xref: /freebsd/sys/vm/uma_core.c (revision f0157ce528a128e2abb181a5c766033a2ce49a5f)
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, "4 Bucket", BUCKET_SIZE(4), 4096 },
210 	{ NULL, "8 Bucket", BUCKET_SIZE(8), 2048 },
211 	{ NULL, "16 Bucket", BUCKET_SIZE(16), 1024 },
212 	{ NULL, "32 Bucket", BUCKET_SIZE(32), 512 },
213 	{ NULL, "64 Bucket", BUCKET_SIZE(64), 256 },
214 	{ NULL, "128 Bucket", BUCKET_SIZE(128), 128 },
215 	{ NULL, NULL, 0}
216 };
217 
218 /*
219  * Flags and enumerations to be passed to internal functions.
220  */
221 enum zfreeskip { SKIP_NONE = 0, SKIP_DTOR, SKIP_FINI };
222 
223 /* Prototypes.. */
224 
225 static void *noobj_alloc(uma_zone_t, int, uint8_t *, int);
226 static void *page_alloc(uma_zone_t, int, uint8_t *, int);
227 static void *startup_alloc(uma_zone_t, int, uint8_t *, int);
228 static void page_free(void *, int, uint8_t);
229 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int);
230 static void cache_drain(uma_zone_t);
231 static void bucket_drain(uma_zone_t, uma_bucket_t);
232 static void bucket_cache_drain(uma_zone_t zone);
233 static int keg_ctor(void *, int, void *, int);
234 static void keg_dtor(void *, int, void *);
235 static int zone_ctor(void *, int, void *, int);
236 static void zone_dtor(void *, int, void *);
237 static int zero_init(void *, int, int);
238 static void keg_small_init(uma_keg_t keg);
239 static void keg_large_init(uma_keg_t keg);
240 static void zone_foreach(void (*zfunc)(uma_zone_t));
241 static void zone_timeout(uma_zone_t zone);
242 static int hash_alloc(struct uma_hash *);
243 static int hash_expand(struct uma_hash *, struct uma_hash *);
244 static void hash_free(struct uma_hash *hash);
245 static void uma_timeout(void *);
246 static void uma_startup3(void);
247 static void *zone_alloc_item(uma_zone_t, void *, int);
248 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip);
249 static void bucket_enable(void);
250 static void bucket_init(void);
251 static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int);
252 static void bucket_free(uma_zone_t zone, uma_bucket_t, void *);
253 static void bucket_zone_drain(void);
254 static uma_bucket_t zone_alloc_bucket(uma_zone_t zone, void *, int flags);
255 static uma_slab_t zone_fetch_slab(uma_zone_t zone, uma_keg_t last, int flags);
256 static uma_slab_t zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int flags);
257 static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab);
258 static void slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item);
259 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit,
260     uma_fini fini, int align, uint32_t flags);
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_MTXCLASS | UMA_ZFLAG_BUCKET);
310 	}
311 }
312 
313 /*
314  * Given a desired number of entries for a bucket, return the zone from which
315  * to allocate the bucket.
316  */
317 static struct uma_bucket_zone *
318 bucket_zone_lookup(int entries)
319 {
320 	struct uma_bucket_zone *ubz;
321 
322 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
323 		if (ubz->ubz_entries >= entries)
324 			return (ubz);
325 	ubz--;
326 	return (ubz);
327 }
328 
329 static int
330 bucket_select(int size)
331 {
332 	struct uma_bucket_zone *ubz;
333 
334 	ubz = &bucket_zones[0];
335 	if (size > ubz->ubz_maxsize)
336 		return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1);
337 
338 	for (; ubz->ubz_entries != 0; ubz++)
339 		if (ubz->ubz_maxsize < size)
340 			break;
341 	ubz--;
342 	return (ubz->ubz_entries);
343 }
344 
345 static uma_bucket_t
346 bucket_alloc(uma_zone_t zone, void *udata, int flags)
347 {
348 	struct uma_bucket_zone *ubz;
349 	uma_bucket_t bucket;
350 
351 	/*
352 	 * This is to stop us from allocating per cpu buckets while we're
353 	 * running out of vm.boot_pages.  Otherwise, we would exhaust the
354 	 * boot pages.  This also prevents us from allocating buckets in
355 	 * low memory situations.
356 	 */
357 	if (bucketdisable)
358 		return (NULL);
359 	/*
360 	 * To limit bucket recursion we store the original zone flags
361 	 * in a cookie passed via zalloc_arg/zfree_arg.  This allows the
362 	 * NOVM flag to persist even through deep recursions.  We also
363 	 * store ZFLAG_BUCKET once we have recursed attempting to allocate
364 	 * a bucket for a bucket zone so we do not allow infinite bucket
365 	 * recursion.  This cookie will even persist to frees of unused
366 	 * buckets via the allocation path or bucket allocations in the
367 	 * free path.
368 	 */
369 	if ((uintptr_t)udata & UMA_ZFLAG_BUCKET)
370 		return (NULL);
371 	if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
372 		udata = (void *)(uintptr_t)zone->uz_flags;
373 	else
374 		udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET);
375 	if ((uintptr_t)udata & UMA_ZFLAG_CACHEONLY)
376 		flags |= M_NOVM;
377 	ubz = bucket_zone_lookup(zone->uz_count);
378 	bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags);
379 	if (bucket) {
380 #ifdef INVARIANTS
381 		bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
382 #endif
383 		bucket->ub_cnt = 0;
384 		bucket->ub_entries = ubz->ubz_entries;
385 	}
386 
387 	return (bucket);
388 }
389 
390 static void
391 bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata)
392 {
393 	struct uma_bucket_zone *ubz;
394 
395 	KASSERT(bucket->ub_cnt == 0,
396 	    ("bucket_free: Freeing a non free bucket."));
397 	if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
398 		udata = (void *)(uintptr_t)zone->uz_flags;
399 	ubz = bucket_zone_lookup(bucket->ub_entries);
400 	uma_zfree_arg(ubz->ubz_zone, bucket, udata);
401 }
402 
403 static void
404 bucket_zone_drain(void)
405 {
406 	struct uma_bucket_zone *ubz;
407 
408 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
409 		zone_drain(ubz->ubz_zone);
410 }
411 
412 static void
413 zone_log_warning(uma_zone_t zone)
414 {
415 	static const struct timeval warninterval = { 300, 0 };
416 
417 	if (!zone_warnings || zone->uz_warning == NULL)
418 		return;
419 
420 	if (ratecheck(&zone->uz_ratecheck, &warninterval))
421 		printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning);
422 }
423 
424 static void
425 zone_foreach_keg(uma_zone_t zone, void (*kegfn)(uma_keg_t))
426 {
427 	uma_klink_t klink;
428 
429 	LIST_FOREACH(klink, &zone->uz_kegs, kl_link)
430 		kegfn(klink->kl_keg);
431 }
432 
433 /*
434  * Routine called by timeout which is used to fire off some time interval
435  * based calculations.  (stats, hash size, etc.)
436  *
437  * Arguments:
438  *	arg   Unused
439  *
440  * Returns:
441  *	Nothing
442  */
443 static void
444 uma_timeout(void *unused)
445 {
446 	bucket_enable();
447 	zone_foreach(zone_timeout);
448 
449 	/* Reschedule this event */
450 	callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
451 }
452 
453 /*
454  * Routine to perform timeout driven calculations.  This expands the
455  * hashes and does per cpu statistics aggregation.
456  *
457  *  Returns nothing.
458  */
459 static void
460 keg_timeout(uma_keg_t keg)
461 {
462 
463 	KEG_LOCK(keg);
464 	/*
465 	 * Expand the keg hash table.
466 	 *
467 	 * This is done if the number of slabs is larger than the hash size.
468 	 * What I'm trying to do here is completely reduce collisions.  This
469 	 * may be a little aggressive.  Should I allow for two collisions max?
470 	 */
471 	if (keg->uk_flags & UMA_ZONE_HASH &&
472 	    keg->uk_pages / keg->uk_ppera >= keg->uk_hash.uh_hashsize) {
473 		struct uma_hash newhash;
474 		struct uma_hash oldhash;
475 		int ret;
476 
477 		/*
478 		 * This is so involved because allocating and freeing
479 		 * while the keg lock is held will lead to deadlock.
480 		 * I have to do everything in stages and check for
481 		 * races.
482 		 */
483 		newhash = keg->uk_hash;
484 		KEG_UNLOCK(keg);
485 		ret = hash_alloc(&newhash);
486 		KEG_LOCK(keg);
487 		if (ret) {
488 			if (hash_expand(&keg->uk_hash, &newhash)) {
489 				oldhash = keg->uk_hash;
490 				keg->uk_hash = newhash;
491 			} else
492 				oldhash = newhash;
493 
494 			KEG_UNLOCK(keg);
495 			hash_free(&oldhash);
496 			return;
497 		}
498 	}
499 	KEG_UNLOCK(keg);
500 }
501 
502 static void
503 zone_timeout(uma_zone_t zone)
504 {
505 
506 	zone_foreach_keg(zone, &keg_timeout);
507 }
508 
509 /*
510  * Allocate and zero fill the next sized hash table from the appropriate
511  * backing store.
512  *
513  * Arguments:
514  *	hash  A new hash structure with the old hash size in uh_hashsize
515  *
516  * Returns:
517  *	1 on sucess and 0 on failure.
518  */
519 static int
520 hash_alloc(struct uma_hash *hash)
521 {
522 	int oldsize;
523 	int alloc;
524 
525 	oldsize = hash->uh_hashsize;
526 
527 	/* We're just going to go to a power of two greater */
528 	if (oldsize)  {
529 		hash->uh_hashsize = oldsize * 2;
530 		alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
531 		hash->uh_slab_hash = (struct slabhead *)malloc(alloc,
532 		    M_UMAHASH, M_NOWAIT);
533 	} else {
534 		alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
535 		hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
536 		    M_WAITOK);
537 		hash->uh_hashsize = UMA_HASH_SIZE_INIT;
538 	}
539 	if (hash->uh_slab_hash) {
540 		bzero(hash->uh_slab_hash, alloc);
541 		hash->uh_hashmask = hash->uh_hashsize - 1;
542 		return (1);
543 	}
544 
545 	return (0);
546 }
547 
548 /*
549  * Expands the hash table for HASH zones.  This is done from zone_timeout
550  * to reduce collisions.  This must not be done in the regular allocation
551  * path, otherwise, we can recurse on the vm while allocating pages.
552  *
553  * Arguments:
554  *	oldhash  The hash you want to expand
555  *	newhash  The hash structure for the new table
556  *
557  * Returns:
558  *	Nothing
559  *
560  * Discussion:
561  */
562 static int
563 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
564 {
565 	uma_slab_t slab;
566 	int hval;
567 	int i;
568 
569 	if (!newhash->uh_slab_hash)
570 		return (0);
571 
572 	if (oldhash->uh_hashsize >= newhash->uh_hashsize)
573 		return (0);
574 
575 	/*
576 	 * I need to investigate hash algorithms for resizing without a
577 	 * full rehash.
578 	 */
579 
580 	for (i = 0; i < oldhash->uh_hashsize; i++)
581 		while (!SLIST_EMPTY(&oldhash->uh_slab_hash[i])) {
582 			slab = SLIST_FIRST(&oldhash->uh_slab_hash[i]);
583 			SLIST_REMOVE_HEAD(&oldhash->uh_slab_hash[i], us_hlink);
584 			hval = UMA_HASH(newhash, slab->us_data);
585 			SLIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
586 			    slab, us_hlink);
587 		}
588 
589 	return (1);
590 }
591 
592 /*
593  * Free the hash bucket to the appropriate backing store.
594  *
595  * Arguments:
596  *	slab_hash  The hash bucket we're freeing
597  *	hashsize   The number of entries in that hash bucket
598  *
599  * Returns:
600  *	Nothing
601  */
602 static void
603 hash_free(struct uma_hash *hash)
604 {
605 	if (hash->uh_slab_hash == NULL)
606 		return;
607 	if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
608 		zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE);
609 	else
610 		free(hash->uh_slab_hash, M_UMAHASH);
611 }
612 
613 /*
614  * Frees all outstanding items in a bucket
615  *
616  * Arguments:
617  *	zone   The zone to free to, must be unlocked.
618  *	bucket The free/alloc bucket with items, cpu queue must be locked.
619  *
620  * Returns:
621  *	Nothing
622  */
623 
624 static void
625 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
626 {
627 	int i;
628 
629 	if (bucket == NULL)
630 		return;
631 
632 	if (zone->uz_fini)
633 		for (i = 0; i < bucket->ub_cnt; i++)
634 			zone->uz_fini(bucket->ub_bucket[i], zone->uz_size);
635 	zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt);
636 	bucket->ub_cnt = 0;
637 }
638 
639 /*
640  * Drains the per cpu caches for a zone.
641  *
642  * NOTE: This may only be called while the zone is being turn down, and not
643  * during normal operation.  This is necessary in order that we do not have
644  * to migrate CPUs to drain the per-CPU caches.
645  *
646  * Arguments:
647  *	zone     The zone to drain, must be unlocked.
648  *
649  * Returns:
650  *	Nothing
651  */
652 static void
653 cache_drain(uma_zone_t zone)
654 {
655 	uma_cache_t cache;
656 	int cpu;
657 
658 	/*
659 	 * XXX: It is safe to not lock the per-CPU caches, because we're
660 	 * tearing down the zone anyway.  I.e., there will be no further use
661 	 * of the caches at this point.
662 	 *
663 	 * XXX: It would good to be able to assert that the zone is being
664 	 * torn down to prevent improper use of cache_drain().
665 	 *
666 	 * XXX: We lock the zone before passing into bucket_cache_drain() as
667 	 * it is used elsewhere.  Should the tear-down path be made special
668 	 * there in some form?
