xref: /freebsd/sys/vm/uma_core.c (revision 4ef0123eb70fb892d6b6b511c67e94abb3301c96)
1 /*-
2  * Copyright (c) 2002-2005, 2009, 2013 Jeffrey Roberson <jeff@FreeBSD.org>
3  * Copyright (c) 2004, 2005 Bosko Milekic <bmilekic@FreeBSD.org>
4  * Copyright (c) 2004-2006 Robert N. M. Watson
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice unmodified, this list of conditions, and the following
12  *    disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
18  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
19  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
20  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
21  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
22  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
23  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
24  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
25  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
26  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
27  */
28 
29 /*
30  * uma_core.c  Implementation of the Universal Memory allocator
31  *
32  * This allocator is intended to replace the multitude of similar object caches
33  * in the standard FreeBSD kernel.  The intent is to be flexible as well as
34  * effecient.  A primary design goal is to return unused memory to the rest of
35  * the system.  This will make the system as a whole more flexible due to the
36  * ability to move memory to subsystems which most need it instead of leaving
37  * pools of reserved memory unused.
38  *
39  * The basic ideas stem from similar slab/zone based allocators whose algorithms
40  * are well known.
41  *
42  */
43 
44 /*
45  * TODO:
46  *	- Improve memory usage for large allocations
47  *	- Investigate cache size adjustments
48  */
49 
50 #include <sys/cdefs.h>
51 __FBSDID("$FreeBSD$");
52 
53 /* I should really use ktr.. */
54 /*
55 #define UMA_DEBUG 1
56 #define UMA_DEBUG_ALLOC 1
57 #define UMA_DEBUG_ALLOC_1 1
58 */
59 
60 #include "opt_ddb.h"
61 #include "opt_param.h"
62 #include "opt_vm.h"
63 
64 #include <sys/param.h>
65 #include <sys/systm.h>
66 #include <sys/bitset.h>
67 #include <sys/kernel.h>
68 #include <sys/types.h>
69 #include <sys/queue.h>
70 #include <sys/malloc.h>
71 #include <sys/ktr.h>
72 #include <sys/lock.h>
73 #include <sys/sysctl.h>
74 #include <sys/mutex.h>
75 #include <sys/proc.h>
76 #include <sys/rwlock.h>
77 #include <sys/sbuf.h>
78 #include <sys/smp.h>
79 #include <sys/vmmeter.h>
80 
81 #include <vm/vm.h>
82 #include <vm/vm_object.h>
83 #include <vm/vm_page.h>
84 #include <vm/vm_pageout.h>
85 #include <vm/vm_param.h>
86 #include <vm/vm_map.h>
87 #include <vm/vm_kern.h>
88 #include <vm/vm_extern.h>
89 #include <vm/uma.h>
90 #include <vm/uma_int.h>
91 #include <vm/uma_dbg.h>
92 
93 #include <ddb/ddb.h>
94 
95 #ifdef DEBUG_MEMGUARD
96 #include <vm/memguard.h>
97 #endif
98 
99 /*
100  * This is the zone and keg from which all zones are spawned.  The idea is that
101  * even the zone & keg heads are allocated from the allocator, so we use the
102  * bss section to bootstrap us.
103  */
104 static struct uma_keg masterkeg;
105 static struct uma_zone masterzone_k;
106 static struct uma_zone masterzone_z;
107 static uma_zone_t kegs = &masterzone_k;
108 static uma_zone_t zones = &masterzone_z;
109 
110 /* This is the zone from which all of uma_slab_t's are allocated. */
111 static uma_zone_t slabzone;
112 static uma_zone_t slabrefzone;	/* With refcounters (for UMA_ZONE_REFCNT) */
113 
114 /*
115  * The initial hash tables come out of this zone so they can be allocated
116  * prior to malloc coming up.
117  */
118 static uma_zone_t hashzone;
119 
120 /* The boot-time adjusted value for cache line alignment. */
121 int uma_align_cache = 64 - 1;
122 
123 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets");
124 
125 /*
126  * Are we allowed to allocate buckets?
127  */
128 static int bucketdisable = 1;
129 
130 /* Linked list of all kegs in the system */
131 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs);
132 
133 /* This mutex protects the keg list */
134 static struct mtx_padalign uma_mtx;
135 
136 /* Linked list of boot time pages */
137 static LIST_HEAD(,uma_slab) uma_boot_pages =
138     LIST_HEAD_INITIALIZER(uma_boot_pages);
139 
140 /* This mutex protects the boot time pages list */
141 static struct mtx_padalign uma_boot_pages_mtx;
142 
143 /* Is the VM done starting up? */
144 static int booted = 0;
145 #define	UMA_STARTUP	1
146 #define	UMA_STARTUP2	2
147 
148 /* Maximum number of allowed items-per-slab if the slab header is OFFPAGE */
149 static const u_int uma_max_ipers = SLAB_SETSIZE;
150 
151 /*
152  * Only mbuf clusters use ref zones.  Just provide enough references
153  * to support the one user.  New code should not use the ref facility.
154  */
155 static const u_int uma_max_ipers_ref = PAGE_SIZE / MCLBYTES;
156 
157 /*
158  * This is the handle used to schedule events that need to happen
159  * outside of the allocation fast path.
160  */
161 static struct callout uma_callout;
162 #define	UMA_TIMEOUT	20		/* Seconds for callout interval. */
163 
164 /*
165  * This structure is passed as the zone ctor arg so that I don't have to create
166  * a special allocation function just for zones.
167  */
168 struct uma_zctor_args {
169 	const char *name;
170 	size_t size;
171 	uma_ctor ctor;
172 	uma_dtor dtor;
173 	uma_init uminit;
174 	uma_fini fini;
175 	uma_import import;
176 	uma_release release;
177 	void *arg;
178 	uma_keg_t keg;
179 	int align;
180 	uint32_t flags;
181 };
182 
183 struct uma_kctor_args {
184 	uma_zone_t zone;
185 	size_t size;
186 	uma_init uminit;
187 	uma_fini fini;
188 	int align;
189 	uint32_t flags;
190 };
191 
192 struct uma_bucket_zone {
193 	uma_zone_t	ubz_zone;
194 	char		*ubz_name;
195 	int		ubz_entries;	/* Number of items it can hold. */
196 	int		ubz_maxsize;	/* Maximum allocation size per-item. */
197 };
198 
199 /*
200  * Compute the actual number of bucket entries to pack them in power
201  * of two sizes for more efficient space utilization.
202  */
203 #define	BUCKET_SIZE(n)						\
204     (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *))
205 
206 #define	BUCKET_MAX	BUCKET_SIZE(128)
207 
208 struct uma_bucket_zone bucket_zones[] = {
209 	{ NULL, "32 Bucket", BUCKET_SIZE(32), 512 },
210 	{ NULL, "64 Bucket", BUCKET_SIZE(64), 256 },
211 	{ NULL, "128 Bucket", BUCKET_SIZE(128), 128 },
212 	{ NULL, NULL, 0}
213 };
214 static uma_zone_t largebucket;
215 
216 /*
217  * Flags and enumerations to be passed to internal functions.
218  */
219 enum zfreeskip { SKIP_NONE = 0, SKIP_DTOR, SKIP_FINI };
220 
221 /* Prototypes.. */
222 
223 static void *noobj_alloc(uma_zone_t, int, uint8_t *, int);
224 static void *page_alloc(uma_zone_t, int, uint8_t *, int);
225 static void *startup_alloc(uma_zone_t, int, uint8_t *, int);
226 static void page_free(void *, int, uint8_t);
227 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int);
228 static void cache_drain(uma_zone_t);
229 static void bucket_drain(uma_zone_t, uma_bucket_t);
230 static void bucket_cache_drain(uma_zone_t zone);
231 static int keg_ctor(void *, int, void *, int);
232 static void keg_dtor(void *, int, void *);
233 static int zone_ctor(void *, int, void *, int);
234 static void zone_dtor(void *, int, void *);
235 static int zero_init(void *, int, int);
236 static void keg_small_init(uma_keg_t keg);
237 static void keg_large_init(uma_keg_t keg);
238 static void zone_foreach(void (*zfunc)(uma_zone_t));
239 static void zone_timeout(uma_zone_t zone);
240 static int hash_alloc(struct uma_hash *);
241 static int hash_expand(struct uma_hash *, struct uma_hash *);
242 static void hash_free(struct uma_hash *hash);
243 static void uma_timeout(void *);
244 static void uma_startup3(void);
245 static void *zone_alloc_item(uma_zone_t, void *, int);
246 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip);
247 static void bucket_enable(void);
248 static void bucket_init(void);
249 static uma_bucket_t bucket_alloc(uma_zone_t zone, int);
250 static void bucket_free(uma_zone_t zone, uma_bucket_t);
251 static void bucket_zone_drain(void);
252 static uma_bucket_t zone_alloc_bucket(uma_zone_t zone, int flags);
253 static uma_slab_t zone_fetch_slab(uma_zone_t zone, uma_keg_t last, int flags);
254 static uma_slab_t zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int flags);
255 static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab);
256 static void slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item);
257 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit,
258     uma_fini fini, int align, uint32_t flags);
259 static int zone_import(uma_zone_t zone, void **bucket, int max, int flags);
260 static void zone_release(uma_zone_t zone, void **bucket, int cnt);
261 
262 void uma_print_zone(uma_zone_t);
263 void uma_print_stats(void);
264 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS);
265 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS);
266 
267 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
268 
269 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLTYPE_INT,
270     0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones");
271 
272 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLTYPE_STRUCT,
273     0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats");
274 
275 static int zone_warnings = 1;
276 TUNABLE_INT("vm.zone_warnings", &zone_warnings);
277 SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RW, &zone_warnings, 0,
278     "Warn when UMA zones becomes full");
279 
280 /*
281  * This routine checks to see whether or not it's safe to enable buckets.
282  */
283 static void
284 bucket_enable(void)
285 {
286 	bucketdisable = vm_page_count_min();
287 }
288 
289 /*
290  * Initialize bucket_zones, the array of zones of buckets of various sizes.
291  *
292  * For each zone, calculate the memory required for each bucket, consisting
293  * of the header and an array of pointers.
294  */
295 static void
296 bucket_init(void)
297 {
298 	struct uma_bucket_zone *ubz;
299 	int size;
300 	int i;
301 
302 	for (i = 0, ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) {
303 		size = roundup(sizeof(struct uma_bucket), sizeof(void *));
304 		size += sizeof(void *) * ubz->ubz_entries;
305 		ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
306 		    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
307 		    UMA_ZONE_MAXBUCKET | UMA_ZONE_MTXCLASS);
308 	}
309 	/*
310 	 * To avoid recursive bucket allocation loops we disable buckets
311 	 * on the smallest bucket zone and use it for the largest zone.
312 	 * The remainder of the zones all use the largest zone.
313 	 */
314 	ubz--;
315 	ubz->ubz_zone->uz_count = bucket_zones[0].ubz_entries;
316 	bucket_zones[0].ubz_zone->uz_count = 0;
317 	largebucket = ubz->ubz_zone;
318 }
319 
320 /*
321  * Given a desired number of entries for a bucket, return the zone from which
322  * to allocate the bucket.
323  */
324 static struct uma_bucket_zone *
325 bucket_zone_lookup(int entries)
326 {
327 	struct uma_bucket_zone *ubz;
328 
329 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
330 		if (ubz->ubz_entries >= entries)
331 			return (ubz);
332 	ubz--;
333 	return (ubz);
334 }
335 
336 static int
337 bucket_select(int size)
338 {
339 	struct uma_bucket_zone *ubz;
340 
341 	ubz = &bucket_zones[0];
342 	if (size > ubz->ubz_maxsize)
343 		return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1);
344 
345 	for (; ubz->ubz_entries != 0; ubz++)
346 		if (ubz->ubz_maxsize < size)
347 			break;
348 	ubz--;
349 	return (ubz->ubz_entries);
350 }
351 
352 static uma_bucket_t
353 bucket_alloc(uma_zone_t zone, int flags)
354 {
355 	struct uma_bucket_zone *ubz;
356 	uma_bucket_t bucket;
357 
358 	/*
359 	 * This is to stop us from allocating per cpu buckets while we're
360 	 * running out of vm.boot_pages.  Otherwise, we would exhaust the
361 	 * boot pages.  This also prevents us from allocating buckets in
362 	 * low memory situations.
363 	 */
364 	if (bucketdisable)
365 		return (NULL);
366 
367 	if (zone->uz_flags & UMA_ZFLAG_CACHEONLY)
368 		flags |= M_NOVM;
369 	ubz = bucket_zone_lookup(zone->uz_count);
370 	bucket = uma_zalloc(ubz->ubz_zone, flags);
371 	if (bucket) {
372 #ifdef INVARIANTS
373 		bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
374 #endif
375 		bucket->ub_cnt = 0;
376 		bucket->ub_entries = ubz->ubz_entries;
377 	}
378 
379 	return (bucket);
380 }
381 
382 static void
383 bucket_free(uma_zone_t zone, uma_bucket_t bucket)
384 {
385 	struct uma_bucket_zone *ubz;
386 
387 	KASSERT(bucket->ub_cnt == 0,
388 	    ("bucket_free: Freeing a non free bucket."));
389 	ubz = bucket_zone_lookup(bucket->ub_entries);
390 	uma_zfree(ubz->ubz_zone, bucket);
391 }
392 
393 static void
394 bucket_zone_drain(void)
395 {
396 	struct uma_bucket_zone *ubz;
397 
398 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
399 		zone_drain(ubz->ubz_zone);
400 }
401 
402 static void
403 zone_log_warning(uma_zone_t zone)
404 {
405 	static const struct timeval warninterval = { 300, 0 };
406 
407 	if (!zone_warnings || zone->uz_warning == NULL)
408 		return;
409 
410 	if (ratecheck(&zone->uz_ratecheck, &warninterval))
411 		printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning);
412 }
413 
414 static void
415 zone_foreach_keg(uma_zone_t zone, void (*kegfn)(uma_keg_t))
416 {
417 	uma_klink_t klink;
418 
419 	LIST_FOREACH(klink, &zone->uz_kegs, kl_link)
420 		kegfn(klink->kl_keg);
421 }
422 
423 /*
424  * Routine called by timeout which is used to fire off some time interval
425  * based calculations.  (stats, hash size, etc.)
426  *
427  * Arguments:
428  *	arg   Unused
429  *
430  * Returns:
431  *	Nothing
432  */
433 static void
434 uma_timeout(void *unused)
435 {
436 	bucket_enable();
437 	zone_foreach(zone_timeout);
438 
439 	/* Reschedule this event */
440 	callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
441 }
442 
443 /*
444  * Routine to perform timeout driven calculations.  This expands the
445  * hashes and does per cpu statistics aggregation.
446  *
447  *  Returns nothing.
448  */
449 static void
450 keg_timeout(uma_keg_t keg)
451 {
452 
453 	KEG_LOCK(keg);
454 	/*
455 	 * Expand the keg hash table.
456 	 *
457 	 * This is done if the number of slabs is larger than the hash size.
