xref: /freebsd/sys/vm/uma_core.c (revision 95fd7f26154281528fc8fc01fd29e961a054555f)
1 /*-
2  * Copyright (c) 2002-2005, 2009, 2013 Jeffrey Roberson <jeff@FreeBSD.org>
3  * Copyright (c) 2004, 2005 Bosko Milekic <bmilekic@FreeBSD.org>
4  * Copyright (c) 2004-2006 Robert N. M. Watson
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice unmodified, this list of conditions, and the following
12  *    disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
18  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
19  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
20  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
21  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
22  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
23  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
24  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
25  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
26  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
27  */
28 
29 /*
30  * uma_core.c  Implementation of the Universal Memory allocator
31  *
32  * This allocator is intended to replace the multitude of similar object caches
33  * in the standard FreeBSD kernel.  The intent is to be flexible as well as
34  * effecient.  A primary design goal is to return unused memory to the rest of
35  * the system.  This will make the system as a whole more flexible due to the
36  * ability to move memory to subsystems which most need it instead of leaving
37  * pools of reserved memory unused.
38  *
39  * The basic ideas stem from similar slab/zone based allocators whose algorithms
40  * are well known.
41  *
42  */
43 
44 /*
45  * TODO:
46  *	- Improve memory usage for large allocations
47  *	- Investigate cache size adjustments
48  */
49 
50 #include <sys/cdefs.h>
51 __FBSDID("$FreeBSD$");
52 
53 /* I should really use ktr.. */
54 /*
55 #define UMA_DEBUG 1
56 #define UMA_DEBUG_ALLOC 1
57 #define UMA_DEBUG_ALLOC_1 1
58 */
59 
60 #include "opt_ddb.h"
61 #include "opt_param.h"
62 #include "opt_vm.h"
63 
64 #include <sys/param.h>
65 #include <sys/systm.h>
66 #include <sys/bitset.h>
67 #include <sys/kernel.h>
68 #include <sys/types.h>
69 #include <sys/queue.h>
70 #include <sys/malloc.h>
71 #include <sys/ktr.h>
72 #include <sys/lock.h>
73 #include <sys/sysctl.h>
74 #include <sys/mutex.h>
75 #include <sys/proc.h>
76 #include <sys/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/vmmeter.h>
82 
83 #include <vm/vm.h>
84 #include <vm/vm_object.h>
85 #include <vm/vm_page.h>
86 #include <vm/vm_pageout.h>
87 #include <vm/vm_param.h>
88 #include <vm/vm_map.h>
89 #include <vm/vm_kern.h>
90 #include <vm/vm_extern.h>
91 #include <vm/uma.h>
92 #include <vm/uma_int.h>
93 #include <vm/uma_dbg.h>
94 
95 #include <ddb/ddb.h>
96 
97 #ifdef DEBUG_MEMGUARD
98 #include <vm/memguard.h>
99 #endif
100 
101 /*
102  * This is the zone and keg from which all zones are spawned.  The idea is that
103  * even the zone & keg heads are allocated from the allocator, so we use the
104  * bss section to bootstrap us.
105  */
106 static struct uma_keg masterkeg;
107 static struct uma_zone masterzone_k;
108 static struct uma_zone masterzone_z;
109 static uma_zone_t kegs = &masterzone_k;
110 static uma_zone_t zones = &masterzone_z;
111 
112 /* This is the zone from which all of uma_slab_t's are allocated. */
113 static uma_zone_t slabzone;
114 static uma_zone_t slabrefzone;	/* With refcounters (for UMA_ZONE_REFCNT) */
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  * Only mbuf clusters use ref zones.  Just provide enough references
158  * to support the one user.  New code should not use the ref facility.
159  */
160 static const u_int uma_max_ipers_ref = PAGE_SIZE / MCLBYTES;
161 
162 /*
163  * This is the handle used to schedule events that need to happen
164  * outside of the allocation fast path.
165  */
166 static struct callout uma_callout;
167 #define	UMA_TIMEOUT	20		/* Seconds for callout interval. */
168 
169 /*
170  * This structure is passed as the zone ctor arg so that I don't have to create
171  * a special allocation function just for zones.
172  */
173 struct uma_zctor_args {
174 	const char *name;
175 	size_t size;
176 	uma_ctor ctor;
177 	uma_dtor dtor;
178 	uma_init uminit;
179 	uma_fini fini;
180 	uma_import import;
181 	uma_release release;
182 	void *arg;
183 	uma_keg_t keg;
184 	int align;
185 	uint32_t flags;
186 };
187 
188 struct uma_kctor_args {
189 	uma_zone_t zone;
190 	size_t size;
191 	uma_init uminit;
192 	uma_fini fini;
193 	int align;
194 	uint32_t flags;
195 };
196 
197 struct uma_bucket_zone {
198 	uma_zone_t	ubz_zone;
199 	char		*ubz_name;
200 	int		ubz_entries;	/* Number of items it can hold. */
201 	int		ubz_maxsize;	/* Maximum allocation size per-item. */
202 };
203 
204 /*
205  * Compute the actual number of bucket entries to pack them in power
206  * of two sizes for more efficient space utilization.
207  */
208 #define	BUCKET_SIZE(n)						\
209     (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *))
210 
211 #define	BUCKET_MAX	BUCKET_SIZE(256)
212 
213 struct uma_bucket_zone bucket_zones[] = {
214 	{ NULL, "4 Bucket", BUCKET_SIZE(4), 4096 },
215 	{ NULL, "6 Bucket", BUCKET_SIZE(6), 3072 },
216 	{ NULL, "8 Bucket", BUCKET_SIZE(8), 2048 },
217 	{ NULL, "12 Bucket", BUCKET_SIZE(12), 1536 },
218 	{ NULL, "16 Bucket", BUCKET_SIZE(16), 1024 },
219 	{ NULL, "32 Bucket", BUCKET_SIZE(32), 512 },
220 	{ NULL, "64 Bucket", BUCKET_SIZE(64), 256 },
221 	{ NULL, "128 Bucket", BUCKET_SIZE(128), 128 },
222 	{ NULL, "256 Bucket", BUCKET_SIZE(256), 64 },
223 	{ NULL, NULL, 0}
224 };
225 
226 /*
227  * Flags and enumerations to be passed to internal functions.
228  */
229 enum zfreeskip { SKIP_NONE = 0, SKIP_DTOR, SKIP_FINI };
230 
231 /* Prototypes.. */
232 
233 static void *noobj_alloc(uma_zone_t, vm_size_t, uint8_t *, int);
234 static void *page_alloc(uma_zone_t, vm_size_t, uint8_t *, int);
235 static void *startup_alloc(uma_zone_t, vm_size_t, uint8_t *, int);
236 static void page_free(void *, vm_size_t, uint8_t);
237 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int);
238 static void cache_drain(uma_zone_t);
239 static void bucket_drain(uma_zone_t, uma_bucket_t);
240 static void bucket_cache_drain(uma_zone_t zone);
241 static int keg_ctor(void *, int, void *, int);
242 static void keg_dtor(void *, int, void *);
243 static int zone_ctor(void *, int, void *, int);
244 static void zone_dtor(void *, int, void *);
245 static int zero_init(void *, int, int);
246 static void keg_small_init(uma_keg_t keg);
247 static void keg_large_init(uma_keg_t keg);
248 static void zone_foreach(void (*zfunc)(uma_zone_t));
249 static void zone_timeout(uma_zone_t zone);
250 static int hash_alloc(struct uma_hash *);
251 static int hash_expand(struct uma_hash *, struct uma_hash *);
252 static void hash_free(struct uma_hash *hash);
253 static void uma_timeout(void *);
254 static void uma_startup3(void);
255 static void *zone_alloc_item(uma_zone_t, void *, int);
256 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip);
257 static void bucket_enable(void);
258 static void bucket_init(void);
259 static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int);
260 static void bucket_free(uma_zone_t zone, uma_bucket_t, void *);
261 static void bucket_zone_drain(void);
262 static uma_bucket_t zone_alloc_bucket(uma_zone_t zone, void *, int flags);
263 static uma_slab_t zone_fetch_slab(uma_zone_t zone, uma_keg_t last, int flags);
264 static uma_slab_t zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int flags);
265 static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab);
266 static void slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item);
267 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit,
268     uma_fini fini, int align, uint32_t flags);
269 static int zone_import(uma_zone_t zone, void **bucket, int max, int flags);
270 static void zone_release(uma_zone_t zone, void **bucket, int cnt);
271 static void uma_zero_item(void *item, uma_zone_t zone);
272 
273 void uma_print_zone(uma_zone_t);
274 void uma_print_stats(void);
275 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS);
276 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS);
277 
278 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
279 
280 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLTYPE_INT,
281     0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones");
282 
283 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLTYPE_STRUCT,
284     0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats");
285 
286 static int zone_warnings = 1;
287 SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RWTUN, &zone_warnings, 0,
288     "Warn when UMA zones becomes full");
289 
290 /*
291  * This routine checks to see whether or not it's safe to enable buckets.
292  */
293 static void
294 bucket_enable(void)
295 {
296 	bucketdisable = vm_page_count_min();
297 }
298 
299 /*
300  * Initialize bucket_zones, the array of zones of buckets of various sizes.
301  *
302  * For each zone, calculate the memory required for each bucket, consisting
303  * of the header and an array of pointers.
304  */
305 static void
306 bucket_init(void)
307 {
308 	struct uma_bucket_zone *ubz;
309 	int size;
310 
311 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) {
312 		size = roundup(sizeof(struct uma_bucket), sizeof(void *));
313 		size += sizeof(void *) * ubz->ubz_entries;
314 		ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
315 		    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
316 		    UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET);
317 	}
318 }
319 
320 /*
321  * Given a desired number of entries for a bucket, return the zone from which
322  * to allocate the bucket.
323  */
324 static struct uma_bucket_zone *
325 bucket_zone_lookup(int entries)
326 {
327 	struct uma_bucket_zone *ubz;
328 
329 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
330 		if (ubz->ubz_entries >= entries)
331 			return (ubz);
332 	ubz--;
333 	return (ubz);
334 }
335 
336 static int
337 bucket_select(int size)
338 {
339 	struct uma_bucket_zone *ubz;
340 
341 	ubz = &bucket_zones[0];
342 	if (size > ubz->ubz_maxsize)
343 		return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1);
344 
345 	for (; ubz->ubz_entries != 0; ubz++)
346 		if (ubz->ubz_maxsize < size)
347 			break;
348 	ubz--;
349 	return (ubz->ubz_entries);
350 }
351 
352 static uma_bucket_t
353 bucket_alloc(uma_zone_t zone, void *udata, int flags)
354 {
355 	struct uma_bucket_zone *ubz;
356 	uma_bucket_t bucket;
357 
358 	/*
359 	 * This is to stop us from allocating per cpu buckets while we're
360 	 * running out of vm.boot_pages.  Otherwise, we would exhaust the
361 	 * boot pages.  This also prevents us from allocating buckets in
362 	 * low memory situations.
363 	 */
364 	if (bucketdisable)
365 		return (NULL);
366 	/*
367 	 * To limit bucket recursion we store the original zone flags
368 	 * in a cookie passed via zalloc_arg/zfree_arg.  This allows the
369 	 * NOVM flag to persist even through deep recursions.  We also
370 	 * store ZFLAG_BUCKET once we have recursed attempting to allocate
371 	 * a bucket for a bucket zone so we do not allow infinite bucket
372 	 * recursion.  This cookie will even persist to frees of unused
373 	 * buckets via the allocation path or bucket allocations in the
374 	 * free path.
375 	 */
376 	if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
377 		udata = (void *)(uintptr_t)zone->uz_flags;
378 	else {
379 		if ((uintptr_t)udata & UMA_ZFLAG_BUCKET)
380 			return (NULL);
381 		udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET);
382 	}
383 	if ((uintptr_t)udata & UMA_ZFLAG_CACHEONLY)
384 		flags |= M_NOVM;
385 	ubz = bucket_zone_lookup(zone->uz_count);
386 	if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0)
387 		ubz++;
388 	bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags);
389 	if (bucket) {
390 #ifdef INVARIANTS
391 		bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
392 #endif
393 		bucket->ub_cnt = 0;
394 		bucket->ub_entries = ubz->ubz_entries;
395 	}
396 
397 	return (bucket);
398 }
399 
400 static void
401 bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata)
402 {
403 	struct uma_bucket_zone *ubz;
404 
405 	KASSERT(bucket->ub_cnt == 0,
406 	    ("bucket_free: Freeing a non free bucket."));
407 	if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
408 		udata = (void *)(uintptr_t)zone->uz_flags;
409 	ubz = bucket_zone_lookup(bucket->ub_entries);
410 	uma_zfree_arg(ubz->ubz_zone, bucket, udata);
411 }
412 
413 static void
414 bucket_zone_drain(void)
415 {
416 	struct uma_bucket_zone *ubz;
417 
418 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
419 		zone_drain(ubz->ubz_zone);
420 }
421 
422 static void
423 zone_log_warning(uma_zone_t zone)
424 {
425 	static const struct timeval warninterval = { 300, 0 };
426 
427 	if (!zone_warnings || zone->uz_warning == NULL)
428 		return;
429 
430 	if (ratecheck(&zone->uz_ratecheck, &warninterval))
431 		printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning);
432 }
433 
434 static void
435 zone_foreach_keg(uma_zone_t zone, void (*kegfn)(uma_keg_t))
436 {
437 	uma_klink_t klink;
438 
439 	LIST_FOREACH(klink, &zone->uz_kegs, kl_link)
440 		kegfn(klink->kl_keg);
441 }
442 
443 /*
444  * Routine called by timeout which is used to fire off some time interval
445  * based calculations.  (stats, hash size, etc.)
446  *
447  * Arguments:
448  *	arg   Unused
449  *
450  * Returns:
451  *	Nothing
452  */
453 static void
454 uma_timeout(void *unused)
455 {
456 	bucket_enable();
457 	zone_foreach(zone_timeout);
458 
459 	/* Reschedule this event */
460 	callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
461 }
462 
463 /*
464  * Routine to perform timeout driven calculations.  This expands the
465  * hashes and does per cpu statistics aggregation.
466  *
467  *  Returns nothing.
468  */
469 static void
470 keg_timeout(uma_keg_t keg)
471 {
472 
473 	KEG_LOCK(keg);
474 	/*
475 	 * Expand the keg hash table.
476 	 *
477 	 * This is done if the number of slabs is larger than the hash size.
478 	 * What I'm trying to do here is completely reduce collisions.  This
479 	 * may be a little aggressive.  Should I allow for two collisions max?
480 	 */
481 	if (keg->uk_flags & UMA_ZONE_HASH &&
482 	    keg->uk_pages / keg->uk_ppera >= keg->uk_hash.uh_hashsize) {
483 		struct uma_hash newhash;
484 		struct uma_hash oldhash;
485 		int ret;
486 
487 		/*
488 		 * This is so involved because allocating and freeing
489 		 * while the keg lock is held will lead to deadlock.
490 		 * I have to do everything in stages and check for
491 		 * races.
492 		 */
493 		newhash = keg->uk_hash;
494 		KEG_UNLOCK(keg);
495 		ret = hash_alloc(&newhash);
496 		KEG_LOCK(keg);
497 		if (ret) {
498 			if (hash_expand(&keg->uk_hash, &newhash)) {
499 				oldhash = keg->uk_hash;
500 				keg->uk_hash = newhash;
501 			} else
502 				oldhash = newhash;
503 
504 			KEG_UNLOCK(keg);
505 			hash_free(&oldhash);
506 			return;
507 		}
508 	}
509 	KEG_UNLOCK(keg);
510 }
511 
512 static void
513 zone_timeout(uma_zone_t zone)
514 {
515 
516 	zone_foreach_keg(zone, &keg_timeout);
517 }
518 
519 /*
520  * Allocate and zero fill the next sized hash table from the appropriate
521  * backing store.
522  *
523  * Arguments:
524  *	hash  A new hash structure with the old hash size in uh_hashsize
525  *
526  * Returns:
527  *	1 on sucess and 0 on failure.
