713 
714 		slab = n;
715 	}
716 finished:
717 	KEG_UNLOCK(keg);
718 
719 	while ((slab = SLIST_FIRST(&freeslabs)) != NULL) {
720 		SLIST_REMOVE(&freeslabs, slab, uma_slab, us_hlink);
721 		if (keg->uk_fini)
722 			for (i = 0; i < keg->uk_ipers; i++)
723 				keg->uk_fini(
724 				    slab->us_data + (keg->uk_rsize * i),
725 				    keg->uk_size);
726 		flags = slab->us_flags;
727 		mem = slab->us_data;
728 
729 		if (keg->uk_flags & UMA_ZONE_VTOSLAB) {
730 			vm_object_t obj;
731 
732 			if (flags & UMA_SLAB_KMEM)
733 				obj = kmem_object;
734 			else if (flags & UMA_SLAB_KERNEL)
735 				obj = kernel_object;
736 			else
737 				obj = NULL;
738 			for (i = 0; i < keg->uk_ppera; i++)
739 				vsetobj((vm_offset_t)mem + (i * PAGE_SIZE),
740 				    obj);
741 		}
742 		if (keg->uk_flags & UMA_ZONE_OFFPAGE)
743 			zone_free_item(keg->uk_slabzone, slab, NULL,
744 			    SKIP_NONE, ZFREE_STATFREE);
745 #ifdef UMA_DEBUG
746 		printf("%s: Returning %d bytes.\n",
747 		    keg->uk_name, UMA_SLAB_SIZE * keg->uk_ppera);
748 #endif
749 		keg->uk_freef(mem, UMA_SLAB_SIZE * keg->uk_ppera, flags);
750 	}
751 }
752 
753 static void
754 zone_drain_wait(uma_zone_t zone, int waitok)
755 {
756 
757 	/*
758 	 * Set draining to interlock with zone_dtor() so we can release our
759 	 * locks as we go.  Only dtor() should do a WAITOK call since it
760 	 * is the only call that knows the structure will still be available
761 	 * when it wakes up.
762 	 */
763 	ZONE_LOCK(zone);
764 	while (zone->uz_flags & UMA_ZFLAG_DRAINING) {
765 		if (waitok == M_NOWAIT)
766 			goto out;
767 		mtx_unlock(&uma_mtx);
768 		msleep(zone, zone->uz_lock, PVM, "zonedrain", 1);
769 		mtx_lock(&uma_mtx);
770 	}
771 	zone->uz_flags |= UMA_ZFLAG_DRAINING;
772 	bucket_cache_drain(zone);
773 	ZONE_UNLOCK(zone);
774 	/*
775 	 * The DRAINING flag protects us from being freed while
776 	 * we're running.  Normally the uma_mtx would protect us but we
777 	 * must be able to release and acquire the right lock for each keg.
778 	 */
779 	zone_foreach_keg(zone, &keg_drain);
780 	ZONE_LOCK(zone);
781 	zone->uz_flags &= ~UMA_ZFLAG_DRAINING;
782 	wakeup(zone);
783 out:
784 	ZONE_UNLOCK(zone);
785 }
786 
787 void
788 zone_drain(uma_zone_t zone)
789 {
790 
791 	zone_drain_wait(zone, M_NOWAIT);
792 }
793 
794 /*
795  * Allocate a new slab for a keg.  This does not insert the slab onto a list.
796  *
797  * Arguments:
798  *	wait  Shall we wait?
799  *
800  * Returns:
801  *	The slab that was allocated or NULL if there is no memory and the
802  *	caller specified M_NOWAIT.
803  */
804 static uma_slab_t
805 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int wait)
806 {
807 	uma_slabrefcnt_t slabref;
808 	uma_alloc allocf;
809 	uma_slab_t slab;
810 	u_int8_t *mem;
811 	u_int8_t flags;
812 	int i;
813 
814 	mtx_assert(&keg->uk_lock, MA_OWNED);
815 	slab = NULL;
816 
817 #ifdef UMA_DEBUG
818 	printf("slab_zalloc:  Allocating a new slab for %s\n", keg->uk_name);
819 #endif
820 	allocf = keg->uk_allocf;
821 	KEG_UNLOCK(keg);
822 
823 	if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
824 		slab = zone_alloc_item(keg->uk_slabzone, NULL, wait);
825 		if (slab == NULL) {
826 			KEG_LOCK(keg);
827 			return NULL;
828 		}
829 	}
830 
831 	/*
832 	 * This reproduces the old vm_zone behavior of zero filling pages the
833 	 * first time they are added to a zone.
834 	 *
835 	 * Malloced items are zeroed in uma_zalloc.
836 	 */
837 
838 	if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
839 		wait |= M_ZERO;
840 	else
841 		wait &= ~M_ZERO;
842 
843 	/* zone is passed for legacy reasons. */
844 	mem = allocf(zone, keg->uk_ppera * UMA_SLAB_SIZE, &flags, wait);
845 	if (mem == NULL) {
846 		if (keg->uk_flags & UMA_ZONE_OFFPAGE)
847 			zone_free_item(keg->uk_slabzone, slab, NULL,
848 			    SKIP_NONE, ZFREE_STATFREE);
849 		KEG_LOCK(keg);
850 		return (NULL);
851 	}
852 
853 	/* Point the slab into the allocated memory */
854 	if (!(keg->uk_flags & UMA_ZONE_OFFPAGE))
855 		slab = (uma_slab_t )(mem + keg->uk_pgoff);
856 
857 	if (keg->uk_flags & UMA_ZONE_VTOSLAB)
858 		for (i = 0; i < keg->uk_ppera; i++)
859 			vsetslab((vm_offset_t)mem + (i * PAGE_SIZE), slab);
860 
861 	slab->us_keg = keg;
862 	slab->us_data = mem;
863 	slab->us_freecount = keg->uk_ipers;
864 	slab->us_firstfree = 0;
865 	slab->us_flags = flags;
866 
867 	if (keg->uk_flags & UMA_ZONE_REFCNT) {
868 		slabref = (uma_slabrefcnt_t)slab;
869 		for (i = 0; i < keg->uk_ipers; i++) {
870 			slabref->us_freelist[i].us_refcnt = 0;
871 			slabref->us_freelist[i].us_item = i+1;
872 		}
873 	} else {
874 		for (i = 0; i < keg->uk_ipers; i++)
875 			slab->us_freelist[i].us_item = i+1;
876 	}
877 
878 	if (keg->uk_init != NULL) {
879 		for (i = 0; i < keg->uk_ipers; i++)
880 			if (keg->uk_init(slab->us_data + (keg->uk_rsize * i),
881 			    keg->uk_size, wait) != 0)
882 				break;
883 		if (i != keg->uk_ipers) {
884 			if (keg->uk_fini != NULL) {
885 				for (i--; i > -1; i--)
886 					keg->uk_fini(slab->us_data +
887 					    (keg->uk_rsize * i),
888 					    keg->uk_size);
889 			}
890 			if (keg->uk_flags & UMA_ZONE_VTOSLAB) {
891 				vm_object_t obj;
892 
893 				if (flags & UMA_SLAB_KMEM)
894 					obj = kmem_object;
895 				else if (flags & UMA_SLAB_KERNEL)
896 					obj = kernel_object;
897 				else
898 					obj = NULL;
899 				for (i = 0; i < keg->uk_ppera; i++)
900 					vsetobj((vm_offset_t)mem +
901 					    (i * PAGE_SIZE), obj);
902 			}
903 			if (keg->uk_flags & UMA_ZONE_OFFPAGE)
904 				zone_free_item(keg->uk_slabzone, slab,
905 				    NULL, SKIP_NONE, ZFREE_STATFREE);
906 			keg->uk_freef(mem, UMA_SLAB_SIZE * keg->uk_ppera,
907 			    flags);
908 			KEG_LOCK(keg);
909 			return (NULL);
910 		}
911 	}
912 	KEG_LOCK(keg);
913 
914 	if (keg->uk_flags & UMA_ZONE_HASH)
915 		UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
916 
917 	keg->uk_pages += keg->uk_ppera;
918 	keg->uk_free += keg->uk_ipers;
919 
920 	return (slab);
921 }
922 
923 /*
924  * This function is intended to be used early on in place of page_alloc() so
925  * that we may use the boot time page cache to satisfy allocations before
926  * the VM is ready.
927  */
928 static void *
929 startup_alloc(uma_zone_t zone, int bytes, u_int8_t *pflag, int wait)
930 {
931 	uma_keg_t keg;
932 	uma_slab_t tmps;
933 	int pages, check_pages;
934 
935 	keg = zone_first_keg(zone);
936 	pages = howmany(bytes, PAGE_SIZE);
937 	check_pages = pages - 1;
938 	KASSERT(pages > 0, ("startup_alloc can't reserve 0 pages\n"));
939 
940 	/*
941 	 * Check our small startup cache to see if it has pages remaining.
942 	 */
943 	mtx_lock(&uma_boot_pages_mtx);
944 
945 	/* First check if we have enough room. */
946 	tmps = LIST_FIRST(&uma_boot_pages);
947 	while (tmps != NULL && check_pages-- > 0)
948 		tmps = LIST_NEXT(tmps, us_link);
949 	if (tmps != NULL) {
950 		/*
951 		 * It's ok to lose tmps references.  The last one will
952 		 * have tmps->us_data pointing to the start address of
953 		 * "pages" contiguous pages of memory.
954 		 */
955 		while (pages-- > 0) {
956 			tmps = LIST_FIRST(&uma_boot_pages);
957 			LIST_REMOVE(tmps, us_link);
958 		}
959 		mtx_unlock(&uma_boot_pages_mtx);
960 		*pflag = tmps->us_flags;
961 		return (tmps->us_data);
962 	}
963 	mtx_unlock(&uma_boot_pages_mtx);
964 	if (booted < UMA_STARTUP2)
965 		panic("UMA: Increase vm.boot_pages");
966 	/*
967 	 * Now that we've booted reset these users to their real allocator.
968 	 */
969 #ifdef UMA_MD_SMALL_ALLOC
970 	keg->uk_allocf = (keg->uk_ppera > 1) ? page_alloc : uma_small_alloc;
971 #else
972 	keg->uk_allocf = page_alloc;
973 #endif
974 	return keg->uk_allocf(zone, bytes, pflag, wait);
975 }
976 
977 /*
978  * Allocates a number of pages from the system
979  *
980  * Arguments:
981  *	bytes  The number of bytes requested
982  *	wait  Shall we wait?
983  *
984  * Returns:
985  *	A pointer to the alloced memory or possibly
986  *	NULL if M_NOWAIT is set.
987  */
988 static void *
989 page_alloc(uma_zone_t zone, int bytes, u_int8_t *pflag, int wait)
990 {
991 	void *p;	/* Returned page */
992 
993 	*pflag = UMA_SLAB_KMEM;
994 	p = (void *) kmem_malloc(kmem_map, bytes, wait);
995 
996 	return (p);
997 }
998 
999 /*
1000  * Allocates a number of pages from within an object
1001  *
1002  * Arguments:
1003  *	bytes  The number of bytes requested
1004  *	wait   Shall we wait?
1005  *
1006  * Returns:
1007  *	A pointer to the alloced memory or possibly
1008  *	NULL if M_NOWAIT is set.
1009  */
1010 static void *
1011 obj_alloc(uma_zone_t zone, int bytes, u_int8_t *flags, int wait)
1012 {
1013 	vm_object_t object;
1014 	vm_offset_t retkva, zkva;
1015 	vm_page_t p;
1016 	int pages, startpages;
1017 	uma_keg_t keg;
1018 
1019 	keg = zone_first_keg(zone);
1020 	object = keg->uk_obj;
1021 	retkva = 0;
1022 
1023 	/*
1024 	 * This looks a little weird since we're getting one page at a time.
