xref: /freebsd/sys/vm/uma_core.c (revision 3b1f7d9e5d6f44b50ff07fde6fd0e1135f213762)
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(int, int);
250 static void bucket_free(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 inline void zone_relock(uma_zone_t zone, uma_keg_t keg);
260 static inline void keg_relock(uma_keg_t keg, uma_zone_t zone);
261 static int zone_import(uma_zone_t zone, void **bucket, int max, int flags);
262 static void zone_release(uma_zone_t zone, void **bucket, int cnt);
263 
264 void uma_print_zone(uma_zone_t);
265 void uma_print_stats(void);
266 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS);
267 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS);
268 
269 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
270 
271 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLTYPE_INT,
272     0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones");
273 
274 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLTYPE_STRUCT,
275     0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats");
276 
277 static int zone_warnings = 1;
278 TUNABLE_INT("vm.zone_warnings", &zone_warnings);
279 SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RW, &zone_warnings, 0,
280     "Warn when UMA zones becomes full");
281 
282 /*
283  * This routine checks to see whether or not it's safe to enable buckets.
284  */
285 static void
286 bucket_enable(void)
287 {
288 	bucketdisable = vm_page_count_min();
289 }
290 
291 /*
292  * Initialize bucket_zones, the array of zones of buckets of various sizes.
293  *
294  * For each zone, calculate the memory required for each bucket, consisting
295  * of the header and an array of pointers.
296  */
297 static void
298 bucket_init(void)
299 {
300 	struct uma_bucket_zone *ubz;
301 	int size;
302 	int i;
303 
304 	for (i = 0, ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) {
305 		size = roundup(sizeof(struct uma_bucket), sizeof(void *));
306 		size += sizeof(void *) * ubz->ubz_entries;
307 		ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
308 		    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
309 		    UMA_ZONE_MAXBUCKET | UMA_ZONE_MTXCLASS);
310 	}
311 	/*
312 	 * To avoid recursive bucket allocation loops we disable buckets
313 	 * on the smallest bucket zone and use it for the largest zone.
314 	 * The remainder of the zones all use the largest zone.
315 	 */
316 	ubz--;
317 	ubz->ubz_zone->uz_count = bucket_zones[0].ubz_entries;
318 	bucket_zones[0].ubz_zone->uz_count = 0;
319 	largebucket = ubz->ubz_zone;
320 }
321 
322 /*
323  * Given a desired number of entries for a bucket, return the zone from which
324  * to allocate the bucket.
325  */
326 static struct uma_bucket_zone *
327 bucket_zone_lookup(int entries)
328 {
329 	struct uma_bucket_zone *ubz;
330 
331 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
332 		if (ubz->ubz_entries >= entries)
333 			return (ubz);
334 	ubz--;
335 	return (ubz);
336 }
337 
338 static int
339 bucket_select(int size)
340 {
341 	struct uma_bucket_zone *ubz;
342 
343 	ubz = &bucket_zones[0];
344 	if (size > ubz->ubz_maxsize)
345 		return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1);
346 
347 	for (; ubz->ubz_entries != 0; ubz++)
348 		if (ubz->ubz_maxsize < size)
349 			break;
350 	ubz--;
351 	return (ubz->ubz_entries);
352 }
353 
354 static uma_bucket_t
355 bucket_alloc(int entries, int bflags)
356 {
357 	struct uma_bucket_zone *ubz;
358 	uma_bucket_t bucket;
359 
360 	/*
361 	 * This is to stop us from allocating per cpu buckets while we're
362 	 * running out of vm.boot_pages.  Otherwise, we would exhaust the
363 	 * boot pages.  This also prevents us from allocating buckets in
364 	 * low memory situations.
365 	 */
366 	if (bucketdisable)
367 		return (NULL);
368 
369 	ubz = bucket_zone_lookup(entries);
370 	bucket = uma_zalloc(ubz->ubz_zone, bflags);
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_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(cache->uc_allocbucket);
666 		if (cache->uc_freebucket != NULL)
667 			bucket_free(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(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_lock, 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 (shared among zones).
1379 	 */
1380 	if (arg->flags & UMA_ZONE_MTXCLASS)
1381 		KEG_LOCK_INIT(keg, 1);
1382 	else
1383 		KEG_LOCK_INIT(keg, 0);
1384 
1385 	/*
1386 	 * If we're putting the slab header in the actual page we need to
1387 	 * figure out where in each page it goes.  This calculates a right
1388 	 * justified offset into the memory on an ALIGN_PTR boundary.
1389 	 */
1390 	if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) {
1391 		u_int totsize;
1392 
1393 		/* Size of the slab struct and free list */
1394 		totsize = sizeof(struct uma_slab);
1395 
1396 		/* Size of the reference counts. */
1397 		if (keg->uk_flags & UMA_ZONE_REFCNT)
1398 			totsize += keg->uk_ipers * sizeof(uint32_t);
1399 
1400 		if (totsize & UMA_ALIGN_PTR)
1401 			totsize = (totsize & ~UMA_ALIGN_PTR) +
1402 			    (UMA_ALIGN_PTR + 1);
1403 		keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - totsize;
1404 
1405 		/*
1406 		 * The only way the following is possible is if with our
1407 		 * UMA_ALIGN_PTR adjustments we are now bigger than
1408 		 * UMA_SLAB_SIZE.  I haven't checked whether this is
1409 		 * mathematically possible for all cases, so we make
1410 		 * sure here anyway.
1411 		 */
1412 		totsize = keg->uk_pgoff + sizeof(struct uma_slab);
1413 		if (keg->uk_flags & UMA_ZONE_REFCNT)
1414 			totsize += keg->uk_ipers * sizeof(uint32_t);
1415 		if (totsize > PAGE_SIZE * keg->uk_ppera) {
1416 			printf("zone %s ipers %d rsize %d size %d\n",
1417 			    zone->uz_name, keg->uk_ipers, keg->uk_rsize,
1418 			    keg->uk_size);
1419 			panic("UMA slab won't fit.");
1420 		}
1421 	}
1422 
1423 	if (keg->uk_flags & UMA_ZONE_HASH)
1424 		hash_alloc(&keg->uk_hash);
1425 
1426 #ifdef UMA_DEBUG
1427 	printf("UMA: %s(%p) size %d(%d) flags %#x ipers %d ppera %d out %d free %d\n",
1428 	    zone->uz_name, zone, keg->uk_size, keg->uk_rsize, keg->uk_flags,
1429 	    keg->uk_ipers, keg->uk_ppera,
1430 	    (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free);
1431 #endif
1432 
1433 	LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
1434 
1435 	mtx_lock(&uma_mtx);
1436 	LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
1437 	mtx_unlock(&uma_mtx);
1438 	return (0);
1439 }
1440 
1441 /*
1442  * Zone header ctor.  This initializes all fields, locks, etc.
1443  *
1444  * Arguments/Returns follow uma_ctor specifications
1445  *	udata  Actually uma_zctor_args
1446  */
1447 static int
1448 zone_ctor(void *mem, int size, void *udata, int flags)
1449 {
1450 	struct uma_zctor_args *arg = udata;
1451 	uma_zone_t zone = mem;
1452 	uma_zone_t z;
1453 	uma_keg_t keg;
1454 
1455 	bzero(zone, size);
1456 	zone->uz_name = arg->name;
1457 	zone->uz_ctor = arg->ctor;
1458 	zone->uz_dtor = arg->dtor;
1459 	zone->uz_slab = zone_fetch_slab;
1460 	zone->uz_init = NULL;
1461 	zone->uz_fini = NULL;
1462 	zone->uz_allocs = 0;
1463 	zone->uz_frees = 0;
1464 	zone->uz_fails = 0;
1465 	zone->uz_sleeps = 0;
1466 	zone->uz_count = 0;
1467 	zone->uz_flags = 0;
1468 	zone->uz_warning = NULL;
1469 	timevalclear(&zone->uz_ratecheck);
1470 	keg = arg->keg;
1471 
1472 	/*
1473 	 * This is a pure cache zone, no kegs.
1474 	 */
1475 	if (arg->import) {
1476 		zone->uz_import = arg->import;
1477 		zone->uz_release = arg->release;
1478 		zone->uz_arg = arg->arg;
1479 		zone->uz_count = BUCKET_MAX;
1480 		return (0);
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_lock = &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_lock = &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 	if ((keg->uk_flags & UMA_ZONE_MAXBUCKET) == 0)
1548 		zone->uz_count = bucket_select(keg->uk_rsize);
1549 	else
1550 		zone->uz_count = BUCKET_MAX;
1551 
1552 	return (0);
1553 }
1554 
1555 /*
1556  * Keg header dtor.  This frees all data, destroys locks, frees the hash
1557  * table and removes the keg from the global list.
1558  *
1559  * Arguments/Returns follow uma_dtor specifications
1560  *	udata  unused
1561  */
1562 static void
1563 keg_dtor(void *arg, int size, void *udata)
1564 {
1565 	uma_keg_t keg;
1566 
1567 	keg = (uma_keg_t)arg;
1568 	KEG_LOCK(keg);
1569 	if (keg->uk_free != 0) {
1570 		printf("Freed UMA keg was not empty (%d items). "
1571 		    " Lost %d pages of memory.\n",
1572 		    keg->uk_free, keg->uk_pages);
1573 	}
1574 	KEG_UNLOCK(keg);
1575 
1576 	hash_free(&keg->uk_hash);
1577 
1578 	KEG_LOCK_FINI(keg);
1579 }
1580 
1581 /*
1582  * Zone header dtor.