528  */
529 static int
530 hash_alloc(struct uma_hash *hash)
531 {
532 	int oldsize;
533 	int alloc;
534 
535 	oldsize = hash->uh_hashsize;
536 
537 	/* We're just going to go to a power of two greater */
538 	if (oldsize)  {
539 		hash->uh_hashsize = oldsize * 2;
540 		alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
541 		hash->uh_slab_hash = (struct slabhead *)malloc(alloc,
542 		    M_UMAHASH, M_NOWAIT);
543 	} else {
544 		alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
545 		hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
546 		    M_WAITOK);
547 		hash->uh_hashsize = UMA_HASH_SIZE_INIT;
548 	}
549 	if (hash->uh_slab_hash) {
550 		bzero(hash->uh_slab_hash, alloc);
551 		hash->uh_hashmask = hash->uh_hashsize - 1;
552 		return (1);
553 	}
554 
555 	return (0);
556 }
557 
558 /*
559  * Expands the hash table for HASH zones.  This is done from zone_timeout
560  * to reduce collisions.  This must not be done in the regular allocation
561  * path, otherwise, we can recurse on the vm while allocating pages.
562  *
563  * Arguments:
564  *	oldhash  The hash you want to expand
565  *	newhash  The hash structure for the new table
566  *
567  * Returns:
568  *	Nothing
569  *
570  * Discussion:
571  */
572 static int
573 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
574 {
575 	uma_slab_t slab;
576 	int hval;
577 	int i;
578 
579 	if (!newhash->uh_slab_hash)
580 		return (0);
581 
582 	if (oldhash->uh_hashsize >= newhash->uh_hashsize)
583 		return (0);
584 
585 	/*
586 	 * I need to investigate hash algorithms for resizing without a
587 	 * full rehash.
588 	 */
589 
590 	for (i = 0; i < oldhash->uh_hashsize; i++)
591 		while (!SLIST_EMPTY(&oldhash->uh_slab_hash[i])) {
592 			slab = SLIST_FIRST(&oldhash->uh_slab_hash[i]);
593 			SLIST_REMOVE_HEAD(&oldhash->uh_slab_hash[i], us_hlink);
594 			hval = UMA_HASH(newhash, slab->us_data);
595 			SLIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
596 			    slab, us_hlink);
597 		}
598 
599 	return (1);
600 }
601 
602 /*
603  * Free the hash bucket to the appropriate backing store.
604  *
605  * Arguments:
606  *	slab_hash  The hash bucket we're freeing
607  *	hashsize   The number of entries in that hash bucket
608  *
609  * Returns:
610  *	Nothing
611  */
612 static void
613 hash_free(struct uma_hash *hash)
614 {
615 	if (hash->uh_slab_hash == NULL)
616 		return;
617 	if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
618 		zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE);
619 	else
620 		free(hash->uh_slab_hash, M_UMAHASH);
621 }
622 
623 /*
624  * Frees all outstanding items in a bucket
625  *
626  * Arguments:
627  *	zone   The zone to free to, must be unlocked.
628  *	bucket The free/alloc bucket with items, cpu queue must be locked.
629  *
630  * Returns:
631  *	Nothing
632  */
633 
634 static void
635 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
636 {
637 	int i;
638 
639 	if (bucket == NULL)
640 		return;
641 
642 	if (zone->uz_fini)
643 		for (i = 0; i < bucket->ub_cnt; i++)
644 			zone->uz_fini(bucket->ub_bucket[i], zone->uz_size);
645 	zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt);
646 	bucket->ub_cnt = 0;
647 }
648 
649 /*
650  * Drains the per cpu caches for a zone.
651  *
652  * NOTE: This may only be called while the zone is being turn down, and not
653  * during normal operation.  This is necessary in order that we do not have
654  * to migrate CPUs to drain the per-CPU caches.
655  *
656  * Arguments:
657  *	zone     The zone to drain, must be unlocked.
658  *
659  * Returns:
660  *	Nothing
661  */
662 static void
663 cache_drain(uma_zone_t zone)
664 {
665 	uma_cache_t cache;
666 	int cpu;
667 
668 	/*
669 	 * XXX: It is safe to not lock the per-CPU caches, because we're
670 	 * tearing down the zone anyway.  I.e., there will be no further use
671 	 * of the caches at this point.
672 	 *
673 	 * XXX: It would good to be able to assert that the zone is being
674 	 * torn down to prevent improper use of cache_drain().
675 	 *
676 	 * XXX: We lock the zone before passing into bucket_cache_drain() as
677 	 * it is used elsewhere.  Should the tear-down path be made special
678 	 * there in some form?
679 	 */
680 	CPU_FOREACH(cpu) {
681 		cache = &zone->uz_cpu[cpu];
682 		bucket_drain(zone, cache->uc_allocbucket);
683 		bucket_drain(zone, cache->uc_freebucket);
684 		if (cache->uc_allocbucket != NULL)
685 			bucket_free(zone, cache->uc_allocbucket, NULL);
686 		if (cache->uc_freebucket != NULL)
687 			bucket_free(zone, cache->uc_freebucket, NULL);
688 		cache->uc_allocbucket = cache->uc_freebucket = NULL;
689 	}
690 	ZONE_LOCK(zone);
691 	bucket_cache_drain(zone);
692 	ZONE_UNLOCK(zone);
693 }
694 
695 static void
696 cache_shrink(uma_zone_t zone)
697 {
698 
699 	if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
700 		return;
701 
702 	ZONE_LOCK(zone);
703 	zone->uz_count = (zone->uz_count_min + zone->uz_count) / 2;
704 	ZONE_UNLOCK(zone);
705 }
706 
707 static void
708 cache_drain_safe_cpu(uma_zone_t zone)
709 {
710 	uma_cache_t cache;
711 	uma_bucket_t b1, b2;
712 
713 	if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
714 		return;
715 
716 	b1 = b2 = NULL;
717 	ZONE_LOCK(zone);
718 	critical_enter();
719 	cache = &zone->uz_cpu[curcpu];
720 	if (cache->uc_allocbucket) {
721 		if (cache->uc_allocbucket->ub_cnt != 0)
722 			LIST_INSERT_HEAD(&zone->uz_buckets,
723 			    cache->uc_allocbucket, ub_link);
724 		else
725 			b1 = cache->uc_allocbucket;
726 		cache->uc_allocbucket = NULL;
727 	}
728 	if (cache->uc_freebucket) {
729 		if (cache->uc_freebucket->ub_cnt != 0)
730 			LIST_INSERT_HEAD(&zone->uz_buckets,
731 			    cache->uc_freebucket, ub_link);
732 		else
733 			b2 = cache->uc_freebucket;
734 		cache->uc_freebucket = NULL;
735 	}
736 	critical_exit();
737 	ZONE_UNLOCK(zone);
738 	if (b1)
739 		bucket_free(zone, b1, NULL);
740 	if (b2)
741 		bucket_free(zone, b2, NULL);
742 }
743 
744 /*
745  * Safely drain per-CPU caches of a zone(s) to alloc bucket.
746  * This is an expensive call because it needs to bind to all CPUs
747  * one by one and enter a critical section on each of them in order
748  * to safely access their cache buckets.
749  * Zone lock must not be held on call this function.
750  */
751 static void
752 cache_drain_safe(uma_zone_t zone)
753 {
754 	int cpu;
755 
756 	/*
757 	 * Polite bucket sizes shrinking was not enouth, shrink aggressively.
758 	 */
759 	if (zone)
760 		cache_shrink(zone);
761 	else
762 		zone_foreach(cache_shrink);
763 
764 	CPU_FOREACH(cpu) {
765 		thread_lock(curthread);
766 		sched_bind(curthread, cpu);
767 		thread_unlock(curthread);
768 
769 		if (zone)
770 			cache_drain_safe_cpu(zone);
771 		else
772 			zone_foreach(cache_drain_safe_cpu);
773 	}
774 	thread_lock(curthread);
775 	sched_unbind(curthread);
776 	thread_unlock(curthread);
777 }
778 
779 /*
780  * Drain the cached buckets from a zone.  Expects a locked zone on entry.
781  */
782 static void
783 bucket_cache_drain(uma_zone_t zone)
784 {
785 	uma_bucket_t bucket;
786 
787 	/*
788 	 * Drain the bucket queues and free the buckets, we just keep two per
789 	 * cpu (alloc/free).
790 	 */
791 	while ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) {
792 		LIST_REMOVE(bucket, ub_link);
793 		ZONE_UNLOCK(zone);
794 		bucket_drain(zone, bucket);
795 		bucket_free(zone, bucket, NULL);
796 		ZONE_LOCK(zone);
797 	}
798 
799 	/*
800 	 * Shrink further bucket sizes.  Price of single zone lock collision
801 	 * is probably lower then price of global cache drain.
802 	 */
803 	if (zone->uz_count > zone->uz_count_min)
804 		zone->uz_count--;
805 }
806 
807 static void
808 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start)
809 {
810 	uint8_t *mem;
811 	int i;
812 	uint8_t flags;
813 
814 	mem = slab->us_data;
815 	flags = slab->us_flags;
816 	i = start;
817 	if (keg->uk_fini != NULL) {
818 		for (i--; i > -1; i--)
819 			keg->uk_fini(slab->us_data + (keg->uk_rsize * i),
820 			    keg->uk_size);
821 	}
822 	if (keg->uk_flags & UMA_ZONE_OFFPAGE)
823 		zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE);
824 #ifdef UMA_DEBUG
825 	printf("%s: Returning %d bytes.\n", keg->uk_name,
826 	    PAGE_SIZE * keg->uk_ppera);
827 #endif
828 	keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags);
829 }
830 
831 /*
832  * Frees pages from a keg back to the system.  This is done on demand from
833  * the pageout daemon.
834  *
835  * Returns nothing.
836  */
837 static void
838 keg_drain(uma_keg_t keg)
839 {
840 	struct slabhead freeslabs = { 0 };
841 	uma_slab_t slab;
842 	uma_slab_t n;
843 
844 	/*
845 	 * We don't want to take pages from statically allocated kegs at this
846 	 * time
847 	 */
848 	if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL)
849 		return;
850 
851 #ifdef UMA_DEBUG
852 	printf("%s free items: %u\n", keg->uk_name, keg->uk_free);
853 #endif
854 	KEG_LOCK(keg);
855 	if (keg->uk_free == 0)
856 		goto finished;
857 
858 	slab = LIST_FIRST(&keg->uk_free_slab);
859 	while (slab) {
860 		n = LIST_NEXT(slab, us_link);
861 
862 		/* We have no where to free these to */
863 		if (slab->us_flags & UMA_SLAB_BOOT) {
864 			slab = n;
865 			continue;
866 		}
867 
868 		LIST_REMOVE(slab, us_link);
869 		keg->uk_pages -= keg->uk_ppera;
870 		keg->uk_free -= keg->uk_ipers;
871 
872 		if (keg->uk_flags & UMA_ZONE_HASH)
873 			UMA_HASH_REMOVE(&keg->uk_hash, slab, slab->us_data);
874 
875 		SLIST_INSERT_HEAD(&freeslabs, slab, us_hlink);
876 
877 		slab = n;
878 	}
879 finished:
880 	KEG_UNLOCK(keg);
881 
882 	while ((slab = SLIST_FIRST(&freeslabs)) != NULL) {
883 		SLIST_REMOVE(&freeslabs, slab, uma_slab, us_hlink);
884 		keg_free_slab(keg, slab, keg->uk_ipers);
885 	}
886 }
887 
888 static void
889 zone_drain_wait(uma_zone_t zone, int waitok)
890 {
891 
892 	/*
893 	 * Set draining to interlock with zone_dtor() so we can release our
894 	 * locks as we go.  Only dtor() should do a WAITOK call since it
895 	 * is the only call that knows the structure will still be available
896 	 * when it wakes up.
897 	 */
898 	ZONE_LOCK(zone);
899 	while (zone->uz_flags & UMA_ZFLAG_DRAINING) {
900 		if (waitok == M_NOWAIT)
901 			goto out;
902 		msleep(zone, zone->uz_lockptr, PVM, "zonedrain", 1);
903 	}
904 	zone->uz_flags |= UMA_ZFLAG_DRAINING;
905 	bucket_cache_drain(zone);
906 	ZONE_UNLOCK(zone);
907 	/*
908 	 * The DRAINING flag protects us from being freed while
909 	 * we're running.  Normally the uma_rwlock would protect us but we
910 	 * must be able to release and acquire the right lock for each keg.
911 	 */
912 	zone_foreach_keg(zone, &keg_drain);
913 	ZONE_LOCK(zone);
914 	zone->uz_flags &= ~UMA_ZFLAG_DRAINING;
915 	wakeup(zone);
916 out:
917 	ZONE_UNLOCK(zone);
918 }
919 
920 void
921 zone_drain(uma_zone_t zone)
922 {
923 
924 	zone_drain_wait(zone, M_NOWAIT);
925 }
926 
927 /*
928  * Allocate a new slab for a keg.  This does not insert the slab onto a list.
929  *
930  * Arguments:
931  *	wait  Shall we wait?
932  *
933  * Returns:
934  *	The slab that was allocated or NULL if there is no memory and the
935  *	caller specified M_NOWAIT.
936  */
937 static uma_slab_t
938 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int wait)
939 {
940 	uma_slabrefcnt_t slabref;
941 	uma_alloc allocf;
942 	uma_slab_t slab;
943 	uint8_t *mem;
944 	uint8_t flags;
945 	int i;
946 
947 	mtx_assert(&keg->uk_lock, MA_OWNED);
948 	slab = NULL;
949 	mem = NULL;
950 
951 #ifdef UMA_DEBUG
952 	printf("alloc_slab:  Allocating a new slab for %s\n", keg->uk_name);
953 #endif
954 	allocf = keg->uk_allocf;
955 	KEG_UNLOCK(keg);
956 
957 	if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
958 		slab = zone_alloc_item(keg->uk_slabzone, NULL, wait);
959 		if (slab == NULL)
960 			goto out;
961 	}
962 
963 	/*
964 	 * This reproduces the old vm_zone behavior of zero filling pages the
965 	 * first time they are added to a zone.
966 	 *
967 	 * Malloced items are zeroed in uma_zalloc.
968 	 */
969 
970 	if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
971 		wait |= M_ZERO;
972 	else
973 		wait &= ~M_ZERO;
974 
975 	if (keg->uk_flags & UMA_ZONE_NODUMP)
976 		wait |= M_NODUMP;
977 
978 	/* zone is passed for legacy reasons. */
979 	mem = allocf(zone, keg->uk_ppera * PAGE_SIZE, &flags, wait);
980 	if (mem == NULL) {
981 		if (keg->uk_flags & UMA_ZONE_OFFPAGE)
982 			zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE);
983 		slab = NULL;
984 		goto out;
985 	}
986 
987 	/* Point the slab into the allocated memory */
988 	if (!(keg->uk_flags & UMA_ZONE_OFFPAGE))
989 		slab = (uma_slab_t )(mem + keg->uk_pgoff);
990 
991 	if (keg->uk_flags & UMA_ZONE_VTOSLAB)
992 		for (i = 0; i < keg->uk_ppera; i++)
993 			vsetslab((vm_offset_t)mem + (i * PAGE_SIZE), slab);
994 
995 	slab->us_keg = keg;
996 	slab->us_data = mem;
997 	slab->us_freecount = keg->uk_ipers;
998 	slab->us_flags = flags;
999 	BIT_FILL(SLAB_SETSIZE, &slab->us_free);
1000 #ifdef INVARIANTS
1001 	BIT_ZERO(SLAB_SETSIZE, &slab->us_debugfree);
1002 #endif
1003 	if (keg->uk_flags & UMA_ZONE_REFCNT) {
1004 		slabref = (uma_slabrefcnt_t)slab;
1005 		for (i = 0; i < keg->uk_ipers; i++)
1006 			slabref->us_refcnt[i] = 0;
1007 	}
1008 
1009 	if (keg->uk_init != NULL) {
1010 		for (i = 0; i < keg->uk_ipers; i++)
1011 			if (keg->uk_init(slab->us_data + (keg->uk_rsize * i),
1012 			    keg->uk_size, wait) != 0)
1013 				break;
1014 		if (i != keg->uk_ipers) {
1015 			keg_free_slab(keg, slab, i);
1016 			slab = NULL;
1017 			goto out;
1018 		}
1019 	}
1020 out:
1021 	KEG_LOCK(keg);
1022 
1023 	if (slab != NULL) {
1024 		if (keg->uk_flags & UMA_ZONE_HASH)
1025 			UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
1026 
1027 		keg->uk_pages += keg->uk_ppera;
1028 		keg->uk_free += keg->uk_ipers;
1029 	}
1030 
1031 	return (slab);
1032 }
1033 
1034 /*
1035  * This function is intended to be used early on in place of page_alloc() so
1036  * that we may use the boot time page cache to satisfy allocations before
1037  * the VM is ready.