1025 	 */
1026 	VM_OBJECT_LOCK(object);
1027 	p = TAILQ_LAST(&object->memq, pglist);
1028 	pages = p != NULL ? p->pindex + 1 : 0;
1029 	startpages = pages;
1030 	zkva = keg->uk_kva + pages * PAGE_SIZE;
1031 	for (; bytes > 0; bytes -= PAGE_SIZE) {
1032 		p = vm_page_alloc(object, pages,
1033 		    VM_ALLOC_INTERRUPT | VM_ALLOC_WIRED);
1034 		if (p == NULL) {
1035 			if (pages != startpages)
1036 				pmap_qremove(retkva, pages - startpages);
1037 			while (pages != startpages) {
1038 				pages--;
1039 				p = TAILQ_LAST(&object->memq, pglist);
1040 				vm_page_unwire(p, 0);
1041 				vm_page_free(p);
1042 			}
1043 			retkva = 0;
1044 			goto done;
1045 		}
1046 		pmap_qenter(zkva, &p, 1);
1047 		if (retkva == 0)
1048 			retkva = zkva;
1049 		zkva += PAGE_SIZE;
1050 		pages += 1;
1051 	}
1052 done:
1053 	VM_OBJECT_UNLOCK(object);
1054 	*flags = UMA_SLAB_PRIV;
1055 
1056 	return ((void *)retkva);
1057 }
1058 
1059 /*
1060  * Frees a number of pages to the system
1061  *
1062  * Arguments:
1063  *	mem   A pointer to the memory to be freed
1064  *	size  The size of the memory being freed
1065  *	flags The original p->us_flags field
1066  *
1067  * Returns:
1068  *	Nothing
1069  */
1070 static void
1071 page_free(void *mem, int size, u_int8_t flags)
1072 {
1073 	vm_map_t map;
1074 
1075 	if (flags & UMA_SLAB_KMEM)
1076 		map = kmem_map;
1077 	else if (flags & UMA_SLAB_KERNEL)
1078 		map = kernel_map;
1079 	else
1080 		panic("UMA: page_free used with invalid flags %d", flags);
1081 
1082 	kmem_free(map, (vm_offset_t)mem, size);
1083 }
1084 
1085 /*
1086  * Zero fill initializer
1087  *
1088  * Arguments/Returns follow uma_init specifications
1089  */
1090 static int
1091 zero_init(void *mem, int size, int flags)
1092 {
1093 	bzero(mem, size);
1094 	return (0);
1095 }
1096 
1097 /*
1098  * Finish creating a small uma keg.  This calculates ipers, and the keg size.
1099  *
1100  * Arguments
1101  *	keg  The zone we should initialize
1102  *
1103  * Returns
1104  *	Nothing
1105  */
1106 static void
1107 keg_small_init(uma_keg_t keg)
1108 {
1109 	u_int rsize;
1110 	u_int memused;
1111 	u_int wastedspace;
1112 	u_int shsize;
1113 
1114 	KASSERT(keg != NULL, ("Keg is null in keg_small_init"));
1115 	rsize = keg->uk_size;
1116 
1117 	if (rsize < UMA_SMALLEST_UNIT)
1118 		rsize = UMA_SMALLEST_UNIT;
1119 	if (rsize & keg->uk_align)
1120 		rsize = (rsize & ~keg->uk_align) + (keg->uk_align + 1);
1121 
1122 	keg->uk_rsize = rsize;
1123 	keg->uk_ppera = 1;
1124 
1125 	if (keg->uk_flags & UMA_ZONE_REFCNT) {
1126 		rsize += UMA_FRITMREF_SZ;	/* linkage & refcnt */
1127 		shsize = sizeof(struct uma_slab_refcnt);
1128 	} else {
1129 		rsize += UMA_FRITM_SZ;	/* Account for linkage */
1130 		shsize = sizeof(struct uma_slab);
1131 	}
1132 
1133 	keg->uk_ipers = (UMA_SLAB_SIZE - shsize) / rsize;
1134 	KASSERT(keg->uk_ipers != 0, ("keg_small_init: ipers is 0"));
1135 	memused = keg->uk_ipers * rsize + shsize;
1136 	wastedspace = UMA_SLAB_SIZE - memused;
1137 
1138 	/*
1139 	 * We can't do OFFPAGE if we're internal or if we've been
1140 	 * asked to not go to the VM for buckets.  If we do this we
1141 	 * may end up going to the VM (kmem_map) for slabs which we
1142 	 * do not want to do if we're UMA_ZFLAG_CACHEONLY as a
1143 	 * result of UMA_ZONE_VM, which clearly forbids it.
1144 	 */
1145 	if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) ||
1146 	    (keg->uk_flags & UMA_ZFLAG_CACHEONLY))
1147 		return;
1148 
1149 	if ((wastedspace >= UMA_MAX_WASTE) &&
1150 	    (keg->uk_ipers < (UMA_SLAB_SIZE / keg->uk_rsize))) {
1151 		keg->uk_ipers = UMA_SLAB_SIZE / keg->uk_rsize;
1152 		KASSERT(keg->uk_ipers <= 255,
1153 		    ("keg_small_init: keg->uk_ipers too high!"));
1154 #ifdef UMA_DEBUG
1155 		printf("UMA decided we need offpage slab headers for "
1156 		    "keg: %s, calculated wastedspace = %d, "
1157 		    "maximum wasted space allowed = %d, "
1158 		    "calculated ipers = %d, "
1159 		    "new wasted space = %d\n", keg->uk_name, wastedspace,
1160 		    UMA_MAX_WASTE, keg->uk_ipers,
1161 		    UMA_SLAB_SIZE - keg->uk_ipers * keg->uk_rsize);
1162 #endif
1163 		keg->uk_flags |= UMA_ZONE_OFFPAGE;
1164 		if ((keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1165 			keg->uk_flags |= UMA_ZONE_HASH;
1166 	}
1167 }
1168 
1169 /*
1170  * Finish creating a large (> UMA_SLAB_SIZE) uma kegs.  Just give in and do
1171  * OFFPAGE for now.  When I can allow for more dynamic slab sizes this will be
1172  * more complicated.
1173  *
1174  * Arguments
1175  *	keg  The keg we should initialize
1176  *
1177  * Returns
1178  *	Nothing
1179  */
1180 static void
1181 keg_large_init(uma_keg_t keg)
1182 {
1183 	int pages;
1184 
1185 	KASSERT(keg != NULL, ("Keg is null in keg_large_init"));
1186 	KASSERT((keg->uk_flags & UMA_ZFLAG_CACHEONLY) == 0,
1187 	    ("keg_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY keg"));
1188 
1189 	pages = keg->uk_size / UMA_SLAB_SIZE;
1190 
1191 	/* Account for remainder */
1192 	if ((pages * UMA_SLAB_SIZE) < keg->uk_size)
1193 		pages++;
1194 
1195 	keg->uk_ppera = pages;
1196 	keg->uk_ipers = 1;
1197 	keg->uk_rsize = keg->uk_size;
1198 
1199 	/* We can't do OFFPAGE if we're internal, bail out here. */
1200 	if (keg->uk_flags & UMA_ZFLAG_INTERNAL)
1201 		return;
1202 
1203 	keg->uk_flags |= UMA_ZONE_OFFPAGE;
1204 	if ((keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1205 		keg->uk_flags |= UMA_ZONE_HASH;
1206 }
1207 
1208 static void
1209 keg_cachespread_init(uma_keg_t keg)
1210 {
1211 	int alignsize;
1212 	int trailer;
1213 	int pages;
1214 	int rsize;
1215 
1216 	alignsize = keg->uk_align + 1;
1217 	rsize = keg->uk_size;
1218 	/*
1219 	 * We want one item to start on every align boundary in a page.  To
1220 	 * do this we will span pages.  We will also extend the item by the
1221 	 * size of align if it is an even multiple of align.  Otherwise, it
1222 	 * would fall on the same boundary every time.
1223 	 */
1224 	if (rsize & keg->uk_align)
1225 		rsize = (rsize & ~keg->uk_align) + alignsize;
1226 	if ((rsize & alignsize) == 0)
1227 		rsize += alignsize;
1228 	trailer = rsize - keg->uk_size;
1229 	pages = (rsize * (PAGE_SIZE / alignsize)) / PAGE_SIZE;
1230 	pages = MIN(pages, (128 * 1024) / PAGE_SIZE);
1231 	keg->uk_rsize = rsize;
1232 	keg->uk_ppera = pages;
1233 	keg->uk_ipers = ((pages * PAGE_SIZE) + trailer) / rsize;
1234 	keg->uk_flags |= UMA_ZONE_OFFPAGE | UMA_ZONE_VTOSLAB;
1235 	KASSERT(keg->uk_ipers <= uma_max_ipers,
1236 	    ("keg_small_init: keg->uk_ipers too high(%d) increase max_ipers",
1237 	    keg->uk_ipers));
1238 }
1239 
1240 /*
1241  * Keg header ctor.  This initializes all fields, locks, etc.  And inserts
1242  * the keg onto the global keg list.
1243  *
1244  * Arguments/Returns follow uma_ctor specifications
1245  *	udata  Actually uma_kctor_args
1246  */
1247 static int
1248 keg_ctor(void *mem, int size, void *udata, int flags)
1249 {
1250 	struct uma_kctor_args *arg = udata;
1251 	uma_keg_t keg = mem;
1252 	uma_zone_t zone;
1253 
1254 	bzero(keg, size);
1255 	keg->uk_size = arg->size;
1256 	keg->uk_init = arg->uminit;
1257 	keg->uk_fini = arg->fini;
1258 	keg->uk_align = arg->align;
1259 	keg->uk_free = 0;
1260 	keg->uk_pages = 0;
1261 	keg->uk_flags = arg->flags;
1262 	keg->uk_allocf = page_alloc;
1263 	keg->uk_freef = page_free;
1264 	keg->uk_recurse = 0;
1265 	keg->uk_slabzone = NULL;
1266 
1267 	/*
1268 	 * The master zone is passed to us at keg-creation time.
1269 	 */
1270 	zone = arg->zone;
1271 	keg->uk_name = zone->uz_name;
1272 
1273 	if (arg->flags & UMA_ZONE_VM)
1274 		keg->uk_flags |= UMA_ZFLAG_CACHEONLY;
1275 
1276 	if (arg->flags & UMA_ZONE_ZINIT)
1277 		keg->uk_init = zero_init;
1278 
1279 	if (arg->flags & UMA_ZONE_REFCNT || arg->flags & UMA_ZONE_MALLOC)
1280 		keg->uk_flags |= UMA_ZONE_VTOSLAB;
1281 
1282 	/*
1283 	 * The +UMA_FRITM_SZ added to uk_size is to account for the
1284 	 * linkage that is added to the size in keg_small_init().  If
1285 	 * we don't account for this here then we may end up in
1286 	 * keg_small_init() with a calculated 'ipers' of 0.
1287 	 */
1288 	if (keg->uk_flags & UMA_ZONE_REFCNT) {
1289 		if (keg->uk_flags & UMA_ZONE_CACHESPREAD)
1290 			keg_cachespread_init(keg);
1291 		else if ((keg->uk_size+UMA_FRITMREF_SZ) >
1292 		    (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt)))
1293 			keg_large_init(keg);
1294 		else
1295 			keg_small_init(keg);
1296 	} else {
1297 		if (keg->uk_flags & UMA_ZONE_CACHESPREAD)
1298 			keg_cachespread_init(keg);
1299 		else if ((keg->uk_size+UMA_FRITM_SZ) >
1300 		    (UMA_SLAB_SIZE - sizeof(struct uma_slab)))
1301 			keg_large_init(keg);
1302 		else
1303 			keg_small_init(keg);
1304 	}
1305 
1306 	if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
1307 		if (keg->uk_flags & UMA_ZONE_REFCNT)
1308 			keg->uk_slabzone = slabrefzone;
1309 		else
1310 			keg->uk_slabzone = slabzone;
1311 	}
1312 
1313 	/*
1314 	 * If we haven't booted yet we need allocations to go through the
1315 	 * startup cache until the vm is ready.
1316 	 */
1317 	if (keg->uk_ppera == 1) {
1318 #ifdef UMA_MD_SMALL_ALLOC
1319 		keg->uk_allocf = uma_small_alloc;
1320 		keg->uk_freef = uma_small_free;
1321 
1322 		if (booted < UMA_STARTUP)
1323 			keg->uk_allocf = startup_alloc;
1324 #else
1325 		if (booted < UMA_STARTUP2)
1326 			keg->uk_allocf = startup_alloc;
1327 #endif
1328 	} else if (booted < UMA_STARTUP2 &&
1329 	    (keg->uk_flags & UMA_ZFLAG_INTERNAL))
1330 		keg->uk_allocf = startup_alloc;
1331 
1332 	/*
1333 	 * Initialize keg's lock (shared among zones).
1334 	 */
1335 	if (arg->flags & UMA_ZONE_MTXCLASS)
1336 		KEG_LOCK_INIT(keg, 1);
1337 	else
1338 		KEG_LOCK_INIT(keg, 0);
1339 
1340 	/*
1341 	 * If we're putting the slab header in the actual page we need to
1342 	 * figure out where in each page it goes.  This calculates a right
1343 	 * justified offset into the memory on an ALIGN_PTR boundary.