1583  *
1584  * Arguments/Returns follow uma_dtor specifications
1585  *	udata  unused
1586  */
1587 static void
1588 zone_dtor(void *arg, int size, void *udata)
1589 {
1590 	uma_klink_t klink;
1591 	uma_zone_t zone;
1592 	uma_keg_t keg;
1593 
1594 	zone = (uma_zone_t)arg;
1595 	keg = zone_first_keg(zone);
1596 
1597 	if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
1598 		cache_drain(zone);
1599 
1600 	mtx_lock(&uma_mtx);
1601 	LIST_REMOVE(zone, uz_link);
1602 	mtx_unlock(&uma_mtx);
1603 	/*
1604 	 * XXX there are some races here where
1605 	 * the zone can be drained but zone lock
1606 	 * released and then refilled before we
1607 	 * remove it... we dont care for now
1608 	 */
1609 	zone_drain_wait(zone, M_WAITOK);
1610 	/*
1611 	 * Unlink all of our kegs.
1612 	 */
1613 	while ((klink = LIST_FIRST(&zone->uz_kegs)) != NULL) {
1614 		klink->kl_keg = NULL;
1615 		LIST_REMOVE(klink, kl_link);
1616 		if (klink == &zone->uz_klink)
1617 			continue;
1618 		free(klink, M_TEMP);
1619 	}
1620 	/*
1621 	 * We only destroy kegs from non secondary zones.
1622 	 */
1623 	if (keg != NULL && (zone->uz_flags & UMA_ZONE_SECONDARY) == 0)  {
1624 		mtx_lock(&uma_mtx);
1625 		LIST_REMOVE(keg, uk_link);
1626 		mtx_unlock(&uma_mtx);
1627 		zone_free_item(kegs, keg, NULL, SKIP_NONE);
1628 	}
1629 }
1630 
1631 /*
1632  * Traverses every zone in the system and calls a callback
1633  *
1634  * Arguments:
1635  *	zfunc  A pointer to a function which accepts a zone
1636  *		as an argument.
1637  *
1638  * Returns:
1639  *	Nothing
1640  */
1641 static void
1642 zone_foreach(void (*zfunc)(uma_zone_t))
1643 {
1644 	uma_keg_t keg;
1645 	uma_zone_t zone;
1646 
1647 	mtx_lock(&uma_mtx);
1648 	LIST_FOREACH(keg, &uma_kegs, uk_link) {
1649 		LIST_FOREACH(zone, &keg->uk_zones, uz_link)
1650 			zfunc(zone);
1651 	}
1652 	mtx_unlock(&uma_mtx);
1653 }
1654 
1655 /* Public functions */
1656 /* See uma.h */
1657 void
1658 uma_startup(void *bootmem, int boot_pages)
1659 {
1660 	struct uma_zctor_args args;
1661 	uma_slab_t slab;
1662 	u_int slabsize;
1663 	int i;
1664 
1665 #ifdef UMA_DEBUG
1666 	printf("Creating uma keg headers zone and keg.\n");
1667 #endif
1668 	mtx_init(&uma_mtx, "UMA lock", NULL, MTX_DEF);
1669 
1670 	/* "manually" create the initial zone */
1671 	memset(&args, 0, sizeof(args));
1672 	args.name = "UMA Kegs";
1673 	args.size = sizeof(struct uma_keg);
1674 	args.ctor = keg_ctor;
1675 	args.dtor = keg_dtor;
1676 	args.uminit = zero_init;
1677 	args.fini = NULL;
1678 	args.keg = &masterkeg;
1679 	args.align = 32 - 1;
1680 	args.flags = UMA_ZFLAG_INTERNAL;
1681 	/* The initial zone has no Per cpu queues so it's smaller */
1682 	zone_ctor(kegs, sizeof(struct uma_zone), &args, M_WAITOK);
1683 
1684 #ifdef UMA_DEBUG
1685 	printf("Filling boot free list.\n");
1686 #endif
1687 	for (i = 0; i < boot_pages; i++) {
1688 		slab = (uma_slab_t)((uint8_t *)bootmem + (i * UMA_SLAB_SIZE));
1689 		slab->us_data = (uint8_t *)slab;
1690 		slab->us_flags = UMA_SLAB_BOOT;
1691 		LIST_INSERT_HEAD(&uma_boot_pages, slab, us_link);
1692 	}
1693 	mtx_init(&uma_boot_pages_mtx, "UMA boot pages", NULL, MTX_DEF);
1694 
1695 #ifdef UMA_DEBUG
1696 	printf("Creating uma zone headers zone and keg.\n");
1697 #endif
1698 	args.name = "UMA Zones";
1699 	args.size = sizeof(struct uma_zone) +
1700 	    (sizeof(struct uma_cache) * (mp_maxid + 1));
1701 	args.ctor = zone_ctor;
1702 	args.dtor = zone_dtor;
1703 	args.uminit = zero_init;
1704 	args.fini = NULL;
1705 	args.keg = NULL;
1706 	args.align = 32 - 1;
1707 	args.flags = UMA_ZFLAG_INTERNAL;
1708 	/* The initial zone has no Per cpu queues so it's smaller */
1709 	zone_ctor(zones, sizeof(struct uma_zone), &args, M_WAITOK);
1710 
1711 #ifdef UMA_DEBUG
1712 	printf("Initializing pcpu cache locks.\n");
1713 #endif
1714 #ifdef UMA_DEBUG
1715 	printf("Creating slab and hash zones.\n");
1716 #endif
1717 
1718 	/* Now make a zone for slab headers */
1719 	slabzone = uma_zcreate("UMA Slabs",
1720 				sizeof(struct uma_slab),
1721 				NULL, NULL, NULL, NULL,
1722 				UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1723 
1724 	/*
1725 	 * We also create a zone for the bigger slabs with reference
1726 	 * counts in them, to accomodate UMA_ZONE_REFCNT zones.
1727 	 */
1728 	slabsize = sizeof(struct uma_slab_refcnt);
1729 	slabsize += uma_max_ipers_ref * sizeof(uint32_t);
1730 	slabrefzone = uma_zcreate("UMA RCntSlabs",
1731 				  slabsize,
1732 				  NULL, NULL, NULL, NULL,
1733 				  UMA_ALIGN_PTR,
1734 				  UMA_ZFLAG_INTERNAL);
1735 
1736 	hashzone = uma_zcreate("UMA Hash",
1737 	    sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
1738 	    NULL, NULL, NULL, NULL,
1739 	    UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1740 
1741 	bucket_init();
1742 
1743 	booted = UMA_STARTUP;
1744 
1745 #ifdef UMA_DEBUG
1746 	printf("UMA startup complete.\n");
1747 #endif
1748 }
1749 
1750 /* see uma.h */
1751 void
1752 uma_startup2(void)
1753 {
1754 	booted = UMA_STARTUP2;
1755 	bucket_enable();
1756 #ifdef UMA_DEBUG
1757 	printf("UMA startup2 complete.\n");
1758 #endif
1759 }
1760 
1761 /*
1762  * Initialize our callout handle
1763  *
1764  */
1765 
1766 static void
1767 uma_startup3(void)
1768 {
1769 #ifdef UMA_DEBUG
1770 	printf("Starting callout.\n");
1771 #endif
1772 	callout_init(&uma_callout, CALLOUT_MPSAFE);
1773 	callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
1774 #ifdef UMA_DEBUG
1775 	printf("UMA startup3 complete.\n");
1776 #endif
1777 }
1778 
1779 static uma_keg_t
1780 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
1781 		int align, uint32_t flags)
1782 {
1783 	struct uma_kctor_args args;
1784 
1785 	args.size = size;
1786 	args.uminit = uminit;
1787 	args.fini = fini;
1788 	args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
1789 	args.flags = flags;
1790 	args.zone = zone;
1791 	return (zone_alloc_item(kegs, &args, M_WAITOK));
1792 }
1793 
1794 /* See uma.h */
1795 void
1796 uma_set_align(int align)
1797 {
1798 
1799 	if (align != UMA_ALIGN_CACHE)
1800 		uma_align_cache = align;
1801 }
1802 
1803 /* See uma.h */
1804 uma_zone_t
1805 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
1806 		uma_init uminit, uma_fini fini, int align, uint32_t flags)
1807 
1808 {
1809 	struct uma_zctor_args args;
1810 
1811 	/* This stuff is essential for the zone ctor */
1812 	memset(&args, 0, sizeof(args));
1813 	args.name = name;
1814 	args.size = size;
1815 	args.ctor = ctor;
1816 	args.dtor = dtor;
1817 	args.uminit = uminit;
1818 	args.fini = fini;
1819 	args.align = align;
1820 	args.flags = flags;
1821 	args.keg = NULL;
1822 
1823 	return (zone_alloc_item(zones, &args, M_WAITOK));
1824 }
1825 
1826 /* See uma.h */
1827 uma_zone_t
1828 uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor,
1829 		    uma_init zinit, uma_fini zfini, uma_zone_t master)
1830 {
1831 	struct uma_zctor_args args;
1832 	uma_keg_t keg;
1833 
1834 	keg = zone_first_keg(master);
1835 	memset(&args, 0, sizeof(args));
1836 	args.name = name;
1837 	args.size = keg->uk_size;
1838 	args.ctor = ctor;
1839 	args.dtor = dtor;
1840 	args.uminit = zinit;
1841 	args.fini = zfini;
1842 	args.align = keg->uk_align;
1843 	args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
1844 	args.keg = keg;
1845 
1846 	/* XXX Attaches only one keg of potentially many. */
1847 	return (zone_alloc_item(zones, &args, M_WAITOK));
1848 }
1849 
1850 /* See uma.h */
1851 uma_zone_t
1852 uma_zcache_create(char *name, uma_ctor ctor, uma_dtor dtor, uma_init zinit,
1853 		    uma_fini zfini, uma_import zimport, uma_release zrelease,
1854 		    void *arg, int flags)
1855 {
1856 	struct uma_zctor_args args;
1857 
1858 	memset(&args, 0, sizeof(args));
1859 	args.name = name;
1860 	args.size = 0;
1861 	args.ctor = ctor;
1862 	args.dtor = dtor;
1863 	args.uminit = zinit;
1864 	args.fini = zfini;
1865 	args.import = zimport;
1866 	args.release = zrelease;
1867 	args.arg = arg;
1868 	args.align = 0;
1869 	args.flags = flags;
1870 
1871 	return (zone_alloc_item(zones, &args, M_WAITOK));
1872 }
1873 
1874 static void
1875 zone_lock_pair(uma_zone_t a, uma_zone_t b)
1876 {
1877 	if (a < b) {
1878 		ZONE_LOCK(a);
1879 		mtx_lock_flags(b->uz_lock, MTX_DUPOK);
1880 	} else {
1881 		ZONE_LOCK(b);
1882 		mtx_lock_flags(a->uz_lock, MTX_DUPOK);
1883 	}
1884 }
1885 
1886 static void
1887 zone_unlock_pair(uma_zone_t a, uma_zone_t b)
1888 {
1889 
1890 	ZONE_UNLOCK(a);
1891 	ZONE_UNLOCK(b);
1892 }
1893 
1894 int
1895 uma_zsecond_add(uma_zone_t zone, uma_zone_t master)
1896 {
1897 	uma_klink_t klink;
1898 	uma_klink_t kl;
1899 	int error;
1900 
1901 	error = 0;
1902 	klink = malloc(sizeof(*klink), M_TEMP, M_WAITOK | M_ZERO);
1903 
1904 	zone_lock_pair(zone, master);
1905 	/*
1906 	 * zone must use vtoslab() to resolve objects and must already be
1907 	 * a secondary.
