xref: /freebsd/sys/vm/uma_core.c (revision 830940567b49bb0c08dfaed40418999e76616909)
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 <machine/vmparam.h>
90 
91 #include <ddb/ddb.h>
92 
93 /*
94  * This is the zone and keg from which all zones are spawned.  The idea is that
95  * even the zone & keg heads are allocated from the allocator, so we use the
96  * bss section to bootstrap us.
97  */
98 static struct uma_keg masterkeg;
99 static struct uma_zone masterzone_k;
100 static struct uma_zone masterzone_z;
101 static uma_zone_t kegs = &masterzone_k;
102 static uma_zone_t zones = &masterzone_z;
103 
104 /* This is the zone from which all of uma_slab_t's are allocated. */
105 static uma_zone_t slabzone;
106 static uma_zone_t slabrefzone;	/* With refcounters (for UMA_ZONE_REFCNT) */
107 
108 /*
109  * The initial hash tables come out of this zone so they can be allocated
110  * prior to malloc coming up.
111  */
112 static uma_zone_t hashzone;
113 
114 /* The boot-time adjusted value for cache line alignment. */
115 static int uma_align_cache = 64 - 1;
116 
117 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets");
118 
119 /*
120  * Are we allowed to allocate buckets?
121  */
122 static int bucketdisable = 1;
123 
124 /* Linked list of all kegs in the system */
125 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(&uma_kegs);
126 
127 /* This mutex protects the keg list */
128 static struct mtx uma_mtx;
129 
130 /* Linked list of boot time pages */
131 static LIST_HEAD(,uma_slab) uma_boot_pages =
132     LIST_HEAD_INITIALIZER(&uma_boot_pages);
133 
134 /* This mutex protects the boot time pages list */
135 static struct mtx uma_boot_pages_mtx;
136 
137 /* Is the VM done starting up? */
138 static int booted = 0;
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 	for (cpu = 0; cpu <= mp_maxid; cpu++) {
624 		if (CPU_ABSENT(cpu))
625 			continue;
626 		cache = &zone->uz_cpu[cpu];
627 		bucket_drain(zone, cache->uc_allocbucket);
628 		bucket_drain(zone, cache->uc_freebucket);
629 		if (cache->uc_allocbucket != NULL)
630 			bucket_free(cache->uc_allocbucket);
631 		if (cache->uc_freebucket != NULL)
632 			bucket_free(cache->uc_freebucket);
633 		cache->uc_allocbucket = cache->uc_freebucket = NULL;
634 	}
635 	ZONE_LOCK(zone);
636 	bucket_cache_drain(zone);
637 	ZONE_UNLOCK(zone);
638 }
639 
640 /*
641  * Drain the cached buckets from a zone.  Expects a locked zone on entry.
642  */
643 static void
644 bucket_cache_drain(uma_zone_t zone)
645 {
646 	uma_bucket_t bucket;
647 
648 	/*
649 	 * Drain the bucket queues and free the buckets, we just keep two per
650 	 * cpu (alloc/free).
651 	 */
652 	while ((bucket = LIST_FIRST(&zone->uz_full_bucket)) != NULL) {
653 		LIST_REMOVE(bucket, ub_link);
654 		ZONE_UNLOCK(zone);
655 		bucket_drain(zone, bucket);
656 		bucket_free(bucket);
657 		ZONE_LOCK(zone);
658 	}
659 
660 	/* Now we do the free queue.. */
661 	while ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) {
662 		LIST_REMOVE(bucket, ub_link);
663 		bucket_free(bucket);
664 	}
665 }
666 
667 /*
668  * Frees pages from a keg back to the system.  This is done on demand from
669  * the pageout daemon.
670  *
671  * Returns nothing.
672  */
673 static void
674 keg_drain(uma_keg_t keg)
675 {
676 	struct slabhead freeslabs = { 0 };
677 	uma_slab_t slab;
678 	uma_slab_t n;
679 	u_int8_t flags;
680 	u_int8_t *mem;
681 	int i;
682 
683 	/*
684 	 * We don't want to take pages from statically allocated kegs at this
685 	 * time
686 	 */
687 	if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL)
688 		return;
689 
690 #ifdef UMA_DEBUG
691 	printf("%s free items: %u\n", keg->uk_name, keg->uk_free);
692 #endif
693 	KEG_LOCK(keg);
694 	if (keg->uk_free == 0)
695 		goto finished;
696 
697 	slab = LIST_FIRST(&keg->uk_free_slab);
698 	while (slab) {
699 		n = LIST_NEXT(slab, us_link);
700 
701 		/* We have no where to free these to */
702 		if (slab->us_flags & UMA_SLAB_BOOT) {
703 			slab = n;
704 			continue;
705 		}
706 
707 		LIST_REMOVE(slab, us_link);
708 		keg->uk_pages -= keg->uk_ppera;
709 		keg->uk_free -= keg->uk_ipers;
710 
711 		if (keg->uk_flags & UMA_ZONE_HASH)
712 			UMA_HASH_REMOVE(&keg->uk_hash, slab, slab->us_data);
713 
714 		SLIST_INSERT_HEAD(&freeslabs, slab, us_hlink);
715 
716 		slab = n;
717 	}
718 finished:
719 	KEG_UNLOCK(keg);
720 
721 	while ((slab = SLIST_FIRST(&freeslabs)) != NULL) {
722 		SLIST_REMOVE(&freeslabs, slab, uma_slab, us_hlink);
723 		if (keg->uk_fini)
724 			for (i = 0; i < keg->uk_ipers; i++)
725 				keg->uk_fini(
726 				    slab->us_data + (keg->uk_rsize * i),
727 				    keg->uk_size);
728 		flags = slab->us_flags;
729 		mem = slab->us_data;
730 
731 		if (keg->uk_flags & UMA_ZONE_VTOSLAB) {
732 			vm_object_t obj;
733 
734 			if (flags & UMA_SLAB_KMEM)
735 				obj = kmem_object;
736 			else if (flags & UMA_SLAB_KERNEL)
737 				obj = kernel_object;
738 			else
739 				obj = NULL;
740 			for (i = 0; i < keg->uk_ppera; i++)
741 				vsetobj((vm_offset_t)mem + (i * PAGE_SIZE),
742 				    obj);
743 		}
744 		if (keg->uk_flags & UMA_ZONE_OFFPAGE)
745 			zone_free_item(keg->uk_slabzone, slab, NULL,
746 			    SKIP_NONE, ZFREE_STATFREE);
747 #ifdef UMA_DEBUG
748 		printf("%s: Returning %d bytes.\n",
749 		    keg->uk_name, UMA_SLAB_SIZE * keg->uk_ppera);
750 #endif
751 		keg->uk_freef(mem, UMA_SLAB_SIZE * keg->uk_ppera, flags);
752 	}
753 }
754 
755 static void
756 zone_drain_wait(uma_zone_t zone, int waitok)
757 {
758 
759 	/*
760 	 * Set draining to interlock with zone_dtor() so we can release our
761 	 * locks as we go.  Only dtor() should do a WAITOK call since it
762 	 * is the only call that knows the structure will still be available
763 	 * when it wakes up.
764 	 */
765 	ZONE_LOCK(zone);
766 	while (zone->uz_flags & UMA_ZFLAG_DRAINING) {
767 		if (waitok == M_NOWAIT)
768 			goto out;
769 		mtx_unlock(&uma_mtx);
770 		msleep(zone, zone->uz_lock, PVM, "zonedrain", 1);
771 		mtx_lock(&uma_mtx);
772 	}
773 	zone->uz_flags |= UMA_ZFLAG_DRAINING;
774 	bucket_cache_drain(zone);
775 	ZONE_UNLOCK(zone);
776 	/*
777 	 * The DRAINING flag protects us from being freed while
778 	 * we're running.  Normally the uma_mtx would protect us but we
779 	 * must be able to release and acquire the right lock for each keg.
780 	 */
781 	zone_foreach_keg(zone, &keg_drain);
782 	ZONE_LOCK(zone);
783 	zone->uz_flags &= ~UMA_ZFLAG_DRAINING;
784 	wakeup(zone);
785 out:
786 	ZONE_UNLOCK(zone);
787 }
788 
789 void
790 zone_drain(uma_zone_t zone)
791 {
792 
793 	zone_drain_wait(zone, M_NOWAIT);
794 }
795 
796 /*
797  * Allocate a new slab for a keg.  This does not insert the slab onto a list.
798  *
799  * Arguments:
800  *	wait  Shall we wait?
801  *
802  * Returns:
803  *	The slab that was allocated or NULL if there is no memory and the
804  *	caller specified M_NOWAIT.
805  */
806 static uma_slab_t
807 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int wait)
808 {
809 	uma_slabrefcnt_t slabref;
810 	uma_alloc allocf;
811 	uma_slab_t slab;
812 	u_int8_t *mem;
813 	u_int8_t flags;
814 	int i;
815 
816 	mtx_assert(&keg->uk_lock, MA_OWNED);
817 	slab = NULL;
818 
819 #ifdef UMA_DEBUG
820 	printf("slab_zalloc:  Allocating a new slab for %s\n", keg->uk_name);
821 #endif
822 	allocf = keg->uk_allocf;
823 	KEG_UNLOCK(keg);
824 
825 	if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
826 		slab = zone_alloc_item(keg->uk_slabzone, NULL, wait);
827 		if (slab == NULL) {
828 			KEG_LOCK(keg);
829 			return NULL;
830 		}
831 	}
832 
833 	/*
834 	 * This reproduces the old vm_zone behavior of zero filling pages the
835 	 * first time they are added to a zone.
836 	 *
837 	 * Malloced items are zeroed in uma_zalloc.
838 	 */
839 
840 	if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
841 		wait |= M_ZERO;
842 	else
843 		wait &= ~M_ZERO;
844 
845 	/* zone is passed for legacy reasons. */
846 	mem = allocf(zone, keg->uk_ppera * UMA_SLAB_SIZE, &flags, wait);
847 	if (mem == NULL) {
848 		if (keg->uk_flags & UMA_ZONE_OFFPAGE)
849 			zone_free_item(keg->uk_slabzone, slab, NULL,
850 			    SKIP_NONE, ZFREE_STATFREE);
851 		KEG_LOCK(keg);
852 		return (NULL);
853 	}
854 
855 	/* Point the slab into the allocated memory */
856 	if (!(keg->uk_flags & UMA_ZONE_OFFPAGE))
857 		slab = (uma_slab_t )(mem + keg->uk_pgoff);
858 
859 	if (keg->uk_flags & UMA_ZONE_VTOSLAB)
860 		for (i = 0; i < keg->uk_ppera; i++)
861 			vsetslab((vm_offset_t)mem + (i * PAGE_SIZE), slab);
862 
863 	slab->us_keg = keg;
864 	slab->us_data = mem;
865 	slab->us_freecount = keg->uk_ipers;
866 	slab->us_firstfree = 0;
867 	slab->us_flags = flags;
868 
869 	if (keg->uk_flags & UMA_ZONE_REFCNT) {
870 		slabref = (uma_slabrefcnt_t)slab;
871 		for (i = 0; i < keg->uk_ipers; i++) {
872 			slabref->us_freelist[i].us_refcnt = 0;
873 			slabref->us_freelist[i].us_item = i+1;
874 		}
875 	} else {
876 		for (i = 0; i < keg->uk_ipers; i++)
877 			slab->us_freelist[i].us_item = i+1;
878 	}
879 
880 	if (keg->uk_init != NULL) {
881 		for (i = 0; i < keg->uk_ipers; i++)
882 			if (keg->uk_init(slab->us_data + (keg->uk_rsize * i),
883 			    keg->uk_size, wait) != 0)
884 				break;
885 		if (i != keg->uk_ipers) {
886 			if (keg->uk_fini != NULL) {
887 				for (i--; i > -1; i--)
888 					keg->uk_fini(slab->us_data +
889 					    (keg->uk_rsize * i),
890 					    keg->uk_size);
891 			}
892 			if (keg->uk_flags & UMA_ZONE_VTOSLAB) {
893 				vm_object_t obj;
894 
895 				if (flags & UMA_SLAB_KMEM)
896 					obj = kmem_object;
897 				else if (flags & UMA_SLAB_KERNEL)
898 					obj = kernel_object;
899 				else
900 					obj = NULL;
901 				for (i = 0; i < keg->uk_ppera; i++)
902 					vsetobj((vm_offset_t)mem +
903 					    (i * PAGE_SIZE), obj);
904 			}
905 			if (keg->uk_flags & UMA_ZONE_OFFPAGE)
906 				zone_free_item(keg->uk_slabzone, slab,
907 				    NULL, SKIP_NONE, ZFREE_STATFREE);
908 			keg->uk_freef(mem, UMA_SLAB_SIZE * keg->uk_ppera,
909 			    flags);
910 			KEG_LOCK(keg);
911 			return (NULL);
912 		}
913 	}
914 	KEG_LOCK(keg);
915 
916 	if (keg->uk_flags & UMA_ZONE_HASH)
917 		UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
918 
919 	keg->uk_pages += keg->uk_ppera;
920 	keg->uk_free += keg->uk_ipers;
921 
922 	return (slab);
923 }
924 
925 /*
926  * This function is intended to be used early on in place of page_alloc() so
927  * that we may use the boot time page cache to satisfy allocations before
928  * the VM is ready.