1344 	 */
1345 	if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) {
1346 		u_int totsize;
1347 
1348 		/* Size of the slab struct and free list */
1349 		if (keg->uk_flags & UMA_ZONE_REFCNT)
1350 			totsize = sizeof(struct uma_slab_refcnt) +
1351 			    keg->uk_ipers * UMA_FRITMREF_SZ;
1352 		else
1353 			totsize = sizeof(struct uma_slab) +
1354 			    keg->uk_ipers * UMA_FRITM_SZ;
1355 
1356 		if (totsize & UMA_ALIGN_PTR)
1357 			totsize = (totsize & ~UMA_ALIGN_PTR) +
1358 			    (UMA_ALIGN_PTR + 1);
1359 		keg->uk_pgoff = (UMA_SLAB_SIZE * keg->uk_ppera) - totsize;
1360 
1361 		if (keg->uk_flags & UMA_ZONE_REFCNT)
1362 			totsize = keg->uk_pgoff + sizeof(struct uma_slab_refcnt)
1363 			    + keg->uk_ipers * UMA_FRITMREF_SZ;
1364 		else
1365 			totsize = keg->uk_pgoff + sizeof(struct uma_slab)
1366 			    + keg->uk_ipers * UMA_FRITM_SZ;
1367 
1368 		/*
1369 		 * The only way the following is possible is if with our
1370 		 * UMA_ALIGN_PTR adjustments we are now bigger than
1371 		 * UMA_SLAB_SIZE.  I haven't checked whether this is
1372 		 * mathematically possible for all cases, so we make
1373 		 * sure here anyway.
1374 		 */
1375 		if (totsize > UMA_SLAB_SIZE * keg->uk_ppera) {
1376 			printf("zone %s ipers %d rsize %d size %d\n",
1377 			    zone->uz_name, keg->uk_ipers, keg->uk_rsize,
1378 			    keg->uk_size);
1379 			panic("UMA slab won't fit.");
1380 		}
1381 	}
1382 
1383 	if (keg->uk_flags & UMA_ZONE_HASH)
1384 		hash_alloc(&keg->uk_hash);
1385 
1386 #ifdef UMA_DEBUG
1387 	printf("UMA: %s(%p) size %d(%d) flags %d ipers %d ppera %d out %d free %d\n",
1388 	    zone->uz_name, zone, keg->uk_size, keg->uk_rsize, keg->uk_flags,
1389 	    keg->uk_ipers, keg->uk_ppera,
1390 	    (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free);
1391 #endif
1392 
1393 	LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
1394 
1395 	mtx_lock(&uma_mtx);
1396 	LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
1397 	mtx_unlock(&uma_mtx);
1398 	return (0);
1399 }
1400 
1401 /*
1402  * Zone header ctor.  This initializes all fields, locks, etc.
1403  *
1404  * Arguments/Returns follow uma_ctor specifications
1405  *	udata  Actually uma_zctor_args
1406  */
1407 static int
1408 zone_ctor(void *mem, int size, void *udata, int flags)
1409 {
1410 	struct uma_zctor_args *arg = udata;
1411 	uma_zone_t zone = mem;
1412 	uma_zone_t z;
1413 	uma_keg_t keg;
1414 
1415 	bzero(zone, size);
1416 	zone->uz_name = arg->name;
1417 	zone->uz_ctor = arg->ctor;
1418 	zone->uz_dtor = arg->dtor;
1419 	zone->uz_slab = zone_fetch_slab;
1420 	zone->uz_init = NULL;
1421 	zone->uz_fini = NULL;
1422 	zone->uz_allocs = 0;
1423 	zone->uz_frees = 0;
1424 	zone->uz_fails = 0;
1425 	zone->uz_sleeps = 0;
1426 	zone->uz_fills = zone->uz_count = 0;
1427 	zone->uz_flags = 0;
1428 	keg = arg->keg;
1429 
1430 	if (arg->flags & UMA_ZONE_SECONDARY) {
1431 		KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
1432 		zone->uz_init = arg->uminit;
1433 		zone->uz_fini = arg->fini;
1434 		zone->uz_lock = &keg->uk_lock;
1435 		zone->uz_flags |= UMA_ZONE_SECONDARY;
1436 		mtx_lock(&uma_mtx);
1437 		ZONE_LOCK(zone);
1438 		LIST_FOREACH(z, &keg->uk_zones, uz_link) {
1439 			if (LIST_NEXT(z, uz_link) == NULL) {
1440 				LIST_INSERT_AFTER(z, zone, uz_link);
1441 				break;
1442 			}
1443 		}
1444 		ZONE_UNLOCK(zone);
1445 		mtx_unlock(&uma_mtx);
1446 	} else if (keg == NULL) {
1447 		if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
1448 		    arg->align, arg->flags)) == NULL)
1449 			return (ENOMEM);
1450 	} else {
1451 		struct uma_kctor_args karg;
1452 		int error;
1453 
1454 		/* We should only be here from uma_startup() */
1455 		karg.size = arg->size;
1456 		karg.uminit = arg->uminit;
1457 		karg.fini = arg->fini;
1458 		karg.align = arg->align;
1459 		karg.flags = arg->flags;
1460 		karg.zone = zone;
1461 		error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
1462 		    flags);
1463 		if (error)
1464 			return (error);
1465 	}
1466 	/*
1467 	 * Link in the first keg.
1468 	 */
1469 	zone->uz_klink.kl_keg = keg;
1470 	LIST_INSERT_HEAD(&zone->uz_kegs, &zone->uz_klink, kl_link);
1471 	zone->uz_lock = &keg->uk_lock;
1472 	zone->uz_size = keg->uk_size;
1473 	zone->uz_flags |= (keg->uk_flags &
1474 	    (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
1475 
1476 	/*
1477 	 * Some internal zones don't have room allocated for the per cpu
1478 	 * caches.  If we're internal, bail out here.
1479 	 */
1480 	if (keg->uk_flags & UMA_ZFLAG_INTERNAL) {
1481 		KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
1482 		    ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
1483 		return (0);
1484 	}
1485 
1486 	if (keg->uk_flags & UMA_ZONE_MAXBUCKET)
1487 		zone->uz_count = BUCKET_MAX;
1488 	else if (keg->uk_ipers <= BUCKET_MAX)
1489 		zone->uz_count = keg->uk_ipers;
1490 	else
1491 		zone->uz_count = BUCKET_MAX;
1492 	return (0);
1493 }
1494 
1495 /*
1496  * Keg header dtor.  This frees all data, destroys locks, frees the hash
1497  * table and removes the keg from the global list.
1498  *
1499  * Arguments/Returns follow uma_dtor specifications
1500  *	udata  unused
1501  */
1502 static void
1503 keg_dtor(void *arg, int size, void *udata)
1504 {
1505 	uma_keg_t keg;
1506 
1507 	keg = (uma_keg_t)arg;
1508 	KEG_LOCK(keg);
1509 	if (keg->uk_free != 0) {
1510 		printf("Freed UMA keg was not empty (%d items). "
1511 		    " Lost %d pages of memory.\n",
1512 		    keg->uk_free, keg->uk_pages);
1513 	}
1514 	KEG_UNLOCK(keg);
1515 
1516 	hash_free(&keg->uk_hash);
1517 
1518 	KEG_LOCK_FINI(keg);
1519 }
1520 
1521 /*
1522  * Zone header dtor.
1523  *
1524  * Arguments/Returns follow uma_dtor specifications
1525  *	udata  unused
1526  */
1527 static void
1528 zone_dtor(void *arg, int size, void *udata)
1529 {
1530 	uma_klink_t klink;
1531 	uma_zone_t zone;
1532 	uma_keg_t keg;
1533 
1534 	zone = (uma_zone_t)arg;
1535 	keg = zone_first_keg(zone);
1536 
1537 	if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
1538 		cache_drain(zone);
1539 
1540 	mtx_lock(&uma_mtx);
1541 	LIST_REMOVE(zone, uz_link);
1542 	mtx_unlock(&uma_mtx);
1543 	/*
1544 	 * XXX there are some races here where
1545 	 * the zone can be drained but zone lock
1546 	 * released and then refilled before we
1547 	 * remove it... we dont care for now
1548 	 */
1549 	zone_drain_wait(zone, M_WAITOK);
1550 	/*
1551 	 * Unlink all of our kegs.
1552 	 */
1553 	while ((klink = LIST_FIRST(&zone->uz_kegs)) != NULL) {
1554 		klink->kl_keg = NULL;
1555 		LIST_REMOVE(klink, kl_link);
1556 		if (klink == &zone->uz_klink)
1557 			continue;
1558 		free(klink, M_TEMP);
1559 	}
1560 	/*
1561 	 * We only destroy kegs from non secondary zones.
1562 	 */
1563 	if ((zone->uz_flags & UMA_ZONE_SECONDARY) == 0)  {
1564 		mtx_lock(&uma_mtx);
1565 		LIST_REMOVE(keg, uk_link);
1566 		mtx_unlock(&uma_mtx);
1567 		zone_free_item(kegs, keg, NULL, SKIP_NONE,
1568 		    ZFREE_STATFREE);
1569 	}
1570 }
1571 
1572 /*
1573  * Traverses every zone in the system and calls a callback
1574  *
1575  * Arguments:
1576  *	zfunc  A pointer to a function which accepts a zone
1577  *		as an argument.
1578  *
1579  * Returns:
1580  *	Nothing
1581  */
1582 static void
1583 zone_foreach(void (*zfunc)(uma_zone_t))
1584 {
1585 	uma_keg_t keg;
1586 	uma_zone_t zone;
1587 
1588 	mtx_lock(&uma_mtx);
1589 	LIST_FOREACH(keg, &uma_kegs, uk_link) {
1590 		LIST_FOREACH(zone, &keg->uk_zones, uz_link)
1591 			zfunc(zone);
1592 	}
1593 	mtx_unlock(&uma_mtx);
1594 }
1595 
1596 /* Public functions */
1597 /* See uma.h */
1598 void
1599 uma_startup(void *bootmem, int boot_pages)
1600 {
1601 	struct uma_zctor_args args;
1602 	uma_slab_t slab;
1603 	u_int slabsize;
1604 	u_int objsize, totsize, wsize;
1605 	int i;
1606 
1607 #ifdef UMA_DEBUG
1608 	printf("Creating uma keg headers zone and keg.\n");
1609 #endif
1610 	mtx_init(&uma_mtx, "UMA lock", NULL, MTX_DEF);
1611 
1612 	/*
1613 	 * Figure out the maximum number of items-per-slab we'll have if
1614 	 * we're using the OFFPAGE slab header to track free items, given
1615 	 * all possible object sizes and the maximum desired wastage
1616 	 * (UMA_MAX_WASTE).
1617 	 *
1618 	 * We iterate until we find an object size for
1619 	 * which the calculated wastage in keg_small_init() will be
1620 	 * enough to warrant OFFPAGE.  Since wastedspace versus objsize
1621 	 * is an overall increasing see-saw function, we find the smallest
1622 	 * objsize such that the wastage is always acceptable for objects
1623 	 * with that objsize or smaller.  Since a smaller objsize always
1624 	 * generates a larger possible uma_max_ipers, we use this computed
1625 	 * objsize to calculate the largest ipers possible.  Since the
1626 	 * ipers calculated for OFFPAGE slab headers is always larger than
1627 	 * the ipers initially calculated in keg_small_init(), we use
1628 	 * the former's equation (UMA_SLAB_SIZE / keg->uk_rsize) to
1629 	 * obtain the maximum ipers possible for offpage slab headers.
1630 	 *
1631 	 * It should be noted that ipers versus objsize is an inversly
1632 	 * proportional function which drops off rather quickly so as
1633 	 * long as our UMA_MAX_WASTE is such that the objsize we calculate
1634 	 * falls into the portion of the inverse relation AFTER the steep
1635 	 * falloff, then uma_max_ipers shouldn't be too high (~10 on i386).
1636 	 *
1637 	 * Note that we have 8-bits (1 byte) to use as a freelist index
1638 	 * inside the actual slab header itself and this is enough to
1639 	 * accomodate us.  In the worst case, a UMA_SMALLEST_UNIT sized
1640 	 * object with offpage slab header would have ipers =
1641 	 * UMA_SLAB_SIZE / UMA_SMALLEST_UNIT (currently = 256), which is
1642 	 * 1 greater than what our byte-integer freelist index can
1643 	 * accomodate, but we know that this situation never occurs as
1644 	 * for UMA_SMALLEST_UNIT-sized objects, we will never calculate
1645 	 * that we need to go to offpage slab headers.  Or, if we do,
1646 	 * then we trap that condition below and panic in the INVARIANTS case.
