xref: /freebsd/sys/vm/uma_core.c (revision b1f9167f94059fd55c630891d359bcff987bd7eb)
1 /*-
2  * Copyright (c) 2002-2005, 2009, 2013 Jeffrey Roberson <jeff@FreeBSD.org>
3  * Copyright (c) 2004, 2005 Bosko Milekic <bmilekic@FreeBSD.org>
4  * Copyright (c) 2004-2006 Robert N. M. Watson
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice unmodified, this list of conditions, and the following
12  *    disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
18  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
19  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
20  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
21  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
22  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
23  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
24  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
25  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
26  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
27  */
28 
29 /*
30  * uma_core.c  Implementation of the Universal Memory allocator
31  *
32  * This allocator is intended to replace the multitude of similar object caches
33  * in the standard FreeBSD kernel.  The intent is to be flexible as well as
34  * effecient.  A primary design goal is to return unused memory to the rest of
35  * the system.  This will make the system as a whole more flexible due to the
36  * ability to move memory to subsystems which most need it instead of leaving
37  * pools of reserved memory unused.
38  *
39  * The basic ideas stem from similar slab/zone based allocators whose algorithms
40  * are well known.
41  *
42  */
43 
44 /*
45  * TODO:
46  *	- Improve memory usage for large allocations
47  *	- Investigate cache size adjustments
48  */
49 
50 #include <sys/cdefs.h>
51 __FBSDID("$FreeBSD$");
52 
53 /* I should really use ktr.. */
54 /*
55 #define UMA_DEBUG 1
56 #define UMA_DEBUG_ALLOC 1
57 #define UMA_DEBUG_ALLOC_1 1
58 */
59 
60 #include "opt_ddb.h"
61 #include "opt_param.h"
62 #include "opt_vm.h"
63 
64 #include <sys/param.h>
65 #include <sys/systm.h>
66 #include <sys/bitset.h>
67 #include <sys/kernel.h>
68 #include <sys/types.h>
69 #include <sys/queue.h>
70 #include <sys/malloc.h>
71 #include <sys/ktr.h>
72 #include <sys/lock.h>
73 #include <sys/sysctl.h>
74 #include <sys/mutex.h>
75 #include <sys/proc.h>
76 #include <sys/rwlock.h>
77 #include <sys/sbuf.h>
78 #include <sys/sched.h>
79 #include <sys/smp.h>
80 #include <sys/vmmeter.h>
81 
82 #include <vm/vm.h>
83 #include <vm/vm_object.h>
84 #include <vm/vm_page.h>
85 #include <vm/vm_pageout.h>
86 #include <vm/vm_param.h>
87 #include <vm/vm_map.h>
88 #include <vm/vm_kern.h>
89 #include <vm/vm_extern.h>
90 #include <vm/uma.h>
91 #include <vm/uma_int.h>
92 #include <vm/uma_dbg.h>
93 
94 #include <ddb/ddb.h>
95 
96 #ifdef DEBUG_MEMGUARD
97 #include <vm/memguard.h>
98 #endif
99 
100 /*
101  * This is the zone and keg from which all zones are spawned.  The idea is that
102  * even the zone & keg heads are allocated from the allocator, so we use the
103  * bss section to bootstrap us.
104  */
105 static struct uma_keg masterkeg;
106 static struct uma_zone masterzone_k;
107 static struct uma_zone masterzone_z;
108 static uma_zone_t kegs = &masterzone_k;
109 static uma_zone_t zones = &masterzone_z;
110 
111 /* This is the zone from which all of uma_slab_t's are allocated. */
112 static uma_zone_t slabzone;
113 static uma_zone_t slabrefzone;	/* With refcounters (for UMA_ZONE_REFCNT) */
114 
115 /*
116  * The initial hash tables come out of this zone so they can be allocated
117  * prior to malloc coming up.
118  */
119 static uma_zone_t hashzone;
120 
121 /* The boot-time adjusted value for cache line alignment. */
122 int uma_align_cache = 64 - 1;
123 
124 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets");
125 
126 /*
127  * Are we allowed to allocate buckets?
128  */
129 static int bucketdisable = 1;
130 
131 /* Linked list of all kegs in the system */
132 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs);
133 
134 /* Linked list of all cache-only zones in the system */
135 static LIST_HEAD(,uma_zone) uma_cachezones =
136     LIST_HEAD_INITIALIZER(uma_cachezones);
137 
138 /* This mutex protects the keg list */
139 static struct mtx_padalign uma_mtx;
140 
141 /* Linked list of boot time pages */
142 static LIST_HEAD(,uma_slab) uma_boot_pages =
143     LIST_HEAD_INITIALIZER(uma_boot_pages);
144 
145 /* This mutex protects the boot time pages list */
146 static struct mtx_padalign uma_boot_pages_mtx;
147 
148 /* Is the VM done starting up? */
149 static int booted = 0;
150 #define	UMA_STARTUP	1
151 #define	UMA_STARTUP2	2
152 
153 /*
154  * Only mbuf clusters use ref zones.  Just provide enough references
155  * to support the one user.  New code should not use the ref facility.
156  */
157 static const u_int uma_max_ipers_ref = PAGE_SIZE / MCLBYTES;
158 
159 /*
160  * This is the handle used to schedule events that need to happen
161  * outside of the allocation fast path.
162  */
163 static struct callout uma_callout;
164 #define	UMA_TIMEOUT	20		/* Seconds for callout interval. */
165 
166 /*
167  * This structure is passed as the zone ctor arg so that I don't have to create
168  * a special allocation function just for zones.
169  */
170 struct uma_zctor_args {
171 	const char *name;
172 	size_t size;
173 	uma_ctor ctor;
174 	uma_dtor dtor;
175 	uma_init uminit;
176 	uma_fini fini;
177 	uma_import import;
178 	uma_release release;
179 	void *arg;
180 	uma_keg_t keg;
181 	int align;
182 	uint32_t flags;
183 };
184 
185 struct uma_kctor_args {
186 	uma_zone_t zone;
187 	size_t size;
188 	uma_init uminit;
189 	uma_fini fini;
190 	int align;
191 	uint32_t flags;
192 };
193 
194 struct uma_bucket_zone {
195 	uma_zone_t	ubz_zone;
196 	char		*ubz_name;
197 	int		ubz_entries;	/* Number of items it can hold. */
198 	int		ubz_maxsize;	/* Maximum allocation size per-item. */
199 };
200 
201 /*
202  * Compute the actual number of bucket entries to pack them in power
203  * of two sizes for more efficient space utilization.