1038  */
1039 static void *
1040 startup_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *pflag, int wait)
1041 {
1042 	uma_keg_t keg;
1043 	uma_slab_t tmps;
1044 	int pages, check_pages;
1045 
1046 	keg = zone_first_keg(zone);
1047 	pages = howmany(bytes, PAGE_SIZE);
1048 	check_pages = pages - 1;
1049 	KASSERT(pages > 0, ("startup_alloc can't reserve 0 pages\n"));
1050 
1051 	/*
1052 	 * Check our small startup cache to see if it has pages remaining.
1053 	 */
1054 	mtx_lock(&uma_boot_pages_mtx);
1055 
1056 	/* First check if we have enough room. */
1057 	tmps = LIST_FIRST(&uma_boot_pages);
1058 	while (tmps != NULL && check_pages-- > 0)
1059 		tmps = LIST_NEXT(tmps, us_link);
1060 	if (tmps != NULL) {
1061 		/*
1062 		 * It's ok to lose tmps references.  The last one will
1063 		 * have tmps->us_data pointing to the start address of
1064 		 * "pages" contiguous pages of memory.
1065 		 */
1066 		while (pages-- > 0) {
1067 			tmps = LIST_FIRST(&uma_boot_pages);
1068 			LIST_REMOVE(tmps, us_link);
1069 		}
1070 		mtx_unlock(&uma_boot_pages_mtx);
1071 		*pflag = tmps->us_flags;
1072 		return (tmps->us_data);
1073 	}
1074 	mtx_unlock(&uma_boot_pages_mtx);
1075 	if (booted < UMA_STARTUP2)
1076 		panic("UMA: Increase vm.boot_pages");
1077 	/*
1078 	 * Now that we've booted reset these users to their real allocator.
1079 	 */
1080 #ifdef UMA_MD_SMALL_ALLOC
1081 	keg->uk_allocf = (keg->uk_ppera > 1) ? page_alloc : uma_small_alloc;
1082 #else
1083 	keg->uk_allocf = page_alloc;
1084 #endif
1085 	return keg->uk_allocf(zone, bytes, pflag, wait);
1086 }
1087 
1088 /*
1089  * Allocates a number of pages from the system
1090  *
1091  * Arguments:
1092  *	bytes  The number of bytes requested
1093  *	wait  Shall we wait?
1094  *
1095  * Returns:
1096  *	A pointer to the alloced memory or possibly
1097  *	NULL if M_NOWAIT is set.
1098  */
1099 static void *
1100 page_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *pflag, int wait)
1101 {
1102 	void *p;	/* Returned page */
1103 
1104 	*pflag = UMA_SLAB_KMEM;
1105 	p = (void *) kmem_malloc(kmem_arena, bytes, wait);
1106 
1107 	return (p);
1108 }
1109 
1110 /*
1111  * Allocates a number of pages from within an object
1112  *
1113  * Arguments:
1114  *	bytes  The number of bytes requested
1115  *	wait   Shall we wait?
1116  *
1117  * Returns:
1118  *	A pointer to the alloced memory or possibly
1119  *	NULL if M_NOWAIT is set.
1120  */
1121 static void *
1122 noobj_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *flags, int wait)
1123 {
1124 	TAILQ_HEAD(, vm_page) alloctail;
1125 	u_long npages;
1126 	vm_offset_t retkva, zkva;
1127 	vm_page_t p, p_next;
1128 	uma_keg_t keg;
1129 
1130 	TAILQ_INIT(&alloctail);
1131 	keg = zone_first_keg(zone);
1132 
1133 	npages = howmany(bytes, PAGE_SIZE);
1134 	while (npages > 0) {
1135 		p = vm_page_alloc(NULL, 0, VM_ALLOC_INTERRUPT |
1136 		    VM_ALLOC_WIRED | VM_ALLOC_NOOBJ);
1137 		if (p != NULL) {
1138 			/*
1139 			 * Since the page does not belong to an object, its
1140 			 * listq is unused.
1141 			 */
1142 			TAILQ_INSERT_TAIL(&alloctail, p, listq);
1143 			npages--;
1144 			continue;
1145 		}
1146 		if (wait & M_WAITOK) {
1147 			VM_WAIT;
1148 			continue;
1149 		}
1150 
1151 		/*
1152 		 * Page allocation failed, free intermediate pages and
1153 		 * exit.
1154 		 */
1155 		TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1156 			vm_page_unwire(p, PQ_NONE);
1157 			vm_page_free(p);
1158 		}
1159 		return (NULL);
1160 	}
1161 	*flags = UMA_SLAB_PRIV;
1162 	zkva = keg->uk_kva +
1163 	    atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
1164 	retkva = zkva;
1165 	TAILQ_FOREACH(p, &alloctail, listq) {
1166 		pmap_qenter(zkva, &p, 1);
1167 		zkva += PAGE_SIZE;
1168 	}
1169 
1170 	return ((void *)retkva);
1171 }
1172 
1173 /*
1174  * Frees a number of pages to the system
1175  *
1176  * Arguments:
1177  *	mem   A pointer to the memory to be freed
1178  *	size  The size of the memory being freed
1179  *	flags The original p->us_flags field
1180  *
1181  * Returns:
1182  *	Nothing
1183  */
1184 static void
1185 page_free(void *mem, vm_size_t size, uint8_t flags)
1186 {
1187 	struct vmem *vmem;
1188 
1189 	if (flags & UMA_SLAB_KMEM)
1190 		vmem = kmem_arena;
1191 	else if (flags & UMA_SLAB_KERNEL)
1192 		vmem = kernel_arena;
1193 	else
1194 		panic("UMA: page_free used with invalid flags %d", flags);
1195 
1196 	kmem_free(vmem, (vm_offset_t)mem, size);
1197 }
1198 
1199 /*
1200  * Zero fill initializer
1201  *
1202  * Arguments/Returns follow uma_init specifications
1203  */
1204 static int
1205 zero_init(void *mem, int size, int flags)
1206 {
1207 	bzero(mem, size);
1208 	return (0);
1209 }
1210 
1211 /*
1212  * Finish creating a small uma keg.  This calculates ipers, and the keg size.
1213  *
1214  * Arguments
1215  *	keg  The zone we should initialize
1216  *
1217  * Returns
1218  *	Nothing
1219  */
1220 static void
1221 keg_small_init(uma_keg_t keg)
1222 {
1223 	u_int rsize;
1224 	u_int memused;
1225 	u_int wastedspace;
1226 	u_int shsize;
1227 
1228 	if (keg->uk_flags & UMA_ZONE_PCPU) {
1229 		u_int ncpus = mp_ncpus ? mp_ncpus : MAXCPU;
1230 
1231 		keg->uk_slabsize = sizeof(struct pcpu);
1232 		keg->uk_ppera = howmany(ncpus * sizeof(struct pcpu),
1233 		    PAGE_SIZE);
1234 	} else {
1235 		keg->uk_slabsize = UMA_SLAB_SIZE;
1236 		keg->uk_ppera = 1;
1237 	}
1238 
1239 	/*
1240 	 * Calculate the size of each allocation (rsize) according to
1241 	 * alignment.  If the requested size is smaller than we have
1242 	 * allocation bits for we round it up.
1243 	 */
1244 	rsize = keg->uk_size;
1245 	if (rsize < keg->uk_slabsize / SLAB_SETSIZE)
1246 		rsize = keg->uk_slabsize / SLAB_SETSIZE;
1247 	if (rsize & keg->uk_align)
1248 		rsize = (rsize & ~keg->uk_align) + (keg->uk_align + 1);
1249 	keg->uk_rsize = rsize;
1250 
1251 	KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
1252 	    keg->uk_rsize < sizeof(struct pcpu),
1253 	    ("%s: size %u too large", __func__, keg->uk_rsize));
1254 
1255 	if (keg->uk_flags & UMA_ZONE_REFCNT)
1256 		rsize += sizeof(uint32_t);
1257 
1258 	if (keg->uk_flags & UMA_ZONE_OFFPAGE)
1259 		shsize = 0;
1260 	else
1261 		shsize = sizeof(struct uma_slab);
1262 
1263 	keg->uk_ipers = (keg->uk_slabsize - shsize) / rsize;
1264 	KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE,
1265 	    ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers));
1266 
1267 	memused = keg->uk_ipers * rsize + shsize;
1268 	wastedspace = keg->uk_slabsize - memused;
1269 
1270 	/*
1271 	 * We can't do OFFPAGE if we're internal or if we've been
1272 	 * asked to not go to the VM for buckets.  If we do this we
1273 	 * may end up going to the VM  for slabs which we do not
1274 	 * want to do if we're UMA_ZFLAG_CACHEONLY as a result
1275 	 * of UMA_ZONE_VM, which clearly forbids it.
1276 	 */
1277 	if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) ||
1278 	    (keg->uk_flags & UMA_ZFLAG_CACHEONLY))
1279 		return;
1280 
1281 	/*
1282 	 * See if using an OFFPAGE slab will limit our waste.  Only do
1283 	 * this if it permits more items per-slab.
1284 	 *
1285 	 * XXX We could try growing slabsize to limit max waste as well.
1286 	 * Historically this was not done because the VM could not
1287 	 * efficiently handle contiguous allocations.
1288 	 */
1289 	if ((wastedspace >= keg->uk_slabsize / UMA_MAX_WASTE) &&
1290 	    (keg->uk_ipers < (keg->uk_slabsize / keg->uk_rsize))) {
1291 		keg->uk_ipers = keg->uk_slabsize / keg->uk_rsize;
1292 		KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE,
1293 		    ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers));
1294 #ifdef UMA_DEBUG
1295 		printf("UMA decided we need offpage slab headers for "
1296 		    "keg: %s, calculated wastedspace = %d, "
1297 		    "maximum wasted space allowed = %d, "
1298 		    "calculated ipers = %d, "
1299 		    "new wasted space = %d\n", keg->uk_name, wastedspace,
1300 		    keg->uk_slabsize / UMA_MAX_WASTE, keg->uk_ipers,
1301 		    keg->uk_slabsize - keg->uk_ipers * keg->uk_rsize);
1302 #endif
1303 		keg->uk_flags |= UMA_ZONE_OFFPAGE;
1304 	}
1305 
1306 	if ((keg->uk_flags & UMA_ZONE_OFFPAGE) &&
1307 	    (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1308 		keg->uk_flags |= UMA_ZONE_HASH;
1309 }
1310 
1311 /*
1312  * Finish creating a large (> UMA_SLAB_SIZE) uma kegs.  Just give in and do
1313  * OFFPAGE for now.  When I can allow for more dynamic slab sizes this will be
1314  * more complicated.
1315  *
1316  * Arguments
1317  *	keg  The keg we should initialize
1318  *
1319  * Returns
1320  *	Nothing
1321  */
1322 static void
1323 keg_large_init(uma_keg_t keg)
1324 {
1325 	u_int shsize;
1326 
1327 	KASSERT(keg != NULL, ("Keg is null in keg_large_init"));
1328 	KASSERT((keg->uk_flags & UMA_ZFLAG_CACHEONLY) == 0,
1329 	    ("keg_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY keg"));
1330 	KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0,
1331 	    ("%s: Cannot large-init a UMA_ZONE_PCPU keg", __func__));
1332 
1333 	keg->uk_ppera = howmany(keg->uk_size, PAGE_SIZE);
1334 	keg->uk_slabsize = keg->uk_ppera * PAGE_SIZE;
1335 	keg->uk_ipers = 1;
1336 	keg->uk_rsize = keg->uk_size;
1337 
1338 	/* We can't do OFFPAGE if we're internal, bail out here. */
1339 	if (keg->uk_flags & UMA_ZFLAG_INTERNAL)
1340 		return;
1341 
1342 	/* Check whether we have enough space to not do OFFPAGE. */
1343 	if ((keg->uk_flags & UMA_ZONE_OFFPAGE) == 0) {
1344 		shsize = sizeof(struct uma_slab);
1345 		if (keg->uk_flags & UMA_ZONE_REFCNT)
1346 			shsize += keg->uk_ipers * sizeof(uint32_t);
1347 		if (shsize & UMA_ALIGN_PTR)
1348 			shsize = (shsize & ~UMA_ALIGN_PTR) +
1349 			    (UMA_ALIGN_PTR + 1);
1350 
1351 		if ((PAGE_SIZE * keg->uk_ppera) - keg->uk_rsize < shsize)
1352 			keg->uk_flags |= UMA_ZONE_OFFPAGE;
1353 	}
1354 
1355 	if ((keg->uk_flags & UMA_ZONE_OFFPAGE) &&
1356 	    (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1357 		keg->uk_flags |= UMA_ZONE_HASH;
1358 }
1359 
1360 static void
1361 keg_cachespread_init(uma_keg_t keg)
1362 {
1363 	int alignsize;
1364 	int trailer;
1365 	int pages;
1366 	int rsize;
1367 
1368 	KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0,
1369 	    ("%s: Cannot cachespread-init a UMA_ZONE_PCPU keg", __func__));
1370 
1371 	alignsize = keg->uk_align + 1;
1372 	rsize = keg->uk_size;
1373 	/*
1374 	 * We want one item to start on every align boundary in a page.  To
1375 	 * do this we will span pages.  We will also extend the item by the
1376 	 * size of align if it is an even multiple of align.  Otherwise, it
1377 	 * would fall on the same boundary every time.
1378 	 */
1379 	if (rsize & keg->uk_align)
1380 		rsize = (rsize & ~keg->uk_align) + alignsize;
1381 	if ((rsize & alignsize) == 0)
1382 		rsize += alignsize;
1383 	trailer = rsize - keg->uk_size;
1384 	pages = (rsize * (PAGE_SIZE / alignsize)) / PAGE_SIZE;
1385 	pages = MIN(pages, (128 * 1024) / PAGE_SIZE);
1386 	keg->uk_rsize = rsize;
1387 	keg->uk_ppera = pages;
1388 	keg->uk_slabsize = UMA_SLAB_SIZE;
1389 	keg->uk_ipers = ((pages * PAGE_SIZE) + trailer) / rsize;
1390 	keg->uk_flags |= UMA_ZONE_OFFPAGE | UMA_ZONE_VTOSLAB;
1391 	KASSERT(keg->uk_ipers <= SLAB_SETSIZE,
1392 	    ("%s: keg->uk_ipers too high(%d) increase max_ipers", __func__,
1393 	    keg->uk_ipers));
1394 }
1395 
1396 /*
1397  * Keg header ctor.  This initializes all fields, locks, etc.  And inserts
1398  * the keg onto the global keg list.
1399  *
1400  * Arguments/Returns follow uma_ctor specifications
1401  *	udata  Actually uma_kctor_args
1402  */
1403 static int
1404 keg_ctor(void *mem, int size, void *udata, int flags)
1405 {
1406 	struct uma_kctor_args *arg = udata;
1407 	uma_keg_t keg = mem;
1408 	uma_zone_t zone;
1409 
1410 	bzero(keg, size);
1411 	keg->uk_size = arg->size;
1412 	keg->uk_init = arg->uminit;
1413 	keg->uk_fini = arg->fini;
1414 	keg->uk_align = arg->align;
1415 	keg->uk_free = 0;
1416 	keg->uk_reserve = 0;
1417 	keg->uk_pages = 0;
1418 	keg->uk_flags = arg->flags;
1419 	keg->uk_allocf = page_alloc;
1420 	keg->uk_freef = page_free;
1421 	keg->uk_slabzone = NULL;
1422 
1423 	/*
1424 	 * The master zone is passed to us at keg-creation time.