929  */
930 static void *
931 startup_alloc(uma_zone_t zone, int bytes, u_int8_t *pflag, int wait)
932 {
933 	uma_keg_t keg;
934 	uma_slab_t tmps;
935 
936 	keg = zone_first_keg(zone);
937 
938 	/*
939 	 * Check our small startup cache to see if it has pages remaining.
940 	 */
941 	mtx_lock(&uma_boot_pages_mtx);
942 	if ((tmps = LIST_FIRST(&uma_boot_pages)) != NULL) {
943 		LIST_REMOVE(tmps, us_link);
944 		mtx_unlock(&uma_boot_pages_mtx);
945 		*pflag = tmps->us_flags;
946 		return (tmps->us_data);
947 	}
948 	mtx_unlock(&uma_boot_pages_mtx);
949 	if (booted == 0)
950 		panic("UMA: Increase vm.boot_pages");
951 	/*
952 	 * Now that we've booted reset these users to their real allocator.
953 	 */
954 #ifdef UMA_MD_SMALL_ALLOC
955 	keg->uk_allocf = uma_small_alloc;
956 #else
957 	keg->uk_allocf = page_alloc;
958 #endif
959 	return keg->uk_allocf(zone, bytes, pflag, wait);
960 }
961 
962 /*
963  * Allocates a number of pages from the system
964  *
965  * Arguments:
966  *	bytes  The number of bytes requested
967  *	wait  Shall we wait?
968  *
969  * Returns:
970  *	A pointer to the alloced memory or possibly
971  *	NULL if M_NOWAIT is set.
972  */
973 static void *
974 page_alloc(uma_zone_t zone, int bytes, u_int8_t *pflag, int wait)
975 {
976 	void *p;	/* Returned page */
977 
978 	*pflag = UMA_SLAB_KMEM;
979 	p = (void *) kmem_malloc(kmem_map, bytes, wait);
980 
981 	return (p);
982 }
983 
984 /*
985  * Allocates a number of pages from within an object
986  *
987  * Arguments:
988  *	bytes  The number of bytes requested
989  *	wait   Shall we wait?
990  *
991  * Returns:
992  *	A pointer to the alloced memory or possibly
993  *	NULL if M_NOWAIT is set.
994  */
995 static void *
996 obj_alloc(uma_zone_t zone, int bytes, u_int8_t *flags, int wait)
997 {
998 	vm_object_t object;
999 	vm_offset_t retkva, zkva;
1000 	vm_page_t p;
1001 	int pages, startpages;
1002 	uma_keg_t keg;
1003 
1004 	keg = zone_first_keg(zone);
1005 	object = keg->uk_obj;
1006 	retkva = 0;
1007 
1008 	/*
1009 	 * This looks a little weird since we're getting one page at a time.
1010 	 */
1011 	VM_OBJECT_LOCK(object);
1012 	p = TAILQ_LAST(&object->memq, pglist);
1013 	pages = p != NULL ? p->pindex + 1 : 0;
1014 	startpages = pages;
1015 	zkva = keg->uk_kva + pages * PAGE_SIZE;
1016 	for (; bytes > 0; bytes -= PAGE_SIZE) {
1017 		p = vm_page_alloc(object, pages,
1018 		    VM_ALLOC_INTERRUPT | VM_ALLOC_WIRED);
1019 		if (p == NULL) {
1020 			if (pages != startpages)
1021 				pmap_qremove(retkva, pages - startpages);
1022 			while (pages != startpages) {
1023 				pages--;
1024 				p = TAILQ_LAST(&object->memq, pglist);
1025 				vm_page_lock_queues();
1026 				vm_page_unwire(p, 0);
1027 				vm_page_free(p);
1028 				vm_page_unlock_queues();
1029 			}
1030 			retkva = 0;
1031 			goto done;
1032 		}
1033 		pmap_qenter(zkva, &p, 1);
1034 		if (retkva == 0)
1035 			retkva = zkva;
1036 		zkva += PAGE_SIZE;
1037 		pages += 1;
1038 	}
1039 done:
1040 	VM_OBJECT_UNLOCK(object);
1041 	*flags = UMA_SLAB_PRIV;
1042 
1043 	return ((void *)retkva);
1044 }
1045 
1046 /*
1047  * Frees a number of pages to the system
1048  *
1049  * Arguments:
1050  *	mem   A pointer to the memory to be freed
1051  *	size  The size of the memory being freed
1052  *	flags The original p->us_flags field
1053  *
1054  * Returns:
1055  *	Nothing
1056  */
1057 static void
1058 page_free(void *mem, int size, u_int8_t flags)
1059 {
1060 	vm_map_t map;
1061 
1062 	if (flags & UMA_SLAB_KMEM)
1063 		map = kmem_map;
1064 	else if (flags & UMA_SLAB_KERNEL)
1065 		map = kernel_map;
1066 	else
1067 		panic("UMA: page_free used with invalid flags %d", flags);
1068 
1069 	kmem_free(map, (vm_offset_t)mem, size);
1070 }
1071 
1072 /*
1073  * Zero fill initializer
1074  *
1075  * Arguments/Returns follow uma_init specifications
1076  */
1077 static int
1078 zero_init(void *mem, int size, int flags)
1079 {
1080 	bzero(mem, size);
1081 	return (0);
1082 }
1083 
1084 /*
1085  * Finish creating a small uma keg.  This calculates ipers, and the keg size.
1086  *
1087  * Arguments
1088  *	keg  The zone we should initialize
1089  *
1090  * Returns
1091  *	Nothing
1092  */
1093 static void
1094 keg_small_init(uma_keg_t keg)
1095 {
1096 	u_int rsize;
1097 	u_int memused;
1098 	u_int wastedspace;
1099 	u_int shsize;
1100 
1101 	KASSERT(keg != NULL, ("Keg is null in keg_small_init"));
1102 	rsize = keg->uk_size;
1103 
1104 	if (rsize < UMA_SMALLEST_UNIT)
1105 		rsize = UMA_SMALLEST_UNIT;
1106 	if (rsize & keg->uk_align)
1107 		rsize = (rsize & ~keg->uk_align) + (keg->uk_align + 1);
1108 
1109 	keg->uk_rsize = rsize;
1110 	keg->uk_ppera = 1;
1111 
1112 	if (keg->uk_flags & UMA_ZONE_REFCNT) {
1113 		rsize += UMA_FRITMREF_SZ;	/* linkage & refcnt */
1114 		shsize = sizeof(struct uma_slab_refcnt);
1115 	} else {
1116 		rsize += UMA_FRITM_SZ;	/* Account for linkage */
1117 		shsize = sizeof(struct uma_slab);
1118 	}
1119 
1120 	keg->uk_ipers = (UMA_SLAB_SIZE - shsize) / rsize;
1121 	KASSERT(keg->uk_ipers != 0, ("keg_small_init: ipers is 0"));
1122 	memused = keg->uk_ipers * rsize + shsize;
1123 	wastedspace = UMA_SLAB_SIZE - memused;
1124 
1125 	/*
1126 	 * We can't do OFFPAGE if we're internal or if we've been
1127 	 * asked to not go to the VM for buckets.  If we do this we
1128 	 * may end up going to the VM (kmem_map) for slabs which we
1129 	 * do not want to do if we're UMA_ZFLAG_CACHEONLY as a
1130 	 * result of UMA_ZONE_VM, which clearly forbids it.
1131 	 */
1132 	if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) ||
1133 	    (keg->uk_flags & UMA_ZFLAG_CACHEONLY))
1134 		return;
1135 
1136 	if ((wastedspace >= UMA_MAX_WASTE) &&
1137 	    (keg->uk_ipers < (UMA_SLAB_SIZE / keg->uk_rsize))) {
1138 		keg->uk_ipers = UMA_SLAB_SIZE / keg->uk_rsize;
1139 		KASSERT(keg->uk_ipers <= 255,
1140 		    ("keg_small_init: keg->uk_ipers too high!"));
1141 #ifdef UMA_DEBUG
1142 		printf("UMA decided we need offpage slab headers for "
1143 		    "keg: %s, calculated wastedspace = %d, "
1144 		    "maximum wasted space allowed = %d, "
1145 		    "calculated ipers = %d, "
1146 		    "new wasted space = %d\n", keg->uk_name, wastedspace,
1147 		    UMA_MAX_WASTE, keg->uk_ipers,
1148 		    UMA_SLAB_SIZE - keg->uk_ipers * keg->uk_rsize);
1149 #endif
1150 		keg->uk_flags |= UMA_ZONE_OFFPAGE;
1151 		if ((keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1152 			keg->uk_flags |= UMA_ZONE_HASH;
1153 	}
1154 }
1155 
1156 /*
1157  * Finish creating a large (> UMA_SLAB_SIZE) uma kegs.  Just give in and do
1158  * OFFPAGE for now.  When I can allow for more dynamic slab sizes this will be
1159  * more complicated.
1160  *
1161  * Arguments
1162  *	keg  The keg we should initialize
1163  *
1164  * Returns
1165  *	Nothing
1166  */
1167 static void
1168 keg_large_init(uma_keg_t keg)
1169 {
1170 	int pages;
1171 
1172 	KASSERT(keg != NULL, ("Keg is null in keg_large_init"));
1173 	KASSERT((keg->uk_flags & UMA_ZFLAG_CACHEONLY) == 0,
1174 	    ("keg_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY keg"));
1175 
1176 	pages = keg->uk_size / UMA_SLAB_SIZE;
1177 
1178 	/* Account for remainder */
1179 	if ((pages * UMA_SLAB_SIZE) < keg->uk_size)
1180 		pages++;
1181 
1182 	keg->uk_ppera = pages;
1183 	keg->uk_ipers = 1;
1184 
1185 	keg->uk_flags |= UMA_ZONE_OFFPAGE;
1186 	if ((keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1187 		keg->uk_flags |= UMA_ZONE_HASH;
1188 
1189 	keg->uk_rsize = keg->uk_size;
1190 }
1191 
1192 static void
1193 keg_cachespread_init(uma_keg_t keg)
1194 {
1195 	int alignsize;
1196 	int trailer;
1197 	int pages;
1198 	int rsize;
1199 
1200 	alignsize = keg->uk_align + 1;
1201 	rsize = keg->uk_size;
1202 	/*
1203 	 * We want one item to start on every align boundary in a page.  To
1204 	 * do this we will span pages.  We will also extend the item by the
1205 	 * size of align if it is an even multiple of align.  Otherwise, it
1206 	 * would fall on the same boundary every time.
1207 	 */
1208 	if (rsize & keg->uk_align)
1209 		rsize = (rsize & ~keg->uk_align) + alignsize;
1210 	if ((rsize & alignsize) == 0)
1211 		rsize += alignsize;
1212 	trailer = rsize - keg->uk_size;
1213 	pages = (rsize * (PAGE_SIZE / alignsize)) / PAGE_SIZE;
1214 	pages = MIN(pages, (128 * 1024) / PAGE_SIZE);
1215 	keg->uk_rsize = rsize;
1216 	keg->uk_ppera = pages;
1217 	keg->uk_ipers = ((pages * PAGE_SIZE) + trailer) / rsize;
1218 	keg->uk_flags |= UMA_ZONE_OFFPAGE | UMA_ZONE_VTOSLAB;
1219 	KASSERT(keg->uk_ipers <= uma_max_ipers,
1220 	    ("keg_small_init: keg->uk_ipers too high(%d) increase max_ipers",
1221 	    keg->uk_ipers));
1222 }
1223 
1224 /*
1225  * Keg header ctor.  This initializes all fields, locks, etc.  And inserts
1226  * the keg onto the global keg list.
1227  *
1228  * Arguments/Returns follow uma_ctor specifications
1229  *	udata  Actually uma_kctor_args
1230  */
1231 static int
1232 keg_ctor(void *mem, int size, void *udata, int flags)
1233 {
1234 	struct uma_kctor_args *arg = udata;
1235 	uma_keg_t keg = mem;
1236 	uma_zone_t zone;
1237 
1238 	bzero(keg, size);
1239 	keg->uk_size = arg->size;
1240 	keg->uk_init = arg->uminit;
1241 	keg->uk_fini = arg->fini;
1242 	keg->uk_align = arg->align;
1243 	keg->uk_free = 0;
1244 	keg->uk_pages = 0;
1245 	keg->uk_flags = arg->flags;
1246 	keg->uk_allocf = page_alloc;
1247 	keg->uk_freef = page_free;
1248 	keg->uk_recurse = 0;
1249 	keg->uk_slabzone = NULL;
1250 
1251 	/*
1252 	 * The master zone is passed to us at keg-creation time.
1253 	 */
1254 	zone = arg->zone;
1255 	keg->uk_name = zone->uz_name;
1256 
1257 	if (arg->flags & UMA_ZONE_VM)
1258 		keg->uk_flags |= UMA_ZFLAG_CACHEONLY;
1259 
1260 	if (arg->flags & UMA_ZONE_ZINIT)
1261 		keg->uk_init = zero_init;
1262 
1263 	if (arg->flags & UMA_ZONE_REFCNT || arg->flags & UMA_ZONE_MALLOC)
1264 		keg->uk_flags |= UMA_ZONE_VTOSLAB;
1265 
1266 	/*
1267 	 * The +UMA_FRITM_SZ added to uk_size is to account for the
1268 	 * linkage that is added to the size in keg_small_init().  If
1269 	 * we don't account for this here then we may end up in
1270 	 * keg_small_init() with a calculated 'ipers' of 0.