204  */
205 #define	BUCKET_SIZE(n)						\
206     (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *))
207 
208 #define	BUCKET_MAX	BUCKET_SIZE(256)
209 
210 struct uma_bucket_zone bucket_zones[] = {
211 	{ NULL, "4 Bucket", BUCKET_SIZE(4), 4096 },
212 	{ NULL, "6 Bucket", BUCKET_SIZE(6), 3072 },
213 	{ NULL, "8 Bucket", BUCKET_SIZE(8), 2048 },
214 	{ NULL, "12 Bucket", BUCKET_SIZE(12), 1536 },
215 	{ NULL, "16 Bucket", BUCKET_SIZE(16), 1024 },
216 	{ NULL, "32 Bucket", BUCKET_SIZE(32), 512 },
217 	{ NULL, "64 Bucket", BUCKET_SIZE(64), 256 },
218 	{ NULL, "128 Bucket", BUCKET_SIZE(128), 128 },
219 	{ NULL, "256 Bucket", BUCKET_SIZE(256), 64 },
220 	{ NULL, NULL, 0}
221 };
222 
223 /*
224  * Flags and enumerations to be passed to internal functions.
225  */
226 enum zfreeskip { SKIP_NONE = 0, SKIP_DTOR, SKIP_FINI };
227 
228 /* Prototypes.. */
229 
230 static void *noobj_alloc(uma_zone_t, int, uint8_t *, int);
231 static void *page_alloc(uma_zone_t, int, uint8_t *, int);
232 static void *startup_alloc(uma_zone_t, int, uint8_t *, int);
233 static void page_free(void *, int, uint8_t);
234 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int);
235 static void cache_drain(uma_zone_t);
236 static void bucket_drain(uma_zone_t, uma_bucket_t);
237 static void bucket_cache_drain(uma_zone_t zone);
238 static int keg_ctor(void *, int, void *, int);
239 static void keg_dtor(void *, int, void *);
240 static int zone_ctor(void *, int, void *, int);
241 static void zone_dtor(void *, int, void *);
242 static int zero_init(void *, int, int);
243 static void keg_small_init(uma_keg_t keg);
244 static void keg_large_init(uma_keg_t keg);
245 static void zone_foreach(void (*zfunc)(uma_zone_t));
246 static void zone_timeout(uma_zone_t zone);
247 static int hash_alloc(struct uma_hash *);
248 static int hash_expand(struct uma_hash *, struct uma_hash *);
249 static void hash_free(struct uma_hash *hash);
250 static void uma_timeout(void *);
251 static void uma_startup3(void);
252 static void *zone_alloc_item(uma_zone_t, void *, int);
253 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip);
254 static void bucket_enable(void);
255 static void bucket_init(void);
256 static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int);
257 static void bucket_free(uma_zone_t zone, uma_bucket_t, void *);
258 static void bucket_zone_drain(void);
259 static uma_bucket_t zone_alloc_bucket(uma_zone_t zone, void *, int flags);
260 static uma_slab_t zone_fetch_slab(uma_zone_t zone, uma_keg_t last, int flags);
261 static uma_slab_t zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int flags);
262 static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab);
263 static void slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item);
264 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit,
265     uma_fini fini, int align, uint32_t flags);
266 static int zone_import(uma_zone_t zone, void **bucket, int max, int flags);
267 static void zone_release(uma_zone_t zone, void **bucket, int cnt);
268 static void uma_zero_item(void *item, uma_zone_t zone);
269 
270 void uma_print_zone(uma_zone_t);
271 void uma_print_stats(void);
272 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS);
273 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS);
274 
275 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
276 
277 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLTYPE_INT,
278     0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones");
279 
280 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLTYPE_STRUCT,
281     0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats");
282 
283 static int zone_warnings = 1;
284 TUNABLE_INT("vm.zone_warnings", &zone_warnings);
285 SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RW, &zone_warnings, 0,
286     "Warn when UMA zones becomes full");
287 
288 /*
289  * This routine checks to see whether or not it's safe to enable buckets.
290  */
291 static void
292 bucket_enable(void)
293 {
294 	bucketdisable = vm_page_count_min();
295 }
296 
297 /*
298  * Initialize bucket_zones, the array of zones of buckets of various sizes.
299  *
300  * For each zone, calculate the memory required for each bucket, consisting
301  * of the header and an array of pointers.
302  */
303 static void
304 bucket_init(void)
305 {
306 	struct uma_bucket_zone *ubz;
307 	int size;
308 	int i;
309 
310 	for (i = 0, ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) {
311 		size = roundup(sizeof(struct uma_bucket), sizeof(void *));
312 		size += sizeof(void *) * ubz->ubz_entries;
313 		ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
314 		    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
315 		    UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET);
316 	}
317 }
318 
319 /*
320  * Given a desired number of entries for a bucket, return the zone from which
321  * to allocate the bucket.
322  */
323 static struct uma_bucket_zone *
324 bucket_zone_lookup(int entries)
325 {
326 	struct uma_bucket_zone *ubz;
327 
328 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
329 		if (ubz->ubz_entries >= entries)
330 			return (ubz);
331 	ubz--;
332 	return (ubz);
333 }
334 
335 static int
336 bucket_select(int size)
337 {
338 	struct uma_bucket_zone *ubz;
339 
340 	ubz = &bucket_zones[0];
341 	if (size > ubz->ubz_maxsize)
342 		return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1);
343 
344 	for (; ubz->ubz_entries != 0; ubz++)
345 		if (ubz->ubz_maxsize < size)
346 			break;
347 	ubz--;
348 	return (ubz->ubz_entries);
349 }
350 
351 static uma_bucket_t
352 bucket_alloc(uma_zone_t zone, void *udata, int flags)
353 {
354 	struct uma_bucket_zone *ubz;
355 	uma_bucket_t bucket;
356 
357 	/*
358 	 * This is to stop us from allocating per cpu buckets while we're
359 	 * running out of vm.boot_pages.  Otherwise, we would exhaust the
360 	 * boot pages.  This also prevents us from allocating buckets in
361 	 * low memory situations.
362 	 */
363 	if (bucketdisable)
364 		return (NULL);
365 	/*
366 	 * To limit bucket recursion we store the original zone flags
367 	 * in a cookie passed via zalloc_arg/zfree_arg.  This allows the
368 	 * NOVM flag to persist even through deep recursions.  We also
369 	 * store ZFLAG_BUCKET once we have recursed attempting to allocate
370 	 * a bucket for a bucket zone so we do not allow infinite bucket
371 	 * recursion.  This cookie will even persist to frees of unused
372 	 * buckets via the allocation path or bucket allocations in the
373 	 * free path.