1908 	 */
1909 	if ((zone->uz_flags & (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY))
1910 	    != (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) {
1911 		error = EINVAL;
1912 		goto out;
1913 	}
1914 	/*
1915 	 * The new master must also use vtoslab().
1916 	 */
1917 	if ((zone->uz_flags & UMA_ZONE_VTOSLAB) != UMA_ZONE_VTOSLAB) {
1918 		error = EINVAL;
1919 		goto out;
1920 	}
1921 	/*
1922 	 * Both must either be refcnt, or not be refcnt.
1923 	 */
1924 	if ((zone->uz_flags & UMA_ZONE_REFCNT) !=
1925 	    (master->uz_flags & UMA_ZONE_REFCNT)) {
1926 		error = EINVAL;
1927 		goto out;
1928 	}
1929 	/*
1930 	 * The underlying object must be the same size.  rsize
1931 	 * may be different.
1932 	 */
1933 	if (master->uz_size != zone->uz_size) {
1934 		error = E2BIG;
1935 		goto out;
1936 	}
1937 	/*
1938 	 * Put it at the end of the list.
1939 	 */
1940 	klink->kl_keg = zone_first_keg(master);
1941 	LIST_FOREACH(kl, &zone->uz_kegs, kl_link) {
1942 		if (LIST_NEXT(kl, kl_link) == NULL) {
1943 			LIST_INSERT_AFTER(kl, klink, kl_link);
1944 			break;
1945 		}
1946 	}
1947 	klink = NULL;
1948 	zone->uz_flags |= UMA_ZFLAG_MULTI;
1949 	zone->uz_slab = zone_fetch_slab_multi;
1950 
1951 out:
1952 	zone_unlock_pair(zone, master);
1953 	if (klink != NULL)
1954 		free(klink, M_TEMP);
1955 
1956 	return (error);
1957 }
1958 
1959 
1960 /* See uma.h */
1961 void
1962 uma_zdestroy(uma_zone_t zone)
1963 {
1964 
1965 	zone_free_item(zones, zone, NULL, SKIP_NONE);
1966 }
1967 
1968 /* See uma.h */
1969 void *
1970 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
1971 {
1972 	void *item;
1973 	uma_cache_t cache;
1974 	uma_bucket_t bucket;
1975 	int lockfail;
1976 	int cpu;
1977 
1978 	/* This is the fast path allocation */
1979 #ifdef UMA_DEBUG_ALLOC_1
1980 	printf("Allocating one item from %s(%p)\n", zone->uz_name, zone);
1981 #endif
1982 	CTR3(KTR_UMA, "uma_zalloc_arg thread %x zone %s flags %d", curthread,
1983 	    zone->uz_name, flags);
1984 
1985 	if (flags & M_WAITOK) {
1986 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
1987 		    "uma_zalloc_arg: zone \"%s\"", zone->uz_name);
1988 	}
1989 #ifdef DEBUG_MEMGUARD
1990 	if (memguard_cmp_zone(zone)) {
1991 		item = memguard_alloc(zone->uz_size, flags);
1992 		if (item != NULL) {
1993 			/*
1994 			 * Avoid conflict with the use-after-free
1995 			 * protecting infrastructure from INVARIANTS.
1996 			 */
1997 			if (zone->uz_init != NULL &&
1998 			    zone->uz_init != mtrash_init &&
1999 			    zone->uz_init(item, zone->uz_size, flags) != 0)
2000 				return (NULL);
2001 			if (zone->uz_ctor != NULL &&
2002 			    zone->uz_ctor != mtrash_ctor &&
2003 			    zone->uz_ctor(item, zone->uz_size, udata,
2004 			    flags) != 0) {
2005 			    	zone->uz_fini(item, zone->uz_size);
2006 				return (NULL);
2007 			}
2008 			return (item);
2009 		}
2010 		/* This is unfortunate but should not be fatal. */
2011 	}
2012 #endif
2013 	/*
2014 	 * If possible, allocate from the per-CPU cache.  There are two
2015 	 * requirements for safe access to the per-CPU cache: (1) the thread
2016 	 * accessing the cache must not be preempted or yield during access,
2017 	 * and (2) the thread must not migrate CPUs without switching which
2018 	 * cache it accesses.  We rely on a critical section to prevent
2019 	 * preemption and migration.  We release the critical section in
2020 	 * order to acquire the zone mutex if we are unable to allocate from
2021 	 * the current cache; when we re-acquire the critical section, we
2022 	 * must detect and handle migration if it has occurred.
2023 	 */
2024 	critical_enter();
2025 	cpu = curcpu;
2026 	cache = &zone->uz_cpu[cpu];
2027 
2028 zalloc_start:
2029 	bucket = cache->uc_allocbucket;
2030 	if (bucket != NULL && bucket->ub_cnt > 0) {
2031 		bucket->ub_cnt--;
2032 		item = bucket->ub_bucket[bucket->ub_cnt];
2033 #ifdef INVARIANTS
2034 		bucket->ub_bucket[bucket->ub_cnt] = NULL;
2035 #endif
2036 		KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled."));
2037 		cache->uc_allocs++;
2038 		critical_exit();
2039 		if (zone->uz_ctor != NULL &&
2040 		    zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2041 			atomic_add_long(&zone->uz_fails, 1);
2042 			zone_free_item(zone, item, udata, SKIP_DTOR);
2043 			return (NULL);
2044 		}
2045 #ifdef INVARIANTS
2046 		uma_dbg_alloc(zone, NULL, item);
2047 #endif
2048 		if (flags & M_ZERO)
2049 			bzero(item, zone->uz_size);
2050 		return (item);
2051 	}
2052 
2053 	/*
2054 	 * We have run out of items in our alloc bucket.
2055 	 * See if we can switch with our free bucket.
2056 	 */
2057 	bucket = cache->uc_freebucket;
2058 	if (bucket != NULL && bucket->ub_cnt > 0) {
2059 #ifdef UMA_DEBUG_ALLOC
2060 		printf("uma_zalloc: Swapping empty with alloc.\n");
2061 #endif
2062 		cache->uc_freebucket = cache->uc_allocbucket;
2063 		cache->uc_allocbucket = bucket;
2064 		goto zalloc_start;
2065 	}
2066 
2067 	/*
2068 	 * Discard any empty allocation bucket while we hold no locks.
2069 	 */
2070 	bucket = cache->uc_allocbucket;
2071 	cache->uc_allocbucket = NULL;
2072 	critical_exit();
2073 	if (bucket != NULL)
2074 		bucket_free(bucket);
2075 
2076 	/* Short-circuit for zones without buckets and low memory. */
2077 	if (zone->uz_count == 0 || bucketdisable)
2078 		goto zalloc_item;
2079 
2080 	/*
2081 	 * Attempt to retrieve the item from the per-CPU cache has failed, so
2082 	 * we must go back to the zone.  This requires the zone lock, so we
2083 	 * must drop the critical section, then re-acquire it when we go back
2084 	 * to the cache.  Since the critical section is released, we may be
2085 	 * preempted or migrate.  As such, make sure not to maintain any
2086 	 * thread-local state specific to the cache from prior to releasing
2087 	 * the critical section.