669 	 */
670 	CPU_FOREACH(cpu) {
671 		cache = &zone->uz_cpu[cpu];
672 		bucket_drain(zone, cache->uc_allocbucket);
673 		bucket_drain(zone, cache->uc_freebucket);
674 		if (cache->uc_allocbucket != NULL)
675 			bucket_free(zone, cache->uc_allocbucket, NULL);
676 		if (cache->uc_freebucket != NULL)
677 			bucket_free(zone, cache->uc_freebucket, NULL);
678 		cache->uc_allocbucket = cache->uc_freebucket = NULL;
679 	}
680 	ZONE_LOCK(zone);
681 	bucket_cache_drain(zone);
682 	ZONE_UNLOCK(zone);
683 }
684 
685 /*
686  * Drain the cached buckets from a zone.  Expects a locked zone on entry.
687  */
688 static void
689 bucket_cache_drain(uma_zone_t zone)
690 {
691 	uma_bucket_t bucket;
692 
693 	/*
694 	 * Drain the bucket queues and free the buckets, we just keep two per
695 	 * cpu (alloc/free).
696 	 */
697 	while ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) {
698 		LIST_REMOVE(bucket, ub_link);
699 		ZONE_UNLOCK(zone);
700 		bucket_drain(zone, bucket);
701 		bucket_free(zone, bucket, NULL);
702 		ZONE_LOCK(zone);
703 	}
704 }
705 
706 static void
707 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start)
708 {
709 	uint8_t *mem;
710 	int i;
711 	uint8_t flags;
712 
713 	mem = slab->us_data;
714 	flags = slab->us_flags;
715 	i = start;
716 	if (keg->uk_fini != NULL) {
717 		for (i--; i > -1; i--)
718 			keg->uk_fini(slab->us_data + (keg->uk_rsize * i),
719 			    keg->uk_size);
720 	}
721 	if (keg->uk_flags & UMA_ZONE_VTOSLAB) {
722 		vm_object_t obj;
723 
724 		if (flags & UMA_SLAB_KMEM)
725 			obj = kmem_object;
726 		else if (flags & UMA_SLAB_KERNEL)
727 			obj = kernel_object;
728 		else
729 			obj = NULL;
730 		for (i = 0; i < keg->uk_ppera; i++)
731 			vsetobj((vm_offset_t)mem + (i * PAGE_SIZE), obj);
732 	}
733 	if (keg->uk_flags & UMA_ZONE_OFFPAGE)
734 		zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE);
735 #ifdef UMA_DEBUG
736 	printf("%s: Returning %d bytes.\n", keg->uk_name,
737 	    PAGE_SIZE * keg->uk_ppera);
738 #endif
739 	keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags);
740 }
741 
742 /*
743  * Frees pages from a keg back to the system.  This is done on demand from
744  * the pageout daemon.
745  *
746  * Returns nothing.
747  */
748 static void
749 keg_drain(uma_keg_t keg)
750 {
751 	struct slabhead freeslabs = { 0 };
752 	uma_slab_t slab;
753 	uma_slab_t n;
754 
755 	/*
756 	 * We don't want to take pages from statically allocated kegs at this
757 	 * time
758 	 */
759 	if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL)
760 		return;
761 
762 #ifdef UMA_DEBUG
763 	printf("%s free items: %u\n", keg->uk_name, keg->uk_free);
764 #endif
765 	KEG_LOCK(keg);
766 	if (keg->uk_free == 0)
767 		goto finished;
768 
769 	slab = LIST_FIRST(&keg->uk_free_slab);
770 	while (slab) {
771 		n = LIST_NEXT(slab, us_link);
772 
773 		/* We have no where to free these to */
774 		if (slab->us_flags & UMA_SLAB_BOOT) {
775 			slab = n;
776 			continue;
777 		}
778 
779 		LIST_REMOVE(slab, us_link);
780 		keg->uk_pages -= keg->uk_ppera;
781 		keg->uk_free -= keg->uk_ipers;
782 
783 		if (keg->uk_flags & UMA_ZONE_HASH)
784 			UMA_HASH_REMOVE(&keg->uk_hash, slab, slab->us_data);
785 
786 		SLIST_INSERT_HEAD(&freeslabs, slab, us_hlink);
787 
788 		slab = n;
789 	}
790 finished:
791 	KEG_UNLOCK(keg);
792 
793 	while ((slab = SLIST_FIRST(&freeslabs)) != NULL) {
794 		SLIST_REMOVE(&freeslabs, slab, uma_slab, us_hlink);
795 		keg_free_slab(keg, slab, 0);
796 	}
797 }
798 
799 static void
800 zone_drain_wait(uma_zone_t zone, int waitok)
801 {
802 
803 	/*
804 	 * Set draining to interlock with zone_dtor() so we can release our
805 	 * locks as we go.  Only dtor() should do a WAITOK call since it
806 	 * is the only call that knows the structure will still be available
807 	 * when it wakes up.
808 	 */
809 	ZONE_LOCK(zone);
810 	while (zone->uz_flags & UMA_ZFLAG_DRAINING) {
811 		if (waitok == M_NOWAIT)
812 			goto out;
813 		mtx_unlock(&uma_mtx);
814 		msleep(zone, zone->uz_lockptr, PVM, "zonedrain", 1);
815 		mtx_lock(&uma_mtx);
816 	}
817 	zone->uz_flags |= UMA_ZFLAG_DRAINING;
818 	bucket_cache_drain(zone);
819 	ZONE_UNLOCK(zone);
820 	/*
821 	 * The DRAINING flag protects us from being freed while
822 	 * we're running.  Normally the uma_mtx would protect us but we
823 	 * must be able to release and acquire the right lock for each keg.
824 	 */
825 	zone_foreach_keg(zone, &keg_drain);
826 	ZONE_LOCK(zone);
827 	zone->uz_flags &= ~UMA_ZFLAG_DRAINING;
828 	wakeup(zone);
829 out:
830 	ZONE_UNLOCK(zone);
831 }
832 
833 void
834 zone_drain(uma_zone_t zone)
835 {
836 
837 	zone_drain_wait(zone, M_NOWAIT);
838 }
839 
840 /*
841  * Allocate a new slab for a keg.  This does not insert the slab onto a list.
842  *
843  * Arguments:
844  *	wait  Shall we wait?
845  *
846  * Returns:
847  *	The slab that was allocated or NULL if there is no memory and the
848  *	caller specified M_NOWAIT.
849  */
850 static uma_slab_t
851 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int wait)
852 {
853 	uma_slabrefcnt_t slabref;
854 	uma_alloc allocf;
855 	uma_slab_t slab;
856 	uint8_t *mem;
857 	uint8_t flags;
858 	int i;
859 
860 	mtx_assert(&keg->uk_lock, MA_OWNED);
861 	slab = NULL;
862 	mem = NULL;
863 
864 #ifdef UMA_DEBUG
865 	printf("alloc_slab:  Allocating a new slab for %s\n", keg->uk_name);
866 #endif
867 	allocf = keg->uk_allocf;
868 	KEG_UNLOCK(keg);
869 
870 	if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
871 		slab = zone_alloc_item(keg->uk_slabzone, NULL, wait);
872 		if (slab == NULL)
873 			goto out;
874 	}
875 
876 	/*
877 	 * This reproduces the old vm_zone behavior of zero filling pages the
878 	 * first time they are added to a zone.
879 	 *
880 	 * Malloced items are zeroed in uma_zalloc.
881 	 */
882 
883 	if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
884 		wait |= M_ZERO;
885 	else
886 		wait &= ~M_ZERO;
887 
888 	if (keg->uk_flags & UMA_ZONE_NODUMP)
889 		wait |= M_NODUMP;
890 
891 	/* zone is passed for legacy reasons. */
892 	mem = allocf(zone, keg->uk_ppera * PAGE_SIZE, &flags, wait);
893 	if (mem == NULL) {
894 		if (keg->uk_flags & UMA_ZONE_OFFPAGE)
895 			zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE);
896 		slab = NULL;
897 		goto out;
898 	}
899 
900 	/* Point the slab into the allocated memory */
901 	if (!(keg->uk_flags & UMA_ZONE_OFFPAGE))
902 		slab = (uma_slab_t )(mem + keg->uk_pgoff);
903 
904 	if (keg->uk_flags & UMA_ZONE_VTOSLAB)
905 		for (i = 0; i < keg->uk_ppera; i++)
906 			vsetslab((vm_offset_t)mem + (i * PAGE_SIZE), slab);
907 
908 	slab->us_keg = keg;
909 	slab->us_data = mem;
910 	slab->us_freecount = keg->uk_ipers;
911 	slab->us_flags = flags;
912 	BIT_FILL(SLAB_SETSIZE, &slab->us_free);
913 #ifdef INVARIANTS
914 	BIT_ZERO(SLAB_SETSIZE, &slab->us_debugfree);
915 #endif
916 	if (keg->uk_flags & UMA_ZONE_REFCNT) {
917 		slabref = (uma_slabrefcnt_t)slab;
918 		for (i = 0; i < keg->uk_ipers; i++)
919 			slabref->us_refcnt[i] = 0;
920 	}
921 
922 	if (keg->uk_init != NULL) {
923 		for (i = 0; i < keg->uk_ipers; i++)
924 			if (keg->uk_init(slab->us_data + (keg->uk_rsize * i),
925 			    keg->uk_size, wait) != 0)
926 				break;
927 		if (i != keg->uk_ipers) {
928 			keg_free_slab(keg, slab, i);
929 			slab = NULL;
930 			goto out;
931 		}
932 	}
933 out:
934 	KEG_LOCK(keg);
935 
936 	if (slab != NULL) {
937 		if (keg->uk_flags & UMA_ZONE_HASH)
938 			UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
939 
940 		keg->uk_pages += keg->uk_ppera;
941 		keg->uk_free += keg->uk_ipers;
942 	}
943 
944 	return (slab);
945 }
946 
947 /*
948  * This function is intended to be used early on in place of page_alloc() so
949  * that we may use the boot time page cache to satisfy allocations before
950  * the VM is ready.
951  */
952 static void *
953 startup_alloc(uma_zone_t zone, int bytes, uint8_t *pflag, int wait)
954 {
955 	uma_keg_t keg;
956 	uma_slab_t tmps;
957 	int pages, check_pages;
958 
959 	keg = zone_first_keg(zone);
960 	pages = howmany(bytes, PAGE_SIZE);
961 	check_pages = pages - 1;
962 	KASSERT(pages > 0, ("startup_alloc can't reserve 0 pages\n"));
963 
964 	/*
965 	 * Check our small startup cache to see if it has pages remaining.
966 	 */
967 	mtx_lock(&uma_boot_pages_mtx);
968 
969 	/* First check if we have enough room. */
970 	tmps = LIST_FIRST(&uma_boot_pages);
971 	while (tmps != NULL && check_pages-- > 0)
972 		tmps = LIST_NEXT(tmps, us_link);
973 	if (tmps != NULL) {
974 		/*
975 		 * It's ok to lose tmps references.  The last one will
976 		 * have tmps->us_data pointing to the start address of
977 		 * "pages" contiguous pages of memory.
978 		 */
979 		while (pages-- > 0) {
980 			tmps = LIST_FIRST(&uma_boot_pages);
981 			LIST_REMOVE(tmps, us_link);
982 		}
983 		mtx_unlock(&uma_boot_pages_mtx);
984 		*pflag = tmps->us_flags;
985 		return (tmps->us_data);
986 	}
987 	mtx_unlock(&uma_boot_pages_mtx);
988 	if (booted < UMA_STARTUP2)
989 		panic("UMA: Increase vm.boot_pages");
990 	/*
991 	 * Now that we've booted reset these users to their real allocator.
992 	 */
993 #ifdef UMA_MD_SMALL_ALLOC
994 	keg->uk_allocf = (keg->uk_ppera > 1) ? page_alloc : uma_small_alloc;
995 #else
996 	keg->uk_allocf = page_alloc;
997 #endif
998 	return keg->uk_allocf(zone, bytes, pflag, wait);
999 }
1000 
1001 /*
1002  * Allocates a number of pages from the system
1003  *
1004  * Arguments:
1005  *	bytes  The number of bytes requested
1006  *	wait  Shall we wait?
1007  *
1008  * Returns:
1009  *	A pointer to the alloced memory or possibly
1010  *	NULL if M_NOWAIT is set.
1011  */
1012 static void *
1013 page_alloc(uma_zone_t zone, int bytes, uint8_t *pflag, int wait)
1014 {
1015 	void *p;	/* Returned page */
1016 
1017 	*pflag = UMA_SLAB_KMEM;
1018 	p = (void *) kmem_malloc(kmem_map, bytes, wait);
1019 
1020 	return (p);
1021 }
1022 
1023 /*
1024  * Allocates a number of pages from within an object
1025  *
1026  * Arguments:
1027  *	bytes  The number of bytes requested
1028  *	wait   Shall we wait?
1029  *
1030  * Returns:
1031  *	A pointer to the alloced memory or possibly
1032  *	NULL if M_NOWAIT is set.
1033  */
1034 static void *
1035 noobj_alloc(uma_zone_t zone, int bytes, uint8_t *flags, int wait)
1036 {
1037 	TAILQ_HEAD(, vm_page) alloctail;
1038 	u_long npages;
1039 	vm_offset_t retkva, zkva;
1040 	vm_page_t p, p_next;
1041 	uma_keg_t keg;
1042 
1043 	TAILQ_INIT(&alloctail);
1044 	keg = zone_first_keg(zone);
1045 
1046 	npages = howmany(bytes, PAGE_SIZE);
1047 	while (npages > 0) {
1048 		p = vm_page_alloc(NULL, 0, VM_ALLOC_INTERRUPT |
1049 		    VM_ALLOC_WIRED | VM_ALLOC_NOOBJ);
1050 		if (p != NULL) {
1051 			/*
1052 			 * Since the page does not belong to an object, its
1053 			 * listq is unused.