1647 	 */
1648 	wsize = UMA_SLAB_SIZE - sizeof(struct uma_slab) - UMA_MAX_WASTE;
1649 	totsize = wsize;
1650 	objsize = UMA_SMALLEST_UNIT;
1651 	while (totsize >= wsize) {
1652 		totsize = (UMA_SLAB_SIZE - sizeof(struct uma_slab)) /
1653 		    (objsize + UMA_FRITM_SZ);
1654 		totsize *= (UMA_FRITM_SZ + objsize);
1655 		objsize++;
1656 	}
1657 	if (objsize > UMA_SMALLEST_UNIT)
1658 		objsize--;
1659 	uma_max_ipers = MAX(UMA_SLAB_SIZE / objsize, 64);
1660 
1661 	wsize = UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt) - UMA_MAX_WASTE;
1662 	totsize = wsize;
1663 	objsize = UMA_SMALLEST_UNIT;
1664 	while (totsize >= wsize) {
1665 		totsize = (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt)) /
1666 		    (objsize + UMA_FRITMREF_SZ);
1667 		totsize *= (UMA_FRITMREF_SZ + objsize);
1668 		objsize++;
1669 	}
1670 	if (objsize > UMA_SMALLEST_UNIT)
1671 		objsize--;
1672 	uma_max_ipers_ref = MAX(UMA_SLAB_SIZE / objsize, 64);
1673 
1674 	KASSERT((uma_max_ipers_ref <= 255) && (uma_max_ipers <= 255),
1675 	    ("uma_startup: calculated uma_max_ipers values too large!"));
1676 
1677 #ifdef UMA_DEBUG
1678 	printf("Calculated uma_max_ipers (for OFFPAGE) is %d\n", uma_max_ipers);
1679 	printf("Calculated uma_max_ipers_slab (for OFFPAGE) is %d\n",
1680 	    uma_max_ipers_ref);
1681 #endif
1682 
1683 	/* "manually" create the initial zone */
1684 	args.name = "UMA Kegs";
1685 	args.size = sizeof(struct uma_keg);
1686 	args.ctor = keg_ctor;
1687 	args.dtor = keg_dtor;
1688 	args.uminit = zero_init;
1689 	args.fini = NULL;
1690 	args.keg = &masterkeg;
1691 	args.align = 32 - 1;
1692 	args.flags = UMA_ZFLAG_INTERNAL;
1693 	/* The initial zone has no Per cpu queues so it's smaller */
1694 	zone_ctor(kegs, sizeof(struct uma_zone), &args, M_WAITOK);
1695 
1696 #ifdef UMA_DEBUG
1697 	printf("Filling boot free list.\n");
1698 #endif
1699 	for (i = 0; i < boot_pages; i++) {
1700 		slab = (uma_slab_t)((u_int8_t *)bootmem + (i * UMA_SLAB_SIZE));
1701 		slab->us_data = (u_int8_t *)slab;
1702 		slab->us_flags = UMA_SLAB_BOOT;
1703 		LIST_INSERT_HEAD(&uma_boot_pages, slab, us_link);
1704 	}
1705 	mtx_init(&uma_boot_pages_mtx, "UMA boot pages", NULL, MTX_DEF);
1706 
1707 #ifdef UMA_DEBUG
1708 	printf("Creating uma zone headers zone and keg.\n");
1709 #endif
1710 	args.name = "UMA Zones";
1711 	args.size = sizeof(struct uma_zone) +
1712 	    (sizeof(struct uma_cache) * (mp_maxid + 1));
1713 	args.ctor = zone_ctor;
1714 	args.dtor = zone_dtor;
1715 	args.uminit = zero_init;
1716 	args.fini = NULL;
1717 	args.keg = NULL;
1718 	args.align = 32 - 1;
1719 	args.flags = UMA_ZFLAG_INTERNAL;
1720 	/* The initial zone has no Per cpu queues so it's smaller */
1721 	zone_ctor(zones, sizeof(struct uma_zone), &args, M_WAITOK);
1722 
1723 #ifdef UMA_DEBUG
1724 	printf("Initializing pcpu cache locks.\n");
1725 #endif
1726 #ifdef UMA_DEBUG
1727 	printf("Creating slab and hash zones.\n");
1728 #endif
1729 
1730 	/*
1731 	 * This is the max number of free list items we'll have with
1732 	 * offpage slabs.
1733 	 */
1734 	slabsize = uma_max_ipers * UMA_FRITM_SZ;
1735 	slabsize += sizeof(struct uma_slab);
1736 
1737 	/* Now make a zone for slab headers */
1738 	slabzone = uma_zcreate("UMA Slabs",
1739 				slabsize,
1740 				NULL, NULL, NULL, NULL,
1741 				UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1742 
1743 	/*
1744 	 * We also create a zone for the bigger slabs with reference
1745 	 * counts in them, to accomodate UMA_ZONE_REFCNT zones.
1746 	 */
1747 	slabsize = uma_max_ipers_ref * UMA_FRITMREF_SZ;
1748 	slabsize += sizeof(struct uma_slab_refcnt);
1749 	slabrefzone = uma_zcreate("UMA RCntSlabs",
1750 				  slabsize,
1751 				  NULL, NULL, NULL, NULL,
1752 				  UMA_ALIGN_PTR,
1753 				  UMA_ZFLAG_INTERNAL);
1754 
1755 	hashzone = uma_zcreate("UMA Hash",
1756 	    sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
1757 	    NULL, NULL, NULL, NULL,
1758 	    UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1759 
1760 	bucket_init();
1761 
1762 	booted = UMA_STARTUP;
1763 
1764 #ifdef UMA_DEBUG
1765 	printf("UMA startup complete.\n");
1766 #endif
1767 }
1768 
1769 /* see uma.h */
1770 void
1771 uma_startup2(void)
1772 {
1773 	booted = UMA_STARTUP2;
1774 	bucket_enable();
1775 #ifdef UMA_DEBUG
1776 	printf("UMA startup2 complete.\n");
1777 #endif
1778 }
1779 
1780 /*
1781  * Initialize our callout handle
1782  *
1783  */
1784 
1785 static void
1786 uma_startup3(void)
1787 {
1788 #ifdef UMA_DEBUG
1789 	printf("Starting callout.\n");
1790 #endif
1791 	callout_init(&uma_callout, CALLOUT_MPSAFE);
1792 	callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
1793 #ifdef UMA_DEBUG
1794 	printf("UMA startup3 complete.\n");
1795 #endif
1796 }
1797 
1798 static uma_keg_t
1799 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
1800 		int align, u_int32_t flags)
1801 {
1802 	struct uma_kctor_args args;
1803 
1804 	args.size = size;
1805 	args.uminit = uminit;
1806 	args.fini = fini;
1807 	args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
1808 	args.flags = flags;
1809 	args.zone = zone;
1810 	return (zone_alloc_item(kegs, &args, M_WAITOK));
1811 }
1812 
1813 /* See uma.h */
1814 void
1815 uma_set_align(int align)
1816 {
1817 
1818 	if (align != UMA_ALIGN_CACHE)
1819 		uma_align_cache = align;
1820 }
1821 
1822 /* See uma.h */
1823 uma_zone_t
1824 uma_zcreate(char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
1825 		uma_init uminit, uma_fini fini, int align, u_int32_t flags)
1826 
1827 {
1828 	struct uma_zctor_args args;
1829 
1830 	/* This stuff is essential for the zone ctor */
1831 	args.name = name;
1832 	args.size = size;
1833 	args.ctor = ctor;
1834 	args.dtor = dtor;
1835 	args.uminit = uminit;
1836 	args.fini = fini;
1837 	args.align = align;
1838 	args.flags = flags;
1839 	args.keg = NULL;
1840 
1841 	return (zone_alloc_item(zones, &args, M_WAITOK));
1842 }
1843 
1844 /* See uma.h */
1845 uma_zone_t
1846 uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor,
1847 		    uma_init zinit, uma_fini zfini, uma_zone_t master)
1848 {
1849 	struct uma_zctor_args args;
1850 	uma_keg_t keg;
1851 
1852 	keg = zone_first_keg(master);
1853 	args.name = name;
1854 	args.size = keg->uk_size;
1855 	args.ctor = ctor;
1856 	args.dtor = dtor;
1857 	args.uminit = zinit;
1858 	args.fini = zfini;
1859 	args.align = keg->uk_align;
1860 	args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
1861 	args.keg = keg;
1862 
1863 	/* XXX Attaches only one keg of potentially many. */
1864 	return (zone_alloc_item(zones, &args, M_WAITOK));
1865 }
1866 
1867 static void
1868 zone_lock_pair(uma_zone_t a, uma_zone_t b)
1869 {
1870 	if (a < b) {
1871 		ZONE_LOCK(a);
1872 		mtx_lock_flags(b->uz_lock, MTX_DUPOK);
1873 	} else {
1874 		ZONE_LOCK(b);
1875 		mtx_lock_flags(a->uz_lock, MTX_DUPOK);
1876 	}
1877 }
1878 
1879 static void
1880 zone_unlock_pair(uma_zone_t a, uma_zone_t b)
1881 {
1882 
1883 	ZONE_UNLOCK(a);
1884 	ZONE_UNLOCK(b);
1885 }
1886 
1887 int
1888 uma_zsecond_add(uma_zone_t zone, uma_zone_t master)
1889 {
1890 	uma_klink_t klink;
1891 	uma_klink_t kl;
1892 	int error;
1893 
1894 	error = 0;
1895 	klink = malloc(sizeof(*klink), M_TEMP, M_WAITOK | M_ZERO);
1896 
1897 	zone_lock_pair(zone, master);
1898 	/*
1899 	 * zone must use vtoslab() to resolve objects and must already be
1900 	 * a secondary.
1901 	 */
1902 	if ((zone->uz_flags & (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY))
1903 	    != (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) {
1904 		error = EINVAL;
1905 		goto out;
1906 	}
1907 	/*
1908 	 * The new master must also use vtoslab().
1909 	 */
1910 	if ((zone->uz_flags & UMA_ZONE_VTOSLAB) != UMA_ZONE_VTOSLAB) {
1911 		error = EINVAL;
1912 		goto out;
1913 	}
1914 	/*
1915 	 * Both must either be refcnt, or not be refcnt.
1916 	 */
1917 	if ((zone->uz_flags & UMA_ZONE_REFCNT) !=
1918 	    (master->uz_flags & UMA_ZONE_REFCNT)) {
1919 		error = EINVAL;
1920 		goto out;
1921 	}
1922 	/*
1923 	 * The underlying object must be the same size.  rsize
1924 	 * may be different.
1925 	 */
1926 	if (master->uz_size != zone->uz_size) {
1927 		error = E2BIG;
1928 		goto out;
1929 	}
1930 	/*
1931 	 * Put it at the end of the list.
1932 	 */
1933 	klink->kl_keg = zone_first_keg(master);
1934 	LIST_FOREACH(kl, &zone->uz_kegs, kl_link) {
1935 		if (LIST_NEXT(kl, kl_link) == NULL) {
1936 			LIST_INSERT_AFTER(kl, klink, kl_link);
1937 			break;
1938 		}
1939 	}
1940 	klink = NULL;
1941 	zone->uz_flags |= UMA_ZFLAG_MULTI;
1942 	zone->uz_slab = zone_fetch_slab_multi;
1943 
1944 out:
1945 	zone_unlock_pair(zone, master);
1946 	if (klink != NULL)
1947 		free(klink, M_TEMP);
1948 
1949 	return (error);
1950 }
1951 
1952 
1953 /* See uma.h */
1954 void
1955 uma_zdestroy(uma_zone_t zone)
1956 {
1957 
1958 	zone_free_item(zones, zone, NULL, SKIP_NONE, ZFREE_STATFREE);
1959 }
1960 
1961 /* See uma.h */
1962 void *
1963 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
1964 {
1965 	void *item;
1966 	uma_cache_t cache;
1967 	uma_bucket_t bucket;
1968 	int cpu;
1969 
1970 	/* This is the fast path allocation */
1971 #ifdef UMA_DEBUG_ALLOC_1
1972 	printf("Allocating one item from %s(%p)\n", zone->uz_name, zone);
1973 #endif
1974 	CTR3(KTR_UMA, "uma_zalloc_arg thread %x zone %s flags %d", curthread,
1975 	    zone->uz_name, flags);
1976 
1977 	if (flags & M_WAITOK) {
1978 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
1979 		    "uma_zalloc_arg: zone \"%s\"", zone->uz_name);
1980 	}
1981 
1982 	/*
1983 	 * If possible, allocate from the per-CPU cache.  There are two
1984 	 * requirements for safe access to the per-CPU cache: (1) the thread
1985 	 * accessing the cache must not be preempted or yield during access,
1986 	 * and (2) the thread must not migrate CPUs without switching which
1987 	 * cache it accesses.  We rely on a critical section to prevent
1988 	 * preemption and migration.  We release the critical section in
1989 	 * order to acquire the zone mutex if we are unable to allocate from
1990 	 * the current cache; when we re-acquire the critical section, we
1991 	 * must detect and handle migration if it has occurred.