2088 	 */
2089 	lockfail = 0;
2090 	if (ZONE_TRYLOCK(zone) == 0) {
2091 		/* Record contention to size the buckets. */
2092 		ZONE_LOCK(zone);
2093 		lockfail = 1;
2094 	}
2095 	critical_enter();
2096 	cpu = curcpu;
2097 	cache = &zone->uz_cpu[cpu];
2098 
2099 	/*
2100 	 * Since we have locked the zone we may as well send back our stats.
2101 	 */
2102 	atomic_add_long(&zone->uz_allocs, cache->uc_allocs);
2103 	atomic_add_long(&zone->uz_frees, cache->uc_frees);
2104 	cache->uc_allocs = 0;
2105 	cache->uc_frees = 0;
2106 
2107 	/* See if we lost the race to fill the cache. */
2108 	if (cache->uc_allocbucket != NULL) {
2109 		ZONE_UNLOCK(zone);
2110 		goto zalloc_start;
2111 	}
2112 
2113 	/*
2114 	 * Check the zone's cache of buckets.
2115 	 */
2116 	if ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) {
2117 		KASSERT(bucket->ub_cnt != 0,
2118 		    ("uma_zalloc_arg: Returning an empty bucket."));
2119 
2120 		LIST_REMOVE(bucket, ub_link);
2121 		cache->uc_allocbucket = bucket;
2122 		ZONE_UNLOCK(zone);
2123 		goto zalloc_start;
2124 	}
2125 	/* We are no longer associated with this CPU. */
2126 	critical_exit();
2127 
2128 	/*
2129 	 * We bump the uz count when the cache size is insufficient to
2130 	 * handle the working set.
2131 	 */
2132 	if (lockfail && zone->uz_count < BUCKET_MAX && zone->uz_count != 0 &&
2133 	    zone != largebucket)
2134 		zone->uz_count++;
2135 	ZONE_UNLOCK(zone);
2136 
2137 	/*
2138 	 * Now lets just fill a bucket and put it on the free list.  If that
2139 	 * works we'll restart the allocation from the begining and it
2140 	 * will use the just filled bucket.
2141 	 */
2142 	bucket = zone_alloc_bucket(zone, flags);
2143 	if (bucket != NULL) {
2144 		ZONE_LOCK(zone);
2145 		critical_enter();
2146 		cpu = curcpu;
2147 		cache = &zone->uz_cpu[cpu];
2148 		/*
2149 		 * See if we lost the race or were migrated.  Cache the
2150 		 * initialized bucket to make this less likely or claim
2151 		 * the memory directly.
2152 		 */
2153 		if (cache->uc_allocbucket == NULL)
2154 			cache->uc_allocbucket = bucket;
2155 		else
2156 			LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link);
2157 		ZONE_UNLOCK(zone);
2158 		goto zalloc_start;
2159 	}
2160 
2161 	/*
2162 	 * We may not be able to get a bucket so return an actual item.
2163 	 */
2164 #ifdef UMA_DEBUG
2165 	printf("uma_zalloc_arg: Bucketzone returned NULL\n");
2166 #endif
2167 
2168 zalloc_item:
2169 	item = zone_alloc_item(zone, udata, flags);
2170 
2171 	return (item);
2172 }
2173 
2174 static uma_slab_t
2175 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int flags)
2176 {
2177 	uma_slab_t slab;
2178 
2179 	mtx_assert(&keg->uk_lock, MA_OWNED);
2180 	slab = NULL;
2181 
2182 	for (;;) {
2183 		/*
2184 		 * Find a slab with some space.  Prefer slabs that are partially
2185 		 * used over those that are totally full.  This helps to reduce
2186 		 * fragmentation.
2187 		 */
2188 		if (keg->uk_free != 0) {
2189 			if (!LIST_EMPTY(&keg->uk_part_slab)) {
2190 				slab = LIST_FIRST(&keg->uk_part_slab);
2191 			} else {
2192 				slab = LIST_FIRST(&keg->uk_free_slab);
2193 				LIST_REMOVE(slab, us_link);
2194 				LIST_INSERT_HEAD(&keg->uk_part_slab, slab,
2195 				    us_link);
2196 			}
2197 			MPASS(slab->us_keg == keg);
2198 			return (slab);
2199 		}
2200 
2201 		/*
2202 		 * M_NOVM means don't ask at all!
2203 		 */
2204 		if (flags & M_NOVM)
2205 			break;
2206 
2207 		if (keg->uk_maxpages && keg->uk_pages >= keg->uk_maxpages) {
2208 			keg->uk_flags |= UMA_ZFLAG_FULL;
2209 			/*
2210 			 * If this is not a multi-zone, set the FULL bit.
2211 			 * Otherwise slab_multi() takes care of it.
2212 			 */
2213 			if ((zone->uz_flags & UMA_ZFLAG_MULTI) == 0) {
2214 				zone->uz_flags |= UMA_ZFLAG_FULL;
2215 				zone_log_warning(zone);
2216 			}
2217 			if (flags & M_NOWAIT)
2218 				break;
2219 			zone->uz_sleeps++;
2220 			msleep(keg, &keg->uk_lock, PVM, "keglimit", 0);
2221 			continue;
2222 		}
2223 		slab = keg_alloc_slab(keg, zone, flags);
2224 		/*
2225 		 * If we got a slab here it's safe to mark it partially used
2226 		 * and return.  We assume that the caller is going to remove
2227 		 * at least one item.
2228 		 */
2229 		if (slab) {
2230 			MPASS(slab->us_keg == keg);
2231 			LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2232 			return (slab);
2233 		}
2234 		/*
2235 		 * We might not have been able to get a slab but another cpu
2236 		 * could have while we were unlocked.  Check again before we
2237 		 * fail.
2238 		 */
2239 		flags |= M_NOVM;
2240 	}
2241 	return (slab);
2242 }
2243 
2244 static inline void
2245 zone_relock(uma_zone_t zone, uma_keg_t keg)
2246 {
2247 	if (zone->uz_lock != &keg->uk_lock) {
2248 		KEG_UNLOCK(keg);
2249 		ZONE_LOCK(zone);
2250 	}
2251 }
2252 
2253 static inline void
2254 keg_relock(uma_keg_t keg, uma_zone_t zone)
2255 {
2256 	if (zone->uz_lock != &keg->uk_lock) {
2257 		ZONE_UNLOCK(zone);
2258 		KEG_LOCK(keg);
2259 	}
2260 }
2261 
2262 static uma_slab_t
2263 zone_fetch_slab(uma_zone_t zone, uma_keg_t keg, int flags)
2264 {
2265 	uma_slab_t slab;
2266 
2267 	if (keg == NULL)
2268 		keg = zone_first_keg(zone);
2269 
2270 	for (;;) {
2271 		slab = keg_fetch_slab(keg, zone, flags);
2272 		if (slab)
2273 			return (slab);
2274 		if (flags & (M_NOWAIT | M_NOVM))
2275 			break;
2276 	}
2277 	return (NULL);
2278 }
2279 
2280 /*
2281  * uma_zone_fetch_slab_multi:  Fetches a slab from one available keg.  Returns
2282  * with the keg locked.  Caller must call zone_relock() afterwards if the
2283  * zone lock is required.  On NULL the zone lock is held.
2284  *
2285  * The last pointer is used to seed the search.  It is not required.
2286  */
2287 static uma_slab_t
2288 zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int rflags)
2289 {
2290 	uma_klink_t klink;
2291 	uma_slab_t slab;
2292 	uma_keg_t keg;
2293 	int flags;
2294 	int empty;
2295 	int full;
2296 
2297 	/*
2298 	 * Don't wait on the first pass.  This will skip limit tests
2299 	 * as well.  We don't want to block if we can find a provider
2300 	 * without blocking.
2301 	 */
2302 	flags = (rflags & ~M_WAITOK) | M_NOWAIT;
2303 	/*
2304 	 * Use the last slab allocated as a hint for where to start
2305 	 * the search.
2306 	 */
2307 	if (last) {
2308 		slab = keg_fetch_slab(last, zone, flags);
2309 		if (slab)
2310 			return (slab);
2311 		zone_relock(zone, last);
2312 		last = NULL;
2313 	}
2314 	/*
2315 	 * Loop until we have a slab incase of transient failures
2316 	 * while M_WAITOK is specified.  I'm not sure this is 100%
2317 	 * required but we've done it for so long now.
2318 	 */
2319 	for (;;) {
2320 		empty = 0;
2321 		full = 0;
2322 		/*
2323 		 * Search the available kegs for slabs.  Be careful to hold the
2324 		 * correct lock while calling into the keg layer.
2325 		 */
2326 		LIST_FOREACH(klink, &zone->uz_kegs, kl_link) {
2327 			keg = klink->kl_keg;
2328 			keg_relock(keg, zone);
2329 			if ((keg->uk_flags & UMA_ZFLAG_FULL) == 0) {
2330 				slab = keg_fetch_slab(keg, zone, flags);
2331 				if (slab)
2332 					return (slab);
2333 			}
2334 			if (keg->uk_flags & UMA_ZFLAG_FULL)
2335 				full++;
2336 			else
2337 				empty++;
2338 			zone_relock(zone, keg);
2339 		}
2340 		if (rflags & (M_NOWAIT | M_NOVM))
2341 			break;
2342 		flags = rflags;
2343 		/*
2344 		 * All kegs are full.  XXX We can't atomically check all kegs
2345 		 * and sleep so just sleep for a short period and retry.