1054 			 */
1055 			TAILQ_INSERT_TAIL(&alloctail, p, listq);
1056 			npages--;
1057 			continue;
1058 		}
1059 		if (wait & M_WAITOK) {
1060 			VM_WAIT;
1061 			continue;
1062 		}
1063 
1064 		/*
1065 		 * Page allocation failed, free intermediate pages and
1066 		 * exit.
1067 		 */
1068 		TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1069 			vm_page_unwire(p, 0);
1070 			vm_page_free(p);
1071 		}
1072 		return (NULL);
1073 	}
1074 	*flags = UMA_SLAB_PRIV;
1075 	zkva = keg->uk_kva +
1076 	    atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
1077 	retkva = zkva;
1078 	TAILQ_FOREACH(p, &alloctail, listq) {
1079 		pmap_qenter(zkva, &p, 1);
1080 		zkva += PAGE_SIZE;
1081 	}
1082 
1083 	return ((void *)retkva);
1084 }
1085 
1086 /*
1087  * Frees a number of pages to the system
1088  *
1089  * Arguments:
1090  *	mem   A pointer to the memory to be freed
1091  *	size  The size of the memory being freed
1092  *	flags The original p->us_flags field
1093  *
1094  * Returns:
1095  *	Nothing
1096  */
1097 static void
1098 page_free(void *mem, int size, uint8_t flags)
1099 {
1100 	vm_map_t map;
1101 
1102 	if (flags & UMA_SLAB_KMEM)
1103 		map = kmem_map;
1104 	else if (flags & UMA_SLAB_KERNEL)
1105 		map = kernel_map;
1106 	else
1107 		panic("UMA: page_free used with invalid flags %d", flags);
1108 
1109 	kmem_free(map, (vm_offset_t)mem, size);
1110 }
1111 
1112 /*
1113  * Zero fill initializer
1114  *
1115  * Arguments/Returns follow uma_init specifications
1116  */
1117 static int
1118 zero_init(void *mem, int size, int flags)
1119 {
1120 	bzero(mem, size);
1121 	return (0);
1122 }
1123 
1124 /*
1125  * Finish creating a small uma keg.  This calculates ipers, and the keg size.
1126  *
1127  * Arguments
1128  *	keg  The zone we should initialize
1129  *
1130  * Returns
1131  *	Nothing
1132  */
1133 static void
1134 keg_small_init(uma_keg_t keg)
1135 {
1136 	u_int rsize;
1137 	u_int memused;
1138 	u_int wastedspace;
1139 	u_int shsize;
1140 
1141 	if (keg->uk_flags & UMA_ZONE_PCPU) {
1142 		KASSERT(mp_ncpus > 0, ("%s: ncpus %d\n", __func__, mp_ncpus));
1143 		keg->uk_slabsize = sizeof(struct pcpu);
1144 		keg->uk_ppera = howmany(mp_ncpus * sizeof(struct pcpu),
1145 		    PAGE_SIZE);
1146 	} else {
1147 		keg->uk_slabsize = UMA_SLAB_SIZE;
1148 		keg->uk_ppera = 1;
1149 	}
1150 
1151 	/*
1152 	 * Calculate the size of each allocation (rsize) according to
1153 	 * alignment.  If the requested size is smaller than we have
1154 	 * allocation bits for we round it up.
1155 	 */
1156 	rsize = keg->uk_size;
1157 	if (rsize < keg->uk_slabsize / SLAB_SETSIZE)
1158 		rsize = keg->uk_slabsize / SLAB_SETSIZE;
1159 	if (rsize & keg->uk_align)
1160 		rsize = (rsize & ~keg->uk_align) + (keg->uk_align + 1);
1161 	keg->uk_rsize = rsize;
1162 
1163 	KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
1164 	    keg->uk_rsize < sizeof(struct pcpu),
1165 	    ("%s: size %u too large", __func__, keg->uk_rsize));
1166 
1167 	if (keg->uk_flags & UMA_ZONE_REFCNT)
1168 		rsize += sizeof(uint32_t);
1169 
1170 	if (keg->uk_flags & UMA_ZONE_OFFPAGE)
1171 		shsize = 0;
1172 	else
1173 		shsize = sizeof(struct uma_slab);
1174 
1175 	keg->uk_ipers = (keg->uk_slabsize - shsize) / rsize;
1176 	KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE,
1177 	    ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers));
1178 
1179 	memused = keg->uk_ipers * rsize + shsize;
1180 	wastedspace = keg->uk_slabsize - memused;
1181 
1182 	/*
1183 	 * We can't do OFFPAGE if we're internal or if we've been
1184 	 * asked to not go to the VM for buckets.  If we do this we
1185 	 * may end up going to the VM  for slabs which we do not
1186 	 * want to do if we're UMA_ZFLAG_CACHEONLY as a result
1187 	 * of UMA_ZONE_VM, which clearly forbids it.
1188 	 */
1189 	if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) ||
1190 	    (keg->uk_flags & UMA_ZFLAG_CACHEONLY))
1191 		return;
1192 
1193 	/*
1194 	 * See if using an OFFPAGE slab will limit our waste.  Only do
1195 	 * this if it permits more items per-slab.
1196 	 *
1197 	 * XXX We could try growing slabsize to limit max waste as well.
1198 	 * Historically this was not done because the VM could not
1199 	 * efficiently handle contiguous allocations.
1200 	 */
1201 	if ((wastedspace >= keg->uk_slabsize / UMA_MAX_WASTE) &&
1202 	    (keg->uk_ipers < (keg->uk_slabsize / keg->uk_rsize))) {
1203 		keg->uk_ipers = keg->uk_slabsize / keg->uk_rsize;
1204 		KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE,
1205 		    ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers));
1206 #ifdef UMA_DEBUG
1207 		printf("UMA decided we need offpage slab headers for "
1208 		    "keg: %s, calculated wastedspace = %d, "
1209 		    "maximum wasted space allowed = %d, "
1210 		    "calculated ipers = %d, "
1211 		    "new wasted space = %d\n", keg->uk_name, wastedspace,
1212 		    keg->uk_slabsize / UMA_MAX_WASTE, keg->uk_ipers,
1213 		    keg->uk_slabsize - keg->uk_ipers * keg->uk_rsize);
1214 #endif
1215 		keg->uk_flags |= UMA_ZONE_OFFPAGE;
1216 	}
1217 
1218 	if ((keg->uk_flags & UMA_ZONE_OFFPAGE) &&
1219 	    (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1220 		keg->uk_flags |= UMA_ZONE_HASH;
1221 }
1222 
1223 /*
1224  * Finish creating a large (> UMA_SLAB_SIZE) uma kegs.  Just give in and do
1225  * OFFPAGE for now.  When I can allow for more dynamic slab sizes this will be
1226  * more complicated.
1227  *
1228  * Arguments
1229  *	keg  The keg we should initialize
1230  *
1231  * Returns
1232  *	Nothing
1233  */
1234 static void
1235 keg_large_init(uma_keg_t keg)
1236 {
1237 
1238 	KASSERT(keg != NULL, ("Keg is null in keg_large_init"));
1239 	KASSERT((keg->uk_flags & UMA_ZFLAG_CACHEONLY) == 0,
1240 	    ("keg_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY keg"));
1241 	KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0,
1242 	    ("%s: Cannot large-init a UMA_ZONE_PCPU keg", __func__));
1243 
1244 	keg->uk_ppera = howmany(keg->uk_size, PAGE_SIZE);
1245 	keg->uk_slabsize = keg->uk_ppera * PAGE_SIZE;
1246 	keg->uk_ipers = 1;
1247 	keg->uk_rsize = keg->uk_size;
1248 
1249 	/* We can't do OFFPAGE if we're internal, bail out here. */
1250 	if (keg->uk_flags & UMA_ZFLAG_INTERNAL)
1251 		return;
1252 
1253 	keg->uk_flags |= UMA_ZONE_OFFPAGE;
1254 	if ((keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1255 		keg->uk_flags |= UMA_ZONE_HASH;
1256 }
1257 
1258 static void
1259 keg_cachespread_init(uma_keg_t keg)
1260 {
1261 	int alignsize;
1262 	int trailer;
1263 	int pages;
1264 	int rsize;
1265 
1266 	KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0,
1267 	    ("%s: Cannot cachespread-init a UMA_ZONE_PCPU keg", __func__));
1268 
1269 	alignsize = keg->uk_align + 1;
1270 	rsize = keg->uk_size;
1271 	/*
1272 	 * We want one item to start on every align boundary in a page.  To
1273 	 * do this we will span pages.  We will also extend the item by the
1274 	 * size of align if it is an even multiple of align.  Otherwise, it
1275 	 * would fall on the same boundary every time.
1276 	 */
1277 	if (rsize & keg->uk_align)
1278 		rsize = (rsize & ~keg->uk_align) + alignsize;
1279 	if ((rsize & alignsize) == 0)
1280 		rsize += alignsize;
1281 	trailer = rsize - keg->uk_size;
1282 	pages = (rsize * (PAGE_SIZE / alignsize)) / PAGE_SIZE;
1283 	pages = MIN(pages, (128 * 1024) / PAGE_SIZE);
1284 	keg->uk_rsize = rsize;
1285 	keg->uk_ppera = pages;
1286 	keg->uk_slabsize = UMA_SLAB_SIZE;
1287 	keg->uk_ipers = ((pages * PAGE_SIZE) + trailer) / rsize;
1288 	keg->uk_flags |= UMA_ZONE_OFFPAGE | UMA_ZONE_VTOSLAB;
1289 	KASSERT(keg->uk_ipers <= uma_max_ipers,
1290 	    ("%s: keg->uk_ipers too high(%d) increase max_ipers", __func__,
1291 	    keg->uk_ipers));
1292 }
1293 
1294 /*
1295  * Keg header ctor.  This initializes all fields, locks, etc.  And inserts
1296  * the keg onto the global keg list.
1297  *
1298  * Arguments/Returns follow uma_ctor specifications
1299  *	udata  Actually uma_kctor_args
1300  */
1301 static int
1302 keg_ctor(void *mem, int size, void *udata, int flags)
1303 {
1304 	struct uma_kctor_args *arg = udata;
1305 	uma_keg_t keg = mem;
1306 	uma_zone_t zone;
1307 
1308 	bzero(keg, size);
1309 	keg->uk_size = arg->size;
1310 	keg->uk_init = arg->uminit;
1311 	keg->uk_fini = arg->fini;
1312 	keg->uk_align = arg->align;
1313 	keg->uk_free = 0;
1314 	keg->uk_reserve = 0;
1315 	keg->uk_pages = 0;
1316 	keg->uk_flags = arg->flags;
1317 	keg->uk_allocf = page_alloc;
1318 	keg->uk_freef = page_free;
1319 	keg->uk_slabzone = NULL;
1320 
1321 	/*
1322 	 * The master zone is passed to us at keg-creation time.
1323 	 */
1324 	zone = arg->zone;
1325 	keg->uk_name = zone->uz_name;
1326 
1327 	if (arg->flags & UMA_ZONE_VM)
1328 		keg->uk_flags |= UMA_ZFLAG_CACHEONLY;
1329 
1330 	if (arg->flags & UMA_ZONE_ZINIT)
1331 		keg->uk_init = zero_init;
1332 
1333 	if (arg->flags & UMA_ZONE_REFCNT || arg->flags & UMA_ZONE_MALLOC)
1334 		keg->uk_flags |= UMA_ZONE_VTOSLAB;
1335 
1336 	if (arg->flags & UMA_ZONE_PCPU)
1337 #ifdef SMP
1338 		keg->uk_flags |= UMA_ZONE_OFFPAGE;
1339 #else
1340 		keg->uk_flags &= ~UMA_ZONE_PCPU;
1341 #endif
1342 
1343 	if (keg->uk_flags & UMA_ZONE_CACHESPREAD) {
1344 		keg_cachespread_init(keg);
1345 	} else if (keg->uk_flags & UMA_ZONE_REFCNT) {
1346 		if (keg->uk_size >
1347 		    (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt) -
1348 		    sizeof(uint32_t)))
1349 			keg_large_init(keg);
1350 		else
1351 			keg_small_init(keg);
1352 	} else {
1353 		if (keg->uk_size > (UMA_SLAB_SIZE - sizeof(struct uma_slab)))
1354 			keg_large_init(keg);
1355 		else
1356 			keg_small_init(keg);
1357 	}
1358 
1359 	if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
1360 		if (keg->uk_flags & UMA_ZONE_REFCNT) {
1361 			if (keg->uk_ipers > uma_max_ipers_ref)
1362 				panic("Too many ref items per zone: %d > %d\n",
1363 				    keg->uk_ipers, uma_max_ipers_ref);
1364 			keg->uk_slabzone = slabrefzone;
1365 		} else
1366 			keg->uk_slabzone = slabzone;
1367 	}
1368 
1369 	/*
1370 	 * If we haven't booted yet we need allocations to go through the
1371 	 * startup cache until the vm is ready.
1372 	 */
1373 	if (keg->uk_ppera == 1) {
1374 #ifdef UMA_MD_SMALL_ALLOC
1375 		keg->uk_allocf = uma_small_alloc;
1376 		keg->uk_freef = uma_small_free;
1377 
1378 		if (booted < UMA_STARTUP)
1379 			keg->uk_allocf = startup_alloc;
1380 #else
1381 		if (booted < UMA_STARTUP2)
1382 			keg->uk_allocf = startup_alloc;
1383 #endif
1384 	} else if (booted < UMA_STARTUP2 &&
1385 	    (keg->uk_flags & UMA_ZFLAG_INTERNAL))
1386 		keg->uk_allocf = startup_alloc;
1387 
1388 	/*
1389 	 * Initialize keg's lock
1390 	 */
1391 	KEG_LOCK_INIT(keg, (arg->flags & UMA_ZONE_MTXCLASS));
1392 
1393 	/*
1394 	 * If we're putting the slab header in the actual page we need to
1395 	 * figure out where in each page it goes.  This calculates a right
1396 	 * justified offset into the memory on an ALIGN_PTR boundary.