1271 	 */
1272 	if (keg->uk_flags & UMA_ZONE_REFCNT) {
1273 		if (keg->uk_flags & UMA_ZONE_CACHESPREAD)
1274 			keg_cachespread_init(keg);
1275 		else if ((keg->uk_size+UMA_FRITMREF_SZ) >
1276 		    (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt)))
1277 			keg_large_init(keg);
1278 		else
1279 			keg_small_init(keg);
1280 	} else {
1281 		if (keg->uk_flags & UMA_ZONE_CACHESPREAD)
1282 			keg_cachespread_init(keg);
1283 		else if ((keg->uk_size+UMA_FRITM_SZ) >
1284 		    (UMA_SLAB_SIZE - sizeof(struct uma_slab)))
1285 			keg_large_init(keg);
1286 		else
1287 			keg_small_init(keg);
1288 	}
1289 
1290 	if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
1291 		if (keg->uk_flags & UMA_ZONE_REFCNT)
1292 			keg->uk_slabzone = slabrefzone;
1293 		else
1294 			keg->uk_slabzone = slabzone;
1295 	}
1296 
1297 	/*
1298 	 * If we haven't booted yet we need allocations to go through the
1299 	 * startup cache until the vm is ready.
1300 	 */
1301 	if (keg->uk_ppera == 1) {
1302 #ifdef UMA_MD_SMALL_ALLOC
1303 		keg->uk_allocf = uma_small_alloc;
1304 		keg->uk_freef = uma_small_free;
1305 #endif
1306 		if (booted == 0)
1307 			keg->uk_allocf = startup_alloc;
1308 	}
1309 
1310 	/*
1311 	 * Initialize keg's lock (shared among zones).
1312 	 */
1313 	if (arg->flags & UMA_ZONE_MTXCLASS)
1314 		KEG_LOCK_INIT(keg, 1);
1315 	else
1316 		KEG_LOCK_INIT(keg, 0);
1317 
1318 	/*
1319 	 * If we're putting the slab header in the actual page we need to
1320 	 * figure out where in each page it goes.  This calculates a right
1321 	 * justified offset into the memory on an ALIGN_PTR boundary.
1322 	 */
1323 	if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) {
1324 		u_int totsize;
1325 
1326 		/* Size of the slab struct and free list */
1327 		if (keg->uk_flags & UMA_ZONE_REFCNT)
1328 			totsize = sizeof(struct uma_slab_refcnt) +
1329 			    keg->uk_ipers * UMA_FRITMREF_SZ;
1330 		else
1331 			totsize = sizeof(struct uma_slab) +
1332 			    keg->uk_ipers * UMA_FRITM_SZ;
1333 
1334 		if (totsize & UMA_ALIGN_PTR)
1335 			totsize = (totsize & ~UMA_ALIGN_PTR) +
1336 			    (UMA_ALIGN_PTR + 1);
1337 		keg->uk_pgoff = UMA_SLAB_SIZE - totsize;
1338 
1339 		if (keg->uk_flags & UMA_ZONE_REFCNT)
1340 			totsize = keg->uk_pgoff + sizeof(struct uma_slab_refcnt)
1341 			    + keg->uk_ipers * UMA_FRITMREF_SZ;
1342 		else
1343 			totsize = keg->uk_pgoff + sizeof(struct uma_slab)
1344 			    + keg->uk_ipers * UMA_FRITM_SZ;
1345 
1346 		/*
1347 		 * The only way the following is possible is if with our
1348 		 * UMA_ALIGN_PTR adjustments we are now bigger than
1349 		 * UMA_SLAB_SIZE.  I haven't checked whether this is
1350 		 * mathematically possible for all cases, so we make
1351 		 * sure here anyway.
1352 		 */
1353 		if (totsize > UMA_SLAB_SIZE) {
1354 			printf("zone %s ipers %d rsize %d size %d\n",
1355 			    zone->uz_name, keg->uk_ipers, keg->uk_rsize,
1356 			    keg->uk_size);
1357 			panic("UMA slab won't fit.");
1358 		}
1359 	}
1360 
1361 	if (keg->uk_flags & UMA_ZONE_HASH)
1362 		hash_alloc(&keg->uk_hash);
1363 
1364 #ifdef UMA_DEBUG
1365 	printf("UMA: %s(%p) size %d(%d) flags %d ipers %d ppera %d out %d free %d\n",
1366 	    zone->uz_name, zone, keg->uk_size, keg->uk_rsize, keg->uk_flags,
1367 	    keg->uk_ipers, keg->uk_ppera,
1368 	    (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free);
1369 #endif
1370 
1371 	LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
1372 
1373 	mtx_lock(&uma_mtx);
1374 	LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
1375 	mtx_unlock(&uma_mtx);
1376 	return (0);
1377 }
1378 
1379 /*
1380  * Zone header ctor.  This initializes all fields, locks, etc.
1381  *
1382  * Arguments/Returns follow uma_ctor specifications
1383  *	udata  Actually uma_zctor_args
1384  */
1385 static int
1386 zone_ctor(void *mem, int size, void *udata, int flags)
1387 {
1388 	struct uma_zctor_args *arg = udata;
1389 	uma_zone_t zone = mem;
1390 	uma_zone_t z;
1391 	uma_keg_t keg;
1392 
1393 	bzero(zone, size);
1394 	zone->uz_name = arg->name;
1395 	zone->uz_ctor = arg->ctor;
1396 	zone->uz_dtor = arg->dtor;
1397 	zone->uz_slab = zone_fetch_slab;
1398 	zone->uz_init = NULL;
1399 	zone->uz_fini = NULL;
1400 	zone->uz_allocs = 0;
1401 	zone->uz_frees = 0;
1402 	zone->uz_fails = 0;
1403 	zone->uz_fills = zone->uz_count = 0;
1404 	zone->uz_flags = 0;
1405 	keg = arg->keg;
1406 
1407 	if (arg->flags & UMA_ZONE_SECONDARY) {
1408 		KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
1409 		zone->uz_init = arg->uminit;
1410 		zone->uz_fini = arg->fini;
1411 		zone->uz_lock = &keg->uk_lock;
1412 		zone->uz_flags |= UMA_ZONE_SECONDARY;
1413 		mtx_lock(&uma_mtx);
1414 		ZONE_LOCK(zone);
1415 		LIST_FOREACH(z, &keg->uk_zones, uz_link) {
1416 			if (LIST_NEXT(z, uz_link) == NULL) {
1417 				LIST_INSERT_AFTER(z, zone, uz_link);
1418 				break;
1419 			}
1420 		}
1421 		ZONE_UNLOCK(zone);
1422 		mtx_unlock(&uma_mtx);
1423 	} else if (keg == NULL) {
1424 		if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
1425 		    arg->align, arg->flags)) == NULL)
1426 			return (ENOMEM);
1427 	} else {
1428 		struct uma_kctor_args karg;
1429 		int error;
1430 
1431 		/* We should only be here from uma_startup() */
1432 		karg.size = arg->size;
1433 		karg.uminit = arg->uminit;
1434 		karg.fini = arg->fini;
1435 		karg.align = arg->align;
1436 		karg.flags = arg->flags;
1437 		karg.zone = zone;
1438 		error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
1439 		    flags);
1440 		if (error)
1441 			return (error);
1442 	}
1443 	/*
1444 	 * Link in the first keg.
1445 	 */
1446 	zone->uz_klink.kl_keg = keg;
1447 	LIST_INSERT_HEAD(&zone->uz_kegs, &zone->uz_klink, kl_link);
1448 	zone->uz_lock = &keg->uk_lock;
1449 	zone->uz_size = keg->uk_size;
1450 	zone->uz_flags |= (keg->uk_flags &
1451 	    (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
1452 
1453 	/*
1454 	 * Some internal zones don't have room allocated for the per cpu
1455 	 * caches.  If we're internal, bail out here.
1456 	 */
1457 	if (keg->uk_flags & UMA_ZFLAG_INTERNAL) {
1458 		KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
1459 		    ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
1460 		return (0);
1461 	}
1462 
1463 	if (keg->uk_flags & UMA_ZONE_MAXBUCKET)
1464 		zone->uz_count = BUCKET_MAX;
1465 	else if (keg->uk_ipers <= BUCKET_MAX)
1466 		zone->uz_count = keg->uk_ipers;
1467 	else
1468 		zone->uz_count = BUCKET_MAX;
1469 	return (0);
1470 }
1471 
1472 /*
1473  * Keg header dtor.  This frees all data, destroys locks, frees the hash
1474  * table and removes the keg from the global list.
1475  *
1476  * Arguments/Returns follow uma_dtor specifications
1477  *	udata  unused
1478  */
1479 static void
1480 keg_dtor(void *arg, int size, void *udata)
1481 {
1482 	uma_keg_t keg;
1483 
1484 	keg = (uma_keg_t)arg;
1485 	KEG_LOCK(keg);
1486 	if (keg->uk_free != 0) {
1487 		printf("Freed UMA keg was not empty (%d items). "
1488 		    " Lost %d pages of memory.\n",
1489 		    keg->uk_free, keg->uk_pages);
1490 	}
1491 	KEG_UNLOCK(keg);
1492 
1493 	hash_free(&keg->uk_hash);
1494 
1495 	KEG_LOCK_FINI(keg);
1496 }
1497 
1498 /*
1499  * Zone header dtor.
1500  *
1501  * Arguments/Returns follow uma_dtor specifications
1502  *	udata  unused
1503  */
1504 static void
1505 zone_dtor(void *arg, int size, void *udata)
1506 {
1507 	uma_klink_t klink;
1508 	uma_zone_t zone;
1509 	uma_keg_t keg;
1510 
1511 	zone = (uma_zone_t)arg;
1512 	keg = zone_first_keg(zone);
1513 
1514 	if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
1515 		cache_drain(zone);
1516 
1517 	mtx_lock(&uma_mtx);
1518 	LIST_REMOVE(zone, uz_link);
1519 	mtx_unlock(&uma_mtx);
1520 	/*
1521 	 * XXX there are some races here where
1522 	 * the zone can be drained but zone lock
1523 	 * released and then refilled before we
1524 	 * remove it... we dont care for now
1525 	 */
1526 	zone_drain_wait(zone, M_WAITOK);
1527 	/*
1528 	 * Unlink all of our kegs.
1529 	 */
1530 	while ((klink = LIST_FIRST(&zone->uz_kegs)) != NULL) {
1531 		klink->kl_keg = NULL;
1532 		LIST_REMOVE(klink, kl_link);
1533 		if (klink == &zone->uz_klink)
1534 			continue;
1535 		free(klink, M_TEMP);
1536 	}
1537 	/*
1538 	 * We only destroy kegs from non secondary zones.
1539 	 */
1540 	if ((zone->uz_flags & UMA_ZONE_SECONDARY) == 0)  {
1541 		mtx_lock(&uma_mtx);
1542 		LIST_REMOVE(keg, uk_link);
1543 		mtx_unlock(&uma_mtx);
1544 		zone_free_item(kegs, keg, NULL, SKIP_NONE,
1545 		    ZFREE_STATFREE);
1546 	}
1547 }
1548 
1549 /*
1550  * Traverses every zone in the system and calls a callback
1551  *
1552  * Arguments:
1553  *	zfunc  A pointer to a function which accepts a zone
1554  *		as an argument.
1555  *
1556  * Returns:
1557  *	Nothing
1558  */
1559 static void
1560 zone_foreach(void (*zfunc)(uma_zone_t))
1561 {
1562 	uma_keg_t keg;
1563 	uma_zone_t zone;
1564 
1565 	mtx_lock(&uma_mtx);
1566 	LIST_FOREACH(keg, &uma_kegs, uk_link) {
1567 		LIST_FOREACH(zone, &keg->uk_zones, uz_link)
1568 			zfunc(zone);
1569 	}
1570 	mtx_unlock(&uma_mtx);
1571 }
1572 
1573 /* Public functions */
1574 /* See uma.h */
1575 void
1576 uma_startup(void *bootmem, int boot_pages)
1577 {
1578 	struct uma_zctor_args args;
1579 	uma_slab_t slab;
1580 	u_int slabsize;
1581 	u_int objsize, totsize, wsize;
1582 	int i;
1583 
1584 #ifdef UMA_DEBUG
1585 	printf("Creating uma keg headers zone and keg.\n");
1586 #endif
1587 	mtx_init(&uma_mtx, "UMA lock", NULL, MTX_DEF);
1588 
1589 	/*
1590 	 * Figure out the maximum number of items-per-slab we'll have if
1591 	 * we're using the OFFPAGE slab header to track free items, given
1592 	 * all possible object sizes and the maximum desired wastage
1593 	 * (UMA_MAX_WASTE).
