xref: /freebsd/sys/vm/uma_core.c (revision 5dae51da3da0cc94d17bd67b308fad304ebec7e0)
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  * efficient.  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/random.h>
77 #include <sys/rwlock.h>
78 #include <sys/sbuf.h>
79 #include <sys/sched.h>
80 #include <sys/smp.h>
81 #include <sys/taskqueue.h>
82 #include <sys/vmmeter.h>
83 
84 #include <vm/vm.h>
85 #include <vm/vm_object.h>
86 #include <vm/vm_page.h>
87 #include <vm/vm_pageout.h>
88 #include <vm/vm_param.h>
89 #include <vm/vm_map.h>
90 #include <vm/vm_kern.h>
91 #include <vm/vm_extern.h>
92 #include <vm/uma.h>
93 #include <vm/uma_int.h>
94 #include <vm/uma_dbg.h>
95 
96 #include <ddb/ddb.h>
97 
98 #ifdef DEBUG_MEMGUARD
99 #include <vm/memguard.h>
100 #endif
101 
102 /*
103  * This is the zone and keg from which all zones are spawned.  The idea is that
104  * even the zone & keg heads are allocated from the allocator, so we use the
105  * bss section to bootstrap us.
106  */
107 static struct uma_keg masterkeg;
108 static struct uma_zone masterzone_k;
109 static struct uma_zone masterzone_z;
110 static uma_zone_t kegs = &masterzone_k;
111 static uma_zone_t zones = &masterzone_z;
112 
113 /* This is the zone from which all of uma_slab_t's are allocated. */
114 static uma_zone_t slabzone;
115 
116 /*
117  * The initial hash tables come out of this zone so they can be allocated
118  * prior to malloc coming up.
119  */
120 static uma_zone_t hashzone;
121 
122 /* The boot-time adjusted value for cache line alignment. */
123 int uma_align_cache = 64 - 1;
124 
125 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets");
126 
127 /*
128  * Are we allowed to allocate buckets?
129  */
130 static int bucketdisable = 1;
131 
132 /* Linked list of all kegs in the system */
133 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs);
134 
135 /* Linked list of all cache-only zones in the system */
136 static LIST_HEAD(,uma_zone) uma_cachezones =
137     LIST_HEAD_INITIALIZER(uma_cachezones);
138 
139 /* This RW lock protects the keg list */
140 static struct rwlock_padalign uma_rwlock;
141 
142 /* Linked list of boot time pages */
143 static LIST_HEAD(,uma_slab) uma_boot_pages =
144     LIST_HEAD_INITIALIZER(uma_boot_pages);
145 
146 /* This mutex protects the boot time pages list */
147 static struct mtx_padalign uma_boot_pages_mtx;
148 
149 static struct sx uma_drain_lock;
150 
151 /* Is the VM done starting up? */
152 static int booted = 0;
153 #define	UMA_STARTUP	1
154 #define	UMA_STARTUP2	2
155 
156 /*
157  * This is the handle used to schedule events that need to happen
158  * outside of the allocation fast path.
159  */
160 static struct callout uma_callout;
161 #define	UMA_TIMEOUT	20		/* Seconds for callout interval. */
162 
163 /*
164  * This structure is passed as the zone ctor arg so that I don't have to create
165  * a special allocation function just for zones.
166  */
167 struct uma_zctor_args {
168 	const char *name;
169 	size_t size;
170 	uma_ctor ctor;
171 	uma_dtor dtor;
172 	uma_init uminit;
173 	uma_fini fini;
174 	uma_import import;
175 	uma_release release;
176 	void *arg;
177 	uma_keg_t keg;
178 	int align;
179 	uint32_t flags;
180 };
181 
182 struct uma_kctor_args {
183 	uma_zone_t zone;
184 	size_t size;
185 	uma_init uminit;
186 	uma_fini fini;
187 	int align;
188 	uint32_t flags;
189 };
190 
191 struct uma_bucket_zone {
192 	uma_zone_t	ubz_zone;
193 	char		*ubz_name;
194 	int		ubz_entries;	/* Number of items it can hold. */
195 	int		ubz_maxsize;	/* Maximum allocation size per-item. */
196 };
197 
198 /*
199  * Compute the actual number of bucket entries to pack them in power
200  * of two sizes for more efficient space utilization.
201  */
202 #define	BUCKET_SIZE(n)						\
203     (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *))
204 
205 #define	BUCKET_MAX	BUCKET_SIZE(256)
206 
207 struct uma_bucket_zone bucket_zones[] = {
208 	{ NULL, "4 Bucket", BUCKET_SIZE(4), 4096 },
209 	{ NULL, "6 Bucket", BUCKET_SIZE(6), 3072 },
210 	{ NULL, "8 Bucket", BUCKET_SIZE(8), 2048 },
211 	{ NULL, "12 Bucket", BUCKET_SIZE(12), 1536 },
212 	{ NULL, "16 Bucket", BUCKET_SIZE(16), 1024 },
213 	{ NULL, "32 Bucket", BUCKET_SIZE(32), 512 },
214 	{ NULL, "64 Bucket", BUCKET_SIZE(64), 256 },
215 	{ NULL, "128 Bucket", BUCKET_SIZE(128), 128 },
216 	{ NULL, "256 Bucket", BUCKET_SIZE(256), 64 },
217 	{ NULL, NULL, 0}
218 };
219 
220 /*
221  * Flags and enumerations to be passed to internal functions.
222  */
223 enum zfreeskip { SKIP_NONE = 0, SKIP_DTOR, SKIP_FINI };
224 
225 /* Prototypes.. */
226 
227 static void *noobj_alloc(uma_zone_t, vm_size_t, uint8_t *, int);
228 static void *page_alloc(uma_zone_t, vm_size_t, uint8_t *, int);
229 static void *startup_alloc(uma_zone_t, vm_size_t, uint8_t *, int);
230 static void page_free(void *, vm_size_t, uint8_t);
231 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int);
232 static void cache_drain(uma_zone_t);
233 static void bucket_drain(uma_zone_t, uma_bucket_t);
234 static void bucket_cache_drain(uma_zone_t zone);
235 static int keg_ctor(void *, int, void *, int);
236 static void keg_dtor(void *, int, void *);
237 static int zone_ctor(void *, int, void *, int);
238 static void zone_dtor(void *, int, void *);
239 static int zero_init(void *, int, int);
240 static void keg_small_init(uma_keg_t keg);
241 static void keg_large_init(uma_keg_t keg);
242 static void zone_foreach(void (*zfunc)(uma_zone_t));
243 static void zone_timeout(uma_zone_t zone);
244 static int hash_alloc(struct uma_hash *);
245 static int hash_expand(struct uma_hash *, struct uma_hash *);
246 static void hash_free(struct uma_hash *hash);
247 static void uma_timeout(void *);
248 static void uma_startup3(void);
249 static void *zone_alloc_item(uma_zone_t, void *, int);
250 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip);
251 static void bucket_enable(void);
252 static void bucket_init(void);
253 static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int);
254 static void bucket_free(uma_zone_t zone, uma_bucket_t, void *);
255 static void bucket_zone_drain(void);
256 static uma_bucket_t zone_alloc_bucket(uma_zone_t zone, void *, int flags);
257 static uma_slab_t zone_fetch_slab(uma_zone_t zone, uma_keg_t last, int flags);
258 static uma_slab_t zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int flags);
259 static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab);
260 static void slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item);
261 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit,
262     uma_fini fini, int align, uint32_t flags);
263 static int zone_import(uma_zone_t zone, void **bucket, int max, int flags);
264 static void zone_release(uma_zone_t zone, void **bucket, int cnt);
265 static void uma_zero_item(void *item, uma_zone_t zone);
266 
267 void uma_print_zone(uma_zone_t);
268 void uma_print_stats(void);
269 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS);
270 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS);
271 
272 #ifdef INVARIANTS
273 static void uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item);
274 static void uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item);
275 #endif
276 
277 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
278 
279 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLTYPE_INT,
280     0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones");
281 
282 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLTYPE_STRUCT,
283     0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats");
284 
285 static int zone_warnings = 1;
286 SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RWTUN, &zone_warnings, 0,
287     "Warn when UMA zones becomes full");
288 
289 /*
290  * This routine checks to see whether or not it's safe to enable buckets.
291  */
292 static void
293 bucket_enable(void)
294 {
295 	bucketdisable = vm_page_count_min();
296 }
297 
298 /*
299  * Initialize bucket_zones, the array of zones of buckets of various sizes.
300  *
301  * For each zone, calculate the memory required for each bucket, consisting
302  * of the header and an array of pointers.
303  */
304 static void
305 bucket_init(void)
306 {
307 	struct uma_bucket_zone *ubz;
308 	int size;
309 
310 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) {
311 		size = roundup(sizeof(struct uma_bucket), sizeof(void *));
312 		size += sizeof(void *) * ubz->ubz_entries;
313 		ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
314 		    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
315 		    UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET);
316 	}
317 }
318 
319 /*
320  * Given a desired number of entries for a bucket, return the zone from which
321  * to allocate the bucket.
322  */
323 static struct uma_bucket_zone *
324 bucket_zone_lookup(int entries)
325 {
326 	struct uma_bucket_zone *ubz;
327 
328 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
329 		if (ubz->ubz_entries >= entries)
330 			return (ubz);
331 	ubz--;
332 	return (ubz);
333 }
334 
335 static int
336 bucket_select(int size)
337 {
338 	struct uma_bucket_zone *ubz;
339 
340 	ubz = &bucket_zones[0];
341 	if (size > ubz->ubz_maxsize)
342 		return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1);
343 
344 	for (; ubz->ubz_entries != 0; ubz++)
345 		if (ubz->ubz_maxsize < size)
346 			break;
347 	ubz--;
348 	return (ubz->ubz_entries);
349 }
350 
351 static uma_bucket_t
352 bucket_alloc(uma_zone_t zone, void *udata, int flags)
353 {
354 	struct uma_bucket_zone *ubz;
355 	uma_bucket_t bucket;
356 
357 	/*
358 	 * This is to stop us from allocating per cpu buckets while we're
359 	 * running out of vm.boot_pages.  Otherwise, we would exhaust the
360 	 * boot pages.  This also prevents us from allocating buckets in
361 	 * low memory situations.
362 	 */
363 	if (bucketdisable)
364 		return (NULL);
365 	/*
366 	 * To limit bucket recursion we store the original zone flags
367 	 * in a cookie passed via zalloc_arg/zfree_arg.  This allows the
368 	 * NOVM flag to persist even through deep recursions.  We also
369 	 * store ZFLAG_BUCKET once we have recursed attempting to allocate
370 	 * a bucket for a bucket zone so we do not allow infinite bucket
371 	 * recursion.  This cookie will even persist to frees of unused
372 	 * buckets via the allocation path or bucket allocations in the
373 	 * free path.
374 	 */
375 	if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
376 		udata = (void *)(uintptr_t)zone->uz_flags;
377 	else {
378 		if ((uintptr_t)udata & UMA_ZFLAG_BUCKET)
379 			return (NULL);
380 		udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET);
381 	}
382 	if ((uintptr_t)udata & UMA_ZFLAG_CACHEONLY)
383 		flags |= M_NOVM;
384 	ubz = bucket_zone_lookup(zone->uz_count);
385 	if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0)
386 		ubz++;
387 	bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags);
388 	if (bucket) {
389 #ifdef INVARIANTS
390 		bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
391 #endif
392 		bucket->ub_cnt = 0;
393 		bucket->ub_entries = ubz->ubz_entries;
394 	}
395 
396 	return (bucket);
397 }
398 
399 static void
400 bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata)
401 {
402 	struct uma_bucket_zone *ubz;
403 
404 	KASSERT(bucket->ub_cnt == 0,
405 	    ("bucket_free: Freeing a non free bucket."));
406 	if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
407 		udata = (void *)(uintptr_t)zone->uz_flags;
408 	ubz = bucket_zone_lookup(bucket->ub_entries);
409 	uma_zfree_arg(ubz->ubz_zone, bucket, udata);
410 }
411 
412 static void
413 bucket_zone_drain(void)
414 {
415 	struct uma_bucket_zone *ubz;
416 
417 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
418 		zone_drain(ubz->ubz_zone);
419 }
420 
421 static void
422 zone_log_warning(uma_zone_t zone)
423 {
424 	static const struct timeval warninterval = { 300, 0 };
425 
426 	if (!zone_warnings || zone->uz_warning == NULL)
427 		return;
428 
429 	if (ratecheck(&zone->uz_ratecheck, &warninterval))
430 		printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning);
431 }
432 
433 static inline void
434 zone_maxaction(uma_zone_t zone)
435 {
436 
437 	if (zone->uz_maxaction.ta_func != NULL)
438 		taskqueue_enqueue(taskqueue_thread, &zone->uz_maxaction);
439 }
440 
441 static void
442 zone_foreach_keg(uma_zone_t zone, void (*kegfn)(uma_keg_t))
443 {
444 	uma_klink_t klink;
445 
446 	LIST_FOREACH(klink, &zone->uz_kegs, kl_link)
447 		kegfn(klink->kl_keg);
448 }
449 
450 /*
451  * Routine called by timeout which is used to fire off some time interval
452  * based calculations.  (stats, hash size, etc.)
453  *
454  * Arguments:
455  *	arg   Unused
456  *
457  * Returns:
458  *	Nothing
459  */
460 static void
461 uma_timeout(void *unused)
462 {
463 	bucket_enable();
464 	zone_foreach(zone_timeout);
465 
466 	/* Reschedule this event */
467 	callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
468 }
469 
470 /*
471  * Routine to perform timeout driven calculations.  This expands the
472  * hashes and does per cpu statistics aggregation.
473  *
474  *  Returns nothing.
475  */
476 static void
477 keg_timeout(uma_keg_t keg)
478 {
479 
480 	KEG_LOCK(keg);
481 	/*
482 	 * Expand the keg hash table.
483 	 *
484 	 * This is done if the number of slabs is larger than the hash size.
485 	 * What I'm trying to do here is completely reduce collisions.  This
486 	 * may be a little aggressive.  Should I allow for two collisions max?
487 	 */
488 	if (keg->uk_flags & UMA_ZONE_HASH &&
489 	    keg->uk_pages / keg->uk_ppera >= keg->uk_hash.uh_hashsize) {
490 		struct uma_hash newhash;
491 		struct uma_hash oldhash;
492 		int ret;
493 
494 		/*
495 		 * This is so involved because allocating and freeing
496 		 * while the keg lock is held will lead to deadlock.
497 		 * I have to do everything in stages and check for
498 		 * races.
499 		 */
500 		newhash = keg->uk_hash;
501 		KEG_UNLOCK(keg);
502 		ret = hash_alloc(&newhash);
503 		KEG_LOCK(keg);
504 		if (ret) {
505 			if (hash_expand(&keg->uk_hash, &newhash)) {
506 				oldhash = keg->uk_hash;
507 				keg->uk_hash = newhash;
508 			} else
509 				oldhash = newhash;
510 
511 			KEG_UNLOCK(keg);
512 			hash_free(&oldhash);
513 			return;
514 		}
515 	}
516 	KEG_UNLOCK(keg);
517 }
518 
519 static void
520 zone_timeout(uma_zone_t zone)
521 {
522 
523 	zone_foreach_keg(zone, &keg_timeout);
524 }
525 
526 /*
527  * Allocate and zero fill the next sized hash table from the appropriate
528  * backing store.
529  *
530  * Arguments:
531  *	hash  A new hash structure with the old hash size in uh_hashsize
532  *
533  * Returns:
534  *	1 on success and 0 on failure.
535  */
536 static int
537 hash_alloc(struct uma_hash *hash)
538 {
539 	int oldsize;
540 	int alloc;
541 
542 	oldsize = hash->uh_hashsize;
543 
544 	/* We're just going to go to a power of two greater */
545 	if (oldsize)  {
546 		hash->uh_hashsize = oldsize * 2;
547 		alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
548 		hash->uh_slab_hash = (struct slabhead *)malloc(alloc,
549 		    M_UMAHASH, M_NOWAIT);
550 	} else {
551 		alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
552 		hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
553 		    M_WAITOK);
554 		hash->uh_hashsize = UMA_HASH_SIZE_INIT;
555 	}
556 	if (hash->uh_slab_hash) {
557 		bzero(hash->uh_slab_hash, alloc);
558 		hash->uh_hashmask = hash->uh_hashsize - 1;
559 		return (1);
560 	}
561 
562 	return (0);
563 }
564 
565 /*
566  * Expands the hash table for HASH zones.  This is done from zone_timeout
567  * to reduce collisions.  This must not be done in the regular allocation
568  * path, otherwise, we can recurse on the vm while allocating pages.