1425 	 */
1426 	zone = arg->zone;
1427 	keg->uk_name = zone->uz_name;
1428 
1429 	if (arg->flags & UMA_ZONE_VM)
1430 		keg->uk_flags |= UMA_ZFLAG_CACHEONLY;
1431 
1432 	if (arg->flags & UMA_ZONE_ZINIT)
1433 		keg->uk_init = zero_init;
1434 
1435 	if (arg->flags & UMA_ZONE_REFCNT || arg->flags & UMA_ZONE_MALLOC)
1436 		keg->uk_flags |= UMA_ZONE_VTOSLAB;
1437 
1438 	if (arg->flags & UMA_ZONE_PCPU)
1439 #ifdef SMP
1440 		keg->uk_flags |= UMA_ZONE_OFFPAGE;
1441 #else
1442 		keg->uk_flags &= ~UMA_ZONE_PCPU;
1443 #endif
1444 
1445 	if (keg->uk_flags & UMA_ZONE_CACHESPREAD) {
1446 		keg_cachespread_init(keg);
1447 	} else if (keg->uk_flags & UMA_ZONE_REFCNT) {
1448 		if (keg->uk_size >
1449 		    (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt) -
1450 		    sizeof(uint32_t)))
1451 			keg_large_init(keg);
1452 		else
1453 			keg_small_init(keg);
1454 	} else {
1455 		if (keg->uk_size > (UMA_SLAB_SIZE - sizeof(struct uma_slab)))
1456 			keg_large_init(keg);
1457 		else
1458 			keg_small_init(keg);
1459 	}
1460 
1461 	if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
1462 		if (keg->uk_flags & UMA_ZONE_REFCNT) {
1463 			if (keg->uk_ipers > uma_max_ipers_ref)
1464 				panic("Too many ref items per zone: %d > %d\n",
1465 				    keg->uk_ipers, uma_max_ipers_ref);
1466 			keg->uk_slabzone = slabrefzone;
1467 		} else
1468 			keg->uk_slabzone = slabzone;
1469 	}
1470 
1471 	/*
1472 	 * If we haven't booted yet we need allocations to go through the
1473 	 * startup cache until the vm is ready.
1474 	 */
1475 	if (keg->uk_ppera == 1) {
1476 #ifdef UMA_MD_SMALL_ALLOC
1477 		keg->uk_allocf = uma_small_alloc;
1478 		keg->uk_freef = uma_small_free;
1479 
1480 		if (booted < UMA_STARTUP)
1481 			keg->uk_allocf = startup_alloc;
1482 #else
1483 		if (booted < UMA_STARTUP2)
1484 			keg->uk_allocf = startup_alloc;
1485 #endif
1486 	} else if (booted < UMA_STARTUP2 &&
1487 	    (keg->uk_flags & UMA_ZFLAG_INTERNAL))
1488 		keg->uk_allocf = startup_alloc;
1489 
1490 	/*
1491 	 * Initialize keg's lock
1492 	 */
1493 	KEG_LOCK_INIT(keg, (arg->flags & UMA_ZONE_MTXCLASS));
1494 
1495 	/*
1496 	 * If we're putting the slab header in the actual page we need to
1497 	 * figure out where in each page it goes.  This calculates a right
1498 	 * justified offset into the memory on an ALIGN_PTR boundary.
1499 	 */
1500 	if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) {
1501 		u_int totsize;
1502 
1503 		/* Size of the slab struct and free list */
1504 		totsize = sizeof(struct uma_slab);
1505 
1506 		/* Size of the reference counts. */
1507 		if (keg->uk_flags & UMA_ZONE_REFCNT)
1508 			totsize += keg->uk_ipers * sizeof(uint32_t);
1509 
1510 		if (totsize & UMA_ALIGN_PTR)
1511 			totsize = (totsize & ~UMA_ALIGN_PTR) +
1512 			    (UMA_ALIGN_PTR + 1);
1513 		keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - totsize;
1514 
1515 		/*
1516 		 * The only way the following is possible is if with our
1517 		 * UMA_ALIGN_PTR adjustments we are now bigger than
1518 		 * UMA_SLAB_SIZE.  I haven't checked whether this is
1519 		 * mathematically possible for all cases, so we make
1520 		 * sure here anyway.
1521 		 */
1522 		totsize = keg->uk_pgoff + sizeof(struct uma_slab);
1523 		if (keg->uk_flags & UMA_ZONE_REFCNT)
1524 			totsize += keg->uk_ipers * sizeof(uint32_t);
1525 		if (totsize > PAGE_SIZE * keg->uk_ppera) {
1526 			printf("zone %s ipers %d rsize %d size %d\n",
1527 			    zone->uz_name, keg->uk_ipers, keg->uk_rsize,
1528 			    keg->uk_size);
1529 			panic("UMA slab won't fit.");
1530 		}
1531 	}
1532 
1533 	if (keg->uk_flags & UMA_ZONE_HASH)
1534 		hash_alloc(&keg->uk_hash);
1535 
1536 #ifdef UMA_DEBUG
1537 	printf("UMA: %s(%p) size %d(%d) flags %#x ipers %d ppera %d out %d free %d\n",
1538 	    zone->uz_name, zone, keg->uk_size, keg->uk_rsize, keg->uk_flags,
1539 	    keg->uk_ipers, keg->uk_ppera,
1540 	    (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free);
1541 #endif
1542 
1543 	LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
1544 
1545 	rw_wlock(&uma_rwlock);
1546 	LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
1547 	rw_wunlock(&uma_rwlock);
1548 	return (0);
1549 }
1550 
1551 /*
1552  * Zone header ctor.  This initializes all fields, locks, etc.
1553  *
1554  * Arguments/Returns follow uma_ctor specifications
1555  *	udata  Actually uma_zctor_args
1556  */
1557 static int
1558 zone_ctor(void *mem, int size, void *udata, int flags)
1559 {
1560 	struct uma_zctor_args *arg = udata;
1561 	uma_zone_t zone = mem;
1562 	uma_zone_t z;
1563 	uma_keg_t keg;
1564 
1565 	bzero(zone, size);
1566 	zone->uz_name = arg->name;
1567 	zone->uz_ctor = arg->ctor;
1568 	zone->uz_dtor = arg->dtor;
1569 	zone->uz_slab = zone_fetch_slab;
1570 	zone->uz_init = NULL;
1571 	zone->uz_fini = NULL;
1572 	zone->uz_allocs = 0;
1573 	zone->uz_frees = 0;
1574 	zone->uz_fails = 0;
1575 	zone->uz_sleeps = 0;
1576 	zone->uz_count = 0;
1577 	zone->uz_count_min = 0;
1578 	zone->uz_flags = 0;
1579 	zone->uz_warning = NULL;
1580 	timevalclear(&zone->uz_ratecheck);
1581 	keg = arg->keg;
1582 
1583 	ZONE_LOCK_INIT(zone, (arg->flags & UMA_ZONE_MTXCLASS));
1584 
1585 	/*
1586 	 * This is a pure cache zone, no kegs.
1587 	 */
1588 	if (arg->import) {
1589 		if (arg->flags & UMA_ZONE_VM)
1590 			arg->flags |= UMA_ZFLAG_CACHEONLY;
1591 		zone->uz_flags = arg->flags;
1592 		zone->uz_size = arg->size;
1593 		zone->uz_import = arg->import;
1594 		zone->uz_release = arg->release;
1595 		zone->uz_arg = arg->arg;
1596 		zone->uz_lockptr = &zone->uz_lock;
1597 		rw_wlock(&uma_rwlock);
1598 		LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link);
1599 		rw_wunlock(&uma_rwlock);
1600 		goto out;
1601 	}
1602 
1603 	/*
1604 	 * Use the regular zone/keg/slab allocator.
1605 	 */
1606 	zone->uz_import = (uma_import)zone_import;
1607 	zone->uz_release = (uma_release)zone_release;
1608 	zone->uz_arg = zone;
1609 
1610 	if (arg->flags & UMA_ZONE_SECONDARY) {
1611 		KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
1612 		zone->uz_init = arg->uminit;
1613 		zone->uz_fini = arg->fini;
1614 		zone->uz_lockptr = &keg->uk_lock;
1615 		zone->uz_flags |= UMA_ZONE_SECONDARY;
1616 		rw_wlock(&uma_rwlock);
1617 		ZONE_LOCK(zone);
1618 		LIST_FOREACH(z, &keg->uk_zones, uz_link) {
1619 			if (LIST_NEXT(z, uz_link) == NULL) {
1620 				LIST_INSERT_AFTER(z, zone, uz_link);
1621 				break;
1622 			}
1623 		}
1624 		ZONE_UNLOCK(zone);
1625 		rw_wunlock(&uma_rwlock);
1626 	} else if (keg == NULL) {
1627 		if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
1628 		    arg->align, arg->flags)) == NULL)
1629 			return (ENOMEM);
1630 	} else {
1631 		struct uma_kctor_args karg;
1632 		int error;
1633 
1634 		/* We should only be here from uma_startup() */
1635 		karg.size = arg->size;
1636 		karg.uminit = arg->uminit;
1637 		karg.fini = arg->fini;
1638 		karg.align = arg->align;
1639 		karg.flags = arg->flags;
1640 		karg.zone = zone;
1641 		error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
1642 		    flags);
1643 		if (error)
1644 			return (error);
1645 	}
1646 
1647 	/*
1648 	 * Link in the first keg.
1649 	 */
1650 	zone->uz_klink.kl_keg = keg;
1651 	LIST_INSERT_HEAD(&zone->uz_kegs, &zone->uz_klink, kl_link);
1652 	zone->uz_lockptr = &keg->uk_lock;
1653 	zone->uz_size = keg->uk_size;
1654 	zone->uz_flags |= (keg->uk_flags &
1655 	    (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
1656 
1657 	/*
1658 	 * Some internal zones don't have room allocated for the per cpu
1659 	 * caches.  If we're internal, bail out here.
1660 	 */
1661 	if (keg->uk_flags & UMA_ZFLAG_INTERNAL) {
1662 		KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
1663 		    ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
1664 		return (0);
1665 	}
1666 
1667 out:
1668 	if ((arg->flags & UMA_ZONE_MAXBUCKET) == 0)
1669 		zone->uz_count = bucket_select(zone->uz_size);
1670 	else
1671 		zone->uz_count = BUCKET_MAX;
1672 	zone->uz_count_min = zone->uz_count;
1673 
1674 	return (0);
1675 }
1676 
1677 /*
1678  * Keg header dtor.  This frees all data, destroys locks, frees the hash
1679  * table and removes the keg from the global list.
1680  *
1681  * Arguments/Returns follow uma_dtor specifications
1682  *	udata  unused
1683  */
1684 static void
1685 keg_dtor(void *arg, int size, void *udata)
1686 {
1687 	uma_keg_t keg;
1688 
1689 	keg = (uma_keg_t)arg;
1690 	KEG_LOCK(keg);
1691 	if (keg->uk_free != 0) {
1692 		printf("Freed UMA keg (%s) was not empty (%d items). "
1693 		    " Lost %d pages of memory.\n",
1694 		    keg->uk_name ? keg->uk_name : "",
1695 		    keg->uk_free, keg->uk_pages);
1696 	}
1697 	KEG_UNLOCK(keg);
1698 
1699 	hash_free(&keg->uk_hash);
1700 
1701 	KEG_LOCK_FINI(keg);
1702 }
1703 
1704 /*
1705  * Zone header dtor.
1706  *
1707  * Arguments/Returns follow uma_dtor specifications
1708  *	udata  unused
1709  */
1710 static void
1711 zone_dtor(void *arg, int size, void *udata)
1712 {
1713 	uma_klink_t klink;
1714 	uma_zone_t zone;
1715 	uma_keg_t keg;
1716 
1717 	zone = (uma_zone_t)arg;
1718 	keg = zone_first_keg(zone);
1719 
1720 	if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
1721 		cache_drain(zone);
1722 
1723 	rw_wlock(&uma_rwlock);
1724 	LIST_REMOVE(zone, uz_link);
1725 	rw_wunlock(&uma_rwlock);
1726 	/*
1727 	 * XXX there are some races here where
1728 	 * the zone can be drained but zone lock
1729 	 * released and then refilled before we
1730 	 * remove it... we dont care for now
1731 	 */
1732 	zone_drain_wait(zone, M_WAITOK);
1733 	/*
1734 	 * Unlink all of our kegs.
1735 	 */
1736 	while ((klink = LIST_FIRST(&zone->uz_kegs)) != NULL) {
1737 		klink->kl_keg = NULL;
1738 		LIST_REMOVE(klink, kl_link);
1739 		if (klink == &zone->uz_klink)
1740 			continue;
1741 		free(klink, M_TEMP);
1742 	}
1743 	/*
1744 	 * We only destroy kegs from non secondary zones.
1745 	 */
1746 	if (keg != NULL && (zone->uz_flags & UMA_ZONE_SECONDARY) == 0)  {
1747 		rw_wlock(&uma_rwlock);
1748 		LIST_REMOVE(keg, uk_link);
1749 		rw_wunlock(&uma_rwlock);
1750 		zone_free_item(kegs, keg, NULL, SKIP_NONE);
1751 	}
1752 	ZONE_LOCK_FINI(zone);
1753 }
1754 
1755 /*
1756  * Traverses every zone in the system and calls a callback
1757  *
1758  * Arguments:
1759  *	zfunc  A pointer to a function which accepts a zone
1760  *		as an argument.
1761  *
1762  * Returns:
1763  *	Nothing
1764  */
1765 static void
1766 zone_foreach(void (*zfunc)(uma_zone_t))
1767 {
1768 	uma_keg_t keg;
1769 	uma_zone_t zone;
1770 
1771 	rw_rlock(&uma_rwlock);
1772 	LIST_FOREACH(keg, &uma_kegs, uk_link) {
1773 		LIST_FOREACH(zone, &keg->uk_zones, uz_link)
1774 			zfunc(zone);
1775 	}
1776 	rw_runlock(&uma_rwlock);
1777 }
1778 
1779 /* Public functions */
1780 /* See uma.h */
1781 void
1782 uma_startup(void *bootmem, int boot_pages)
1783 {
1784 	struct uma_zctor_args args;
1785 	uma_slab_t slab;
1786 	u_int slabsize;
1787 	int i;
1788 
1789 #ifdef UMA_DEBUG
1790 	printf("Creating uma keg headers zone and keg.\n");
1791 #endif
1792 	rw_init(&uma_rwlock, "UMA lock");
1793 
1794 	/* "manually" create the initial zone */
1795 	memset(&args, 0, sizeof(args));
1796 	args.name = "UMA Kegs";
1797 	args.size = sizeof(struct uma_keg);
1798 	args.ctor = keg_ctor;
1799 	args.dtor = keg_dtor;
1800 	args.uminit = zero_init;
1801 	args.fini = NULL;
1802 	args.keg = &masterkeg;
1803 	args.align = 32 - 1;
1804 	args.flags = UMA_ZFLAG_INTERNAL;
1805 	/* The initial zone has no Per cpu queues so it's smaller */
1806 	zone_ctor(kegs, sizeof(struct uma_zone), &args, M_WAITOK);
1807 
1808 #ifdef UMA_DEBUG
1809 	printf("Filling boot free list.\n");
1810 #endif
1811 	for (i = 0; i < boot_pages; i++) {
1812 		slab = (uma_slab_t)((uint8_t *)bootmem + (i * UMA_SLAB_SIZE));
1813 		slab->us_data = (uint8_t *)slab;
1814 		slab->us_flags = UMA_SLAB_BOOT;
1815 		LIST_INSERT_HEAD(&uma_boot_pages, slab, us_link);
1816 	}
1817 	mtx_init(&uma_boot_pages_mtx, "UMA boot pages", NULL, MTX_DEF);
1818 
1819 #ifdef UMA_DEBUG
1820 	printf("Creating uma zone headers zone and keg.\n");
1821 #endif
1822 	args.name = "UMA Zones";
1823 	args.size = sizeof(struct uma_zone) +
1824 	    (sizeof(struct uma_cache) * (mp_maxid + 1));
1825 	args.ctor = zone_ctor;
1826 	args.dtor = zone_dtor;
1827 	args.uminit = zero_init;
1828 	args.fini = NULL;
1829 	args.keg = NULL;
1830 	args.align = 32 - 1;
1831 	args.flags = UMA_ZFLAG_INTERNAL;
1832 	/* The initial zone has no Per cpu queues so it's smaller */
1833 	zone_ctor(zones, sizeof(struct uma_zone), &args, M_WAITOK);
1834 
1835 #ifdef UMA_DEBUG
1836 	printf("Creating slab and hash zones.\n");
1837 #endif
1838 
1839 	/* Now make a zone for slab headers */
1840 	slabzone = uma_zcreate("UMA Slabs",
1841 				sizeof(struct uma_slab),
1842 				NULL, NULL, NULL, NULL,
1843 				UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1844 
1845 	/*
1846 	 * We also create a zone for the bigger slabs with reference
1847 	 * counts in them, to accomodate UMA_ZONE_REFCNT zones.