1594 	 *
1595 	 * We iterate until we find an object size for
1596 	 * which the calculated wastage in keg_small_init() will be
1597 	 * enough to warrant OFFPAGE.  Since wastedspace versus objsize
1598 	 * is an overall increasing see-saw function, we find the smallest
1599 	 * objsize such that the wastage is always acceptable for objects
1600 	 * with that objsize or smaller.  Since a smaller objsize always
1601 	 * generates a larger possible uma_max_ipers, we use this computed
1602 	 * objsize to calculate the largest ipers possible.  Since the
1603 	 * ipers calculated for OFFPAGE slab headers is always larger than
1604 	 * the ipers initially calculated in keg_small_init(), we use
1605 	 * the former's equation (UMA_SLAB_SIZE / keg->uk_rsize) to
1606 	 * obtain the maximum ipers possible for offpage slab headers.
1607 	 *
1608 	 * It should be noted that ipers versus objsize is an inversly
1609 	 * proportional function which drops off rather quickly so as
1610 	 * long as our UMA_MAX_WASTE is such that the objsize we calculate
1611 	 * falls into the portion of the inverse relation AFTER the steep
1612 	 * falloff, then uma_max_ipers shouldn't be too high (~10 on i386).
1613 	 *
1614 	 * Note that we have 8-bits (1 byte) to use as a freelist index
1615 	 * inside the actual slab header itself and this is enough to
1616 	 * accomodate us.  In the worst case, a UMA_SMALLEST_UNIT sized
1617 	 * object with offpage slab header would have ipers =
1618 	 * UMA_SLAB_SIZE / UMA_SMALLEST_UNIT (currently = 256), which is
1619 	 * 1 greater than what our byte-integer freelist index can
1620 	 * accomodate, but we know that this situation never occurs as
1621 	 * for UMA_SMALLEST_UNIT-sized objects, we will never calculate
1622 	 * that we need to go to offpage slab headers.  Or, if we do,
1623 	 * then we trap that condition below and panic in the INVARIANTS case.
1624 	 */
1625 	wsize = UMA_SLAB_SIZE - sizeof(struct uma_slab) - UMA_MAX_WASTE;
1626 	totsize = wsize;
1627 	objsize = UMA_SMALLEST_UNIT;
1628 	while (totsize >= wsize) {
1629 		totsize = (UMA_SLAB_SIZE - sizeof(struct uma_slab)) /
1630 		    (objsize + UMA_FRITM_SZ);
1631 		totsize *= (UMA_FRITM_SZ + objsize);
1632 		objsize++;
1633 	}
1634 	if (objsize > UMA_SMALLEST_UNIT)
1635 		objsize--;
1636 	uma_max_ipers = MAX(UMA_SLAB_SIZE / objsize, 64);
1637 
1638 	wsize = UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt) - UMA_MAX_WASTE;
1639 	totsize = wsize;
1640 	objsize = UMA_SMALLEST_UNIT;
1641 	while (totsize >= wsize) {
1642 		totsize = (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt)) /
1643 		    (objsize + UMA_FRITMREF_SZ);
1644 		totsize *= (UMA_FRITMREF_SZ + objsize);
1645 		objsize++;
1646 	}
1647 	if (objsize > UMA_SMALLEST_UNIT)
1648 		objsize--;
1649 	uma_max_ipers_ref = MAX(UMA_SLAB_SIZE / objsize, 64);
1650 
1651 	KASSERT((uma_max_ipers_ref <= 255) && (uma_max_ipers <= 255),
1652 	    ("uma_startup: calculated uma_max_ipers values too large!"));
1653 
1654 #ifdef UMA_DEBUG
1655 	printf("Calculated uma_max_ipers (for OFFPAGE) is %d\n", uma_max_ipers);
1656 	printf("Calculated uma_max_ipers_slab (for OFFPAGE) is %d\n",
1657 	    uma_max_ipers_ref);
1658 #endif
1659 
1660 	/* "manually" create the initial zone */
1661 	args.name = "UMA Kegs";
1662 	args.size = sizeof(struct uma_keg);
1663 	args.ctor = keg_ctor;
1664 	args.dtor = keg_dtor;
1665 	args.uminit = zero_init;
1666 	args.fini = NULL;
1667 	args.keg = &masterkeg;
1668 	args.align = 32 - 1;
1669 	args.flags = UMA_ZFLAG_INTERNAL;
1670 	/* The initial zone has no Per cpu queues so it's smaller */
1671 	zone_ctor(kegs, sizeof(struct uma_zone), &args, M_WAITOK);
1672 
1673 #ifdef UMA_DEBUG
1674 	printf("Filling boot free list.\n");
1675 #endif
1676 	for (i = 0; i < boot_pages; i++) {
1677 		slab = (uma_slab_t)((u_int8_t *)bootmem + (i * UMA_SLAB_SIZE));
1678 		slab->us_data = (u_int8_t *)slab;
1679 		slab->us_flags = UMA_SLAB_BOOT;
1680 		LIST_INSERT_HEAD(&uma_boot_pages, slab, us_link);
1681 	}
1682 	mtx_init(&uma_boot_pages_mtx, "UMA boot pages", NULL, MTX_DEF);
1683 
1684 #ifdef UMA_DEBUG
1685 	printf("Creating uma zone headers zone and keg.\n");
1686 #endif
1687 	args.name = "UMA Zones";
1688 	args.size = sizeof(struct uma_zone) +
1689 	    (sizeof(struct uma_cache) * (mp_maxid + 1));
1690 	args.ctor = zone_ctor;
1691 	args.dtor = zone_dtor;
1692 	args.uminit = zero_init;
1693 	args.fini = NULL;
1694 	args.keg = NULL;
1695 	args.align = 32 - 1;
1696 	args.flags = UMA_ZFLAG_INTERNAL;
1697 	/* The initial zone has no Per cpu queues so it's smaller */
1698 	zone_ctor(zones, sizeof(struct uma_zone), &args, M_WAITOK);
1699 
1700 #ifdef UMA_DEBUG
1701 	printf("Initializing pcpu cache locks.\n");
1702 #endif
1703 #ifdef UMA_DEBUG
1704 	printf("Creating slab and hash zones.\n");
1705 #endif
1706 
1707 	/*
1708 	 * This is the max number of free list items we'll have with
1709 	 * offpage slabs.
1710 	 */
1711 	slabsize = uma_max_ipers * UMA_FRITM_SZ;
1712 	slabsize += sizeof(struct uma_slab);
1713 
1714 	/* Now make a zone for slab headers */
1715 	slabzone = uma_zcreate("UMA Slabs",
1716 				slabsize,
1717 				NULL, NULL, NULL, NULL,
1718 				UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1719 
1720 	/*
1721 	 * We also create a zone for the bigger slabs with reference
1722 	 * counts in them, to accomodate UMA_ZONE_REFCNT zones.
1723 	 */
1724 	slabsize = uma_max_ipers_ref * UMA_FRITMREF_SZ;
1725 	slabsize += sizeof(struct uma_slab_refcnt);
1726 	slabrefzone = uma_zcreate("UMA RCntSlabs",
1727 				  slabsize,
1728 				  NULL, NULL, NULL, NULL,
1729 				  UMA_ALIGN_PTR,
1730 				  UMA_ZFLAG_INTERNAL);
1731 
1732 	hashzone = uma_zcreate("UMA Hash",
1733 	    sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
1734 	    NULL, NULL, NULL, NULL,
1735 	    UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1736 
1737 	bucket_init();
1738 
1739 #if defined(UMA_MD_SMALL_ALLOC) && !defined(UMA_MD_SMALL_ALLOC_NEEDS_VM)
1740 	booted = 1;
1741 #endif
1742 
1743 #ifdef UMA_DEBUG
1744 	printf("UMA startup complete.\n");
1745 #endif
1746 }
1747 
1748 /* see uma.h */
1749 void
1750 uma_startup2(void)
1751 {
1752 	booted = 1;
1753 	bucket_enable();
1754 #ifdef UMA_DEBUG
1755 	printf("UMA startup2 complete.\n");
1756 #endif
1757 }
1758 
1759 /*
1760  * Initialize our callout handle
1761  *
1762  */
1763 
1764 static void
1765 uma_startup3(void)
1766 {
1767 #ifdef UMA_DEBUG
1768 	printf("Starting callout.\n");
1769 #endif
1770 	callout_init(&uma_callout, CALLOUT_MPSAFE);
1771 	callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
1772 #ifdef UMA_DEBUG
1773 	printf("UMA startup3 complete.\n");
1774 #endif
1775 }
1776 
1777 static uma_keg_t
1778 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
1779 		int align, u_int32_t flags)
1780 {
1781 	struct uma_kctor_args args;
1782 
1783 	args.size = size;
1784 	args.uminit = uminit;
1785 	args.fini = fini;
1786 	args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
1787 	args.flags = flags;
1788 	args.zone = zone;
1789 	return (zone_alloc_item(kegs, &args, M_WAITOK));
1790 }
1791 
1792 /* See uma.h */
1793 void
1794 uma_set_align(int align)
1795 {
1796 
1797 	if (align != UMA_ALIGN_CACHE)
1798 		uma_align_cache = align;
1799 }
1800 
1801 /* See uma.h */
1802 uma_zone_t
1803 uma_zcreate(char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
1804 		uma_init uminit, uma_fini fini, int align, u_int32_t flags)
1805 
1806 {
1807 	struct uma_zctor_args args;
1808 
1809 	/* This stuff is essential for the zone ctor */
1810 	args.name = name;
1811 	args.size = size;
1812 	args.ctor = ctor;
1813 	args.dtor = dtor;
1814 	args.uminit = uminit;
1815 	args.fini = fini;
1816 	args.align = align;
1817 	args.flags = flags;
1818 	args.keg = NULL;
1819 
1820 	return (zone_alloc_item(zones, &args, M_WAITOK));
1821 }
1822 
1823 /* See uma.h */
1824 uma_zone_t
1825 uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor,
1826 		    uma_init zinit, uma_fini zfini, uma_zone_t master)
1827 {
1828 	struct uma_zctor_args args;
1829 	uma_keg_t keg;
1830 
1831 	keg = zone_first_keg(master);
1832 	args.name = name;
1833 	args.size = keg->uk_size;
1834 	args.ctor = ctor;
1835 	args.dtor = dtor;
1836 	args.uminit = zinit;
1837 	args.fini = zfini;
1838 	args.align = keg->uk_align;
1839 	args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
1840 	args.keg = keg;
1841 
1842 	/* XXX Attaches only one keg of potentially many. */
1843 	return (zone_alloc_item(zones, &args, M_WAITOK));
1844 }
1845 
1846 static void
1847 zone_lock_pair(uma_zone_t a, uma_zone_t b)
1848 {
1849 	if (a < b) {
1850 		ZONE_LOCK(a);
1851 		mtx_lock_flags(b->uz_lock, MTX_DUPOK);
1852 	} else {
1853 		ZONE_LOCK(b);
1854 		mtx_lock_flags(a->uz_lock, MTX_DUPOK);
1855 	}
1856 }
1857 
1858 static void
1859 zone_unlock_pair(uma_zone_t a, uma_zone_t b)
1860 {
1861 
1862 	ZONE_UNLOCK(a);
1863 	ZONE_UNLOCK(b);
1864 }
1865 
1866 int
1867 uma_zsecond_add(uma_zone_t zone, uma_zone_t master)
1868 {
1869 	uma_klink_t klink;
1870 	uma_klink_t kl;
1871 	int error;
1872 
1873 	error = 0;
1874 	klink = malloc(sizeof(*klink), M_TEMP, M_WAITOK | M_ZERO);
1875 
1876 	zone_lock_pair(zone, master);
1877 	/*
1878 	 * zone must use vtoslab() to resolve objects and must already be
1879 	 * a secondary.
1880 	 */
1881 	if ((zone->uz_flags & (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY))
1882 	    != (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) {
1883 		error = EINVAL;
1884 		goto out;
1885 	}
1886 	/*
1887 	 * The new master must also use vtoslab().
1888 	 */
1889 	if ((zone->uz_flags & UMA_ZONE_VTOSLAB) != UMA_ZONE_VTOSLAB) {
1890 		error = EINVAL;
1891 		goto out;
1892 	}
1893 	/*
1894 	 * Both must either be refcnt, or not be refcnt.
1895 	 */
1896 	if ((zone->uz_flags & UMA_ZONE_REFCNT) !=
1897 	    (master->uz_flags & UMA_ZONE_REFCNT)) {
1898 		error = EINVAL;
1899 		goto out;
1900 	}
1901 	/*
1902 	 * The underlying object must be the same size.  rsize
1903 	 * may be different.
1904 	 */
1905 	if (master->uz_size != zone->uz_size) {
1906 		error = E2BIG;
1907 		goto out;
1908 	}
1909 	/*
1910 	 * Put it at the end of the list.