569  *
570  * Arguments:
571  *	oldhash  The hash you want to expand
572  *	newhash  The hash structure for the new table
573  *
574  * Returns:
575  *	Nothing
576  *
577  * Discussion:
578  */
579 static int
580 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
581 {
582 	uma_slab_t slab;
583 	int hval;
584 	int i;
585 
586 	if (!newhash->uh_slab_hash)
587 		return (0);
588 
589 	if (oldhash->uh_hashsize >= newhash->uh_hashsize)
590 		return (0);
591 
592 	/*
593 	 * I need to investigate hash algorithms for resizing without a
594 	 * full rehash.
595 	 */
596 
597 	for (i = 0; i < oldhash->uh_hashsize; i++)
598 		while (!SLIST_EMPTY(&oldhash->uh_slab_hash[i])) {
599 			slab = SLIST_FIRST(&oldhash->uh_slab_hash[i]);
600 			SLIST_REMOVE_HEAD(&oldhash->uh_slab_hash[i], us_hlink);
601 			hval = UMA_HASH(newhash, slab->us_data);
602 			SLIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
603 			    slab, us_hlink);
604 		}
605 
606 	return (1);
607 }
608 
609 /*
610  * Free the hash bucket to the appropriate backing store.
611  *
612  * Arguments:
613  *	slab_hash  The hash bucket we're freeing
614  *	hashsize   The number of entries in that hash bucket
615  *
616  * Returns:
617  *	Nothing
618  */
619 static void
620 hash_free(struct uma_hash *hash)
621 {
622 	if (hash->uh_slab_hash == NULL)
623 		return;
624 	if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
625 		zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE);
626 	else
627 		free(hash->uh_slab_hash, M_UMAHASH);
628 }
629 
630 /*
631  * Frees all outstanding items in a bucket
632  *
633  * Arguments:
634  *	zone   The zone to free to, must be unlocked.
635  *	bucket The free/alloc bucket with items, cpu queue must be locked.
636  *
637  * Returns:
638  *	Nothing
639  */
640 
641 static void
642 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
643 {
644 	int i;
645 
646 	if (bucket == NULL)
647 		return;
648 
649 	if (zone->uz_fini)
650 		for (i = 0; i < bucket->ub_cnt; i++)
651 			zone->uz_fini(bucket->ub_bucket[i], zone->uz_size);
652 	zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt);
653 	bucket->ub_cnt = 0;
654 }
655 
656 /*
657  * Drains the per cpu caches for a zone.
658  *
659  * NOTE: This may only be called while the zone is being turn down, and not
660  * during normal operation.  This is necessary in order that we do not have
661  * to migrate CPUs to drain the per-CPU caches.
662  *
663  * Arguments:
664  *	zone     The zone to drain, must be unlocked.
665  *
666  * Returns:
667  *	Nothing
668  */
669 static void
670 cache_drain(uma_zone_t zone)
671 {
672 	uma_cache_t cache;
673 	int cpu;
674 
675 	/*
676 	 * XXX: It is safe to not lock the per-CPU caches, because we're
677 	 * tearing down the zone anyway.  I.e., there will be no further use
678 	 * of the caches at this point.
679 	 *
680 	 * XXX: It would good to be able to assert that the zone is being
681 	 * torn down to prevent improper use of cache_drain().
682 	 *
683 	 * XXX: We lock the zone before passing into bucket_cache_drain() as
684 	 * it is used elsewhere.  Should the tear-down path be made special
685 	 * there in some form?
686 	 */
687 	CPU_FOREACH(cpu) {
688 		cache = &zone->uz_cpu[cpu];
689 		bucket_drain(zone, cache->uc_allocbucket);
690 		bucket_drain(zone, cache->uc_freebucket);
691 		if (cache->uc_allocbucket != NULL)
692 			bucket_free(zone, cache->uc_allocbucket, NULL);
693 		if (cache->uc_freebucket != NULL)
694 			bucket_free(zone, cache->uc_freebucket, NULL);
695 		cache->uc_allocbucket = cache->uc_freebucket = NULL;
696 	}
697 	ZONE_LOCK(zone);
698 	bucket_cache_drain(zone);
699 	ZONE_UNLOCK(zone);
700 }
701 
702 static void
703 cache_shrink(uma_zone_t zone)
704 {
705 
706 	if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
707 		return;
708 
709 	ZONE_LOCK(zone);
710 	zone->uz_count = (zone->uz_count_min + zone->uz_count) / 2;
711 	ZONE_UNLOCK(zone);
712 }
713 
714 static void
715 cache_drain_safe_cpu(uma_zone_t zone)
716 {
717 	uma_cache_t cache;
718 	uma_bucket_t b1, b2;
719 
720 	if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
721 		return;
722 
723 	b1 = b2 = NULL;
724 	ZONE_LOCK(zone);
725 	critical_enter();
726 	cache = &zone->uz_cpu[curcpu];
727 	if (cache->uc_allocbucket) {
728 		if (cache->uc_allocbucket->ub_cnt != 0)
729 			LIST_INSERT_HEAD(&zone->uz_buckets,
730 			    cache->uc_allocbucket, ub_link);
731 		else
732 			b1 = cache->uc_allocbucket;
733 		cache->uc_allocbucket = NULL;
734 	}
735 	if (cache->uc_freebucket) {
736 		if (cache->uc_freebucket->ub_cnt != 0)
737 			LIST_INSERT_HEAD(&zone->uz_buckets,
738 			    cache->uc_freebucket, ub_link);
739 		else
740 			b2 = cache->uc_freebucket;
741 		cache->uc_freebucket = NULL;
742 	}
743 	critical_exit();
744 	ZONE_UNLOCK(zone);
745 	if (b1)
746 		bucket_free(zone, b1, NULL);
747 	if (b2)
748 		bucket_free(zone, b2, NULL);
749 }
750 
751 /*
752  * Safely drain per-CPU caches of a zone(s) to alloc bucket.
753  * This is an expensive call because it needs to bind to all CPUs
754  * one by one and enter a critical section on each of them in order
755  * to safely access their cache buckets.
756  * Zone lock must not be held on call this function.
757  */
758 static void
759 cache_drain_safe(uma_zone_t zone)
760 {
761 	int cpu;
762 
763 	/*
764 	 * Polite bucket sizes shrinking was not enouth, shrink aggressively.
765 	 */
766 	if (zone)
767 		cache_shrink(zone);
768 	else
769 		zone_foreach(cache_shrink);
770 
771 	CPU_FOREACH(cpu) {
772 		thread_lock(curthread);
773 		sched_bind(curthread, cpu);
774 		thread_unlock(curthread);
775 
776 		if (zone)
777 			cache_drain_safe_cpu(zone);
778 		else
779 			zone_foreach(cache_drain_safe_cpu);
780 	}
781 	thread_lock(curthread);
782 	sched_unbind(curthread);
783 	thread_unlock(curthread);
784 }
785 
786 /*
787  * Drain the cached buckets from a zone.  Expects a locked zone on entry.
788  */
789 static void
790 bucket_cache_drain(uma_zone_t zone)
791 {
792 	uma_bucket_t bucket;
793 
794 	/*
795 	 * Drain the bucket queues and free the buckets, we just keep two per
796 	 * cpu (alloc/free).
797 	 */
798 	while ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) {
799 		LIST_REMOVE(bucket, ub_link);
800 		ZONE_UNLOCK(zone);
801 		bucket_drain(zone, bucket);
802 		bucket_free(zone, bucket, NULL);
803 		ZONE_LOCK(zone);
804 	}
805 
806 	/*
807 	 * Shrink further bucket sizes.  Price of single zone lock collision
808 	 * is probably lower then price of global cache drain.
809 	 */
810 	if (zone->uz_count > zone->uz_count_min)
811 		zone->uz_count--;
812 }
813 
814 static void
815 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start)
816 {
817 	uint8_t *mem;
818 	int i;
819 	uint8_t flags;
820 
821 	mem = slab->us_data;
822 	flags = slab->us_flags;
823 	i = start;
824 	if (keg->uk_fini != NULL) {
825 		for (i--; i > -1; i--)
826 			keg->uk_fini(slab->us_data + (keg->uk_rsize * i),
827 			    keg->uk_size);
828 	}
829 	if (keg->uk_flags & UMA_ZONE_OFFPAGE)
830 		zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE);
831 #ifdef UMA_DEBUG
832 	printf("%s: Returning %d bytes.\n", keg->uk_name,
833 	    PAGE_SIZE * keg->uk_ppera);
834 #endif
835 	keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags);
836 }
837 
838 /*
839  * Frees pages from a keg back to the system.  This is done on demand from
840  * the pageout daemon.
841  *
842  * Returns nothing.
843  */
844 static void
845 keg_drain(uma_keg_t keg)
846 {
847 	struct slabhead freeslabs = { 0 };
848 	uma_slab_t slab, tmp;
849 
850 	/*
851 	 * We don't want to take pages from statically allocated kegs at this
852 	 * time
853 	 */
854 	if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL)
855 		return;
856 
857 #ifdef UMA_DEBUG
858 	printf("%s free items: %u\n", keg->uk_name, keg->uk_free);
859 #endif
860 	KEG_LOCK(keg);
861 	if (keg->uk_free == 0)
862 		goto finished;
863 
864 	LIST_FOREACH_SAFE(slab, &keg->uk_free_slab, us_link, tmp) {
865 		/* We have nowhere to free these to. */
866 		if (slab->us_flags & UMA_SLAB_BOOT)
867 			continue;
868 
869 		LIST_REMOVE(slab, us_link);
870 		keg->uk_pages -= keg->uk_ppera;
871 		keg->uk_free -= keg->uk_ipers;
872 
873 		if (keg->uk_flags & UMA_ZONE_HASH)
874 			UMA_HASH_REMOVE(&keg->uk_hash, slab, slab->us_data);
875 
876 		SLIST_INSERT_HEAD(&freeslabs, slab, us_hlink);
877 	}
878 finished:
879 	KEG_UNLOCK(keg);
880 
881 	while ((slab = SLIST_FIRST(&freeslabs)) != NULL) {
882 		SLIST_REMOVE(&freeslabs, slab, uma_slab, us_hlink);
883 		keg_free_slab(keg, slab, keg->uk_ipers);
884 	}
885 }
886 
887 static void
888 zone_drain_wait(uma_zone_t zone, int waitok)
889 {
890 
891 	/*
892 	 * Set draining to interlock with zone_dtor() so we can release our
893 	 * locks as we go.  Only dtor() should do a WAITOK call since it
894 	 * is the only call that knows the structure will still be available
895 	 * when it wakes up.
896 	 */
897 	ZONE_LOCK(zone);
898 	while (zone->uz_flags & UMA_ZFLAG_DRAINING) {
899 		if (waitok == M_NOWAIT)
900 			goto out;
901 		msleep(zone, zone->uz_lockptr, PVM, "zonedrain", 1);
902 	}
903 	zone->uz_flags |= UMA_ZFLAG_DRAINING;
904 	bucket_cache_drain(zone);
905 	ZONE_UNLOCK(zone);
906 	/*
907 	 * The DRAINING flag protects us from being freed while
908 	 * we're running.  Normally the uma_rwlock would protect us but we
909 	 * must be able to release and acquire the right lock for each keg.
910 	 */
911 	zone_foreach_keg(zone, &keg_drain);
912 	ZONE_LOCK(zone);
913 	zone->uz_flags &= ~UMA_ZFLAG_DRAINING;
914 	wakeup(zone);
915 out:
916 	ZONE_UNLOCK(zone);
917 }
918 
919 void
920 zone_drain(uma_zone_t zone)
921 {
922 
923 	zone_drain_wait(zone, M_NOWAIT);
924 }
925 
926 /*
927  * Allocate a new slab for a keg.  This does not insert the slab onto a list.
928  *
929  * Arguments:
930  *	wait  Shall we wait?
931  *
932  * Returns:
933  *	The slab that was allocated or NULL if there is no memory and the
934  *	caller specified M_NOWAIT.
935  */
936 static uma_slab_t
937 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int wait)
938 {
939 	uma_alloc allocf;
940 	uma_slab_t slab;
941 	uint8_t *mem;
942 	uint8_t flags;
943 	int i;
944 
945 	mtx_assert(&keg->uk_lock, MA_OWNED);
946 	slab = NULL;
947 	mem = NULL;
948 
949 #ifdef UMA_DEBUG
950 	printf("alloc_slab:  Allocating a new slab for %s\n", keg->uk_name);
951 #endif
952 	allocf = keg->uk_allocf;
953 	KEG_UNLOCK(keg);
954 
955 	if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
956 		slab = zone_alloc_item(keg->uk_slabzone, NULL, wait);
957 		if (slab == NULL)
958 			goto out;
959 	}
960 
961 	/*
962 	 * This reproduces the old vm_zone behavior of zero filling pages the
963 	 * first time they are added to a zone.
964 	 *
965 	 * Malloced items are zeroed in uma_zalloc.
966 	 */
967 
968 	if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
969 		wait |= M_ZERO;
970 	else
971 		wait &= ~M_ZERO;
972 
973 	if (keg->uk_flags & UMA_ZONE_NODUMP)
974 		wait |= M_NODUMP;
975 
976 	/* zone is passed for legacy reasons. */
977 	mem = allocf(zone, keg->uk_ppera * PAGE_SIZE, &flags, wait);
978 	if (mem == NULL) {
979 		if (keg->uk_flags & UMA_ZONE_OFFPAGE)
980 			zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE);
981 		slab = NULL;
982 		goto out;
983 	}
984 
985 	/* Point the slab into the allocated memory */
986 	if (!(keg->uk_flags & UMA_ZONE_OFFPAGE))
987 		slab = (uma_slab_t )(mem + keg->uk_pgoff);
988 
989 	if (keg->uk_flags & UMA_ZONE_VTOSLAB)
990 		for (i = 0; i < keg->uk_ppera; i++)
991 			vsetslab((vm_offset_t)mem + (i * PAGE_SIZE), slab);
992 
993 	slab->us_keg = keg;
994 	slab->us_data = mem;
995 	slab->us_freecount = keg->uk_ipers;
996 	slab->us_flags = flags;
997 	BIT_FILL(SLAB_SETSIZE, &slab->us_free);
998 #ifdef INVARIANTS
999 	BIT_ZERO(SLAB_SETSIZE, &slab->us_debugfree);
1000 #endif
1001 
1002 	if (keg->uk_init != NULL) {
1003 		for (i = 0; i < keg->uk_ipers; i++)
1004 			if (keg->uk_init(slab->us_data + (keg->uk_rsize * i),
1005 			    keg->uk_size, wait) != 0)
1006 				break;
1007 		if (i != keg->uk_ipers) {
1008 			keg_free_slab(keg, slab, i);
1009 			slab = NULL;
1010 			goto out;
1011 		}
1012 	}
1013 out:
1014 	KEG_LOCK(keg);
1015 
1016 	if (slab != NULL) {
1017 		if (keg->uk_flags & UMA_ZONE_HASH)
1018 			UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
1019 
1020 		keg->uk_pages += keg->uk_ppera;
1021 		keg->uk_free += keg->uk_ipers;
1022 	}
1023 
1024 	return (slab);
1025 }
1026 
1027 /*
1028  * This function is intended to be used early on in place of page_alloc() so
1029  * that we may use the boot time page cache to satisfy allocations before
1030  * the VM is ready.
1031  */
1032 static void *
1033 startup_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *pflag, int wait)
1034 {
1035 	uma_keg_t keg;
1036 	uma_slab_t tmps;
1037 	int pages, check_pages;
1038 
1039 	keg = zone_first_keg(zone);
1040 	pages = howmany(bytes, PAGE_SIZE);
1041 	check_pages = pages - 1;
1042 	KASSERT(pages > 0, ("startup_alloc can't reserve 0 pages\n"));
1043 
1044 	/*
1045 	 * Check our small startup cache to see if it has pages remaining.