1848 	 */
1849 	slabsize = sizeof(struct uma_slab_refcnt);
1850 	slabsize += uma_max_ipers_ref * sizeof(uint32_t);
1851 	slabrefzone = uma_zcreate("UMA RCntSlabs",
1852 				  slabsize,
1853 				  NULL, NULL, NULL, NULL,
1854 				  UMA_ALIGN_PTR,
1855 				  UMA_ZFLAG_INTERNAL);
1856 
1857 	hashzone = uma_zcreate("UMA Hash",
1858 	    sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
1859 	    NULL, NULL, NULL, NULL,
1860 	    UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1861 
1862 	bucket_init();
1863 
1864 	booted = UMA_STARTUP;
1865 
1866 #ifdef UMA_DEBUG
1867 	printf("UMA startup complete.\n");
1868 #endif
1869 }
1870 
1871 /* see uma.h */
1872 void
1873 uma_startup2(void)
1874 {
1875 	booted = UMA_STARTUP2;
1876 	bucket_enable();
1877 	sx_init(&uma_drain_lock, "umadrain");
1878 #ifdef UMA_DEBUG
1879 	printf("UMA startup2 complete.\n");
1880 #endif
1881 }
1882 
1883 /*
1884  * Initialize our callout handle
1885  *
1886  */
1887 
1888 static void
1889 uma_startup3(void)
1890 {
1891 #ifdef UMA_DEBUG
1892 	printf("Starting callout.\n");
1893 #endif
1894 	callout_init(&uma_callout, 1);
1895 	callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
1896 #ifdef UMA_DEBUG
1897 	printf("UMA startup3 complete.\n");
1898 #endif
1899 }
1900 
1901 static uma_keg_t
1902 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
1903 		int align, uint32_t flags)
1904 {
1905 	struct uma_kctor_args args;
1906 
1907 	args.size = size;
1908 	args.uminit = uminit;
1909 	args.fini = fini;
1910 	args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
1911 	args.flags = flags;
1912 	args.zone = zone;
1913 	return (zone_alloc_item(kegs, &args, M_WAITOK));
1914 }
1915 
1916 /* See uma.h */
1917 void
1918 uma_set_align(int align)
1919 {
1920 
1921 	if (align != UMA_ALIGN_CACHE)
1922 		uma_align_cache = align;
1923 }
1924 
1925 /* See uma.h */
1926 uma_zone_t
1927 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
1928 		uma_init uminit, uma_fini fini, int align, uint32_t flags)
1929 
1930 {
1931 	struct uma_zctor_args args;
1932 	uma_zone_t res;
1933 	bool locked;
1934 
1935 	/* This stuff is essential for the zone ctor */
1936 	memset(&args, 0, sizeof(args));
1937 	args.name = name;
1938 	args.size = size;
1939 	args.ctor = ctor;
1940 	args.dtor = dtor;
1941 	args.uminit = uminit;
1942 	args.fini = fini;
1943 #ifdef  INVARIANTS
1944 	/*
1945 	 * If a zone is being created with an empty constructor and
1946 	 * destructor, pass UMA constructor/destructor which checks for
1947 	 * memory use after free.
1948 	 */
1949 	if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOFREE))) &&
1950 	    ctor == NULL && dtor == NULL && uminit == NULL && fini == NULL) {
1951 		args.ctor = trash_ctor;
1952 		args.dtor = trash_dtor;
1953 		args.uminit = trash_init;
1954 		args.fini = trash_fini;
1955 	}
1956 #endif
1957 	args.align = align;
1958 	args.flags = flags;
1959 	args.keg = NULL;
1960 
1961 	if (booted < UMA_STARTUP2) {
1962 		locked = false;
1963 	} else {
1964 		sx_slock(&uma_drain_lock);
1965 		locked = true;
1966 	}
1967 	res = zone_alloc_item(zones, &args, M_WAITOK);
1968 	if (locked)
1969 		sx_sunlock(&uma_drain_lock);
1970 	return (res);
1971 }
1972 
1973 /* See uma.h */
1974 uma_zone_t
1975 uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor,
1976 		    uma_init zinit, uma_fini zfini, uma_zone_t master)
1977 {
1978 	struct uma_zctor_args args;
1979 	uma_keg_t keg;
1980 	uma_zone_t res;
1981 	bool locked;
1982 
1983 	keg = zone_first_keg(master);
1984 	memset(&args, 0, sizeof(args));
1985 	args.name = name;
1986 	args.size = keg->uk_size;
1987 	args.ctor = ctor;
1988 	args.dtor = dtor;
1989 	args.uminit = zinit;
1990 	args.fini = zfini;
1991 	args.align = keg->uk_align;
1992 	args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
1993 	args.keg = keg;
1994 
1995 	if (booted < UMA_STARTUP2) {
1996 		locked = false;
1997 	} else {
1998 		sx_slock(&uma_drain_lock);
1999 		locked = true;
2000 	}
2001 	/* XXX Attaches only one keg of potentially many. */
2002 	res = zone_alloc_item(zones, &args, M_WAITOK);
2003 	if (locked)
2004 		sx_sunlock(&uma_drain_lock);
2005 	return (res);
2006 }
2007 
2008 /* See uma.h */
2009 uma_zone_t
2010 uma_zcache_create(char *name, int size, uma_ctor ctor, uma_dtor dtor,
2011 		    uma_init zinit, uma_fini zfini, uma_import zimport,
2012 		    uma_release zrelease, void *arg, int flags)
2013 {
2014 	struct uma_zctor_args args;
2015 
2016 	memset(&args, 0, sizeof(args));
2017 	args.name = name;
2018 	args.size = size;
2019 	args.ctor = ctor;
2020 	args.dtor = dtor;
2021 	args.uminit = zinit;
2022 	args.fini = zfini;
2023 	args.import = zimport;
2024 	args.release = zrelease;
2025 	args.arg = arg;
2026 	args.align = 0;
2027 	args.flags = flags;
2028 
2029 	return (zone_alloc_item(zones, &args, M_WAITOK));
2030 }
2031 
2032 static void
2033 zone_lock_pair(uma_zone_t a, uma_zone_t b)
2034 {
2035 	if (a < b) {
2036 		ZONE_LOCK(a);
2037 		mtx_lock_flags(b->uz_lockptr, MTX_DUPOK);
2038 	} else {
2039 		ZONE_LOCK(b);
2040 		mtx_lock_flags(a->uz_lockptr, MTX_DUPOK);
2041 	}
2042 }
2043 
2044 static void
2045 zone_unlock_pair(uma_zone_t a, uma_zone_t b)
2046 {
2047 
2048 	ZONE_UNLOCK(a);
2049 	ZONE_UNLOCK(b);
2050 }
2051 
2052 int
2053 uma_zsecond_add(uma_zone_t zone, uma_zone_t master)
2054 {
2055 	uma_klink_t klink;
2056 	uma_klink_t kl;
2057 	int error;
2058 
2059 	error = 0;
2060 	klink = malloc(sizeof(*klink), M_TEMP, M_WAITOK | M_ZERO);
2061 
2062 	zone_lock_pair(zone, master);
2063 	/*
2064 	 * zone must use vtoslab() to resolve objects and must already be
2065 	 * a secondary.
2066 	 */
2067 	if ((zone->uz_flags & (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY))
2068 	    != (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) {
2069 		error = EINVAL;
2070 		goto out;
2071 	}
2072 	/*
2073 	 * The new master must also use vtoslab().
2074 	 */
2075 	if ((zone->uz_flags & UMA_ZONE_VTOSLAB) != UMA_ZONE_VTOSLAB) {
2076 		error = EINVAL;
2077 		goto out;
2078 	}
2079 	/*
2080 	 * Both must either be refcnt, or not be refcnt.
2081 	 */
2082 	if ((zone->uz_flags & UMA_ZONE_REFCNT) !=
2083 	    (master->uz_flags & UMA_ZONE_REFCNT)) {
2084 		error = EINVAL;
2085 		goto out;
2086 	}
2087 	/*
2088 	 * The underlying object must be the same size.  rsize
2089 	 * may be different.
2090 	 */
2091 	if (master->uz_size != zone->uz_size) {
2092 		error = E2BIG;
2093 		goto out;
2094 	}
2095 	/*
2096 	 * Put it at the end of the list.
2097 	 */
2098 	klink->kl_keg = zone_first_keg(master);
2099 	LIST_FOREACH(kl, &zone->uz_kegs, kl_link) {
2100 		if (LIST_NEXT(kl, kl_link) == NULL) {
2101 			LIST_INSERT_AFTER(kl, klink, kl_link);
2102 			break;
2103 		}
2104 	}
2105 	klink = NULL;
2106 	zone->uz_flags |= UMA_ZFLAG_MULTI;
2107 	zone->uz_slab = zone_fetch_slab_multi;
2108 
2109 out:
2110 	zone_unlock_pair(zone, master);
2111 	if (klink != NULL)
2112 		free(klink, M_TEMP);
2113 
2114 	return (error);
2115 }
2116 
2117 
2118 /* See uma.h */
2119 void
2120 uma_zdestroy(uma_zone_t zone)
2121 {
2122 
2123 	sx_slock(&uma_drain_lock);
2124 	zone_free_item(zones, zone, NULL, SKIP_NONE);
2125 	sx_sunlock(&uma_drain_lock);
2126 }
2127 
2128 /* See uma.h */
2129 void *
2130 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
2131 {
2132 	void *item;
2133 	uma_cache_t cache;
2134 	uma_bucket_t bucket;
2135 	int lockfail;
2136 	int cpu;
2137 
2138 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
2139 	random_harvest_fast_uma(&zone, sizeof(zone), 1, RANDOM_UMA);
2140 
2141 	/* This is the fast path allocation */
2142 #ifdef UMA_DEBUG_ALLOC_1
2143 	printf("Allocating one item from %s(%p)\n", zone->uz_name, zone);
2144 #endif
2145 	CTR3(KTR_UMA, "uma_zalloc_arg thread %x zone %s flags %d", curthread,
2146 	    zone->uz_name, flags);
2147 
2148 	if (flags & M_WAITOK) {
2149 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
2150 		    "uma_zalloc_arg: zone \"%s\"", zone->uz_name);
2151 	}
2152 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
2153 	    ("uma_zalloc_arg: called with spinlock or critical section held"));
2154 
2155 #ifdef DEBUG_MEMGUARD
2156 	if (memguard_cmp_zone(zone)) {
2157 		item = memguard_alloc(zone->uz_size, flags);
2158 		if (item != NULL) {
2159 			/*
2160 			 * Avoid conflict with the use-after-free
2161 			 * protecting infrastructure from INVARIANTS.
2162 			 */
2163 			if (zone->uz_init != NULL &&
2164 			    zone->uz_init != mtrash_init &&
2165 			    zone->uz_init(item, zone->uz_size, flags) != 0)
2166 				return (NULL);
2167 			if (zone->uz_ctor != NULL &&
2168 			    zone->uz_ctor != mtrash_ctor &&
2169 			    zone->uz_ctor(item, zone->uz_size, udata,
2170 			    flags) != 0) {
2171 			    	zone->uz_fini(item, zone->uz_size);
2172 				return (NULL);
2173 			}
2174 			return (item);
2175 		}
2176 		/* This is unfortunate but should not be fatal. */
2177 	}
2178 #endif
2179 	/*
2180 	 * If possible, allocate from the per-CPU cache.  There are two
2181 	 * requirements for safe access to the per-CPU cache: (1) the thread
2182 	 * accessing the cache must not be preempted or yield during access,
2183 	 * and (2) the thread must not migrate CPUs without switching which
2184 	 * cache it accesses.  We rely on a critical section to prevent
2185 	 * preemption and migration.  We release the critical section in
2186 	 * order to acquire the zone mutex if we are unable to allocate from
2187 	 * the current cache; when we re-acquire the critical section, we
2188 	 * must detect and handle migration if it has occurred.
2189 	 */
2190 	critical_enter();
2191 	cpu = curcpu;
2192 	cache = &zone->uz_cpu[cpu];
2193 
2194 zalloc_start:
2195 	bucket = cache->uc_allocbucket;
2196 	if (bucket != NULL && bucket->ub_cnt > 0) {
2197 		bucket->ub_cnt--;
2198 		item = bucket->ub_bucket[bucket->ub_cnt];
2199 #ifdef INVARIANTS
2200 		bucket->ub_bucket[bucket->ub_cnt] = NULL;
2201 #endif
2202 		KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled."));
2203 		cache->uc_allocs++;
2204 		critical_exit();
2205 		if (zone->uz_ctor != NULL &&
2206 		    zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2207 			atomic_add_long(&zone->uz_fails, 1);
2208 			zone_free_item(zone, item, udata, SKIP_DTOR);
2209 			return (NULL);
2210 		}
2211 #ifdef INVARIANTS
2212 		uma_dbg_alloc(zone, NULL, item);
2213 #endif
2214 		if (flags & M_ZERO)
2215 			uma_zero_item(item, zone);
2216 		return (item);
2217 	}
2218 
2219 	/*
2220 	 * We have run out of items in our alloc bucket.
2221 	 * See if we can switch with our free bucket.
2222 	 */
2223 	bucket = cache->uc_freebucket;
2224 	if (bucket != NULL && bucket->ub_cnt > 0) {
2225 #ifdef UMA_DEBUG_ALLOC
2226 		printf("uma_zalloc: Swapping empty with alloc.\n");
2227 #endif
2228 		cache->uc_freebucket = cache->uc_allocbucket;
2229 		cache->uc_allocbucket = bucket;
2230 		goto zalloc_start;
2231 	}
2232 
2233 	/*
2234 	 * Discard any empty allocation bucket while we hold no locks.
2235 	 */
2236 	bucket = cache->uc_allocbucket;
2237 	cache->uc_allocbucket = NULL;
2238 	critical_exit();
2239 	if (bucket != NULL)
2240 		bucket_free(zone, bucket, udata);
2241 
2242 	/* Short-circuit for zones without buckets and low memory. */
2243 	if (zone->uz_count == 0 || bucketdisable)
2244 		goto zalloc_item;
2245 
2246 	/*
2247 	 * Attempt to retrieve the item from the per-CPU cache has failed, so
2248 	 * we must go back to the zone.  This requires the zone lock, so we
2249 	 * must drop the critical section, then re-acquire it when we go back
2250 	 * to the cache.  Since the critical section is released, we may be
2251 	 * preempted or migrate.  As such, make sure not to maintain any
2252 	 * thread-local state specific to the cache from prior to releasing
2253 	 * the critical section.
2254 	 */
2255 	lockfail = 0;
2256 	if (ZONE_TRYLOCK(zone) == 0) {
2257 		/* Record contention to size the buckets. */
2258 		ZONE_LOCK(zone);
2259 		lockfail = 1;
2260 	}
2261 	critical_enter();
2262 	cpu = curcpu;
2263 	cache = &zone->uz_cpu[cpu];
2264 
2265 	/*
2266 	 * Since we have locked the zone we may as well send back our stats.
2267 	 */
2268 	atomic_add_long(&zone->uz_allocs, cache->uc_allocs);
2269 	atomic_add_long(&zone->uz_frees, cache->uc_frees);
2270 	cache->uc_allocs = 0;
2271 	cache->uc_frees = 0;
2272 
2273 	/* See if we lost the race to fill the cache. */
2274 	if (cache->uc_allocbucket != NULL) {
2275 		ZONE_UNLOCK(zone);
2276 		goto zalloc_start;
2277 	}
2278 
2279 	/*
2280 	 * Check the zone's cache of buckets.