1911 	 */
1912 	klink->kl_keg = zone_first_keg(master);
1913 	LIST_FOREACH(kl, &zone->uz_kegs, kl_link) {
1914 		if (LIST_NEXT(kl, kl_link) == NULL) {
1915 			LIST_INSERT_AFTER(kl, klink, kl_link);
1916 			break;
1917 		}
1918 	}
1919 	klink = NULL;
1920 	zone->uz_flags |= UMA_ZFLAG_MULTI;
1921 	zone->uz_slab = zone_fetch_slab_multi;
1922 
1923 out:
1924 	zone_unlock_pair(zone, master);
1925 	if (klink != NULL)
1926 		free(klink, M_TEMP);
1927 
1928 	return (error);
1929 }
1930 
1931 
1932 /* See uma.h */
1933 void
1934 uma_zdestroy(uma_zone_t zone)
1935 {
1936 
1937 	zone_free_item(zones, zone, NULL, SKIP_NONE, ZFREE_STATFREE);
1938 }
1939 
1940 /* See uma.h */
1941 void *
1942 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
1943 {
1944 	void *item;
1945 	uma_cache_t cache;
1946 	uma_bucket_t bucket;
1947 	int cpu;
1948 
1949 	/* This is the fast path allocation */
1950 #ifdef UMA_DEBUG_ALLOC_1
1951 	printf("Allocating one item from %s(%p)\n", zone->uz_name, zone);
1952 #endif
1953 	CTR3(KTR_UMA, "uma_zalloc_arg thread %x zone %s flags %d", curthread,
1954 	    zone->uz_name, flags);
1955 
1956 	if (flags & M_WAITOK) {
1957 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
1958 		    "uma_zalloc_arg: zone \"%s\"", zone->uz_name);
1959 	}
1960 
1961 	/*
1962 	 * If possible, allocate from the per-CPU cache.  There are two
1963 	 * requirements for safe access to the per-CPU cache: (1) the thread
1964 	 * accessing the cache must not be preempted or yield during access,
1965 	 * and (2) the thread must not migrate CPUs without switching which
1966 	 * cache it accesses.  We rely on a critical section to prevent
1967 	 * preemption and migration.  We release the critical section in
1968 	 * order to acquire the zone mutex if we are unable to allocate from
1969 	 * the current cache; when we re-acquire the critical section, we
1970 	 * must detect and handle migration if it has occurred.
1971 	 */
1972 zalloc_restart:
1973 	critical_enter();
1974 	cpu = curcpu;
1975 	cache = &zone->uz_cpu[cpu];
1976 
1977 zalloc_start:
1978 	bucket = cache->uc_allocbucket;
1979 
1980 	if (bucket) {
1981 		if (bucket->ub_cnt > 0) {
1982 			bucket->ub_cnt--;
1983 			item = bucket->ub_bucket[bucket->ub_cnt];
1984 #ifdef INVARIANTS
1985 			bucket->ub_bucket[bucket->ub_cnt] = NULL;
1986 #endif
1987 			KASSERT(item != NULL,
1988 			    ("uma_zalloc: Bucket pointer mangled."));
1989 			cache->uc_allocs++;
1990 			critical_exit();
1991 #ifdef INVARIANTS
1992 			ZONE_LOCK(zone);
1993 			uma_dbg_alloc(zone, NULL, item);
1994 			ZONE_UNLOCK(zone);
1995 #endif
1996 			if (zone->uz_ctor != NULL) {
1997 				if (zone->uz_ctor(item, zone->uz_size,
1998 				    udata, flags) != 0) {
1999 					zone_free_item(zone, item, udata,
2000 					    SKIP_DTOR, ZFREE_STATFAIL |
2001 					    ZFREE_STATFREE);
2002 					return (NULL);
2003 				}
2004 			}
2005 			if (flags & M_ZERO)
2006 				bzero(item, zone->uz_size);
2007 			return (item);
2008 		} else if (cache->uc_freebucket) {
2009 			/*
2010 			 * We have run out of items in our allocbucket.
2011 			 * See if we can switch with our free bucket.
2012 			 */
2013 			if (cache->uc_freebucket->ub_cnt > 0) {
2014 #ifdef UMA_DEBUG_ALLOC
2015 				printf("uma_zalloc: Swapping empty with"
2016 				    " alloc.\n");
2017 #endif
2018 				bucket = cache->uc_freebucket;
2019 				cache->uc_freebucket = cache->uc_allocbucket;
2020 				cache->uc_allocbucket = bucket;
2021 
2022 				goto zalloc_start;
2023 			}
2024 		}
2025 	}
2026 	/*
2027 	 * Attempt to retrieve the item from the per-CPU cache has failed, so
2028 	 * we must go back to the zone.  This requires the zone lock, so we
2029 	 * must drop the critical section, then re-acquire it when we go back
2030 	 * to the cache.  Since the critical section is released, we may be
2031 	 * preempted or migrate.  As such, make sure not to maintain any
2032 	 * thread-local state specific to the cache from prior to releasing
2033 	 * the critical section.
2034 	 */
2035 	critical_exit();
2036 	ZONE_LOCK(zone);
2037 	critical_enter();
2038 	cpu = curcpu;
2039 	cache = &zone->uz_cpu[cpu];
2040 	bucket = cache->uc_allocbucket;
2041 	if (bucket != NULL) {
2042 		if (bucket->ub_cnt > 0) {
2043 			ZONE_UNLOCK(zone);
2044 			goto zalloc_start;
2045 		}
2046 		bucket = cache->uc_freebucket;
2047 		if (bucket != NULL && bucket->ub_cnt > 0) {
2048 			ZONE_UNLOCK(zone);
2049 			goto zalloc_start;
2050 		}
2051 	}
2052 
2053 	/* Since we have locked the zone we may as well send back our stats */
2054 	zone->uz_allocs += cache->uc_allocs;
2055 	cache->uc_allocs = 0;
2056 	zone->uz_frees += cache->uc_frees;
2057 	cache->uc_frees = 0;
2058 
2059 	/* Our old one is now a free bucket */
2060 	if (cache->uc_allocbucket) {
2061 		KASSERT(cache->uc_allocbucket->ub_cnt == 0,
2062 		    ("uma_zalloc_arg: Freeing a non free bucket."));
2063 		LIST_INSERT_HEAD(&zone->uz_free_bucket,
2064 		    cache->uc_allocbucket, ub_link);
2065 		cache->uc_allocbucket = NULL;
2066 	}
2067 
2068 	/* Check the free list for a new alloc bucket */
2069 	if ((bucket = LIST_FIRST(&zone->uz_full_bucket)) != NULL) {
2070 		KASSERT(bucket->ub_cnt != 0,
2071 		    ("uma_zalloc_arg: Returning an empty bucket."));
2072 
2073 		LIST_REMOVE(bucket, ub_link);
2074 		cache->uc_allocbucket = bucket;
2075 		ZONE_UNLOCK(zone);
2076 		goto zalloc_start;
2077 	}
2078 	/* We are no longer associated with this CPU. */
2079 	critical_exit();
2080 
2081 	/* Bump up our uz_count so we get here less */
2082 	if (zone->uz_count < BUCKET_MAX)
2083 		zone->uz_count++;
2084 
2085 	/*
2086 	 * Now lets just fill a bucket and put it on the free list.  If that
2087 	 * works we'll restart the allocation from the begining.
2088 	 */
2089 	if (zone_alloc_bucket(zone, flags)) {
2090 		ZONE_UNLOCK(zone);
2091 		goto zalloc_restart;
2092 	}
2093 	ZONE_UNLOCK(zone);
2094 	/*
2095 	 * We may not be able to get a bucket so return an actual item.
2096 	 */
2097 #ifdef UMA_DEBUG
2098 	printf("uma_zalloc_arg: Bucketzone returned NULL\n");
2099 #endif
2100 
2101 	item = zone_alloc_item(zone, udata, flags);
2102 	return (item);
2103 }
2104 
2105 static uma_slab_t
2106 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int flags)
2107 {
2108 	uma_slab_t slab;
2109 
2110 	mtx_assert(&keg->uk_lock, MA_OWNED);
2111 	slab = NULL;
2112 
2113 	for (;;) {
2114 		/*
2115 		 * Find a slab with some space.  Prefer slabs that are partially
2116 		 * used over those that are totally full.  This helps to reduce
2117 		 * fragmentation.
2118 		 */
2119 		if (keg->uk_free != 0) {
2120 			if (!LIST_EMPTY(&keg->uk_part_slab)) {
2121 				slab = LIST_FIRST(&keg->uk_part_slab);
2122 			} else {
2123 				slab = LIST_FIRST(&keg->uk_free_slab);
2124 				LIST_REMOVE(slab, us_link);
2125 				LIST_INSERT_HEAD(&keg->uk_part_slab, slab,
2126 				    us_link);
2127 			}
2128 			MPASS(slab->us_keg == keg);
2129 			return (slab);
2130 		}
2131 
2132 		/*
2133 		 * M_NOVM means don't ask at all!
2134 		 */
2135 		if (flags & M_NOVM)
2136 			break;
2137 
2138 		if (keg->uk_maxpages && keg->uk_pages >= keg->uk_maxpages) {
2139 			keg->uk_flags |= UMA_ZFLAG_FULL;
2140 			/*
2141 			 * If this is not a multi-zone, set the FULL bit.
2142 			 * Otherwise slab_multi() takes care of it.
2143 			 */
2144 			if ((zone->uz_flags & UMA_ZFLAG_MULTI) == 0)
2145 				zone->uz_flags |= UMA_ZFLAG_FULL;
2146 			if (flags & M_NOWAIT)
2147 				break;
2148 			msleep(keg, &keg->uk_lock, PVM, "keglimit", 0);
2149 			continue;
2150 		}
2151 		keg->uk_recurse++;
2152 		slab = keg_alloc_slab(keg, zone, flags);
2153 		keg->uk_recurse--;
2154 		/*
2155 		 * If we got a slab here it's safe to mark it partially used
2156 		 * and return.  We assume that the caller is going to remove
2157 		 * at least one item.
2158 		 */
2159 		if (slab) {
2160 			MPASS(slab->us_keg == keg);
2161 			LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2162 			return (slab);
2163 		}
2164 		/*
2165 		 * We might not have been able to get a slab but another cpu
2166 		 * could have while we were unlocked.  Check again before we
2167 		 * fail.
2168 		 */
2169 		flags |= M_NOVM;
2170 	}
2171 	return (slab);
2172 }
2173 
2174 static inline void
2175 zone_relock(uma_zone_t zone, uma_keg_t keg)
2176 {
2177 	if (zone->uz_lock != &keg->uk_lock) {
2178 		KEG_UNLOCK(keg);
2179 		ZONE_LOCK(zone);
2180 	}
2181 }
2182 
2183 static inline void
2184 keg_relock(uma_keg_t keg, uma_zone_t zone)
2185 {
2186 	if (zone->uz_lock != &keg->uk_lock) {
2187 		ZONE_UNLOCK(zone);
2188 		KEG_LOCK(keg);
2189 	}
2190 }
2191 
2192 static uma_slab_t
2193 zone_fetch_slab(uma_zone_t zone, uma_keg_t keg, int flags)
2194 {
2195 	uma_slab_t slab;
2196 
2197 	if (keg == NULL)
2198 		keg = zone_first_keg(zone);
2199 	/*
2200 	 * This is to prevent us from recursively trying to allocate
2201 	 * buckets.  The problem is that if an allocation forces us to
2202 	 * grab a new bucket we will call page_alloc, which will go off
2203 	 * and cause the vm to allocate vm_map_entries.  If we need new
2204 	 * buckets there too we will recurse in kmem_alloc and bad
2205 	 * things happen.  So instead we return a NULL bucket, and make
2206 	 * the code that allocates buckets smart enough to deal with it
2207 	 */
2208 	if (keg->uk_flags & UMA_ZFLAG_BUCKET && keg->uk_recurse != 0)
2209 		return (NULL);
2210 
2211 	for (;;) {
2212 		slab = keg_fetch_slab(keg, zone, flags);
2213 		if (slab)
2214 			return (slab);
2215 		if (flags & (M_NOWAIT | M_NOVM))
2216 			break;
2217 	}
2218 	return (NULL);
2219 }
2220 
2221 /*
2222  * uma_zone_fetch_slab_multi:  Fetches a slab from one available keg.  Returns
2223  * with the keg locked.  Caller must call zone_relock() afterwards if the
2224  * zone lock is required.  On NULL the zone lock is held.
2225  *
2226  * The last pointer is used to seed the search.  It is not required.
2227  */
2228 static uma_slab_t
2229 zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int rflags)
2230 {
2231 	uma_klink_t klink;
2232 	uma_slab_t slab;
2233 	uma_keg_t keg;
2234 	int flags;
2235 	int empty;
2236 	int full;
2237 
2238 	/*
2239 	 * Don't wait on the first pass.  This will skip limit tests
2240 	 * as well.  We don't want to block if we can find a provider
2241 	 * without blocking.
2242 	 */
2243 	flags = (rflags & ~M_WAITOK) | M_NOWAIT;
2244 	/*
2245 	 * Use the last slab allocated as a hint for where to start
2246 	 * the search.
2247 	 */
2248 	if (last) {
2249 		slab = keg_fetch_slab(last, zone, flags);
2250 		if (slab)
2251 			return (slab);
2252 		zone_relock(zone, last);
2253 		last = NULL;
2254 	}
2255 	/*
2256 	 * Loop until we have a slab incase of transient failures
2257 	 * while M_WAITOK is specified.  I'm not sure this is 100%
2258 	 * required but we've done it for so long now.
2259 	 */
2260 	for (;;) {
2261 		empty = 0;
2262 		full = 0;
2263 		/*
2264 		 * Search the available kegs for slabs.  Be careful to hold the
2265 		 * correct lock while calling into the keg layer.