458 	 * What I'm trying to do here is completely reduce collisions.  This
459 	 * may be a little aggressive.  Should I allow for two collisions max?
460 	 */
461 	if (keg->uk_flags & UMA_ZONE_HASH &&
462 	    keg->uk_pages / keg->uk_ppera >= keg->uk_hash.uh_hashsize) {
463 		struct uma_hash newhash;
464 		struct uma_hash oldhash;
465 		int ret;
466 
467 		/*
468 		 * This is so involved because allocating and freeing
469 		 * while the keg lock is held will lead to deadlock.
470 		 * I have to do everything in stages and check for
471 		 * races.
472 		 */
473 		newhash = keg->uk_hash;
474 		KEG_UNLOCK(keg);
475 		ret = hash_alloc(&newhash);
476 		KEG_LOCK(keg);
477 		if (ret) {
478 			if (hash_expand(&keg->uk_hash, &newhash)) {
479 				oldhash = keg->uk_hash;
480 				keg->uk_hash = newhash;
481 			} else
482 				oldhash = newhash;
483 
484 			KEG_UNLOCK(keg);
485 			hash_free(&oldhash);
486 			KEG_LOCK(keg);
487 		}
488 	}
489 	KEG_UNLOCK(keg);
490 }
491 
492 static void
493 zone_timeout(uma_zone_t zone)
494 {
495 
496 	zone_foreach_keg(zone, &keg_timeout);
497 }
498 
499 /*
500  * Allocate and zero fill the next sized hash table from the appropriate
501  * backing store.
502  *
503  * Arguments:
504  *	hash  A new hash structure with the old hash size in uh_hashsize
505  *
506  * Returns:
507  *	1 on sucess and 0 on failure.
508  */
509 static int
510 hash_alloc(struct uma_hash *hash)
511 {
512 	int oldsize;
513 	int alloc;
514 
515 	oldsize = hash->uh_hashsize;
516 
517 	/* We're just going to go to a power of two greater */
518 	if (oldsize)  {
519 		hash->uh_hashsize = oldsize * 2;
520 		alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
521 		hash->uh_slab_hash = (struct slabhead *)malloc(alloc,
522 		    M_UMAHASH, M_NOWAIT);
523 	} else {
524 		alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
525 		hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
526 		    M_WAITOK);
527 		hash->uh_hashsize = UMA_HASH_SIZE_INIT;
528 	}
529 	if (hash->uh_slab_hash) {
530 		bzero(hash->uh_slab_hash, alloc);
531 		hash->uh_hashmask = hash->uh_hashsize - 1;
532 		return (1);
533 	}
534 
535 	return (0);
536 }
537 
538 /*
539  * Expands the hash table for HASH zones.  This is done from zone_timeout
540  * to reduce collisions.  This must not be done in the regular allocation
541  * path, otherwise, we can recurse on the vm while allocating pages.
542  *
543  * Arguments:
544  *	oldhash  The hash you want to expand
545  *	newhash  The hash structure for the new table
546  *
547  * Returns:
548  *	Nothing
549  *
550  * Discussion:
551  */
552 static int
553 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
554 {
555 	uma_slab_t slab;
556 	int hval;
557 	int i;
558 
559 	if (!newhash->uh_slab_hash)
560 		return (0);
561 
562 	if (oldhash->uh_hashsize >= newhash->uh_hashsize)
563 		return (0);
564 
565 	/*
566 	 * I need to investigate hash algorithms for resizing without a
567 	 * full rehash.
568 	 */
569 
570 	for (i = 0; i < oldhash->uh_hashsize; i++)
571 		while (!SLIST_EMPTY(&oldhash->uh_slab_hash[i])) {
572 			slab = SLIST_FIRST(&oldhash->uh_slab_hash[i]);
573 			SLIST_REMOVE_HEAD(&oldhash->uh_slab_hash[i], us_hlink);
574 			hval = UMA_HASH(newhash, slab->us_data);
575 			SLIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
576 			    slab, us_hlink);
577 		}
578 
579 	return (1);
580 }
581 
582 /*
583  * Free the hash bucket to the appropriate backing store.
584  *
585  * Arguments:
586  *	slab_hash  The hash bucket we're freeing
587  *	hashsize   The number of entries in that hash bucket
588  *
589  * Returns:
590  *	Nothing
591  */
592 static void
593 hash_free(struct uma_hash *hash)
594 {
595 	if (hash->uh_slab_hash == NULL)
596 		return;
597 	if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
598 		zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE);
599 	else
600 		free(hash->uh_slab_hash, M_UMAHASH);
601 }
602 
603 /*
604  * Frees all outstanding items in a bucket
605  *
606  * Arguments:
607  *	zone   The zone to free to, must be unlocked.
608  *	bucket The free/alloc bucket with items, cpu queue must be locked.
609  *
610  * Returns:
611  *	Nothing
612  */
613 
614 static void
615 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
616 {
617 	int i;
618 
619 	if (bucket == NULL)
620 		return;
621 
622 	if (zone->uz_fini)
623 		for (i = 0; i < bucket->ub_cnt; i++)
624 			zone->uz_fini(bucket->ub_bucket[i], zone->uz_size);
625 	zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt);
626 	bucket->ub_cnt = 0;
627 }
628 
629 /*
630  * Drains the per cpu caches for a zone.
631  *
632  * NOTE: This may only be called while the zone is being turn down, and not
633  * during normal operation.  This is necessary in order that we do not have
634  * to migrate CPUs to drain the per-CPU caches.
635  *
636  * Arguments:
637  *	zone     The zone to drain, must be unlocked.
638  *
639  * Returns:
640  *	Nothing
641  */
642 static void
643 cache_drain(uma_zone_t zone)
644 {
645 	uma_cache_t cache;
646 	int cpu;
647 
648 	/*
649 	 * XXX: It is safe to not lock the per-CPU caches, because we're
650 	 * tearing down the zone anyway.  I.e., there will be no further use
651 	 * of the caches at this point.
652 	 *
653 	 * XXX: It would good to be able to assert that the zone is being
654 	 * torn down to prevent improper use of cache_drain().
655 	 *
656 	 * XXX: We lock the zone before passing into bucket_cache_drain() as
657 	 * it is used elsewhere.  Should the tear-down path be made special
658 	 * there in some form?
659 	 */
660 	CPU_FOREACH(cpu) {
661 		cache = &zone->uz_cpu[cpu];
662 		bucket_drain(zone, cache->uc_allocbucket);
663 		bucket_drain(zone, cache->uc_freebucket);
664 		if (cache->uc_allocbucket != NULL)
665 			bucket_free(zone, cache->uc_allocbucket);
666 		if (cache->uc_freebucket != NULL)
667 			bucket_free(zone, cache->uc_freebucket);
668 		cache->uc_allocbucket = cache->uc_freebucket = NULL;
669 	}
670 	ZONE_LOCK(zone);
671 	bucket_cache_drain(zone);
672 	ZONE_UNLOCK(zone);
673 }
674 
675 /*
676  * Drain the cached buckets from a zone.  Expects a locked zone on entry.
677  */
678 static void
679 bucket_cache_drain(uma_zone_t zone)
680 {
681 	uma_bucket_t bucket;
682 
683 	/*
684 	 * Drain the bucket queues and free the buckets, we just keep two per
685 	 * cpu (alloc/free).
686 	 */
687 	while ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) {
688 		LIST_REMOVE(bucket, ub_link);
689 		ZONE_UNLOCK(zone);
690 		bucket_drain(zone, bucket);
691 		bucket_free(zone, bucket);
692 		ZONE_LOCK(zone);
693 	}
694 }
695 
696 static void
697 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start)
698 {
699 	uint8_t *mem;
700 	int i;
701 	uint8_t flags;
702 
703 	mem = slab->us_data;
704 	flags = slab->us_flags;
705 	i = start;
706 	if (keg->uk_fini != NULL) {
707 		for (i--; i > -1; i--)
708 			keg->uk_fini(slab->us_data + (keg->uk_rsize * i),
709 			    keg->uk_size);
710 	}
711 	if (keg->uk_flags & UMA_ZONE_VTOSLAB) {
712 		vm_object_t obj;
713 
714 		if (flags & UMA_SLAB_KMEM)
715 			obj = kmem_object;
716 		else if (flags & UMA_SLAB_KERNEL)
717 			obj = kernel_object;
718 		else
719 			obj = NULL;
720 		for (i = 0; i < keg->uk_ppera; i++)
721 			vsetobj((vm_offset_t)mem + (i * PAGE_SIZE), obj);
722 	}
723 	if (keg->uk_flags & UMA_ZONE_OFFPAGE)
724 		zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE);
725 #ifdef UMA_DEBUG
726 	printf("%s: Returning %d bytes.\n", keg->uk_name,
727 	    PAGE_SIZE * keg->uk_ppera);
728 #endif
729 	keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags);
730 }
731 
732 /*
733  * Frees pages from a keg back to the system.  This is done on demand from
734  * the pageout daemon.
735  *
736  * Returns nothing.
737  */
738 static void
739 keg_drain(uma_keg_t keg)
740 {
741 	struct slabhead freeslabs = { 0 };
742 	uma_slab_t slab;
743 	uma_slab_t n;
744 
745 	/*
746 	 * We don't want to take pages from statically allocated kegs at this
747 	 * time
748 	 */
749 	if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL)
750 		return;
751 
752 #ifdef UMA_DEBUG
753 	printf("%s free items: %u\n", keg->uk_name, keg->uk_free);
754 #endif
755 	KEG_LOCK(keg);
756 	if (keg->uk_free == 0)
757 		goto finished;
758 
759 	slab = LIST_FIRST(&keg->uk_free_slab);
760 	while (slab) {
761 		n = LIST_NEXT(slab, us_link);
762 
763 		/* We have no where to free these to */
764 		if (slab->us_flags & UMA_SLAB_BOOT) {
765 			slab = n;
766 			continue;
767 		}
768 
769 		LIST_REMOVE(slab, us_link);
770 		keg->uk_pages -= keg->uk_ppera;
771 		keg->uk_free -= keg->uk_ipers;
772 
773 		if (keg->uk_flags & UMA_ZONE_HASH)
774 			UMA_HASH_REMOVE(&keg->uk_hash, slab, slab->us_data);
775 
776 		SLIST_INSERT_HEAD(&freeslabs, slab, us_hlink);
777 
778 		slab = n;
779 	}
780 finished:
781 	KEG_UNLOCK(keg);
782 
783 	while ((slab = SLIST_FIRST(&freeslabs)) != NULL) {
784 		SLIST_REMOVE(&freeslabs, slab, uma_slab, us_hlink);
785 		keg_free_slab(keg, slab, 0);
786 	}
787 }
788 
789 static void
790 zone_drain_wait(uma_zone_t zone, int waitok)
791 {
792 
793 	/*
794 	 * Set draining to interlock with zone_dtor() so we can release our
795 	 * locks as we go.  Only dtor() should do a WAITOK call since it
796 	 * is the only call that knows the structure will still be available
797 	 * when it wakes up.
798 	 */
799 	ZONE_LOCK(zone);
800 	while (zone->uz_flags & UMA_ZFLAG_DRAINING) {
801 		if (waitok == M_NOWAIT)
802 			goto out;
803 		mtx_unlock(&uma_mtx);
804 		msleep(zone, zone->uz_lockptr, PVM, "zonedrain", 1);
805 		mtx_lock(&uma_mtx);
806 	}
807 	zone->uz_flags |= UMA_ZFLAG_DRAINING;
808 	bucket_cache_drain(zone);
809 	ZONE_UNLOCK(zone);
810 	/*
811 	 * The DRAINING flag protects us from being freed while
812 	 * we're running.  Normally the uma_mtx would protect us but we
813 	 * must be able to release and acquire the right lock for each keg.
814 	 */
815 	zone_foreach_keg(zone, &keg_drain);
816 	ZONE_LOCK(zone);
817 	zone->uz_flags &= ~UMA_ZFLAG_DRAINING;
818 	wakeup(zone);
819 out:
820 	ZONE_UNLOCK(zone);
821 }
822 
823 void
824 zone_drain(uma_zone_t zone)
825 {
826 
827 	zone_drain_wait(zone, M_NOWAIT);
828 }
829 
830 /*
831  * Allocate a new slab for a keg.  This does not insert the slab onto a list.
832  *
833  * Arguments:
834  *	wait  Shall we wait?
835  *
836  * Returns:
837  *	The slab that was allocated or NULL if there is no memory and the
838  *	caller specified M_NOWAIT.
839  */
840 static uma_slab_t
841 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int wait)
842 {
843 	uma_slabrefcnt_t slabref;
844 	uma_alloc allocf;
845 	uma_slab_t slab;
846 	uint8_t *mem;
847 	uint8_t flags;
848 	int i;
849 
850 	mtx_assert(&keg->uk_lock, MA_OWNED);
851 	slab = NULL;
852 	mem = NULL;
853 
854 #ifdef UMA_DEBUG
855 	printf("alloc_slab:  Allocating a new slab for %s\n", keg->uk_name);
856 #endif
857 	allocf = keg->uk_allocf;
858 	KEG_UNLOCK(keg);
859 
860 	if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
861 		slab = zone_alloc_item(keg->uk_slabzone, NULL, wait);
862 		if (slab == NULL)
863 			goto out;
864 	}
865 
866 	/*
867 	 * This reproduces the old vm_zone behavior of zero filling pages the
868 	 * first time they are added to a zone.
869 	 *
870 	 * Malloced items are zeroed in uma_zalloc.
871 	 */
872 
873 	if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
874 		wait |= M_ZERO;
875 	else
876 		wait &= ~M_ZERO;
877 
878 	if (keg->uk_flags & UMA_ZONE_NODUMP)
879 		wait |= M_NODUMP;
880 
881 	/* zone is passed for legacy reasons. */
882 	mem = allocf(zone, keg->uk_ppera * PAGE_SIZE, &flags, wait);
883 	if (mem == NULL) {
884 		if (keg->uk_flags & UMA_ZONE_OFFPAGE)
885 			zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE);
886 		slab = NULL;
887 		goto out;
888 	}
889 
890 	/* Point the slab into the allocated memory */
891 	if (!(keg->uk_flags & UMA_ZONE_OFFPAGE))
892 		slab = (uma_slab_t )(mem + keg->uk_pgoff);
893 
894 	if (keg->uk_flags & UMA_ZONE_VTOSLAB)
895 		for (i = 0; i < keg->uk_ppera; i++)
896 			vsetslab((vm_offset_t)mem + (i * PAGE_SIZE), slab);
897 
898 	slab->us_keg = keg;
899 	slab->us_data = mem;
900 	slab->us_freecount = keg->uk_ipers;
901 	slab->us_flags = flags;
902 	BIT_FILL(SLAB_SETSIZE, &slab->us_free);
903 #ifdef INVARIANTS
904 	BIT_ZERO(SLAB_SETSIZE, &slab->us_debugfree);
905 #endif
906 	if (keg->uk_flags & UMA_ZONE_REFCNT) {
907 		slabref = (uma_slabrefcnt_t)slab;
908 		for (i = 0; i < keg->uk_ipers; i++)
909 			slabref->us_refcnt[i] = 0;
910 	}
911 
912 	if (keg->uk_init != NULL) {
913 		for (i = 0; i < keg->uk_ipers; i++)
914 			if (keg->uk_init(slab->us_data + (keg->uk_rsize * i),
915 			    keg->uk_size, wait) != 0)
916 				break;
917 		if (i != keg->uk_ipers) {
918 			keg_free_slab(keg, slab, i);
919 			slab = NULL;
920 			goto out;
921 		}
922 	}
923 out:
924 	KEG_LOCK(keg);
925 
926 	if (slab != NULL) {
927 		if (keg->uk_flags & UMA_ZONE_HASH)
928 			UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
929 
930 		keg->uk_pages += keg->uk_ppera;
931 		keg->uk_free += keg->uk_ipers;
932 	}
933 
934 	return (slab);
935 }
936 
937 /*
938  * This function is intended to be used early on in place of page_alloc() so
939  * that we may use the boot time page cache to satisfy allocations before
940  * the VM is ready.