2281 	 */
2282 	if ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) {
2283 		KASSERT(bucket->ub_cnt != 0,
2284 		    ("uma_zalloc_arg: Returning an empty bucket."));
2285 
2286 		LIST_REMOVE(bucket, ub_link);
2287 		cache->uc_allocbucket = bucket;
2288 		ZONE_UNLOCK(zone);
2289 		goto zalloc_start;
2290 	}
2291 	/* We are no longer associated with this CPU. */
2292 	critical_exit();
2293 
2294 	/*
2295 	 * We bump the uz count when the cache size is insufficient to
2296 	 * handle the working set.
2297 	 */
2298 	if (lockfail && zone->uz_count < BUCKET_MAX)
2299 		zone->uz_count++;
2300 	ZONE_UNLOCK(zone);
2301 
2302 	/*
2303 	 * Now lets just fill a bucket and put it on the free list.  If that
2304 	 * works we'll restart the allocation from the begining and it
2305 	 * will use the just filled bucket.
2306 	 */
2307 	bucket = zone_alloc_bucket(zone, udata, flags);
2308 	if (bucket != NULL) {
2309 		ZONE_LOCK(zone);
2310 		critical_enter();
2311 		cpu = curcpu;
2312 		cache = &zone->uz_cpu[cpu];
2313 		/*
2314 		 * See if we lost the race or were migrated.  Cache the
2315 		 * initialized bucket to make this less likely or claim
2316 		 * the memory directly.
2317 		 */
2318 		if (cache->uc_allocbucket == NULL)
2319 			cache->uc_allocbucket = bucket;
2320 		else
2321 			LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link);
2322 		ZONE_UNLOCK(zone);
2323 		goto zalloc_start;
2324 	}
2325 
2326 	/*
2327 	 * We may not be able to get a bucket so return an actual item.
2328 	 */
2329 #ifdef UMA_DEBUG
2330 	printf("uma_zalloc_arg: Bucketzone returned NULL\n");
2331 #endif
2332 
2333 zalloc_item:
2334 	item = zone_alloc_item(zone, udata, flags);
2335 
2336 	return (item);
2337 }
2338 
2339 static uma_slab_t
2340 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int flags)
2341 {
2342 	uma_slab_t slab;
2343 	int reserve;
2344 
2345 	mtx_assert(&keg->uk_lock, MA_OWNED);
2346 	slab = NULL;
2347 	reserve = 0;
2348 	if ((flags & M_USE_RESERVE) == 0)
2349 		reserve = keg->uk_reserve;
2350 
2351 	for (;;) {
2352 		/*
2353 		 * Find a slab with some space.  Prefer slabs that are partially
2354 		 * used over those that are totally full.  This helps to reduce
2355 		 * fragmentation.
2356 		 */
2357 		if (keg->uk_free > reserve) {
2358 			if (!LIST_EMPTY(&keg->uk_part_slab)) {
2359 				slab = LIST_FIRST(&keg->uk_part_slab);
2360 			} else {
2361 				slab = LIST_FIRST(&keg->uk_free_slab);
2362 				LIST_REMOVE(slab, us_link);
2363 				LIST_INSERT_HEAD(&keg->uk_part_slab, slab,
2364 				    us_link);
2365 			}
2366 			MPASS(slab->us_keg == keg);
2367 			return (slab);
2368 		}
2369 
2370 		/*
2371 		 * M_NOVM means don't ask at all!
2372 		 */
2373 		if (flags & M_NOVM)
2374 			break;
2375 
2376 		if (keg->uk_maxpages && keg->uk_pages >= keg->uk_maxpages) {
2377 			keg->uk_flags |= UMA_ZFLAG_FULL;
2378 			/*
2379 			 * If this is not a multi-zone, set the FULL bit.
2380 			 * Otherwise slab_multi() takes care of it.
2381 			 */
2382 			if ((zone->uz_flags & UMA_ZFLAG_MULTI) == 0) {
2383 				zone->uz_flags |= UMA_ZFLAG_FULL;
2384 				zone_log_warning(zone);
2385 			}
2386 			if (flags & M_NOWAIT)
2387 				break;
2388 			zone->uz_sleeps++;
2389 			msleep(keg, &keg->uk_lock, PVM, "keglimit", 0);
2390 			continue;
2391 		}
2392 		slab = keg_alloc_slab(keg, zone, flags);
2393 		/*
2394 		 * If we got a slab here it's safe to mark it partially used
2395 		 * and return.  We assume that the caller is going to remove
2396 		 * at least one item.
2397 		 */
2398 		if (slab) {
2399 			MPASS(slab->us_keg == keg);
2400 			LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2401 			return (slab);
2402 		}
2403 		/*
2404 		 * We might not have been able to get a slab but another cpu
2405 		 * could have while we were unlocked.  Check again before we
2406 		 * fail.
2407 		 */
2408 		flags |= M_NOVM;
2409 	}
2410 	return (slab);
2411 }
2412 
2413 static uma_slab_t
2414 zone_fetch_slab(uma_zone_t zone, uma_keg_t keg, int flags)
2415 {
2416 	uma_slab_t slab;
2417 
2418 	if (keg == NULL) {
2419 		keg = zone_first_keg(zone);
2420 		KEG_LOCK(keg);
2421 	}
2422 
2423 	for (;;) {
2424 		slab = keg_fetch_slab(keg, zone, flags);
2425 		if (slab)
2426 			return (slab);
2427 		if (flags & (M_NOWAIT | M_NOVM))
2428 			break;
2429 	}
2430 	KEG_UNLOCK(keg);
2431 	return (NULL);
2432 }
2433 
2434 /*
2435  * uma_zone_fetch_slab_multi:  Fetches a slab from one available keg.  Returns
2436  * with the keg locked.  On NULL no lock is held.
2437  *
2438  * The last pointer is used to seed the search.  It is not required.
2439  */
2440 static uma_slab_t
2441 zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int rflags)
2442 {
2443 	uma_klink_t klink;
2444 	uma_slab_t slab;
2445 	uma_keg_t keg;
2446 	int flags;
2447 	int empty;
2448 	int full;
2449 
2450 	/*
2451 	 * Don't wait on the first pass.  This will skip limit tests
2452 	 * as well.  We don't want to block if we can find a provider
2453 	 * without blocking.
2454 	 */
2455 	flags = (rflags & ~M_WAITOK) | M_NOWAIT;
2456 	/*
2457 	 * Use the last slab allocated as a hint for where to start
2458 	 * the search.
2459 	 */
2460 	if (last != NULL) {
2461 		slab = keg_fetch_slab(last, zone, flags);
2462 		if (slab)
2463 			return (slab);
2464 		KEG_UNLOCK(last);
2465 	}
2466 	/*
2467 	 * Loop until we have a slab incase of transient failures
2468 	 * while M_WAITOK is specified.  I'm not sure this is 100%
2469 	 * required but we've done it for so long now.
2470 	 */
2471 	for (;;) {
2472 		empty = 0;
2473 		full = 0;
2474 		/*
2475 		 * Search the available kegs for slabs.  Be careful to hold the
2476 		 * correct lock while calling into the keg layer.
2477 		 */
2478 		LIST_FOREACH(klink, &zone->uz_kegs, kl_link) {
2479 			keg = klink->kl_keg;
2480 			KEG_LOCK(keg);
2481 			if ((keg->uk_flags & UMA_ZFLAG_FULL) == 0) {
2482 				slab = keg_fetch_slab(keg, zone, flags);
2483 				if (slab)
2484 					return (slab);
2485 			}
2486 			if (keg->uk_flags & UMA_ZFLAG_FULL)
2487 				full++;
2488 			else
2489 				empty++;
2490 			KEG_UNLOCK(keg);
2491 		}
2492 		if (rflags & (M_NOWAIT | M_NOVM))
2493 			break;
2494 		flags = rflags;
2495 		/*
2496 		 * All kegs are full.  XXX We can't atomically check all kegs
2497 		 * and sleep so just sleep for a short period and retry.
2498 		 */
2499 		if (full && !empty) {
2500 			ZONE_LOCK(zone);
2501 			zone->uz_flags |= UMA_ZFLAG_FULL;
2502 			zone->uz_sleeps++;
2503 			zone_log_warning(zone);
2504 			msleep(zone, zone->uz_lockptr, PVM,
2505 			    "zonelimit", hz/100);
2506 			zone->uz_flags &= ~UMA_ZFLAG_FULL;
2507 			ZONE_UNLOCK(zone);
2508 			continue;
2509 		}
2510 	}
2511 	return (NULL);
2512 }
2513 
2514 static void *
2515 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
2516 {
2517 	void *item;
2518 	uint8_t freei;
2519 
2520 	MPASS(keg == slab->us_keg);
2521 	mtx_assert(&keg->uk_lock, MA_OWNED);
2522 
2523 	freei = BIT_FFS(SLAB_SETSIZE, &slab->us_free) - 1;
2524 	BIT_CLR(SLAB_SETSIZE, freei, &slab->us_free);
2525 	item = slab->us_data + (keg->uk_rsize * freei);
2526 	slab->us_freecount--;
2527 	keg->uk_free--;
2528 
2529 	/* Move this slab to the full list */
2530 	if (slab->us_freecount == 0) {
2531 		LIST_REMOVE(slab, us_link);
2532 		LIST_INSERT_HEAD(&keg->uk_full_slab, slab, us_link);
2533 	}
2534 
2535 	return (item);
2536 }
2537 
2538 static int
2539 zone_import(uma_zone_t zone, void **bucket, int max, int flags)
2540 {
2541 	uma_slab_t slab;
2542 	uma_keg_t keg;
2543 	int i;
2544 
2545 	slab = NULL;
2546 	keg = NULL;
2547 	/* Try to keep the buckets totally full */
2548 	for (i = 0; i < max; ) {
2549 		if ((slab = zone->uz_slab(zone, keg, flags)) == NULL)
2550 			break;
2551 		keg = slab->us_keg;
2552 		while (slab->us_freecount && i < max) {
2553 			bucket[i++] = slab_alloc_item(keg, slab);
2554 			if (keg->uk_free <= keg->uk_reserve)
2555 				break;
2556 		}
2557 		/* Don't grab more than one slab at a time. */
2558 		flags &= ~M_WAITOK;
2559 		flags |= M_NOWAIT;
2560 	}
2561 	if (slab != NULL)
2562 		KEG_UNLOCK(keg);
2563 
2564 	return i;
2565 }
2566 
2567 static uma_bucket_t
2568 zone_alloc_bucket(uma_zone_t zone, void *udata, int flags)
2569 {
2570 	uma_bucket_t bucket;
2571 	int max;
2572 
2573 	/* Don't wait for buckets, preserve caller's NOVM setting. */
2574 	bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
2575 	if (bucket == NULL)
2576 		return (NULL);
2577 
2578 	max = MIN(bucket->ub_entries, zone->uz_count);
2579 	bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
2580 	    max, flags);
2581 
2582 	/*
2583 	 * Initialize the memory if necessary.
2584 	 */
2585 	if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
2586 		int i;
2587 
2588 		for (i = 0; i < bucket->ub_cnt; i++)
2589 			if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size,
2590 			    flags) != 0)
2591 				break;
2592 		/*
2593 		 * If we couldn't initialize the whole bucket, put the
2594 		 * rest back onto the freelist.
2595 		 */
2596 		if (i != bucket->ub_cnt) {
2597 			zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
2598 			    bucket->ub_cnt - i);
2599 #ifdef INVARIANTS
2600 			bzero(&bucket->ub_bucket[i],
2601 			    sizeof(void *) * (bucket->ub_cnt - i));
2602 #endif
2603 			bucket->ub_cnt = i;
2604 		}
2605 	}
2606 
2607 	if (bucket->ub_cnt == 0) {
2608 		bucket_free(zone, bucket, udata);
2609 		atomic_add_long(&zone->uz_fails, 1);
2610 		return (NULL);
2611 	}
2612 
2613 	return (bucket);
2614 }
2615 
2616 /*
2617  * Allocates a single item from a zone.
2618  *
2619  * Arguments
2620  *	zone   The zone to alloc for.
2621  *	udata  The data to be passed to the constructor.
2622  *	flags  M_WAITOK, M_NOWAIT, M_ZERO.
2623  *
2624  * Returns
2625  *	NULL if there is no memory and M_NOWAIT is set
2626  *	An item if successful
2627  */
2628 
2629 static void *
2630 zone_alloc_item(uma_zone_t zone, void *udata, int flags)
2631 {
2632 	void *item;
2633 
2634 	item = NULL;
2635 
2636 #ifdef UMA_DEBUG_ALLOC
2637 	printf("INTERNAL: Allocating one item from %s(%p)\n", zone->uz_name, zone);
2638 #endif
2639 	if (zone->uz_import(zone->uz_arg, &item, 1, flags) != 1)
2640 		goto fail;
2641 	atomic_add_long(&zone->uz_allocs, 1);
2642 
2643 	/*
2644 	 * We have to call both the zone's init (not the keg's init)
2645 	 * and the zone's ctor.  This is because the item is going from
2646 	 * a keg slab directly to the user, and the user is expecting it
2647 	 * to be both zone-init'd as well as zone-ctor'd.
2648 	 */
2649 	if (zone->uz_init != NULL) {
2650 		if (zone->uz_init(item, zone->uz_size, flags) != 0) {
2651 			zone_free_item(zone, item, udata, SKIP_FINI);
2652 			goto fail;
2653 		}
2654 	}
2655 	if (zone->uz_ctor != NULL) {
2656 		if (zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2657 			zone_free_item(zone, item, udata, SKIP_DTOR);
2658 			goto fail;
2659 		}
2660 	}
2661 #ifdef INVARIANTS
2662 	uma_dbg_alloc(zone, NULL, item);
2663 #endif
2664 	if (flags & M_ZERO)
2665 		uma_zero_item(item, zone);
2666 
2667 	return (item);
2668 
2669 fail:
2670 	atomic_add_long(&zone->uz_fails, 1);
2671 	return (NULL);
2672 }
2673 
2674 /* See uma.h */
2675 void
2676 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
2677 {
2678 	uma_cache_t cache;
2679 	uma_bucket_t bucket;
2680 	int lockfail;
2681 	int cpu;
2682 
2683 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
2684 	random_harvest_fast_uma(&zone, sizeof(zone), 1, RANDOM_UMA);
2685 
2686 #ifdef UMA_DEBUG_ALLOC_1
2687 	printf("Freeing item %p to %s(%p)\n", item, zone->uz_name, zone);
2688 #endif
2689 	CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread,
2690 	    zone->uz_name);
2691 
2692 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
2693 	    ("uma_zfree_arg: called with spinlock or critical section held"));
2694 
2695         /* uma_zfree(..., NULL) does nothing, to match free(9). */
2696         if (item == NULL)
2697                 return;
2698 #ifdef DEBUG_MEMGUARD
2699 	if (is_memguard_addr(item)) {
2700 		if (zone->uz_dtor != NULL && zone->uz_dtor != mtrash_dtor)
2701 			zone->uz_dtor(item, zone->uz_size, udata);
2702 		if (zone->uz_fini != NULL && zone->uz_fini != mtrash_fini)
2703 			zone->uz_fini(item, zone->uz_size);
2704 		memguard_free(item);
2705 		return;
2706 	}
2707 #endif
2708 #ifdef INVARIANTS
2709 	if (zone->uz_flags & UMA_ZONE_MALLOC)
2710 		uma_dbg_free(zone, udata, item);
2711 	else
2712 		uma_dbg_free(zone, NULL, item);
2713 #endif
2714 	if (zone->uz_dtor != NULL)
2715 		zone->uz_dtor(item, zone->uz_size, udata);
2716 
2717 	/*
2718 	 * The race here is acceptable.  If we miss it we'll just have to wait
2719 	 * a little longer for the limits to be reset.
2720 	 */
2721 	if (zone->uz_flags & UMA_ZFLAG_FULL)
2722 		goto zfree_item;
2723 
2724 	/*
2725 	 * If possible, free to the per-CPU cache.  There are two
2726 	 * requirements for safe access to the per-CPU cache: (1) the thread
2727 	 * accessing the cache must not be preempted or yield during access,
2728 	 * and (2) the thread must not migrate CPUs without switching which
2729 	 * cache it accesses.  We rely on a critical section to prevent
2730 	 * preemption and migration.  We release the critical section in
2731 	 * order to acquire the zone mutex if we are unable to free to the
2732 	 * current cache; when we re-acquire the critical section, we must
2733 	 * detect and handle migration if it has occurred.