374 	 */
375 	if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
376 		udata = (void *)(uintptr_t)zone->uz_flags;
377 	else {
378 		if ((uintptr_t)udata & UMA_ZFLAG_BUCKET)
379 			return (NULL);
380 		udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET);
381 	}
382 	if ((uintptr_t)udata & UMA_ZFLAG_CACHEONLY)
383 		flags |= M_NOVM;
384 	ubz = bucket_zone_lookup(zone->uz_count);
385 	if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0)
386 		ubz++;
387 	bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags);
388 	if (bucket) {
389 #ifdef INVARIANTS
390 		bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
391 #endif
392 		bucket->ub_cnt = 0;
393 		bucket->ub_entries = ubz->ubz_entries;
394 	}
395 
396 	return (bucket);
397 }
398 
399 static void
400 bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata)
401 {
402 	struct uma_bucket_zone *ubz;
403 
404 	KASSERT(bucket->ub_cnt == 0,
405 	    ("bucket_free: Freeing a non free bucket."));
406 	if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
407 		udata = (void *)(uintptr_t)zone->uz_flags;
408 	ubz = bucket_zone_lookup(bucket->ub_entries);
409 	uma_zfree_arg(ubz->ubz_zone, bucket, udata);
410 }
411 
412 static void
413 bucket_zone_drain(void)
414 {
415 	struct uma_bucket_zone *ubz;
416 
417 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
418 		zone_drain(ubz->ubz_zone);
419 }
420 
421 static void
422 zone_log_warning(uma_zone_t zone)
423 {
424 	static const struct timeval warninterval = { 300, 0 };
425 
426 	if (!zone_warnings || zone->uz_warning == NULL)
427 		return;
428 
429 	if (ratecheck(&zone->uz_ratecheck, &warninterval))
430 		printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning);
431 }
432 
433 static void
434 zone_foreach_keg(uma_zone_t zone, void (*kegfn)(uma_keg_t))
435 {
436 	uma_klink_t klink;
437 
438 	LIST_FOREACH(klink, &zone->uz_kegs, kl_link)
439 		kegfn(klink->kl_keg);
440 }
441 
442 /*
443  * Routine called by timeout which is used to fire off some time interval
444  * based calculations.  (stats, hash size, etc.)
445  *
446  * Arguments:
447  *	arg   Unused
448  *
449  * Returns:
450  *	Nothing
451  */
452 static void
453 uma_timeout(void *unused)
454 {
455 	bucket_enable();
456 	zone_foreach(zone_timeout);
457 
458 	/* Reschedule this event */
459 	callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
460 }
461 
462 /*
463  * Routine to perform timeout driven calculations.  This expands the
464  * hashes and does per cpu statistics aggregation.
465  *
466  *  Returns nothing.
467  */
468 static void
469 keg_timeout(uma_keg_t keg)
470 {
471 
472 	KEG_LOCK(keg);
473 	/*
474 	 * Expand the keg hash table.
475 	 *
476 	 * This is done if the number of slabs is larger than the hash size.
477 	 * What I'm trying to do here is completely reduce collisions.  This
478 	 * may be a little aggressive.  Should I allow for two collisions max?
479 	 */
480 	if (keg->uk_flags & UMA_ZONE_HASH &&
481 	    keg->uk_pages / keg->uk_ppera >= keg->uk_hash.uh_hashsize) {
482 		struct uma_hash newhash;
483 		struct uma_hash oldhash;
484 		int ret;
485 
486 		/*
487 		 * This is so involved because allocating and freeing
488 		 * while the keg lock is held will lead to deadlock.
489 		 * I have to do everything in stages and check for
490 		 * races.
491 		 */
492 		newhash = keg->uk_hash;
493 		KEG_UNLOCK(keg);
494 		ret = hash_alloc(&newhash);
495 		KEG_LOCK(keg);
496 		if (ret) {
497 			if (hash_expand(&keg->uk_hash, &newhash)) {
498 				oldhash = keg->uk_hash;
499 				keg->uk_hash = newhash;
500 			} else
501 				oldhash = newhash;
502 
503 			KEG_UNLOCK(keg);
504 			hash_free(&oldhash);
505 			return;
506 		}
507 	}
508 	KEG_UNLOCK(keg);
509 }
510 
511 static void
512 zone_timeout(uma_zone_t zone)
513 {
514 
515 	zone_foreach_keg(zone, &keg_timeout);
516 }
517 
518 /*
519  * Allocate and zero fill the next sized hash table from the appropriate
520  * backing store.
521  *
522  * Arguments:
523  *	hash  A new hash structure with the old hash size in uh_hashsize
524  *
525  * Returns:
526  *	1 on sucess and 0 on failure.
527  */
528 static int
529 hash_alloc(struct uma_hash *hash)
530 {
531 	int oldsize;
532 	int alloc;
533 
534 	oldsize = hash->uh_hashsize;
535 
536 	/* We're just going to go to a power of two greater */
537 	if (oldsize)  {
538 		hash->uh_hashsize = oldsize * 2;
539 		alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
540 		hash->uh_slab_hash = (struct slabhead *)malloc(alloc,
541 		    M_UMAHASH, M_NOWAIT);
542 	} else {
543 		alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
544 		hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
545 		    M_WAITOK);
546 		hash->uh_hashsize = UMA_HASH_SIZE_INIT;
547 	}
548 	if (hash->uh_slab_hash) {
549 		bzero(hash->uh_slab_hash, alloc);
550 		hash->uh_hashmask = hash->uh_hashsize - 1;
551 		return (1);
552 	}
553 
554 	return (0);
555 }
556 
557 /*
558  * Expands the hash table for HASH zones.  This is done from zone_timeout
559  * to reduce collisions.  This must not be done in the regular allocation
560  * path, otherwise, we can recurse on the vm while allocating pages.
561  *
562  * Arguments:
563  *	oldhash  The hash you want to expand
564  *	newhash  The hash structure for the new table
565  *
566  * Returns:
567  *	Nothing
568  *
569  * Discussion:
570  */
571 static int
572 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
573 {
574 	uma_slab_t slab;
575 	int hval;
576 	int i;
577 
578 	if (!newhash->uh_slab_hash)
579 		return (0);
580 
581 	if (oldhash->uh_hashsize >= newhash->uh_hashsize)
582 		return (0);
583 
584 	/*
585 	 * I need to investigate hash algorithms for resizing without a
586 	 * full rehash.
587 	 */
588 
589 	for (i = 0; i < oldhash->uh_hashsize; i++)
590 		while (!SLIST_EMPTY(&oldhash->uh_slab_hash[i])) {
591 			slab = SLIST_FIRST(&oldhash->uh_slab_hash[i]);
592 			SLIST_REMOVE_HEAD(&oldhash->uh_slab_hash[i], us_hlink);
593 			hval = UMA_HASH(newhash, slab->us_data);
594 			SLIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
595 			    slab, us_hlink);
596 		}
597 
598 	return (1);
599 }
600 
601 /*
602  * Free the hash bucket to the appropriate backing store.