2266 		 */
2267 		LIST_FOREACH(klink, &zone->uz_kegs, kl_link) {
2268 			keg = klink->kl_keg;
2269 			keg_relock(keg, zone);
2270 			if ((keg->uk_flags & UMA_ZFLAG_FULL) == 0) {
2271 				slab = keg_fetch_slab(keg, zone, flags);
2272 				if (slab)
2273 					return (slab);
2274 			}
2275 			if (keg->uk_flags & UMA_ZFLAG_FULL)
2276 				full++;
2277 			else
2278 				empty++;
2279 			zone_relock(zone, keg);
2280 		}
2281 		if (rflags & (M_NOWAIT | M_NOVM))
2282 			break;
2283 		flags = rflags;
2284 		/*
2285 		 * All kegs are full.  XXX We can't atomically check all kegs
2286 		 * and sleep so just sleep for a short period and retry.
2287 		 */
2288 		if (full && !empty) {
2289 			zone->uz_flags |= UMA_ZFLAG_FULL;
2290 			msleep(zone, zone->uz_lock, PVM, "zonelimit", hz/100);
2291 			zone->uz_flags &= ~UMA_ZFLAG_FULL;
2292 			continue;
2293 		}
2294 	}
2295 	return (NULL);
2296 }
2297 
2298 static void *
2299 slab_alloc_item(uma_zone_t zone, uma_slab_t slab)
2300 {
2301 	uma_keg_t keg;
2302 	uma_slabrefcnt_t slabref;
2303 	void *item;
2304 	u_int8_t freei;
2305 
2306 	keg = slab->us_keg;
2307 	mtx_assert(&keg->uk_lock, MA_OWNED);
2308 
2309 	freei = slab->us_firstfree;
2310 	if (keg->uk_flags & UMA_ZONE_REFCNT) {
2311 		slabref = (uma_slabrefcnt_t)slab;
2312 		slab->us_firstfree = slabref->us_freelist[freei].us_item;
2313 	} else {
2314 		slab->us_firstfree = slab->us_freelist[freei].us_item;
2315 	}
2316 	item = slab->us_data + (keg->uk_rsize * freei);
2317 
2318 	slab->us_freecount--;
2319 	keg->uk_free--;
2320 #ifdef INVARIANTS
2321 	uma_dbg_alloc(zone, slab, item);
2322 #endif
2323 	/* Move this slab to the full list */
2324 	if (slab->us_freecount == 0) {
2325 		LIST_REMOVE(slab, us_link);
2326 		LIST_INSERT_HEAD(&keg->uk_full_slab, slab, us_link);
2327 	}
2328 
2329 	return (item);
2330 }
2331 
2332 static int
2333 zone_alloc_bucket(uma_zone_t zone, int flags)
2334 {
2335 	uma_bucket_t bucket;
2336 	uma_slab_t slab;
2337 	uma_keg_t keg;
2338 	int16_t saved;
2339 	int max, origflags = flags;
2340 
2341 	/*
2342 	 * Try this zone's free list first so we don't allocate extra buckets.
2343 	 */
2344 	if ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) {
2345 		KASSERT(bucket->ub_cnt == 0,
2346 		    ("zone_alloc_bucket: Bucket on free list is not empty."));
2347 		LIST_REMOVE(bucket, ub_link);
2348 	} else {
2349 		int bflags;
2350 
2351 		bflags = (flags & ~M_ZERO);
2352 		if (zone->uz_flags & UMA_ZFLAG_CACHEONLY)
2353 			bflags |= M_NOVM;
2354 
2355 		ZONE_UNLOCK(zone);
2356 		bucket = bucket_alloc(zone->uz_count, bflags);
2357 		ZONE_LOCK(zone);
2358 	}
2359 
2360 	if (bucket == NULL) {
2361 		return (0);
2362 	}
2363 
2364 #ifdef SMP
2365 	/*
2366 	 * This code is here to limit the number of simultaneous bucket fills
2367 	 * for any given zone to the number of per cpu caches in this zone. This
2368 	 * is done so that we don't allocate more memory than we really need.
2369 	 */
2370 	if (zone->uz_fills >= mp_ncpus)
2371 		goto done;
2372 
2373 #endif
2374 	zone->uz_fills++;
2375 
2376 	max = MIN(bucket->ub_entries, zone->uz_count);
2377 	/* Try to keep the buckets totally full */
2378 	saved = bucket->ub_cnt;
2379 	slab = NULL;
2380 	keg = NULL;
2381 	while (bucket->ub_cnt < max &&
2382 	    (slab = zone->uz_slab(zone, keg, flags)) != NULL) {
2383 		keg = slab->us_keg;
2384 		while (slab->us_freecount && bucket->ub_cnt < max) {
2385 			bucket->ub_bucket[bucket->ub_cnt++] =
2386 			    slab_alloc_item(zone, slab);
2387 		}
2388 
2389 		/* Don't block on the next fill */
2390 		flags |= M_NOWAIT;
2391 	}
2392 	if (slab)
2393 		zone_relock(zone, keg);
2394 
2395 	/*
2396 	 * We unlock here because we need to call the zone's init.
2397 	 * It should be safe to unlock because the slab dealt with
2398 	 * above is already on the appropriate list within the keg
2399 	 * and the bucket we filled is not yet on any list, so we
2400 	 * own it.
2401 	 */
2402 	if (zone->uz_init != NULL) {
2403 		int i;
2404 
2405 		ZONE_UNLOCK(zone);
2406 		for (i = saved; i < bucket->ub_cnt; i++)
2407 			if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size,
2408 			    origflags) != 0)
2409 				break;
2410 		/*
2411 		 * If we couldn't initialize the whole bucket, put the
2412 		 * rest back onto the freelist.
2413 		 */
2414 		if (i != bucket->ub_cnt) {
2415 			int j;
2416 
2417 			for (j = i; j < bucket->ub_cnt; j++) {
2418 				zone_free_item(zone, bucket->ub_bucket[j],
2419 				    NULL, SKIP_FINI, 0);
2420 #ifdef INVARIANTS
2421 				bucket->ub_bucket[j] = NULL;
2422 #endif
2423 			}
2424 			bucket->ub_cnt = i;
2425 		}
2426 		ZONE_LOCK(zone);
2427 	}
2428 
2429 	zone->uz_fills--;
2430 	if (bucket->ub_cnt != 0) {
2431 		LIST_INSERT_HEAD(&zone->uz_full_bucket,
2432 		    bucket, ub_link);
2433 		return (1);
2434 	}
2435 #ifdef SMP
2436 done:
2437 #endif
2438 	bucket_free(bucket);
2439 
2440 	return (0);
2441 }
2442 /*
2443  * Allocates an item for an internal zone
2444  *
2445  * Arguments
2446  *	zone   The zone to alloc for.
2447  *	udata  The data to be passed to the constructor.
2448  *	flags  M_WAITOK, M_NOWAIT, M_ZERO.
2449  *
2450  * Returns
2451  *	NULL if there is no memory and M_NOWAIT is set
2452  *	An item if successful
2453  */
2454 
2455 static void *
2456 zone_alloc_item(uma_zone_t zone, void *udata, int flags)
2457 {
2458 	uma_slab_t slab;
2459 	void *item;
2460 
2461 	item = NULL;
2462 
2463 #ifdef UMA_DEBUG_ALLOC
2464 	printf("INTERNAL: Allocating one item from %s(%p)\n", zone->uz_name, zone);
2465 #endif
2466 	ZONE_LOCK(zone);
2467 
2468 	slab = zone->uz_slab(zone, NULL, flags);
2469 	if (slab == NULL) {
2470 		zone->uz_fails++;
2471 		ZONE_UNLOCK(zone);
2472 		return (NULL);
2473 	}
2474 
2475 	item = slab_alloc_item(zone, slab);
2476 
2477 	zone_relock(zone, slab->us_keg);
2478 	zone->uz_allocs++;
2479 	ZONE_UNLOCK(zone);
2480 
2481 	/*
2482 	 * We have to call both the zone's init (not the keg's init)
2483 	 * and the zone's ctor.  This is because the item is going from
2484 	 * a keg slab directly to the user, and the user is expecting it
2485 	 * to be both zone-init'd as well as zone-ctor'd.
2486 	 */
2487 	if (zone->uz_init != NULL) {
2488 		if (zone->uz_init(item, zone->uz_size, flags) != 0) {
2489 			zone_free_item(zone, item, udata, SKIP_FINI,
2490 			    ZFREE_STATFAIL | ZFREE_STATFREE);
2491 			return (NULL);
2492 		}
2493 	}
2494 	if (zone->uz_ctor != NULL) {
2495 		if (zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2496 			zone_free_item(zone, item, udata, SKIP_DTOR,
2497 			    ZFREE_STATFAIL | ZFREE_STATFREE);
2498 			return (NULL);
2499 		}
2500 	}
2501 	if (flags & M_ZERO)
2502 		bzero(item, zone->uz_size);
2503 
2504 	return (item);
2505 }
2506 
2507 /* See uma.h */
2508 void
2509 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
2510 {
2511 	uma_cache_t cache;
2512 	uma_bucket_t bucket;
2513 	int bflags;
2514 	int cpu;
2515 
2516 #ifdef UMA_DEBUG_ALLOC_1
2517 	printf("Freeing item %p to %s(%p)\n", item, zone->uz_name, zone);
2518 #endif
2519 	CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread,
2520 	    zone->uz_name);
2521 
2522 	if (zone->uz_dtor)
2523 		zone->uz_dtor(item, zone->uz_size, udata);
2524 
2525 #ifdef INVARIANTS
2526 	ZONE_LOCK(zone);
2527 	if (zone->uz_flags & UMA_ZONE_MALLOC)
2528 		uma_dbg_free(zone, udata, item);
2529 	else
2530 		uma_dbg_free(zone, NULL, item);
2531 	ZONE_UNLOCK(zone);
2532 #endif
2533 	/*
2534 	 * The race here is acceptable.  If we miss it we'll just have to wait
2535 	 * a little longer for the limits to be reset.
2536 	 */
2537 	if (zone->uz_flags & UMA_ZFLAG_FULL)
2538 		goto zfree_internal;
2539 
2540 	/*
2541 	 * If possible, free to the per-CPU cache.  There are two
2542 	 * requirements for safe access to the per-CPU cache: (1) the thread
2543 	 * accessing the cache must not be preempted or yield during access,
2544 	 * and (2) the thread must not migrate CPUs without switching which
2545 	 * cache it accesses.  We rely on a critical section to prevent
2546 	 * preemption and migration.  We release the critical section in
2547 	 * order to acquire the zone mutex if we are unable to free to the
2548 	 * current cache; when we re-acquire the critical section, we must
2549 	 * detect and handle migration if it has occurred.
2550 	 */
2551 zfree_restart:
2552 	critical_enter();
2553 	cpu = curcpu;
2554 	cache = &zone->uz_cpu[cpu];
2555 
2556 zfree_start:
2557 	bucket = cache->uc_freebucket;
2558 
2559 	if (bucket) {
2560 		/*
2561 		 * Do we have room in our bucket? It is OK for this uz count
2562 		 * check to be slightly out of sync.
2563 		 */
2564 
2565 		if (bucket->ub_cnt < bucket->ub_entries) {
2566 			KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL,
2567 			    ("uma_zfree: Freeing to non free bucket index."));
2568 			bucket->ub_bucket[bucket->ub_cnt] = item;
2569 			bucket->ub_cnt++;
2570 			cache->uc_frees++;
2571 			critical_exit();
2572 			return;
2573 		} else if (cache->uc_allocbucket) {
2574 #ifdef UMA_DEBUG_ALLOC
2575 			printf("uma_zfree: Swapping buckets.\n");
2576 #endif
2577 			/*
2578 			 * We have run out of space in our freebucket.
2579 			 * See if we can switch with our alloc bucket.
2580 			 */
2581 			if (cache->uc_allocbucket->ub_cnt <
2582 			    cache->uc_freebucket->ub_cnt) {
2583 				bucket = cache->uc_freebucket;
2584 				cache->uc_freebucket = cache->uc_allocbucket;
2585 				cache->uc_allocbucket = bucket;
2586 				goto zfree_start;
2587 			}
2588 		}
2589 	}
2590 	/*
2591 	 * We can get here for two reasons:
2592 	 *
2593 	 * 1) The buckets are NULL
2594 	 * 2) The alloc and free buckets are both somewhat full.
2595 	 *
2596 	 * We must go back the zone, which requires acquiring the zone lock,
2597 	 * which in turn means we must release and re-acquire the critical
2598 	 * section.  Since the critical section is released, we may be
2599 	 * preempted or migrate.  As such, make sure not to maintain any
2600 	 * thread-local state specific to the cache from prior to releasing
2601 	 * the critical section.