2734 	 */
2735 zfree_restart:
2736 	critical_enter();
2737 	cpu = curcpu;
2738 	cache = &zone->uz_cpu[cpu];
2739 
2740 zfree_start:
2741 	/*
2742 	 * Try to free into the allocbucket first to give LIFO ordering
2743 	 * for cache-hot datastructures.  Spill over into the freebucket
2744 	 * if necessary.  Alloc will swap them if one runs dry.
2745 	 */
2746 	bucket = cache->uc_allocbucket;
2747 	if (bucket == NULL || bucket->ub_cnt >= bucket->ub_entries)
2748 		bucket = cache->uc_freebucket;
2749 	if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) {
2750 		KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL,
2751 		    ("uma_zfree: Freeing to non free bucket index."));
2752 		bucket->ub_bucket[bucket->ub_cnt] = item;
2753 		bucket->ub_cnt++;
2754 		cache->uc_frees++;
2755 		critical_exit();
2756 		return;
2757 	}
2758 
2759 	/*
2760 	 * We must go back the zone, which requires acquiring the zone lock,
2761 	 * which in turn means we must release and re-acquire the critical
2762 	 * section.  Since the critical section is released, we may be
2763 	 * preempted or migrate.  As such, make sure not to maintain any
2764 	 * thread-local state specific to the cache from prior to releasing
2765 	 * the critical section.
2766 	 */
2767 	critical_exit();
2768 	if (zone->uz_count == 0 || bucketdisable)
2769 		goto zfree_item;
2770 
2771 	lockfail = 0;
2772 	if (ZONE_TRYLOCK(zone) == 0) {
2773 		/* Record contention to size the buckets. */
2774 		ZONE_LOCK(zone);
2775 		lockfail = 1;
2776 	}
2777 	critical_enter();
2778 	cpu = curcpu;
2779 	cache = &zone->uz_cpu[cpu];
2780 
2781 	/*
2782 	 * Since we have locked the zone we may as well send back our stats.
2783 	 */
2784 	atomic_add_long(&zone->uz_allocs, cache->uc_allocs);
2785 	atomic_add_long(&zone->uz_frees, cache->uc_frees);
2786 	cache->uc_allocs = 0;
2787 	cache->uc_frees = 0;
2788 
2789 	bucket = cache->uc_freebucket;
2790 	if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) {
2791 		ZONE_UNLOCK(zone);
2792 		goto zfree_start;
2793 	}
2794 	cache->uc_freebucket = NULL;
2795 
2796 	/* Can we throw this on the zone full list? */
2797 	if (bucket != NULL) {
2798 #ifdef UMA_DEBUG_ALLOC
2799 		printf("uma_zfree: Putting old bucket on the free list.\n");
2800 #endif
2801 		/* ub_cnt is pointing to the last free item */
2802 		KASSERT(bucket->ub_cnt != 0,
2803 		    ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n"));
2804 		LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link);
2805 	}
2806 
2807 	/* We are no longer associated with this CPU. */
2808 	critical_exit();
2809 
2810 	/*
2811 	 * We bump the uz count when the cache size is insufficient to
2812 	 * handle the working set.
2813 	 */
2814 	if (lockfail && zone->uz_count < BUCKET_MAX)
2815 		zone->uz_count++;
2816 	ZONE_UNLOCK(zone);
2817 
2818 #ifdef UMA_DEBUG_ALLOC
2819 	printf("uma_zfree: Allocating new free bucket.\n");
2820 #endif
2821 	bucket = bucket_alloc(zone, udata, M_NOWAIT);
2822 	if (bucket) {
2823 		critical_enter();
2824 		cpu = curcpu;
2825 		cache = &zone->uz_cpu[cpu];
2826 		if (cache->uc_freebucket == NULL) {
2827 			cache->uc_freebucket = bucket;
2828 			goto zfree_start;
2829 		}
2830 		/*
2831 		 * We lost the race, start over.  We have to drop our
2832 		 * critical section to free the bucket.
2833 		 */
2834 		critical_exit();
2835 		bucket_free(zone, bucket, udata);
2836 		goto zfree_restart;
2837 	}
2838 
2839 	/*
2840 	 * If nothing else caught this, we'll just do an internal free.
2841 	 */
2842 zfree_item:
2843 	zone_free_item(zone, item, udata, SKIP_DTOR);
2844 
2845 	return;
2846 }
2847 
2848 static void
2849 slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item)
2850 {
2851 	uint8_t freei;
2852 
2853 	mtx_assert(&keg->uk_lock, MA_OWNED);
2854 	MPASS(keg == slab->us_keg);
2855 
2856 	/* Do we need to remove from any lists? */
2857 	if (slab->us_freecount+1 == keg->uk_ipers) {
2858 		LIST_REMOVE(slab, us_link);
2859 		LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
2860 	} else if (slab->us_freecount == 0) {
2861 		LIST_REMOVE(slab, us_link);
2862 		LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2863 	}
2864 
2865 	/* Slab management. */
2866 	freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
2867 	BIT_SET(SLAB_SETSIZE, freei, &slab->us_free);
2868 	slab->us_freecount++;
2869 
2870 	/* Keg statistics. */
2871 	keg->uk_free++;
2872 }
2873 
2874 static void
2875 zone_release(uma_zone_t zone, void **bucket, int cnt)
2876 {
2877 	void *item;
2878 	uma_slab_t slab;
2879 	uma_keg_t keg;
2880 	uint8_t *mem;
2881 	int clearfull;
2882 	int i;
2883 
2884 	clearfull = 0;
2885 	keg = zone_first_keg(zone);
2886 	KEG_LOCK(keg);
2887 	for (i = 0; i < cnt; i++) {
2888 		item = bucket[i];
2889 		if (!(zone->uz_flags & UMA_ZONE_VTOSLAB)) {
2890 			mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
2891 			if (zone->uz_flags & UMA_ZONE_HASH) {
2892 				slab = hash_sfind(&keg->uk_hash, mem);
2893 			} else {
2894 				mem += keg->uk_pgoff;
2895 				slab = (uma_slab_t)mem;
2896 			}
2897 		} else {
2898 			slab = vtoslab((vm_offset_t)item);
2899 			if (slab->us_keg != keg) {
2900 				KEG_UNLOCK(keg);
2901 				keg = slab->us_keg;
2902 				KEG_LOCK(keg);
2903 			}
2904 		}
2905 		slab_free_item(keg, slab, item);
2906 		if (keg->uk_flags & UMA_ZFLAG_FULL) {
2907 			if (keg->uk_pages < keg->uk_maxpages) {
2908 				keg->uk_flags &= ~UMA_ZFLAG_FULL;
2909 				clearfull = 1;
2910 			}
2911 
2912 			/*
2913 			 * We can handle one more allocation. Since we're
2914 			 * clearing ZFLAG_FULL, wake up all procs blocked
2915 			 * on pages. This should be uncommon, so keeping this
2916 			 * simple for now (rather than adding count of blocked
2917 			 * threads etc).
2918 			 */
2919 			wakeup(keg);
2920 		}
2921 	}
2922 	KEG_UNLOCK(keg);
2923 	if (clearfull) {
2924 		ZONE_LOCK(zone);
2925 		zone->uz_flags &= ~UMA_ZFLAG_FULL;
2926 		wakeup(zone);
2927 		ZONE_UNLOCK(zone);
2928 	}
2929 
2930 }
2931 
2932 /*
2933  * Frees a single item to any zone.
2934  *
2935  * Arguments:
2936  *	zone   The zone to free to
2937  *	item   The item we're freeing
2938  *	udata  User supplied data for the dtor
2939  *	skip   Skip dtors and finis
2940  */
2941 static void
2942 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
2943 {
2944 
2945 #ifdef INVARIANTS
2946 	if (skip == SKIP_NONE) {
2947 		if (zone->uz_flags & UMA_ZONE_MALLOC)
2948 			uma_dbg_free(zone, udata, item);
2949 		else
2950 			uma_dbg_free(zone, NULL, item);
2951 	}
2952 #endif
2953 	if (skip < SKIP_DTOR && zone->uz_dtor)
2954 		zone->uz_dtor(item, zone->uz_size, udata);
2955 
2956 	if (skip < SKIP_FINI && zone->uz_fini)
2957 		zone->uz_fini(item, zone->uz_size);
2958 
2959 	atomic_add_long(&zone->uz_frees, 1);
2960 	zone->uz_release(zone->uz_arg, &item, 1);
2961 }
2962 
2963 /* See uma.h */
2964 int
2965 uma_zone_set_max(uma_zone_t zone, int nitems)
2966 {
2967 	uma_keg_t keg;
2968 
2969 	keg = zone_first_keg(zone);
2970 	if (keg == NULL)
2971 		return (0);
2972 	KEG_LOCK(keg);
2973 	keg->uk_maxpages = (nitems / keg->uk_ipers) * keg->uk_ppera;
2974 	if (keg->uk_maxpages * keg->uk_ipers < nitems)
2975 		keg->uk_maxpages += keg->uk_ppera;
2976 	nitems = keg->uk_maxpages * keg->uk_ipers;
2977 	KEG_UNLOCK(keg);
2978 
2979 	return (nitems);
2980 }
2981 
2982 /* See uma.h */
2983 int
2984 uma_zone_get_max(uma_zone_t zone)
2985 {
2986 	int nitems;
2987 	uma_keg_t keg;
2988 
2989 	keg = zone_first_keg(zone);
2990 	if (keg == NULL)
2991 		return (0);
2992 	KEG_LOCK(keg);
2993 	nitems = keg->uk_maxpages * keg->uk_ipers;
2994 	KEG_UNLOCK(keg);
2995 
2996 	return (nitems);
2997 }
2998 
2999 /* See uma.h */
3000 void
3001 uma_zone_set_warning(uma_zone_t zone, const char *warning)
3002 {
3003 
3004 	ZONE_LOCK(zone);
3005 	zone->uz_warning = warning;
3006 	ZONE_UNLOCK(zone);
3007 }
3008 
3009 /* See uma.h */
3010 int
3011 uma_zone_get_cur(uma_zone_t zone)
3012 {
3013 	int64_t nitems;
3014 	u_int i;
3015 
3016 	ZONE_LOCK(zone);
3017 	nitems = zone->uz_allocs - zone->uz_frees;
3018 	CPU_FOREACH(i) {
3019 		/*
3020 		 * See the comment in sysctl_vm_zone_stats() regarding the
3021 		 * safety of accessing the per-cpu caches. With the zone lock
3022 		 * held, it is safe, but can potentially result in stale data.
3023 		 */
3024 		nitems += zone->uz_cpu[i].uc_allocs -
3025 		    zone->uz_cpu[i].uc_frees;
3026 	}
3027 	ZONE_UNLOCK(zone);
3028 
3029 	return (nitems < 0 ? 0 : nitems);
3030 }
3031 
3032 /* See uma.h */
3033 void
3034 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
3035 {
3036 	uma_keg_t keg;
3037 
3038 	keg = zone_first_keg(zone);
3039 	KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type"));
3040 	KEG_LOCK(keg);
3041 	KASSERT(keg->uk_pages == 0,
3042 	    ("uma_zone_set_init on non-empty keg"));
3043 	keg->uk_init = uminit;
3044 	KEG_UNLOCK(keg);
3045 }
3046 
3047 /* See uma.h */
3048 void
3049 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
3050 {
3051 	uma_keg_t keg;
3052 
3053 	keg = zone_first_keg(zone);
3054 	KASSERT(keg != NULL, ("uma_zone_set_fini: Invalid zone type"));
3055 	KEG_LOCK(keg);
3056 	KASSERT(keg->uk_pages == 0,
3057 	    ("uma_zone_set_fini on non-empty keg"));
3058 	keg->uk_fini = fini;
3059 	KEG_UNLOCK(keg);
3060 }
3061 
3062 /* See uma.h */
3063 void
3064 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
3065 {
3066 
3067 	ZONE_LOCK(zone);
3068 	KASSERT(zone_first_keg(zone)->uk_pages == 0,
3069 	    ("uma_zone_set_zinit on non-empty keg"));
3070 	zone->uz_init = zinit;
3071 	ZONE_UNLOCK(zone);
3072 }
3073 
3074 /* See uma.h */
3075 void
3076 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
3077 {
3078 
3079 	ZONE_LOCK(zone);
3080 	KASSERT(zone_first_keg(zone)->uk_pages == 0,
3081 	    ("uma_zone_set_zfini on non-empty keg"));
3082 	zone->uz_fini = zfini;
3083 	ZONE_UNLOCK(zone);
3084 }
3085 
3086 /* See uma.h */
3087 /* XXX uk_freef is not actually used with the zone locked */
3088 void
3089 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
3090 {
3091 	uma_keg_t keg;
3092 
3093 	keg = zone_first_keg(zone);
3094 	KASSERT(keg != NULL, ("uma_zone_set_freef: Invalid zone type"));
3095 	KEG_LOCK(keg);
3096 	keg->uk_freef = freef;
3097 	KEG_UNLOCK(keg);
3098 }
3099 
3100 /* See uma.h */
3101 /* XXX uk_allocf is not actually used with the zone locked */
3102 void
3103 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
3104 {
3105 	uma_keg_t keg;
3106 
3107 	keg = zone_first_keg(zone);
3108 	KEG_LOCK(keg);
3109 	keg->uk_allocf = allocf;
3110 	KEG_UNLOCK(keg);
3111 }
3112 
3113 /* See uma.h */
3114 void
3115 uma_zone_reserve(uma_zone_t zone, int items)
3116 {
3117 	uma_keg_t keg;
3118 
3119 	keg = zone_first_keg(zone);
3120 	if (keg == NULL)
3121 		return;
3122 	KEG_LOCK(keg);
3123 	keg->uk_reserve = items;
3124 	KEG_UNLOCK(keg);
3125 
3126 	return;
3127 }
3128 
3129 /* See uma.h */
3130 int
3131 uma_zone_reserve_kva(uma_zone_t zone, int count)
3132 {
3133 	uma_keg_t keg;
3134 	vm_offset_t kva;
3135 	u_int pages;
3136 
3137 	keg = zone_first_keg(zone);
3138 	if (keg == NULL)
3139 		return (0);
3140 	pages = count / keg->uk_ipers;
3141 
3142 	if (pages * keg->uk_ipers < count)
3143 		pages++;
3144 
3145 #ifdef UMA_MD_SMALL_ALLOC
3146 	if (keg->uk_ppera > 1) {
3147 #else
3148 	if (1) {
3149 #endif
3150 		kva = kva_alloc((vm_size_t)pages * UMA_SLAB_SIZE);
3151 		if (kva == 0)
3152 			return (0);
3153 	} else
3154 		kva = 0;
3155 	KEG_LOCK(keg);
3156 	keg->uk_kva = kva;
3157 	keg->uk_offset = 0;
3158 	keg->uk_maxpages = pages;
3159 #ifdef UMA_MD_SMALL_ALLOC
3160 	keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
3161 #else
3162 	keg->uk_allocf = noobj_alloc;
3163 #endif
3164 	keg->uk_flags |= UMA_ZONE_NOFREE;
3165 	KEG_UNLOCK(keg);
3166 
3167 	return (1);
3168 }
3169 
3170 /* See uma.h */
3171 void
3172 uma_prealloc(uma_zone_t zone, int items)
3173 {
3174 	int slabs;
3175 	uma_slab_t slab;
3176 	uma_keg_t keg;
3177 
3178 	keg = zone_first_keg(zone);
3179 	if (keg == NULL)
3180 		return;
3181 	KEG_LOCK(keg);
3182 	slabs = items / keg->uk_ipers;
3183 	if (slabs * keg->uk_ipers < items)
3184 		slabs++;
3185 	while (slabs > 0) {
3186 		slab = keg_alloc_slab(keg, zone, M_WAITOK);
3187 		if (slab == NULL)
3188 			break;
3189 		MPASS(slab->us_keg == keg);
3190 		LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
3191 		slabs--;
3192 	}
3193 	KEG_UNLOCK(keg);
3194 }
3195 
3196 /* See uma.h */
3197 uint32_t *
3198 uma_find_refcnt(uma_zone_t zone, void *item)
3199 {
3200 	uma_slabrefcnt_t slabref;
3201 	uma_slab_t slab;
3202 	uma_keg_t keg;
3203 	uint32_t *refcnt;
3204 	int idx;
3205 
3206 	slab = vtoslab((vm_offset_t)item & (~UMA_SLAB_MASK));
3207 	slabref = (uma_slabrefcnt_t)slab;
3208 	keg = slab->us_keg;
3209 	KASSERT(keg->uk_flags & UMA_ZONE_REFCNT,
3210 	    ("uma_find_refcnt(): zone possibly not UMA_ZONE_REFCNT"));
3211 	idx = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
3212 	refcnt = &slabref->us_refcnt[idx];
3213 	return refcnt;
3214 }
3215 
3216 /* See uma.h */
3217 static void
3218 uma_reclaim_locked(bool kmem_danger)
3219 {
3220 
3221 #ifdef UMA_DEBUG
3222 	printf("UMA: vm asked us to release pages!\n");
3223 #endif
3224 	sx_assert(&uma_drain_lock, SA_XLOCKED);
3225 	bucket_enable();
3226 	zone_foreach(zone_drain);
3227 	if (vm_page_count_min() || kmem_danger) {
3228 		cache_drain_safe(NULL);
3229 		zone_foreach(zone_drain);
3230 	}
3231 	/*
3232 	 * Some slabs may have been freed but this zone will be visited early
3233 	 * we visit again so that we can free pages that are empty once other
3234 	 * zones are drained.  We have to do the same for buckets.