2346 		 */
2347 		if (full && !empty) {
2348 			zone->uz_flags |= UMA_ZFLAG_FULL;
2349 			zone->uz_sleeps++;
2350 			zone_log_warning(zone);
2351 			msleep(zone, zone->uz_lock, PVM, "zonelimit", hz/100);
2352 			zone->uz_flags &= ~UMA_ZFLAG_FULL;
2353 			continue;
2354 		}
2355 	}
2356 	return (NULL);
2357 }
2358 
2359 static void *
2360 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
2361 {
2362 	void *item;
2363 	uint8_t freei;
2364 
2365 	MPASS(keg == slab->us_keg);
2366 	mtx_assert(&keg->uk_lock, MA_OWNED);
2367 
2368 	freei = BIT_FFS(SLAB_SETSIZE, &slab->us_free) - 1;
2369 	BIT_CLR(SLAB_SETSIZE, freei, &slab->us_free);
2370 	item = slab->us_data + (keg->uk_rsize * freei);
2371 	slab->us_freecount--;
2372 	keg->uk_free--;
2373 
2374 	/* Move this slab to the full list */
2375 	if (slab->us_freecount == 0) {
2376 		LIST_REMOVE(slab, us_link);
2377 		LIST_INSERT_HEAD(&keg->uk_full_slab, slab, us_link);
2378 	}
2379 
2380 	return (item);
2381 }
2382 
2383 static int
2384 zone_import(uma_zone_t zone, void **bucket, int max, int flags)
2385 {
2386 	uma_slab_t slab;
2387 	uma_keg_t keg;
2388 	int i;
2389 
2390 	ZONE_LOCK(zone);
2391 	/* Try to keep the buckets totally full */
2392 	slab = NULL;
2393 	keg = NULL;
2394 	for (i = 0; i < max; ) {
2395 		if ((slab = zone->uz_slab(zone, keg, flags)) == NULL)
2396 			break;
2397 		keg = slab->us_keg;
2398 		while (slab->us_freecount && i < max)
2399 			bucket[i++] = slab_alloc_item(keg, slab);
2400 
2401 		/* Don't block on the next fill */
2402 		flags &= ~M_WAITOK;
2403 		flags |= M_NOWAIT;
2404 	}
2405 	if (slab != NULL)
2406 		KEG_UNLOCK(keg);
2407 	else
2408 		ZONE_UNLOCK(zone);
2409 
2410 	return i;
2411 }
2412 
2413 static uma_bucket_t
2414 zone_alloc_bucket(uma_zone_t zone, int flags)
2415 {
2416 	uma_bucket_t bucket;
2417 	int bflags;
2418 	int max;
2419 
2420 	max = zone->uz_count;
2421 	bflags = (flags & ~M_WAITOK) | M_NOWAIT;
2422 	if (zone->uz_flags & UMA_ZFLAG_CACHEONLY)
2423 		bflags |= M_NOVM;
2424 	bucket = bucket_alloc(zone->uz_count, bflags);
2425 	if (bucket == NULL)
2426 		goto out;
2427 
2428 	max = MIN(bucket->ub_entries, max);
2429 	bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
2430 	    max, flags);
2431 
2432 	/*
2433 	 * Initialize the memory if necessary.
2434 	 */
2435 	if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
2436 		int i;
2437 
2438 		for (i = 0; i < bucket->ub_cnt; i++)
2439 			if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size,
2440 			    flags) != 0)
2441 				break;
2442 		/*
2443 		 * If we couldn't initialize the whole bucket, put the
2444 		 * rest back onto the freelist.
2445 		 */
2446 		if (i != bucket->ub_cnt) {
2447 			zone->uz_release(zone->uz_arg, bucket->ub_bucket[i],
2448 			    bucket->ub_cnt - i);
2449 #ifdef INVARIANTS
2450 			bzero(&bucket->ub_bucket[i],
2451 			    sizeof(void *) * (bucket->ub_cnt - i));
2452 #endif
2453 			bucket->ub_cnt = i;
2454 		}
2455 	}
2456 
2457 out:
2458 	if (bucket == NULL || bucket->ub_cnt == 0) {
2459 		if (bucket != NULL)
2460 			bucket_free(bucket);
2461 		atomic_add_long(&zone->uz_fails, 1);
2462 		return (NULL);
2463 	}
2464 
2465 	return (bucket);
2466 }
2467 
2468 /*
2469  * Allocates a single item from a zone.
2470  *
2471  * Arguments
2472  *	zone   The zone to alloc for.
2473  *	udata  The data to be passed to the constructor.
2474  *	flags  M_WAITOK, M_NOWAIT, M_ZERO.
2475  *
2476  * Returns
2477  *	NULL if there is no memory and M_NOWAIT is set
2478  *	An item if successful
2479  */
2480 
2481 static void *
2482 zone_alloc_item(uma_zone_t zone, void *udata, int flags)
2483 {
2484 	void *item;
2485 
2486 	item = NULL;
2487 
2488 #ifdef UMA_DEBUG_ALLOC
2489 	printf("INTERNAL: Allocating one item from %s(%p)\n", zone->uz_name, zone);
2490 #endif
2491 	if (zone->uz_import(zone->uz_arg, &item, 1, flags) != 1)
2492 		goto fail;
2493 	atomic_add_long(&zone->uz_allocs, 1);
2494 
2495 	/*
2496 	 * We have to call both the zone's init (not the keg's init)
2497 	 * and the zone's ctor.  This is because the item is going from
2498 	 * a keg slab directly to the user, and the user is expecting it
2499 	 * to be both zone-init'd as well as zone-ctor'd.
2500 	 */
2501 	if (zone->uz_init != NULL) {
2502 		if (zone->uz_init(item, zone->uz_size, flags) != 0) {
2503 			zone_free_item(zone, item, udata, SKIP_FINI);
2504 			goto fail;
2505 		}
2506 	}
2507 	if (zone->uz_ctor != NULL) {
2508 		if (zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2509 			zone_free_item(zone, item, udata, SKIP_DTOR);
2510 			goto fail;
2511 		}
2512 	}
2513 #ifdef INVARIANTS
2514 	uma_dbg_alloc(zone, NULL, item);
2515 #endif
2516 	if (flags & M_ZERO)
2517 		bzero(item, zone->uz_size);
2518 
2519 	return (item);
2520 
2521 fail:
2522 	atomic_add_long(&zone->uz_fails, 1);
2523 	return (NULL);
2524 }
2525 
2526 /* See uma.h */
2527 void
2528 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
2529 {
2530 	uma_cache_t cache;
2531 	uma_bucket_t bucket;
2532 	int bflags;
2533 	int cpu;
2534 
2535 #ifdef UMA_DEBUG_ALLOC_1
2536 	printf("Freeing item %p to %s(%p)\n", item, zone->uz_name, zone);
2537 #endif
2538 	CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread,
2539 	    zone->uz_name);
2540 
2541         /* uma_zfree(..., NULL) does nothing, to match free(9). */
2542         if (item == NULL)
2543                 return;
2544 #ifdef DEBUG_MEMGUARD
2545 	if (is_memguard_addr(item)) {
2546 		if (zone->uz_dtor != NULL && zone->uz_dtor != mtrash_dtor)
2547 			zone->uz_dtor(item, zone->uz_size, udata);
2548 		if (zone->uz_fini != NULL && zone->uz_fini != mtrash_fini)
2549 			zone->uz_fini(item, zone->uz_size);
2550 		memguard_free(item);
2551 		return;
2552 	}
2553 #endif
2554 #ifdef INVARIANTS
2555 	if (zone->uz_flags & UMA_ZONE_MALLOC)
2556 		uma_dbg_free(zone, udata, item);
2557 	else
2558 		uma_dbg_free(zone, NULL, item);
2559 #endif
2560 	if (zone->uz_dtor != NULL)
2561 		zone->uz_dtor(item, zone->uz_size, udata);
2562 
2563 	/*
2564 	 * The race here is acceptable.  If we miss it we'll just have to wait
2565 	 * a little longer for the limits to be reset.
2566 	 */
2567 	if (zone->uz_flags & UMA_ZFLAG_FULL)
2568 		goto zfree_item;
2569 
2570 	/*
2571 	 * If possible, free to the per-CPU cache.  There are two
2572 	 * requirements for safe access to the per-CPU cache: (1) the thread
2573 	 * accessing the cache must not be preempted or yield during access,
2574 	 * and (2) the thread must not migrate CPUs without switching which
2575 	 * cache it accesses.  We rely on a critical section to prevent
2576 	 * preemption and migration.  We release the critical section in
2577 	 * order to acquire the zone mutex if we are unable to free to the
2578 	 * current cache; when we re-acquire the critical section, we must
2579 	 * detect and handle migration if it has occurred.
2580 	 */
2581 zfree_restart:
2582 	critical_enter();
2583 	cpu = curcpu;
2584 	cache = &zone->uz_cpu[cpu];
2585 
2586 zfree_start:
2587 	/*
2588 	 * Try to free into the allocbucket first to give LIFO ordering
2589 	 * for cache-hot datastructures.  Spill over into the freebucket
2590 	 * if necessary.  Alloc will swap them if one runs dry.