941  */
942 static void *
943 startup_alloc(uma_zone_t zone, int bytes, uint8_t *pflag, int wait)
944 {
945 	uma_keg_t keg;
946 	uma_slab_t tmps;
947 	int pages, check_pages;
948 
949 	keg = zone_first_keg(zone);
950 	pages = howmany(bytes, PAGE_SIZE);
951 	check_pages = pages - 1;
952 	KASSERT(pages > 0, ("startup_alloc can't reserve 0 pages\n"));
953 
954 	/*
955 	 * Check our small startup cache to see if it has pages remaining.
956 	 */
957 	mtx_lock(&uma_boot_pages_mtx);
958 
959 	/* First check if we have enough room. */
960 	tmps = LIST_FIRST(&uma_boot_pages);
961 	while (tmps != NULL && check_pages-- > 0)
962 		tmps = LIST_NEXT(tmps, us_link);
963 	if (tmps != NULL) {
964 		/*
965 		 * It's ok to lose tmps references.  The last one will
966 		 * have tmps->us_data pointing to the start address of
967 		 * "pages" contiguous pages of memory.
968 		 */
969 		while (pages-- > 0) {
970 			tmps = LIST_FIRST(&uma_boot_pages);
971 			LIST_REMOVE(tmps, us_link);
972 		}
973 		mtx_unlock(&uma_boot_pages_mtx);
974 		*pflag = tmps->us_flags;
975 		return (tmps->us_data);
976 	}
977 	mtx_unlock(&uma_boot_pages_mtx);
978 	if (booted < UMA_STARTUP2)
979 		panic("UMA: Increase vm.boot_pages");
980 	/*
981 	 * Now that we've booted reset these users to their real allocator.
982 	 */
983 #ifdef UMA_MD_SMALL_ALLOC
984 	keg->uk_allocf = (keg->uk_ppera > 1) ? page_alloc : uma_small_alloc;
985 #else
986 	keg->uk_allocf = page_alloc;
987 #endif
988 	return keg->uk_allocf(zone, bytes, pflag, wait);
989 }
990 
991 /*
992  * Allocates a number of pages from the system
993  *
994  * Arguments:
995  *	bytes  The number of bytes requested
996  *	wait  Shall we wait?
997  *
998  * Returns:
999  *	A pointer to the alloced memory or possibly
1000  *	NULL if M_NOWAIT is set.
1001  */
1002 static void *
1003 page_alloc(uma_zone_t zone, int bytes, uint8_t *pflag, int wait)
1004 {
1005 	void *p;	/* Returned page */
1006 
1007 	*pflag = UMA_SLAB_KMEM;
1008 	p = (void *) kmem_malloc(kmem_map, bytes, wait);
1009 
1010 	return (p);
1011 }
1012 
1013 /*
1014  * Allocates a number of pages from within an object
1015  *
1016  * Arguments:
1017  *	bytes  The number of bytes requested
1018  *	wait   Shall we wait?
1019  *
1020  * Returns:
1021  *	A pointer to the alloced memory or possibly
1022  *	NULL if M_NOWAIT is set.
1023  */
1024 static void *
1025 noobj_alloc(uma_zone_t zone, int bytes, uint8_t *flags, int wait)
1026 {
1027 	TAILQ_HEAD(, vm_page) alloctail;
1028 	u_long npages;
1029 	vm_offset_t retkva, zkva;
1030 	vm_page_t p, p_next;
1031 	uma_keg_t keg;
1032 
1033 	TAILQ_INIT(&alloctail);
1034 	keg = zone_first_keg(zone);
1035 
1036 	npages = howmany(bytes, PAGE_SIZE);
1037 	while (npages > 0) {
1038 		p = vm_page_alloc(NULL, 0, VM_ALLOC_INTERRUPT |
1039 		    VM_ALLOC_WIRED | VM_ALLOC_NOOBJ);
1040 		if (p != NULL) {
1041 			/*
1042 			 * Since the page does not belong to an object, its
1043 			 * listq is unused.
1044 			 */
1045 			TAILQ_INSERT_TAIL(&alloctail, p, listq);
1046 			npages--;
1047 			continue;
1048 		}
1049 		if (wait & M_WAITOK) {
1050 			VM_WAIT;
1051 			continue;
1052 		}
1053 
1054 		/*
1055 		 * Page allocation failed, free intermediate pages and
1056 		 * exit.
1057 		 */
1058 		TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1059 			vm_page_unwire(p, 0);
1060 			vm_page_free(p);
1061 		}
1062 		return (NULL);
1063 	}
1064 	*flags = UMA_SLAB_PRIV;
1065 	zkva = keg->uk_kva +
1066 	    atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
1067 	retkva = zkva;
1068 	TAILQ_FOREACH(p, &alloctail, listq) {
1069 		pmap_qenter(zkva, &p, 1);
1070 		zkva += PAGE_SIZE;
1071 	}
1072 
1073 	return ((void *)retkva);
1074 }
1075 
1076 /*
1077  * Frees a number of pages to the system
1078  *
1079  * Arguments:
1080  *	mem   A pointer to the memory to be freed
1081  *	size  The size of the memory being freed
1082  *	flags The original p->us_flags field
1083  *
1084  * Returns:
1085  *	Nothing
1086  */
1087 static void
1088 page_free(void *mem, int size, uint8_t flags)
1089 {
1090 	vm_map_t map;
1091 
1092 	if (flags & UMA_SLAB_KMEM)
1093 		map = kmem_map;
1094 	else if (flags & UMA_SLAB_KERNEL)
1095 		map = kernel_map;
1096 	else
1097 		panic("UMA: page_free used with invalid flags %d", flags);
1098 
1099 	kmem_free(map, (vm_offset_t)mem, size);
1100 }
1101 
1102 /*
1103  * Zero fill initializer
1104  *
1105  * Arguments/Returns follow uma_init specifications
1106  */
1107 static int
1108 zero_init(void *mem, int size, int flags)
1109 {
1110 	bzero(mem, size);
1111 	return (0);
1112 }
1113 
1114 /*
1115  * Finish creating a small uma keg.  This calculates ipers, and the keg size.
1116  *
1117  * Arguments
1118  *	keg  The zone we should initialize
1119  *
1120  * Returns
1121  *	Nothing
1122  */
1123 static void
1124 keg_small_init(uma_keg_t keg)
1125 {
1126 	u_int rsize;
1127 	u_int memused;
1128 	u_int wastedspace;
1129 	u_int shsize;
1130 
1131 	if (keg->uk_flags & UMA_ZONE_PCPU) {
1132 		KASSERT(mp_ncpus > 0, ("%s: ncpus %d\n", __func__, mp_ncpus));
1133 		keg->uk_slabsize = sizeof(struct pcpu);
1134 		keg->uk_ppera = howmany(mp_ncpus * sizeof(struct pcpu),
1135 		    PAGE_SIZE);
1136 	} else {
1137 		keg->uk_slabsize = UMA_SLAB_SIZE;
1138 		keg->uk_ppera = 1;
1139 	}
1140 
1141 	/*
1142 	 * Calculate the size of each allocation (rsize) according to
1143 	 * alignment.  If the requested size is smaller than we have
1144 	 * allocation bits for we round it up.
1145 	 */
1146 	rsize = keg->uk_size;
1147 	if (rsize < keg->uk_slabsize / SLAB_SETSIZE)
1148 		rsize = keg->uk_slabsize / SLAB_SETSIZE;
1149 	if (rsize & keg->uk_align)
1150 		rsize = (rsize & ~keg->uk_align) + (keg->uk_align + 1);
1151 	keg->uk_rsize = rsize;
1152 
1153 	KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
1154 	    keg->uk_rsize < sizeof(struct pcpu),
1155 	    ("%s: size %u too large", __func__, keg->uk_rsize));
1156 
1157 	if (keg->uk_flags & UMA_ZONE_REFCNT)
1158 		rsize += sizeof(uint32_t);
1159 
1160 	if (keg->uk_flags & UMA_ZONE_OFFPAGE)
1161 		shsize = 0;
1162 	else
1163 		shsize = sizeof(struct uma_slab);
1164 
1165 	keg->uk_ipers = (keg->uk_slabsize - shsize) / rsize;
1166 	KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE,
1167 	    ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers));
1168 
1169 	memused = keg->uk_ipers * rsize + shsize;
1170 	wastedspace = keg->uk_slabsize - memused;
1171 
1172 	/*
1173 	 * We can't do OFFPAGE if we're internal or if we've been
1174 	 * asked to not go to the VM for buckets.  If we do this we
1175 	 * may end up going to the VM (kmem_map) for slabs which we
1176 	 * do not want to do if we're UMA_ZFLAG_CACHEONLY as a
1177 	 * result of UMA_ZONE_VM, which clearly forbids it.
1178 	 */
1179 	if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) ||
1180 	    (keg->uk_flags & UMA_ZFLAG_CACHEONLY))
1181 		return;
1182 
1183 	/*
1184 	 * See if using an OFFPAGE slab will limit our waste.  Only do
1185 	 * this if it permits more items per-slab.
1186 	 *
1187 	 * XXX We could try growing slabsize to limit max waste as well.
1188 	 * Historically this was not done because the VM could not
1189 	 * efficiently handle contiguous allocations.
1190 	 */
1191 	if ((wastedspace >= keg->uk_slabsize / UMA_MAX_WASTE) &&
1192 	    (keg->uk_ipers < (keg->uk_slabsize / keg->uk_rsize))) {
1193 		keg->uk_ipers = keg->uk_slabsize / keg->uk_rsize;
1194 		KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE,
1195 		    ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers));
1196 #ifdef UMA_DEBUG
1197 		printf("UMA decided we need offpage slab headers for "
1198 		    "keg: %s, calculated wastedspace = %d, "
1199 		    "maximum wasted space allowed = %d, "
1200 		    "calculated ipers = %d, "
1201 		    "new wasted space = %d\n", keg->uk_name, wastedspace,
1202 		    keg->uk_slabsize / UMA_MAX_WASTE, keg->uk_ipers,
1203 		    keg->uk_slabsize - keg->uk_ipers * keg->uk_rsize);
1204 #endif
1205 		keg->uk_flags |= UMA_ZONE_OFFPAGE;
1206 	}
1207 
1208 	if ((keg->uk_flags & UMA_ZONE_OFFPAGE) &&
1209 	    (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1210 		keg->uk_flags |= UMA_ZONE_HASH;
1211 }
1212 
1213 /*
1214  * Finish creating a large (> UMA_SLAB_SIZE) uma kegs.  Just give in and do
1215  * OFFPAGE for now.  When I can allow for more dynamic slab sizes this will be
1216  * more complicated.
1217  *
1218  * Arguments
1219  *	keg  The keg we should initialize
1220  *
1221  * Returns
1222  *	Nothing
1223  */
1224 static void
1225 keg_large_init(uma_keg_t keg)
1226 {
1227 
1228 	KASSERT(keg != NULL, ("Keg is null in keg_large_init"));
1229 	KASSERT((keg->uk_flags & UMA_ZFLAG_CACHEONLY) == 0,
1230 	    ("keg_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY keg"));
1231 	KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0,
1232 	    ("%s: Cannot large-init a UMA_ZONE_PCPU keg", __func__));
1233 
1234 	keg->uk_ppera = howmany(keg->uk_size, PAGE_SIZE);
1235 	keg->uk_slabsize = keg->uk_ppera * PAGE_SIZE;
1236 	keg->uk_ipers = 1;
1237 	keg->uk_rsize = keg->uk_size;
1238 
1239 	/* We can't do OFFPAGE if we're internal, bail out here. */
1240 	if (keg->uk_flags & UMA_ZFLAG_INTERNAL)
1241 		return;
1242 
1243 	keg->uk_flags |= UMA_ZONE_OFFPAGE;
1244 	if ((keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1245 		keg->uk_flags |= UMA_ZONE_HASH;
1246 }
1247 
1248 static void
1249 keg_cachespread_init(uma_keg_t keg)
1250 {
1251 	int alignsize;
1252 	int trailer;
1253 	int pages;
1254 	int rsize;
1255 
1256 	KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0,
1257 	    ("%s: Cannot cachespread-init a UMA_ZONE_PCPU keg", __func__));
1258 
1259 	alignsize = keg->uk_align + 1;
1260 	rsize = keg->uk_size;
1261 	/*
1262 	 * We want one item to start on every align boundary in a page.  To
1263 	 * do this we will span pages.  We will also extend the item by the
1264 	 * size of align if it is an even multiple of align.  Otherwise, it
1265 	 * would fall on the same boundary every time.
1266 	 */
1267 	if (rsize & keg->uk_align)
1268 		rsize = (rsize & ~keg->uk_align) + alignsize;
1269 	if ((rsize & alignsize) == 0)
1270 		rsize += alignsize;
1271 	trailer = rsize - keg->uk_size;
1272 	pages = (rsize * (PAGE_SIZE / alignsize)) / PAGE_SIZE;
1273 	pages = MIN(pages, (128 * 1024) / PAGE_SIZE);
1274 	keg->uk_rsize = rsize;
1275 	keg->uk_ppera = pages;
1276 	keg->uk_slabsize = UMA_SLAB_SIZE;
1277 	keg->uk_ipers = ((pages * PAGE_SIZE) + trailer) / rsize;
1278 	keg->uk_flags |= UMA_ZONE_OFFPAGE | UMA_ZONE_VTOSLAB;
1279 	KASSERT(keg->uk_ipers <= uma_max_ipers,
1280 	    ("%s: keg->uk_ipers too high(%d) increase max_ipers", __func__,
1281 	    keg->uk_ipers));
1282 }
1283 
1284 /*
1285  * Keg header ctor.  This initializes all fields, locks, etc.  And inserts
1286  * the keg onto the global keg list.