1046 	 */
1047 	mtx_lock(&uma_boot_pages_mtx);
1048 
1049 	/* First check if we have enough room. */
1050 	tmps = LIST_FIRST(&uma_boot_pages);
1051 	while (tmps != NULL && check_pages-- > 0)
1052 		tmps = LIST_NEXT(tmps, us_link);
1053 	if (tmps != NULL) {
1054 		/*
1055 		 * It's ok to lose tmps references.  The last one will
1056 		 * have tmps->us_data pointing to the start address of
1057 		 * "pages" contiguous pages of memory.
1058 		 */
1059 		while (pages-- > 0) {
1060 			tmps = LIST_FIRST(&uma_boot_pages);
1061 			LIST_REMOVE(tmps, us_link);
1062 		}
1063 		mtx_unlock(&uma_boot_pages_mtx);
1064 		*pflag = tmps->us_flags;
1065 		return (tmps->us_data);
1066 	}
1067 	mtx_unlock(&uma_boot_pages_mtx);
1068 	if (booted < UMA_STARTUP2)
1069 		panic("UMA: Increase vm.boot_pages");
1070 	/*
1071 	 * Now that we've booted reset these users to their real allocator.
1072 	 */
1073 #ifdef UMA_MD_SMALL_ALLOC
1074 	keg->uk_allocf = (keg->uk_ppera > 1) ? page_alloc : uma_small_alloc;
1075 #else
1076 	keg->uk_allocf = page_alloc;
1077 #endif
1078 	return keg->uk_allocf(zone, bytes, pflag, wait);
1079 }
1080 
1081 /*
1082  * Allocates a number of pages from the system
1083  *
1084  * Arguments:
1085  *	bytes  The number of bytes requested
1086  *	wait  Shall we wait?
1087  *
1088  * Returns:
1089  *	A pointer to the alloced memory or possibly
1090  *	NULL if M_NOWAIT is set.
1091  */
1092 static void *
1093 page_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *pflag, int wait)
1094 {
1095 	void *p;	/* Returned page */
1096 
1097 	*pflag = UMA_SLAB_KMEM;
1098 	p = (void *) kmem_malloc(kmem_arena, bytes, wait);
1099 
1100 	return (p);
1101 }
1102 
1103 /*
1104  * Allocates a number of pages from within an object
1105  *
1106  * Arguments:
1107  *	bytes  The number of bytes requested
1108  *	wait   Shall we wait?
1109  *
1110  * Returns:
1111  *	A pointer to the alloced memory or possibly
1112  *	NULL if M_NOWAIT is set.
1113  */
1114 static void *
1115 noobj_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *flags, int wait)
1116 {
1117 	TAILQ_HEAD(, vm_page) alloctail;
1118 	u_long npages;
1119 	vm_offset_t retkva, zkva;
1120 	vm_page_t p, p_next;
1121 	uma_keg_t keg;
1122 
1123 	TAILQ_INIT(&alloctail);
1124 	keg = zone_first_keg(zone);
1125 
1126 	npages = howmany(bytes, PAGE_SIZE);
1127 	while (npages > 0) {
1128 		p = vm_page_alloc(NULL, 0, VM_ALLOC_INTERRUPT |
1129 		    VM_ALLOC_WIRED | VM_ALLOC_NOOBJ);
1130 		if (p != NULL) {
1131 			/*
1132 			 * Since the page does not belong to an object, its
1133 			 * listq is unused.
1134 			 */
1135 			TAILQ_INSERT_TAIL(&alloctail, p, listq);
1136 			npages--;
1137 			continue;
1138 		}
1139 		if (wait & M_WAITOK) {
1140 			VM_WAIT;
1141 			continue;
1142 		}
1143 
1144 		/*
1145 		 * Page allocation failed, free intermediate pages and
1146 		 * exit.
1147 		 */
1148 		TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1149 			vm_page_unwire(p, PQ_NONE);
1150 			vm_page_free(p);
1151 		}
1152 		return (NULL);
1153 	}
1154 	*flags = UMA_SLAB_PRIV;
1155 	zkva = keg->uk_kva +
1156 	    atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
1157 	retkva = zkva;
1158 	TAILQ_FOREACH(p, &alloctail, listq) {
1159 		pmap_qenter(zkva, &p, 1);
1160 		zkva += PAGE_SIZE;
1161 	}
1162 
1163 	return ((void *)retkva);
1164 }
1165 
1166 /*
1167  * Frees a number of pages to the system
1168  *
1169  * Arguments:
1170  *	mem   A pointer to the memory to be freed
1171  *	size  The size of the memory being freed
1172  *	flags The original p->us_flags field
1173  *
1174  * Returns:
1175  *	Nothing
1176  */
1177 static void
1178 page_free(void *mem, vm_size_t size, uint8_t flags)
1179 {
1180 	struct vmem *vmem;
1181 
1182 	if (flags & UMA_SLAB_KMEM)
1183 		vmem = kmem_arena;
1184 	else if (flags & UMA_SLAB_KERNEL)
1185 		vmem = kernel_arena;
1186 	else
1187 		panic("UMA: page_free used with invalid flags %d", flags);
1188 
1189 	kmem_free(vmem, (vm_offset_t)mem, size);
1190 }
1191 
1192 /*
1193  * Zero fill initializer
1194  *
1195  * Arguments/Returns follow uma_init specifications
1196  */
1197 static int
1198 zero_init(void *mem, int size, int flags)
1199 {
1200 	bzero(mem, size);
1201 	return (0);
1202 }
1203 
1204 /*
1205  * Finish creating a small uma keg.  This calculates ipers, and the keg size.
1206  *
1207  * Arguments
1208  *	keg  The zone we should initialize
1209  *
1210  * Returns
1211  *	Nothing
1212  */
1213 static void
1214 keg_small_init(uma_keg_t keg)
1215 {
1216 	u_int rsize;
1217 	u_int memused;
1218 	u_int wastedspace;
1219 	u_int shsize;
1220 
1221 	if (keg->uk_flags & UMA_ZONE_PCPU) {
1222 		u_int ncpus = (mp_maxid + 1) ? (mp_maxid + 1) : MAXCPU;
1223 
1224 		keg->uk_slabsize = sizeof(struct pcpu);
1225 		keg->uk_ppera = howmany(ncpus * sizeof(struct pcpu),
1226 		    PAGE_SIZE);
1227 	} else {
1228 		keg->uk_slabsize = UMA_SLAB_SIZE;
1229 		keg->uk_ppera = 1;
1230 	}
1231 
1232 	/*
1233 	 * Calculate the size of each allocation (rsize) according to
1234 	 * alignment.  If the requested size is smaller than we have
1235 	 * allocation bits for we round it up.
1236 	 */
1237 	rsize = keg->uk_size;
1238 	if (rsize < keg->uk_slabsize / SLAB_SETSIZE)
1239 		rsize = keg->uk_slabsize / SLAB_SETSIZE;
1240 	if (rsize & keg->uk_align)
1241 		rsize = (rsize & ~keg->uk_align) + (keg->uk_align + 1);
1242 	keg->uk_rsize = rsize;
1243 
1244 	KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
1245 	    keg->uk_rsize < sizeof(struct pcpu),
1246 	    ("%s: size %u too large", __func__, keg->uk_rsize));
1247 
1248 	if (keg->uk_flags & UMA_ZONE_OFFPAGE)
1249 		shsize = 0;
1250 	else
1251 		shsize = sizeof(struct uma_slab);
1252 
1253 	keg->uk_ipers = (keg->uk_slabsize - shsize) / rsize;
1254 	KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE,
1255 	    ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers));
1256 
1257 	memused = keg->uk_ipers * rsize + shsize;
1258 	wastedspace = keg->uk_slabsize - memused;
1259 
1260 	/*
1261 	 * We can't do OFFPAGE if we're internal or if we've been
1262 	 * asked to not go to the VM for buckets.  If we do this we
1263 	 * may end up going to the VM  for slabs which we do not
1264 	 * want to do if we're UMA_ZFLAG_CACHEONLY as a result
1265 	 * of UMA_ZONE_VM, which clearly forbids it.
1266 	 */
1267 	if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) ||
1268 	    (keg->uk_flags & UMA_ZFLAG_CACHEONLY))
1269 		return;
1270 
1271 	/*
1272 	 * See if using an OFFPAGE slab will limit our waste.  Only do
1273 	 * this if it permits more items per-slab.
1274 	 *
1275 	 * XXX We could try growing slabsize to limit max waste as well.
1276 	 * Historically this was not done because the VM could not
1277 	 * efficiently handle contiguous allocations.
1278 	 */
1279 	if ((wastedspace >= keg->uk_slabsize / UMA_MAX_WASTE) &&
1280 	    (keg->uk_ipers < (keg->uk_slabsize / keg->uk_rsize))) {
1281 		keg->uk_ipers = keg->uk_slabsize / keg->uk_rsize;
1282 		KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE,
1283 		    ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers));
1284 #ifdef UMA_DEBUG
1285 		printf("UMA decided we need offpage slab headers for "
1286 		    "keg: %s, calculated wastedspace = %d, "
1287 		    "maximum wasted space allowed = %d, "
1288 		    "calculated ipers = %d, "
1289 		    "new wasted space = %d\n", keg->uk_name, wastedspace,
1290 		    keg->uk_slabsize / UMA_MAX_WASTE, keg->uk_ipers,
1291 		    keg->uk_slabsize - keg->uk_ipers * keg->uk_rsize);
1292 #endif
1293 		keg->uk_flags |= UMA_ZONE_OFFPAGE;
1294 	}
1295 
1296 	if ((keg->uk_flags & UMA_ZONE_OFFPAGE) &&
1297 	    (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1298 		keg->uk_flags |= UMA_ZONE_HASH;
1299 }
1300 
1301 /*
1302  * Finish creating a large (> UMA_SLAB_SIZE) uma kegs.  Just give in and do
1303  * OFFPAGE for now.  When I can allow for more dynamic slab sizes this will be
1304  * more complicated.
1305  *
1306  * Arguments
1307  *	keg  The keg we should initialize
1308  *
1309  * Returns
1310  *	Nothing
1311  */
1312 static void
1313 keg_large_init(uma_keg_t keg)
1314 {
1315 	u_int shsize;
1316 
1317 	KASSERT(keg != NULL, ("Keg is null in keg_large_init"));
1318 	KASSERT((keg->uk_flags & UMA_ZFLAG_CACHEONLY) == 0,
1319 	    ("keg_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY keg"));
1320 	KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0,
1321 	    ("%s: Cannot large-init a UMA_ZONE_PCPU keg", __func__));
1322 
1323 	keg->uk_ppera = howmany(keg->uk_size, PAGE_SIZE);
1324 	keg->uk_slabsize = keg->uk_ppera * PAGE_SIZE;
1325 	keg->uk_ipers = 1;
1326 	keg->uk_rsize = keg->uk_size;
1327 
1328 	/* We can't do OFFPAGE if we're internal, bail out here. */
1329 	if (keg->uk_flags & UMA_ZFLAG_INTERNAL)
1330 		return;
1331 
1332 	/* Check whether we have enough space to not do OFFPAGE. */
1333 	if ((keg->uk_flags & UMA_ZONE_OFFPAGE) == 0) {
1334 		shsize = sizeof(struct uma_slab);
1335 		if (shsize & UMA_ALIGN_PTR)
1336 			shsize = (shsize & ~UMA_ALIGN_PTR) +
1337 			    (UMA_ALIGN_PTR + 1);
1338 
1339 		if ((PAGE_SIZE * keg->uk_ppera) - keg->uk_rsize < shsize)
1340 			keg->uk_flags |= UMA_ZONE_OFFPAGE;
1341 	}
1342 
1343 	if ((keg->uk_flags & UMA_ZONE_OFFPAGE) &&
1344 	    (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1345 		keg->uk_flags |= UMA_ZONE_HASH;
1346 }
1347 
1348 static void
1349 keg_cachespread_init(uma_keg_t keg)
1350 {
1351 	int alignsize;
1352 	int trailer;
1353 	int pages;
1354 	int rsize;
1355 
1356 	KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0,
1357 	    ("%s: Cannot cachespread-init a UMA_ZONE_PCPU keg", __func__));
1358 
1359 	alignsize = keg->uk_align + 1;
1360 	rsize = keg->uk_size;
1361 	/*
1362 	 * We want one item to start on every align boundary in a page.  To
1363 	 * do this we will span pages.  We will also extend the item by the
1364 	 * size of align if it is an even multiple of align.  Otherwise, it
1365 	 * would fall on the same boundary every time.
1366 	 */
1367 	if (rsize & keg->uk_align)
1368 		rsize = (rsize & ~keg->uk_align) + alignsize;
1369 	if ((rsize & alignsize) == 0)
1370 		rsize += alignsize;
1371 	trailer = rsize - keg->uk_size;
1372 	pages = (rsize * (PAGE_SIZE / alignsize)) / PAGE_SIZE;
1373 	pages = MIN(pages, (128 * 1024) / PAGE_SIZE);
1374 	keg->uk_rsize = rsize;
1375 	keg->uk_ppera = pages;
1376 	keg->uk_slabsize = UMA_SLAB_SIZE;
1377 	keg->uk_ipers = ((pages * PAGE_SIZE) + trailer) / rsize;
1378 	keg->uk_flags |= UMA_ZONE_OFFPAGE | UMA_ZONE_VTOSLAB;
1379 	KASSERT(keg->uk_ipers <= SLAB_SETSIZE,
1380 	    ("%s: keg->uk_ipers too high(%d) increase max_ipers", __func__,
1381 	    keg->uk_ipers));
1382 }
1383 
1384 /*
1385  * Keg header ctor.  This initializes all fields, locks, etc.  And inserts
1386  * the keg onto the global keg list.
1387  *
1388  * Arguments/Returns follow uma_ctor specifications
1389  *	udata  Actually uma_kctor_args
1390  */
1391 static int
1392 keg_ctor(void *mem, int size, void *udata, int flags)
1393 {
1394 	struct uma_kctor_args *arg = udata;
1395 	uma_keg_t keg = mem;
1396 	uma_zone_t zone;
1397 
1398 	bzero(keg, size);
1399 	keg->uk_size = arg->size;
1400 	keg->uk_init = arg->uminit;
1401 	keg->uk_fini = arg->fini;
1402 	keg->uk_align = arg->align;
1403 	keg->uk_free = 0;
1404 	keg->uk_reserve = 0;
1405 	keg->uk_pages = 0;
1406 	keg->uk_flags = arg->flags;
1407 	keg->uk_allocf = page_alloc;
1408 	keg->uk_freef = page_free;
1409 	keg->uk_slabzone = NULL;
1410 
1411 	/*
1412 	 * The master zone is passed to us at keg-creation time.
1413 	 */
1414 	zone = arg->zone;
1415 	keg->uk_name = zone->uz_name;
1416 
1417 	if (arg->flags & UMA_ZONE_VM)
1418 		keg->uk_flags |= UMA_ZFLAG_CACHEONLY;
1419 
1420 	if (arg->flags & UMA_ZONE_ZINIT)
1421 		keg->uk_init = zero_init;
1422 
1423 	if (arg->flags & UMA_ZONE_MALLOC)
1424 		keg->uk_flags |= UMA_ZONE_VTOSLAB;
1425 
1426 	if (arg->flags & UMA_ZONE_PCPU)
1427 #ifdef SMP
1428 		keg->uk_flags |= UMA_ZONE_OFFPAGE;
1429 #else
1430 		keg->uk_flags &= ~UMA_ZONE_PCPU;
1431 #endif
1432 
1433 	if (keg->uk_flags & UMA_ZONE_CACHESPREAD) {
1434 		keg_cachespread_init(keg);
1435 	} else {
1436 		if (keg->uk_size > (UMA_SLAB_SIZE - sizeof(struct uma_slab)))
1437 			keg_large_init(keg);
1438 		else
1439 			keg_small_init(keg);
1440 	}
1441 
1442 	if (keg->uk_flags & UMA_ZONE_OFFPAGE)
1443 		keg->uk_slabzone = slabzone;
1444 
1445 	/*
1446 	 * If we haven't booted yet we need allocations to go through the
1447 	 * startup cache until the vm is ready.