2602 	 */
2603 	critical_exit();
2604 	ZONE_LOCK(zone);
2605 	critical_enter();
2606 	cpu = curcpu;
2607 	cache = &zone->uz_cpu[cpu];
2608 	if (cache->uc_freebucket != NULL) {
2609 		if (cache->uc_freebucket->ub_cnt <
2610 		    cache->uc_freebucket->ub_entries) {
2611 			ZONE_UNLOCK(zone);
2612 			goto zfree_start;
2613 		}
2614 		if (cache->uc_allocbucket != NULL &&
2615 		    (cache->uc_allocbucket->ub_cnt <
2616 		    cache->uc_freebucket->ub_cnt)) {
2617 			ZONE_UNLOCK(zone);
2618 			goto zfree_start;
2619 		}
2620 	}
2621 
2622 	/* Since we have locked the zone we may as well send back our stats */
2623 	zone->uz_allocs += cache->uc_allocs;
2624 	cache->uc_allocs = 0;
2625 	zone->uz_frees += cache->uc_frees;
2626 	cache->uc_frees = 0;
2627 
2628 	bucket = cache->uc_freebucket;
2629 	cache->uc_freebucket = NULL;
2630 
2631 	/* Can we throw this on the zone full list? */
2632 	if (bucket != NULL) {
2633 #ifdef UMA_DEBUG_ALLOC
2634 		printf("uma_zfree: Putting old bucket on the free list.\n");
2635 #endif
2636 		/* ub_cnt is pointing to the last free item */
2637 		KASSERT(bucket->ub_cnt != 0,
2638 		    ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n"));
2639 		LIST_INSERT_HEAD(&zone->uz_full_bucket,
2640 		    bucket, ub_link);
2641 	}
2642 	if ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) {
2643 		LIST_REMOVE(bucket, ub_link);
2644 		ZONE_UNLOCK(zone);
2645 		cache->uc_freebucket = bucket;
2646 		goto zfree_start;
2647 	}
2648 	/* We are no longer associated with this CPU. */
2649 	critical_exit();
2650 
2651 	/* And the zone.. */
2652 	ZONE_UNLOCK(zone);
2653 
2654 #ifdef UMA_DEBUG_ALLOC
2655 	printf("uma_zfree: Allocating new free bucket.\n");
2656 #endif
2657 	bflags = M_NOWAIT;
2658 
2659 	if (zone->uz_flags & UMA_ZFLAG_CACHEONLY)
2660 		bflags |= M_NOVM;
2661 	bucket = bucket_alloc(zone->uz_count, bflags);
2662 	if (bucket) {
2663 		ZONE_LOCK(zone);
2664 		LIST_INSERT_HEAD(&zone->uz_free_bucket,
2665 		    bucket, ub_link);
2666 		ZONE_UNLOCK(zone);
2667 		goto zfree_restart;
2668 	}
2669 
2670 	/*
2671 	 * If nothing else caught this, we'll just do an internal free.
2672 	 */
2673 zfree_internal:
2674 	zone_free_item(zone, item, udata, SKIP_DTOR, ZFREE_STATFREE);
2675 
2676 	return;
2677 }
2678 
2679 /*
2680  * Frees an item to an INTERNAL zone or allocates a free bucket
2681  *
2682  * Arguments:
2683  *	zone   The zone to free to
2684  *	item   The item we're freeing
2685  *	udata  User supplied data for the dtor
2686  *	skip   Skip dtors and finis
2687  */
2688 static void
2689 zone_free_item(uma_zone_t zone, void *item, void *udata,
2690     enum zfreeskip skip, int flags)
2691 {
2692 	uma_slab_t slab;
2693 	uma_slabrefcnt_t slabref;
2694 	uma_keg_t keg;
2695 	u_int8_t *mem;
2696 	u_int8_t freei;
2697 	int clearfull;
2698 
2699 	if (skip < SKIP_DTOR && zone->uz_dtor)
2700 		zone->uz_dtor(item, zone->uz_size, udata);
2701 
2702 	if (skip < SKIP_FINI && zone->uz_fini)
2703 		zone->uz_fini(item, zone->uz_size);
2704 
2705 	ZONE_LOCK(zone);
2706 
2707 	if (flags & ZFREE_STATFAIL)
2708 		zone->uz_fails++;
2709 	if (flags & ZFREE_STATFREE)
2710 		zone->uz_frees++;
2711 
2712 	if (!(zone->uz_flags & UMA_ZONE_VTOSLAB)) {
2713 		mem = (u_int8_t *)((unsigned long)item & (~UMA_SLAB_MASK));
2714 		keg = zone_first_keg(zone); /* Must only be one. */
2715 		if (zone->uz_flags & UMA_ZONE_HASH) {
2716 			slab = hash_sfind(&keg->uk_hash, mem);
2717 		} else {
2718 			mem += keg->uk_pgoff;
2719 			slab = (uma_slab_t)mem;
2720 		}
2721 	} else {
2722 		/* This prevents redundant lookups via free(). */
2723 		if ((zone->uz_flags & UMA_ZONE_MALLOC) && udata != NULL)
2724 			slab = (uma_slab_t)udata;
2725 		else
2726 			slab = vtoslab((vm_offset_t)item);
2727 		keg = slab->us_keg;
2728 		keg_relock(keg, zone);
2729 	}
2730 	MPASS(keg == slab->us_keg);
2731 
2732 	/* Do we need to remove from any lists? */
2733 	if (slab->us_freecount+1 == keg->uk_ipers) {
2734 		LIST_REMOVE(slab, us_link);
2735 		LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
2736 	} else if (slab->us_freecount == 0) {
2737 		LIST_REMOVE(slab, us_link);
2738 		LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2739 	}
2740 
2741 	/* Slab management stuff */
2742 	freei = ((unsigned long)item - (unsigned long)slab->us_data)
2743 		/ keg->uk_rsize;
2744 
2745 #ifdef INVARIANTS
2746 	if (!skip)
2747 		uma_dbg_free(zone, slab, item);
2748 #endif
2749 
2750 	if (keg->uk_flags & UMA_ZONE_REFCNT) {
2751 		slabref = (uma_slabrefcnt_t)slab;
2752 		slabref->us_freelist[freei].us_item = slab->us_firstfree;
2753 	} else {
2754 		slab->us_freelist[freei].us_item = slab->us_firstfree;
2755 	}
2756 	slab->us_firstfree = freei;
2757 	slab->us_freecount++;
2758 
2759 	/* Zone statistics */
2760 	keg->uk_free++;
2761 
2762 	clearfull = 0;
2763 	if (keg->uk_flags & UMA_ZFLAG_FULL) {
2764 		if (keg->uk_pages < keg->uk_maxpages) {
2765 			keg->uk_flags &= ~UMA_ZFLAG_FULL;
2766 			clearfull = 1;
2767 		}
2768 
2769 		/*
2770 		 * We can handle one more allocation. Since we're clearing ZFLAG_FULL,
2771 		 * wake up all procs blocked on pages. This should be uncommon, so
2772 		 * keeping this simple for now (rather than adding count of blocked
2773 		 * threads etc).
2774 		 */
2775 		wakeup(keg);
2776 	}
2777 	if (clearfull) {
2778 		zone_relock(zone, keg);
2779 		zone->uz_flags &= ~UMA_ZFLAG_FULL;
2780 		wakeup(zone);
2781 		ZONE_UNLOCK(zone);
2782 	} else
2783 		KEG_UNLOCK(keg);
2784 }
2785 
2786 /* See uma.h */
2787 void
2788 uma_zone_set_max(uma_zone_t zone, int nitems)
2789 {
2790 	uma_keg_t keg;
2791 
2792 	ZONE_LOCK(zone);
2793 	keg = zone_first_keg(zone);
2794 	keg->uk_maxpages = (nitems / keg->uk_ipers) * keg->uk_ppera;
2795 	if (keg->uk_maxpages * keg->uk_ipers < nitems)
2796 		keg->uk_maxpages += keg->uk_ppera;
2797 
2798 	ZONE_UNLOCK(zone);
2799 }
2800 
2801 /* See uma.h */
2802 void
2803 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
2804 {
2805 	uma_keg_t keg;
2806 
2807 	ZONE_LOCK(zone);
2808 	keg = zone_first_keg(zone);
2809 	KASSERT(keg->uk_pages == 0,
2810 	    ("uma_zone_set_init on non-empty keg"));
2811 	keg->uk_init = uminit;
2812 	ZONE_UNLOCK(zone);
2813 }
2814 
2815 /* See uma.h */
2816 void
2817 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
2818 {
2819 	uma_keg_t keg;
2820 
2821 	ZONE_LOCK(zone);
2822 	keg = zone_first_keg(zone);
2823 	KASSERT(keg->uk_pages == 0,
2824 	    ("uma_zone_set_fini on non-empty keg"));
2825 	keg->uk_fini = fini;
2826 	ZONE_UNLOCK(zone);
2827 }
2828 
2829 /* See uma.h */
2830 void
2831 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
2832 {
2833 	ZONE_LOCK(zone);
2834 	KASSERT(zone_first_keg(zone)->uk_pages == 0,
2835 	    ("uma_zone_set_zinit on non-empty keg"));
2836 	zone->uz_init = zinit;
2837 	ZONE_UNLOCK(zone);
2838 }
2839 
2840 /* See uma.h */
2841 void
2842 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
2843 {
2844 	ZONE_LOCK(zone);
2845 	KASSERT(zone_first_keg(zone)->uk_pages == 0,
2846 	    ("uma_zone_set_zfini on non-empty keg"));
2847 	zone->uz_fini = zfini;
2848 	ZONE_UNLOCK(zone);
2849 }
2850 
2851 /* See uma.h */
2852 /* XXX uk_freef is not actually used with the zone locked */
2853 void
2854 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
2855 {
2856 
2857 	ZONE_LOCK(zone);
2858 	zone_first_keg(zone)->uk_freef = freef;
2859 	ZONE_UNLOCK(zone);
2860 }
2861 
2862 /* See uma.h */
2863 /* XXX uk_allocf is not actually used with the zone locked */
2864 void
2865 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
2866 {
2867 	uma_keg_t keg;
2868 
2869 	ZONE_LOCK(zone);
2870 	keg = zone_first_keg(zone);
2871 	keg->uk_flags |= UMA_ZFLAG_PRIVALLOC;
2872 	keg->uk_allocf = allocf;
2873 	ZONE_UNLOCK(zone);
2874 }
2875 
2876 /* See uma.h */
2877 int
2878 uma_zone_set_obj(uma_zone_t zone, struct vm_object *obj, int count)
2879 {
2880 	uma_keg_t keg;
2881 	vm_offset_t kva;
2882 	int pages;
2883 
2884 	keg = zone_first_keg(zone);
2885 	pages = count / keg->uk_ipers;
2886 
2887 	if (pages * keg->uk_ipers < count)
2888 		pages++;
2889 
2890 	kva = kmem_alloc_nofault(kernel_map, pages * UMA_SLAB_SIZE);
2891 
2892 	if (kva == 0)
2893 		return (0);
2894 	if (obj == NULL) {
2895 		obj = vm_object_allocate(OBJT_DEFAULT,
2896 		    pages);
2897 	} else {
2898 		VM_OBJECT_LOCK_INIT(obj, "uma object");
2899 		_vm_object_allocate(OBJT_DEFAULT,
2900 		    pages, obj);
2901 	}
2902 	ZONE_LOCK(zone);
2903 	keg->uk_kva = kva;
2904 	keg->uk_obj = obj;
2905 	keg->uk_maxpages = pages;
2906 	keg->uk_allocf = obj_alloc;
2907 	keg->uk_flags |= UMA_ZONE_NOFREE | UMA_ZFLAG_PRIVALLOC;
2908 	ZONE_UNLOCK(zone);
2909 	return (1);
2910 }
2911 
2912 /* See uma.h */
2913 void
2914 uma_prealloc(uma_zone_t zone, int items)
2915 {
2916 	int slabs;
2917 	uma_slab_t slab;
2918 	uma_keg_t keg;
2919 
2920 	keg = zone_first_keg(zone);
2921 	ZONE_LOCK(zone);
2922 	slabs = items / keg->uk_ipers;
2923 	if (slabs * keg->uk_ipers < items)
2924 		slabs++;
2925 	while (slabs > 0) {
2926 		slab = keg_alloc_slab(keg, zone, M_WAITOK);
2927 		if (slab == NULL)
2928 			break;
2929 		MPASS(slab->us_keg == keg);
2930 		LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
2931 		slabs--;
2932 	}
2933 	ZONE_UNLOCK(zone);
2934 }
2935 
2936 /* See uma.h */
2937 u_int32_t *
2938 uma_find_refcnt(uma_zone_t zone, void *item)
2939 {
2940 	uma_slabrefcnt_t slabref;
2941 	uma_keg_t keg;
2942 	u_int32_t *refcnt;
2943 	int idx;
2944 
2945 	slabref = (uma_slabrefcnt_t)vtoslab((vm_offset_t)item &
2946 	    (~UMA_SLAB_MASK));
2947 	keg = slabref->us_keg;
2948 	KASSERT(slabref != NULL && slabref->us_keg->uk_flags & UMA_ZONE_REFCNT,
2949 	    ("uma_find_refcnt(): zone possibly not UMA_ZONE_REFCNT"));
2950 	idx = ((unsigned long)item - (unsigned long)slabref->us_data)
2951 	    / keg->uk_rsize;
2952 	refcnt = &slabref->us_freelist[idx].us_refcnt;
2953 	return refcnt;
2954 }
2955 
2956 /* See uma.h */
2957 void
2958 uma_reclaim(void)
2959 {
2960 #ifdef UMA_DEBUG
2961 	printf("UMA: vm asked us to release pages!\n");
2962 #endif
2963 	bucket_enable();
2964 	zone_foreach(zone_drain);
2965 	/*
2966 	 * Some slabs may have been freed but this zone will be visited early
2967 	 * we visit again so that we can free pages that are empty once other
2968 	 * zones are drained.  We have to do the same for buckets.