1397 	 */
1398 	if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) {
1399 		u_int totsize;
1400 
1401 		/* Size of the slab struct and free list */
1402 		totsize = sizeof(struct uma_slab);
1403 
1404 		/* Size of the reference counts. */
1405 		if (keg->uk_flags & UMA_ZONE_REFCNT)
1406 			totsize += keg->uk_ipers * sizeof(uint32_t);
1407 
1408 		if (totsize & UMA_ALIGN_PTR)
1409 			totsize = (totsize & ~UMA_ALIGN_PTR) +
1410 			    (UMA_ALIGN_PTR + 1);
1411 		keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - totsize;
1412 
1413 		/*
1414 		 * The only way the following is possible is if with our
1415 		 * UMA_ALIGN_PTR adjustments we are now bigger than
1416 		 * UMA_SLAB_SIZE.  I haven't checked whether this is
1417 		 * mathematically possible for all cases, so we make
1418 		 * sure here anyway.
1419 		 */
1420 		totsize = keg->uk_pgoff + sizeof(struct uma_slab);
1421 		if (keg->uk_flags & UMA_ZONE_REFCNT)
1422 			totsize += keg->uk_ipers * sizeof(uint32_t);
1423 		if (totsize > PAGE_SIZE * keg->uk_ppera) {
1424 			printf("zone %s ipers %d rsize %d size %d\n",
1425 			    zone->uz_name, keg->uk_ipers, keg->uk_rsize,
1426 			    keg->uk_size);
1427 			panic("UMA slab won't fit.");
1428 		}
1429 	}
1430 
1431 	if (keg->uk_flags & UMA_ZONE_HASH)
1432 		hash_alloc(&keg->uk_hash);
1433 
1434 #ifdef UMA_DEBUG
1435 	printf("UMA: %s(%p) size %d(%d) flags %#x ipers %d ppera %d out %d free %d\n",
1436 	    zone->uz_name, zone, keg->uk_size, keg->uk_rsize, keg->uk_flags,
1437 	    keg->uk_ipers, keg->uk_ppera,
1438 	    (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free);
1439 #endif
1440 
1441 	LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
1442 
1443 	mtx_lock(&uma_mtx);
1444 	LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
1445 	mtx_unlock(&uma_mtx);
1446 	return (0);
1447 }
1448 
1449 /*
1450  * Zone header ctor.  This initializes all fields, locks, etc.
1451  *
1452  * Arguments/Returns follow uma_ctor specifications
1453  *	udata  Actually uma_zctor_args
1454  */
1455 static int
1456 zone_ctor(void *mem, int size, void *udata, int flags)
1457 {
1458 	struct uma_zctor_args *arg = udata;
1459 	uma_zone_t zone = mem;
1460 	uma_zone_t z;
1461 	uma_keg_t keg;
1462 
1463 	bzero(zone, size);
1464 	zone->uz_name = arg->name;
1465 	zone->uz_ctor = arg->ctor;
1466 	zone->uz_dtor = arg->dtor;
1467 	zone->uz_slab = zone_fetch_slab;
1468 	zone->uz_init = NULL;
1469 	zone->uz_fini = NULL;
1470 	zone->uz_allocs = 0;
1471 	zone->uz_frees = 0;
1472 	zone->uz_fails = 0;
1473 	zone->uz_sleeps = 0;
1474 	zone->uz_count = 0;
1475 	zone->uz_flags = 0;
1476 	zone->uz_warning = NULL;
1477 	timevalclear(&zone->uz_ratecheck);
1478 	keg = arg->keg;
1479 
1480 	ZONE_LOCK_INIT(zone, (arg->flags & UMA_ZONE_MTXCLASS));
1481 
1482 	/*
1483 	 * This is a pure cache zone, no kegs.
1484 	 */
1485 	if (arg->import) {
1486 		if (arg->flags & UMA_ZONE_VM)
1487 			arg->flags |= UMA_ZFLAG_CACHEONLY;
1488 		zone->uz_flags = arg->flags;
1489 		zone->uz_size = arg->size;
1490 		zone->uz_import = arg->import;
1491 		zone->uz_release = arg->release;
1492 		zone->uz_arg = arg->arg;
1493 		zone->uz_lockptr = &zone->uz_lock;
1494 		goto out;
1495 	}
1496 
1497 	/*
1498 	 * Use the regular zone/keg/slab allocator.
1499 	 */
1500 	zone->uz_import = (uma_import)zone_import;
1501 	zone->uz_release = (uma_release)zone_release;
1502 	zone->uz_arg = zone;
1503 
1504 	if (arg->flags & UMA_ZONE_SECONDARY) {
1505 		KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
1506 		zone->uz_init = arg->uminit;
1507 		zone->uz_fini = arg->fini;
1508 		zone->uz_lockptr = &keg->uk_lock;
1509 		zone->uz_flags |= UMA_ZONE_SECONDARY;
1510 		mtx_lock(&uma_mtx);
1511 		ZONE_LOCK(zone);
1512 		LIST_FOREACH(z, &keg->uk_zones, uz_link) {
1513 			if (LIST_NEXT(z, uz_link) == NULL) {
1514 				LIST_INSERT_AFTER(z, zone, uz_link);
1515 				break;
1516 			}
1517 		}
1518 		ZONE_UNLOCK(zone);
1519 		mtx_unlock(&uma_mtx);
1520 	} else if (keg == NULL) {
1521 		if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
1522 		    arg->align, arg->flags)) == NULL)
1523 			return (ENOMEM);
1524 	} else {
1525 		struct uma_kctor_args karg;
1526 		int error;
1527 
1528 		/* We should only be here from uma_startup() */
1529 		karg.size = arg->size;
1530 		karg.uminit = arg->uminit;
1531 		karg.fini = arg->fini;
1532 		karg.align = arg->align;
1533 		karg.flags = arg->flags;
1534 		karg.zone = zone;
1535 		error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
1536 		    flags);
1537 		if (error)
1538 			return (error);
1539 	}
1540 
1541 	/*
1542 	 * Link in the first keg.
1543 	 */
1544 	zone->uz_klink.kl_keg = keg;
1545 	LIST_INSERT_HEAD(&zone->uz_kegs, &zone->uz_klink, kl_link);
1546 	zone->uz_lockptr = &keg->uk_lock;
1547 	zone->uz_size = keg->uk_size;
1548 	zone->uz_flags |= (keg->uk_flags &
1549 	    (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
1550 
1551 	/*
1552 	 * Some internal zones don't have room allocated for the per cpu
1553 	 * caches.  If we're internal, bail out here.
1554 	 */
1555 	if (keg->uk_flags & UMA_ZFLAG_INTERNAL) {
1556 		KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
1557 		    ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
1558 		return (0);
1559 	}
1560 
1561 out:
1562 	if ((arg->flags & UMA_ZONE_MAXBUCKET) == 0)
1563 		zone->uz_count = bucket_select(zone->uz_size);
1564 	else
1565 		zone->uz_count = BUCKET_MAX;
1566 
1567 	return (0);
1568 }
1569 
1570 /*
1571  * Keg header dtor.  This frees all data, destroys locks, frees the hash
1572  * table and removes the keg from the global list.
1573  *
1574  * Arguments/Returns follow uma_dtor specifications
1575  *	udata  unused
1576  */
1577 static void
1578 keg_dtor(void *arg, int size, void *udata)
1579 {
1580 	uma_keg_t keg;
1581 
1582 	keg = (uma_keg_t)arg;
1583 	KEG_LOCK(keg);
1584 	if (keg->uk_free != 0) {
1585 		printf("Freed UMA keg was not empty (%d items). "
1586 		    " Lost %d pages of memory.\n",
1587 		    keg->uk_free, keg->uk_pages);
1588 	}
1589 	KEG_UNLOCK(keg);
1590 
1591 	hash_free(&keg->uk_hash);
1592 
1593 	KEG_LOCK_FINI(keg);
1594 }
1595 
1596 /*
1597  * Zone header dtor.
1598  *
1599  * Arguments/Returns follow uma_dtor specifications
1600  *	udata  unused
1601  */
1602 static void
1603 zone_dtor(void *arg, int size, void *udata)
1604 {
1605 	uma_klink_t klink;
1606 	uma_zone_t zone;
1607 	uma_keg_t keg;
1608 
1609 	zone = (uma_zone_t)arg;
1610 	keg = zone_first_keg(zone);
1611 
1612 	if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
1613 		cache_drain(zone);
1614 
1615 	mtx_lock(&uma_mtx);
1616 	LIST_REMOVE(zone, uz_link);
1617 	mtx_unlock(&uma_mtx);
1618 	/*
1619 	 * XXX there are some races here where
1620 	 * the zone can be drained but zone lock
1621 	 * released and then refilled before we
1622 	 * remove it... we dont care for now
1623 	 */
1624 	zone_drain_wait(zone, M_WAITOK);
1625 	/*
1626 	 * Unlink all of our kegs.
1627 	 */
1628 	while ((klink = LIST_FIRST(&zone->uz_kegs)) != NULL) {
1629 		klink->kl_keg = NULL;
1630 		LIST_REMOVE(klink, kl_link);
1631 		if (klink == &zone->uz_klink)
1632 			continue;
1633 		free(klink, M_TEMP);
1634 	}
1635 	/*
1636 	 * We only destroy kegs from non secondary zones.
1637 	 */
1638 	if (keg != NULL && (zone->uz_flags & UMA_ZONE_SECONDARY) == 0)  {
1639 		mtx_lock(&uma_mtx);
1640 		LIST_REMOVE(keg, uk_link);
1641 		mtx_unlock(&uma_mtx);
1642 		zone_free_item(kegs, keg, NULL, SKIP_NONE);
1643 	}
1644 	ZONE_LOCK_FINI(zone);
1645 }
1646 
1647 /*
1648  * Traverses every zone in the system and calls a callback
1649  *
1650  * Arguments:
1651  *	zfunc  A pointer to a function which accepts a zone
1652  *		as an argument.
1653  *
1654  * Returns:
1655  *	Nothing
1656  */
1657 static void
1658 zone_foreach(void (*zfunc)(uma_zone_t))
1659 {
1660 	uma_keg_t keg;
1661 	uma_zone_t zone;
1662 
1663 	mtx_lock(&uma_mtx);
1664 	LIST_FOREACH(keg, &uma_kegs, uk_link) {
1665 		LIST_FOREACH(zone, &keg->uk_zones, uz_link)
1666 			zfunc(zone);
1667 	}
1668 	mtx_unlock(&uma_mtx);
1669 }
1670 
1671 /* Public functions */
1672 /* See uma.h */
1673 void
1674 uma_startup(void *bootmem, int boot_pages)
1675 {
1676 	struct uma_zctor_args args;
1677 	uma_slab_t slab;
1678 	u_int slabsize;
1679 	int i;
1680 
1681 #ifdef UMA_DEBUG
1682 	printf("Creating uma keg headers zone and keg.\n");
1683 #endif
1684 	mtx_init(&uma_mtx, "UMA lock", NULL, MTX_DEF);
1685 
1686 	/* "manually" create the initial zone */
1687 	memset(&args, 0, sizeof(args));
1688 	args.name = "UMA Kegs";
1689 	args.size = sizeof(struct uma_keg);
1690 	args.ctor = keg_ctor;
1691 	args.dtor = keg_dtor;
1692 	args.uminit = zero_init;
1693 	args.fini = NULL;
1694 	args.keg = &masterkeg;
1695 	args.align = 32 - 1;
1696 	args.flags = UMA_ZFLAG_INTERNAL;
1697 	/* The initial zone has no Per cpu queues so it's smaller */
1698 	zone_ctor(kegs, sizeof(struct uma_zone), &args, M_WAITOK);
1699 
1700 #ifdef UMA_DEBUG
1701 	printf("Filling boot free list.\n");
1702 #endif
1703 	for (i = 0; i < boot_pages; i++) {
1704 		slab = (uma_slab_t)((uint8_t *)bootmem + (i * UMA_SLAB_SIZE));
1705 		slab->us_data = (uint8_t *)slab;
1706 		slab->us_flags = UMA_SLAB_BOOT;
1707 		LIST_INSERT_HEAD(&uma_boot_pages, slab, us_link);
1708 	}
1709 	mtx_init(&uma_boot_pages_mtx, "UMA boot pages", NULL, MTX_DEF);
1710 
1711 #ifdef UMA_DEBUG
1712 	printf("Creating uma zone headers zone and keg.\n");
1713 #endif
1714 	args.name = "UMA Zones";
1715 	args.size = sizeof(struct uma_zone) +
1716 	    (sizeof(struct uma_cache) * (mp_maxid + 1));
1717 	args.ctor = zone_ctor;
1718 	args.dtor = zone_dtor;
1719 	args.uminit = zero_init;
1720 	args.fini = NULL;
1721 	args.keg = NULL;
1722 	args.align = 32 - 1;
1723 	args.flags = UMA_ZFLAG_INTERNAL;
1724 	/* The initial zone has no Per cpu queues so it's smaller */
1725 	zone_ctor(zones, sizeof(struct uma_zone), &args, M_WAITOK);
1726 
1727 #ifdef UMA_DEBUG
1728 	printf("Initializing pcpu cache locks.\n");
1729 #endif
1730 #ifdef UMA_DEBUG
1731 	printf("Creating slab and hash zones.\n");
1732 #endif
1733 
1734 	/* Now make a zone for slab headers */
1735 	slabzone = uma_zcreate("UMA Slabs",
1736 				sizeof(struct uma_slab),
1737 				NULL, NULL, NULL, NULL,
1738 				UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1739 
1740 	/*
1741 	 * We also create a zone for the bigger slabs with reference
1742 	 * counts in them, to accomodate UMA_ZONE_REFCNT zones.