2591 	 */
2592 	bucket = cache->uc_allocbucket;
2593 	if (bucket == NULL || bucket->ub_cnt >= bucket->ub_entries)
2594 		bucket = cache->uc_freebucket;
2595 	if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) {
2596 		KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL,
2597 		    ("uma_zfree: Freeing to non free bucket index."));
2598 		bucket->ub_bucket[bucket->ub_cnt] = item;
2599 		bucket->ub_cnt++;
2600 		cache->uc_frees++;
2601 		critical_exit();
2602 		return;
2603 	}
2604 
2605 	/*
2606 	 * We must go back the zone, which requires acquiring the zone lock,
2607 	 * which in turn means we must release and re-acquire the critical
2608 	 * section.  Since the critical section is released, we may be
2609 	 * preempted or migrate.  As such, make sure not to maintain any
2610 	 * thread-local state specific to the cache from prior to releasing
2611 	 * the critical section.
2612 	 */
2613 	critical_exit();
2614 	if (zone->uz_count == 0 || bucketdisable)
2615 		goto zfree_item;
2616 
2617 	ZONE_LOCK(zone);
2618 	critical_enter();
2619 	cpu = curcpu;
2620 	cache = &zone->uz_cpu[cpu];
2621 
2622 	/*
2623 	 * Since we have locked the zone we may as well send back our stats.
2624 	 */
2625 	atomic_add_long(&zone->uz_allocs, cache->uc_allocs);
2626 	atomic_add_long(&zone->uz_frees, cache->uc_frees);
2627 	cache->uc_allocs = 0;
2628 	cache->uc_frees = 0;
2629 
2630 	bucket = cache->uc_freebucket;
2631 	if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) {
2632 		ZONE_UNLOCK(zone);
2633 		goto zfree_start;
2634 	}
2635 	cache->uc_freebucket = NULL;
2636 
2637 	/* Can we throw this on the zone full list? */
2638 	if (bucket != NULL) {
2639 #ifdef UMA_DEBUG_ALLOC
2640 		printf("uma_zfree: Putting old bucket on the free list.\n");
2641 #endif
2642 		/* ub_cnt is pointing to the last free item */
2643 		KASSERT(bucket->ub_cnt != 0,
2644 		    ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n"));
2645 		LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link);
2646 	}
2647 
2648 	/* We are no longer associated with this CPU. */
2649 	critical_exit();
2650 
2651 	/* And the zone.. */
2652 	ZONE_UNLOCK(zone);
2653 
2654 #ifdef UMA_DEBUG_ALLOC
2655 	printf("uma_zfree: Allocating new free bucket.\n");
2656 #endif
2657 	bflags = M_NOWAIT;
2658 	if (zone->uz_flags & UMA_ZFLAG_CACHEONLY)
2659 		bflags |= M_NOVM;
2660 	bucket = bucket_alloc(zone->uz_count, bflags);
2661 	if (bucket) {
2662 		critical_enter();
2663 		cpu = curcpu;
2664 		cache = &zone->uz_cpu[cpu];
2665 		if (cache->uc_freebucket == NULL) {
2666 			cache->uc_freebucket = bucket;
2667 			goto zfree_start;
2668 		}
2669 		/*
2670 		 * We lost the race, start over.  We have to drop our
2671 		 * critical section to free the bucket.
2672 		 */
2673 		critical_exit();
2674 		bucket_free(bucket);
2675 		goto zfree_restart;
2676 	}
2677 
2678 	/*
2679 	 * If nothing else caught this, we'll just do an internal free.
2680 	 */
2681 zfree_item:
2682 	zone_free_item(zone, item, udata, SKIP_DTOR);
2683 
2684 	return;
2685 }
2686 
2687 static void
2688 slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item)
2689 {
2690 	uint8_t freei;
2691 
2692 	mtx_assert(&keg->uk_lock, MA_OWNED);
2693 	MPASS(keg == slab->us_keg);
2694 
2695 	/* Do we need to remove from any lists? */
2696 	if (slab->us_freecount+1 == keg->uk_ipers) {
2697 		LIST_REMOVE(slab, us_link);
2698 		LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
2699 	} else if (slab->us_freecount == 0) {
2700 		LIST_REMOVE(slab, us_link);
2701 		LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2702 	}
2703 
2704 	/* Slab management. */
2705 	freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
2706 	BIT_SET(SLAB_SETSIZE, freei, &slab->us_free);
2707 	slab->us_freecount++;
2708 
2709 	/* Keg statistics. */
2710 	keg->uk_free++;
2711 }
2712 
2713 static void
2714 zone_release(uma_zone_t zone, void **bucket, int cnt)
2715 {
2716 	void *item;
2717 	uma_slab_t slab;
2718 	uma_keg_t keg;
2719 	uint8_t *mem;
2720 	int clearfull;
2721 	int i;
2722 
2723 	clearfull = 0;
2724 	ZONE_LOCK(zone);
2725 	keg = zone_first_keg(zone);
2726 	for (i = 0; i < cnt; i++) {
2727 		item = bucket[i];
2728 		if (!(zone->uz_flags & UMA_ZONE_VTOSLAB)) {
2729 			mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
2730 			if (zone->uz_flags & UMA_ZONE_HASH) {
2731 				slab = hash_sfind(&keg->uk_hash, mem);
2732 			} else {
2733 				mem += keg->uk_pgoff;
2734 				slab = (uma_slab_t)mem;
2735 			}
2736 		} else {
2737 			slab = vtoslab((vm_offset_t)item);
2738 			if (slab->us_keg != keg) {
2739 				KEG_UNLOCK(keg);
2740 				keg = slab->us_keg;
2741 				KEG_LOCK(keg);
2742 			}
2743 		}
2744 		slab_free_item(keg, slab, item);
2745 		if (keg->uk_flags & UMA_ZFLAG_FULL) {
2746 			if (keg->uk_pages < keg->uk_maxpages) {
2747 				keg->uk_flags &= ~UMA_ZFLAG_FULL;
2748 				clearfull = 1;
2749 			}
2750 
2751 			/*
2752 			 * We can handle one more allocation. Since we're
2753 			 * clearing ZFLAG_FULL, wake up all procs blocked
2754 			 * on pages. This should be uncommon, so keeping this
2755 			 * simple for now (rather than adding count of blocked
2756 			 * threads etc).
2757 			 */
2758 			wakeup(keg);
2759 		}
2760 	}
2761 	zone_relock(zone, keg);
2762 	if (clearfull) {
2763 		zone->uz_flags &= ~UMA_ZFLAG_FULL;
2764 		wakeup(zone);
2765 	}
2766 	ZONE_UNLOCK(zone);
2767 
2768 }
2769 
2770 /*
2771  * Frees a single item to any zone.
2772  *
2773  * Arguments:
2774  *	zone   The zone to free to
2775  *	item   The item we're freeing
2776  *	udata  User supplied data for the dtor
2777  *	skip   Skip dtors and finis
2778  */
2779 static void
2780 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
2781 {
2782 
2783 #ifdef INVARIANTS
2784 	if (skip == SKIP_NONE) {
2785 		if (zone->uz_flags & UMA_ZONE_MALLOC)
2786 			uma_dbg_free(zone, udata, item);
2787 		else
2788 			uma_dbg_free(zone, NULL, item);
2789 	}
2790 #endif
2791 	if (skip < SKIP_DTOR && zone->uz_dtor)
2792 		zone->uz_dtor(item, zone->uz_size, udata);
2793 
2794 	if (skip < SKIP_FINI && zone->uz_fini)
2795 		zone->uz_fini(item, zone->uz_size);
2796 
2797 	atomic_add_long(&zone->uz_frees, 1);
2798 	zone->uz_release(zone->uz_arg, &item, 1);
2799 }
2800 
2801 /* See uma.h */
2802 int
2803 uma_zone_set_max(uma_zone_t zone, int nitems)
2804 {
2805 	uma_keg_t keg;
2806 
2807 	keg = zone_first_keg(zone);
2808 	if (keg == NULL)
2809 		return (0);
2810 	ZONE_LOCK(zone);
2811 	keg->uk_maxpages = (nitems / keg->uk_ipers) * keg->uk_ppera;
2812 	if (keg->uk_maxpages * keg->uk_ipers < nitems)
2813 		keg->uk_maxpages += keg->uk_ppera;
2814 	nitems = keg->uk_maxpages * keg->uk_ipers;
2815 	ZONE_UNLOCK(zone);
2816 
2817 	return (nitems);
2818 }
2819 
2820 /* See uma.h */
2821 int
2822 uma_zone_get_max(uma_zone_t zone)
2823 {
2824 	int nitems;
2825 	uma_keg_t keg;
2826 
2827 	keg = zone_first_keg(zone);
2828 	if (keg == NULL)
2829 		return (0);
2830 	ZONE_LOCK(zone);
2831 	nitems = keg->uk_maxpages * keg->uk_ipers;
2832 	ZONE_UNLOCK(zone);
2833 
2834 	return (nitems);
2835 }
2836 
2837 /* See uma.h */
2838 void
2839 uma_zone_set_warning(uma_zone_t zone, const char *warning)
2840 {
2841 
2842 	ZONE_LOCK(zone);
2843 	zone->uz_warning = warning;
2844 	ZONE_UNLOCK(zone);
2845 }
2846 
2847 /* See uma.h */
2848 int
2849 uma_zone_get_cur(uma_zone_t zone)
2850 {
2851 	int64_t nitems;
2852 	u_int i;
2853 
2854 	ZONE_LOCK(zone);
2855 	nitems = zone->uz_allocs - zone->uz_frees;
2856 	CPU_FOREACH(i) {
2857 		/*
2858 		 * See the comment in sysctl_vm_zone_stats() regarding the
2859 		 * safety of accessing the per-cpu caches. With the zone lock
2860 		 * held, it is safe, but can potentially result in stale data.