1992 	 */
1993 zalloc_restart:
1994 	critical_enter();
1995 	cpu = curcpu;
1996 	cache = &zone->uz_cpu[cpu];
1997 
1998 zalloc_start:
1999 	bucket = cache->uc_allocbucket;
2000 
2001 	if (bucket) {
2002 		if (bucket->ub_cnt > 0) {
2003 			bucket->ub_cnt--;
2004 			item = bucket->ub_bucket[bucket->ub_cnt];
2005 #ifdef INVARIANTS
2006 			bucket->ub_bucket[bucket->ub_cnt] = NULL;
2007 #endif
2008 			KASSERT(item != NULL,
2009 			    ("uma_zalloc: Bucket pointer mangled."));
2010 			cache->uc_allocs++;
2011 			critical_exit();
2012 #ifdef INVARIANTS
2013 			ZONE_LOCK(zone);
2014 			uma_dbg_alloc(zone, NULL, item);
2015 			ZONE_UNLOCK(zone);
2016 #endif
2017 			if (zone->uz_ctor != NULL) {
2018 				if (zone->uz_ctor(item, zone->uz_size,
2019 				    udata, flags) != 0) {
2020 					zone_free_item(zone, item, udata,
2021 					    SKIP_DTOR, ZFREE_STATFAIL |
2022 					    ZFREE_STATFREE);
2023 					return (NULL);
2024 				}
2025 			}
2026 			if (flags & M_ZERO)
2027 				bzero(item, zone->uz_size);
2028 			return (item);
2029 		} else if (cache->uc_freebucket) {
2030 			/*
2031 			 * We have run out of items in our allocbucket.
2032 			 * See if we can switch with our free bucket.
2033 			 */
2034 			if (cache->uc_freebucket->ub_cnt > 0) {
2035 #ifdef UMA_DEBUG_ALLOC
2036 				printf("uma_zalloc: Swapping empty with"
2037 				    " alloc.\n");
2038 #endif
2039 				bucket = cache->uc_freebucket;
2040 				cache->uc_freebucket = cache->uc_allocbucket;
2041 				cache->uc_allocbucket = bucket;
2042 
2043 				goto zalloc_start;
2044 			}
2045 		}
2046 	}
2047 	/*
2048 	 * Attempt to retrieve the item from the per-CPU cache has failed, so
2049 	 * we must go back to the zone.  This requires the zone lock, so we
2050 	 * must drop the critical section, then re-acquire it when we go back
2051 	 * to the cache.  Since the critical section is released, we may be
2052 	 * preempted or migrate.  As such, make sure not to maintain any
2053 	 * thread-local state specific to the cache from prior to releasing
2054 	 * the critical section.
2055 	 */
2056 	critical_exit();
2057 	ZONE_LOCK(zone);
2058 	critical_enter();
2059 	cpu = curcpu;
2060 	cache = &zone->uz_cpu[cpu];
2061 	bucket = cache->uc_allocbucket;
2062 	if (bucket != NULL) {
2063 		if (bucket->ub_cnt > 0) {
2064 			ZONE_UNLOCK(zone);
2065 			goto zalloc_start;
2066 		}
2067 		bucket = cache->uc_freebucket;
2068 		if (bucket != NULL && bucket->ub_cnt > 0) {
2069 			ZONE_UNLOCK(zone);
2070 			goto zalloc_start;
2071 		}
2072 	}
2073 
2074 	/* Since we have locked the zone we may as well send back our stats */
2075 	zone->uz_allocs += cache->uc_allocs;
2076 	cache->uc_allocs = 0;
2077 	zone->uz_frees += cache->uc_frees;
2078 	cache->uc_frees = 0;
2079 
2080 	/* Our old one is now a free bucket */
2081 	if (cache->uc_allocbucket) {
2082 		KASSERT(cache->uc_allocbucket->ub_cnt == 0,
2083 		    ("uma_zalloc_arg: Freeing a non free bucket."));
2084 		LIST_INSERT_HEAD(&zone->uz_free_bucket,
2085 		    cache->uc_allocbucket, ub_link);
2086 		cache->uc_allocbucket = NULL;
2087 	}
2088 
2089 	/* Check the free list for a new alloc bucket */
2090 	if ((bucket = LIST_FIRST(&zone->uz_full_bucket)) != NULL) {
2091 		KASSERT(bucket->ub_cnt != 0,
2092 		    ("uma_zalloc_arg: Returning an empty bucket."));
2093 
2094 		LIST_REMOVE(bucket, ub_link);
2095 		cache->uc_allocbucket = bucket;
2096 		ZONE_UNLOCK(zone);
2097 		goto zalloc_start;
2098 	}
2099 	/* We are no longer associated with this CPU. */
2100 	critical_exit();
2101 
2102 	/* Bump up our uz_count so we get here less */
2103 	if (zone->uz_count < BUCKET_MAX)
2104 		zone->uz_count++;
2105 
2106 	/*
2107 	 * Now lets just fill a bucket and put it on the free list.  If that
2108 	 * works we'll restart the allocation from the begining.
2109 	 */
2110 	if (zone_alloc_bucket(zone, flags)) {
2111 		ZONE_UNLOCK(zone);
2112 		goto zalloc_restart;
2113 	}
2114 	ZONE_UNLOCK(zone);
2115 	/*
2116 	 * We may not be able to get a bucket so return an actual item.
2117 	 */
2118 #ifdef UMA_DEBUG
2119 	printf("uma_zalloc_arg: Bucketzone returned NULL\n");
2120 #endif
2121 
2122 	item = zone_alloc_item(zone, udata, flags);
2123 	return (item);
2124 }
2125 
2126 static uma_slab_t
2127 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int flags)
2128 {
2129 	uma_slab_t slab;
2130 
2131 	mtx_assert(&keg->uk_lock, MA_OWNED);
2132 	slab = NULL;
2133 
2134 	for (;;) {
2135 		/*
2136 		 * Find a slab with some space.  Prefer slabs that are partially
2137 		 * used over those that are totally full.  This helps to reduce
2138 		 * fragmentation.
2139 		 */
2140 		if (keg->uk_free != 0) {
2141 			if (!LIST_EMPTY(&keg->uk_part_slab)) {
2142 				slab = LIST_FIRST(&keg->uk_part_slab);
2143 			} else {
2144 				slab = LIST_FIRST(&keg->uk_free_slab);
2145 				LIST_REMOVE(slab, us_link);
2146 				LIST_INSERT_HEAD(&keg->uk_part_slab, slab,
2147 				    us_link);
2148 			}
2149 			MPASS(slab->us_keg == keg);
2150 			return (slab);
2151 		}
2152 
2153 		/*
2154 		 * M_NOVM means don't ask at all!
2155 		 */
2156 		if (flags & M_NOVM)
2157 			break;
2158 
2159 		if (keg->uk_maxpages && keg->uk_pages >= keg->uk_maxpages) {
2160 			keg->uk_flags |= UMA_ZFLAG_FULL;
2161 			/*
2162 			 * If this is not a multi-zone, set the FULL bit.
2163 			 * Otherwise slab_multi() takes care of it.
2164 			 */
2165 			if ((zone->uz_flags & UMA_ZFLAG_MULTI) == 0)
2166 				zone->uz_flags |= UMA_ZFLAG_FULL;
2167 			if (flags & M_NOWAIT)
2168 				break;
2169 			msleep(keg, &keg->uk_lock, PVM, "keglimit", 0);
2170 			continue;
2171 		}
2172 		keg->uk_recurse++;
2173 		slab = keg_alloc_slab(keg, zone, flags);
2174 		keg->uk_recurse--;
2175 		/*
2176 		 * If we got a slab here it's safe to mark it partially used
2177 		 * and return.  We assume that the caller is going to remove
2178 		 * at least one item.
2179 		 */
2180 		if (slab) {
2181 			MPASS(slab->us_keg == keg);
2182 			LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2183 			return (slab);
2184 		}
2185 		/*
2186 		 * We might not have been able to get a slab but another cpu
2187 		 * could have while we were unlocked.  Check again before we
2188 		 * fail.
2189 		 */
2190 		flags |= M_NOVM;
2191 	}
2192 	return (slab);
2193 }
2194 
2195 static inline void
2196 zone_relock(uma_zone_t zone, uma_keg_t keg)
2197 {
2198 	if (zone->uz_lock != &keg->uk_lock) {
2199 		KEG_UNLOCK(keg);
2200 		ZONE_LOCK(zone);
2201 	}
2202 }
2203 
2204 static inline void
2205 keg_relock(uma_keg_t keg, uma_zone_t zone)
2206 {
2207 	if (zone->uz_lock != &keg->uk_lock) {
2208 		ZONE_UNLOCK(zone);
2209 		KEG_LOCK(keg);
2210 	}
2211 }
2212 
2213 static uma_slab_t
2214 zone_fetch_slab(uma_zone_t zone, uma_keg_t keg, int flags)
2215 {
2216 	uma_slab_t slab;
2217 
2218 	if (keg == NULL)
2219 		keg = zone_first_keg(zone);
2220 	/*
2221 	 * This is to prevent us from recursively trying to allocate
2222 	 * buckets.  The problem is that if an allocation forces us to
2223 	 * grab a new bucket we will call page_alloc, which will go off
2224 	 * and cause the vm to allocate vm_map_entries.  If we need new
2225 	 * buckets there too we will recurse in kmem_alloc and bad
2226 	 * things happen.  So instead we return a NULL bucket, and make
2227 	 * the code that allocates buckets smart enough to deal with it
2228 	 */
2229 	if (keg->uk_flags & UMA_ZFLAG_BUCKET && keg->uk_recurse != 0)
2230 		return (NULL);
2231 
2232 	for (;;) {
2233 		slab = keg_fetch_slab(keg, zone, flags);
2234 		if (slab)
2235 			return (slab);
2236 		if (flags & (M_NOWAIT | M_NOVM))
2237 			break;
2238 	}
2239 	return (NULL);
2240 }
2241 
2242 /*
2243  * uma_zone_fetch_slab_multi:  Fetches a slab from one available keg.  Returns
2244  * with the keg locked.  Caller must call zone_relock() afterwards if the
2245  * zone lock is required.  On NULL the zone lock is held.
2246  *
2247  * The last pointer is used to seed the search.  It is not required.
2248  */
2249 static uma_slab_t
2250 zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int rflags)
2251 {
2252 	uma_klink_t klink;
2253 	uma_slab_t slab;
2254 	uma_keg_t keg;
2255 	int flags;
2256 	int empty;
2257 	int full;
2258 
2259 	/*
2260 	 * Don't wait on the first pass.  This will skip limit tests
2261 	 * as well.  We don't want to block if we can find a provider
2262 	 * without blocking.
2263 	 */
2264 	flags = (rflags & ~M_WAITOK) | M_NOWAIT;
2265 	/*
2266 	 * Use the last slab allocated as a hint for where to start
2267 	 * the search.
2268 	 */
2269 	if (last) {
2270 		slab = keg_fetch_slab(last, zone, flags);
2271 		if (slab)
2272 			return (slab);
2273 		zone_relock(zone, last);
2274 		last = NULL;
2275 	}
2276 	/*
2277 	 * Loop until we have a slab incase of transient failures
2278 	 * while M_WAITOK is specified.  I'm not sure this is 100%
2279 	 * required but we've done it for so long now.
2280 	 */
2281 	for (;;) {
2282 		empty = 0;
2283 		full = 0;
2284 		/*
2285 		 * Search the available kegs for slabs.  Be careful to hold the
2286 		 * correct lock while calling into the keg layer.
2287 		 */
2288 		LIST_FOREACH(klink, &zone->uz_kegs, kl_link) {
2289 			keg = klink->kl_keg;
2290 			keg_relock(keg, zone);
2291 			if ((keg->uk_flags & UMA_ZFLAG_FULL) == 0) {
2292 				slab = keg_fetch_slab(keg, zone, flags);
2293 				if (slab)
2294 					return (slab);
2295 			}
2296 			if (keg->uk_flags & UMA_ZFLAG_FULL)
2297 				full++;
2298 			else
2299 				empty++;
2300 			zone_relock(zone, keg);
2301 		}
2302 		if (rflags & (M_NOWAIT | M_NOVM))
2303 			break;
2304 		flags = rflags;
2305 		/*
2306 		 * All kegs are full.  XXX We can't atomically check all kegs
2307 		 * and sleep so just sleep for a short period and retry.