1743 	 */
1744 	slabsize = sizeof(struct uma_slab_refcnt);
1745 	slabsize += uma_max_ipers_ref * sizeof(uint32_t);
1746 	slabrefzone = uma_zcreate("UMA RCntSlabs",
1747 				  slabsize,
1748 				  NULL, NULL, NULL, NULL,
1749 				  UMA_ALIGN_PTR,
1750 				  UMA_ZFLAG_INTERNAL);
1751 
1752 	hashzone = uma_zcreate("UMA Hash",
1753 	    sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
1754 	    NULL, NULL, NULL, NULL,
1755 	    UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1756 
1757 	bucket_init();
1758 
1759 	booted = UMA_STARTUP;
1760 
1761 #ifdef UMA_DEBUG
1762 	printf("UMA startup complete.\n");
1763 #endif
1764 }
1765 
1766 /* see uma.h */
1767 void
1768 uma_startup2(void)
1769 {
1770 	booted = UMA_STARTUP2;
1771 	bucket_enable();
1772 #ifdef UMA_DEBUG
1773 	printf("UMA startup2 complete.\n");
1774 #endif
1775 }
1776 
1777 /*
1778  * Initialize our callout handle
1779  *
1780  */
1781 
1782 static void
1783 uma_startup3(void)
1784 {
1785 #ifdef UMA_DEBUG
1786 	printf("Starting callout.\n");
1787 #endif
1788 	callout_init(&uma_callout, CALLOUT_MPSAFE);
1789 	callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
1790 #ifdef UMA_DEBUG
1791 	printf("UMA startup3 complete.\n");
1792 #endif
1793 }
1794 
1795 static uma_keg_t
1796 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
1797 		int align, uint32_t flags)
1798 {
1799 	struct uma_kctor_args args;
1800 
1801 	args.size = size;
1802 	args.uminit = uminit;
1803 	args.fini = fini;
1804 	args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
1805 	args.flags = flags;
1806 	args.zone = zone;
1807 	return (zone_alloc_item(kegs, &args, M_WAITOK));
1808 }
1809 
1810 /* See uma.h */
1811 void
1812 uma_set_align(int align)
1813 {
1814 
1815 	if (align != UMA_ALIGN_CACHE)
1816 		uma_align_cache = align;
1817 }
1818 
1819 /* See uma.h */
1820 uma_zone_t
1821 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
1822 		uma_init uminit, uma_fini fini, int align, uint32_t flags)
1823 
1824 {
1825 	struct uma_zctor_args args;
1826 
1827 	/* This stuff is essential for the zone ctor */
1828 	memset(&args, 0, sizeof(args));
1829 	args.name = name;
1830 	args.size = size;
1831 	args.ctor = ctor;
1832 	args.dtor = dtor;
1833 	args.uminit = uminit;
1834 	args.fini = fini;
1835 	args.align = align;
1836 	args.flags = flags;
1837 	args.keg = NULL;
1838 
1839 	return (zone_alloc_item(zones, &args, M_WAITOK));
1840 }
1841 
1842 /* See uma.h */
1843 uma_zone_t
1844 uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor,
1845 		    uma_init zinit, uma_fini zfini, uma_zone_t master)
1846 {
1847 	struct uma_zctor_args args;
1848 	uma_keg_t keg;
1849 
1850 	keg = zone_first_keg(master);
1851 	memset(&args, 0, sizeof(args));
1852 	args.name = name;
1853 	args.size = keg->uk_size;
1854 	args.ctor = ctor;
1855 	args.dtor = dtor;
1856 	args.uminit = zinit;
1857 	args.fini = zfini;
1858 	args.align = keg->uk_align;
1859 	args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
1860 	args.keg = keg;
1861 
1862 	/* XXX Attaches only one keg of potentially many. */
1863 	return (zone_alloc_item(zones, &args, M_WAITOK));
1864 }
1865 
1866 /* See uma.h */
1867 uma_zone_t
1868 uma_zcache_create(char *name, int size, uma_ctor ctor, uma_dtor dtor,
1869 		    uma_init zinit, uma_fini zfini, uma_import zimport,
1870 		    uma_release zrelease, void *arg, int flags)
1871 {
1872 	struct uma_zctor_args args;
1873 
1874 	memset(&args, 0, sizeof(args));
1875 	args.name = name;
1876 	args.size = size;
1877 	args.ctor = ctor;
1878 	args.dtor = dtor;
1879 	args.uminit = zinit;
1880 	args.fini = zfini;
1881 	args.import = zimport;
1882 	args.release = zrelease;
1883 	args.arg = arg;
1884 	args.align = 0;
1885 	args.flags = flags;
1886 
1887 	return (zone_alloc_item(zones, &args, M_WAITOK));
1888 }
1889 
1890 static void
1891 zone_lock_pair(uma_zone_t a, uma_zone_t b)
1892 {
1893 	if (a < b) {
1894 		ZONE_LOCK(a);
1895 		mtx_lock_flags(b->uz_lockptr, MTX_DUPOK);
1896 	} else {
1897 		ZONE_LOCK(b);
1898 		mtx_lock_flags(a->uz_lockptr, MTX_DUPOK);
1899 	}
1900 }
1901 
1902 static void
1903 zone_unlock_pair(uma_zone_t a, uma_zone_t b)
1904 {
1905 
1906 	ZONE_UNLOCK(a);
1907 	ZONE_UNLOCK(b);
1908 }
1909 
1910 int
1911 uma_zsecond_add(uma_zone_t zone, uma_zone_t master)
1912 {
1913 	uma_klink_t klink;
1914 	uma_klink_t kl;
1915 	int error;
1916 
1917 	error = 0;
1918 	klink = malloc(sizeof(*klink), M_TEMP, M_WAITOK | M_ZERO);
1919 
1920 	zone_lock_pair(zone, master);
1921 	/*
1922 	 * zone must use vtoslab() to resolve objects and must already be
1923 	 * a secondary.
1924 	 */
1925 	if ((zone->uz_flags & (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY))
1926 	    != (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) {
1927 		error = EINVAL;
1928 		goto out;
1929 	}
1930 	/*
1931 	 * The new master must also use vtoslab().
1932 	 */
1933 	if ((zone->uz_flags & UMA_ZONE_VTOSLAB) != UMA_ZONE_VTOSLAB) {
1934 		error = EINVAL;
1935 		goto out;
1936 	}
1937 	/*
1938 	 * Both must either be refcnt, or not be refcnt.
1939 	 */
1940 	if ((zone->uz_flags & UMA_ZONE_REFCNT) !=
1941 	    (master->uz_flags & UMA_ZONE_REFCNT)) {
1942 		error = EINVAL;
1943 		goto out;
1944 	}
1945 	/*
1946 	 * The underlying object must be the same size.  rsize
1947 	 * may be different.
1948 	 */
1949 	if (master->uz_size != zone->uz_size) {
1950 		error = E2BIG;
1951 		goto out;
1952 	}
1953 	/*
1954 	 * Put it at the end of the list.
1955 	 */
1956 	klink->kl_keg = zone_first_keg(master);
1957 	LIST_FOREACH(kl, &zone->uz_kegs, kl_link) {
1958 		if (LIST_NEXT(kl, kl_link) == NULL) {
1959 			LIST_INSERT_AFTER(kl, klink, kl_link);
1960 			break;
1961 		}
1962 	}
1963 	klink = NULL;
1964 	zone->uz_flags |= UMA_ZFLAG_MULTI;
1965 	zone->uz_slab = zone_fetch_slab_multi;
1966 
1967 out:
1968 	zone_unlock_pair(zone, master);
1969 	if (klink != NULL)
1970 		free(klink, M_TEMP);
1971 
1972 	return (error);
1973 }
1974 
1975 
1976 /* See uma.h */
1977 void
1978 uma_zdestroy(uma_zone_t zone)
1979 {
1980 
1981 	zone_free_item(zones, zone, NULL, SKIP_NONE);
1982 }
1983 
1984 /* See uma.h */
1985 void *
1986 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
1987 {
1988 	void *item;
1989 	uma_cache_t cache;
1990 	uma_bucket_t bucket;
1991 	int lockfail;
1992 	int cpu;
1993 
1994 	/* This is the fast path allocation */
1995 #ifdef UMA_DEBUG_ALLOC_1
1996 	printf("Allocating one item from %s(%p)\n", zone->uz_name, zone);
1997 #endif
1998 	CTR3(KTR_UMA, "uma_zalloc_arg thread %x zone %s flags %d", curthread,
1999 	    zone->uz_name, flags);
2000 
2001 	if (flags & M_WAITOK) {
2002 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
2003 		    "uma_zalloc_arg: zone \"%s\"", zone->uz_name);
2004 	}
2005 #ifdef DEBUG_MEMGUARD
2006 	if (memguard_cmp_zone(zone)) {
2007 		item = memguard_alloc(zone->uz_size, flags);
2008 		if (item != NULL) {
2009 			/*
2010 			 * Avoid conflict with the use-after-free
2011 			 * protecting infrastructure from INVARIANTS.
2012 			 */
2013 			if (zone->uz_init != NULL &&
2014 			    zone->uz_init != mtrash_init &&
2015 			    zone->uz_init(item, zone->uz_size, flags) != 0)
2016 				return (NULL);
2017 			if (zone->uz_ctor != NULL &&
2018 			    zone->uz_ctor != mtrash_ctor &&
2019 			    zone->uz_ctor(item, zone->uz_size, udata,
2020 			    flags) != 0) {
2021 			    	zone->uz_fini(item, zone->uz_size);
2022 				return (NULL);
2023 			}
2024 			return (item);
2025 		}
2026 		/* This is unfortunate but should not be fatal. */
2027 	}
2028 #endif
2029 	/*
2030 	 * If possible, allocate from the per-CPU cache.  There are two
2031 	 * requirements for safe access to the per-CPU cache: (1) the thread
2032 	 * accessing the cache must not be preempted or yield during access,
2033 	 * and (2) the thread must not migrate CPUs without switching which
2034 	 * cache it accesses.  We rely on a critical section to prevent
2035 	 * preemption and migration.  We release the critical section in
2036 	 * order to acquire the zone mutex if we are unable to allocate from
2037 	 * the current cache; when we re-acquire the critical section, we
2038 	 * must detect and handle migration if it has occurred.
2039 	 */
2040 	critical_enter();
2041 	cpu = curcpu;
2042 	cache = &zone->uz_cpu[cpu];
2043 
2044 zalloc_start:
2045 	bucket = cache->uc_allocbucket;
2046 	if (bucket != NULL && bucket->ub_cnt > 0) {
2047 		bucket->ub_cnt--;
2048 		item = bucket->ub_bucket[bucket->ub_cnt];
2049 #ifdef INVARIANTS
2050 		bucket->ub_bucket[bucket->ub_cnt] = NULL;
2051 #endif
2052 		KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled."));
2053 		cache->uc_allocs++;
2054 		critical_exit();
2055 		if (zone->uz_ctor != NULL &&
2056 		    zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2057 			atomic_add_long(&zone->uz_fails, 1);
2058 			zone_free_item(zone, item, udata, SKIP_DTOR);
2059 			return (NULL);
2060 		}
2061 #ifdef INVARIANTS
2062 		uma_dbg_alloc(zone, NULL, item);
2063 #endif
2064 		if (flags & M_ZERO)
2065 			bzero(item, zone->uz_size);
2066 		return (item);
2067 	}
2068 
2069 	/*
2070 	 * We have run out of items in our alloc bucket.
2071 	 * See if we can switch with our free bucket.
2072 	 */
2073 	bucket = cache->uc_freebucket;
2074 	if (bucket != NULL && bucket->ub_cnt > 0) {
2075 #ifdef UMA_DEBUG_ALLOC
2076 		printf("uma_zalloc: Swapping empty with alloc.\n");
2077 #endif
2078 		cache->uc_freebucket = cache->uc_allocbucket;
2079 		cache->uc_allocbucket = bucket;
2080 		goto zalloc_start;
2081 	}
2082 
2083 	/*
2084 	 * Discard any empty allocation bucket while we hold no locks.
2085 	 */
2086 	bucket = cache->uc_allocbucket;
2087 	cache->uc_allocbucket = NULL;
2088 	critical_exit();
2089 	if (bucket != NULL)
2090 		bucket_free(zone, bucket, udata);
2091 
2092 	/* Short-circuit for zones without buckets and low memory. */
2093 	if (zone->uz_count == 0 || bucketdisable)
2094 		goto zalloc_item;
2095 
2096 	/*
2097 	 * Attempt to retrieve the item from the per-CPU cache has failed, so
2098 	 * we must go back to the zone.  This requires the zone lock, so we
2099 	 * must drop the critical section, then re-acquire it when we go back
2100 	 * to the cache.  Since the critical section is released, we may be
2101 	 * preempted or migrate.  As such, make sure not to maintain any
2102 	 * thread-local state specific to the cache from prior to releasing
2103 	 * the critical section.
2104 	 */
2105 	lockfail = 0;
2106 	if (ZONE_TRYLOCK(zone) == 0) {
2107 		/* Record contention to size the buckets. */
2108 		ZONE_LOCK(zone);
2109 		lockfail = 1;
2110 	}
2111 	critical_enter();
2112 	cpu = curcpu;
2113 	cache = &zone->uz_cpu[cpu];
2114 
2115 	/*
2116 	 * Since we have locked the zone we may as well send back our stats.
2117 	 */
2118 	atomic_add_long(&zone->uz_allocs, cache->uc_allocs);
2119 	atomic_add_long(&zone->uz_frees, cache->uc_frees);
2120 	cache->uc_allocs = 0;
2121 	cache->uc_frees = 0;
2122 
2123 	/* See if we lost the race to fill the cache. */
2124 	if (cache->uc_allocbucket != NULL) {
2125 		ZONE_UNLOCK(zone);
2126 		goto zalloc_start;
2127 	}
2128 
2129 	/*
2130 	 * Check the zone's cache of buckets.