2861 		 */
2862 		nitems += zone->uz_cpu[i].uc_allocs -
2863 		    zone->uz_cpu[i].uc_frees;
2864 	}
2865 	ZONE_UNLOCK(zone);
2866 
2867 	return (nitems < 0 ? 0 : nitems);
2868 }
2869 
2870 /* See uma.h */
2871 void
2872 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
2873 {
2874 	uma_keg_t keg;
2875 
2876 	ZONE_LOCK(zone);
2877 	keg = zone_first_keg(zone);
2878 	KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type"));
2879 	KASSERT(keg->uk_pages == 0,
2880 	    ("uma_zone_set_init on non-empty keg"));
2881 	keg->uk_init = uminit;
2882 	ZONE_UNLOCK(zone);
2883 }
2884 
2885 /* See uma.h */
2886 void
2887 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
2888 {
2889 	uma_keg_t keg;
2890 
2891 	ZONE_LOCK(zone);
2892 	keg = zone_first_keg(zone);
2893 	KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type"));
2894 	KASSERT(keg->uk_pages == 0,
2895 	    ("uma_zone_set_fini on non-empty keg"));
2896 	keg->uk_fini = fini;
2897 	ZONE_UNLOCK(zone);
2898 }
2899 
2900 /* See uma.h */
2901 void
2902 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
2903 {
2904 	ZONE_LOCK(zone);
2905 	KASSERT(zone_first_keg(zone)->uk_pages == 0,
2906 	    ("uma_zone_set_zinit on non-empty keg"));
2907 	zone->uz_init = zinit;
2908 	ZONE_UNLOCK(zone);
2909 }
2910 
2911 /* See uma.h */
2912 void
2913 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
2914 {
2915 	ZONE_LOCK(zone);
2916 	KASSERT(zone_first_keg(zone)->uk_pages == 0,
2917 	    ("uma_zone_set_zfini on non-empty keg"));
2918 	zone->uz_fini = zfini;
2919 	ZONE_UNLOCK(zone);
2920 }
2921 
2922 /* See uma.h */
2923 /* XXX uk_freef is not actually used with the zone locked */
2924 void
2925 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
2926 {
2927 	uma_keg_t keg;
2928 
2929 	ZONE_LOCK(zone);
2930 	keg = zone_first_keg(zone);
2931 	KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type"));
2932 	keg->uk_freef = freef;
2933 	ZONE_UNLOCK(zone);
2934 }
2935 
2936 /* See uma.h */
2937 /* XXX uk_allocf is not actually used with the zone locked */
2938 void
2939 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
2940 {
2941 	uma_keg_t keg;
2942 
2943 	ZONE_LOCK(zone);
2944 	keg = zone_first_keg(zone);
2945 	keg->uk_flags |= UMA_ZFLAG_PRIVALLOC;
2946 	keg->uk_allocf = allocf;
2947 	ZONE_UNLOCK(zone);
2948 }
2949 
2950 /* See uma.h */
2951 int
2952 uma_zone_reserve_kva(uma_zone_t zone, int count)
2953 {
2954 	uma_keg_t keg;
2955 	vm_offset_t kva;
2956 	int pages;
2957 
2958 	keg = zone_first_keg(zone);
2959 	if (keg == NULL)
2960 		return (0);
2961 	pages = count / keg->uk_ipers;
2962 
2963 	if (pages * keg->uk_ipers < count)
2964 		pages++;
2965 
2966 #ifdef UMA_MD_SMALL_ALLOC
2967 	if (keg->uk_ppera > 1) {
2968 #else
2969 	if (1) {
2970 #endif
2971 		kva = kmem_alloc_nofault(kernel_map, pages * UMA_SLAB_SIZE);
2972 		if (kva == 0)
2973 			return (0);
2974 	} else
2975 		kva = 0;
2976 	ZONE_LOCK(zone);
2977 	keg->uk_kva = kva;
2978 	keg->uk_offset = 0;
2979 	keg->uk_maxpages = pages;
2980 #ifdef UMA_MD_SMALL_ALLOC
2981 	keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
2982 #else
2983 	keg->uk_allocf = noobj_alloc;
2984 #endif
2985 	keg->uk_flags |= UMA_ZONE_NOFREE | UMA_ZFLAG_PRIVALLOC;
2986 	ZONE_UNLOCK(zone);
2987 	return (1);
2988 }
2989 
2990 /* See uma.h */
2991 void
2992 uma_prealloc(uma_zone_t zone, int items)
2993 {
2994 	int slabs;
2995 	uma_slab_t slab;
2996 	uma_keg_t keg;
2997 
2998 	keg = zone_first_keg(zone);
2999 	if (keg == NULL)
3000 		return;
3001 	ZONE_LOCK(zone);
3002 	slabs = items / keg->uk_ipers;
3003 	if (slabs * keg->uk_ipers < items)
3004 		slabs++;
3005 	while (slabs > 0) {
3006 		slab = keg_alloc_slab(keg, zone, M_WAITOK);
3007 		if (slab == NULL)
3008 			break;
3009 		MPASS(slab->us_keg == keg);
3010 		LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
3011 		slabs--;
3012 	}
3013 	ZONE_UNLOCK(zone);
3014 }
3015 
3016 /* See uma.h */
3017 uint32_t *
3018 uma_find_refcnt(uma_zone_t zone, void *item)
3019 {
3020 	uma_slabrefcnt_t slabref;
3021 	uma_slab_t slab;
3022 	uma_keg_t keg;
3023 	uint32_t *refcnt;
3024 	int idx;
3025 
3026 	slab = vtoslab((vm_offset_t)item & (~UMA_SLAB_MASK));
3027 	slabref = (uma_slabrefcnt_t)slab;
3028 	keg = slab->us_keg;
3029 	KASSERT(keg->uk_flags & UMA_ZONE_REFCNT,
3030 	    ("uma_find_refcnt(): zone possibly not UMA_ZONE_REFCNT"));
3031 	idx = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
3032 	refcnt = &slabref->us_refcnt[idx];
3033 	return refcnt;
3034 }
3035 
3036 /* See uma.h */
3037 void
3038 uma_reclaim(void)
3039 {
3040 #ifdef UMA_DEBUG
3041 	printf("UMA: vm asked us to release pages!\n");
3042 #endif
3043 	bucket_enable();
3044 	zone_foreach(zone_drain);
3045 	/*
3046 	 * Some slabs may have been freed but this zone will be visited early
3047 	 * we visit again so that we can free pages that are empty once other
3048 	 * zones are drained.  We have to do the same for buckets.
3049 	 */
3050 	zone_drain(slabzone);
3051 	zone_drain(slabrefzone);
3052 	bucket_zone_drain();
3053 }
3054 
3055 /* See uma.h */
3056 int
3057 uma_zone_exhausted(uma_zone_t zone)
3058 {
3059 	int full;
3060 
3061 	ZONE_LOCK(zone);
3062 	full = (zone->uz_flags & UMA_ZFLAG_FULL);
3063 	ZONE_UNLOCK(zone);
3064 	return (full);
3065 }
3066 
3067 int
3068 uma_zone_exhausted_nolock(uma_zone_t zone)
3069 {
3070 	return (zone->uz_flags & UMA_ZFLAG_FULL);
3071 }
3072 
3073 void *
3074 uma_large_malloc(int size, int wait)
3075 {
3076 	void *mem;
3077 	uma_slab_t slab;
3078 	uint8_t flags;
3079 
3080 	slab = zone_alloc_item(slabzone, NULL, wait);
3081 	if (slab == NULL)
3082 		return (NULL);
3083 	mem = page_alloc(NULL, size, &flags, wait);
3084 	if (mem) {
3085 		vsetslab((vm_offset_t)mem, slab);
3086 		slab->us_data = mem;
3087 		slab->us_flags = flags | UMA_SLAB_MALLOC;
3088 		slab->us_size = size;
3089 	} else {
3090 		zone_free_item(slabzone, slab, NULL, SKIP_NONE);
3091 	}
3092 
3093 	return (mem);
3094 }
3095 
3096 void
3097 uma_large_free(uma_slab_t slab)
3098 {
3099 	vsetobj((vm_offset_t)slab->us_data, kmem_object);
3100 	page_free(slab->us_data, slab->us_size, slab->us_flags);
3101 	zone_free_item(slabzone, slab, NULL, SKIP_NONE);
3102 }
3103 
3104 void
3105 uma_print_stats(void)
3106 {
3107 	zone_foreach(uma_print_zone);
3108 }
3109 
3110 static void
3111 slab_print(uma_slab_t slab)
3112 {
3113 	printf("slab: keg %p, data %p, freecount %d\n",
3114 		slab->us_keg, slab->us_data, slab->us_freecount);
3115 }
3116 
3117 static void
3118 cache_print(uma_cache_t cache)
3119 {
3120 	printf("alloc: %p(%d), free: %p(%d)\n",
3121 		cache->uc_allocbucket,
3122 		cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0,
3123 		cache->uc_freebucket,
3124 		cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0);
3125 }
3126 
3127 static void
3128 uma_print_keg(uma_keg_t keg)
3129 {
3130 	uma_slab_t slab;
3131 
3132 	printf("keg: %s(%p) size %d(%d) flags %#x ipers %d ppera %d "
3133 	    "out %d free %d limit %d\n",
3134 	    keg->uk_name, keg, keg->uk_size, keg->uk_rsize, keg->uk_flags,
3135 	    keg->uk_ipers, keg->uk_ppera,
3136 	    (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free,
3137 	    (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers);
3138 	printf("Part slabs:\n");
3139 	LIST_FOREACH(slab, &keg->uk_part_slab, us_link)
3140 		slab_print(slab);
3141 	printf("Free slabs:\n");
3142 	LIST_FOREACH(slab, &keg->uk_free_slab, us_link)
3143 		slab_print(slab);
3144 	printf("Full slabs:\n");
3145 	LIST_FOREACH(slab, &keg->uk_full_slab, us_link)
3146 		slab_print(slab);
3147 }
3148 
3149 void
3150 uma_print_zone(uma_zone_t zone)
3151 {
3152 	uma_cache_t cache;
3153 	uma_klink_t kl;
3154 	int i;
3155 
3156 	printf("zone: %s(%p) size %d flags %#x\n",
3157 	    zone->uz_name, zone, zone->uz_size, zone->uz_flags);
3158 	LIST_FOREACH(kl, &zone->uz_kegs, kl_link)
3159 		uma_print_keg(kl->kl_keg);
3160 	CPU_FOREACH(i) {
3161 		cache = &zone->uz_cpu[i];
3162 		printf("CPU %d Cache:\n", i);
3163 		cache_print(cache);
3164 	}
3165 }
3166 
3167 #ifdef DDB
3168 /*
3169  * Generate statistics across both the zone and its per-cpu cache's.  Return
3170  * desired statistics if the pointer is non-NULL for that statistic.