1448 	 */
1449 	if (keg->uk_ppera == 1) {
1450 #ifdef UMA_MD_SMALL_ALLOC
1451 		keg->uk_allocf = uma_small_alloc;
1452 		keg->uk_freef = uma_small_free;
1453 
1454 		if (booted < UMA_STARTUP)
1455 			keg->uk_allocf = startup_alloc;
1456 #else
1457 		if (booted < UMA_STARTUP2)
1458 			keg->uk_allocf = startup_alloc;
1459 #endif
1460 	} else if (booted < UMA_STARTUP2 &&
1461 	    (keg->uk_flags & UMA_ZFLAG_INTERNAL))
1462 		keg->uk_allocf = startup_alloc;
1463 
1464 	/*
1465 	 * Initialize keg's lock
1466 	 */
1467 	KEG_LOCK_INIT(keg, (arg->flags & UMA_ZONE_MTXCLASS));
1468 
1469 	/*
1470 	 * If we're putting the slab header in the actual page we need to
1471 	 * figure out where in each page it goes.  This calculates a right
1472 	 * justified offset into the memory on an ALIGN_PTR boundary.
1473 	 */
1474 	if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) {
1475 		u_int totsize;
1476 
1477 		/* Size of the slab struct and free list */
1478 		totsize = sizeof(struct uma_slab);
1479 
1480 		if (totsize & UMA_ALIGN_PTR)
1481 			totsize = (totsize & ~UMA_ALIGN_PTR) +
1482 			    (UMA_ALIGN_PTR + 1);
1483 		keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - totsize;
1484 
1485 		/*
1486 		 * The only way the following is possible is if with our
1487 		 * UMA_ALIGN_PTR adjustments we are now bigger than
1488 		 * UMA_SLAB_SIZE.  I haven't checked whether this is
1489 		 * mathematically possible for all cases, so we make
1490 		 * sure here anyway.
1491 		 */
1492 		totsize = keg->uk_pgoff + sizeof(struct uma_slab);
1493 		if (totsize > PAGE_SIZE * keg->uk_ppera) {
1494 			printf("zone %s ipers %d rsize %d size %d\n",
1495 			    zone->uz_name, keg->uk_ipers, keg->uk_rsize,
1496 			    keg->uk_size);
1497 			panic("UMA slab won't fit.");
1498 		}
1499 	}
1500 
1501 	if (keg->uk_flags & UMA_ZONE_HASH)
1502 		hash_alloc(&keg->uk_hash);
1503 
1504 #ifdef UMA_DEBUG
1505 	printf("UMA: %s(%p) size %d(%d) flags %#x ipers %d ppera %d out %d free %d\n",
1506 	    zone->uz_name, zone, keg->uk_size, keg->uk_rsize, keg->uk_flags,
1507 	    keg->uk_ipers, keg->uk_ppera,
1508 	    (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free);
1509 #endif
1510 
1511 	LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
1512 
1513 	rw_wlock(&uma_rwlock);
1514 	LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
1515 	rw_wunlock(&uma_rwlock);
1516 	return (0);
1517 }
1518 
1519 /*
1520  * Zone header ctor.  This initializes all fields, locks, etc.
1521  *
1522  * Arguments/Returns follow uma_ctor specifications
1523  *	udata  Actually uma_zctor_args
1524  */
1525 static int
1526 zone_ctor(void *mem, int size, void *udata, int flags)
1527 {
1528 	struct uma_zctor_args *arg = udata;
1529 	uma_zone_t zone = mem;
1530 	uma_zone_t z;
1531 	uma_keg_t keg;
1532 
1533 	bzero(zone, size);
1534 	zone->uz_name = arg->name;
1535 	zone->uz_ctor = arg->ctor;
1536 	zone->uz_dtor = arg->dtor;
1537 	zone->uz_slab = zone_fetch_slab;
1538 	zone->uz_init = NULL;
1539 	zone->uz_fini = NULL;
1540 	zone->uz_allocs = 0;
1541 	zone->uz_frees = 0;
1542 	zone->uz_fails = 0;
1543 	zone->uz_sleeps = 0;
1544 	zone->uz_count = 0;
1545 	zone->uz_count_min = 0;
1546 	zone->uz_flags = 0;
1547 	zone->uz_warning = NULL;
1548 	timevalclear(&zone->uz_ratecheck);
1549 	keg = arg->keg;
1550 
1551 	ZONE_LOCK_INIT(zone, (arg->flags & UMA_ZONE_MTXCLASS));
1552 
1553 	/*
1554 	 * This is a pure cache zone, no kegs.
1555 	 */
1556 	if (arg->import) {
1557 		if (arg->flags & UMA_ZONE_VM)
1558 			arg->flags |= UMA_ZFLAG_CACHEONLY;
1559 		zone->uz_flags = arg->flags;
1560 		zone->uz_size = arg->size;
1561 		zone->uz_import = arg->import;
1562 		zone->uz_release = arg->release;
1563 		zone->uz_arg = arg->arg;
1564 		zone->uz_lockptr = &zone->uz_lock;
1565 		rw_wlock(&uma_rwlock);
1566 		LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link);
1567 		rw_wunlock(&uma_rwlock);
1568 		goto out;
1569 	}
1570 
1571 	/*
1572 	 * Use the regular zone/keg/slab allocator.
1573 	 */
1574 	zone->uz_import = (uma_import)zone_import;
1575 	zone->uz_release = (uma_release)zone_release;
1576 	zone->uz_arg = zone;
1577 
1578 	if (arg->flags & UMA_ZONE_SECONDARY) {
1579 		KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
1580 		zone->uz_init = arg->uminit;
1581 		zone->uz_fini = arg->fini;
1582 		zone->uz_lockptr = &keg->uk_lock;
1583 		zone->uz_flags |= UMA_ZONE_SECONDARY;
1584 		rw_wlock(&uma_rwlock);
1585 		ZONE_LOCK(zone);
1586 		LIST_FOREACH(z, &keg->uk_zones, uz_link) {
1587 			if (LIST_NEXT(z, uz_link) == NULL) {
1588 				LIST_INSERT_AFTER(z, zone, uz_link);
1589 				break;
1590 			}
1591 		}
1592 		ZONE_UNLOCK(zone);
1593 		rw_wunlock(&uma_rwlock);
1594 	} else if (keg == NULL) {
1595 		if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
1596 		    arg->align, arg->flags)) == NULL)
1597 			return (ENOMEM);
1598 	} else {
1599 		struct uma_kctor_args karg;
1600 		int error;
1601 
1602 		/* We should only be here from uma_startup() */
1603 		karg.size = arg->size;
1604 		karg.uminit = arg->uminit;
1605 		karg.fini = arg->fini;
1606 		karg.align = arg->align;
1607 		karg.flags = arg->flags;
1608 		karg.zone = zone;
1609 		error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
1610 		    flags);
1611 		if (error)
1612 			return (error);
1613 	}
1614 
1615 	/*
1616 	 * Link in the first keg.
1617 	 */
1618 	zone->uz_klink.kl_keg = keg;
1619 	LIST_INSERT_HEAD(&zone->uz_kegs, &zone->uz_klink, kl_link);
1620 	zone->uz_lockptr = &keg->uk_lock;
1621 	zone->uz_size = keg->uk_size;
1622 	zone->uz_flags |= (keg->uk_flags &
1623 	    (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
1624 
1625 	/*
1626 	 * Some internal zones don't have room allocated for the per cpu
1627 	 * caches.  If we're internal, bail out here.
1628 	 */
1629 	if (keg->uk_flags & UMA_ZFLAG_INTERNAL) {
1630 		KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
1631 		    ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
1632 		return (0);
1633 	}
1634 
1635 out:
1636 	if ((arg->flags & UMA_ZONE_MAXBUCKET) == 0)
1637 		zone->uz_count = bucket_select(zone->uz_size);
1638 	else
1639 		zone->uz_count = BUCKET_MAX;
1640 	zone->uz_count_min = zone->uz_count;
1641 
1642 	return (0);
1643 }
1644 
1645 /*
1646  * Keg header dtor.  This frees all data, destroys locks, frees the hash
1647  * table and removes the keg from the global list.
1648  *
1649  * Arguments/Returns follow uma_dtor specifications
1650  *	udata  unused
1651  */
1652 static void
1653 keg_dtor(void *arg, int size, void *udata)
1654 {
1655 	uma_keg_t keg;
1656 
1657 	keg = (uma_keg_t)arg;
1658 	KEG_LOCK(keg);
1659 	if (keg->uk_free != 0) {
1660 		printf("Freed UMA keg (%s) was not empty (%d items). "
1661 		    " Lost %d pages of memory.\n",
1662 		    keg->uk_name ? keg->uk_name : "",
1663 		    keg->uk_free, keg->uk_pages);
1664 	}
1665 	KEG_UNLOCK(keg);
1666 
1667 	hash_free(&keg->uk_hash);
1668 
1669 	KEG_LOCK_FINI(keg);
1670 }
1671 
1672 /*
1673  * Zone header dtor.
1674  *
1675  * Arguments/Returns follow uma_dtor specifications
1676  *	udata  unused
1677  */
1678 static void
1679 zone_dtor(void *arg, int size, void *udata)
1680 {
1681 	uma_klink_t klink;
1682 	uma_zone_t zone;
1683 	uma_keg_t keg;
1684 
1685 	zone = (uma_zone_t)arg;
1686 	keg = zone_first_keg(zone);
1687 
1688 	if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
1689 		cache_drain(zone);
1690 
1691 	rw_wlock(&uma_rwlock);
1692 	LIST_REMOVE(zone, uz_link);
1693 	rw_wunlock(&uma_rwlock);
1694 	/*
1695 	 * XXX there are some races here where
1696 	 * the zone can be drained but zone lock
1697 	 * released and then refilled before we
1698 	 * remove it... we dont care for now
1699 	 */
1700 	zone_drain_wait(zone, M_WAITOK);
1701 	/*
1702 	 * Unlink all of our kegs.
1703 	 */
1704 	while ((klink = LIST_FIRST(&zone->uz_kegs)) != NULL) {
1705 		klink->kl_keg = NULL;
1706 		LIST_REMOVE(klink, kl_link);
1707 		if (klink == &zone->uz_klink)
1708 			continue;
1709 		free(klink, M_TEMP);
1710 	}
1711 	/*
1712 	 * We only destroy kegs from non secondary zones.
1713 	 */
1714 	if (keg != NULL && (zone->uz_flags & UMA_ZONE_SECONDARY) == 0)  {
1715 		rw_wlock(&uma_rwlock);
1716 		LIST_REMOVE(keg, uk_link);
1717 		rw_wunlock(&uma_rwlock);
1718 		zone_free_item(kegs, keg, NULL, SKIP_NONE);
1719 	}
1720 	ZONE_LOCK_FINI(zone);
1721 }
1722 
1723 /*
1724  * Traverses every zone in the system and calls a callback
1725  *
1726  * Arguments:
1727  *	zfunc  A pointer to a function which accepts a zone
1728  *		as an argument.
1729  *
1730  * Returns:
1731  *	Nothing
1732  */
1733 static void
1734 zone_foreach(void (*zfunc)(uma_zone_t))
1735 {
1736 	uma_keg_t keg;
1737 	uma_zone_t zone;
1738 
1739 	rw_rlock(&uma_rwlock);
1740 	LIST_FOREACH(keg, &uma_kegs, uk_link) {
1741 		LIST_FOREACH(zone, &keg->uk_zones, uz_link)
1742 			zfunc(zone);
1743 	}
1744 	rw_runlock(&uma_rwlock);
1745 }
1746 
1747 /* Public functions */
1748 /* See uma.h */
1749 void
1750 uma_startup(void *bootmem, int boot_pages)
1751 {
1752 	struct uma_zctor_args args;
1753 	uma_slab_t slab;
1754 	int i;
1755 
1756 #ifdef UMA_DEBUG
1757 	printf("Creating uma keg headers zone and keg.\n");
1758 #endif
1759 	rw_init(&uma_rwlock, "UMA lock");
1760 
1761 	/* "manually" create the initial zone */
1762 	memset(&args, 0, sizeof(args));
1763 	args.name = "UMA Kegs";
1764 	args.size = sizeof(struct uma_keg);
1765 	args.ctor = keg_ctor;
1766 	args.dtor = keg_dtor;
1767 	args.uminit = zero_init;
1768 	args.fini = NULL;
1769 	args.keg = &masterkeg;
1770 	args.align = 32 - 1;
1771 	args.flags = UMA_ZFLAG_INTERNAL;
1772 	/* The initial zone has no Per cpu queues so it's smaller */
1773 	zone_ctor(kegs, sizeof(struct uma_zone), &args, M_WAITOK);
1774 
1775 #ifdef UMA_DEBUG
1776 	printf("Filling boot free list.\n");
1777 #endif
1778 	for (i = 0; i < boot_pages; i++) {
1779 		slab = (uma_slab_t)((uint8_t *)bootmem + (i * UMA_SLAB_SIZE));
1780 		slab->us_data = (uint8_t *)slab;
1781 		slab->us_flags = UMA_SLAB_BOOT;
1782 		LIST_INSERT_HEAD(&uma_boot_pages, slab, us_link);
1783 	}
1784 	mtx_init(&uma_boot_pages_mtx, "UMA boot pages", NULL, MTX_DEF);
1785 
1786 #ifdef UMA_DEBUG
1787 	printf("Creating uma zone headers zone and keg.\n");
1788 #endif
1789 	args.name = "UMA Zones";
1790 	args.size = sizeof(struct uma_zone) +
1791 	    (sizeof(struct uma_cache) * (mp_maxid + 1));
1792 	args.ctor = zone_ctor;
1793 	args.dtor = zone_dtor;
1794 	args.uminit = zero_init;
1795 	args.fini = NULL;
1796 	args.keg = NULL;
1797 	args.align = 32 - 1;
1798 	args.flags = UMA_ZFLAG_INTERNAL;
1799 	/* The initial zone has no Per cpu queues so it's smaller */
1800 	zone_ctor(zones, sizeof(struct uma_zone), &args, M_WAITOK);
1801 
1802 #ifdef UMA_DEBUG
1803 	printf("Creating slab and hash zones.\n");
1804 #endif
1805 
1806 	/* Now make a zone for slab headers */
1807 	slabzone = uma_zcreate("UMA Slabs",
1808 				sizeof(struct uma_slab),
1809 				NULL, NULL, NULL, NULL,
1810 				UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1811 
1812 	hashzone = uma_zcreate("UMA Hash",
1813 	    sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
1814 	    NULL, NULL, NULL, NULL,
1815 	    UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1816 
1817 	bucket_init();
1818 
1819 	booted = UMA_STARTUP;
1820 
1821 #ifdef UMA_DEBUG
1822 	printf("UMA startup complete.\n");
1823 #endif
1824 }
1825 
1826 /* see uma.h */
1827 void
1828 uma_startup2(void)
1829 {
1830 	booted = UMA_STARTUP2;
1831 	bucket_enable();
1832 	sx_init(&uma_drain_lock, "umadrain");
1833 #ifdef UMA_DEBUG
1834 	printf("UMA startup2 complete.\n");
1835 #endif
1836 }
1837 
1838 /*
1839  * Initialize our callout handle
1840  *
1841  */
1842 
1843 static void
1844 uma_startup3(void)
1845 {
1846 #ifdef UMA_DEBUG
1847 	printf("Starting callout.\n");
1848 #endif
1849 	callout_init(&uma_callout, 1);
1850 	callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
1851 #ifdef UMA_DEBUG
1852 	printf("UMA startup3 complete.\n");
1853 #endif
1854 }
1855 
1856 static uma_keg_t
1857 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
1858 		int align, uint32_t flags)
1859 {
1860 	struct uma_kctor_args args;
1861 
1862 	args.size = size;
1863 	args.uminit = uminit;
1864 	args.fini = fini;
1865 	args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
1866 	args.flags = flags;
1867 	args.zone = zone;
1868 	return (zone_alloc_item(kegs, &args, M_WAITOK));
1869 }
1870 
1871 /* See uma.h */
1872 void
1873 uma_set_align(int align)
1874 {
1875 
1876 	if (align != UMA_ALIGN_CACHE)
1877 		uma_align_cache = align;
1878 }
1879 
1880 /* See uma.h */
1881 uma_zone_t
1882 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
1883 		uma_init uminit, uma_fini fini, int align, uint32_t flags)
1884 
1885 {
1886 	struct uma_zctor_args args;
1887 	uma_zone_t res;
1888 	bool locked;
1889 
1890 	/* This stuff is essential for the zone ctor */
1891 	memset(&args, 0, sizeof(args));
1892 	args.name = name;
1893 	args.size = size;
1894 	args.ctor = ctor;
1895 	args.dtor = dtor;
1896 	args.uminit = uminit;
1897 	args.fini = fini;
1898 #ifdef  INVARIANTS
1899 	/*
1900 	 * If a zone is being created with an empty constructor and
1901 	 * destructor, pass UMA constructor/destructor which checks for
1902 	 * memory use after free.