2308 		 */
2309 		if (full && !empty) {
2310 			zone->uz_flags |= UMA_ZFLAG_FULL;
2311 			zone->uz_sleeps++;
2312 			msleep(zone, zone->uz_lock, PVM, "zonelimit", hz/100);
2313 			zone->uz_flags &= ~UMA_ZFLAG_FULL;
2314 			continue;
2315 		}
2316 	}
2317 	return (NULL);
2318 }
2319 
2320 static void *
2321 slab_alloc_item(uma_zone_t zone, uma_slab_t slab)
2322 {
2323 	uma_keg_t keg;
2324 	uma_slabrefcnt_t slabref;
2325 	void *item;
2326 	u_int8_t freei;
2327 
2328 	keg = slab->us_keg;
2329 	mtx_assert(&keg->uk_lock, MA_OWNED);
2330 
2331 	freei = slab->us_firstfree;
2332 	if (keg->uk_flags & UMA_ZONE_REFCNT) {
2333 		slabref = (uma_slabrefcnt_t)slab;
2334 		slab->us_firstfree = slabref->us_freelist[freei].us_item;
2335 	} else {
2336 		slab->us_firstfree = slab->us_freelist[freei].us_item;
2337 	}
2338 	item = slab->us_data + (keg->uk_rsize * freei);
2339 
2340 	slab->us_freecount--;
2341 	keg->uk_free--;
2342 #ifdef INVARIANTS
2343 	uma_dbg_alloc(zone, slab, item);
2344 #endif
2345 	/* Move this slab to the full list */
2346 	if (slab->us_freecount == 0) {
2347 		LIST_REMOVE(slab, us_link);
2348 		LIST_INSERT_HEAD(&keg->uk_full_slab, slab, us_link);
2349 	}
2350 
2351 	return (item);
2352 }
2353 
2354 static int
2355 zone_alloc_bucket(uma_zone_t zone, int flags)
2356 {
2357 	uma_bucket_t bucket;
2358 	uma_slab_t slab;
2359 	uma_keg_t keg;
2360 	int16_t saved;
2361 	int max, origflags = flags;
2362 
2363 	/*
2364 	 * Try this zone's free list first so we don't allocate extra buckets.
2365 	 */
2366 	if ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) {
2367 		KASSERT(bucket->ub_cnt == 0,
2368 		    ("zone_alloc_bucket: Bucket on free list is not empty."));
2369 		LIST_REMOVE(bucket, ub_link);
2370 	} else {
2371 		int bflags;
2372 
2373 		bflags = (flags & ~M_ZERO);
2374 		if (zone->uz_flags & UMA_ZFLAG_CACHEONLY)
2375 			bflags |= M_NOVM;
2376 
2377 		ZONE_UNLOCK(zone);
2378 		bucket = bucket_alloc(zone->uz_count, bflags);
2379 		ZONE_LOCK(zone);
2380 	}
2381 
2382 	if (bucket == NULL) {
2383 		return (0);
2384 	}
2385 
2386 #ifdef SMP
2387 	/*
2388 	 * This code is here to limit the number of simultaneous bucket fills
2389 	 * for any given zone to the number of per cpu caches in this zone. This
2390 	 * is done so that we don't allocate more memory than we really need.
2391 	 */
2392 	if (zone->uz_fills >= mp_ncpus)
2393 		goto done;
2394 
2395 #endif
2396 	zone->uz_fills++;
2397 
2398 	max = MIN(bucket->ub_entries, zone->uz_count);
2399 	/* Try to keep the buckets totally full */
2400 	saved = bucket->ub_cnt;
2401 	slab = NULL;
2402 	keg = NULL;
2403 	while (bucket->ub_cnt < max &&
2404 	    (slab = zone->uz_slab(zone, keg, flags)) != NULL) {
2405 		keg = slab->us_keg;
2406 		while (slab->us_freecount && bucket->ub_cnt < max) {
2407 			bucket->ub_bucket[bucket->ub_cnt++] =
2408 			    slab_alloc_item(zone, slab);
2409 		}
2410 
2411 		/* Don't block on the next fill */
2412 		flags |= M_NOWAIT;
2413 	}
2414 	if (slab)
2415 		zone_relock(zone, keg);
2416 
2417 	/*
2418 	 * We unlock here because we need to call the zone's init.
2419 	 * It should be safe to unlock because the slab dealt with
2420 	 * above is already on the appropriate list within the keg
2421 	 * and the bucket we filled is not yet on any list, so we
2422 	 * own it.
2423 	 */
2424 	if (zone->uz_init != NULL) {
2425 		int i;
2426 
2427 		ZONE_UNLOCK(zone);
2428 		for (i = saved; i < bucket->ub_cnt; i++)
2429 			if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size,
2430 			    origflags) != 0)
2431 				break;
2432 		/*
2433 		 * If we couldn't initialize the whole bucket, put the
2434 		 * rest back onto the freelist.
2435 		 */
2436 		if (i != bucket->ub_cnt) {
2437 			int j;
2438 
2439 			for (j = i; j < bucket->ub_cnt; j++) {
2440 				zone_free_item(zone, bucket->ub_bucket[j],
2441 				    NULL, SKIP_FINI, 0);
2442 #ifdef INVARIANTS
2443 				bucket->ub_bucket[j] = NULL;
2444 #endif
2445 			}
2446 			bucket->ub_cnt = i;
2447 		}
2448 		ZONE_LOCK(zone);
2449 	}
2450 
2451 	zone->uz_fills--;
2452 	if (bucket->ub_cnt != 0) {
2453 		LIST_INSERT_HEAD(&zone->uz_full_bucket,
2454 		    bucket, ub_link);
2455 		return (1);
2456 	}
2457 #ifdef SMP
2458 done:
2459 #endif
2460 	bucket_free(bucket);
2461 
2462 	return (0);
2463 }
2464 /*
2465  * Allocates an item for an internal zone
2466  *
2467  * Arguments
2468  *	zone   The zone to alloc for.
2469  *	udata  The data to be passed to the constructor.
2470  *	flags  M_WAITOK, M_NOWAIT, M_ZERO.
2471  *
2472  * Returns
2473  *	NULL if there is no memory and M_NOWAIT is set
2474  *	An item if successful
2475  */
2476 
2477 static void *
2478 zone_alloc_item(uma_zone_t zone, void *udata, int flags)
2479 {
2480 	uma_slab_t slab;
2481 	void *item;
2482 
2483 	item = NULL;
2484 
2485 #ifdef UMA_DEBUG_ALLOC
2486 	printf("INTERNAL: Allocating one item from %s(%p)\n", zone->uz_name, zone);
2487 #endif
2488 	ZONE_LOCK(zone);
2489 
2490 	slab = zone->uz_slab(zone, NULL, flags);
2491 	if (slab == NULL) {
2492 		zone->uz_fails++;
2493 		ZONE_UNLOCK(zone);
2494 		return (NULL);
2495 	}
2496 
2497 	item = slab_alloc_item(zone, slab);
2498 
2499 	zone_relock(zone, slab->us_keg);
2500 	zone->uz_allocs++;
2501 	ZONE_UNLOCK(zone);
2502 
2503 	/*
2504 	 * We have to call both the zone's init (not the keg's init)
2505 	 * and the zone's ctor.  This is because the item is going from
2506 	 * a keg slab directly to the user, and the user is expecting it
2507 	 * to be both zone-init'd as well as zone-ctor'd.
2508 	 */
2509 	if (zone->uz_init != NULL) {
2510 		if (zone->uz_init(item, zone->uz_size, flags) != 0) {
2511 			zone_free_item(zone, item, udata, SKIP_FINI,
2512 			    ZFREE_STATFAIL | ZFREE_STATFREE);
2513 			return (NULL);
2514 		}
2515 	}
2516 	if (zone->uz_ctor != NULL) {
2517 		if (zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2518 			zone_free_item(zone, item, udata, SKIP_DTOR,
2519 			    ZFREE_STATFAIL | ZFREE_STATFREE);
2520 			return (NULL);
2521 		}
2522 	}
2523 	if (flags & M_ZERO)
2524 		bzero(item, zone->uz_size);
2525 
2526 	return (item);
2527 }
2528 
2529 /* See uma.h */
2530 void
2531 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
2532 {
2533 	uma_cache_t cache;
2534 	uma_bucket_t bucket;
2535 	int bflags;
2536 	int cpu;
2537 
2538 #ifdef UMA_DEBUG_ALLOC_1
2539 	printf("Freeing item %p to %s(%p)\n", item, zone->uz_name, zone);
2540 #endif
2541 	CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread,
2542 	    zone->uz_name);
2543 
2544         /* uma_zfree(..., NULL) does nothing, to match free(9). */
2545         if (item == NULL)
2546                 return;
2547 
2548 	if (zone->uz_dtor)
2549 		zone->uz_dtor(item, zone->uz_size, udata);
2550 
2551 #ifdef INVARIANTS
2552 	ZONE_LOCK(zone);
2553 	if (zone->uz_flags & UMA_ZONE_MALLOC)
2554 		uma_dbg_free(zone, udata, item);
2555 	else
2556 		uma_dbg_free(zone, NULL, item);
2557 	ZONE_UNLOCK(zone);
2558 #endif
2559 	/*
2560 	 * The race here is acceptable.  If we miss it we'll just have to wait
2561 	 * a little longer for the limits to be reset.
2562 	 */
2563 	if (zone->uz_flags & UMA_ZFLAG_FULL)
2564 		goto zfree_internal;
2565 
2566 	/*
2567 	 * If possible, free to the per-CPU cache.  There are two
2568 	 * requirements for safe access to the per-CPU cache: (1) the thread
2569 	 * accessing the cache must not be preempted or yield during access,
2570 	 * and (2) the thread must not migrate CPUs without switching which
2571 	 * cache it accesses.  We rely on a critical section to prevent
2572 	 * preemption and migration.  We release the critical section in
2573 	 * order to acquire the zone mutex if we are unable to free to the
2574 	 * current cache; when we re-acquire the critical section, we must
2575 	 * detect and handle migration if it has occurred.
2576 	 */
2577 zfree_restart:
2578 	critical_enter();
2579 	cpu = curcpu;
2580 	cache = &zone->uz_cpu[cpu];
2581 
2582 zfree_start:
2583 	bucket = cache->uc_freebucket;
2584 
2585 	if (bucket) {
2586 		/*
2587 		 * Do we have room in our bucket? It is OK for this uz count
2588 		 * check to be slightly out of sync.
2589 		 */
2590 
2591 		if (bucket->ub_cnt < bucket->ub_entries) {
2592 			KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL,
2593 			    ("uma_zfree: Freeing to non free bucket index."));
2594 			bucket->ub_bucket[bucket->ub_cnt] = item;
2595 			bucket->ub_cnt++;
2596 			cache->uc_frees++;
2597 			critical_exit();
2598 			return;
2599 		} else if (cache->uc_allocbucket) {
2600 #ifdef UMA_DEBUG_ALLOC
2601 			printf("uma_zfree: Swapping buckets.\n");
2602 #endif
2603 			/*
2604 			 * We have run out of space in our freebucket.
2605 			 * See if we can switch with our alloc bucket.
2606 			 */
2607 			if (cache->uc_allocbucket->ub_cnt <
2608 			    cache->uc_freebucket->ub_cnt) {
2609 				bucket = cache->uc_freebucket;
2610 				cache->uc_freebucket = cache->uc_allocbucket;
2611 				cache->uc_allocbucket = bucket;
2612 				goto zfree_start;
2613 			}
2614 		}
2615 	}
2616 	/*
2617 	 * We can get here for two reasons:
2618 	 *
2619 	 * 1) The buckets are NULL
2620 	 * 2) The alloc and free buckets are both somewhat full.
2621 	 *
2622 	 * We must go back the zone, which requires acquiring the zone lock,
2623 	 * which in turn means we must release and re-acquire the critical
2624 	 * section.  Since the critical section is released, we may be
2625 	 * preempted or migrate.  As such, make sure not to maintain any
2626 	 * thread-local state specific to the cache from prior to releasing
2627 	 * the critical section.
2628 	 */
2629 	critical_exit();
2630 	ZONE_LOCK(zone);
2631 	critical_enter();
2632 	cpu = curcpu;
2633 	cache = &zone->uz_cpu[cpu];
2634 	if (cache->uc_freebucket != NULL) {
2635 		if (cache->uc_freebucket->ub_cnt <
2636 		    cache->uc_freebucket->ub_entries) {
2637 			ZONE_UNLOCK(zone);
2638 			goto zfree_start;
2639 		}
2640 		if (cache->uc_allocbucket != NULL &&
2641 		    (cache->uc_allocbucket->ub_cnt <
2642 		    cache->uc_freebucket->ub_cnt)) {
2643 			ZONE_UNLOCK(zone);
2644 			goto zfree_start;
2645 		}
2646 	}
2647 
2648 	/* Since we have locked the zone we may as well send back our stats */
2649 	zone->uz_allocs += cache->uc_allocs;
2650 	cache->uc_allocs = 0;
2651 	zone->uz_frees += cache->uc_frees;
2652 	cache->uc_frees = 0;
2653 
2654 	bucket = cache->uc_freebucket;
2655 	cache->uc_freebucket = NULL;
2656 
2657 	/* Can we throw this on the zone full list? */
2658 	if (bucket != NULL) {
2659 #ifdef UMA_DEBUG_ALLOC
2660 		printf("uma_zfree: Putting old bucket on the free list.\n");
2661 #endif
2662 		/* ub_cnt is pointing to the last free item */
2663 		KASSERT(bucket->ub_cnt != 0,
2664 		    ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n"));
2665 		LIST_INSERT_HEAD(&zone->uz_full_bucket,
2666 		    bucket, ub_link);
2667 	}
2668 	if ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) {
2669 		LIST_REMOVE(bucket, ub_link);
2670 		ZONE_UNLOCK(zone);
2671 		cache->uc_freebucket = bucket;
2672 		goto zfree_start;
2673 	}
2674 	/* We are no longer associated with this CPU. */
2675 	critical_exit();
2676 
2677 	/* And the zone.. */
2678 	ZONE_UNLOCK(zone);
2679 
2680 #ifdef UMA_DEBUG_ALLOC
2681 	printf("uma_zfree: Allocating new free bucket.\n");
2682 #endif
2683 	bflags = M_NOWAIT;
2684 
2685 	if (zone->uz_flags & UMA_ZFLAG_CACHEONLY)
2686 		bflags |= M_NOVM;
2687 	bucket = bucket_alloc(zone->uz_count, bflags);
2688 	if (bucket) {
2689 		ZONE_LOCK(zone);
2690 		LIST_INSERT_HEAD(&zone->uz_free_bucket,
2691 		    bucket, ub_link);
2692 		ZONE_UNLOCK(zone);
2693 		goto zfree_restart;
2694 	}
2695 
2696 	/*
2697 	 * If nothing else caught this, we'll just do an internal free.