2131 	 */
2132 	if ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) {
2133 		KASSERT(bucket->ub_cnt != 0,
2134 		    ("uma_zalloc_arg: Returning an empty bucket."));
2135 
2136 		LIST_REMOVE(bucket, ub_link);
2137 		cache->uc_allocbucket = bucket;
2138 		ZONE_UNLOCK(zone);
2139 		goto zalloc_start;
2140 	}
2141 	/* We are no longer associated with this CPU. */
2142 	critical_exit();
2143 
2144 	/*
2145 	 * We bump the uz count when the cache size is insufficient to
2146 	 * handle the working set.
2147 	 */
2148 	if (lockfail && zone->uz_count < BUCKET_MAX)
2149 		zone->uz_count++;
2150 	ZONE_UNLOCK(zone);
2151 
2152 	/*
2153 	 * Now lets just fill a bucket and put it on the free list.  If that
2154 	 * works we'll restart the allocation from the begining and it
2155 	 * will use the just filled bucket.
2156 	 */
2157 	bucket = zone_alloc_bucket(zone, udata, flags);
2158 	if (bucket != NULL) {
2159 		ZONE_LOCK(zone);
2160 		critical_enter();
2161 		cpu = curcpu;
2162 		cache = &zone->uz_cpu[cpu];
2163 		/*
2164 		 * See if we lost the race or were migrated.  Cache the
2165 		 * initialized bucket to make this less likely or claim
2166 		 * the memory directly.
2167 		 */
2168 		if (cache->uc_allocbucket == NULL)
2169 			cache->uc_allocbucket = bucket;
2170 		else
2171 			LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link);
2172 		ZONE_UNLOCK(zone);
2173 		goto zalloc_start;
2174 	}
2175 
2176 	/*
2177 	 * We may not be able to get a bucket so return an actual item.
2178 	 */
2179 #ifdef UMA_DEBUG
2180 	printf("uma_zalloc_arg: Bucketzone returned NULL\n");
2181 #endif
2182 
2183 zalloc_item:
2184 	item = zone_alloc_item(zone, udata, flags);
2185 
2186 	return (item);
2187 }
2188 
2189 static uma_slab_t
2190 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int flags)
2191 {
2192 	uma_slab_t slab;
2193 	int reserve;
2194 
2195 	mtx_assert(&keg->uk_lock, MA_OWNED);
2196 	slab = NULL;
2197 	reserve = 0;
2198 	if ((flags & M_USE_RESERVE) == 0)
2199 		reserve = keg->uk_reserve;
2200 
2201 	for (;;) {
2202 		/*
2203 		 * Find a slab with some space.  Prefer slabs that are partially
2204 		 * used over those that are totally full.  This helps to reduce
2205 		 * fragmentation.
2206 		 */
2207 		if (keg->uk_free > reserve) {
2208 			if (!LIST_EMPTY(&keg->uk_part_slab)) {
2209 				slab = LIST_FIRST(&keg->uk_part_slab);
2210 			} else {
2211 				slab = LIST_FIRST(&keg->uk_free_slab);
2212 				LIST_REMOVE(slab, us_link);
2213 				LIST_INSERT_HEAD(&keg->uk_part_slab, slab,
2214 				    us_link);
2215 			}
2216 			MPASS(slab->us_keg == keg);
2217 			return (slab);
2218 		}
2219 
2220 		/*
2221 		 * M_NOVM means don't ask at all!
2222 		 */
2223 		if (flags & M_NOVM)
2224 			break;
2225 
2226 		if (keg->uk_maxpages && keg->uk_pages >= keg->uk_maxpages) {
2227 			keg->uk_flags |= UMA_ZFLAG_FULL;
2228 			/*
2229 			 * If this is not a multi-zone, set the FULL bit.
2230 			 * Otherwise slab_multi() takes care of it.
2231 			 */
2232 			if ((zone->uz_flags & UMA_ZFLAG_MULTI) == 0) {
2233 				zone->uz_flags |= UMA_ZFLAG_FULL;
2234 				zone_log_warning(zone);
2235 			}
2236 			if (flags & M_NOWAIT)
2237 				break;
2238 			zone->uz_sleeps++;
2239 			msleep(keg, &keg->uk_lock, PVM, "keglimit", 0);
2240 			continue;
2241 		}
2242 		slab = keg_alloc_slab(keg, zone, flags);
2243 		/*
2244 		 * If we got a slab here it's safe to mark it partially used
2245 		 * and return.  We assume that the caller is going to remove
2246 		 * at least one item.
2247 		 */
2248 		if (slab) {
2249 			MPASS(slab->us_keg == keg);
2250 			LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2251 			return (slab);
2252 		}
2253 		/*
2254 		 * We might not have been able to get a slab but another cpu
2255 		 * could have while we were unlocked.  Check again before we
2256 		 * fail.
2257 		 */
2258 		flags |= M_NOVM;
2259 	}
2260 	return (slab);
2261 }
2262 
2263 static uma_slab_t
2264 zone_fetch_slab(uma_zone_t zone, uma_keg_t keg, int flags)
2265 {
2266 	uma_slab_t slab;
2267 
2268 	if (keg == NULL) {
2269 		keg = zone_first_keg(zone);
2270 		KEG_LOCK(keg);
2271 	}
2272 
2273 	for (;;) {
2274 		slab = keg_fetch_slab(keg, zone, flags);
2275 		if (slab)
2276 			return (slab);
2277 		if (flags & (M_NOWAIT | M_NOVM))
2278 			break;
2279 	}
2280 	KEG_UNLOCK(keg);
2281 	return (NULL);
2282 }
2283 
2284 /*
2285  * uma_zone_fetch_slab_multi:  Fetches a slab from one available keg.  Returns
2286  * with the keg locked.  On NULL no lock is held.
2287  *
2288  * The last pointer is used to seed the search.  It is not required.
2289  */
2290 static uma_slab_t
2291 zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int rflags)
2292 {
2293 	uma_klink_t klink;
2294 	uma_slab_t slab;
2295 	uma_keg_t keg;
2296 	int flags;
2297 	int empty;
2298 	int full;
2299 
2300 	/*
2301 	 * Don't wait on the first pass.  This will skip limit tests
2302 	 * as well.  We don't want to block if we can find a provider
2303 	 * without blocking.
2304 	 */
2305 	flags = (rflags & ~M_WAITOK) | M_NOWAIT;
2306 	/*
2307 	 * Use the last slab allocated as a hint for where to start
2308 	 * the search.
2309 	 */
2310 	if (last != NULL) {
2311 		slab = keg_fetch_slab(last, zone, flags);
2312 		if (slab)
2313 			return (slab);
2314 		KEG_UNLOCK(last);
2315 	}
2316 	/*
2317 	 * Loop until we have a slab incase of transient failures
2318 	 * while M_WAITOK is specified.  I'm not sure this is 100%
2319 	 * required but we've done it for so long now.
2320 	 */
2321 	for (;;) {
2322 		empty = 0;
2323 		full = 0;
2324 		/*
2325 		 * Search the available kegs for slabs.  Be careful to hold the
2326 		 * correct lock while calling into the keg layer.
2327 		 */
2328 		LIST_FOREACH(klink, &zone->uz_kegs, kl_link) {
2329 			keg = klink->kl_keg;
2330 			KEG_LOCK(keg);
2331 			if ((keg->uk_flags & UMA_ZFLAG_FULL) == 0) {
2332 				slab = keg_fetch_slab(keg, zone, flags);
2333 				if (slab)
2334 					return (slab);
2335 			}
2336 			if (keg->uk_flags & UMA_ZFLAG_FULL)
2337 				full++;
2338 			else
2339 				empty++;
2340 			KEG_UNLOCK(keg);
2341 		}
2342 		if (rflags & (M_NOWAIT | M_NOVM))
2343 			break;
2344 		flags = rflags;
2345 		/*
2346 		 * All kegs are full.  XXX We can't atomically check all kegs
2347 		 * and sleep so just sleep for a short period and retry.
2348 		 */
2349 		if (full && !empty) {
2350 			ZONE_LOCK(zone);
2351 			zone->uz_flags |= UMA_ZFLAG_FULL;
2352 			zone->uz_sleeps++;
2353 			zone_log_warning(zone);
2354 			msleep(zone, zone->uz_lockptr, PVM,
2355 			    "zonelimit", hz/100);
2356 			zone->uz_flags &= ~UMA_ZFLAG_FULL;
2357 			ZONE_UNLOCK(zone);
2358 			continue;
2359 		}
2360 	}
2361 	return (NULL);
2362 }
2363 
2364 static void *
2365 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
2366 {
2367 	void *item;
2368 	uint8_t freei;
2369 
2370 	MPASS(keg == slab->us_keg);
2371 	mtx_assert(&keg->uk_lock, MA_OWNED);
2372 
2373 	freei = BIT_FFS(SLAB_SETSIZE, &slab->us_free) - 1;
2374 	BIT_CLR(SLAB_SETSIZE, freei, &slab->us_free);
2375 	item = slab->us_data + (keg->uk_rsize * freei);
2376 	slab->us_freecount--;
2377 	keg->uk_free--;
2378 
2379 	/* Move this slab to the full list */
2380 	if (slab->us_freecount == 0) {
2381 		LIST_REMOVE(slab, us_link);
2382 		LIST_INSERT_HEAD(&keg->uk_full_slab, slab, us_link);
2383 	}
2384 
2385 	return (item);
2386 }
2387 
2388 static int
2389 zone_import(uma_zone_t zone, void **bucket, int max, int flags)
2390 {
2391 	uma_slab_t slab;
2392 	uma_keg_t keg;
2393 	int i;
2394 
2395 	slab = NULL;
2396 	keg = NULL;
2397 	/* Try to keep the buckets totally full */
2398 	for (i = 0; i < max; ) {
2399 		if ((slab = zone->uz_slab(zone, keg, flags)) == NULL)
2400 			break;
2401 		keg = slab->us_keg;
2402 		while (slab->us_freecount && i < max) {
2403 			bucket[i++] = slab_alloc_item(keg, slab);
2404 			if (keg->uk_free <= keg->uk_reserve)
2405 				break;
2406 		}
2407 		/* Don't grab more than one slab at a time. */
2408 		flags &= ~M_WAITOK;
2409 		flags |= M_NOWAIT;
2410 	}
2411 	if (slab != NULL)
2412 		KEG_UNLOCK(keg);
2413 
2414 	return i;
2415 }
2416 
2417 static uma_bucket_t
2418 zone_alloc_bucket(uma_zone_t zone, void *udata, int flags)
2419 {
2420 	uma_bucket_t bucket;
2421 	int max;
2422 
2423 	/* Don't wait for buckets, preserve caller's NOVM setting. */
2424 	bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
2425 	if (bucket == NULL)
2426 		goto out;
2427 
2428 	max = MIN(bucket->ub_entries, zone->uz_count);
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(zone, bucket, udata);
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 cpu;
2533 
2534 #ifdef UMA_DEBUG_ALLOC_1
2535 	printf("Freeing item %p to %s(%p)\n", item, zone->uz_name, zone);
2536 #endif
2537 	CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread,
2538 	    zone->uz_name);
2539 
2540         /* uma_zfree(..., NULL) does nothing, to match free(9). */
2541         if (item == NULL)
2542                 return;
2543 #ifdef DEBUG_MEMGUARD
2544 	if (is_memguard_addr(item)) {
2545 		if (zone->uz_dtor != NULL && zone->uz_dtor != mtrash_dtor)
2546 			zone->uz_dtor(item, zone->uz_size, udata);
2547 		if (zone->uz_fini != NULL && zone->uz_fini != mtrash_fini)
2548 			zone->uz_fini(item, zone->uz_size);
2549 		memguard_free(item);
2550 		return;
2551 	}
2552 #endif
2553 #ifdef INVARIANTS
2554 	if (zone->uz_flags & UMA_ZONE_MALLOC)
2555 		uma_dbg_free(zone, udata, item);
2556 	else
2557 		uma_dbg_free(zone, NULL, item);
2558 #endif
2559 	if (zone->uz_dtor != NULL)
2560 		zone->uz_dtor(item, zone->uz_size, udata);
2561 
2562 	/*
2563 	 * The race here is acceptable.  If we miss it we'll just have to wait
2564 	 * a little longer for the limits to be reset.
2565 	 */
2566 	if (zone->uz_flags & UMA_ZFLAG_FULL)
2567 		goto zfree_item;
2568 
2569 	/*
2570 	 * If possible, free to the per-CPU cache.  There are two
2571 	 * requirements for safe access to the per-CPU cache: (1) the thread
2572 	 * accessing the cache must not be preempted or yield during access,
2573 	 * and (2) the thread must not migrate CPUs without switching which
2574 	 * cache it accesses.  We rely on a critical section to prevent
2575 	 * preemption and migration.  We release the critical section in
2576 	 * order to acquire the zone mutex if we are unable to free to the
2577 	 * current cache; when we re-acquire the critical section, we must
2578 	 * detect and handle migration if it has occurred.
2579 	 */
2580 zfree_restart:
2581 	critical_enter();
2582 	cpu = curcpu;
2583 	cache = &zone->uz_cpu[cpu];
2584 
2585 zfree_start:
2586 	/*
2587 	 * Try to free into the allocbucket first to give LIFO ordering
2588 	 * for cache-hot datastructures.  Spill over into the freebucket
2589 	 * if necessary.  Alloc will swap them if one runs dry.