1287  *
1288  * Arguments/Returns follow uma_ctor specifications
1289  *	udata  Actually uma_kctor_args
1290  */
1291 static int
1292 keg_ctor(void *mem, int size, void *udata, int flags)
1293 {
1294 	struct uma_kctor_args *arg = udata;
1295 	uma_keg_t keg = mem;
1296 	uma_zone_t zone;
1297 
1298 	bzero(keg, size);
1299 	keg->uk_size = arg->size;
1300 	keg->uk_init = arg->uminit;
1301 	keg->uk_fini = arg->fini;
1302 	keg->uk_align = arg->align;
1303 	keg->uk_free = 0;
1304 	keg->uk_pages = 0;
1305 	keg->uk_flags = arg->flags;
1306 	keg->uk_allocf = page_alloc;
1307 	keg->uk_freef = page_free;
1308 	keg->uk_slabzone = NULL;
1309 
1310 	/*
1311 	 * The master zone is passed to us at keg-creation time.
1312 	 */
1313 	zone = arg->zone;
1314 	keg->uk_name = zone->uz_name;
1315 
1316 	if (arg->flags & UMA_ZONE_VM)
1317 		keg->uk_flags |= UMA_ZFLAG_CACHEONLY;
1318 
1319 	if (arg->flags & UMA_ZONE_ZINIT)
1320 		keg->uk_init = zero_init;
1321 
1322 	if (arg->flags & UMA_ZONE_REFCNT || arg->flags & UMA_ZONE_MALLOC)
1323 		keg->uk_flags |= UMA_ZONE_VTOSLAB;
1324 
1325 	if (arg->flags & UMA_ZONE_PCPU)
1326 #ifdef SMP
1327 		keg->uk_flags |= UMA_ZONE_OFFPAGE;
1328 #else
1329 		keg->uk_flags &= ~UMA_ZONE_PCPU;
1330 #endif
1331 
1332 	if (keg->uk_flags & UMA_ZONE_CACHESPREAD) {
1333 		keg_cachespread_init(keg);
1334 	} else if (keg->uk_flags & UMA_ZONE_REFCNT) {
1335 		if (keg->uk_size >
1336 		    (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt) -
1337 		    sizeof(uint32_t)))
1338 			keg_large_init(keg);
1339 		else
1340 			keg_small_init(keg);
1341 	} else {
1342 		if (keg->uk_size > (UMA_SLAB_SIZE - sizeof(struct uma_slab)))
1343 			keg_large_init(keg);
1344 		else
1345 			keg_small_init(keg);
1346 	}
1347 
1348 	if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
1349 		if (keg->uk_flags & UMA_ZONE_REFCNT) {
1350 			if (keg->uk_ipers > uma_max_ipers_ref)
1351 				panic("Too many ref items per zone: %d > %d\n",
1352 				    keg->uk_ipers, uma_max_ipers_ref);
1353 			keg->uk_slabzone = slabrefzone;
1354 		} else
1355 			keg->uk_slabzone = slabzone;
1356 	}
1357 
1358 	/*
1359 	 * If we haven't booted yet we need allocations to go through the
1360 	 * startup cache until the vm is ready.
1361 	 */
1362 	if (keg->uk_ppera == 1) {
1363 #ifdef UMA_MD_SMALL_ALLOC
1364 		keg->uk_allocf = uma_small_alloc;
1365 		keg->uk_freef = uma_small_free;
1366 
1367 		if (booted < UMA_STARTUP)
1368 			keg->uk_allocf = startup_alloc;
1369 #else
1370 		if (booted < UMA_STARTUP2)
1371 			keg->uk_allocf = startup_alloc;
1372 #endif
1373 	} else if (booted < UMA_STARTUP2 &&
1374 	    (keg->uk_flags & UMA_ZFLAG_INTERNAL))
1375 		keg->uk_allocf = startup_alloc;
1376 
1377 	/*
1378 	 * Initialize keg's lock
1379 	 */
1380 	KEG_LOCK_INIT(keg, (arg->flags & UMA_ZONE_MTXCLASS));
1381 
1382 	/*
1383 	 * If we're putting the slab header in the actual page we need to
1384 	 * figure out where in each page it goes.  This calculates a right
1385 	 * justified offset into the memory on an ALIGN_PTR boundary.
1386 	 */
1387 	if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) {
1388 		u_int totsize;
1389 
1390 		/* Size of the slab struct and free list */
1391 		totsize = sizeof(struct uma_slab);
1392 
1393 		/* Size of the reference counts. */
1394 		if (keg->uk_flags & UMA_ZONE_REFCNT)
1395 			totsize += keg->uk_ipers * sizeof(uint32_t);
1396 
1397 		if (totsize & UMA_ALIGN_PTR)
1398 			totsize = (totsize & ~UMA_ALIGN_PTR) +
1399 			    (UMA_ALIGN_PTR + 1);
1400 		keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - totsize;
1401 
1402 		/*
1403 		 * The only way the following is possible is if with our
1404 		 * UMA_ALIGN_PTR adjustments we are now bigger than
1405 		 * UMA_SLAB_SIZE.  I haven't checked whether this is
1406 		 * mathematically possible for all cases, so we make
1407 		 * sure here anyway.
1408 		 */
1409 		totsize = keg->uk_pgoff + sizeof(struct uma_slab);
1410 		if (keg->uk_flags & UMA_ZONE_REFCNT)
1411 			totsize += keg->uk_ipers * sizeof(uint32_t);
1412 		if (totsize > PAGE_SIZE * keg->uk_ppera) {
1413 			printf("zone %s ipers %d rsize %d size %d\n",
1414 			    zone->uz_name, keg->uk_ipers, keg->uk_rsize,
1415 			    keg->uk_size);
1416 			panic("UMA slab won't fit.");
1417 		}
1418 	}
1419 
1420 	if (keg->uk_flags & UMA_ZONE_HASH)
1421 		hash_alloc(&keg->uk_hash);
1422 
1423 #ifdef UMA_DEBUG
1424 	printf("UMA: %s(%p) size %d(%d) flags %#x ipers %d ppera %d out %d free %d\n",
1425 	    zone->uz_name, zone, keg->uk_size, keg->uk_rsize, keg->uk_flags,
1426 	    keg->uk_ipers, keg->uk_ppera,
1427 	    (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free);
1428 #endif
1429 
1430 	LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
1431 
1432 	mtx_lock(&uma_mtx);
1433 	LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
1434 	mtx_unlock(&uma_mtx);
1435 	return (0);
1436 }
1437 
1438 /*
1439  * Zone header ctor.  This initializes all fields, locks, etc.
1440  *
1441  * Arguments/Returns follow uma_ctor specifications
1442  *	udata  Actually uma_zctor_args
1443  */
1444 static int
1445 zone_ctor(void *mem, int size, void *udata, int flags)
1446 {
1447 	struct uma_zctor_args *arg = udata;
1448 	uma_zone_t zone = mem;
1449 	uma_zone_t z;
1450 	uma_keg_t keg;
1451 
1452 	bzero(zone, size);
1453 	zone->uz_name = arg->name;
1454 	zone->uz_ctor = arg->ctor;
1455 	zone->uz_dtor = arg->dtor;
1456 	zone->uz_slab = zone_fetch_slab;
1457 	zone->uz_init = NULL;
1458 	zone->uz_fini = NULL;
1459 	zone->uz_allocs = 0;
1460 	zone->uz_frees = 0;
1461 	zone->uz_fails = 0;
1462 	zone->uz_sleeps = 0;
1463 	zone->uz_count = 0;
1464 	zone->uz_flags = 0;
1465 	zone->uz_warning = NULL;
1466 	timevalclear(&zone->uz_ratecheck);
1467 	keg = arg->keg;
1468 
1469 	ZONE_LOCK_INIT(zone, (arg->flags & UMA_ZONE_MTXCLASS));
1470 
1471 	/*
1472 	 * This is a pure cache zone, no kegs.
1473 	 */
1474 	if (arg->import) {
1475 		zone->uz_size = arg->size;
1476 		zone->uz_import = arg->import;
1477 		zone->uz_release = arg->release;
1478 		zone->uz_arg = arg->arg;
1479 		zone->uz_lockptr = &zone->uz_lock;
1480 		goto out;
1481 	}
1482 
1483 	/*
1484 	 * Use the regular zone/keg/slab allocator.
1485 	 */
1486 	zone->uz_import = (uma_import)zone_import;
1487 	zone->uz_release = (uma_release)zone_release;
1488 	zone->uz_arg = zone;
1489 
1490 	if (arg->flags & UMA_ZONE_SECONDARY) {
1491 		KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
1492 		zone->uz_init = arg->uminit;
1493 		zone->uz_fini = arg->fini;
1494 		zone->uz_lockptr = &keg->uk_lock;
1495 		zone->uz_flags |= UMA_ZONE_SECONDARY;
1496 		mtx_lock(&uma_mtx);
1497 		ZONE_LOCK(zone);
1498 		LIST_FOREACH(z, &keg->uk_zones, uz_link) {
1499 			if (LIST_NEXT(z, uz_link) == NULL) {
1500 				LIST_INSERT_AFTER(z, zone, uz_link);
1501 				break;
1502 			}
1503 		}
1504 		ZONE_UNLOCK(zone);
1505 		mtx_unlock(&uma_mtx);
1506 	} else if (keg == NULL) {
1507 		if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
1508 		    arg->align, arg->flags)) == NULL)
1509 			return (ENOMEM);
1510 	} else {
1511 		struct uma_kctor_args karg;
1512 		int error;
1513 
1514 		/* We should only be here from uma_startup() */
1515 		karg.size = arg->size;
1516 		karg.uminit = arg->uminit;
1517 		karg.fini = arg->fini;
1518 		karg.align = arg->align;
1519 		karg.flags = arg->flags;
1520 		karg.zone = zone;
1521 		error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
1522 		    flags);
1523 		if (error)
1524 			return (error);
1525 	}
1526 
1527 	/*
1528 	 * Link in the first keg.
1529 	 */
1530 	zone->uz_klink.kl_keg = keg;
1531 	LIST_INSERT_HEAD(&zone->uz_kegs, &zone->uz_klink, kl_link);
1532 	zone->uz_lockptr = &keg->uk_lock;
1533 	zone->uz_size = keg->uk_size;
1534 	zone->uz_flags |= (keg->uk_flags &
1535 	    (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
1536 
1537 	/*
1538 	 * Some internal zones don't have room allocated for the per cpu
1539 	 * caches.  If we're internal, bail out here.
1540 	 */
1541 	if (keg->uk_flags & UMA_ZFLAG_INTERNAL) {
1542 		KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
1543 		    ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
1544 		return (0);
1545 	}
1546 
1547 out:
1548 	if ((arg->flags & UMA_ZONE_MAXBUCKET) == 0)
1549 		zone->uz_count = bucket_select(zone->uz_size);
1550 	else
1551 		zone->uz_count = BUCKET_MAX;
1552 
1553 	return (0);
1554 }
1555 
1556 /*
1557  * Keg header dtor.  This frees all data, destroys locks, frees the hash
1558  * table and removes the keg from the global list.
1559  *
1560  * Arguments/Returns follow uma_dtor specifications
1561  *	udata  unused
1562  */
1563 static void
1564 keg_dtor(void *arg, int size, void *udata)
1565 {
1566 	uma_keg_t keg;
1567 
1568 	keg = (uma_keg_t)arg;
1569 	KEG_LOCK(keg);
1570 	if (keg->uk_free != 0) {
1571 		printf("Freed UMA keg was not empty (%d items). "
1572 		    " Lost %d pages of memory.\n",
1573 		    keg->uk_free, keg->uk_pages);
1574 	}
1575 	KEG_UNLOCK(keg);
1576 
1577 	hash_free(&keg->uk_hash);
1578 
1579 	KEG_LOCK_FINI(keg);
1580 }
1581 
1582 /*
1583  * Zone header dtor.
1584  *
1585  * Arguments/Returns follow uma_dtor specifications
1586  *	udata  unused
1587  */
1588 static void
1589 zone_dtor(void *arg, int size, void *udata)
1590 {
1591 	uma_klink_t klink;
1592 	uma_zone_t zone;
1593 	uma_keg_t keg;
1594 
1595 	zone = (uma_zone_t)arg;
1596 	keg = zone_first_keg(zone);
1597 
1598 	if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
1599 		cache_drain(zone);
1600 
1601 	mtx_lock(&uma_mtx);
1602 	LIST_REMOVE(zone, uz_link);
1603 	mtx_unlock(&uma_mtx);
1604 	/*
1605 	 * XXX there are some races here where
1606 	 * the zone can be drained but zone lock
1607 	 * released and then refilled before we
1608 	 * remove it... we dont care for now
1609 	 */
1610 	zone_drain_wait(zone, M_WAITOK);
1611 	/*
1612 	 * Unlink all of our kegs.
1613 	 */
1614 	while ((klink = LIST_FIRST(&zone->uz_kegs)) != NULL) {
1615 		klink->kl_keg = NULL;
1616 		LIST_REMOVE(klink, kl_link);
1617 		if (klink == &zone->uz_klink)
1618 			continue;
1619 		free(klink, M_TEMP);
1620 	}
1621 	/*
1622 	 * We only destroy kegs from non secondary zones.
1623 	 */
1624 	if (keg != NULL && (zone->uz_flags & UMA_ZONE_SECONDARY) == 0)  {
1625 		mtx_lock(&uma_mtx);
1626 		LIST_REMOVE(keg, uk_link);
1627 		mtx_unlock(&uma_mtx);
1628 		zone_free_item(kegs, keg, NULL, SKIP_NONE);
1629 	}
1630 	ZONE_LOCK_FINI(zone);
1631 }
1632 
1633 /*
1634  * Traverses every zone in the system and calls a callback
1635  *
1636  * Arguments:
1637  *	zfunc  A pointer to a function which accepts a zone
1638  *		as an argument.