603  *
604  * Arguments:
605  *	slab_hash  The hash bucket we're freeing
606  *	hashsize   The number of entries in that hash bucket
607  *
608  * Returns:
609  *	Nothing
610  */
611 static void
612 hash_free(struct uma_hash *hash)
613 {
614 	if (hash->uh_slab_hash == NULL)
615 		return;
616 	if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
617 		zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE);
618 	else
619 		free(hash->uh_slab_hash, M_UMAHASH);
620 }
621 
622 /*
623  * Frees all outstanding items in a bucket
624  *
625  * Arguments:
626  *	zone   The zone to free to, must be unlocked.
627  *	bucket The free/alloc bucket with items, cpu queue must be locked.
628  *
629  * Returns:
630  *	Nothing
631  */
632 
633 static void
634 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
635 {
636 	int i;
637 
638 	if (bucket == NULL)
639 		return;
640 
641 	if (zone->uz_fini)
642 		for (i = 0; i < bucket->ub_cnt; i++)
643 			zone->uz_fini(bucket->ub_bucket[i], zone->uz_size);
644 	zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt);
645 	bucket->ub_cnt = 0;
646 }
647 
648 /*
649  * Drains the per cpu caches for a zone.
650  *
651  * NOTE: This may only be called while the zone is being turn down, and not
652  * during normal operation.  This is necessary in order that we do not have
653  * to migrate CPUs to drain the per-CPU caches.
654  *
655  * Arguments:
656  *	zone     The zone to drain, must be unlocked.
657  *
658  * Returns:
659  *	Nothing
660  */
661 static void
662 cache_drain(uma_zone_t zone)
663 {
664 	uma_cache_t cache;
665 	int cpu;
666 
667 	/*
668 	 * XXX: It is safe to not lock the per-CPU caches, because we're
669 	 * tearing down the zone anyway.  I.e., there will be no further use
670 	 * of the caches at this point.
671 	 *
672 	 * XXX: It would good to be able to assert that the zone is being
673 	 * torn down to prevent improper use of cache_drain().
674 	 *
675 	 * XXX: We lock the zone before passing into bucket_cache_drain() as
676 	 * it is used elsewhere.  Should the tear-down path be made special
677 	 * there in some form?
678 	 */
679 	CPU_FOREACH(cpu) {
680 		cache = &zone->uz_cpu[cpu];
681 		bucket_drain(zone, cache->uc_allocbucket);
682 		bucket_drain(zone, cache->uc_freebucket);
683 		if (cache->uc_allocbucket != NULL)
684 			bucket_free(zone, cache->uc_allocbucket, NULL);
685 		if (cache->uc_freebucket != NULL)
686 			bucket_free(zone, cache->uc_freebucket, NULL);
687 		cache->uc_allocbucket = cache->uc_freebucket = NULL;
688 	}
689 	ZONE_LOCK(zone);
690 	bucket_cache_drain(zone);
691 	ZONE_UNLOCK(zone);
692 }
693 
694 static void
695 cache_shrink(uma_zone_t zone)
696 {
697 
698 	if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
699 		return;
700 
701 	ZONE_LOCK(zone);
702 	zone->uz_count = (zone->uz_count_min + zone->uz_count) / 2;
703 	ZONE_UNLOCK(zone);
704 }
705 
706 static void
707 cache_drain_safe_cpu(uma_zone_t zone)
708 {
709 	uma_cache_t cache;
710 	uma_bucket_t b1, b2;
711 
712 	if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
713 		return;
714 
715 	b1 = b2 = NULL;
716 	ZONE_LOCK(zone);
717 	critical_enter();
718 	cache = &zone->uz_cpu[curcpu];
719 	if (cache->uc_allocbucket) {
720 		if (cache->uc_allocbucket->ub_cnt != 0)
721 			LIST_INSERT_HEAD(&zone->uz_buckets,
722 			    cache->uc_allocbucket, ub_link);
723 		else
724 			b1 = cache->uc_allocbucket;
725 		cache->uc_allocbucket = NULL;
726 	}
727 	if (cache->uc_freebucket) {
728 		if (cache->uc_freebucket->ub_cnt != 0)
729 			LIST_INSERT_HEAD(&zone->uz_buckets,
730 			    cache->uc_freebucket, ub_link);
731 		else
732 			b2 = cache->uc_freebucket;
733 		cache->uc_freebucket = NULL;
734 	}
735 	critical_exit();
736 	ZONE_UNLOCK(zone);
737 	if (b1)
738 		bucket_free(zone, b1, NULL);
739 	if (b2)
740 		bucket_free(zone, b2, NULL);
741 }
742 
743 /*
744  * Safely drain per-CPU caches of a zone(s) to alloc bucket.
745  * This is an expensive call because it needs to bind to all CPUs
746  * one by one and enter a critical section on each of them in order
747  * to safely access their cache buckets.
748  * Zone lock must not be held on call this function.
749  */
750 static void
751 cache_drain_safe(uma_zone_t zone)
752 {
753 	int cpu;
754 
755 	/*
756 	 * Polite bucket sizes shrinking was not enouth, shrink aggressively.
757 	 */
758 	if (zone)
759 		cache_shrink(zone);
760 	else
761 		zone_foreach(cache_shrink);
762 
763 	CPU_FOREACH(cpu) {
764 		thread_lock(curthread);
765 		sched_bind(curthread, cpu);
766 		thread_unlock(curthread);
767 
768 		if (zone)
769 			cache_drain_safe_cpu(zone);
770 		else
771 			zone_foreach(cache_drain_safe_cpu);
772 	}
773 	thread_lock(curthread);
774 	sched_unbind(curthread);
775 	thread_unlock(curthread);
776 }
777 
778 /*
779  * Drain the cached buckets from a zone.  Expects a locked zone on entry.
780  */
781 static void
782 bucket_cache_drain(uma_zone_t zone)
783 {
784 	uma_bucket_t bucket;
785 
786 	/*
787 	 * Drain the bucket queues and free the buckets, we just keep two per
788 	 * cpu (alloc/free).
789 	 */
790 	while ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) {
791 		LIST_REMOVE(bucket, ub_link);
792 		ZONE_UNLOCK(zone);
793 		bucket_drain(zone, bucket);
794 		bucket_free(zone, bucket, NULL);
795 		ZONE_LOCK(zone);
796 	}
797 
798 	/*
799 	 * Shrink further bucket sizes.  Price of single zone lock collision
800 	 * is probably lower then price of global cache drain.