1903 	 */
1904 	if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOFREE))) &&
1905 	    ctor == NULL && dtor == NULL && uminit == NULL && fini == NULL) {
1906 		args.ctor = trash_ctor;
1907 		args.dtor = trash_dtor;
1908 		args.uminit = trash_init;
1909 		args.fini = trash_fini;
1910 	}
1911 #endif
1912 	args.align = align;
1913 	args.flags = flags;
1914 	args.keg = NULL;
1915 
1916 	if (booted < UMA_STARTUP2) {
1917 		locked = false;
1918 	} else {
1919 		sx_slock(&uma_drain_lock);
1920 		locked = true;
1921 	}
1922 	res = zone_alloc_item(zones, &args, M_WAITOK);
1923 	if (locked)
1924 		sx_sunlock(&uma_drain_lock);
1925 	return (res);
1926 }
1927 
1928 /* See uma.h */
1929 uma_zone_t
1930 uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor,
1931 		    uma_init zinit, uma_fini zfini, uma_zone_t master)
1932 {
1933 	struct uma_zctor_args args;
1934 	uma_keg_t keg;
1935 	uma_zone_t res;
1936 	bool locked;
1937 
1938 	keg = zone_first_keg(master);
1939 	memset(&args, 0, sizeof(args));
1940 	args.name = name;
1941 	args.size = keg->uk_size;
1942 	args.ctor = ctor;
1943 	args.dtor = dtor;
1944 	args.uminit = zinit;
1945 	args.fini = zfini;
1946 	args.align = keg->uk_align;
1947 	args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
1948 	args.keg = keg;
1949 
1950 	if (booted < UMA_STARTUP2) {
1951 		locked = false;
1952 	} else {
1953 		sx_slock(&uma_drain_lock);
1954 		locked = true;
1955 	}
1956 	/* XXX Attaches only one keg of potentially many. */
1957 	res = zone_alloc_item(zones, &args, M_WAITOK);
1958 	if (locked)
1959 		sx_sunlock(&uma_drain_lock);
1960 	return (res);
1961 }
1962 
1963 /* See uma.h */
1964 uma_zone_t
1965 uma_zcache_create(char *name, int size, uma_ctor ctor, uma_dtor dtor,
1966 		    uma_init zinit, uma_fini zfini, uma_import zimport,
1967 		    uma_release zrelease, void *arg, int flags)
1968 {
1969 	struct uma_zctor_args args;
1970 
1971 	memset(&args, 0, sizeof(args));
1972 	args.name = name;
1973 	args.size = size;
1974 	args.ctor = ctor;
1975 	args.dtor = dtor;
1976 	args.uminit = zinit;
1977 	args.fini = zfini;
1978 	args.import = zimport;
1979 	args.release = zrelease;
1980 	args.arg = arg;
1981 	args.align = 0;
1982 	args.flags = flags;
1983 
1984 	return (zone_alloc_item(zones, &args, M_WAITOK));
1985 }
1986 
1987 static void
1988 zone_lock_pair(uma_zone_t a, uma_zone_t b)
1989 {
1990 	if (a < b) {
1991 		ZONE_LOCK(a);
1992 		mtx_lock_flags(b->uz_lockptr, MTX_DUPOK);
1993 	} else {
1994 		ZONE_LOCK(b);
1995 		mtx_lock_flags(a->uz_lockptr, MTX_DUPOK);
1996 	}
1997 }
1998 
1999 static void
2000 zone_unlock_pair(uma_zone_t a, uma_zone_t b)
2001 {
2002 
2003 	ZONE_UNLOCK(a);
2004 	ZONE_UNLOCK(b);
2005 }
2006 
2007 int
2008 uma_zsecond_add(uma_zone_t zone, uma_zone_t master)
2009 {
2010 	uma_klink_t klink;
2011 	uma_klink_t kl;
2012 	int error;
2013 
2014 	error = 0;
2015 	klink = malloc(sizeof(*klink), M_TEMP, M_WAITOK | M_ZERO);
2016 
2017 	zone_lock_pair(zone, master);
2018 	/*
2019 	 * zone must use vtoslab() to resolve objects and must already be
2020 	 * a secondary.
2021 	 */
2022 	if ((zone->uz_flags & (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY))
2023 	    != (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) {
2024 		error = EINVAL;
2025 		goto out;
2026 	}
2027 	/*
2028 	 * The new master must also use vtoslab().
2029 	 */
2030 	if ((zone->uz_flags & UMA_ZONE_VTOSLAB) != UMA_ZONE_VTOSLAB) {
2031 		error = EINVAL;
2032 		goto out;
2033 	}
2034 
2035 	/*
2036 	 * The underlying object must be the same size.  rsize
2037 	 * may be different.
2038 	 */
2039 	if (master->uz_size != zone->uz_size) {
2040 		error = E2BIG;
2041 		goto out;
2042 	}
2043 	/*
2044 	 * Put it at the end of the list.
2045 	 */
2046 	klink->kl_keg = zone_first_keg(master);
2047 	LIST_FOREACH(kl, &zone->uz_kegs, kl_link) {
2048 		if (LIST_NEXT(kl, kl_link) == NULL) {
2049 			LIST_INSERT_AFTER(kl, klink, kl_link);
2050 			break;
2051 		}
2052 	}
2053 	klink = NULL;
2054 	zone->uz_flags |= UMA_ZFLAG_MULTI;
2055 	zone->uz_slab = zone_fetch_slab_multi;
2056 
2057 out:
2058 	zone_unlock_pair(zone, master);
2059 	if (klink != NULL)
2060 		free(klink, M_TEMP);
2061 
2062 	return (error);
2063 }
2064 
2065 
2066 /* See uma.h */
2067 void
2068 uma_zdestroy(uma_zone_t zone)
2069 {
2070 
2071 	sx_slock(&uma_drain_lock);
2072 	zone_free_item(zones, zone, NULL, SKIP_NONE);
2073 	sx_sunlock(&uma_drain_lock);
2074 }
2075 
2076 /* See uma.h */
2077 void *
2078 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
2079 {
2080 	void *item;
2081 	uma_cache_t cache;
2082 	uma_bucket_t bucket;
2083 	int lockfail;
2084 	int cpu;
2085 
2086 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
2087 	random_harvest_fast_uma(&zone, sizeof(zone), 1, RANDOM_UMA);
2088 
2089 	/* This is the fast path allocation */
2090 #ifdef UMA_DEBUG_ALLOC_1
2091 	printf("Allocating one item from %s(%p)\n", zone->uz_name, zone);
2092 #endif
2093 	CTR3(KTR_UMA, "uma_zalloc_arg thread %x zone %s flags %d", curthread,
2094 	    zone->uz_name, flags);
2095 
2096 	if (flags & M_WAITOK) {
2097 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
2098 		    "uma_zalloc_arg: zone \"%s\"", zone->uz_name);
2099 	}
2100 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
2101 	    ("uma_zalloc_arg: called with spinlock or critical section held"));
2102 
2103 #ifdef DEBUG_MEMGUARD
2104 	if (memguard_cmp_zone(zone)) {
2105 		item = memguard_alloc(zone->uz_size, flags);
2106 		if (item != NULL) {
2107 			if (zone->uz_init != NULL &&
2108 			    zone->uz_init(item, zone->uz_size, flags) != 0)
2109 				return (NULL);
2110 			if (zone->uz_ctor != NULL &&
2111 			    zone->uz_ctor(item, zone->uz_size, udata,
2112 			    flags) != 0) {
2113 			    	zone->uz_fini(item, zone->uz_size);
2114 				return (NULL);
2115 			}
2116 			return (item);
2117 		}
2118 		/* This is unfortunate but should not be fatal. */
2119 	}
2120 #endif
2121 	/*
2122 	 * If possible, allocate from the per-CPU cache.  There are two
2123 	 * requirements for safe access to the per-CPU cache: (1) the thread
2124 	 * accessing the cache must not be preempted or yield during access,
2125 	 * and (2) the thread must not migrate CPUs without switching which
2126 	 * cache it accesses.  We rely on a critical section to prevent
2127 	 * preemption and migration.  We release the critical section in
2128 	 * order to acquire the zone mutex if we are unable to allocate from
2129 	 * the current cache; when we re-acquire the critical section, we
2130 	 * must detect and handle migration if it has occurred.
2131 	 */
2132 	critical_enter();
2133 	cpu = curcpu;
2134 	cache = &zone->uz_cpu[cpu];
2135 
2136 zalloc_start:
2137 	bucket = cache->uc_allocbucket;
2138 	if (bucket != NULL && bucket->ub_cnt > 0) {
2139 		bucket->ub_cnt--;
2140 		item = bucket->ub_bucket[bucket->ub_cnt];
2141 #ifdef INVARIANTS
2142 		bucket->ub_bucket[bucket->ub_cnt] = NULL;
2143 #endif
2144 		KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled."));
2145 		cache->uc_allocs++;
2146 		critical_exit();
2147 		if (zone->uz_ctor != NULL &&
2148 		    zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2149 			atomic_add_long(&zone->uz_fails, 1);
2150 			zone_free_item(zone, item, udata, SKIP_DTOR);
2151 			return (NULL);
2152 		}
2153 #ifdef INVARIANTS
2154 		uma_dbg_alloc(zone, NULL, item);
2155 #endif
2156 		if (flags & M_ZERO)
2157 			uma_zero_item(item, zone);
2158 		return (item);
2159 	}
2160 
2161 	/*
2162 	 * We have run out of items in our alloc bucket.
2163 	 * See if we can switch with our free bucket.
2164 	 */
2165 	bucket = cache->uc_freebucket;
2166 	if (bucket != NULL && bucket->ub_cnt > 0) {
2167 #ifdef UMA_DEBUG_ALLOC
2168 		printf("uma_zalloc: Swapping empty with alloc.\n");
2169 #endif
2170 		cache->uc_freebucket = cache->uc_allocbucket;
2171 		cache->uc_allocbucket = bucket;
2172 		goto zalloc_start;
2173 	}
2174 
2175 	/*
2176 	 * Discard any empty allocation bucket while we hold no locks.
2177 	 */
2178 	bucket = cache->uc_allocbucket;
2179 	cache->uc_allocbucket = NULL;
2180 	critical_exit();
2181 	if (bucket != NULL)
2182 		bucket_free(zone, bucket, udata);
2183 
2184 	/* Short-circuit for zones without buckets and low memory. */
2185 	if (zone->uz_count == 0 || bucketdisable)
2186 		goto zalloc_item;
2187 
2188 	/*
2189 	 * Attempt to retrieve the item from the per-CPU cache has failed, so
2190 	 * we must go back to the zone.  This requires the zone lock, so we
2191 	 * must drop the critical section, then re-acquire it when we go back
2192 	 * to the cache.  Since the critical section is released, we may be
2193 	 * preempted or migrate.  As such, make sure not to maintain any
2194 	 * thread-local state specific to the cache from prior to releasing
2195 	 * the critical section.
2196 	 */
2197 	lockfail = 0;
2198 	if (ZONE_TRYLOCK(zone) == 0) {
2199 		/* Record contention to size the buckets. */
2200 		ZONE_LOCK(zone);
2201 		lockfail = 1;
2202 	}
2203 	critical_enter();
2204 	cpu = curcpu;
2205 	cache = &zone->uz_cpu[cpu];
2206 
2207 	/*
2208 	 * Since we have locked the zone we may as well send back our stats.
2209 	 */
2210 	atomic_add_long(&zone->uz_allocs, cache->uc_allocs);
2211 	atomic_add_long(&zone->uz_frees, cache->uc_frees);
2212 	cache->uc_allocs = 0;
2213 	cache->uc_frees = 0;
2214 
2215 	/* See if we lost the race to fill the cache. */
2216 	if (cache->uc_allocbucket != NULL) {
2217 		ZONE_UNLOCK(zone);
2218 		goto zalloc_start;
2219 	}
2220 
2221 	/*
2222 	 * Check the zone's cache of buckets.
2223 	 */
2224 	if ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) {
2225 		KASSERT(bucket->ub_cnt != 0,
2226 		    ("uma_zalloc_arg: Returning an empty bucket."));
2227 
2228 		LIST_REMOVE(bucket, ub_link);
2229 		cache->uc_allocbucket = bucket;
2230 		ZONE_UNLOCK(zone);
2231 		goto zalloc_start;
2232 	}
2233 	/* We are no longer associated with this CPU. */
2234 	critical_exit();
2235 
2236 	/*
2237 	 * We bump the uz count when the cache size is insufficient to
2238 	 * handle the working set.
2239 	 */
2240 	if (lockfail && zone->uz_count < BUCKET_MAX)
2241 		zone->uz_count++;
2242 	ZONE_UNLOCK(zone);
2243 
2244 	/*
2245 	 * Now lets just fill a bucket and put it on the free list.  If that
2246 	 * works we'll restart the allocation from the beginning and it
2247 	 * will use the just filled bucket.
2248 	 */
2249 	bucket = zone_alloc_bucket(zone, udata, flags);
2250 	if (bucket != NULL) {
2251 		ZONE_LOCK(zone);
2252 		critical_enter();
2253 		cpu = curcpu;
2254 		cache = &zone->uz_cpu[cpu];
2255 		/*
2256 		 * See if we lost the race or were migrated.  Cache the
2257 		 * initialized bucket to make this less likely or claim
2258 		 * the memory directly.
2259 		 */
2260 		if (cache->uc_allocbucket == NULL)
2261 			cache->uc_allocbucket = bucket;
2262 		else
2263 			LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link);
2264 		ZONE_UNLOCK(zone);
2265 		goto zalloc_start;
2266 	}
2267 
2268 	/*
2269 	 * We may not be able to get a bucket so return an actual item.
2270 	 */
2271 #ifdef UMA_DEBUG
2272 	printf("uma_zalloc_arg: Bucketzone returned NULL\n");
2273 #endif
2274 
2275 zalloc_item:
2276 	item = zone_alloc_item(zone, udata, flags);
2277 
2278 	return (item);
2279 }
2280 
2281 static uma_slab_t
2282 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int flags)
2283 {
2284 	uma_slab_t slab;
2285 	int reserve;
2286 
2287 	mtx_assert(&keg->uk_lock, MA_OWNED);
2288 	slab = NULL;
2289 	reserve = 0;
2290 	if ((flags & M_USE_RESERVE) == 0)
2291 		reserve = keg->uk_reserve;
2292 
2293 	for (;;) {
2294 		/*
2295 		 * Find a slab with some space.  Prefer slabs that are partially
2296 		 * used over those that are totally full.  This helps to reduce
2297 		 * fragmentation.
2298 		 */
2299 		if (keg->uk_free > reserve) {
2300 			if (!LIST_EMPTY(&keg->uk_part_slab)) {
2301 				slab = LIST_FIRST(&keg->uk_part_slab);
2302 			} else {
2303 				slab = LIST_FIRST(&keg->uk_free_slab);
2304 				LIST_REMOVE(slab, us_link);
2305 				LIST_INSERT_HEAD(&keg->uk_part_slab, slab,
2306 				    us_link);
2307 			}
2308 			MPASS(slab->us_keg == keg);
2309 			return (slab);
2310 		}
2311 
2312 		/*
2313 		 * M_NOVM means don't ask at all!