1639  *
1640  * Returns:
1641  *	Nothing
1642  */
1643 static void
1644 zone_foreach(void (*zfunc)(uma_zone_t))
1645 {
1646 	uma_keg_t keg;
1647 	uma_zone_t zone;
1648 
1649 	mtx_lock(&uma_mtx);
1650 	LIST_FOREACH(keg, &uma_kegs, uk_link) {
1651 		LIST_FOREACH(zone, &keg->uk_zones, uz_link)
1652 			zfunc(zone);
1653 	}
1654 	mtx_unlock(&uma_mtx);
1655 }
1656 
1657 /* Public functions */
1658 /* See uma.h */
1659 void
1660 uma_startup(void *bootmem, int boot_pages)
1661 {
1662 	struct uma_zctor_args args;
1663 	uma_slab_t slab;
1664 	u_int slabsize;
1665 	int i;
1666 
1667 #ifdef UMA_DEBUG
1668 	printf("Creating uma keg headers zone and keg.\n");
1669 #endif
1670 	mtx_init(&uma_mtx, "UMA lock", NULL, MTX_DEF);
1671 
1672 	/* "manually" create the initial zone */
1673 	memset(&args, 0, sizeof(args));
1674 	args.name = "UMA Kegs";
1675 	args.size = sizeof(struct uma_keg);
1676 	args.ctor = keg_ctor;
1677 	args.dtor = keg_dtor;
1678 	args.uminit = zero_init;
1679 	args.fini = NULL;
1680 	args.keg = &masterkeg;
1681 	args.align = 32 - 1;
1682 	args.flags = UMA_ZFLAG_INTERNAL;
1683 	/* The initial zone has no Per cpu queues so it's smaller */
1684 	zone_ctor(kegs, sizeof(struct uma_zone), &args, M_WAITOK);
1685 
1686 #ifdef UMA_DEBUG
1687 	printf("Filling boot free list.\n");
1688 #endif
1689 	for (i = 0; i < boot_pages; i++) {
1690 		slab = (uma_slab_t)((uint8_t *)bootmem + (i * UMA_SLAB_SIZE));
1691 		slab->us_data = (uint8_t *)slab;
1692 		slab->us_flags = UMA_SLAB_BOOT;
1693 		LIST_INSERT_HEAD(&uma_boot_pages, slab, us_link);
1694 	}
1695 	mtx_init(&uma_boot_pages_mtx, "UMA boot pages", NULL, MTX_DEF);
1696 
1697 #ifdef UMA_DEBUG
1698 	printf("Creating uma zone headers zone and keg.\n");
1699 #endif
1700 	args.name = "UMA Zones";
1701 	args.size = sizeof(struct uma_zone) +
1702 	    (sizeof(struct uma_cache) * (mp_maxid + 1));
1703 	args.ctor = zone_ctor;
1704 	args.dtor = zone_dtor;
1705 	args.uminit = zero_init;
1706 	args.fini = NULL;
1707 	args.keg = NULL;
1708 	args.align = 32 - 1;
1709 	args.flags = UMA_ZFLAG_INTERNAL;
1710 	/* The initial zone has no Per cpu queues so it's smaller */
1711 	zone_ctor(zones, sizeof(struct uma_zone), &args, M_WAITOK);
1712 
1713 #ifdef UMA_DEBUG
1714 	printf("Initializing pcpu cache locks.\n");
1715 #endif
1716 #ifdef UMA_DEBUG
1717 	printf("Creating slab and hash zones.\n");
1718 #endif
1719 
1720 	/* Now make a zone for slab headers */
1721 	slabzone = uma_zcreate("UMA Slabs",
1722 				sizeof(struct uma_slab),
1723 				NULL, NULL, NULL, NULL,
1724 				UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1725 
1726 	/*
1727 	 * We also create a zone for the bigger slabs with reference
1728 	 * counts in them, to accomodate UMA_ZONE_REFCNT zones.
1729 	 */
1730 	slabsize = sizeof(struct uma_slab_refcnt);
1731 	slabsize += uma_max_ipers_ref * sizeof(uint32_t);
1732 	slabrefzone = uma_zcreate("UMA RCntSlabs",
1733 				  slabsize,
1734 				  NULL, NULL, NULL, NULL,
1735 				  UMA_ALIGN_PTR,
1736 				  UMA_ZFLAG_INTERNAL);
1737 
1738 	hashzone = uma_zcreate("UMA Hash",
1739 	    sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
1740 	    NULL, NULL, NULL, NULL,
1741 	    UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1742 
1743 	bucket_init();
1744 
1745 	booted = UMA_STARTUP;
1746 
1747 #ifdef UMA_DEBUG
1748 	printf("UMA startup complete.\n");
1749 #endif
1750 }
1751 
1752 /* see uma.h */
1753 void
1754 uma_startup2(void)
1755 {
1756 	booted = UMA_STARTUP2;
1757 	bucket_enable();
1758 #ifdef UMA_DEBUG
1759 	printf("UMA startup2 complete.\n");
1760 #endif
1761 }
1762 
1763 /*
1764  * Initialize our callout handle
1765  *
1766  */
1767 
1768 static void
1769 uma_startup3(void)
1770 {
1771 #ifdef UMA_DEBUG
1772 	printf("Starting callout.\n");
1773 #endif
1774 	callout_init(&uma_callout, CALLOUT_MPSAFE);
1775 	callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
1776 #ifdef UMA_DEBUG
1777 	printf("UMA startup3 complete.\n");
1778 #endif
1779 }
1780 
1781 static uma_keg_t
1782 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
1783 		int align, uint32_t flags)
1784 {
1785 	struct uma_kctor_args args;
1786 
1787 	args.size = size;
1788 	args.uminit = uminit;
1789 	args.fini = fini;
1790 	args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
1791 	args.flags = flags;
1792 	args.zone = zone;
1793 	return (zone_alloc_item(kegs, &args, M_WAITOK));
1794 }
1795 
1796 /* See uma.h */
1797 void
1798 uma_set_align(int align)
1799 {
1800 
1801 	if (align != UMA_ALIGN_CACHE)
1802 		uma_align_cache = align;
1803 }
1804 
1805 /* See uma.h */
1806 uma_zone_t
1807 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
1808 		uma_init uminit, uma_fini fini, int align, uint32_t flags)
1809 
1810 {
1811 	struct uma_zctor_args args;
1812 
1813 	/* This stuff is essential for the zone ctor */
1814 	memset(&args, 0, sizeof(args));
1815 	args.name = name;
1816 	args.size = size;
1817 	args.ctor = ctor;
1818 	args.dtor = dtor;
1819 	args.uminit = uminit;
1820 	args.fini = fini;
1821 	args.align = align;
1822 	args.flags = flags;
1823 	args.keg = NULL;
1824 
1825 	return (zone_alloc_item(zones, &args, M_WAITOK));
1826 }
1827 
1828 /* See uma.h */
1829 uma_zone_t
1830 uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor,
1831 		    uma_init zinit, uma_fini zfini, uma_zone_t master)
1832 {
1833 	struct uma_zctor_args args;
1834 	uma_keg_t keg;
1835 
1836 	keg = zone_first_keg(master);
1837 	memset(&args, 0, sizeof(args));
1838 	args.name = name;
1839 	args.size = keg->uk_size;
1840 	args.ctor = ctor;
1841 	args.dtor = dtor;
1842 	args.uminit = zinit;
1843 	args.fini = zfini;
1844 	args.align = keg->uk_align;
1845 	args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
1846 	args.keg = keg;
1847 
1848 	/* XXX Attaches only one keg of potentially many. */
1849 	return (zone_alloc_item(zones, &args, M_WAITOK));
1850 }
1851 
1852 /* See uma.h */
1853 uma_zone_t
1854 uma_zcache_create(char *name, int size, uma_ctor ctor, uma_dtor dtor,
1855 		    uma_init zinit, uma_fini zfini, uma_import zimport,
1856 		    uma_release zrelease, void *arg, int flags)
1857 {
1858 	struct uma_zctor_args args;
1859 
1860 	memset(&args, 0, sizeof(args));
1861 	args.name = name;
1862 	args.size = size;
1863 	args.ctor = ctor;
1864 	args.dtor = dtor;
1865 	args.uminit = zinit;
1866 	args.fini = zfini;
1867 	args.import = zimport;
1868 	args.release = zrelease;
1869 	args.arg = arg;
1870 	args.align = 0;
1871 	args.flags = flags;
1872 
1873 	return (zone_alloc_item(zones, &args, M_WAITOK));
1874 }
1875 
1876 static void
1877 zone_lock_pair(uma_zone_t a, uma_zone_t b)
1878 {
1879 	if (a < b) {
1880 		ZONE_LOCK(a);
1881 		mtx_lock_flags(b->uz_lockptr, MTX_DUPOK);
1882 	} else {
1883 		ZONE_LOCK(b);
1884 		mtx_lock_flags(a->uz_lockptr, MTX_DUPOK);
1885 	}
1886 }
1887 
1888 static void
1889 zone_unlock_pair(uma_zone_t a, uma_zone_t b)
1890 {
1891 
1892 	ZONE_UNLOCK(a);
1893 	ZONE_UNLOCK(b);
1894 }
1895 
1896 int
1897 uma_zsecond_add(uma_zone_t zone, uma_zone_t master)
1898 {
1899 	uma_klink_t klink;
1900 	uma_klink_t kl;
1901 	int error;
1902 
1903 	error = 0;
1904 	klink = malloc(sizeof(*klink), M_TEMP, M_WAITOK | M_ZERO);
1905 
1906 	zone_lock_pair(zone, master);
1907 	/*
1908 	 * zone must use vtoslab() to resolve objects and must already be
1909 	 * a secondary.
1910 	 */
1911 	if ((zone->uz_flags & (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY))
1912 	    != (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) {
1913 		error = EINVAL;
1914 		goto out;
1915 	}
1916 	/*
1917 	 * The new master must also use vtoslab().
1918 	 */
1919 	if ((zone->uz_flags & UMA_ZONE_VTOSLAB) != UMA_ZONE_VTOSLAB) {
1920 		error = EINVAL;
1921 		goto out;
1922 	}
1923 	/*
1924 	 * Both must either be refcnt, or not be refcnt.
1925 	 */
1926 	if ((zone->uz_flags & UMA_ZONE_REFCNT) !=
1927 	    (master->uz_flags & UMA_ZONE_REFCNT)) {
1928 		error = EINVAL;
1929 		goto out;
1930 	}
1931 	/*
1932 	 * The underlying object must be the same size.  rsize
1933 	 * may be different.
1934 	 */
1935 	if (master->uz_size != zone->uz_size) {
1936 		error = E2BIG;
1937 		goto out;
1938 	}
1939 	/*
1940 	 * Put it at the end of the list.
1941 	 */
1942 	klink->kl_keg = zone_first_keg(master);
1943 	LIST_FOREACH(kl, &zone->uz_kegs, kl_link) {
1944 		if (LIST_NEXT(kl, kl_link) == NULL) {
1945 			LIST_INSERT_AFTER(kl, klink, kl_link);
1946 			break;
1947 		}
1948 	}
1949 	klink = NULL;
1950 	zone->uz_flags |= UMA_ZFLAG_MULTI;
1951 	zone->uz_slab = zone_fetch_slab_multi;
1952 
1953 out:
1954 	zone_unlock_pair(zone, master);
1955 	if (klink != NULL)
1956 		free(klink, M_TEMP);
1957 
1958 	return (error);
1959 }
1960 
1961 
1962 /* See uma.h */
1963 void
1964 uma_zdestroy(uma_zone_t zone)
1965 {
1966 
1967 	zone_free_item(zones, zone, NULL, SKIP_NONE);
1968 }
1969 
1970 /* See uma.h */
1971 void *
1972 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
1973 {
1974 	void *item;
1975 	uma_cache_t cache;
1976 	uma_bucket_t bucket;
1977 	int lockfail;
1978 	int cpu;
1979 
1980 	/* This is the fast path allocation */
1981 #ifdef UMA_DEBUG_ALLOC_1
1982 	printf("Allocating one item from %s(%p)\n", zone->uz_name, zone);
1983 #endif
1984 	CTR3(KTR_UMA, "uma_zalloc_arg thread %x zone %s flags %d", curthread,
1985 	    zone->uz_name, flags);
1986 
1987 	if (flags & M_WAITOK) {
1988 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
1989 		    "uma_zalloc_arg: zone \"%s\"", zone->uz_name);
1990 	}
1991 #ifdef DEBUG_MEMGUARD
1992 	if (memguard_cmp_zone(zone)) {
1993 		item = memguard_alloc(zone->uz_size, flags);
1994 		if (item != NULL) {
1995 			/*
1996 			 * Avoid conflict with the use-after-free
1997 			 * protecting infrastructure from INVARIANTS.
1998 			 */
1999 			if (zone->uz_init != NULL &&
2000 			    zone->uz_init != mtrash_init &&
2001 			    zone->uz_init(item, zone->uz_size, flags) != 0)
2002 				return (NULL);
2003 			if (zone->uz_ctor != NULL &&
2004 			    zone->uz_ctor != mtrash_ctor &&
2005 			    zone->uz_ctor(item, zone->uz_size, udata,
2006 			    flags) != 0) {
2007 			    	zone->uz_fini(item, zone->uz_size);
2008 				return (NULL);
2009 			}
2010 			return (item);
2011 		}
2012 		/* This is unfortunate but should not be fatal. */
2013 	}
2014 #endif
2015 	/*
2016 	 * If possible, allocate from the per-CPU cache.  There are two
2017 	 * requirements for safe access to the per-CPU cache: (1) the thread
2018 	 * accessing the cache must not be preempted or yield during access,
2019 	 * and (2) the thread must not migrate CPUs without switching which
2020 	 * cache it accesses.  We rely on a critical section to prevent
2021 	 * preemption and migration.  We release the critical section in
2022 	 * order to acquire the zone mutex if we are unable to allocate from
2023 	 * the current cache; when we re-acquire the critical section, we
2024 	 * must detect and handle migration if it has occurred.
2025 	 */
2026 	critical_enter();
2027 	cpu = curcpu;
2028 	cache = &zone->uz_cpu[cpu];
2029 
2030 zalloc_start:
2031 	bucket = cache->uc_allocbucket;
2032 	if (bucket != NULL && bucket->ub_cnt > 0) {
2033 		bucket->ub_cnt--;
2034 		item = bucket->ub_bucket[bucket->ub_cnt];
2035 #ifdef INVARIANTS
2036 		bucket->ub_bucket[bucket->ub_cnt] = NULL;
2037 #endif
2038 		KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled."));
2039 		cache->uc_allocs++;
2040 		critical_exit();
2041 		if (zone->uz_ctor != NULL &&
2042 		    zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2043 			atomic_add_long(&zone->uz_fails, 1);
2044 			zone_free_item(zone, item, udata, SKIP_DTOR);
2045 			return (NULL);
2046 		}
2047 #ifdef INVARIANTS
2048 		uma_dbg_alloc(zone, NULL, item);
2049 #endif
2050 		if (flags & M_ZERO)
2051 			bzero(item, zone->uz_size);
2052 		return (item);
2053 	}
2054 
2055 	/*
2056 	 * We have run out of items in our alloc bucket.
2057 	 * See if we can switch with our free bucket.
2058 	 */
2059 	bucket = cache->uc_freebucket;
2060 	if (bucket != NULL && bucket->ub_cnt > 0) {
2061 #ifdef UMA_DEBUG_ALLOC
2062 		printf("uma_zalloc: Swapping empty with alloc.\n");
2063 #endif
2064 		cache->uc_freebucket = cache->uc_allocbucket;
2065 		cache->uc_allocbucket = bucket;
2066 		goto zalloc_start;
2067 	}
2068 
2069 	/*
2070 	 * Discard any empty allocation bucket while we hold no locks.
2071 	 */
2072 	bucket = cache->uc_allocbucket;
2073 	cache->uc_allocbucket = NULL;
2074 	critical_exit();
2075 	if (bucket != NULL)
2076 		bucket_free(zone, bucket);
2077 
2078 	/* Short-circuit for zones without buckets and low memory. */
2079 	if (zone->uz_count == 0 || bucketdisable)
2080 		goto zalloc_item;
2081 
2082 	/*
2083 	 * Attempt to retrieve the item from the per-CPU cache has failed, so
2084 	 * we must go back to the zone.  This requires the zone lock, so we
2085 	 * must drop the critical section, then re-acquire it when we go back
2086 	 * to the cache.  Since the critical section is released, we may be
2087 	 * preempted or migrate.  As such, make sure not to maintain any
2088 	 * thread-local state specific to the cache from prior to releasing
2089 	 * the critical section.