801 	 */
802 	if (zone->uz_count > zone->uz_count_min)
803 		zone->uz_count--;
804 }
805 
806 static void
807 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start)
808 {
809 	uint8_t *mem;
810 	int i;
811 	uint8_t flags;
812 
813 	mem = slab->us_data;
814 	flags = slab->us_flags;
815 	i = start;
816 	if (keg->uk_fini != NULL) {
817 		for (i--; i > -1; i--)
818 			keg->uk_fini(slab->us_data + (keg->uk_rsize * i),
819 			    keg->uk_size);
820 	}
821 	if (keg->uk_flags & UMA_ZONE_OFFPAGE)
822 		zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE);
823 #ifdef UMA_DEBUG
824 	printf("%s: Returning %d bytes.\n", keg->uk_name,
825 	    PAGE_SIZE * keg->uk_ppera);
826 #endif
827 	keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags);
828 }
829 
830 /*
831  * Frees pages from a keg back to the system.  This is done on demand from
832  * the pageout daemon.
833  *
834  * Returns nothing.
835  */
836 static void
837 keg_drain(uma_keg_t keg)
838 {
839 	struct slabhead freeslabs = { 0 };
840 	uma_slab_t slab;
841 	uma_slab_t n;
842 
843 	/*
844 	 * We don't want to take pages from statically allocated kegs at this
845 	 * time
846 	 */
847 	if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL)
848 		return;
849 
850 #ifdef UMA_DEBUG
851 	printf("%s free items: %u\n", keg->uk_name, keg->uk_free);
852 #endif
853 	KEG_LOCK(keg);
854 	if (keg->uk_free == 0)
855 		goto finished;
856 
857 	slab = LIST_FIRST(&keg->uk_free_slab);
858 	while (slab) {
859 		n = LIST_NEXT(slab, us_link);
860 
861 		/* We have no where to free these to */
862 		if (slab->us_flags & UMA_SLAB_BOOT) {
863 			slab = n;
864 			continue;
865 		}
866 
867 		LIST_REMOVE(slab, us_link);
868 		keg->uk_pages -= keg->uk_ppera;
869 		keg->uk_free -= keg->uk_ipers;
870 
871 		if (keg->uk_flags & UMA_ZONE_HASH)
872 			UMA_HASH_REMOVE(&keg->uk_hash, slab, slab->us_data);
873 
874 		SLIST_INSERT_HEAD(&freeslabs, slab, us_hlink);
875 
876 		slab = n;
877 	}
878 finished:
879 	KEG_UNLOCK(keg);
880 
881 	while ((slab = SLIST_FIRST(&freeslabs)) != NULL) {
882 		SLIST_REMOVE(&freeslabs, slab, uma_slab, us_hlink);
883 		keg_free_slab(keg, slab, keg->uk_ipers);
884 	}
885 }
886 
887 static void
888 zone_drain_wait(uma_zone_t zone, int waitok)
889 {
890 
891 	/*
892 	 * Set draining to interlock with zone_dtor() so we can release our
893 	 * locks as we go.  Only dtor() should do a WAITOK call since it
894 	 * is the only call that knows the structure will still be available
895 	 * when it wakes up.
896 	 */
897 	ZONE_LOCK(zone);
898 	while (zone->uz_flags & UMA_ZFLAG_DRAINING) {
899 		if (waitok == M_NOWAIT)
900 			goto out;
901 		mtx_unlock(&uma_mtx);
902 		msleep(zone, zone->uz_lockptr, PVM, "zonedrain", 1);
903 		mtx_lock(&uma_mtx);
904 	}
905 	zone->uz_flags |= UMA_ZFLAG_DRAINING;
906 	bucket_cache_drain(zone);
907 	ZONE_UNLOCK(zone);
908 	/*
909 	 * The DRAINING flag protects us from being freed while
910 	 * we're running.  Normally the uma_mtx would protect us but we
911 	 * must be able to release and acquire the right lock for each keg.
912 	 */
913 	zone_foreach_keg(zone, &keg_drain);
914 	ZONE_LOCK(zone);
915 	zone->uz_flags &= ~UMA_ZFLAG_DRAINING;
916 	wakeup(zone);
917 out:
918 	ZONE_UNLOCK(zone);
919 }
920 
921 void
922 zone_drain(uma_zone_t zone)
923 {
924 
925 	zone_drain_wait(zone, M_NOWAIT);
926 }
927 
928 /*
929  * Allocate a new slab for a keg.  This does not insert the slab onto a list.
930  *
931  * Arguments:
932  *	wait  Shall we wait?
933  *
934  * Returns:
935  *	The slab that was allocated or NULL if there is no memory and the
936  *	caller specified M_NOWAIT.
937  */
938 static uma_slab_t
939 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int wait)
940 {
941 	uma_slabrefcnt_t slabref;
942 	uma_alloc allocf;
943 	uma_slab_t slab;
944 	uint8_t *mem;
945 	uint8_t flags;
946 	int i;
947 
948 	mtx_assert(&keg->uk_lock, MA_OWNED);
949 	slab = NULL;
950 	mem = NULL;
951 
952 #ifdef UMA_DEBUG
953 	printf("alloc_slab:  Allocating a new slab for %s\n", keg->uk_name);
954 #endif
955 	allocf = keg->uk_allocf;
956 	KEG_UNLOCK(keg);
957 
958 	if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
959 		slab = zone_alloc_item(keg->uk_slabzone, NULL, wait);
960 		if (slab == NULL)
961 			goto out;
962 	}
963 
964 	/*
965 	 * This reproduces the old vm_zone behavior of zero filling pages the
966 	 * first time they are added to a zone.
967 	 *
968 	 * Malloced items are zeroed in uma_zalloc.