2969 	 */
2970 	zone_drain(slabzone);
2971 	zone_drain(slabrefzone);
2972 	bucket_zone_drain();
2973 }
2974 
2975 /* See uma.h */
2976 int
2977 uma_zone_exhausted(uma_zone_t zone)
2978 {
2979 	int full;
2980 
2981 	ZONE_LOCK(zone);
2982 	full = (zone->uz_flags & UMA_ZFLAG_FULL);
2983 	ZONE_UNLOCK(zone);
2984 	return (full);
2985 }
2986 
2987 int
2988 uma_zone_exhausted_nolock(uma_zone_t zone)
2989 {
2990 	return (zone->uz_flags & UMA_ZFLAG_FULL);
2991 }
2992 
2993 void *
2994 uma_large_malloc(int size, int wait)
2995 {
2996 	void *mem;
2997 	uma_slab_t slab;
2998 	u_int8_t flags;
2999 
3000 	slab = zone_alloc_item(slabzone, NULL, wait);
3001 	if (slab == NULL)
3002 		return (NULL);
3003 	mem = page_alloc(NULL, size, &flags, wait);
3004 	if (mem) {
3005 		vsetslab((vm_offset_t)mem, slab);
3006 		slab->us_data = mem;
3007 		slab->us_flags = flags | UMA_SLAB_MALLOC;
3008 		slab->us_size = size;
3009 	} else {
3010 		zone_free_item(slabzone, slab, NULL, SKIP_NONE,
3011 		    ZFREE_STATFAIL | ZFREE_STATFREE);
3012 	}
3013 
3014 	return (mem);
3015 }
3016 
3017 void
3018 uma_large_free(uma_slab_t slab)
3019 {
3020 	vsetobj((vm_offset_t)slab->us_data, kmem_object);
3021 	page_free(slab->us_data, slab->us_size, slab->us_flags);
3022 	zone_free_item(slabzone, slab, NULL, SKIP_NONE, ZFREE_STATFREE);
3023 }
3024 
3025 void
3026 uma_print_stats(void)
3027 {
3028 	zone_foreach(uma_print_zone);
3029 }
3030 
3031 static void
3032 slab_print(uma_slab_t slab)
3033 {
3034 	printf("slab: keg %p, data %p, freecount %d, firstfree %d\n",
3035 		slab->us_keg, slab->us_data, slab->us_freecount,
3036 		slab->us_firstfree);
3037 }
3038 
3039 static void
3040 cache_print(uma_cache_t cache)
3041 {
3042 	printf("alloc: %p(%d), free: %p(%d)\n",
3043 		cache->uc_allocbucket,
3044 		cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0,
3045 		cache->uc_freebucket,
3046 		cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0);
3047 }
3048 
3049 static void
3050 uma_print_keg(uma_keg_t keg)
3051 {
3052 	uma_slab_t slab;
3053 
3054 	printf("keg: %s(%p) size %d(%d) flags %d ipers %d ppera %d "
3055 	    "out %d free %d limit %d\n",
3056 	    keg->uk_name, keg, keg->uk_size, keg->uk_rsize, keg->uk_flags,
3057 	    keg->uk_ipers, keg->uk_ppera,
3058 	    (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free,
3059 	    (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers);
3060 	printf("Part slabs:\n");
3061 	LIST_FOREACH(slab, &keg->uk_part_slab, us_link)
3062 		slab_print(slab);
3063 	printf("Free slabs:\n");
3064 	LIST_FOREACH(slab, &keg->uk_free_slab, us_link)
3065 		slab_print(slab);
3066 	printf("Full slabs:\n");
3067 	LIST_FOREACH(slab, &keg->uk_full_slab, us_link)
3068 		slab_print(slab);
3069 }
3070 
3071 void
3072 uma_print_zone(uma_zone_t zone)
3073 {
3074 	uma_cache_t cache;
3075 	uma_klink_t kl;
3076 	int i;
3077 
3078 	printf("zone: %s(%p) size %d flags %d\n",
3079 	    zone->uz_name, zone, zone->uz_size, zone->uz_flags);
3080 	LIST_FOREACH(kl, &zone->uz_kegs, kl_link)
3081 		uma_print_keg(kl->kl_keg);
3082 	for (i = 0; i <= mp_maxid; i++) {
3083 		if (CPU_ABSENT(i))
3084 			continue;
3085 		cache = &zone->uz_cpu[i];
3086 		printf("CPU %d Cache:\n", i);
3087 		cache_print(cache);
3088 	}
3089 }
3090 
3091 #ifdef DDB
3092 /*
3093  * Generate statistics across both the zone and its per-cpu cache's.  Return
3094  * desired statistics if the pointer is non-NULL for that statistic.
3095  *
3096  * Note: does not update the zone statistics, as it can't safely clear the
3097  * per-CPU cache statistic.
3098  *
3099  * XXXRW: Following the uc_allocbucket and uc_freebucket pointers here isn't
3100  * safe from off-CPU; we should modify the caches to track this information
3101  * directly so that we don't have to.
3102  */
3103 static void
3104 uma_zone_sumstat(uma_zone_t z, int *cachefreep, u_int64_t *allocsp,
3105     u_int64_t *freesp)
3106 {
3107 	uma_cache_t cache;
3108 	u_int64_t allocs, frees;
3109 	int cachefree, cpu;
3110 
3111 	allocs = frees = 0;
3112 	cachefree = 0;
3113 	for (cpu = 0; cpu <= mp_maxid; cpu++) {
3114 		if (CPU_ABSENT(cpu))
3115 			continue;
3116 		cache = &z->uz_cpu[cpu];
3117 		if (cache->uc_allocbucket != NULL)
3118 			cachefree += cache->uc_allocbucket->ub_cnt;
3119 		if (cache->uc_freebucket != NULL)
3120 			cachefree += cache->uc_freebucket->ub_cnt;
3121 		allocs += cache->uc_allocs;
3122 		frees += cache->uc_frees;
3123 	}
3124 	allocs += z->uz_allocs;
3125 	frees += z->uz_frees;
3126 	if (cachefreep != NULL)
3127 		*cachefreep = cachefree;
3128 	if (allocsp != NULL)
3129 		*allocsp = allocs;
3130 	if (freesp != NULL)
3131 		*freesp = frees;
3132 }
3133 #endif /* DDB */
3134 
3135 static int
3136 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
3137 {
3138 	uma_keg_t kz;
3139 	uma_zone_t z;
3140 	int count;
3141 
3142 	count = 0;
3143 	mtx_lock(&uma_mtx);
3144 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3145 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
3146 			count++;
3147 	}
3148 	mtx_unlock(&uma_mtx);
3149 	return (sysctl_handle_int(oidp, &count, 0, req));
3150 }
3151 
3152 static int
3153 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
3154 {
3155 	struct uma_stream_header ush;
3156 	struct uma_type_header uth;
3157 	struct uma_percpu_stat ups;
3158 	uma_bucket_t bucket;
3159 	struct sbuf sbuf;
3160 	uma_cache_t cache;
3161 	uma_klink_t kl;
3162 	uma_keg_t kz;
3163 	uma_zone_t z;
3164 	uma_keg_t k;
3165 	char *buffer;
3166 	int buflen, count, error, i;
3167 
3168 	mtx_lock(&uma_mtx);
3169 restart:
3170 	mtx_assert(&uma_mtx, MA_OWNED);
3171 	count = 0;
3172 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3173 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
3174 			count++;
3175 	}
3176 	mtx_unlock(&uma_mtx);
3177 
3178 	buflen = sizeof(ush) + count * (sizeof(uth) + sizeof(ups) *
3179 	    (mp_maxid + 1)) + 1;
3180 	buffer = malloc(buflen, M_TEMP, M_WAITOK | M_ZERO);
3181 
3182 	mtx_lock(&uma_mtx);
3183 	i = 0;
3184 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3185 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
3186 			i++;
3187 	}
3188 	if (i > count) {
3189 		free(buffer, M_TEMP);
3190 		goto restart;
3191 	}
3192 	count =  i;
3193 
3194 	sbuf_new(&sbuf, buffer, buflen, SBUF_FIXEDLEN);
3195 
3196 	/*
3197 	 * Insert stream header.
3198 	 */
3199 	bzero(&ush, sizeof(ush));
3200 	ush.ush_version = UMA_STREAM_VERSION;
3201 	ush.ush_maxcpus = (mp_maxid + 1);
3202 	ush.ush_count = count;
3203 	if (sbuf_bcat(&sbuf, &ush, sizeof(ush)) < 0) {
3204 		mtx_unlock(&uma_mtx);
3205 		error = ENOMEM;
3206 		goto out;
3207 	}
3208 
3209 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3210 		LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3211 			bzero(&uth, sizeof(uth));
3212 			ZONE_LOCK(z);
3213 			strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
3214 			uth.uth_align = kz->uk_align;
3215 			uth.uth_size = kz->uk_size;
3216 			uth.uth_rsize = kz->uk_rsize;
3217 			LIST_FOREACH(kl, &z->uz_kegs, kl_link) {
3218 				k = kl->kl_keg;
3219 				uth.uth_maxpages += k->uk_maxpages;
3220 				uth.uth_pages += k->uk_pages;
3221 				uth.uth_keg_free += k->uk_free;
3222 				uth.uth_limit = (k->uk_maxpages / k->uk_ppera)
3223 				    * k->uk_ipers;
3224 			}
3225 
3226 			/*
3227 			 * A zone is secondary is it is not the first entry
3228 			 * on the keg's zone list.
3229 			 */
3230 			if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
3231 			    (LIST_FIRST(&kz->uk_zones) != z))
3232 				uth.uth_zone_flags = UTH_ZONE_SECONDARY;
3233 
3234 			LIST_FOREACH(bucket, &z->uz_full_bucket, ub_link)
3235 				uth.uth_zone_free += bucket->ub_cnt;
3236 			uth.uth_allocs = z->uz_allocs;
3237 			uth.uth_frees = z->uz_frees;
3238 			uth.uth_fails = z->uz_fails;
3239 			if (sbuf_bcat(&sbuf, &uth, sizeof(uth)) < 0) {
3240 				ZONE_UNLOCK(z);
3241 				mtx_unlock(&uma_mtx);
3242 				error = ENOMEM;
3243 				goto out;
3244 			}
3245 			/*
3246 			 * While it is not normally safe to access the cache
3247 			 * bucket pointers while not on the CPU that owns the
3248 			 * cache, we only allow the pointers to be exchanged
3249 			 * without the zone lock held, not invalidated, so
3250 			 * accept the possible race associated with bucket
3251 			 * exchange during monitoring.
3252 			 */
3253 			for (i = 0; i < (mp_maxid + 1); i++) {
3254 				bzero(&ups, sizeof(ups));
3255 				if (kz->uk_flags & UMA_ZFLAG_INTERNAL)
3256 					goto skip;
3257 				if (CPU_ABSENT(i))
3258 					goto skip;
3259 				cache = &z->uz_cpu[i];
3260 				if (cache->uc_allocbucket != NULL)
3261 					ups.ups_cache_free +=
3262 					    cache->uc_allocbucket->ub_cnt;
3263 				if (cache->uc_freebucket != NULL)
3264 					ups.ups_cache_free +=
3265 					    cache->uc_freebucket->ub_cnt;
3266 				ups.ups_allocs = cache->uc_allocs;
3267 				ups.ups_frees = cache->uc_frees;
3268 skip:
3269 				if (sbuf_bcat(&sbuf, &ups, sizeof(ups)) < 0) {
3270 					ZONE_UNLOCK(z);
3271 					mtx_unlock(&uma_mtx);
3272 					error = ENOMEM;
3273 					goto out;
3274 				}
3275 			}
3276 			ZONE_UNLOCK(z);
3277 		}
3278 	}
3279 	mtx_unlock(&uma_mtx);
3280 	sbuf_finish(&sbuf);
3281 	error = SYSCTL_OUT(req, sbuf_data(&sbuf), sbuf_len(&sbuf));
3282 out:
3283 	free(buffer, M_TEMP);
3284 	return (error);
3285 }
3286 
3287 #ifdef DDB
3288 DB_SHOW_COMMAND(uma, db_show_uma)
3289 {
3290 	u_int64_t allocs, frees;
3291 	uma_bucket_t bucket;
3292 	uma_keg_t kz;
3293 	uma_zone_t z;
3294 	int cachefree;
3295 
3296 	db_printf("%18s %8s %8s %8s %12s\n", "Zone", "Size", "Used", "Free",
3297 	    "Requests");
3298 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3299 		LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3300 			if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
3301 				allocs = z->uz_allocs;
3302 				frees = z->uz_frees;
3303 				cachefree = 0;
3304 			} else
3305 				uma_zone_sumstat(z, &cachefree, &allocs,
3306 				    &frees);
3307 			if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
3308 			    (LIST_FIRST(&kz->uk_zones) != z)))
3309 				cachefree += kz->uk_free;
3310 			LIST_FOREACH(bucket, &z->uz_full_bucket, ub_link)
3311 				cachefree += bucket->ub_cnt;
3312 			db_printf("%18s %8ju %8jd %8d %12ju\n", z->uz_name,
3313 			    (uintmax_t)kz->uk_size,
3314 			    (intmax_t)(allocs - frees), cachefree,
3315 			    (uintmax_t)allocs);
3316 		}
3317 	}
3318 }
3319 #endif
3320