2090 	 */
2091 	lockfail = 0;
2092 	if (ZONE_TRYLOCK(zone) == 0) {
2093 		/* Record contention to size the buckets. */
2094 		ZONE_LOCK(zone);
2095 		lockfail = 1;
2096 	}
2097 	critical_enter();
2098 	cpu = curcpu;
2099 	cache = &zone->uz_cpu[cpu];
2100 
2101 	/*
2102 	 * Since we have locked the zone we may as well send back our stats.
2103 	 */
2104 	atomic_add_long(&zone->uz_allocs, cache->uc_allocs);
2105 	atomic_add_long(&zone->uz_frees, cache->uc_frees);
2106 	cache->uc_allocs = 0;
2107 	cache->uc_frees = 0;
2108 
2109 	/* See if we lost the race to fill the cache. */
2110 	if (cache->uc_allocbucket != NULL) {
2111 		ZONE_UNLOCK(zone);
2112 		goto zalloc_start;
2113 	}
2114 
2115 	/*
2116 	 * Check the zone's cache of buckets.
2117 	 */
2118 	if ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) {
2119 		KASSERT(bucket->ub_cnt != 0,
2120 		    ("uma_zalloc_arg: Returning an empty bucket."));
2121 
2122 		LIST_REMOVE(bucket, ub_link);
2123 		cache->uc_allocbucket = bucket;
2124 		ZONE_UNLOCK(zone);
2125 		goto zalloc_start;
2126 	}
2127 	/* We are no longer associated with this CPU. */
2128 	critical_exit();
2129 
2130 	/*
2131 	 * We bump the uz count when the cache size is insufficient to
2132 	 * handle the working set.
2133 	 */
2134 	if (lockfail && zone->uz_count < BUCKET_MAX && zone->uz_count != 0 &&
2135 	    zone != largebucket)
2136 		zone->uz_count++;
2137 	ZONE_UNLOCK(zone);
2138 
2139 	/*
2140 	 * Now lets just fill a bucket and put it on the free list.  If that
2141 	 * works we'll restart the allocation from the begining and it
2142 	 * will use the just filled bucket.
2143 	 */
2144 	bucket = zone_alloc_bucket(zone, flags);
2145 	if (bucket != NULL) {
2146 		ZONE_LOCK(zone);
2147 		critical_enter();
2148 		cpu = curcpu;
2149 		cache = &zone->uz_cpu[cpu];
2150 		/*
2151 		 * See if we lost the race or were migrated.  Cache the
2152 		 * initialized bucket to make this less likely or claim
2153 		 * the memory directly.
2154 		 */
2155 		if (cache->uc_allocbucket == NULL)
2156 			cache->uc_allocbucket = bucket;
2157 		else
2158 			LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link);
2159 		ZONE_UNLOCK(zone);
2160 		goto zalloc_start;
2161 	}
2162 
2163 	/*
2164 	 * We may not be able to get a bucket so return an actual item.
2165 	 */
2166 #ifdef UMA_DEBUG
2167 	printf("uma_zalloc_arg: Bucketzone returned NULL\n");
2168 #endif
2169 
2170 zalloc_item:
2171 	item = zone_alloc_item(zone, udata, flags);
2172 
2173 	return (item);
2174 }
2175 
2176 static uma_slab_t
2177 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int flags)
2178 {
2179 	uma_slab_t slab;
2180 
2181 	mtx_assert(&keg->uk_lock, MA_OWNED);
2182 	slab = NULL;
2183 
2184 	for (;;) {
2185 		/*
2186 		 * Find a slab with some space.  Prefer slabs that are partially
2187 		 * used over those that are totally full.  This helps to reduce
2188 		 * fragmentation.
2189 		 */
2190 		if (keg->uk_free != 0) {
2191 			if (!LIST_EMPTY(&keg->uk_part_slab)) {
2192 				slab = LIST_FIRST(&keg->uk_part_slab);
2193 			} else {
2194 				slab = LIST_FIRST(&keg->uk_free_slab);
2195 				LIST_REMOVE(slab, us_link);
2196 				LIST_INSERT_HEAD(&keg->uk_part_slab, slab,
2197 				    us_link);
2198 			}
2199 			MPASS(slab->us_keg == keg);
2200 			return (slab);
2201 		}
2202 
2203 		/*
2204 		 * M_NOVM means don't ask at all!
2205 		 */
2206 		if (flags & M_NOVM)
2207 			break;
2208 
2209 		if (keg->uk_maxpages && keg->uk_pages >= keg->uk_maxpages) {
2210 			keg->uk_flags |= UMA_ZFLAG_FULL;
2211 			/*
2212 			 * If this is not a multi-zone, set the FULL bit.
2213 			 * Otherwise slab_multi() takes care of it.
2214 			 */
2215 			if ((zone->uz_flags & UMA_ZFLAG_MULTI) == 0) {
2216 				zone->uz_flags |= UMA_ZFLAG_FULL;
2217 				zone_log_warning(zone);
2218 			}
2219 			if (flags & M_NOWAIT)
2220 				break;
2221 			zone->uz_sleeps++;
2222 			msleep(keg, &keg->uk_lock, PVM, "keglimit", 0);
2223 			continue;
2224 		}
2225 		slab = keg_alloc_slab(keg, zone, flags);
2226 		/*
2227 		 * If we got a slab here it's safe to mark it partially used
2228 		 * and return.  We assume that the caller is going to remove
2229 		 * at least one item.
2230 		 */
2231 		if (slab) {
2232 			MPASS(slab->us_keg == keg);
2233 			LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2234 			return (slab);
2235 		}
2236 		/*
2237 		 * We might not have been able to get a slab but another cpu
2238 		 * could have while we were unlocked.  Check again before we
2239 		 * fail.
2240 		 */
2241 		flags |= M_NOVM;
2242 	}
2243 	return (slab);
2244 }
2245 
2246 static uma_slab_t
2247 zone_fetch_slab(uma_zone_t zone, uma_keg_t keg, int flags)
2248 {
2249 	uma_slab_t slab;
2250 
2251 	if (keg == NULL) {
2252 		keg = zone_first_keg(zone);
2253 		KEG_LOCK(keg);
2254 	}
2255 
2256 	for (;;) {
2257 		slab = keg_fetch_slab(keg, zone, flags);
2258 		if (slab)
2259 			return (slab);
2260 		if (flags & (M_NOWAIT | M_NOVM))
2261 			break;
2262 	}
2263 	KEG_UNLOCK(keg);
2264 	return (NULL);
2265 }
2266 
2267 /*
2268  * uma_zone_fetch_slab_multi:  Fetches a slab from one available keg.  Returns
2269  * with the keg locked.  On NULL no lock is held.
2270  *
2271  * The last pointer is used to seed the search.  It is not required.
2272  */
2273 static uma_slab_t
2274 zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int rflags)
2275 {
2276 	uma_klink_t klink;
2277 	uma_slab_t slab;
2278 	uma_keg_t keg;
2279 	int flags;
2280 	int empty;
2281 	int full;
2282 
2283 	/*
2284 	 * Don't wait on the first pass.  This will skip limit tests
2285 	 * as well.  We don't want to block if we can find a provider
2286 	 * without blocking.
2287 	 */
2288 	flags = (rflags & ~M_WAITOK) | M_NOWAIT;
2289 	/*
2290 	 * Use the last slab allocated as a hint for where to start
2291 	 * the search.
2292 	 */
2293 	if (last != NULL) {
2294 		slab = keg_fetch_slab(last, zone, flags);
2295 		if (slab)
2296 			return (slab);
2297 		KEG_UNLOCK(last);
2298 	}
2299 	/*
2300 	 * Loop until we have a slab incase of transient failures
2301 	 * while M_WAITOK is specified.  I'm not sure this is 100%
2302 	 * required but we've done it for so long now.
2303 	 */
2304 	for (;;) {
2305 		empty = 0;
2306 		full = 0;
2307 		/*
2308 		 * Search the available kegs for slabs.  Be careful to hold the
2309 		 * correct lock while calling into the keg layer.
2310 		 */
2311 		LIST_FOREACH(klink, &zone->uz_kegs, kl_link) {
2312 			keg = klink->kl_keg;
2313 			KEG_LOCK(keg);
2314 			if ((keg->uk_flags & UMA_ZFLAG_FULL) == 0) {
2315 				slab = keg_fetch_slab(keg, zone, flags);
2316 				if (slab)
2317 					return (slab);
2318 			}
2319 			if (keg->uk_flags & UMA_ZFLAG_FULL)
2320 				full++;
2321 			else
2322 				empty++;
2323 			KEG_UNLOCK(keg);
2324 		}
2325 		if (rflags & (M_NOWAIT | M_NOVM))
2326 			break;
2327 		flags = rflags;
2328 		/*
2329 		 * All kegs are full.  XXX We can't atomically check all kegs
2330 		 * and sleep so just sleep for a short period and retry.
2331 		 */
2332 		if (full && !empty) {
2333 			ZONE_LOCK(zone);
2334 			zone->uz_flags |= UMA_ZFLAG_FULL;
2335 			zone->uz_sleeps++;
2336 			zone_log_warning(zone);
2337 			msleep(zone, zone->uz_lockptr, PVM,
2338 			    "zonelimit", hz/100);
2339 			zone->uz_flags &= ~UMA_ZFLAG_FULL;
2340 			ZONE_UNLOCK(zone);
2341 			continue;
2342 		}
2343 	}
2344 	return (NULL);
2345 }
2346 
2347 static void *
2348 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
2349 {
2350 	void *item;
2351 	uint8_t freei;
2352 
2353 	MPASS(keg == slab->us_keg);
2354 	mtx_assert(&keg->uk_lock, MA_OWNED);
2355 
2356 	freei = BIT_FFS(SLAB_SETSIZE, &slab->us_free) - 1;
2357 	BIT_CLR(SLAB_SETSIZE, freei, &slab->us_free);
2358 	item = slab->us_data + (keg->uk_rsize * freei);
2359 	slab->us_freecount--;
2360 	keg->uk_free--;
2361 
2362 	/* Move this slab to the full list */
2363 	if (slab->us_freecount == 0) {
2364 		LIST_REMOVE(slab, us_link);
2365 		LIST_INSERT_HEAD(&keg->uk_full_slab, slab, us_link);
2366 	}
2367 
2368 	return (item);
2369 }
2370 
2371 static int
2372 zone_import(uma_zone_t zone, void **bucket, int max, int flags)
2373 {
2374 	uma_slab_t slab;
2375 	uma_keg_t keg;
2376 	int i;
2377 
2378 	slab = NULL;
2379 	keg = NULL;
2380 	/* Try to keep the buckets totally full */
2381 	for (i = 0; i < max; ) {
2382 		if ((slab = zone->uz_slab(zone, keg, flags)) == NULL)
2383 			break;
2384 		keg = slab->us_keg;
2385 		while (slab->us_freecount && i < max)
2386 			bucket[i++] = slab_alloc_item(keg, slab);
2387 
2388 		/* Don't block on the next fill */
2389 		flags &= ~M_WAITOK;
2390 		flags |= M_NOWAIT;
2391 	}
2392 	if (slab != NULL)
2393 		KEG_UNLOCK(keg);
2394 
2395 	return i;
2396 }
2397 
2398 static uma_bucket_t
2399 zone_alloc_bucket(uma_zone_t zone, int flags)
2400 {
2401 	uma_bucket_t bucket;
2402 	int max;
2403 
2404 	bucket = bucket_alloc(zone, M_NOWAIT | (flags & M_NOVM));
2405 	if (bucket == NULL)
2406 		goto out;
2407 
2408 	max = MIN(bucket->ub_entries, zone->uz_count);
2409 	bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
2410 	    max, flags);
2411 
2412 	/*
2413 	 * Initialize the memory if necessary.
2414 	 */
2415 	if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
2416 		int i;
2417 
2418 		for (i = 0; i < bucket->ub_cnt; i++)
2419 			if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size,
2420 			    flags) != 0)
2421 				break;
2422 		/*
2423 		 * If we couldn't initialize the whole bucket, put the
2424 		 * rest back onto the freelist.
2425 		 */
2426 		if (i != bucket->ub_cnt) {
2427 			zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
2428 			    bucket->ub_cnt - i);
2429 #ifdef INVARIANTS
2430 			bzero(&bucket->ub_bucket[i],
2431 			    sizeof(void *) * (bucket->ub_cnt - i));
2432 #endif
2433 			bucket->ub_cnt = i;
2434 		}
2435 	}
2436 
2437 out:
2438 	if (bucket == NULL || bucket->ub_cnt == 0) {
2439 		if (bucket != NULL)
2440 			bucket_free(zone, bucket);
2441 		atomic_add_long(&zone->uz_fails, 1);
2442 		return (NULL);
2443 	}
2444 
2445 	return (bucket);
2446 }
2447 
2448 /*
2449  * Allocates a single item from a zone.
2450  *
2451  * Arguments
2452  *	zone   The zone to alloc for.
2453  *	udata  The data to be passed to the constructor.
2454  *	flags  M_WAITOK, M_NOWAIT, M_ZERO.
2455  *
2456  * Returns
2457  *	NULL if there is no memory and M_NOWAIT is set
2458  *	An item if successful
2459  */
2460 
2461 static void *
2462 zone_alloc_item(uma_zone_t zone, void *udata, int flags)
2463 {
2464 	void *item;
2465 
2466 	item = NULL;
2467 
2468 #ifdef UMA_DEBUG_ALLOC
2469 	printf("INTERNAL: Allocating one item from %s(%p)\n", zone->uz_name, zone);
2470 #endif
2471 	if (zone->uz_import(zone->uz_arg, &item, 1, flags) != 1)
2472 		goto fail;
2473 	atomic_add_long(&zone->uz_allocs, 1);
2474 
2475 	/*
2476 	 * We have to call both the zone's init (not the keg's init)
2477 	 * and the zone's ctor.  This is because the item is going from
2478 	 * a keg slab directly to the user, and the user is expecting it
2479 	 * to be both zone-init'd as well as zone-ctor'd.