969 	 */
970 
971 	if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
972 		wait |= M_ZERO;
973 	else
974 		wait &= ~M_ZERO;
975 
976 	if (keg->uk_flags & UMA_ZONE_NODUMP)
977 		wait |= M_NODUMP;
978 
979 	/* zone is passed for legacy reasons. */
980 	mem = allocf(zone, keg->uk_ppera * PAGE_SIZE, &flags, wait);
981 	if (mem == NULL) {
982 		if (keg->uk_flags & UMA_ZONE_OFFPAGE)
983 			zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE);
984 		slab = NULL;
985 		goto out;
986 	}
987 
988 	/* Point the slab into the allocated memory */
989 	if (!(keg->uk_flags & UMA_ZONE_OFFPAGE))
990 		slab = (uma_slab_t )(mem + keg->uk_pgoff);
991 
992 	if (keg->uk_flags & UMA_ZONE_VTOSLAB)
993 		for (i = 0; i < keg->uk_ppera; i++)
994 			vsetslab((vm_offset_t)mem + (i * PAGE_SIZE), slab);
995 
996 	slab->us_keg = keg;
997 	slab->us_data = mem;
998 	slab->us_freecount = keg->uk_ipers;
999 	slab->us_flags = flags;
1000 	BIT_FILL(SLAB_SETSIZE, &slab->us_free);
1001 #ifdef INVARIANTS
1002 	BIT_ZERO(SLAB_SETSIZE, &slab->us_debugfree);
1003 #endif
1004 	if (keg->uk_flags & UMA_ZONE_REFCNT) {
1005 		slabref = (uma_slabrefcnt_t)slab;
1006 		for (i = 0; i < keg->uk_ipers; i++)
1007 			slabref->us_refcnt[i] = 0;
1008 	}
1009 
1010 	if (keg->uk_init != NULL) {
1011 		for (i = 0; i < keg->uk_ipers; i++)
1012 			if (keg->uk_init(slab->us_data + (keg->uk_rsize * i),
1013 			    keg->uk_size, wait) != 0)
1014 				break;
1015 		if (i != keg->uk_ipers) {
1016 			keg_free_slab(keg, slab, i);
1017 			slab = NULL;
1018 			goto out;
1019 		}
1020 	}
1021 out:
1022 	KEG_LOCK(keg);
1023 
1024 	if (slab != NULL) {
1025 		if (keg->uk_flags & UMA_ZONE_HASH)
1026 			UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
1027 
1028 		keg->uk_pages += keg->uk_ppera;
1029 		keg->uk_free += keg->uk_ipers;
1030 	}
1031 
1032 	return (slab);
1033 }
1034 
1035 /*
1036  * This function is intended to be used early on in place of page_alloc() so
1037  * that we may use the boot time page cache to satisfy allocations before
1038  * the VM is ready.
1039  */
1040 static void *
1041 startup_alloc(uma_zone_t zone, int bytes, uint8_t *pflag, int wait)
1042 {
1043 	uma_keg_t keg;
1044 	uma_slab_t tmps;
1045 	int pages, check_pages;
1046 
1047 	keg = zone_first_keg(zone);
1048 	pages = howmany(bytes, PAGE_SIZE);
1049 	check_pages = pages - 1;
1050 	KASSERT(pages > 0, ("startup_alloc can't reserve 0 pages\n"));
1051 
1052 	/*
1053 	 * Check our small startup cache to see if it has pages remaining.
1054 	 */
1055 	mtx_lock(&uma_boot_pages_mtx);
1056 
1057 	/* First check if we have enough room. */
1058 	tmps = LIST_FIRST(&uma_boot_pages);
1059 	while (tmps != NULL && check_pages-- > 0)
1060 		tmps = LIST_NEXT(tmps, us_link);
1061 	if (tmps != NULL) {
1062 		/*
1063 		 * It's ok to lose tmps references.  The last one will
1064 		 * have tmps->us_data pointing to the start address of
1065 		 * "pages" contiguous pages of memory.
1066 		 */
1067 		while (pages-- > 0) {
1068 			tmps = LIST_FIRST(&uma_boot_pages);
1069 			LIST_REMOVE(tmps, us_link);
1070 		}
1071 		mtx_unlock(&uma_boot_pages_mtx);
1072 		*pflag = tmps->us_flags;
1073 		return (tmps->us_data);
1074 	}
1075 	mtx_unlock(&uma_boot_pages_mtx);
1076 	if (booted < UMA_STARTUP2)
1077 		panic("UMA: Increase vm.boot_pages");
1078 	/*
1079 	 * Now that we've booted reset these users to their real allocator.
1080 	 */
1081 #ifdef UMA_MD_SMALL_ALLOC
1082 	keg->uk_allocf = (keg->uk_ppera > 1) ? page_alloc : uma_small_alloc;
1083 #else
1084 	keg->uk_allocf = page_alloc;
1085 #endif
1086 	return keg->uk_allocf(zone, bytes, pflag, wait);
1087 }
1088 
1089 /*
1090  * Allocates a number of pages from the system
1091  *
1092  * Arguments:
1093  *	bytes  The number of bytes requested
1094  *	wait  Shall we wait?
1095  *
1096  * Returns:
1097  *	A pointer to the alloced memory or possibly
1098  *	NULL if M_NOWAIT is set.
1099  */
1100 static void *
1101 page_alloc(uma_zone_t zone, int bytes, uint8_t *pflag, int wait)
1102 {
1103 	void *p;	/* Returned page */
1104 
1105 	*pflag = UMA_SLAB_KMEM;
1106 	p = (void *) kmem_malloc(kmem_arena, bytes, wait);
1107 
1108 	return (p);
1109 }
1110 
1111 /*
1112  * Allocates a number of pages from within an object
1113  *
1114  * Arguments:
1115  *	bytes  The number of bytes requested
1116  *	wait   Shall we wait?
1117  *
1118  * Returns:
1119  *	A pointer to the alloced memory or possibly
1120  *	NULL if M_NOWAIT is set.
1121  */
1122 static void *
1123 noobj_alloc(uma_zone_t zone, int bytes, uint8_t *flags, int wait)
1124 {
1125 	TAILQ_HEAD(, vm_page) alloctail;
1126 	u_long npages;
1127 	vm_offset_t retkva, zkva;
1128 	vm_page_t p, p_next;
1129 	uma_keg_t keg;
1130 
1131 	TAILQ_INIT(&alloctail);
1132 	keg = zone_first_keg(zone);
1133 
1134 	npages = howmany(bytes, PAGE_SIZE);
1135 	while (npages > 0) {
1136 		p = vm_page_alloc(NULL, 0, VM_ALLOC_INTERRUPT |
1137 		    VM_ALLOC_WIRED | VM_ALLOC_NOOBJ);
1138 		if (p != NULL) {
1139 			/*
1140 			 * Since the page does not belong to an object, its
1141 			 * listq is unused.
1142 			 */
1143 			TAILQ_INSERT_TAIL(&alloctail, p, listq);
1144 			npages--;
1145 			continue;
1146 		}
1147 		if (wait & M_WAITOK) {
1148 			VM_WAIT;
1149 			continue;
1150 		}
1151 
1152 		/*
1153 		 * Page allocation failed, free intermediate pages and
1154 		 * exit.