2314 		 */
2315 		if (flags & M_NOVM)
2316 			break;
2317 
2318 		if (keg->uk_maxpages && keg->uk_pages >= keg->uk_maxpages) {
2319 			keg->uk_flags |= UMA_ZFLAG_FULL;
2320 			/*
2321 			 * If this is not a multi-zone, set the FULL bit.
2322 			 * Otherwise slab_multi() takes care of it.
2323 			 */
2324 			if ((zone->uz_flags & UMA_ZFLAG_MULTI) == 0) {
2325 				zone->uz_flags |= UMA_ZFLAG_FULL;
2326 				zone_log_warning(zone);
2327 				zone_maxaction(zone);
2328 			}
2329 			if (flags & M_NOWAIT)
2330 				break;
2331 			zone->uz_sleeps++;
2332 			msleep(keg, &keg->uk_lock, PVM, "keglimit", 0);
2333 			continue;
2334 		}
2335 		slab = keg_alloc_slab(keg, zone, flags);
2336 		/*
2337 		 * If we got a slab here it's safe to mark it partially used
2338 		 * and return.  We assume that the caller is going to remove
2339 		 * at least one item.
2340 		 */
2341 		if (slab) {
2342 			MPASS(slab->us_keg == keg);
2343 			LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2344 			return (slab);
2345 		}
2346 		/*
2347 		 * We might not have been able to get a slab but another cpu
2348 		 * could have while we were unlocked.  Check again before we
2349 		 * fail.
2350 		 */
2351 		flags |= M_NOVM;
2352 	}
2353 	return (slab);
2354 }
2355 
2356 static uma_slab_t
2357 zone_fetch_slab(uma_zone_t zone, uma_keg_t keg, int flags)
2358 {
2359 	uma_slab_t slab;
2360 
2361 	if (keg == NULL) {
2362 		keg = zone_first_keg(zone);
2363 		KEG_LOCK(keg);
2364 	}
2365 
2366 	for (;;) {
2367 		slab = keg_fetch_slab(keg, zone, flags);
2368 		if (slab)
2369 			return (slab);
2370 		if (flags & (M_NOWAIT | M_NOVM))
2371 			break;
2372 	}
2373 	KEG_UNLOCK(keg);
2374 	return (NULL);
2375 }
2376 
2377 /*
2378  * uma_zone_fetch_slab_multi:  Fetches a slab from one available keg.  Returns
2379  * with the keg locked.  On NULL no lock is held.
2380  *
2381  * The last pointer is used to seed the search.  It is not required.
2382  */
2383 static uma_slab_t
2384 zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int rflags)
2385 {
2386 	uma_klink_t klink;
2387 	uma_slab_t slab;
2388 	uma_keg_t keg;
2389 	int flags;
2390 	int empty;
2391 	int full;
2392 
2393 	/*
2394 	 * Don't wait on the first pass.  This will skip limit tests
2395 	 * as well.  We don't want to block if we can find a provider
2396 	 * without blocking.
2397 	 */
2398 	flags = (rflags & ~M_WAITOK) | M_NOWAIT;
2399 	/*
2400 	 * Use the last slab allocated as a hint for where to start
2401 	 * the search.
2402 	 */
2403 	if (last != NULL) {
2404 		slab = keg_fetch_slab(last, zone, flags);
2405 		if (slab)
2406 			return (slab);
2407 		KEG_UNLOCK(last);
2408 	}
2409 	/*
2410 	 * Loop until we have a slab incase of transient failures
2411 	 * while M_WAITOK is specified.  I'm not sure this is 100%
2412 	 * required but we've done it for so long now.
2413 	 */
2414 	for (;;) {
2415 		empty = 0;
2416 		full = 0;
2417 		/*
2418 		 * Search the available kegs for slabs.  Be careful to hold the
2419 		 * correct lock while calling into the keg layer.
2420 		 */
2421 		LIST_FOREACH(klink, &zone->uz_kegs, kl_link) {
2422 			keg = klink->kl_keg;
2423 			KEG_LOCK(keg);
2424 			if ((keg->uk_flags & UMA_ZFLAG_FULL) == 0) {
2425 				slab = keg_fetch_slab(keg, zone, flags);
2426 				if (slab)
2427 					return (slab);
2428 			}
2429 			if (keg->uk_flags & UMA_ZFLAG_FULL)
2430 				full++;
2431 			else
2432 				empty++;
2433 			KEG_UNLOCK(keg);
2434 		}
2435 		if (rflags & (M_NOWAIT | M_NOVM))
2436 			break;
2437 		flags = rflags;
2438 		/*
2439 		 * All kegs are full.  XXX We can't atomically check all kegs
2440 		 * and sleep so just sleep for a short period and retry.
2441 		 */
2442 		if (full && !empty) {
2443 			ZONE_LOCK(zone);
2444 			zone->uz_flags |= UMA_ZFLAG_FULL;
2445 			zone->uz_sleeps++;
2446 			zone_log_warning(zone);
2447 			zone_maxaction(zone);
2448 			msleep(zone, zone->uz_lockptr, PVM,
2449 			    "zonelimit", hz/100);
2450 			zone->uz_flags &= ~UMA_ZFLAG_FULL;
2451 			ZONE_UNLOCK(zone);
2452 			continue;
2453 		}
2454 	}
2455 	return (NULL);
2456 }
2457 
2458 static void *
2459 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
2460 {
2461 	void *item;
2462 	uint8_t freei;
2463 
2464 	MPASS(keg == slab->us_keg);
2465 	mtx_assert(&keg->uk_lock, MA_OWNED);
2466 
2467 	freei = BIT_FFS(SLAB_SETSIZE, &slab->us_free) - 1;
2468 	BIT_CLR(SLAB_SETSIZE, freei, &slab->us_free);
2469 	item = slab->us_data + (keg->uk_rsize * freei);
2470 	slab->us_freecount--;
2471 	keg->uk_free--;
2472 
2473 	/* Move this slab to the full list */
2474 	if (slab->us_freecount == 0) {
2475 		LIST_REMOVE(slab, us_link);
2476 		LIST_INSERT_HEAD(&keg->uk_full_slab, slab, us_link);
2477 	}
2478 
2479 	return (item);
2480 }
2481 
2482 static int
2483 zone_import(uma_zone_t zone, void **bucket, int max, int flags)
2484 {
2485 	uma_slab_t slab;
2486 	uma_keg_t keg;
2487 	int i;
2488 
2489 	slab = NULL;
2490 	keg = NULL;
2491 	/* Try to keep the buckets totally full */
2492 	for (i = 0; i < max; ) {
2493 		if ((slab = zone->uz_slab(zone, keg, flags)) == NULL)
2494 			break;
2495 		keg = slab->us_keg;
2496 		while (slab->us_freecount && i < max) {
2497 			bucket[i++] = slab_alloc_item(keg, slab);
2498 			if (keg->uk_free <= keg->uk_reserve)
2499 				break;
2500 		}
2501 		/* Don't grab more than one slab at a time. */
2502 		flags &= ~M_WAITOK;
2503 		flags |= M_NOWAIT;
2504 	}
2505 	if (slab != NULL)
2506 		KEG_UNLOCK(keg);
2507 
2508 	return i;
2509 }
2510 
2511 static uma_bucket_t
2512 zone_alloc_bucket(uma_zone_t zone, void *udata, int flags)
2513 {
2514 	uma_bucket_t bucket;
2515 	int max;
2516 
2517 	/* Don't wait for buckets, preserve caller's NOVM setting. */
2518 	bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
2519 	if (bucket == NULL)
2520 		return (NULL);
2521 
2522 	max = MIN(bucket->ub_entries, zone->uz_count);
2523 	bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
2524 	    max, flags);
2525 
2526 	/*
2527 	 * Initialize the memory if necessary.
2528 	 */
2529 	if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
2530 		int i;
2531 
2532 		for (i = 0; i < bucket->ub_cnt; i++)
2533 			if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size,
2534 			    flags) != 0)
2535 				break;
2536 		/*
2537 		 * If we couldn't initialize the whole bucket, put the
2538 		 * rest back onto the freelist.
2539 		 */
2540 		if (i != bucket->ub_cnt) {
2541 			zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
2542 			    bucket->ub_cnt - i);
2543 #ifdef INVARIANTS
2544 			bzero(&bucket->ub_bucket[i],
2545 			    sizeof(void *) * (bucket->ub_cnt - i));
2546 #endif
2547 			bucket->ub_cnt = i;
2548 		}
2549 	}
2550 
2551 	if (bucket->ub_cnt == 0) {
2552 		bucket_free(zone, bucket, udata);
2553 		atomic_add_long(&zone->uz_fails, 1);
2554 		return (NULL);
2555 	}
2556 
2557 	return (bucket);
2558 }
2559 
2560 /*
2561  * Allocates a single item from a zone.
2562  *
2563  * Arguments
2564  *	zone   The zone to alloc for.
2565  *	udata  The data to be passed to the constructor.
2566  *	flags  M_WAITOK, M_NOWAIT, M_ZERO.
2567  *
2568  * Returns
2569  *	NULL if there is no memory and M_NOWAIT is set
2570  *	An item if successful
2571  */
2572 
2573 static void *
2574 zone_alloc_item(uma_zone_t zone, void *udata, int flags)
2575 {
2576 	void *item;
2577 
2578 	item = NULL;
2579 
2580 #ifdef UMA_DEBUG_ALLOC
2581 	printf("INTERNAL: Allocating one item from %s(%p)\n", zone->uz_name, zone);
2582 #endif
2583 	if (zone->uz_import(zone->uz_arg, &item, 1, flags) != 1)
2584 		goto fail;
2585 	atomic_add_long(&zone->uz_allocs, 1);
2586 
2587 	/*
2588 	 * We have to call both the zone's init (not the keg's init)
2589 	 * and the zone's ctor.  This is because the item is going from
2590 	 * a keg slab directly to the user, and the user is expecting it
2591 	 * to be both zone-init'd as well as zone-ctor'd.
2592 	 */
2593 	if (zone->uz_init != NULL) {
2594 		if (zone->uz_init(item, zone->uz_size, flags) != 0) {
2595 			zone_free_item(zone, item, udata, SKIP_FINI);
2596 			goto fail;
2597 		}
2598 	}
2599 	if (zone->uz_ctor != NULL) {
2600 		if (zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2601 			zone_free_item(zone, item, udata, SKIP_DTOR);
2602 			goto fail;
2603 		}
2604 	}
2605 #ifdef INVARIANTS
2606 	uma_dbg_alloc(zone, NULL, item);
2607 #endif
2608 	if (flags & M_ZERO)
2609 		uma_zero_item(item, zone);
2610 
2611 	return (item);
2612 
2613 fail:
2614 	atomic_add_long(&zone->uz_fails, 1);
2615 	return (NULL);
2616 }
2617 
2618 /* See uma.h */
2619 void
2620 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
2621 {
2622 	uma_cache_t cache;
2623 	uma_bucket_t bucket;
2624 	int lockfail;
2625 	int cpu;
2626 
2627 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
2628 	random_harvest_fast_uma(&zone, sizeof(zone), 1, RANDOM_UMA);
2629 
2630 #ifdef UMA_DEBUG_ALLOC_1
2631 	printf("Freeing item %p to %s(%p)\n", item, zone->uz_name, zone);
2632 #endif
2633 	CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread,
2634 	    zone->uz_name);
2635 
2636 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
2637 	    ("uma_zfree_arg: called with spinlock or critical section held"));
2638 
2639         /* uma_zfree(..., NULL) does nothing, to match free(9). */
2640         if (item == NULL)
2641                 return;
2642 #ifdef DEBUG_MEMGUARD
2643 	if (is_memguard_addr(item)) {
2644 		if (zone->uz_dtor != NULL)
2645 			zone->uz_dtor(item, zone->uz_size, udata);
2646 		if (zone->uz_fini != NULL)
2647 			zone->uz_fini(item, zone->uz_size);
2648 		memguard_free(item);
2649 		return;
2650 	}
2651 #endif
2652 #ifdef INVARIANTS
2653 	if (zone->uz_flags & UMA_ZONE_MALLOC)
2654 		uma_dbg_free(zone, udata, item);
2655 	else
2656 		uma_dbg_free(zone, NULL, item);
2657 #endif
2658 	if (zone->uz_dtor != NULL)
2659 		zone->uz_dtor(item, zone->uz_size, udata);
2660 
2661 	/*
2662 	 * The race here is acceptable.  If we miss it we'll just have to wait
2663 	 * a little longer for the limits to be reset.
2664 	 */
2665 	if (zone->uz_flags & UMA_ZFLAG_FULL)
2666 		goto zfree_item;
2667 
2668 	/*
2669 	 * If possible, free to the per-CPU cache.  There are two
2670 	 * requirements for safe access to the per-CPU cache: (1) the thread
2671 	 * accessing the cache must not be preempted or yield during access,
2672 	 * and (2) the thread must not migrate CPUs without switching which
2673 	 * cache it accesses.  We rely on a critical section to prevent
2674 	 * preemption and migration.  We release the critical section in
2675 	 * order to acquire the zone mutex if we are unable to free to the
2676 	 * current cache; when we re-acquire the critical section, we must
2677 	 * detect and handle migration if it has occurred.
2678 	 */
2679 zfree_restart:
2680 	critical_enter();
2681 	cpu = curcpu;
2682 	cache = &zone->uz_cpu[cpu];
2683 
2684 zfree_start:
2685 	/*
2686 	 * Try to free into the allocbucket first to give LIFO ordering
2687 	 * for cache-hot datastructures.  Spill over into the freebucket
2688 	 * if necessary.  Alloc will swap them if one runs dry.
2689 	 */
2690 	bucket = cache->uc_allocbucket;
2691 	if (bucket == NULL || bucket->ub_cnt >= bucket->ub_entries)
2692 		bucket = cache->uc_freebucket;
2693 	if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) {
2694 		KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL,
2695 		    ("uma_zfree: Freeing to non free bucket index."));
2696 		bucket->ub_bucket[bucket->ub_cnt] = item;
2697 		bucket->ub_cnt++;
2698 		cache->uc_frees++;
2699 		critical_exit();
2700 		return;
2701 	}
2702 
2703 	/*
2704 	 * We must go back the zone, which requires acquiring the zone lock,
2705 	 * which in turn means we must release and re-acquire the critical
2706 	 * section.  Since the critical section is released, we may be
2707 	 * preempted or migrate.  As such, make sure not to maintain any
2708 	 * thread-local state specific to the cache from prior to releasing
2709 	 * the critical section.
2710 	 */
2711 	critical_exit();
2712 	if (zone->uz_count == 0 || bucketdisable)
2713 		goto zfree_item;
2714 
2715 	lockfail = 0;
2716 	if (ZONE_TRYLOCK(zone) == 0) {
2717 		/* Record contention to size the buckets. */
2718 		ZONE_LOCK(zone);
2719 		lockfail = 1;
2720 	}
2721 	critical_enter();
2722 	cpu = curcpu;
2723 	cache = &zone->uz_cpu[cpu];
2724 
2725 	/*
2726 	 * Since we have locked the zone we may as well send back our stats.
2727 	 */
2728 	atomic_add_long(&zone->uz_allocs, cache->uc_allocs);
2729 	atomic_add_long(&zone->uz_frees, cache->uc_frees);
2730 	cache->uc_allocs = 0;
2731 	cache->uc_frees = 0;
2732 
2733 	bucket = cache->uc_freebucket;
2734 	if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) {
2735 		ZONE_UNLOCK(zone);
2736 		goto zfree_start;
2737 	}
2738 	cache->uc_freebucket = NULL;
2739 	/* We are no longer associated with this CPU. */
2740 	critical_exit();
2741 
2742 	/* Can we throw this on the zone full list? */
2743 	if (bucket != NULL) {
2744 #ifdef UMA_DEBUG_ALLOC
2745 		printf("uma_zfree: Putting old bucket on the free list.\n");
2746 #endif
2747 		/* ub_cnt is pointing to the last free item */
2748 		KASSERT(bucket->ub_cnt != 0,
2749 		    ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n"));
2750 		LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link);
2751 	}
2752 
2753 	/*
2754 	 * We bump the uz count when the cache size is insufficient to
2755 	 * handle the working set.