3171  *
3172  * Note: does not update the zone statistics, as it can't safely clear the
3173  * per-CPU cache statistic.
3174  *
3175  * XXXRW: Following the uc_allocbucket and uc_freebucket pointers here isn't
3176  * safe from off-CPU; we should modify the caches to track this information
3177  * directly so that we don't have to.
3178  */
3179 static void
3180 uma_zone_sumstat(uma_zone_t z, int *cachefreep, uint64_t *allocsp,
3181     uint64_t *freesp, uint64_t *sleepsp)
3182 {
3183 	uma_cache_t cache;
3184 	uint64_t allocs, frees, sleeps;
3185 	int cachefree, cpu;
3186 
3187 	allocs = frees = sleeps = 0;
3188 	cachefree = 0;
3189 	CPU_FOREACH(cpu) {
3190 		cache = &z->uz_cpu[cpu];
3191 		if (cache->uc_allocbucket != NULL)
3192 			cachefree += cache->uc_allocbucket->ub_cnt;
3193 		if (cache->uc_freebucket != NULL)
3194 			cachefree += cache->uc_freebucket->ub_cnt;
3195 		allocs += cache->uc_allocs;
3196 		frees += cache->uc_frees;
3197 	}
3198 	allocs += z->uz_allocs;
3199 	frees += z->uz_frees;
3200 	sleeps += z->uz_sleeps;
3201 	if (cachefreep != NULL)
3202 		*cachefreep = cachefree;
3203 	if (allocsp != NULL)
3204 		*allocsp = allocs;
3205 	if (freesp != NULL)
3206 		*freesp = frees;
3207 	if (sleepsp != NULL)
3208 		*sleepsp = sleeps;
3209 }
3210 #endif /* DDB */
3211 
3212 static int
3213 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
3214 {
3215 	uma_keg_t kz;
3216 	uma_zone_t z;
3217 	int count;
3218 
3219 	count = 0;
3220 	mtx_lock(&uma_mtx);
3221 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3222 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
3223 			count++;
3224 	}
3225 	mtx_unlock(&uma_mtx);
3226 	return (sysctl_handle_int(oidp, &count, 0, req));
3227 }
3228 
3229 static int
3230 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
3231 {
3232 	struct uma_stream_header ush;
3233 	struct uma_type_header uth;
3234 	struct uma_percpu_stat ups;
3235 	uma_bucket_t bucket;
3236 	struct sbuf sbuf;
3237 	uma_cache_t cache;
3238 	uma_klink_t kl;
3239 	uma_keg_t kz;
3240 	uma_zone_t z;
3241 	uma_keg_t k;
3242 	int count, error, i;
3243 
3244 	error = sysctl_wire_old_buffer(req, 0);
3245 	if (error != 0)
3246 		return (error);
3247 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
3248 
3249 	count = 0;
3250 	mtx_lock(&uma_mtx);
3251 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3252 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
3253 			count++;
3254 	}
3255 
3256 	/*
3257 	 * Insert stream header.
3258 	 */
3259 	bzero(&ush, sizeof(ush));
3260 	ush.ush_version = UMA_STREAM_VERSION;
3261 	ush.ush_maxcpus = (mp_maxid + 1);
3262 	ush.ush_count = count;
3263 	(void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
3264 
3265 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3266 		LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3267 			bzero(&uth, sizeof(uth));
3268 			ZONE_LOCK(z);
3269 			strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
3270 			uth.uth_align = kz->uk_align;
3271 			uth.uth_size = kz->uk_size;
3272 			uth.uth_rsize = kz->uk_rsize;
3273 			LIST_FOREACH(kl, &z->uz_kegs, kl_link) {
3274 				k = kl->kl_keg;
3275 				uth.uth_maxpages += k->uk_maxpages;
3276 				uth.uth_pages += k->uk_pages;
3277 				uth.uth_keg_free += k->uk_free;
3278 				uth.uth_limit = (k->uk_maxpages / k->uk_ppera)
3279 				    * k->uk_ipers;
3280 			}
3281 
3282 			/*
3283 			 * A zone is secondary is it is not the first entry
3284 			 * on the keg's zone list.
3285 			 */
3286 			if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
3287 			    (LIST_FIRST(&kz->uk_zones) != z))
3288 				uth.uth_zone_flags = UTH_ZONE_SECONDARY;
3289 
3290 			LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3291 				uth.uth_zone_free += bucket->ub_cnt;
3292 			uth.uth_allocs = z->uz_allocs;
3293 			uth.uth_frees = z->uz_frees;
3294 			uth.uth_fails = z->uz_fails;
3295 			uth.uth_sleeps = z->uz_sleeps;
3296 			(void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
3297 			/*
3298 			 * While it is not normally safe to access the cache
3299 			 * bucket pointers while not on the CPU that owns the
3300 			 * cache, we only allow the pointers to be exchanged
3301 			 * without the zone lock held, not invalidated, so
3302 			 * accept the possible race associated with bucket
3303 			 * exchange during monitoring.
3304 			 */
3305 			for (i = 0; i < (mp_maxid + 1); i++) {
3306 				bzero(&ups, sizeof(ups));
3307 				if (kz->uk_flags & UMA_ZFLAG_INTERNAL)
3308 					goto skip;
3309 				if (CPU_ABSENT(i))
3310 					goto skip;
3311 				cache = &z->uz_cpu[i];
3312 				if (cache->uc_allocbucket != NULL)
3313 					ups.ups_cache_free +=
3314 					    cache->uc_allocbucket->ub_cnt;
3315 				if (cache->uc_freebucket != NULL)
3316 					ups.ups_cache_free +=
3317 					    cache->uc_freebucket->ub_cnt;
3318 				ups.ups_allocs = cache->uc_allocs;
3319 				ups.ups_frees = cache->uc_frees;
3320 skip:
3321 				(void)sbuf_bcat(&sbuf, &ups, sizeof(ups));
3322 			}
3323 			ZONE_UNLOCK(z);
3324 		}
3325 	}
3326 	mtx_unlock(&uma_mtx);
3327 	error = sbuf_finish(&sbuf);
3328 	sbuf_delete(&sbuf);
3329 	return (error);
3330 }
3331 
3332 #ifdef DDB
3333 DB_SHOW_COMMAND(uma, db_show_uma)
3334 {
3335 	uint64_t allocs, frees, sleeps;
3336 	uma_bucket_t bucket;
3337 	uma_keg_t kz;
3338 	uma_zone_t z;
3339 	int cachefree;
3340 
3341 	db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
3342 	    "Requests", "Sleeps");
3343 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3344 		LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3345 			if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
3346 				allocs = z->uz_allocs;
3347 				frees = z->uz_frees;
3348 				sleeps = z->uz_sleeps;
3349 				cachefree = 0;
3350 			} else
3351 				uma_zone_sumstat(z, &cachefree, &allocs,
3352 				    &frees, &sleeps);
3353 			if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
3354 			    (LIST_FIRST(&kz->uk_zones) != z)))
3355 				cachefree += kz->uk_free;
3356 			LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3357 				cachefree += bucket->ub_cnt;
3358 			db_printf("%18s %8ju %8jd %8d %12ju %8ju\n", z->uz_name,
3359 			    (uintmax_t)kz->uk_size,
3360 			    (intmax_t)(allocs - frees), cachefree,
3361 			    (uintmax_t)allocs, sleeps);
3362 			if (db_pager_quit)
3363 				return;
3364 		}
3365 	}
3366 }
3367 #endif
3368