1155 		 */
1156 		TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1157 			vm_page_unwire(p, PQ_INACTIVE);
1158 			vm_page_free(p);
1159 		}
1160 		return (NULL);
1161 	}
1162 	*flags = UMA_SLAB_PRIV;
1163 	zkva = keg->uk_kva +
1164 	    atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
1165 	retkva = zkva;
1166 	TAILQ_FOREACH(p, &alloctail, listq) {
1167 		pmap_qenter(zkva, &p, 1);
1168 		zkva += PAGE_SIZE;
1169 	}
1170 
1171 	return ((void *)retkva);
1172 }
1173 
1174 /*
1175  * Frees a number of pages to the system
1176  *
1177  * Arguments:
1178  *	mem   A pointer to the memory to be freed
1179  *	size  The size of the memory being freed
1180  *	flags The original p->us_flags field
1181  *
1182  * Returns:
1183  *	Nothing
1184  */
1185 static void
1186 page_free(void *mem, int size, uint8_t flags)
1187 {
1188 	struct vmem *vmem;
1189 
1190 	if (flags & UMA_SLAB_KMEM)
1191 		vmem = kmem_arena;
1192 	else if (flags & UMA_SLAB_KERNEL)
1193 		vmem = kernel_arena;
1194 	else
1195 		panic("UMA: page_free used with invalid flags %d", flags);
1196 
1197 	kmem_free(vmem, (vm_offset_t)mem, size);
1198 }
1199 
1200 /*
1201  * Zero fill initializer
1202  *
1203  * Arguments/Returns follow uma_init specifications
1204  */
1205 static int
1206 zero_init(void *mem, int size, int flags)
1207 {
1208 	bzero(mem, size);
1209 	return (0);
1210 }
1211 
1212 /*
1213  * Finish creating a small uma keg.  This calculates ipers, and the keg size.
1214  *
1215  * Arguments
1216  *	keg  The zone we should initialize
1217  *
1218  * Returns
1219  *	Nothing
1220  */
1221 static void
1222 keg_small_init(uma_keg_t keg)
1223 {
1224 	u_int rsize;
1225 	u_int memused;
1226 	u_int wastedspace;
1227 	u_int shsize;
1228 
1229 	if (keg->uk_flags & UMA_ZONE_PCPU) {
1230 		u_int ncpus = mp_ncpus ? mp_ncpus : MAXCPU;
1231 
1232 		keg->uk_slabsize = sizeof(struct pcpu);
1233 		keg->uk_ppera = howmany(ncpus * sizeof(struct pcpu),
1234 		    PAGE_SIZE);
1235 	} else {
1236 		keg->uk_slabsize = UMA_SLAB_SIZE;
1237 		keg->uk_ppera = 1;
1238 	}
1239 
1240 	/*
1241 	 * Calculate the size of each allocation (rsize) according to
1242 	 * alignment.  If the requested size is smaller than we have
1243 	 * allocation bits for we round it up.
1244 	 */
1245 	rsize = keg->uk_size;
1246 	if (rsize < keg->uk_slabsize / SLAB_SETSIZE)
1247 		rsize = keg->uk_slabsize / SLAB_SETSIZE;
1248 	if (rsize & keg->uk_align)
1249 		rsize = (rsize & ~keg->uk_align) + (keg->uk_align + 1);
1250 	keg->uk_rsize = rsize;
1251 
1252 	KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
1253 	    keg->uk_rsize < sizeof(struct pcpu),
1254 	    ("%s: size %u too large", __func__, keg->uk_rsize));
1255 
1256 	if (keg->uk_flags & UMA_ZONE_REFCNT)
1257 		rsize += sizeof(uint32_t);
1258 
1259 	if (keg->uk_flags & UMA_ZONE_OFFPAGE)
1260 		shsize = 0;
1261 	else
1262 		shsize = sizeof(struct uma_slab);
1263 
1264 	keg->uk_ipers = (keg->uk_slabsize - shsize) / rsize;
1265 	KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE,
1266 	    ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers));
1267 
1268 	memused = keg->uk_ipers * rsize + shsize;
1269 	wastedspace = keg->uk_slabsize - memused;
1270 
1271 	/*
1272 	 * We can't do OFFPAGE if we're internal or if we've been
1273 	 * asked to not go to the VM for buckets.  If we do this we
1274 	 * may end up going to the VM  for slabs which we do not
1275 	 * want to do if we're UMA_ZFLAG_CACHEONLY as a result
1276 	 * of UMA_ZONE_VM, which clearly forbids it.
1277 	 */
1278 	if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) ||
1279 	    (keg->uk_flags & UMA_ZFLAG_CACHEONLY))
1280 		return;
1281 
1282 	/*
1283 	 * See if using an OFFPAGE slab will limit our waste.  Only do
1284 	 * this if it permits more items per-slab.
1285 	 *
1286 	 * XXX We could try growing slabsize to limit max waste as well.
1287 	 * Historically this was not done because the VM could not
1288 	 * efficiently handle contiguous allocations.
1289 	 */
1290 	if ((wastedspace >= keg->uk_slabsize / UMA_MAX_WASTE) &&
1291 	    (keg->uk_ipers < (keg->uk_slabsize / keg->uk_rsize))) {
1292 		keg->uk_ipers = keg->uk_slabsize / keg->uk_rsize;
1293 		KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE,
1294 		    ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers));
1295 #ifdef UMA_DEBUG
1296 		printf("UMA decided we need offpage slab headers for "
1297 		    "keg: %s, calculated wastedspace = %d, "
1298 		    "maximum wasted space allowed = %d, "
1299 		    "calculated ipers = %d, "
1300 		    "new wasted space = %d\n", keg->uk_name, wastedspace,
1301 		    keg->uk_slabsize / UMA_MAX_WASTE, keg->uk_ipers,
1302 		    keg->uk_slabsize - keg->uk_ipers * keg->uk_rsize);
1303 #endif
1304 		keg->uk_flags |= UMA_ZONE_OFFPAGE;
1305 	}
1306 
1307 	if ((keg->uk_flags & UMA_ZONE_OFFPAGE) &&
1308 	    (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1309 		keg->uk_flags |= UMA_ZONE_HASH;
1310 }
1311 
1312 /*
1313  * Finish creating a large (> UMA_SLAB_SIZE) uma kegs.  Just give in and do
1314  * OFFPAGE for now.  When I can allow for more dynamic slab sizes this will be
1315  * more complicated.