2756 	 */
2757 	if (lockfail && zone->uz_count < BUCKET_MAX)
2758 		zone->uz_count++;
2759 	ZONE_UNLOCK(zone);
2760 
2761 #ifdef UMA_DEBUG_ALLOC
2762 	printf("uma_zfree: Allocating new free bucket.\n");
2763 #endif
2764 	bucket = bucket_alloc(zone, udata, M_NOWAIT);
2765 	if (bucket) {
2766 		critical_enter();
2767 		cpu = curcpu;
2768 		cache = &zone->uz_cpu[cpu];
2769 		if (cache->uc_freebucket == NULL) {
2770 			cache->uc_freebucket = bucket;
2771 			goto zfree_start;
2772 		}
2773 		/*
2774 		 * We lost the race, start over.  We have to drop our
2775 		 * critical section to free the bucket.
2776 		 */
2777 		critical_exit();
2778 		bucket_free(zone, bucket, udata);
2779 		goto zfree_restart;
2780 	}
2781 
2782 	/*
2783 	 * If nothing else caught this, we'll just do an internal free.
2784 	 */
2785 zfree_item:
2786 	zone_free_item(zone, item, udata, SKIP_DTOR);
2787 
2788 	return;
2789 }
2790 
2791 static void
2792 slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item)
2793 {
2794 	uint8_t freei;
2795 
2796 	mtx_assert(&keg->uk_lock, MA_OWNED);
2797 	MPASS(keg == slab->us_keg);
2798 
2799 	/* Do we need to remove from any lists? */
2800 	if (slab->us_freecount+1 == keg->uk_ipers) {
2801 		LIST_REMOVE(slab, us_link);
2802 		LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
2803 	} else if (slab->us_freecount == 0) {
2804 		LIST_REMOVE(slab, us_link);
2805 		LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2806 	}
2807 
2808 	/* Slab management. */
2809 	freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
2810 	BIT_SET(SLAB_SETSIZE, freei, &slab->us_free);
2811 	slab->us_freecount++;
2812 
2813 	/* Keg statistics. */
2814 	keg->uk_free++;
2815 }
2816 
2817 static void
2818 zone_release(uma_zone_t zone, void **bucket, int cnt)
2819 {
2820 	void *item;
2821 	uma_slab_t slab;
2822 	uma_keg_t keg;
2823 	uint8_t *mem;
2824 	int clearfull;
2825 	int i;
2826 
2827 	clearfull = 0;
2828 	keg = zone_first_keg(zone);
2829 	KEG_LOCK(keg);
2830 	for (i = 0; i < cnt; i++) {
2831 		item = bucket[i];
2832 		if (!(zone->uz_flags & UMA_ZONE_VTOSLAB)) {
2833 			mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
2834 			if (zone->uz_flags & UMA_ZONE_HASH) {
2835 				slab = hash_sfind(&keg->uk_hash, mem);
2836 			} else {
2837 				mem += keg->uk_pgoff;
2838 				slab = (uma_slab_t)mem;
2839 			}
2840 		} else {
2841 			slab = vtoslab((vm_offset_t)item);
2842 			if (slab->us_keg != keg) {
2843 				KEG_UNLOCK(keg);
2844 				keg = slab->us_keg;
2845 				KEG_LOCK(keg);
2846 			}
2847 		}
2848 		slab_free_item(keg, slab, item);
2849 		if (keg->uk_flags & UMA_ZFLAG_FULL) {
2850 			if (keg->uk_pages < keg->uk_maxpages) {
2851 				keg->uk_flags &= ~UMA_ZFLAG_FULL;
2852 				clearfull = 1;
2853 			}
2854 
2855 			/*
2856 			 * We can handle one more allocation. Since we're
2857 			 * clearing ZFLAG_FULL, wake up all procs blocked
2858 			 * on pages. This should be uncommon, so keeping this
2859 			 * simple for now (rather than adding count of blocked
2860 			 * threads etc).
2861 			 */
2862 			wakeup(keg);
2863 		}
2864 	}
2865 	KEG_UNLOCK(keg);
2866 	if (clearfull) {
2867 		ZONE_LOCK(zone);
2868 		zone->uz_flags &= ~UMA_ZFLAG_FULL;
2869 		wakeup(zone);
2870 		ZONE_UNLOCK(zone);
2871 	}
2872 
2873 }
2874 
2875 /*
2876  * Frees a single item to any zone.
2877  *
2878  * Arguments:
2879  *	zone   The zone to free to
2880  *	item   The item we're freeing
2881  *	udata  User supplied data for the dtor
2882  *	skip   Skip dtors and finis
2883  */
2884 static void
2885 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
2886 {
2887 
2888 #ifdef INVARIANTS
2889 	if (skip == SKIP_NONE) {
2890 		if (zone->uz_flags & UMA_ZONE_MALLOC)
2891 			uma_dbg_free(zone, udata, item);
2892 		else
2893 			uma_dbg_free(zone, NULL, item);
2894 	}
2895 #endif
2896 	if (skip < SKIP_DTOR && zone->uz_dtor)
2897 		zone->uz_dtor(item, zone->uz_size, udata);
2898 
2899 	if (skip < SKIP_FINI && zone->uz_fini)
2900 		zone->uz_fini(item, zone->uz_size);
2901 
2902 	atomic_add_long(&zone->uz_frees, 1);
2903 	zone->uz_release(zone->uz_arg, &item, 1);
2904 }
2905 
2906 /* See uma.h */
2907 int
2908 uma_zone_set_max(uma_zone_t zone, int nitems)
2909 {
2910 	uma_keg_t keg;
2911 
2912 	keg = zone_first_keg(zone);
2913 	if (keg == NULL)
2914 		return (0);
2915 	KEG_LOCK(keg);
2916 	keg->uk_maxpages = (nitems / keg->uk_ipers) * keg->uk_ppera;
2917 	if (keg->uk_maxpages * keg->uk_ipers < nitems)
2918 		keg->uk_maxpages += keg->uk_ppera;
2919 	nitems = keg->uk_maxpages * keg->uk_ipers;
2920 	KEG_UNLOCK(keg);
2921 
2922 	return (nitems);
2923 }
2924 
2925 /* See uma.h */
2926 int
2927 uma_zone_get_max(uma_zone_t zone)
2928 {
2929 	int nitems;
2930 	uma_keg_t keg;
2931 
2932 	keg = zone_first_keg(zone);
2933 	if (keg == NULL)
2934 		return (0);
2935 	KEG_LOCK(keg);
2936 	nitems = keg->uk_maxpages * keg->uk_ipers;
2937 	KEG_UNLOCK(keg);
2938 
2939 	return (nitems);
2940 }
2941 
2942 /* See uma.h */
2943 void
2944 uma_zone_set_warning(uma_zone_t zone, const char *warning)
2945 {
2946 
2947 	ZONE_LOCK(zone);
2948 	zone->uz_warning = warning;
2949 	ZONE_UNLOCK(zone);
2950 }
2951 
2952 /* See uma.h */
2953 void
2954 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction)
2955 {
2956 
2957 	ZONE_LOCK(zone);
2958 	TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone);
2959 	ZONE_UNLOCK(zone);
2960 }
2961 
2962 /* See uma.h */
2963 int
2964 uma_zone_get_cur(uma_zone_t zone)
2965 {
2966 	int64_t nitems;
2967 	u_int i;
2968 
2969 	ZONE_LOCK(zone);
2970 	nitems = zone->uz_allocs - zone->uz_frees;
2971 	CPU_FOREACH(i) {
2972 		/*
2973 		 * See the comment in sysctl_vm_zone_stats() regarding the
2974 		 * safety of accessing the per-cpu caches. With the zone lock
2975 		 * held, it is safe, but can potentially result in stale data.
2976 		 */
2977 		nitems += zone->uz_cpu[i].uc_allocs -
2978 		    zone->uz_cpu[i].uc_frees;
2979 	}
2980 	ZONE_UNLOCK(zone);
2981 
2982 	return (nitems < 0 ? 0 : nitems);
2983 }
2984 
2985 /* See uma.h */
2986 void
2987 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
2988 {
2989 	uma_keg_t keg;
2990 
2991 	keg = zone_first_keg(zone);
2992 	KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type"));
2993 	KEG_LOCK(keg);
2994 	KASSERT(keg->uk_pages == 0,
2995 	    ("uma_zone_set_init on non-empty keg"));
2996 	keg->uk_init = uminit;
2997 	KEG_UNLOCK(keg);
2998 }
2999 
3000 /* See uma.h */
3001 void
3002 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
3003 {
3004 	uma_keg_t keg;
3005 
3006 	keg = zone_first_keg(zone);
3007 	KASSERT(keg != NULL, ("uma_zone_set_fini: Invalid zone type"));
3008 	KEG_LOCK(keg);
3009 	KASSERT(keg->uk_pages == 0,
3010 	    ("uma_zone_set_fini on non-empty keg"));
3011 	keg->uk_fini = fini;
3012 	KEG_UNLOCK(keg);
3013 }
3014 
3015 /* See uma.h */
3016 void
3017 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
3018 {
3019 
3020 	ZONE_LOCK(zone);
3021 	KASSERT(zone_first_keg(zone)->uk_pages == 0,
3022 	    ("uma_zone_set_zinit on non-empty keg"));
3023 	zone->uz_init = zinit;
3024 	ZONE_UNLOCK(zone);
3025 }
3026 
3027 /* See uma.h */
3028 void
3029 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
3030 {
3031 
3032 	ZONE_LOCK(zone);
3033 	KASSERT(zone_first_keg(zone)->uk_pages == 0,
3034 	    ("uma_zone_set_zfini on non-empty keg"));
3035 	zone->uz_fini = zfini;
3036 	ZONE_UNLOCK(zone);
3037 }
3038 
3039 /* See uma.h */
3040 /* XXX uk_freef is not actually used with the zone locked */
3041 void
3042 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
3043 {
3044 	uma_keg_t keg;
3045 
3046 	keg = zone_first_keg(zone);
3047 	KASSERT(keg != NULL, ("uma_zone_set_freef: Invalid zone type"));
3048 	KEG_LOCK(keg);
3049 	keg->uk_freef = freef;
3050 	KEG_UNLOCK(keg);
3051 }
3052 
3053 /* See uma.h */
3054 /* XXX uk_allocf is not actually used with the zone locked */
3055 void
3056 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
3057 {
3058 	uma_keg_t keg;
3059 
3060 	keg = zone_first_keg(zone);
3061 	KEG_LOCK(keg);
3062 	keg->uk_allocf = allocf;
3063 	KEG_UNLOCK(keg);
3064 }
3065 
3066 /* See uma.h */
3067 void
3068 uma_zone_reserve(uma_zone_t zone, int items)
3069 {
3070 	uma_keg_t keg;
3071 
3072 	keg = zone_first_keg(zone);
3073 	if (keg == NULL)
3074 		return;
3075 	KEG_LOCK(keg);
3076 	keg->uk_reserve = items;
3077 	KEG_UNLOCK(keg);
3078 
3079 	return;
3080 }
3081 
3082 /* See uma.h */
3083 int
3084 uma_zone_reserve_kva(uma_zone_t zone, int count)
3085 {
3086 	uma_keg_t keg;
3087 	vm_offset_t kva;
3088 	u_int pages;
3089 
3090 	keg = zone_first_keg(zone);
3091 	if (keg == NULL)
3092 		return (0);
3093 	pages = count / keg->uk_ipers;
3094 
3095 	if (pages * keg->uk_ipers < count)
3096 		pages++;
3097 
3098 #ifdef UMA_MD_SMALL_ALLOC
3099 	if (keg->uk_ppera > 1) {
3100 #else
3101 	if (1) {
3102 #endif
3103 		kva = kva_alloc((vm_size_t)pages * UMA_SLAB_SIZE);
3104 		if (kva == 0)
3105 			return (0);
3106 	} else
3107 		kva = 0;
3108 	KEG_LOCK(keg);
3109 	keg->uk_kva = kva;
3110 	keg->uk_offset = 0;
3111 	keg->uk_maxpages = pages;
3112 #ifdef UMA_MD_SMALL_ALLOC
3113 	keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
3114 #else
3115 	keg->uk_allocf = noobj_alloc;
3116 #endif
3117 	keg->uk_flags |= UMA_ZONE_NOFREE;
3118 	KEG_UNLOCK(keg);
3119 
3120 	return (1);
3121 }
3122 
3123 /* See uma.h */
3124 void
3125 uma_prealloc(uma_zone_t zone, int items)
3126 {
3127 	int slabs;
3128 	uma_slab_t slab;
3129 	uma_keg_t keg;
3130 
3131 	keg = zone_first_keg(zone);
3132 	if (keg == NULL)
3133 		return;
3134 	KEG_LOCK(keg);
3135 	slabs = items / keg->uk_ipers;
3136 	if (slabs * keg->uk_ipers < items)
3137 		slabs++;
3138 	while (slabs > 0) {
3139 		slab = keg_alloc_slab(keg, zone, M_WAITOK);
3140 		if (slab == NULL)
3141 			break;
3142 		MPASS(slab->us_keg == keg);
3143 		LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
3144 		slabs--;
3145 	}
3146 	KEG_UNLOCK(keg);
3147 }
3148 
3149 /* See uma.h */
3150 static void
3151 uma_reclaim_locked(bool kmem_danger)
3152 {
3153 
3154 #ifdef UMA_DEBUG
3155 	printf("UMA: vm asked us to release pages!\n");
3156 #endif
3157 	sx_assert(&uma_drain_lock, SA_XLOCKED);
3158 	bucket_enable();
3159 	zone_foreach(zone_drain);
3160 	if (vm_page_count_min() || kmem_danger) {
3161 		cache_drain_safe(NULL);
3162 		zone_foreach(zone_drain);
3163 	}
3164 	/*
3165 	 * Some slabs may have been freed but this zone will be visited early
3166 	 * we visit again so that we can free pages that are empty once other
3167 	 * zones are drained.  We have to do the same for buckets.