2480 	 */
2481 	if (zone->uz_init != NULL) {
2482 		if (zone->uz_init(item, zone->uz_size, flags) != 0) {
2483 			zone_free_item(zone, item, udata, SKIP_FINI);
2484 			goto fail;
2485 		}
2486 	}
2487 	if (zone->uz_ctor != NULL) {
2488 		if (zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2489 			zone_free_item(zone, item, udata, SKIP_DTOR);
2490 			goto fail;
2491 		}
2492 	}
2493 #ifdef INVARIANTS
2494 	uma_dbg_alloc(zone, NULL, item);
2495 #endif
2496 	if (flags & M_ZERO)
2497 		bzero(item, zone->uz_size);
2498 
2499 	return (item);
2500 
2501 fail:
2502 	atomic_add_long(&zone->uz_fails, 1);
2503 	return (NULL);
2504 }
2505 
2506 /* See uma.h */
2507 void
2508 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
2509 {
2510 	uma_cache_t cache;
2511 	uma_bucket_t bucket;
2512 	int cpu;
2513 
2514 #ifdef UMA_DEBUG_ALLOC_1
2515 	printf("Freeing item %p to %s(%p)\n", item, zone->uz_name, zone);
2516 #endif
2517 	CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread,
2518 	    zone->uz_name);
2519 
2520         /* uma_zfree(..., NULL) does nothing, to match free(9). */
2521         if (item == NULL)
2522                 return;
2523 #ifdef DEBUG_MEMGUARD
2524 	if (is_memguard_addr(item)) {
2525 		if (zone->uz_dtor != NULL && zone->uz_dtor != mtrash_dtor)
2526 			zone->uz_dtor(item, zone->uz_size, udata);
2527 		if (zone->uz_fini != NULL && zone->uz_fini != mtrash_fini)
2528 			zone->uz_fini(item, zone->uz_size);
2529 		memguard_free(item);
2530 		return;
2531 	}
2532 #endif
2533 #ifdef INVARIANTS
2534 	if (zone->uz_flags & UMA_ZONE_MALLOC)
2535 		uma_dbg_free(zone, udata, item);
2536 	else
2537 		uma_dbg_free(zone, NULL, item);
2538 #endif
2539 	if (zone->uz_dtor != NULL)
2540 		zone->uz_dtor(item, zone->uz_size, udata);
2541 
2542 	/*
2543 	 * The race here is acceptable.  If we miss it we'll just have to wait
2544 	 * a little longer for the limits to be reset.
2545 	 */
2546 	if (zone->uz_flags & UMA_ZFLAG_FULL)
2547 		goto zfree_item;
2548 
2549 	/*
2550 	 * If possible, free to the per-CPU cache.  There are two
2551 	 * requirements for safe access to the per-CPU cache: (1) the thread
2552 	 * accessing the cache must not be preempted or yield during access,
2553 	 * and (2) the thread must not migrate CPUs without switching which
2554 	 * cache it accesses.  We rely on a critical section to prevent
2555 	 * preemption and migration.  We release the critical section in
2556 	 * order to acquire the zone mutex if we are unable to free to the
2557 	 * current cache; when we re-acquire the critical section, we must
2558 	 * detect and handle migration if it has occurred.
2559 	 */
2560 zfree_restart:
2561 	critical_enter();
2562 	cpu = curcpu;
2563 	cache = &zone->uz_cpu[cpu];
2564 
2565 zfree_start:
2566 	/*
2567 	 * Try to free into the allocbucket first to give LIFO ordering
2568 	 * for cache-hot datastructures.  Spill over into the freebucket
2569 	 * if necessary.  Alloc will swap them if one runs dry.
2570 	 */
2571 	bucket = cache->uc_allocbucket;
2572 	if (bucket == NULL || bucket->ub_cnt >= bucket->ub_entries)
2573 		bucket = cache->uc_freebucket;
2574 	if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) {
2575 		KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL,
2576 		    ("uma_zfree: Freeing to non free bucket index."));
2577 		bucket->ub_bucket[bucket->ub_cnt] = item;
2578 		bucket->ub_cnt++;
2579 		cache->uc_frees++;
2580 		critical_exit();
2581 		return;
2582 	}
2583 
2584 	/*
2585 	 * We must go back the zone, which requires acquiring the zone lock,
2586 	 * which in turn means we must release and re-acquire the critical
2587 	 * section.  Since the critical section is released, we may be
2588 	 * preempted or migrate.  As such, make sure not to maintain any
2589 	 * thread-local state specific to the cache from prior to releasing
2590 	 * the critical section.
2591 	 */
2592 	critical_exit();
2593 	if (zone->uz_count == 0 || bucketdisable)
2594 		goto zfree_item;
2595 
2596 	ZONE_LOCK(zone);
2597 	critical_enter();
2598 	cpu = curcpu;
2599 	cache = &zone->uz_cpu[cpu];
2600 
2601 	/*
2602 	 * Since we have locked the zone we may as well send back our stats.
2603 	 */
2604 	atomic_add_long(&zone->uz_allocs, cache->uc_allocs);
2605 	atomic_add_long(&zone->uz_frees, cache->uc_frees);
2606 	cache->uc_allocs = 0;
2607 	cache->uc_frees = 0;
2608 
2609 	bucket = cache->uc_freebucket;
2610 	if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) {
2611 		ZONE_UNLOCK(zone);
2612 		goto zfree_start;
2613 	}
2614 	cache->uc_freebucket = NULL;
2615 
2616 	/* Can we throw this on the zone full list? */
2617 	if (bucket != NULL) {
2618 #ifdef UMA_DEBUG_ALLOC
2619 		printf("uma_zfree: Putting old bucket on the free list.\n");
2620 #endif
2621 		/* ub_cnt is pointing to the last free item */
2622 		KASSERT(bucket->ub_cnt != 0,
2623 		    ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n"));
2624 		LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link);
2625 	}
2626 
2627 	/* We are no longer associated with this CPU. */
2628 	critical_exit();
2629 
2630 	/* And the zone.. */
2631 	ZONE_UNLOCK(zone);
2632 
2633 #ifdef UMA_DEBUG_ALLOC
2634 	printf("uma_zfree: Allocating new free bucket.\n");
2635 #endif
2636 	bucket = bucket_alloc(zone, M_NOWAIT);
2637 	if (bucket) {
2638 		critical_enter();
2639 		cpu = curcpu;
2640 		cache = &zone->uz_cpu[cpu];
2641 		if (cache->uc_freebucket == NULL) {
2642 			cache->uc_freebucket = bucket;
2643 			goto zfree_start;
2644 		}
2645 		/*
2646 		 * We lost the race, start over.  We have to drop our
2647 		 * critical section to free the bucket.
2648 		 */
2649 		critical_exit();
2650 		bucket_free(zone, bucket);
2651 		goto zfree_restart;
2652 	}
2653 
2654 	/*
2655 	 * If nothing else caught this, we'll just do an internal free.
2656 	 */
2657 zfree_item:
2658 	zone_free_item(zone, item, udata, SKIP_DTOR);
2659 
2660 	return;
2661 }
2662 
2663 static void
2664 slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item)
2665 {
2666 	uint8_t freei;
2667 
2668 	mtx_assert(&keg->uk_lock, MA_OWNED);
2669 	MPASS(keg == slab->us_keg);
2670 
2671 	/* Do we need to remove from any lists? */
2672 	if (slab->us_freecount+1 == keg->uk_ipers) {
2673 		LIST_REMOVE(slab, us_link);
2674 		LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
2675 	} else if (slab->us_freecount == 0) {
2676 		LIST_REMOVE(slab, us_link);
2677 		LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2678 	}
2679 
2680 	/* Slab management. */
2681 	freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
2682 	BIT_SET(SLAB_SETSIZE, freei, &slab->us_free);
2683 	slab->us_freecount++;
2684 
2685 	/* Keg statistics. */
2686 	keg->uk_free++;
2687 }
2688 
2689 static void
2690 zone_release(uma_zone_t zone, void **bucket, int cnt)
2691 {
2692 	void *item;
2693 	uma_slab_t slab;
2694 	uma_keg_t keg;
2695 	uint8_t *mem;
2696 	int clearfull;
2697 	int i;
2698 
2699 	clearfull = 0;
2700 	keg = zone_first_keg(zone);
2701 	KEG_LOCK(keg);
2702 	for (i = 0; i < cnt; i++) {
2703 		item = bucket[i];
2704 		if (!(zone->uz_flags & UMA_ZONE_VTOSLAB)) {
2705 			mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
2706 			if (zone->uz_flags & UMA_ZONE_HASH) {
2707 				slab = hash_sfind(&keg->uk_hash, mem);
2708 			} else {
2709 				mem += keg->uk_pgoff;
2710 				slab = (uma_slab_t)mem;
2711 			}
2712 		} else {
2713 			slab = vtoslab((vm_offset_t)item);
2714 			if (slab->us_keg != keg) {
2715 				KEG_UNLOCK(keg);
2716 				keg = slab->us_keg;
2717 				KEG_LOCK(keg);
2718 			}
2719 		}
2720 		slab_free_item(keg, slab, item);
2721 		if (keg->uk_flags & UMA_ZFLAG_FULL) {
2722 			if (keg->uk_pages < keg->uk_maxpages) {
2723 				keg->uk_flags &= ~UMA_ZFLAG_FULL;
2724 				clearfull = 1;
2725 			}
2726 
2727 			/*
2728 			 * We can handle one more allocation. Since we're
2729 			 * clearing ZFLAG_FULL, wake up all procs blocked
2730 			 * on pages. This should be uncommon, so keeping this
2731 			 * simple for now (rather than adding count of blocked
2732 			 * threads etc).
2733 			 */
2734 			wakeup(keg);
2735 		}
2736 	}
2737 	KEG_UNLOCK(keg);
2738 	if (clearfull) {
2739 		ZONE_LOCK(zone);
2740 		zone->uz_flags &= ~UMA_ZFLAG_FULL;
2741 		wakeup(zone);
2742 		ZONE_UNLOCK(zone);
2743 	}
2744 
2745 }
2746 
2747 /*
2748  * Frees a single item to any zone.
2749  *
2750  * Arguments:
2751  *	zone   The zone to free to
2752  *	item   The item we're freeing
2753  *	udata  User supplied data for the dtor
2754  *	skip   Skip dtors and finis
2755  */
2756 static void
2757 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
2758 {
2759 
2760 #ifdef INVARIANTS
2761 	if (skip == SKIP_NONE) {
2762 		if (zone->uz_flags & UMA_ZONE_MALLOC)
2763 			uma_dbg_free(zone, udata, item);
2764 		else
2765 			uma_dbg_free(zone, NULL, item);
2766 	}
2767 #endif
2768 	if (skip < SKIP_DTOR && zone->uz_dtor)
2769 		zone->uz_dtor(item, zone->uz_size, udata);
2770 
2771 	if (skip < SKIP_FINI && zone->uz_fini)
2772 		zone->uz_fini(item, zone->uz_size);
2773 
2774 	atomic_add_long(&zone->uz_frees, 1);
2775 	zone->uz_release(zone->uz_arg, &item, 1);
2776 }
2777 
2778 /* See uma.h */
2779 int
2780 uma_zone_set_max(uma_zone_t zone, int nitems)
2781 {
2782 	uma_keg_t keg;
2783 
2784 	keg = zone_first_keg(zone);
2785 	if (keg == NULL)
2786 		return (0);
2787 	KEG_LOCK(keg);
2788 	keg->uk_maxpages = (nitems / keg->uk_ipers) * keg->uk_ppera;
2789 	if (keg->uk_maxpages * keg->uk_ipers < nitems)
2790 		keg->uk_maxpages += keg->uk_ppera;
2791 	nitems = keg->uk_maxpages * keg->uk_ipers;
2792 	KEG_UNLOCK(keg);
2793 
2794 	return (nitems);
2795 }
2796 
2797 /* See uma.h */
2798 int
2799 uma_zone_get_max(uma_zone_t zone)
2800 {
2801 	int nitems;
2802 	uma_keg_t keg;
2803 
2804 	keg = zone_first_keg(zone);
2805 	if (keg == NULL)
2806 		return (0);
2807 	KEG_LOCK(keg);
2808 	nitems = keg->uk_maxpages * keg->uk_ipers;
2809 	KEG_UNLOCK(keg);
2810 
2811 	return (nitems);
2812 }
2813 
2814 /* See uma.h */
2815 void
2816 uma_zone_set_warning(uma_zone_t zone, const char *warning)
2817 {
2818 
2819 	ZONE_LOCK(zone);
2820 	zone->uz_warning = warning;
2821 	ZONE_UNLOCK(zone);
2822 }
2823 
2824 /* See uma.h */
2825 int
2826 uma_zone_get_cur(uma_zone_t zone)
2827 {
2828 	int64_t nitems;
2829 	u_int i;
2830 
2831 	ZONE_LOCK(zone);
2832 	nitems = zone->uz_allocs - zone->uz_frees;
2833 	CPU_FOREACH(i) {
2834 		/*
2835 		 * See the comment in sysctl_vm_zone_stats() regarding the
2836 		 * safety of accessing the per-cpu caches. With the zone lock
2837 		 * held, it is safe, but can potentially result in stale data.