3235 	 */
3236 	zone_drain(slabzone);
3237 	zone_drain(slabrefzone);
3238 	bucket_zone_drain();
3239 }
3240 
3241 void
3242 uma_reclaim(void)
3243 {
3244 
3245 	sx_xlock(&uma_drain_lock);
3246 	uma_reclaim_locked(false);
3247 	sx_xunlock(&uma_drain_lock);
3248 }
3249 
3250 static int uma_reclaim_needed;
3251 
3252 void
3253 uma_reclaim_wakeup(void)
3254 {
3255 
3256 	uma_reclaim_needed = 1;
3257 	wakeup(&uma_reclaim_needed);
3258 }
3259 
3260 void
3261 uma_reclaim_worker(void *arg __unused)
3262 {
3263 
3264 	sx_xlock(&uma_drain_lock);
3265 	for (;;) {
3266 		sx_sleep(&uma_reclaim_needed, &uma_drain_lock, PVM,
3267 		    "umarcl", 0);
3268 		if (uma_reclaim_needed) {
3269 			uma_reclaim_needed = 0;
3270 			uma_reclaim_locked(true);
3271 		}
3272 	}
3273 }
3274 
3275 /* See uma.h */
3276 int
3277 uma_zone_exhausted(uma_zone_t zone)
3278 {
3279 	int full;
3280 
3281 	ZONE_LOCK(zone);
3282 	full = (zone->uz_flags & UMA_ZFLAG_FULL);
3283 	ZONE_UNLOCK(zone);
3284 	return (full);
3285 }
3286 
3287 int
3288 uma_zone_exhausted_nolock(uma_zone_t zone)
3289 {
3290 	return (zone->uz_flags & UMA_ZFLAG_FULL);
3291 }
3292 
3293 void *
3294 uma_large_malloc(vm_size_t size, int wait)
3295 {
3296 	void *mem;
3297 	uma_slab_t slab;
3298 	uint8_t flags;
3299 
3300 	slab = zone_alloc_item(slabzone, NULL, wait);
3301 	if (slab == NULL)
3302 		return (NULL);
3303 	mem = page_alloc(NULL, size, &flags, wait);
3304 	if (mem) {
3305 		vsetslab((vm_offset_t)mem, slab);
3306 		slab->us_data = mem;
3307 		slab->us_flags = flags | UMA_SLAB_MALLOC;
3308 		slab->us_size = size;
3309 	} else {
3310 		zone_free_item(slabzone, slab, NULL, SKIP_NONE);
3311 	}
3312 
3313 	return (mem);
3314 }
3315 
3316 void
3317 uma_large_free(uma_slab_t slab)
3318 {
3319 
3320 	page_free(slab->us_data, slab->us_size, slab->us_flags);
3321 	zone_free_item(slabzone, slab, NULL, SKIP_NONE);
3322 }
3323 
3324 static void
3325 uma_zero_item(void *item, uma_zone_t zone)
3326 {
3327 
3328 	if (zone->uz_flags & UMA_ZONE_PCPU) {
3329 		for (int i = 0; i < mp_ncpus; i++)
3330 			bzero(zpcpu_get_cpu(item, i), zone->uz_size);
3331 	} else
3332 		bzero(item, zone->uz_size);
3333 }
3334 
3335 void
3336 uma_print_stats(void)
3337 {
3338 	zone_foreach(uma_print_zone);
3339 }
3340 
3341 static void
3342 slab_print(uma_slab_t slab)
3343 {
3344 	printf("slab: keg %p, data %p, freecount %d\n",
3345 		slab->us_keg, slab->us_data, slab->us_freecount);
3346 }
3347 
3348 static void
3349 cache_print(uma_cache_t cache)
3350 {
3351 	printf("alloc: %p(%d), free: %p(%d)\n",
3352 		cache->uc_allocbucket,
3353 		cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0,
3354 		cache->uc_freebucket,
3355 		cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0);
3356 }
3357 
3358 static void
3359 uma_print_keg(uma_keg_t keg)
3360 {
3361 	uma_slab_t slab;
3362 
3363 	printf("keg: %s(%p) size %d(%d) flags %#x ipers %d ppera %d "
3364 	    "out %d free %d limit %d\n",
3365 	    keg->uk_name, keg, keg->uk_size, keg->uk_rsize, keg->uk_flags,
3366 	    keg->uk_ipers, keg->uk_ppera,
3367 	    (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free,
3368 	    (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers);
3369 	printf("Part slabs:\n");
3370 	LIST_FOREACH(slab, &keg->uk_part_slab, us_link)
3371 		slab_print(slab);
3372 	printf("Free slabs:\n");
3373 	LIST_FOREACH(slab, &keg->uk_free_slab, us_link)
3374 		slab_print(slab);
3375 	printf("Full slabs:\n");
3376 	LIST_FOREACH(slab, &keg->uk_full_slab, us_link)
3377 		slab_print(slab);
3378 }
3379 
3380 void
3381 uma_print_zone(uma_zone_t zone)
3382 {
3383 	uma_cache_t cache;
3384 	uma_klink_t kl;
3385 	int i;
3386 
3387 	printf("zone: %s(%p) size %d flags %#x\n",
3388 	    zone->uz_name, zone, zone->uz_size, zone->uz_flags);
3389 	LIST_FOREACH(kl, &zone->uz_kegs, kl_link)
3390 		uma_print_keg(kl->kl_keg);
3391 	CPU_FOREACH(i) {
3392 		cache = &zone->uz_cpu[i];
3393 		printf("CPU %d Cache:\n", i);
3394 		cache_print(cache);
3395 	}
3396 }
3397 
3398 #ifdef DDB
3399 /*
3400  * Generate statistics across both the zone and its per-cpu cache's.  Return
3401  * desired statistics if the pointer is non-NULL for that statistic.
3402  *
3403  * Note: does not update the zone statistics, as it can't safely clear the
3404  * per-CPU cache statistic.
3405  *
3406  * XXXRW: Following the uc_allocbucket and uc_freebucket pointers here isn't
3407  * safe from off-CPU; we should modify the caches to track this information
3408  * directly so that we don't have to.
3409  */
3410 static void
3411 uma_zone_sumstat(uma_zone_t z, int *cachefreep, uint64_t *allocsp,
3412     uint64_t *freesp, uint64_t *sleepsp)
3413 {
3414 	uma_cache_t cache;
3415 	uint64_t allocs, frees, sleeps;
3416 	int cachefree, cpu;
3417 
3418 	allocs = frees = sleeps = 0;
3419 	cachefree = 0;
3420 	CPU_FOREACH(cpu) {
3421 		cache = &z->uz_cpu[cpu];
3422 		if (cache->uc_allocbucket != NULL)
3423 			cachefree += cache->uc_allocbucket->ub_cnt;
3424 		if (cache->uc_freebucket != NULL)
3425 			cachefree += cache->uc_freebucket->ub_cnt;
3426 		allocs += cache->uc_allocs;
3427 		frees += cache->uc_frees;
3428 	}
3429 	allocs += z->uz_allocs;
3430 	frees += z->uz_frees;
3431 	sleeps += z->uz_sleeps;
3432 	if (cachefreep != NULL)
3433 		*cachefreep = cachefree;
3434 	if (allocsp != NULL)
3435 		*allocsp = allocs;
3436 	if (freesp != NULL)
3437 		*freesp = frees;
3438 	if (sleepsp != NULL)
3439 		*sleepsp = sleeps;
3440 }
3441 #endif /* DDB */
3442 
3443 static int
3444 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
3445 {
3446 	uma_keg_t kz;
3447 	uma_zone_t z;
3448 	int count;
3449 
3450 	count = 0;
3451 	rw_rlock(&uma_rwlock);
3452 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3453 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
3454 			count++;
3455 	}
3456 	rw_runlock(&uma_rwlock);
3457 	return (sysctl_handle_int(oidp, &count, 0, req));
3458 }
3459 
3460 static int
3461 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
3462 {
3463 	struct uma_stream_header ush;
3464 	struct uma_type_header uth;
3465 	struct uma_percpu_stat ups;
3466 	uma_bucket_t bucket;
3467 	struct sbuf sbuf;
3468 	uma_cache_t cache;
3469 	uma_klink_t kl;
3470 	uma_keg_t kz;
3471 	uma_zone_t z;
3472 	uma_keg_t k;
3473 	int count, error, i;
3474 
3475 	error = sysctl_wire_old_buffer(req, 0);
3476 	if (error != 0)
3477 		return (error);
3478 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
3479 	sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL);
3480 
3481 	count = 0;
3482 	rw_rlock(&uma_rwlock);
3483 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3484 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
3485 			count++;
3486 	}
3487 
3488 	/*
3489 	 * Insert stream header.
3490 	 */
3491 	bzero(&ush, sizeof(ush));
3492 	ush.ush_version = UMA_STREAM_VERSION;
3493 	ush.ush_maxcpus = (mp_maxid + 1);
3494 	ush.ush_count = count;
3495 	(void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
3496 
3497 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3498 		LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3499 			bzero(&uth, sizeof(uth));
3500 			ZONE_LOCK(z);
3501 			strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
3502 			uth.uth_align = kz->uk_align;
3503 			uth.uth_size = kz->uk_size;
3504 			uth.uth_rsize = kz->uk_rsize;
3505 			LIST_FOREACH(kl, &z->uz_kegs, kl_link) {
3506 				k = kl->kl_keg;
3507 				uth.uth_maxpages += k->uk_maxpages;
3508 				uth.uth_pages += k->uk_pages;
3509 				uth.uth_keg_free += k->uk_free;
3510 				uth.uth_limit = (k->uk_maxpages / k->uk_ppera)
3511 				    * k->uk_ipers;
3512 			}
3513 
3514 			/*
3515 			 * A zone is secondary is it is not the first entry
3516 			 * on the keg's zone list.
3517 			 */
3518 			if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
3519 			    (LIST_FIRST(&kz->uk_zones) != z))
3520 				uth.uth_zone_flags = UTH_ZONE_SECONDARY;
3521 
3522 			LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3523 				uth.uth_zone_free += bucket->ub_cnt;
3524 			uth.uth_allocs = z->uz_allocs;
3525 			uth.uth_frees = z->uz_frees;
3526 			uth.uth_fails = z->uz_fails;
3527 			uth.uth_sleeps = z->uz_sleeps;
3528 			(void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
3529 			/*
3530 			 * While it is not normally safe to access the cache
3531 			 * bucket pointers while not on the CPU that owns the
3532 			 * cache, we only allow the pointers to be exchanged
3533 			 * without the zone lock held, not invalidated, so
3534 			 * accept the possible race associated with bucket
3535 			 * exchange during monitoring.
3536 			 */
3537 			for (i = 0; i < (mp_maxid + 1); i++) {
3538 				bzero(&ups, sizeof(ups));
3539 				if (kz->uk_flags & UMA_ZFLAG_INTERNAL)
3540 					goto skip;
3541 				if (CPU_ABSENT(i))
3542 					goto skip;
3543 				cache = &z->uz_cpu[i];
3544 				if (cache->uc_allocbucket != NULL)
3545 					ups.ups_cache_free +=
3546 					    cache->uc_allocbucket->ub_cnt;
3547 				if (cache->uc_freebucket != NULL)
3548 					ups.ups_cache_free +=
3549 					    cache->uc_freebucket->ub_cnt;
3550 				ups.ups_allocs = cache->uc_allocs;
3551 				ups.ups_frees = cache->uc_frees;
3552 skip:
3553 				(void)sbuf_bcat(&sbuf, &ups, sizeof(ups));
3554 			}
3555 			ZONE_UNLOCK(z);
3556 		}
3557 	}
3558 	rw_runlock(&uma_rwlock);
3559 	error = sbuf_finish(&sbuf);
3560 	sbuf_delete(&sbuf);
3561 	return (error);
3562 }
3563 
3564 int
3565 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS)
3566 {
3567 	uma_zone_t zone = *(uma_zone_t *)arg1;
3568 	int error, max;
3569 
3570 	max = uma_zone_get_max(zone);
3571 	error = sysctl_handle_int(oidp, &max, 0, req);
3572 	if (error || !req->newptr)
3573 		return (error);
3574 
3575 	uma_zone_set_max(zone, max);
3576 
3577 	return (0);
3578 }
3579 
3580 int
3581 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS)
3582 {
3583 	uma_zone_t zone = *(uma_zone_t *)arg1;
3584 	int cur;
3585 
3586 	cur = uma_zone_get_cur(zone);
3587 	return (sysctl_handle_int(oidp, &cur, 0, req));
3588 }
3589 
3590 #ifdef DDB
3591 DB_SHOW_COMMAND(uma, db_show_uma)
3592 {
3593 	uint64_t allocs, frees, sleeps;
3594 	uma_bucket_t bucket;
3595 	uma_keg_t kz;
3596 	uma_zone_t z;
3597 	int cachefree;
3598 
3599 	db_printf("%18s %8s %8s %8s %12s %8s %8s\n", "Zone", "Size", "Used",
3600 	    "Free", "Requests", "Sleeps", "Bucket");
3601 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3602 		LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3603 			if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
3604 				allocs = z->uz_allocs;
3605 				frees = z->uz_frees;
3606 				sleeps = z->uz_sleeps;
3607 				cachefree = 0;
3608 			} else
3609 				uma_zone_sumstat(z, &cachefree, &allocs,
3610 				    &frees, &sleeps);
3611 			if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
3612 			    (LIST_FIRST(&kz->uk_zones) != z)))
3613 				cachefree += kz->uk_free;
3614 			LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3615 				cachefree += bucket->ub_cnt;
3616 			db_printf("%18s %8ju %8jd %8d %12ju %8ju %8u\n",
3617 			    z->uz_name, (uintmax_t)kz->uk_size,
3618 			    (intmax_t)(allocs - frees), cachefree,
3619 			    (uintmax_t)allocs, sleeps, z->uz_count);
3620 			if (db_pager_quit)
3621 				return;
3622 		}
3623 	}
3624 }
3625 
3626 DB_SHOW_COMMAND(umacache, db_show_umacache)
3627 {
3628 	uint64_t allocs, frees;
3629 	uma_bucket_t bucket;
3630 	uma_zone_t z;
3631 	int cachefree;
3632 
3633 	db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
3634 	    "Requests", "Bucket");
3635 	LIST_FOREACH(z, &uma_cachezones, uz_link) {
3636 		uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL);
3637 		LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3638 			cachefree += bucket->ub_cnt;
3639 		db_printf("%18s %8ju %8jd %8d %12ju %8u\n",
3640 		    z->uz_name, (uintmax_t)z->uz_size,
3641 		    (intmax_t)(allocs - frees), cachefree,
3642 		    (uintmax_t)allocs, z->uz_count);
3643 		if (db_pager_quit)
3644 			return;
3645 	}
3646 }
3647 #endif
3648