2698 	 */
2699 zfree_internal:
2700 	zone_free_item(zone, item, udata, SKIP_DTOR, ZFREE_STATFREE);
2701 
2702 	return;
2703 }
2704 
2705 /*
2706  * Frees an item to an INTERNAL zone or allocates a free bucket
2707  *
2708  * Arguments:
2709  *	zone   The zone to free to
2710  *	item   The item we're freeing
2711  *	udata  User supplied data for the dtor
2712  *	skip   Skip dtors and finis
2713  */
2714 static void
2715 zone_free_item(uma_zone_t zone, void *item, void *udata,
2716     enum zfreeskip skip, int flags)
2717 {
2718 	uma_slab_t slab;
2719 	uma_slabrefcnt_t slabref;
2720 	uma_keg_t keg;
2721 	u_int8_t *mem;
2722 	u_int8_t freei;
2723 	int clearfull;
2724 
2725 	if (skip < SKIP_DTOR && zone->uz_dtor)
2726 		zone->uz_dtor(item, zone->uz_size, udata);
2727 
2728 	if (skip < SKIP_FINI && zone->uz_fini)
2729 		zone->uz_fini(item, zone->uz_size);
2730 
2731 	ZONE_LOCK(zone);
2732 
2733 	if (flags & ZFREE_STATFAIL)
2734 		zone->uz_fails++;
2735 	if (flags & ZFREE_STATFREE)
2736 		zone->uz_frees++;
2737 
2738 	if (!(zone->uz_flags & UMA_ZONE_VTOSLAB)) {
2739 		mem = (u_int8_t *)((unsigned long)item & (~UMA_SLAB_MASK));
2740 		keg = zone_first_keg(zone); /* Must only be one. */
2741 		if (zone->uz_flags & UMA_ZONE_HASH) {
2742 			slab = hash_sfind(&keg->uk_hash, mem);
2743 		} else {
2744 			mem += keg->uk_pgoff;
2745 			slab = (uma_slab_t)mem;
2746 		}
2747 	} else {
2748 		/* This prevents redundant lookups via free(). */
2749 		if ((zone->uz_flags & UMA_ZONE_MALLOC) && udata != NULL)
2750 			slab = (uma_slab_t)udata;
2751 		else
2752 			slab = vtoslab((vm_offset_t)item);
2753 		keg = slab->us_keg;
2754 		keg_relock(keg, zone);
2755 	}
2756 	MPASS(keg == slab->us_keg);
2757 
2758 	/* Do we need to remove from any lists? */
2759 	if (slab->us_freecount+1 == keg->uk_ipers) {
2760 		LIST_REMOVE(slab, us_link);
2761 		LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
2762 	} else if (slab->us_freecount == 0) {
2763 		LIST_REMOVE(slab, us_link);
2764 		LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2765 	}
2766 
2767 	/* Slab management stuff */
2768 	freei = ((unsigned long)item - (unsigned long)slab->us_data)
2769 		/ keg->uk_rsize;
2770 
2771 #ifdef INVARIANTS
2772 	if (!skip)
2773 		uma_dbg_free(zone, slab, item);
2774 #endif
2775 
2776 	if (keg->uk_flags & UMA_ZONE_REFCNT) {
2777 		slabref = (uma_slabrefcnt_t)slab;
2778 		slabref->us_freelist[freei].us_item = slab->us_firstfree;
2779 	} else {
2780 		slab->us_freelist[freei].us_item = slab->us_firstfree;
2781 	}
2782 	slab->us_firstfree = freei;
2783 	slab->us_freecount++;
2784 
2785 	/* Zone statistics */
2786 	keg->uk_free++;
2787 
2788 	clearfull = 0;
2789 	if (keg->uk_flags & UMA_ZFLAG_FULL) {
2790 		if (keg->uk_pages < keg->uk_maxpages) {
2791 			keg->uk_flags &= ~UMA_ZFLAG_FULL;
2792 			clearfull = 1;
2793 		}
2794 
2795 		/*
2796 		 * We can handle one more allocation. Since we're clearing ZFLAG_FULL,
2797 		 * wake up all procs blocked on pages. This should be uncommon, so
2798 		 * keeping this simple for now (rather than adding count of blocked
2799 		 * threads etc).
2800 		 */
2801 		wakeup(keg);
2802 	}
2803 	if (clearfull) {
2804 		zone_relock(zone, keg);
2805 		zone->uz_flags &= ~UMA_ZFLAG_FULL;
2806 		wakeup(zone);
2807 		ZONE_UNLOCK(zone);
2808 	} else
2809 		KEG_UNLOCK(keg);
2810 }
2811 
2812 /* See uma.h */
2813 int
2814 uma_zone_set_max(uma_zone_t zone, int nitems)
2815 {
2816 	uma_keg_t keg;
2817 
2818 	ZONE_LOCK(zone);
2819 	keg = zone_first_keg(zone);
2820 	keg->uk_maxpages = (nitems / keg->uk_ipers) * keg->uk_ppera;
2821 	if (keg->uk_maxpages * keg->uk_ipers < nitems)
2822 		keg->uk_maxpages += keg->uk_ppera;
2823 	nitems = keg->uk_maxpages * keg->uk_ipers;
2824 	ZONE_UNLOCK(zone);
2825 
2826 	return (nitems);
2827 }
2828 
2829 /* See uma.h */
2830 int
2831 uma_zone_get_max(uma_zone_t zone)
2832 {
2833 	int nitems;
2834 	uma_keg_t keg;
2835 
2836 	ZONE_LOCK(zone);
2837 	keg = zone_first_keg(zone);
2838 	nitems = keg->uk_maxpages * keg->uk_ipers;
2839 	ZONE_UNLOCK(zone);
2840 
2841 	return (nitems);
2842 }
2843 
2844 /* See uma.h */
2845 int
2846 uma_zone_get_cur(uma_zone_t zone)
2847 {
2848 	int64_t nitems;
2849 	u_int i;
2850 
2851 	ZONE_LOCK(zone);
2852 	nitems = zone->uz_allocs - zone->uz_frees;
2853 	CPU_FOREACH(i) {
2854 		/*
2855 		 * See the comment in sysctl_vm_zone_stats() regarding the
2856 		 * safety of accessing the per-cpu caches. With the zone lock
2857 		 * held, it is safe, but can potentially result in stale data.
2858 		 */
2859 		nitems += zone->uz_cpu[i].uc_allocs -
2860 		    zone->uz_cpu[i].uc_frees;
2861 	}
2862 	ZONE_UNLOCK(zone);
2863 
2864 	return (nitems < 0 ? 0 : nitems);
2865 }
2866 
2867 /* See uma.h */
2868 void
2869 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
2870 {
2871 	uma_keg_t keg;
2872 
2873 	ZONE_LOCK(zone);
2874 	keg = zone_first_keg(zone);
2875 	KASSERT(keg->uk_pages == 0,
2876 	    ("uma_zone_set_init on non-empty keg"));
2877 	keg->uk_init = uminit;
2878 	ZONE_UNLOCK(zone);
2879 }
2880 
2881 /* See uma.h */
2882 void
2883 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
2884 {
2885 	uma_keg_t keg;
2886 
2887 	ZONE_LOCK(zone);
2888 	keg = zone_first_keg(zone);
2889 	KASSERT(keg->uk_pages == 0,
2890 	    ("uma_zone_set_fini on non-empty keg"));
2891 	keg->uk_fini = fini;
2892 	ZONE_UNLOCK(zone);
2893 }
2894 
2895 /* See uma.h */
2896 void
2897 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
2898 {
2899 	ZONE_LOCK(zone);
2900 	KASSERT(zone_first_keg(zone)->uk_pages == 0,
2901 	    ("uma_zone_set_zinit on non-empty keg"));
2902 	zone->uz_init = zinit;
2903 	ZONE_UNLOCK(zone);
2904 }
2905 
2906 /* See uma.h */
2907 void
2908 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
2909 {
2910 	ZONE_LOCK(zone);
2911 	KASSERT(zone_first_keg(zone)->uk_pages == 0,
2912 	    ("uma_zone_set_zfini on non-empty keg"));
2913 	zone->uz_fini = zfini;
2914 	ZONE_UNLOCK(zone);
2915 }
2916 
2917 /* See uma.h */
2918 /* XXX uk_freef is not actually used with the zone locked */
2919 void
2920 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
2921 {
2922 
2923 	ZONE_LOCK(zone);
2924 	zone_first_keg(zone)->uk_freef = freef;
2925 	ZONE_UNLOCK(zone);
2926 }
2927 
2928 /* See uma.h */
2929 /* XXX uk_allocf is not actually used with the zone locked */
2930 void
2931 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
2932 {
2933 	uma_keg_t keg;
2934 
2935 	ZONE_LOCK(zone);
2936 	keg = zone_first_keg(zone);
2937 	keg->uk_flags |= UMA_ZFLAG_PRIVALLOC;
2938 	keg->uk_allocf = allocf;
2939 	ZONE_UNLOCK(zone);
2940 }
2941 
2942 /* See uma.h */
2943 int
2944 uma_zone_set_obj(uma_zone_t zone, struct vm_object *obj, int count)
2945 {
2946 	uma_keg_t keg;
2947 	vm_offset_t kva;
2948 	int pages;
2949 
2950 	keg = zone_first_keg(zone);
2951 	pages = count / keg->uk_ipers;
2952 
2953 	if (pages * keg->uk_ipers < count)
2954 		pages++;
2955 
2956 	kva = kmem_alloc_nofault(kernel_map, pages * UMA_SLAB_SIZE);
2957 
2958 	if (kva == 0)
2959 		return (0);
2960 	if (obj == NULL)
2961 		obj = vm_object_allocate(OBJT_PHYS, pages);
2962 	else {
2963 		VM_OBJECT_LOCK_INIT(obj, "uma object");
2964 		_vm_object_allocate(OBJT_PHYS, pages, obj);
2965 	}
2966 	ZONE_LOCK(zone);
2967 	keg->uk_kva = kva;
2968 	keg->uk_obj = obj;
2969 	keg->uk_maxpages = pages;
2970 	keg->uk_allocf = obj_alloc;
2971 	keg->uk_flags |= UMA_ZONE_NOFREE | UMA_ZFLAG_PRIVALLOC;
2972 	ZONE_UNLOCK(zone);
2973 	return (1);
2974 }
2975 
2976 /* See uma.h */
2977 void
2978 uma_prealloc(uma_zone_t zone, int items)
2979 {
2980 	int slabs;
2981 	uma_slab_t slab;
2982 	uma_keg_t keg;
2983 
2984 	keg = zone_first_keg(zone);
2985 	ZONE_LOCK(zone);
2986 	slabs = items / keg->uk_ipers;
2987 	if (slabs * keg->uk_ipers < items)
2988 		slabs++;
2989 	while (slabs > 0) {
2990 		slab = keg_alloc_slab(keg, zone, M_WAITOK);
2991 		if (slab == NULL)
2992 			break;
2993 		MPASS(slab->us_keg == keg);
2994 		LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
2995 		slabs--;
2996 	}
2997 	ZONE_UNLOCK(zone);
2998 }
2999 
3000 /* See uma.h */
3001 u_int32_t *
3002 uma_find_refcnt(uma_zone_t zone, void *item)
3003 {
3004 	uma_slabrefcnt_t slabref;
3005 	uma_keg_t keg;
3006 	u_int32_t *refcnt;
3007 	int idx;
3008 
3009 	slabref = (uma_slabrefcnt_t)vtoslab((vm_offset_t)item &
3010 	    (~UMA_SLAB_MASK));
3011 	keg = slabref->us_keg;
3012 	KASSERT(slabref != NULL && slabref->us_keg->uk_flags & UMA_ZONE_REFCNT,
3013 	    ("uma_find_refcnt(): zone possibly not UMA_ZONE_REFCNT"));
3014 	idx = ((unsigned long)item - (unsigned long)slabref->us_data)
3015 	    / keg->uk_rsize;
3016 	refcnt = &slabref->us_freelist[idx].us_refcnt;
3017 	return refcnt;
3018 }
3019 
3020 /* See uma.h */
3021 void
3022 uma_reclaim(void)
3023 {
3024 #ifdef UMA_DEBUG
3025 	printf("UMA: vm asked us to release pages!\n");
3026 #endif
3027 	bucket_enable();
3028 	zone_foreach(zone_drain);
3029 	/*
3030 	 * Some slabs may have been freed but this zone will be visited early
3031 	 * we visit again so that we can free pages that are empty once other
3032 	 * zones are drained.  We have to do the same for buckets.