2838 		 */
2839 		nitems += zone->uz_cpu[i].uc_allocs -
2840 		    zone->uz_cpu[i].uc_frees;
2841 	}
2842 	ZONE_UNLOCK(zone);
2843 
2844 	return (nitems < 0 ? 0 : nitems);
2845 }
2846 
2847 /* See uma.h */
2848 void
2849 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
2850 {
2851 	uma_keg_t keg;
2852 
2853 	keg = zone_first_keg(zone);
2854 	KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type"));
2855 	KEG_LOCK(keg);
2856 	KASSERT(keg->uk_pages == 0,
2857 	    ("uma_zone_set_init on non-empty keg"));
2858 	keg->uk_init = uminit;
2859 	KEG_UNLOCK(keg);
2860 }
2861 
2862 /* See uma.h */
2863 void
2864 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
2865 {
2866 	uma_keg_t keg;
2867 
2868 	keg = zone_first_keg(zone);
2869 	KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type"));
2870 	KEG_LOCK(keg);
2871 	KASSERT(keg->uk_pages == 0,
2872 	    ("uma_zone_set_fini on non-empty keg"));
2873 	keg->uk_fini = fini;
2874 	KEG_UNLOCK(keg);
2875 }
2876 
2877 /* See uma.h */
2878 void
2879 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
2880 {
2881 
2882 	ZONE_LOCK(zone);
2883 	KASSERT(zone_first_keg(zone)->uk_pages == 0,
2884 	    ("uma_zone_set_zinit on non-empty keg"));
2885 	zone->uz_init = zinit;
2886 	ZONE_UNLOCK(zone);
2887 }
2888 
2889 /* See uma.h */
2890 void
2891 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
2892 {
2893 
2894 	ZONE_LOCK(zone);
2895 	KASSERT(zone_first_keg(zone)->uk_pages == 0,
2896 	    ("uma_zone_set_zfini on non-empty keg"));
2897 	zone->uz_fini = zfini;
2898 	ZONE_UNLOCK(zone);
2899 }
2900 
2901 /* See uma.h */
2902 /* XXX uk_freef is not actually used with the zone locked */
2903 void
2904 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
2905 {
2906 	uma_keg_t keg;
2907 
2908 	keg = zone_first_keg(zone);
2909 	KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type"));
2910 	KEG_LOCK(keg);
2911 	keg->uk_freef = freef;
2912 	KEG_UNLOCK(keg);
2913 }
2914 
2915 /* See uma.h */
2916 /* XXX uk_allocf is not actually used with the zone locked */
2917 void
2918 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
2919 {
2920 	uma_keg_t keg;
2921 
2922 	keg = zone_first_keg(zone);
2923 	KEG_LOCK(keg);
2924 	keg->uk_flags |= UMA_ZFLAG_PRIVALLOC;
2925 	keg->uk_allocf = allocf;
2926 	KEG_UNLOCK(keg);
2927 }
2928 
2929 /* See uma.h */
2930 int
2931 uma_zone_reserve_kva(uma_zone_t zone, int count)
2932 {
2933 	uma_keg_t keg;
2934 	vm_offset_t kva;
2935 	int pages;
2936 
2937 	keg = zone_first_keg(zone);
2938 	if (keg == NULL)
2939 		return (0);
2940 	pages = count / keg->uk_ipers;
2941 
2942 	if (pages * keg->uk_ipers < count)
2943 		pages++;
2944 
2945 #ifdef UMA_MD_SMALL_ALLOC
2946 	if (keg->uk_ppera > 1) {
2947 #else
2948 	if (1) {
2949 #endif
2950 		kva = kmem_alloc_nofault(kernel_map, pages * UMA_SLAB_SIZE);
2951 		if (kva == 0)
2952 			return (0);
2953 	} else
2954 		kva = 0;
2955 	KEG_LOCK(keg);
2956 	keg->uk_kva = kva;
2957 	keg->uk_offset = 0;
2958 	keg->uk_maxpages = pages;
2959 #ifdef UMA_MD_SMALL_ALLOC
2960 	keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
2961 #else
2962 	keg->uk_allocf = noobj_alloc;
2963 #endif
2964 	keg->uk_flags |= UMA_ZONE_NOFREE | UMA_ZFLAG_PRIVALLOC;
2965 	KEG_UNLOCK(keg);
2966 
2967 	return (1);
2968 }
2969 
2970 /* See uma.h */
2971 void
2972 uma_prealloc(uma_zone_t zone, int items)
2973 {
2974 	int slabs;
2975 	uma_slab_t slab;
2976 	uma_keg_t keg;
2977 
2978 	keg = zone_first_keg(zone);
2979 	if (keg == NULL)
2980 		return;
2981 	KEG_LOCK(keg);
2982 	slabs = items / keg->uk_ipers;
2983 	if (slabs * keg->uk_ipers < items)
2984 		slabs++;
2985 	while (slabs > 0) {
2986 		slab = keg_alloc_slab(keg, zone, M_WAITOK);
2987 		if (slab == NULL)
2988 			break;
2989 		MPASS(slab->us_keg == keg);
2990 		LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
2991 		slabs--;
2992 	}
2993 	KEG_UNLOCK(keg);
2994 }
2995 
2996 /* See uma.h */
2997 uint32_t *
2998 uma_find_refcnt(uma_zone_t zone, void *item)
2999 {
3000 	uma_slabrefcnt_t slabref;
3001 	uma_slab_t slab;
3002 	uma_keg_t keg;
3003 	uint32_t *refcnt;
3004 	int idx;
3005 
3006 	slab = vtoslab((vm_offset_t)item & (~UMA_SLAB_MASK));
3007 	slabref = (uma_slabrefcnt_t)slab;
3008 	keg = slab->us_keg;
3009 	KASSERT(keg->uk_flags & UMA_ZONE_REFCNT,
3010 	    ("uma_find_refcnt(): zone possibly not UMA_ZONE_REFCNT"));
3011 	idx = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
3012 	refcnt = &slabref->us_refcnt[idx];
3013 	return refcnt;
3014 }
3015 
3016 /* See uma.h */
3017 void
3018 uma_reclaim(void)
3019 {
3020 #ifdef UMA_DEBUG
3021 	printf("UMA: vm asked us to release pages!\n");
3022 #endif
3023 	bucket_enable();
3024 	zone_foreach(zone_drain);
3025 	/*
3026 	 * Some slabs may have been freed but this zone will be visited early
3027 	 * we visit again so that we can free pages that are empty once other
3028 	 * zones are drained.  We have to do the same for buckets.
3029 	 */
3030 	zone_drain(slabzone);
3031 	zone_drain(slabrefzone);
3032 	bucket_zone_drain();
3033 }
3034 
3035 /* See uma.h */
3036 int
3037 uma_zone_exhausted(uma_zone_t zone)
3038 {
3039 	int full;
3040 
3041 	ZONE_LOCK(zone);
3042 	full = (zone->uz_flags & UMA_ZFLAG_FULL);
3043 	ZONE_UNLOCK(zone);
3044 	return (full);
3045 }
3046 
3047 int
3048 uma_zone_exhausted_nolock(uma_zone_t zone)
3049 {
3050 	return (zone->uz_flags & UMA_ZFLAG_FULL);
3051 }
3052 
3053 void *
3054 uma_large_malloc(int size, int wait)
3055 {
3056 	void *mem;
3057 	uma_slab_t slab;
3058 	uint8_t flags;
3059 
3060 	slab = zone_alloc_item(slabzone, NULL, wait);
3061 	if (slab == NULL)
3062 		return (NULL);
3063 	mem = page_alloc(NULL, size, &flags, wait);
3064 	if (mem) {
3065 		vsetslab((vm_offset_t)mem, slab);
3066 		slab->us_data = mem;
3067 		slab->us_flags = flags | UMA_SLAB_MALLOC;
3068 		slab->us_size = size;
3069 	} else {
3070 		zone_free_item(slabzone, slab, NULL, SKIP_NONE);
3071 	}
3072 
3073 	return (mem);
3074 }
3075 
3076 void
3077 uma_large_free(uma_slab_t slab)
3078 {
3079 	vsetobj((vm_offset_t)slab->us_data, kmem_object);
3080 	page_free(slab->us_data, slab->us_size, slab->us_flags);
3081 	zone_free_item(slabzone, slab, NULL, SKIP_NONE);
3082 }
3083 
3084 void
3085 uma_print_stats(void)
3086 {
3087 	zone_foreach(uma_print_zone);
3088 }
3089 
3090 static void
3091 slab_print(uma_slab_t slab)
3092 {
3093 	printf("slab: keg %p, data %p, freecount %d\n",
3094 		slab->us_keg, slab->us_data, slab->us_freecount);
3095 }
3096 
3097 static void
3098 cache_print(uma_cache_t cache)
3099 {
3100 	printf("alloc: %p(%d), free: %p(%d)\n",
3101 		cache->uc_allocbucket,
3102 		cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0,
3103 		cache->uc_freebucket,
3104 		cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0);
3105 }
3106 
3107 static void
3108 uma_print_keg(uma_keg_t keg)
3109 {
3110 	uma_slab_t slab;
3111 
3112 	printf("keg: %s(%p) size %d(%d) flags %#x ipers %d ppera %d "
3113 	    "out %d free %d limit %d\n",
3114 	    keg->uk_name, keg, keg->uk_size, keg->uk_rsize, keg->uk_flags,
3115 	    keg->uk_ipers, keg->uk_ppera,
3116 	    (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free,
3117 	    (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers);
3118 	printf("Part slabs:\n");
3119 	LIST_FOREACH(slab, &keg->uk_part_slab, us_link)
3120 		slab_print(slab);
3121 	printf("Free slabs:\n");
3122 	LIST_FOREACH(slab, &keg->uk_free_slab, us_link)
3123 		slab_print(slab);
3124 	printf("Full slabs:\n");
3125 	LIST_FOREACH(slab, &keg->uk_full_slab, us_link)
3126 		slab_print(slab);
3127 }
3128 
3129 void
3130 uma_print_zone(uma_zone_t zone)
3131 {
3132 	uma_cache_t cache;
3133 	uma_klink_t kl;
3134 	int i;
3135 
3136 	printf("zone: %s(%p) size %d flags %#x\n",
3137 	    zone->uz_name, zone, zone->uz_size, zone->uz_flags);
3138 	LIST_FOREACH(kl, &zone->uz_kegs, kl_link)
3139 		uma_print_keg(kl->kl_keg);
3140 	CPU_FOREACH(i) {
3141 		cache = &zone->uz_cpu[i];
3142 		printf("CPU %d Cache:\n", i);
3143 		cache_print(cache);
3144 	}
3145 }
3146 
3147 #ifdef DDB
3148 /*
3149  * Generate statistics across both the zone and its per-cpu cache's.  Return
3150  * desired statistics if the pointer is non-NULL for that statistic.
3151  *
3152  * Note: does not update the zone statistics, as it can't safely clear the
3153  * per-CPU cache statistic.
3154  *
3155  * XXXRW: Following the uc_allocbucket and uc_freebucket pointers here isn't
3156  * safe from off-CPU; we should modify the caches to track this information
3157  * directly so that we don't have to.
3158  */
3159 static void
3160 uma_zone_sumstat(uma_zone_t z, int *cachefreep, uint64_t *allocsp,
3161     uint64_t *freesp, uint64_t *sleepsp)
3162 {
3163 	uma_cache_t cache;
3164 	uint64_t allocs, frees, sleeps;
3165 	int cachefree, cpu;
3166 
3167 	allocs = frees = sleeps = 0;
3168 	cachefree = 0;
3169 	CPU_FOREACH(cpu) {
3170 		cache = &z->uz_cpu[cpu];
3171 		if (cache->uc_allocbucket != NULL)
3172 			cachefree += cache->uc_allocbucket->ub_cnt;
3173 		if (cache->uc_freebucket != NULL)
3174 			cachefree += cache->uc_freebucket->ub_cnt;
3175 		allocs += cache->uc_allocs;
3176 		frees += cache->uc_frees;
3177 	}
3178 	allocs += z->uz_allocs;
3179 	frees += z->uz_frees;
3180 	sleeps += z->uz_sleeps;
3181 	if (cachefreep != NULL)
3182 		*cachefreep = cachefree;
3183 	if (allocsp != NULL)
3184 		*allocsp = allocs;
3185 	if (freesp != NULL)
3186 		*freesp = frees;
3187 	if (sleepsp != NULL)
3188 		*sleepsp = sleeps;
3189 }
3190 #endif /* DDB */
3191 
3192 static int
3193 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
3194 {
3195 	uma_keg_t kz;
3196 	uma_zone_t z;
3197 	int count;
3198 
3199 	count = 0;
3200 	mtx_lock(&uma_mtx);
3201 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3202 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
3203 			count++;
3204 	}
3205 	mtx_unlock(&uma_mtx);
3206 	return (sysctl_handle_int(oidp, &count, 0, req));
3207 }
3208 
3209 static int
3210 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
3211 {
3212 	struct uma_stream_header ush;
3213 	struct uma_type_header uth;
3214 	struct uma_percpu_stat ups;
3215 	uma_bucket_t bucket;
3216 	struct sbuf sbuf;
3217 	uma_cache_t cache;
3218 	uma_klink_t kl;
3219 	uma_keg_t kz;
3220 	uma_zone_t z;
3221 	uma_keg_t k;
3222 	int count, error, i;
3223 
3224 	error = sysctl_wire_old_buffer(req, 0);
3225 	if (error != 0)
3226 		return (error);
3227 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
3228 
3229 	count = 0;
3230 	mtx_lock(&uma_mtx);
3231 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3232 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
3233 			count++;
3234 	}
3235 
3236 	/*
3237 	 * Insert stream header.
3238 	 */
3239 	bzero(&ush, sizeof(ush));
3240 	ush.ush_version = UMA_STREAM_VERSION;
3241 	ush.ush_maxcpus = (mp_maxid + 1);
3242 	ush.ush_count = count;
3243 	(void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
3244 
3245 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3246 		LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3247 			bzero(&uth, sizeof(uth));
3248 			ZONE_LOCK(z);
3249 			strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
3250 			uth.uth_align = kz->uk_align;
3251 			uth.uth_size = kz->uk_size;
3252 			uth.uth_rsize = kz->uk_rsize;
3253 			LIST_FOREACH(kl, &z->uz_kegs, kl_link) {
3254 				k = kl->kl_keg;
3255 				uth.uth_maxpages += k->uk_maxpages;
3256 				uth.uth_pages += k->uk_pages;
3257 				uth.uth_keg_free += k->uk_free;
3258 				uth.uth_limit = (k->uk_maxpages / k->uk_ppera)
3259 				    * k->uk_ipers;
3260 			}
3261 
3262 			/*
3263 			 * A zone is secondary is it is not the first entry
3264 			 * on the keg's zone list.
3265 			 */
3266 			if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
3267 			    (LIST_FIRST(&kz->uk_zones) != z))
3268 				uth.uth_zone_flags = UTH_ZONE_SECONDARY;
3269 
3270 			LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3271 				uth.uth_zone_free += bucket->ub_cnt;
3272 			uth.uth_allocs = z->uz_allocs;
3273 			uth.uth_frees = z->uz_frees;
3274 			uth.uth_fails = z->uz_fails;
3275 			uth.uth_sleeps = z->uz_sleeps;
3276 			(void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
3277 			/*
3278 			 * While it is not normally safe to access the cache
3279 			 * bucket pointers while not on the CPU that owns the
3280 			 * cache, we only allow the pointers to be exchanged
3281 			 * without the zone lock held, not invalidated, so
3282 			 * accept the possible race associated with bucket
3283 			 * exchange during monitoring.
3284 			 */
3285 			for (i = 0; i < (mp_maxid + 1); i++) {
3286 				bzero(&ups, sizeof(ups));
3287 				if (kz->uk_flags & UMA_ZFLAG_INTERNAL)
3288 					goto skip;
3289 				if (CPU_ABSENT(i))
3290 					goto skip;
3291 				cache = &z->uz_cpu[i];
3292 				if (cache->uc_allocbucket != NULL)
3293 					ups.ups_cache_free +=
3294 					    cache->uc_allocbucket->ub_cnt;
3295 				if (cache->uc_freebucket != NULL)
3296 					ups.ups_cache_free +=
3297 					    cache->uc_freebucket->ub_cnt;
3298 				ups.ups_allocs = cache->uc_allocs;
3299 				ups.ups_frees = cache->uc_frees;
3300 skip:
3301 				(void)sbuf_bcat(&sbuf, &ups, sizeof(ups));
3302 			}
3303 			ZONE_UNLOCK(z);
3304 		}
3305 	}
3306 	mtx_unlock(&uma_mtx);
3307 	error = sbuf_finish(&sbuf);
3308 	sbuf_delete(&sbuf);
3309 	return (error);
3310 }
3311 
3312 #ifdef DDB
3313 DB_SHOW_COMMAND(uma, db_show_uma)
3314 {
3315 	uint64_t allocs, frees, sleeps;
3316 	uma_bucket_t bucket;
3317 	uma_keg_t kz;
3318 	uma_zone_t z;
3319 	int cachefree;
3320 
3321 	db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
3322 	    "Requests", "Sleeps");
3323 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3324 		LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3325 			if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
3326 				allocs = z->uz_allocs;
3327 				frees = z->uz_frees;
3328 				sleeps = z->uz_sleeps;
3329 				cachefree = 0;
3330 			} else
3331 				uma_zone_sumstat(z, &cachefree, &allocs,
3332 				    &frees, &sleeps);
3333 			if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
3334 			    (LIST_FIRST(&kz->uk_zones) != z)))
3335 				cachefree += kz->uk_free;
3336 			LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3337 				cachefree += bucket->ub_cnt;
3338 			db_printf("%18s %8ju %8jd %8d %12ju %8ju\n", z->uz_name,
3339 			    (uintmax_t)kz->uk_size,
3340 			    (intmax_t)(allocs - frees), cachefree,
3341 			    (uintmax_t)allocs, sleeps);
3342 			if (db_pager_quit)
3343 				return;
3344 		}
3345 	}
3346 }
3347 #endif
3348