2590 	 */
2591 	bucket = cache->uc_allocbucket;
2592 	if (bucket == NULL || bucket->ub_cnt >= bucket->ub_entries)
2593 		bucket = cache->uc_freebucket;
2594 	if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) {
2595 		KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL,
2596 		    ("uma_zfree: Freeing to non free bucket index."));
2597 		bucket->ub_bucket[bucket->ub_cnt] = item;
2598 		bucket->ub_cnt++;
2599 		cache->uc_frees++;
2600 		critical_exit();
2601 		return;
2602 	}
2603 
2604 	/*
2605 	 * We must go back the zone, which requires acquiring the zone lock,
2606 	 * which in turn means we must release and re-acquire the critical
2607 	 * section.  Since the critical section is released, we may be
2608 	 * preempted or migrate.  As such, make sure not to maintain any
2609 	 * thread-local state specific to the cache from prior to releasing
2610 	 * the critical section.
2611 	 */
2612 	critical_exit();
2613 	if (zone->uz_count == 0 || bucketdisable)
2614 		goto zfree_item;
2615 
2616 	ZONE_LOCK(zone);
2617 	critical_enter();
2618 	cpu = curcpu;
2619 	cache = &zone->uz_cpu[cpu];
2620 
2621 	/*
2622 	 * Since we have locked the zone we may as well send back our stats.
2623 	 */
2624 	atomic_add_long(&zone->uz_allocs, cache->uc_allocs);
2625 	atomic_add_long(&zone->uz_frees, cache->uc_frees);
2626 	cache->uc_allocs = 0;
2627 	cache->uc_frees = 0;
2628 
2629 	bucket = cache->uc_freebucket;
2630 	if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) {
2631 		ZONE_UNLOCK(zone);
2632 		goto zfree_start;
2633 	}
2634 	cache->uc_freebucket = NULL;
2635 
2636 	/* Can we throw this on the zone full list? */
2637 	if (bucket != NULL) {
2638 #ifdef UMA_DEBUG_ALLOC
2639 		printf("uma_zfree: Putting old bucket on the free list.\n");
2640 #endif
2641 		/* ub_cnt is pointing to the last free item */
2642 		KASSERT(bucket->ub_cnt != 0,
2643 		    ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n"));
2644 		LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link);
2645 	}
2646 
2647 	/* We are no longer associated with this CPU. */
2648 	critical_exit();
2649 
2650 	/* And the zone.. */
2651 	ZONE_UNLOCK(zone);
2652 
2653 #ifdef UMA_DEBUG_ALLOC
2654 	printf("uma_zfree: Allocating new free bucket.\n");
2655 #endif
2656 	bucket = bucket_alloc(zone, udata, M_NOWAIT);
2657 	if (bucket) {
2658 		critical_enter();
2659 		cpu = curcpu;
2660 		cache = &zone->uz_cpu[cpu];
2661 		if (cache->uc_freebucket == NULL) {
2662 			cache->uc_freebucket = bucket;
2663 			goto zfree_start;
2664 		}
2665 		/*
2666 		 * We lost the race, start over.  We have to drop our
2667 		 * critical section to free the bucket.
2668 		 */
2669 		critical_exit();
2670 		bucket_free(zone, bucket, udata);
2671 		goto zfree_restart;
2672 	}
2673 
2674 	/*
2675 	 * If nothing else caught this, we'll just do an internal free.
2676 	 */
2677 zfree_item:
2678 	zone_free_item(zone, item, udata, SKIP_DTOR);
2679 
2680 	return;
2681 }
2682 
2683 static void
2684 slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item)
2685 {
2686 	uint8_t freei;
2687 
2688 	mtx_assert(&keg->uk_lock, MA_OWNED);
2689 	MPASS(keg == slab->us_keg);
2690 
2691 	/* Do we need to remove from any lists? */
2692 	if (slab->us_freecount+1 == keg->uk_ipers) {
2693 		LIST_REMOVE(slab, us_link);
2694 		LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
2695 	} else if (slab->us_freecount == 0) {
2696 		LIST_REMOVE(slab, us_link);
2697 		LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2698 	}
2699 
2700 	/* Slab management. */
2701 	freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
2702 	BIT_SET(SLAB_SETSIZE, freei, &slab->us_free);
2703 	slab->us_freecount++;
2704 
2705 	/* Keg statistics. */
2706 	keg->uk_free++;
2707 }
2708 
2709 static void
2710 zone_release(uma_zone_t zone, void **bucket, int cnt)
2711 {
2712 	void *item;
2713 	uma_slab_t slab;
2714 	uma_keg_t keg;
2715 	uint8_t *mem;
2716 	int clearfull;
2717 	int i;
2718 
2719 	clearfull = 0;
2720 	keg = zone_first_keg(zone);
2721 	KEG_LOCK(keg);
2722 	for (i = 0; i < cnt; i++) {
2723 		item = bucket[i];
2724 		if (!(zone->uz_flags & UMA_ZONE_VTOSLAB)) {
2725 			mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
2726 			if (zone->uz_flags & UMA_ZONE_HASH) {
2727 				slab = hash_sfind(&keg->uk_hash, mem);
2728 			} else {
2729 				mem += keg->uk_pgoff;
2730 				slab = (uma_slab_t)mem;
2731 			}
2732 		} else {
2733 			slab = vtoslab((vm_offset_t)item);
2734 			if (slab->us_keg != keg) {
2735 				KEG_UNLOCK(keg);
2736 				keg = slab->us_keg;
2737 				KEG_LOCK(keg);
2738 			}
2739 		}
2740 		slab_free_item(keg, slab, item);
2741 		if (keg->uk_flags & UMA_ZFLAG_FULL) {
2742 			if (keg->uk_pages < keg->uk_maxpages) {
2743 				keg->uk_flags &= ~UMA_ZFLAG_FULL;
2744 				clearfull = 1;
2745 			}
2746 
2747 			/*
2748 			 * We can handle one more allocation. Since we're
2749 			 * clearing ZFLAG_FULL, wake up all procs blocked
2750 			 * on pages. This should be uncommon, so keeping this
2751 			 * simple for now (rather than adding count of blocked
2752 			 * threads etc).
2753 			 */
2754 			wakeup(keg);
2755 		}
2756 	}
2757 	KEG_UNLOCK(keg);
2758 	if (clearfull) {
2759 		ZONE_LOCK(zone);
2760 		zone->uz_flags &= ~UMA_ZFLAG_FULL;
2761 		wakeup(zone);
2762 		ZONE_UNLOCK(zone);
2763 	}
2764 
2765 }
2766 
2767 /*
2768  * Frees a single item to any zone.
2769  *
2770  * Arguments:
2771  *	zone   The zone to free to
2772  *	item   The item we're freeing
2773  *	udata  User supplied data for the dtor
2774  *	skip   Skip dtors and finis
2775  */
2776 static void
2777 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
2778 {
2779 
2780 #ifdef INVARIANTS
2781 	if (skip == SKIP_NONE) {
2782 		if (zone->uz_flags & UMA_ZONE_MALLOC)
2783 			uma_dbg_free(zone, udata, item);
2784 		else
2785 			uma_dbg_free(zone, NULL, item);
2786 	}
2787 #endif
2788 	if (skip < SKIP_DTOR && zone->uz_dtor)
2789 		zone->uz_dtor(item, zone->uz_size, udata);
2790 
2791 	if (skip < SKIP_FINI && zone->uz_fini)
2792 		zone->uz_fini(item, zone->uz_size);
2793 
2794 	atomic_add_long(&zone->uz_frees, 1);
2795 	zone->uz_release(zone->uz_arg, &item, 1);
2796 }
2797 
2798 /* See uma.h */
2799 int
2800 uma_zone_set_max(uma_zone_t zone, int nitems)
2801 {
2802 	uma_keg_t keg;
2803 
2804 	keg = zone_first_keg(zone);
2805 	if (keg == NULL)
2806 		return (0);
2807 	KEG_LOCK(keg);
2808 	keg->uk_maxpages = (nitems / keg->uk_ipers) * keg->uk_ppera;
2809 	if (keg->uk_maxpages * keg->uk_ipers < nitems)
2810 		keg->uk_maxpages += keg->uk_ppera;
2811 	nitems = keg->uk_maxpages * keg->uk_ipers;
2812 	KEG_UNLOCK(keg);
2813 
2814 	return (nitems);
2815 }
2816 
2817 /* See uma.h */
2818 int
2819 uma_zone_get_max(uma_zone_t zone)
2820 {
2821 	int nitems;
2822 	uma_keg_t keg;
2823 
2824 	keg = zone_first_keg(zone);
2825 	if (keg == NULL)
2826 		return (0);
2827 	KEG_LOCK(keg);
2828 	nitems = keg->uk_maxpages * keg->uk_ipers;
2829 	KEG_UNLOCK(keg);
2830 
2831 	return (nitems);
2832 }
2833 
2834 /* See uma.h */
2835 void
2836 uma_zone_set_warning(uma_zone_t zone, const char *warning)
2837 {
2838 
2839 	ZONE_LOCK(zone);
2840 	zone->uz_warning = warning;
2841 	ZONE_UNLOCK(zone);
2842 }
2843 
2844 /* See uma.h */
2845 int
2846 uma_zone_get_cur(uma_zone_t zone)
2847 {
2848 	int64_t nitems;
2849 	u_int i;
2850 
2851 	ZONE_LOCK(zone);
2852 	nitems = zone->uz_allocs - zone->uz_frees;
2853 	CPU_FOREACH(i) {
2854 		/*
2855 		 * See the comment in sysctl_vm_zone_stats() regarding the
2856 		 * safety of accessing the per-cpu caches. With the zone lock
2857 		 * held, it is safe, but can potentially result in stale data.
2858 		 */
2859 		nitems += zone->uz_cpu[i].uc_allocs -
2860 		    zone->uz_cpu[i].uc_frees;
2861 	}
2862 	ZONE_UNLOCK(zone);
2863 
2864 	return (nitems < 0 ? 0 : nitems);
2865 }
2866 
2867 /* See uma.h */
2868 void
2869 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
2870 {
2871 	uma_keg_t keg;
2872 
2873 	keg = zone_first_keg(zone);
2874 	KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type"));
2875 	KEG_LOCK(keg);
2876 	KASSERT(keg->uk_pages == 0,
2877 	    ("uma_zone_set_init on non-empty keg"));
2878 	keg->uk_init = uminit;
2879 	KEG_UNLOCK(keg);
2880 }
2881 
2882 /* See uma.h */
2883 void
2884 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
2885 {
2886 	uma_keg_t keg;
2887 
2888 	keg = zone_first_keg(zone);
2889 	KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type"));
2890 	KEG_LOCK(keg);
2891 	KASSERT(keg->uk_pages == 0,
2892 	    ("uma_zone_set_fini on non-empty keg"));
2893 	keg->uk_fini = fini;
2894 	KEG_UNLOCK(keg);
2895 }
2896 
2897 /* See uma.h */
2898 void
2899 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
2900 {
2901 
2902 	ZONE_LOCK(zone);
2903 	KASSERT(zone_first_keg(zone)->uk_pages == 0,
2904 	    ("uma_zone_set_zinit on non-empty keg"));
2905 	zone->uz_init = zinit;
2906 	ZONE_UNLOCK(zone);
2907 }
2908 
2909 /* See uma.h */
2910 void
2911 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
2912 {
2913 
2914 	ZONE_LOCK(zone);
2915 	KASSERT(zone_first_keg(zone)->uk_pages == 0,
2916 	    ("uma_zone_set_zfini on non-empty keg"));
2917 	zone->uz_fini = zfini;
2918 	ZONE_UNLOCK(zone);
2919 }
2920 
2921 /* See uma.h */
2922 /* XXX uk_freef is not actually used with the zone locked */
2923 void
2924 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
2925 {
2926 	uma_keg_t keg;
2927 
2928 	keg = zone_first_keg(zone);
2929 	KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type"));
2930 	KEG_LOCK(keg);
2931 	keg->uk_freef = freef;
2932 	KEG_UNLOCK(keg);
2933 }
2934 
2935 /* See uma.h */
2936 /* XXX uk_allocf is not actually used with the zone locked */
2937 void
2938 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
2939 {
2940 	uma_keg_t keg;
2941 
2942 	keg = zone_first_keg(zone);
2943 	KEG_LOCK(keg);
2944 	keg->uk_allocf = allocf;
2945 	KEG_UNLOCK(keg);
2946 }
2947 
2948 /* See uma.h */
2949 void
2950 uma_zone_reserve(uma_zone_t zone, int items)
2951 {
2952 	uma_keg_t keg;
2953 
2954 	keg = zone_first_keg(zone);
2955 	if (keg == NULL)
2956 		return;
2957 	KEG_LOCK(keg);
2958 	keg->uk_reserve = items;
2959 	KEG_UNLOCK(keg);
2960 
2961 	return;
2962 }
2963 
2964 /* See uma.h */
2965 int
2966 uma_zone_reserve_kva(uma_zone_t zone, int count)
2967 {
2968 	uma_keg_t keg;
2969 	vm_offset_t kva;
2970 	int pages;
2971 
2972 	keg = zone_first_keg(zone);
2973 	if (keg == NULL)
2974 		return (0);
2975 	pages = count / keg->uk_ipers;
2976 
2977 	if (pages * keg->uk_ipers < count)
2978 		pages++;
2979 
2980 #ifdef UMA_MD_SMALL_ALLOC
2981 	if (keg->uk_ppera > 1) {
2982 #else
2983 	if (1) {
2984 #endif
2985 		kva = kmem_alloc_nofault(kernel_map, pages * UMA_SLAB_SIZE);
2986 		if (kva == 0)
2987 			return (0);
2988 	} else
2989 		kva = 0;
2990 	KEG_LOCK(keg);
2991 	keg->uk_kva = kva;
2992 	keg->uk_offset = 0;
2993 	keg->uk_maxpages = pages;
2994 #ifdef UMA_MD_SMALL_ALLOC
2995 	keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
2996 #else
2997 	keg->uk_allocf = noobj_alloc;
2998 #endif
2999 	keg->uk_flags |= UMA_ZONE_NOFREE;
3000 	KEG_UNLOCK(keg);
3001 
3002 	return (1);
3003 }
3004 
3005 /* See uma.h */
3006 void
3007 uma_prealloc(uma_zone_t zone, int items)
3008 {
3009 	int slabs;
3010 	uma_slab_t slab;
3011 	uma_keg_t keg;
3012 
3013 	keg = zone_first_keg(zone);
3014 	if (keg == NULL)
3015 		return;
3016 	KEG_LOCK(keg);
3017 	slabs = items / keg->uk_ipers;
3018 	if (slabs * keg->uk_ipers < items)
3019 		slabs++;
3020 	while (slabs > 0) {
3021 		slab = keg_alloc_slab(keg, zone, M_WAITOK);
3022 		if (slab == NULL)
3023 			break;
3024 		MPASS(slab->us_keg == keg);
3025 		LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
3026 		slabs--;
3027 	}
3028 	KEG_UNLOCK(keg);
3029 }
3030 
3031 /* See uma.h */
3032 uint32_t *
3033 uma_find_refcnt(uma_zone_t zone, void *item)
3034 {
3035 	uma_slabrefcnt_t slabref;
3036 	uma_slab_t slab;
3037 	uma_keg_t keg;
3038 	uint32_t *refcnt;
3039 	int idx;
3040 
3041 	slab = vtoslab((vm_offset_t)item & (~UMA_SLAB_MASK));
3042 	slabref = (uma_slabrefcnt_t)slab;
3043 	keg = slab->us_keg;
3044 	KASSERT(keg->uk_flags & UMA_ZONE_REFCNT,
3045 	    ("uma_find_refcnt(): zone possibly not UMA_ZONE_REFCNT"));
3046 	idx = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
3047 	refcnt = &slabref->us_refcnt[idx];
3048 	return refcnt;
3049 }
3050 
3051 /* See uma.h */
3052 void
3053 uma_reclaim(void)
3054 {
3055 #ifdef UMA_DEBUG
3056 	printf("UMA: vm asked us to release pages!\n");
3057 #endif
3058 	bucket_enable();
3059 	zone_foreach(zone_drain);
3060 	/*
3061 	 * Some slabs may have been freed but this zone will be visited early
3062 	 * we visit again so that we can free pages that are empty once other
3063 	 * zones are drained.  We have to do the same for buckets.