3168 	 */
3169 	zone_drain(slabzone);
3170 	bucket_zone_drain();
3171 }
3172 
3173 void
3174 uma_reclaim(void)
3175 {
3176 
3177 	sx_xlock(&uma_drain_lock);
3178 	uma_reclaim_locked(false);
3179 	sx_xunlock(&uma_drain_lock);
3180 }
3181 
3182 static int uma_reclaim_needed;
3183 
3184 void
3185 uma_reclaim_wakeup(void)
3186 {
3187 
3188 	uma_reclaim_needed = 1;
3189 	wakeup(&uma_reclaim_needed);
3190 }
3191 
3192 void
3193 uma_reclaim_worker(void *arg __unused)
3194 {
3195 
3196 	sx_xlock(&uma_drain_lock);
3197 	for (;;) {
3198 		sx_sleep(&uma_reclaim_needed, &uma_drain_lock, PVM,
3199 		    "umarcl", 0);
3200 		if (uma_reclaim_needed) {
3201 			uma_reclaim_needed = 0;
3202 			uma_reclaim_locked(true);
3203 		}
3204 	}
3205 }
3206 
3207 /* See uma.h */
3208 int
3209 uma_zone_exhausted(uma_zone_t zone)
3210 {
3211 	int full;
3212 
3213 	ZONE_LOCK(zone);
3214 	full = (zone->uz_flags & UMA_ZFLAG_FULL);
3215 	ZONE_UNLOCK(zone);
3216 	return (full);
3217 }
3218 
3219 int
3220 uma_zone_exhausted_nolock(uma_zone_t zone)
3221 {
3222 	return (zone->uz_flags & UMA_ZFLAG_FULL);
3223 }
3224 
3225 void *
3226 uma_large_malloc(vm_size_t size, int wait)
3227 {
3228 	void *mem;
3229 	uma_slab_t slab;
3230 	uint8_t flags;
3231 
3232 	slab = zone_alloc_item(slabzone, NULL, wait);
3233 	if (slab == NULL)
3234 		return (NULL);
3235 	mem = page_alloc(NULL, size, &flags, wait);
3236 	if (mem) {
3237 		vsetslab((vm_offset_t)mem, slab);
3238 		slab->us_data = mem;
3239 		slab->us_flags = flags | UMA_SLAB_MALLOC;
3240 		slab->us_size = size;
3241 	} else {
3242 		zone_free_item(slabzone, slab, NULL, SKIP_NONE);
3243 	}
3244 
3245 	return (mem);
3246 }
3247 
3248 void
3249 uma_large_free(uma_slab_t slab)
3250 {
3251 
3252 	page_free(slab->us_data, slab->us_size, slab->us_flags);
3253 	zone_free_item(slabzone, slab, NULL, SKIP_NONE);
3254 }
3255 
3256 static void
3257 uma_zero_item(void *item, uma_zone_t zone)
3258 {
3259 	int i;
3260 
3261 	if (zone->uz_flags & UMA_ZONE_PCPU) {
3262 		CPU_FOREACH(i)
3263 			bzero(zpcpu_get_cpu(item, i), zone->uz_size);
3264 	} else
3265 		bzero(item, zone->uz_size);
3266 }
3267 
3268 void
3269 uma_print_stats(void)
3270 {
3271 	zone_foreach(uma_print_zone);
3272 }
3273 
3274 static void
3275 slab_print(uma_slab_t slab)
3276 {
3277 	printf("slab: keg %p, data %p, freecount %d\n",
3278 		slab->us_keg, slab->us_data, slab->us_freecount);
3279 }
3280 
3281 static void
3282 cache_print(uma_cache_t cache)
3283 {
3284 	printf("alloc: %p(%d), free: %p(%d)\n",
3285 		cache->uc_allocbucket,
3286 		cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0,
3287 		cache->uc_freebucket,
3288 		cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0);
3289 }
3290 
3291 static void
3292 uma_print_keg(uma_keg_t keg)
3293 {
3294 	uma_slab_t slab;
3295 
3296 	printf("keg: %s(%p) size %d(%d) flags %#x ipers %d ppera %d "
3297 	    "out %d free %d limit %d\n",
3298 	    keg->uk_name, keg, keg->uk_size, keg->uk_rsize, keg->uk_flags,
3299 	    keg->uk_ipers, keg->uk_ppera,
3300 	    (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free,
3301 	    (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers);
3302 	printf("Part slabs:\n");
3303 	LIST_FOREACH(slab, &keg->uk_part_slab, us_link)
3304 		slab_print(slab);
3305 	printf("Free slabs:\n");
3306 	LIST_FOREACH(slab, &keg->uk_free_slab, us_link)
3307 		slab_print(slab);
3308 	printf("Full slabs:\n");
3309 	LIST_FOREACH(slab, &keg->uk_full_slab, us_link)
3310 		slab_print(slab);
3311 }
3312 
3313 void
3314 uma_print_zone(uma_zone_t zone)
3315 {
3316 	uma_cache_t cache;
3317 	uma_klink_t kl;
3318 	int i;
3319 
3320 	printf("zone: %s(%p) size %d flags %#x\n",
3321 	    zone->uz_name, zone, zone->uz_size, zone->uz_flags);
3322 	LIST_FOREACH(kl, &zone->uz_kegs, kl_link)
3323 		uma_print_keg(kl->kl_keg);
3324 	CPU_FOREACH(i) {
3325 		cache = &zone->uz_cpu[i];
3326 		printf("CPU %d Cache:\n", i);
3327 		cache_print(cache);
3328 	}
3329 }
3330 
3331 #ifdef DDB
3332 /*
3333  * Generate statistics across both the zone and its per-cpu cache's.  Return
3334  * desired statistics if the pointer is non-NULL for that statistic.
3335  *
3336  * Note: does not update the zone statistics, as it can't safely clear the
3337  * per-CPU cache statistic.
3338  *
3339  * XXXRW: Following the uc_allocbucket and uc_freebucket pointers here isn't
3340  * safe from off-CPU; we should modify the caches to track this information
3341  * directly so that we don't have to.
3342  */
3343 static void
3344 uma_zone_sumstat(uma_zone_t z, int *cachefreep, uint64_t *allocsp,
3345     uint64_t *freesp, uint64_t *sleepsp)
3346 {
3347 	uma_cache_t cache;
3348 	uint64_t allocs, frees, sleeps;
3349 	int cachefree, cpu;
3350 
3351 	allocs = frees = sleeps = 0;
3352 	cachefree = 0;
3353 	CPU_FOREACH(cpu) {
3354 		cache = &z->uz_cpu[cpu];
3355 		if (cache->uc_allocbucket != NULL)
3356 			cachefree += cache->uc_allocbucket->ub_cnt;
3357 		if (cache->uc_freebucket != NULL)
3358 			cachefree += cache->uc_freebucket->ub_cnt;
3359 		allocs += cache->uc_allocs;
3360 		frees += cache->uc_frees;
3361 	}
3362 	allocs += z->uz_allocs;
3363 	frees += z->uz_frees;
3364 	sleeps += z->uz_sleeps;
3365 	if (cachefreep != NULL)
3366 		*cachefreep = cachefree;
3367 	if (allocsp != NULL)
3368 		*allocsp = allocs;
3369 	if (freesp != NULL)
3370 		*freesp = frees;
3371 	if (sleepsp != NULL)
3372 		*sleepsp = sleeps;
3373 }
3374 #endif /* DDB */
3375 
3376 static int
3377 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
3378 {
3379 	uma_keg_t kz;
3380 	uma_zone_t z;
3381 	int count;
3382 
3383 	count = 0;
3384 	rw_rlock(&uma_rwlock);
3385 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3386 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
3387 			count++;
3388 	}
3389 	rw_runlock(&uma_rwlock);
3390 	return (sysctl_handle_int(oidp, &count, 0, req));
3391 }
3392 
3393 static int
3394 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
3395 {
3396 	struct uma_stream_header ush;
3397 	struct uma_type_header uth;
3398 	struct uma_percpu_stat ups;
3399 	uma_bucket_t bucket;
3400 	struct sbuf sbuf;
3401 	uma_cache_t cache;
3402 	uma_klink_t kl;
3403 	uma_keg_t kz;
3404 	uma_zone_t z;
3405 	uma_keg_t k;
3406 	int count, error, i;
3407 
3408 	error = sysctl_wire_old_buffer(req, 0);
3409 	if (error != 0)
3410 		return (error);
3411 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
3412 	sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL);
3413 
3414 	count = 0;
3415 	rw_rlock(&uma_rwlock);
3416 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3417 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
3418 			count++;
3419 	}
3420 
3421 	/*
3422 	 * Insert stream header.
3423 	 */
3424 	bzero(&ush, sizeof(ush));
3425 	ush.ush_version = UMA_STREAM_VERSION;
3426 	ush.ush_maxcpus = (mp_maxid + 1);
3427 	ush.ush_count = count;
3428 	(void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
3429 
3430 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3431 		LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3432 			bzero(&uth, sizeof(uth));
3433 			ZONE_LOCK(z);
3434 			strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
3435 			uth.uth_align = kz->uk_align;
3436 			uth.uth_size = kz->uk_size;
3437 			uth.uth_rsize = kz->uk_rsize;
3438 			LIST_FOREACH(kl, &z->uz_kegs, kl_link) {
3439 				k = kl->kl_keg;
3440 				uth.uth_maxpages += k->uk_maxpages;
3441 				uth.uth_pages += k->uk_pages;
3442 				uth.uth_keg_free += k->uk_free;
3443 				uth.uth_limit = (k->uk_maxpages / k->uk_ppera)
3444 				    * k->uk_ipers;
3445 			}
3446 
3447 			/*
3448 			 * A zone is secondary is it is not the first entry
3449 			 * on the keg's zone list.
3450 			 */
3451 			if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
3452 			    (LIST_FIRST(&kz->uk_zones) != z))
3453 				uth.uth_zone_flags = UTH_ZONE_SECONDARY;
3454 
3455 			LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3456 				uth.uth_zone_free += bucket->ub_cnt;
3457 			uth.uth_allocs = z->uz_allocs;
3458 			uth.uth_frees = z->uz_frees;
3459 			uth.uth_fails = z->uz_fails;
3460 			uth.uth_sleeps = z->uz_sleeps;
3461 			(void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
3462 			/*
3463 			 * While it is not normally safe to access the cache
3464 			 * bucket pointers while not on the CPU that owns the
3465 			 * cache, we only allow the pointers to be exchanged
3466 			 * without the zone lock held, not invalidated, so
3467 			 * accept the possible race associated with bucket
3468 			 * exchange during monitoring.
3469 			 */
3470 			for (i = 0; i < (mp_maxid + 1); i++) {
3471 				bzero(&ups, sizeof(ups));
3472 				if (kz->uk_flags & UMA_ZFLAG_INTERNAL)
3473 					goto skip;
3474 				if (CPU_ABSENT(i))
3475 					goto skip;
3476 				cache = &z->uz_cpu[i];
3477 				if (cache->uc_allocbucket != NULL)
3478 					ups.ups_cache_free +=
3479 					    cache->uc_allocbucket->ub_cnt;
3480 				if (cache->uc_freebucket != NULL)
3481 					ups.ups_cache_free +=
3482 					    cache->uc_freebucket->ub_cnt;
3483 				ups.ups_allocs = cache->uc_allocs;
3484 				ups.ups_frees = cache->uc_frees;
3485 skip:
3486 				(void)sbuf_bcat(&sbuf, &ups, sizeof(ups));
3487 			}
3488 			ZONE_UNLOCK(z);
3489 		}
3490 	}
3491 	rw_runlock(&uma_rwlock);
3492 	error = sbuf_finish(&sbuf);
3493 	sbuf_delete(&sbuf);
3494 	return (error);
3495 }
3496 
3497 int
3498 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS)
3499 {
3500 	uma_zone_t zone = *(uma_zone_t *)arg1;
3501 	int error, max;
3502 
3503 	max = uma_zone_get_max(zone);
3504 	error = sysctl_handle_int(oidp, &max, 0, req);
3505 	if (error || !req->newptr)
3506 		return (error);
3507 
3508 	uma_zone_set_max(zone, max);
3509 
3510 	return (0);
3511 }
3512 
3513 int
3514 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS)
3515 {
3516 	uma_zone_t zone = *(uma_zone_t *)arg1;
3517 	int cur;
3518 
3519 	cur = uma_zone_get_cur(zone);
3520 	return (sysctl_handle_int(oidp, &cur, 0, req));
3521 }
3522 
3523 #ifdef INVARIANTS
3524 static uma_slab_t
3525 uma_dbg_getslab(uma_zone_t zone, void *item)
3526 {
3527 	uma_slab_t slab;
3528 	uma_keg_t keg;
3529 	uint8_t *mem;
3530 
3531 	mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
3532 	if (zone->uz_flags & UMA_ZONE_VTOSLAB) {
3533 		slab = vtoslab((vm_offset_t)mem);
3534 	} else {
3535 		/*
3536 		 * It is safe to return the slab here even though the
3537 		 * zone is unlocked because the item's allocation state
3538 		 * essentially holds a reference.
3539 		 */
3540 		ZONE_LOCK(zone);
3541 		keg = LIST_FIRST(&zone->uz_kegs)->kl_keg;
3542 		if (keg->uk_flags & UMA_ZONE_HASH)
3543 			slab = hash_sfind(&keg->uk_hash, mem);
3544 		else
3545 			slab = (uma_slab_t)(mem + keg->uk_pgoff);
3546 		ZONE_UNLOCK(zone);
3547 	}
3548 
3549 	return (slab);
3550 }
3551 
3552 /*
3553  * Set up the slab's freei data such that uma_dbg_free can function.
3554  *
3555  */
3556 static void
3557 uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item)
3558 {
3559 	uma_keg_t keg;
3560 	int freei;
3561 
3562 	if (zone_first_keg(zone) == NULL)
3563 		return;
3564 	if (slab == NULL) {
3565 		slab = uma_dbg_getslab(zone, item);
3566 		if (slab == NULL)
3567 			panic("uma: item %p did not belong to zone %s\n",
3568 			    item, zone->uz_name);
3569 	}
3570 	keg = slab->us_keg;
3571 	freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
3572 
3573 	if (BIT_ISSET(SLAB_SETSIZE, freei, &slab->us_debugfree))
3574 		panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)\n",
3575 		    item, zone, zone->uz_name, slab, freei);
3576 	BIT_SET_ATOMIC(SLAB_SETSIZE, freei, &slab->us_debugfree);
3577 
3578 	return;
3579 }
3580 
3581 /*
3582  * Verifies freed addresses.  Checks for alignment, valid slab membership
3583  * and duplicate frees.
3584  *
3585  */
3586 static void
3587 uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item)
3588 {
3589 	uma_keg_t keg;
3590 	int freei;
3591 
3592 	if (zone_first_keg(zone) == NULL)
3593 		return;
3594 	if (slab == NULL) {
3595 		slab = uma_dbg_getslab(zone, item);
3596 		if (slab == NULL)
3597 			panic("uma: Freed item %p did not belong to zone %s\n",
3598 			    item, zone->uz_name);
3599 	}
3600 	keg = slab->us_keg;
3601 	freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
3602 
3603 	if (freei >= keg->uk_ipers)
3604 		panic("Invalid free of %p from zone %p(%s) slab %p(%d)\n",
3605 		    item, zone, zone->uz_name, slab, freei);
3606 
3607 	if (((freei * keg->uk_rsize) + slab->us_data) != item)
3608 		panic("Unaligned free of %p from zone %p(%s) slab %p(%d)\n",
3609 		    item, zone, zone->uz_name, slab, freei);
3610 
3611 	if (!BIT_ISSET(SLAB_SETSIZE, freei, &slab->us_debugfree))
3612 		panic("Duplicate free of %p from zone %p(%s) slab %p(%d)\n",
3613 		    item, zone, zone->uz_name, slab, freei);
3614 
3615 	BIT_CLR_ATOMIC(SLAB_SETSIZE, freei, &slab->us_debugfree);
3616 }
3617 #endif /* INVARIANTS */
3618 
3619 #ifdef DDB
3620 DB_SHOW_COMMAND(uma, db_show_uma)
3621 {
3622 	uint64_t allocs, frees, sleeps;
3623 	uma_bucket_t bucket;
3624 	uma_keg_t kz;
3625 	uma_zone_t z;
3626 	int cachefree;
3627 
3628 	db_printf("%18s %8s %8s %8s %12s %8s %8s\n", "Zone", "Size", "Used",
3629 	    "Free", "Requests", "Sleeps", "Bucket");
3630 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3631 		LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3632 			if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
3633 				allocs = z->uz_allocs;
3634 				frees = z->uz_frees;
3635 				sleeps = z->uz_sleeps;
3636 				cachefree = 0;
3637 			} else
3638 				uma_zone_sumstat(z, &cachefree, &allocs,
3639 				    &frees, &sleeps);
3640 			if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
3641 			    (LIST_FIRST(&kz->uk_zones) != z)))
3642 				cachefree += kz->uk_free;
3643 			LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3644 				cachefree += bucket->ub_cnt;
3645 			db_printf("%18s %8ju %8jd %8d %12ju %8ju %8u\n",
3646 			    z->uz_name, (uintmax_t)kz->uk_size,
3647 			    (intmax_t)(allocs - frees), cachefree,
3648 			    (uintmax_t)allocs, sleeps, z->uz_count);
3649 			if (db_pager_quit)
3650 				return;
3651 		}
3652 	}
3653 }
3654 
3655 DB_SHOW_COMMAND(umacache, db_show_umacache)
3656 {
3657 	uint64_t allocs, frees;
3658 	uma_bucket_t bucket;
3659 	uma_zone_t z;
3660 	int cachefree;
3661 
3662 	db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
3663 	    "Requests", "Bucket");
3664 	LIST_FOREACH(z, &uma_cachezones, uz_link) {
3665 		uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL);
3666 		LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3667 			cachefree += bucket->ub_cnt;
3668 		db_printf("%18s %8ju %8jd %8d %12ju %8u\n",
3669 		    z->uz_name, (uintmax_t)z->uz_size,
3670 		    (intmax_t)(allocs - frees), cachefree,
3671 		    (uintmax_t)allocs, z->uz_count);
3672 		if (db_pager_quit)
3673 			return;
3674 	}
3675 }
3676 #endif	/* DDB */
3677