3033 	 */
3034 	zone_drain(slabzone);
3035 	zone_drain(slabrefzone);
3036 	bucket_zone_drain();
3037 }
3038 
3039 /* See uma.h */
3040 int
3041 uma_zone_exhausted(uma_zone_t zone)
3042 {
3043 	int full;
3044 
3045 	ZONE_LOCK(zone);
3046 	full = (zone->uz_flags & UMA_ZFLAG_FULL);
3047 	ZONE_UNLOCK(zone);
3048 	return (full);
3049 }
3050 
3051 int
3052 uma_zone_exhausted_nolock(uma_zone_t zone)
3053 {
3054 	return (zone->uz_flags & UMA_ZFLAG_FULL);
3055 }
3056 
3057 void *
3058 uma_large_malloc(int size, int wait)
3059 {
3060 	void *mem;
3061 	uma_slab_t slab;
3062 	u_int8_t flags;
3063 
3064 	slab = zone_alloc_item(slabzone, NULL, wait);
3065 	if (slab == NULL)
3066 		return (NULL);
3067 	mem = page_alloc(NULL, size, &flags, wait);
3068 	if (mem) {
3069 		vsetslab((vm_offset_t)mem, slab);
3070 		slab->us_data = mem;
3071 		slab->us_flags = flags | UMA_SLAB_MALLOC;
3072 		slab->us_size = size;
3073 	} else {
3074 		zone_free_item(slabzone, slab, NULL, SKIP_NONE,
3075 		    ZFREE_STATFAIL | ZFREE_STATFREE);
3076 	}
3077 
3078 	return (mem);
3079 }
3080 
3081 void
3082 uma_large_free(uma_slab_t slab)
3083 {
3084 	vsetobj((vm_offset_t)slab->us_data, kmem_object);
3085 	page_free(slab->us_data, slab->us_size, slab->us_flags);
3086 	zone_free_item(slabzone, slab, NULL, SKIP_NONE, ZFREE_STATFREE);
3087 }
3088 
3089 void
3090 uma_print_stats(void)
3091 {
3092 	zone_foreach(uma_print_zone);
3093 }
3094 
3095 static void
3096 slab_print(uma_slab_t slab)
3097 {
3098 	printf("slab: keg %p, data %p, freecount %d, firstfree %d\n",
3099 		slab->us_keg, slab->us_data, slab->us_freecount,
3100 		slab->us_firstfree);
3101 }
3102 
3103 static void
3104 cache_print(uma_cache_t cache)
3105 {
3106 	printf("alloc: %p(%d), free: %p(%d)\n",
3107 		cache->uc_allocbucket,
3108 		cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0,
3109 		cache->uc_freebucket,
3110 		cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0);
3111 }
3112 
3113 static void
3114 uma_print_keg(uma_keg_t keg)
3115 {
3116 	uma_slab_t slab;
3117 
3118 	printf("keg: %s(%p) size %d(%d) flags %d ipers %d ppera %d "
3119 	    "out %d free %d limit %d\n",
3120 	    keg->uk_name, keg, keg->uk_size, keg->uk_rsize, keg->uk_flags,
3121 	    keg->uk_ipers, keg->uk_ppera,
3122 	    (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free,
3123 	    (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers);
3124 	printf("Part slabs:\n");
3125 	LIST_FOREACH(slab, &keg->uk_part_slab, us_link)
3126 		slab_print(slab);
3127 	printf("Free slabs:\n");
3128 	LIST_FOREACH(slab, &keg->uk_free_slab, us_link)
3129 		slab_print(slab);
3130 	printf("Full slabs:\n");
3131 	LIST_FOREACH(slab, &keg->uk_full_slab, us_link)
3132 		slab_print(slab);
3133 }
3134 
3135 void
3136 uma_print_zone(uma_zone_t zone)
3137 {
3138 	uma_cache_t cache;
3139 	uma_klink_t kl;
3140 	int i;
3141 
3142 	printf("zone: %s(%p) size %d flags %d\n",
3143 	    zone->uz_name, zone, zone->uz_size, zone->uz_flags);
3144 	LIST_FOREACH(kl, &zone->uz_kegs, kl_link)
3145 		uma_print_keg(kl->kl_keg);
3146 	CPU_FOREACH(i) {
3147 		cache = &zone->uz_cpu[i];
3148 		printf("CPU %d Cache:\n", i);
3149 		cache_print(cache);
3150 	}
3151 }
3152 
3153 #ifdef DDB
3154 /*
3155  * Generate statistics across both the zone and its per-cpu cache's.  Return
3156  * desired statistics if the pointer is non-NULL for that statistic.
3157  *
3158  * Note: does not update the zone statistics, as it can't safely clear the
3159  * per-CPU cache statistic.
3160  *
3161  * XXXRW: Following the uc_allocbucket and uc_freebucket pointers here isn't
3162  * safe from off-CPU; we should modify the caches to track this information
3163  * directly so that we don't have to.
3164  */
3165 static void
3166 uma_zone_sumstat(uma_zone_t z, int *cachefreep, u_int64_t *allocsp,
3167     u_int64_t *freesp, u_int64_t *sleepsp)
3168 {
3169 	uma_cache_t cache;
3170 	u_int64_t allocs, frees, sleeps;
3171 	int cachefree, cpu;
3172 
3173 	allocs = frees = sleeps = 0;
3174 	cachefree = 0;
3175 	CPU_FOREACH(cpu) {
3176 		cache = &z->uz_cpu[cpu];
3177 		if (cache->uc_allocbucket != NULL)
3178 			cachefree += cache->uc_allocbucket->ub_cnt;
3179 		if (cache->uc_freebucket != NULL)
3180 			cachefree += cache->uc_freebucket->ub_cnt;
3181 		allocs += cache->uc_allocs;
3182 		frees += cache->uc_frees;
3183 	}
3184 	allocs += z->uz_allocs;
3185 	frees += z->uz_frees;
3186 	sleeps += z->uz_sleeps;
3187 	if (cachefreep != NULL)
3188 		*cachefreep = cachefree;
3189 	if (allocsp != NULL)
3190 		*allocsp = allocs;
3191 	if (freesp != NULL)
3192 		*freesp = frees;
3193 	if (sleepsp != NULL)
3194 		*sleepsp = sleeps;
3195 }
3196 #endif /* DDB */
3197 
3198 static int
3199 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
3200 {
3201 	uma_keg_t kz;
3202 	uma_zone_t z;
3203 	int count;
3204 
3205 	count = 0;
3206 	mtx_lock(&uma_mtx);
3207 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3208 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
3209 			count++;
3210 	}
3211 	mtx_unlock(&uma_mtx);
3212 	return (sysctl_handle_int(oidp, &count, 0, req));
3213 }
3214 
3215 static int
3216 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
3217 {
3218 	struct uma_stream_header ush;
3219 	struct uma_type_header uth;
3220 	struct uma_percpu_stat ups;
3221 	uma_bucket_t bucket;
3222 	struct sbuf sbuf;
3223 	uma_cache_t cache;
3224 	uma_klink_t kl;
3225 	uma_keg_t kz;
3226 	uma_zone_t z;
3227 	uma_keg_t k;
3228 	int count, error, i;
3229 
3230 	error = sysctl_wire_old_buffer(req, 0);
3231 	if (error != 0)
3232 		return (error);
3233 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
3234 
3235 	count = 0;
3236 	mtx_lock(&uma_mtx);
3237 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3238 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
3239 			count++;
3240 	}
3241 
3242 	/*
3243 	 * Insert stream header.
3244 	 */
3245 	bzero(&ush, sizeof(ush));
3246 	ush.ush_version = UMA_STREAM_VERSION;
3247 	ush.ush_maxcpus = (mp_maxid + 1);
3248 	ush.ush_count = count;
3249 	(void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
3250 
3251 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3252 		LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3253 			bzero(&uth, sizeof(uth));
3254 			ZONE_LOCK(z);
3255 			strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
3256 			uth.uth_align = kz->uk_align;
3257 			uth.uth_size = kz->uk_size;
3258 			uth.uth_rsize = kz->uk_rsize;
3259 			LIST_FOREACH(kl, &z->uz_kegs, kl_link) {
3260 				k = kl->kl_keg;
3261 				uth.uth_maxpages += k->uk_maxpages;
3262 				uth.uth_pages += k->uk_pages;
3263 				uth.uth_keg_free += k->uk_free;
3264 				uth.uth_limit = (k->uk_maxpages / k->uk_ppera)
3265 				    * k->uk_ipers;
3266 			}
3267 
3268 			/*
3269 			 * A zone is secondary is it is not the first entry
3270 			 * on the keg's zone list.
3271 			 */
3272 			if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
3273 			    (LIST_FIRST(&kz->uk_zones) != z))
3274 				uth.uth_zone_flags = UTH_ZONE_SECONDARY;
3275 
3276 			LIST_FOREACH(bucket, &z->uz_full_bucket, ub_link)
3277 				uth.uth_zone_free += bucket->ub_cnt;
3278 			uth.uth_allocs = z->uz_allocs;
3279 			uth.uth_frees = z->uz_frees;
3280 			uth.uth_fails = z->uz_fails;
3281 			uth.uth_sleeps = z->uz_sleeps;
3282 			(void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
3283 			/*
3284 			 * While it is not normally safe to access the cache
3285 			 * bucket pointers while not on the CPU that owns the
3286 			 * cache, we only allow the pointers to be exchanged
3287 			 * without the zone lock held, not invalidated, so
3288 			 * accept the possible race associated with bucket
3289 			 * exchange during monitoring.
3290 			 */
3291 			for (i = 0; i < (mp_maxid + 1); i++) {
3292 				bzero(&ups, sizeof(ups));
3293 				if (kz->uk_flags & UMA_ZFLAG_INTERNAL)
3294 					goto skip;
3295 				if (CPU_ABSENT(i))
3296 					goto skip;
3297 				cache = &z->uz_cpu[i];
3298 				if (cache->uc_allocbucket != NULL)
3299 					ups.ups_cache_free +=
3300 					    cache->uc_allocbucket->ub_cnt;
3301 				if (cache->uc_freebucket != NULL)
3302 					ups.ups_cache_free +=
3303 					    cache->uc_freebucket->ub_cnt;
3304 				ups.ups_allocs = cache->uc_allocs;
3305 				ups.ups_frees = cache->uc_frees;
3306 skip:
3307 				(void)sbuf_bcat(&sbuf, &ups, sizeof(ups));
3308 			}
3309 			ZONE_UNLOCK(z);
3310 		}
3311 	}
3312 	mtx_unlock(&uma_mtx);
3313 	error = sbuf_finish(&sbuf);
3314 	sbuf_delete(&sbuf);
3315 	return (error);
3316 }
3317 
3318 #ifdef DDB
3319 DB_SHOW_COMMAND(uma, db_show_uma)
3320 {
3321 	u_int64_t allocs, frees, sleeps;
3322 	uma_bucket_t bucket;
3323 	uma_keg_t kz;
3324 	uma_zone_t z;
3325 	int cachefree;
3326 
3327 	db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
3328 	    "Requests", "Sleeps");
3329 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3330 		LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3331 			if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
3332 				allocs = z->uz_allocs;
3333 				frees = z->uz_frees;
3334 				sleeps = z->uz_sleeps;
3335 				cachefree = 0;
3336 			} else
3337 				uma_zone_sumstat(z, &cachefree, &allocs,
3338 				    &frees, &sleeps);
3339 			if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
3340 			    (LIST_FIRST(&kz->uk_zones) != z)))
3341 				cachefree += kz->uk_free;
3342 			LIST_FOREACH(bucket, &z->uz_full_bucket, ub_link)
3343 				cachefree += bucket->ub_cnt;
3344 			db_printf("%18s %8ju %8jd %8d %12ju %8ju\n", z->uz_name,
3345 			    (uintmax_t)kz->uk_size,
3346 			    (intmax_t)(allocs - frees), cachefree,
3347 			    (uintmax_t)allocs, sleeps);
3348 		}
3349 	}
3350 }
3351 #endif
3352