3064 	 */
3065 	zone_drain(slabzone);
3066 	zone_drain(slabrefzone);
3067 	bucket_zone_drain();
3068 }
3069 
3070 /* See uma.h */
3071 int
3072 uma_zone_exhausted(uma_zone_t zone)
3073 {
3074 	int full;
3075 
3076 	ZONE_LOCK(zone);
3077 	full = (zone->uz_flags & UMA_ZFLAG_FULL);
3078 	ZONE_UNLOCK(zone);
3079 	return (full);
3080 }
3081 
3082 int
3083 uma_zone_exhausted_nolock(uma_zone_t zone)
3084 {
3085 	return (zone->uz_flags & UMA_ZFLAG_FULL);
3086 }
3087 
3088 void *
3089 uma_large_malloc(int size, int wait)
3090 {
3091 	void *mem;
3092 	uma_slab_t slab;
3093 	uint8_t flags;
3094 
3095 	slab = zone_alloc_item(slabzone, NULL, wait);
3096 	if (slab == NULL)
3097 		return (NULL);
3098 	mem = page_alloc(NULL, size, &flags, wait);
3099 	if (mem) {
3100 		vsetslab((vm_offset_t)mem, slab);
3101 		slab->us_data = mem;
3102 		slab->us_flags = flags | UMA_SLAB_MALLOC;
3103 		slab->us_size = size;
3104 	} else {
3105 		zone_free_item(slabzone, slab, NULL, SKIP_NONE);
3106 	}
3107 
3108 	return (mem);
3109 }
3110 
3111 void
3112 uma_large_free(uma_slab_t slab)
3113 {
3114 	vsetobj((vm_offset_t)slab->us_data, kmem_object);
3115 	page_free(slab->us_data, slab->us_size, slab->us_flags);
3116 	zone_free_item(slabzone, slab, NULL, SKIP_NONE);
3117 }
3118 
3119 void
3120 uma_print_stats(void)
3121 {
3122 	zone_foreach(uma_print_zone);
3123 }
3124 
3125 static void
3126 slab_print(uma_slab_t slab)
3127 {
3128 	printf("slab: keg %p, data %p, freecount %d\n",
3129 		slab->us_keg, slab->us_data, slab->us_freecount);
3130 }
3131 
3132 static void
3133 cache_print(uma_cache_t cache)
3134 {
3135 	printf("alloc: %p(%d), free: %p(%d)\n",
3136 		cache->uc_allocbucket,
3137 		cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0,
3138 		cache->uc_freebucket,
3139 		cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0);
3140 }
3141 
3142 static void
3143 uma_print_keg(uma_keg_t keg)
3144 {
3145 	uma_slab_t slab;
3146 
3147 	printf("keg: %s(%p) size %d(%d) flags %#x ipers %d ppera %d "
3148 	    "out %d free %d limit %d\n",
3149 	    keg->uk_name, keg, keg->uk_size, keg->uk_rsize, keg->uk_flags,
3150 	    keg->uk_ipers, keg->uk_ppera,
3151 	    (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free,
3152 	    (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers);
3153 	printf("Part slabs:\n");
3154 	LIST_FOREACH(slab, &keg->uk_part_slab, us_link)
3155 		slab_print(slab);
3156 	printf("Free slabs:\n");
3157 	LIST_FOREACH(slab, &keg->uk_free_slab, us_link)
3158 		slab_print(slab);
3159 	printf("Full slabs:\n");
3160 	LIST_FOREACH(slab, &keg->uk_full_slab, us_link)
3161 		slab_print(slab);
3162 }
3163 
3164 void
3165 uma_print_zone(uma_zone_t zone)
3166 {
3167 	uma_cache_t cache;
3168 	uma_klink_t kl;
3169 	int i;
3170 
3171 	printf("zone: %s(%p) size %d flags %#x\n",
3172 	    zone->uz_name, zone, zone->uz_size, zone->uz_flags);
3173 	LIST_FOREACH(kl, &zone->uz_kegs, kl_link)
3174 		uma_print_keg(kl->kl_keg);
3175 	CPU_FOREACH(i) {
3176 		cache = &zone->uz_cpu[i];
3177 		printf("CPU %d Cache:\n", i);
3178 		cache_print(cache);
3179 	}
3180 }
3181 
3182 #ifdef DDB
3183 /*
3184  * Generate statistics across both the zone and its per-cpu cache's.  Return
3185  * desired statistics if the pointer is non-NULL for that statistic.
3186  *
3187  * Note: does not update the zone statistics, as it can't safely clear the
3188  * per-CPU cache statistic.
3189  *
3190  * XXXRW: Following the uc_allocbucket and uc_freebucket pointers here isn't
3191  * safe from off-CPU; we should modify the caches to track this information
3192  * directly so that we don't have to.
3193  */
3194 static void
3195 uma_zone_sumstat(uma_zone_t z, int *cachefreep, uint64_t *allocsp,
3196     uint64_t *freesp, uint64_t *sleepsp)
3197 {
3198 	uma_cache_t cache;
3199 	uint64_t allocs, frees, sleeps;
3200 	int cachefree, cpu;
3201 
3202 	allocs = frees = sleeps = 0;
3203 	cachefree = 0;
3204 	CPU_FOREACH(cpu) {
3205 		cache = &z->uz_cpu[cpu];
3206 		if (cache->uc_allocbucket != NULL)
3207 			cachefree += cache->uc_allocbucket->ub_cnt;
3208 		if (cache->uc_freebucket != NULL)
3209 			cachefree += cache->uc_freebucket->ub_cnt;
3210 		allocs += cache->uc_allocs;
3211 		frees += cache->uc_frees;
3212 	}
3213 	allocs += z->uz_allocs;
3214 	frees += z->uz_frees;
3215 	sleeps += z->uz_sleeps;
3216 	if (cachefreep != NULL)
3217 		*cachefreep = cachefree;
3218 	if (allocsp != NULL)
3219 		*allocsp = allocs;
3220 	if (freesp != NULL)
3221 		*freesp = frees;
3222 	if (sleepsp != NULL)
3223 		*sleepsp = sleeps;
3224 }
3225 #endif /* DDB */
3226 
3227 static int
3228 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
3229 {
3230 	uma_keg_t kz;
3231 	uma_zone_t z;
3232 	int count;
3233 
3234 	count = 0;
3235 	mtx_lock(&uma_mtx);
3236 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3237 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
3238 			count++;
3239 	}
3240 	mtx_unlock(&uma_mtx);
3241 	return (sysctl_handle_int(oidp, &count, 0, req));
3242 }
3243 
3244 static int
3245 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
3246 {
3247 	struct uma_stream_header ush;
3248 	struct uma_type_header uth;
3249 	struct uma_percpu_stat ups;
3250 	uma_bucket_t bucket;
3251 	struct sbuf sbuf;
3252 	uma_cache_t cache;
3253 	uma_klink_t kl;
3254 	uma_keg_t kz;
3255 	uma_zone_t z;
3256 	uma_keg_t k;
3257 	int count, error, i;
3258 
3259 	error = sysctl_wire_old_buffer(req, 0);
3260 	if (error != 0)
3261 		return (error);
3262 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
3263 
3264 	count = 0;
3265 	mtx_lock(&uma_mtx);
3266 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3267 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
3268 			count++;
3269 	}
3270 
3271 	/*
3272 	 * Insert stream header.
3273 	 */
3274 	bzero(&ush, sizeof(ush));
3275 	ush.ush_version = UMA_STREAM_VERSION;
3276 	ush.ush_maxcpus = (mp_maxid + 1);
3277 	ush.ush_count = count;
3278 	(void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
3279 
3280 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3281 		LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3282 			bzero(&uth, sizeof(uth));
3283 			ZONE_LOCK(z);
3284 			strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
3285 			uth.uth_align = kz->uk_align;
3286 			uth.uth_size = kz->uk_size;
3287 			uth.uth_rsize = kz->uk_rsize;
3288 			LIST_FOREACH(kl, &z->uz_kegs, kl_link) {
3289 				k = kl->kl_keg;
3290 				uth.uth_maxpages += k->uk_maxpages;
3291 				uth.uth_pages += k->uk_pages;
3292 				uth.uth_keg_free += k->uk_free;
3293 				uth.uth_limit = (k->uk_maxpages / k->uk_ppera)
3294 				    * k->uk_ipers;
3295 			}
3296 
3297 			/*
3298 			 * A zone is secondary is it is not the first entry
3299 			 * on the keg's zone list.
3300 			 */
3301 			if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
3302 			    (LIST_FIRST(&kz->uk_zones) != z))
3303 				uth.uth_zone_flags = UTH_ZONE_SECONDARY;
3304 
3305 			LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3306 				uth.uth_zone_free += bucket->ub_cnt;
3307 			uth.uth_allocs = z->uz_allocs;
3308 			uth.uth_frees = z->uz_frees;
3309 			uth.uth_fails = z->uz_fails;
3310 			uth.uth_sleeps = z->uz_sleeps;
3311 			(void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
3312 			/*
3313 			 * While it is not normally safe to access the cache
3314 			 * bucket pointers while not on the CPU that owns the
3315 			 * cache, we only allow the pointers to be exchanged
3316 			 * without the zone lock held, not invalidated, so
3317 			 * accept the possible race associated with bucket
3318 			 * exchange during monitoring.
3319 			 */
3320 			for (i = 0; i < (mp_maxid + 1); i++) {
3321 				bzero(&ups, sizeof(ups));
3322 				if (kz->uk_flags & UMA_ZFLAG_INTERNAL)
3323 					goto skip;
3324 				if (CPU_ABSENT(i))
3325 					goto skip;
3326 				cache = &z->uz_cpu[i];
3327 				if (cache->uc_allocbucket != NULL)
3328 					ups.ups_cache_free +=
3329 					    cache->uc_allocbucket->ub_cnt;
3330 				if (cache->uc_freebucket != NULL)
3331 					ups.ups_cache_free +=
3332 					    cache->uc_freebucket->ub_cnt;
3333 				ups.ups_allocs = cache->uc_allocs;
3334 				ups.ups_frees = cache->uc_frees;
3335 skip:
3336 				(void)sbuf_bcat(&sbuf, &ups, sizeof(ups));
3337 			}
3338 			ZONE_UNLOCK(z);
3339 		}
3340 	}
3341 	mtx_unlock(&uma_mtx);
3342 	error = sbuf_finish(&sbuf);
3343 	sbuf_delete(&sbuf);
3344 	return (error);
3345 }
3346 
3347 #ifdef DDB
3348 DB_SHOW_COMMAND(uma, db_show_uma)
3349 {
3350 	uint64_t allocs, frees, sleeps;
3351 	uma_bucket_t bucket;
3352 	uma_keg_t kz;
3353 	uma_zone_t z;
3354 	int cachefree;
3355 
3356 	db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
3357 	    "Requests", "Sleeps");
3358 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3359 		LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3360 			if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
3361 				allocs = z->uz_allocs;
3362 				frees = z->uz_frees;
3363 				sleeps = z->uz_sleeps;
3364 				cachefree = 0;
3365 			} else
3366 				uma_zone_sumstat(z, &cachefree, &allocs,
3367 				    &frees, &sleeps);
3368 			if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
3369 			    (LIST_FIRST(&kz->uk_zones) != z)))
3370 				cachefree += kz->uk_free;
3371 			LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3372 				cachefree += bucket->ub_cnt;
3373 			db_printf("%18s %8ju %8jd %8d %12ju %8ju\n", z->uz_name,
3374 			    (uintmax_t)kz->uk_size,
3375 			    (intmax_t)(allocs - frees), cachefree,
3376 			    (uintmax_t)allocs, sleeps);
3377 			if (db_pager_quit)
3378 				return;
3379 		}
3380 	}
3381 }
3382 #endif
3383