1316  *
1317  * Arguments
1318  *	keg  The keg we should initialize
1319  *
1320  * Returns
1321  *	Nothing
1322  */
1323 static void
1324 keg_large_init(uma_keg_t keg)
1325 {
1326 	u_int shsize;
1327 
1328 	KASSERT(keg != NULL, ("Keg is null in keg_large_init"));
1329 	KASSERT((keg->uk_flags & UMA_ZFLAG_CACHEONLY) == 0,
1330 	    ("keg_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY keg"));
1331 	KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0,
1332 	    ("%s: Cannot large-init a UMA_ZONE_PCPU keg", __func__));
1333 
1334 	keg->uk_ppera = howmany(keg->uk_size, PAGE_SIZE);
1335 	keg->uk_slabsize = keg->uk_ppera * PAGE_SIZE;
1336 	keg->uk_ipers = 1;
1337 	keg->uk_rsize = keg->uk_size;
1338 
1339 	/* We can't do OFFPAGE if we're internal, bail out here. */
1340 	if (keg->uk_flags & UMA_ZFLAG_INTERNAL)
1341 		return;
1342 
1343 	/* Check whether we have enough space to not do OFFPAGE. */
1344 	if ((keg->uk_flags & UMA_ZONE_OFFPAGE) == 0) {
1345 		shsize = sizeof(struct uma_slab);
1346 		if (keg->uk_flags & UMA_ZONE_REFCNT)
1347 			shsize += keg->uk_ipers * sizeof(uint32_t);
1348 		if (shsize & UMA_ALIGN_PTR)
1349 			shsize = (shsize & ~UMA_ALIGN_PTR) +
1350 			    (UMA_ALIGN_PTR + 1);
1351 
1352 		if ((PAGE_SIZE * keg->uk_ppera) - keg->uk_rsize < shsize)
1353 			keg->uk_flags |= UMA_ZONE_OFFPAGE;
1354 	}
1355 
1356 	if ((keg->uk_flags & UMA_ZONE_OFFPAGE) &&
1357 	    (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1358 		keg->uk_flags |= UMA_ZONE_HASH;
1359 }
1360 
1361 static void
1362 keg_cachespread_init(uma_keg_t keg)
1363 {
1364 	int alignsize;
1365 	int trailer;
1366 	int pages;
1367 	int rsize;
1368 
1369 	KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0,
1370 	    ("%s: Cannot cachespread-init a UMA_ZONE_PCPU keg", __func__));
1371 
1372 	alignsize = keg->uk_align + 1;
1373 	rsize = keg->uk_size;
1374 	/*
1375 	 * We want one item to start on every align boundary in a page.  To
1376 	 * do this we will span pages.  We will also extend the item by the
1377 	 * size of align if it is an even multiple of align.  Otherwise, it
1378 	 * would fall on the same boundary every time.
1379 	 */
1380 	if (rsize & keg->uk_align)
1381 		rsize = (rsize & ~keg->uk_align) + alignsize;
1382 	if ((rsize & alignsize) == 0)
1383 		rsize += alignsize;
1384 	trailer = rsize - keg->uk_size;
1385 	pages = (rsize * (PAGE_SIZE / alignsize)) / PAGE_SIZE;
1386 	pages = MIN(pages, (128 * 1024) / PAGE_SIZE);
1387 	keg->uk_rsize = rsize;
1388 	keg->uk_ppera = pages;
1389 	keg->uk_slabsize = UMA_SLAB_SIZE;
1390 	keg->uk_ipers = ((pages * PAGE_SIZE) + trailer) / rsize;
1391 	keg->uk_flags |= UMA_ZONE_OFFPAGE | UMA_ZONE_VTOSLAB;
1392 	KASSERT(keg->uk_ipers <= SLAB_SETSIZE,
1393 	    ("%s: keg->uk_ipers too high(%d) increase max_ipers", __func__,
1394 	    keg->uk_ipers));
1395 }
1396 
1397 /*
1398  * Keg header ctor.  This initializes all fields, locks, etc.  And inserts
1399  * the keg onto the global keg list.
1400  *
1401  * Arguments/Returns follow uma_ctor specifications
1402  *	udata  Actually uma_kctor_args
1403  */
1404 static int
1405 keg_ctor(void *mem, int size, void *udata, int flags)
1406 {
1407 	struct uma_kctor_args *arg = udata;
1408 	uma_keg_t keg = mem;
1409 	uma_zone_t zone;
1410 
1411 	bzero(keg, size);
1412 	keg->uk_size = arg->size;
1413 	keg->uk_init = arg->uminit;
1414 	keg->uk_fini = arg->fini;
1415 	keg->uk_align = arg->align;
1416 	keg->uk_free = 0;
1417 	keg->uk_reserve = 0;
1418 	keg->uk_pages = 0;
1419 	keg->uk_flags = arg->flags;
1420 	keg->uk_allocf = page_alloc;
1421 	keg->uk_freef = page_free;
1422 	keg->uk_slabzone = NULL;
1423 
1424 	/*
1425 	 * The master zone is passed to us at keg-creation time.
1426 	 */
1427 	zone = arg->zone;
1428 	keg->uk_name = zone->uz_name;
1429 
1430 	if (arg->flags & UMA_ZONE_VM)
1431 		keg->uk_flags |= UMA_ZFLAG_CACHEONLY;
1432 
1433 	if (arg->flags & UMA_ZONE_ZINIT)
1434 		keg->uk_init = zero_init;
1435 
1436 	if (arg->flags & UMA_ZONE_REFCNT || arg->flags & UMA_ZONE_MALLOC)
1437 		keg->uk_flags |= UMA_ZONE_VTOSLAB;
1438 
1439 	if (arg->flags & UMA_ZONE_PCPU)
1440 #ifdef SMP
1441 		keg->uk_flags |= UMA_ZONE_OFFPAGE;
1442 #else
1443 		keg->uk_flags &= ~UMA_ZONE_PCPU;
1444 #endif
1445 
1446 	if (keg->uk_flags & UMA_ZONE_CACHESPREAD) {
1447 		keg_cachespread_init(keg);
1448 	} else if (keg->uk_flags & UMA_ZONE_REFCNT) {
1449 		if (keg->uk_size >
1450 		    (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt) -
1451 		    sizeof(uint32_t)))
1452 			keg_large_init(keg);
1453 		else
1454 			keg_small_init(keg);
1455 	} else {
1456 		if (keg->uk_size > (UMA_SLAB_SIZE - sizeof(struct uma_slab)))
1457 			keg_large_init(keg);
1458 		else
1459 			keg_small_init(keg);
1460 	}
1461 
1462 	if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
1463 		if (keg->uk_flags & UMA_ZONE_REFCNT) {
1464 			if (keg->uk_ipers > uma_max_ipers_ref)
1465 				panic("Too many ref items per zone: %d > %d\n",
1466 				    keg->uk_ipers, uma_max_ipers_ref);
1467 			keg->uk_slabzone = slabrefzone;
1468 		} else
1469 			keg->uk_slabzone = slabzone;
1470 	}
1471 
1472 	/*
1473 	 * If we haven't booted yet we need allocations to go through the
1474 	 * startup cache until the vm is ready.
1475 	 */
1476 	if (keg->uk_ppera == 1) {
1477 #ifdef UMA_MD_SMALL_ALLOC
1478 		keg->uk_allocf = uma_small_alloc;
1479 		keg->uk_freef = uma_small_free;
1480 
1481 		if (booted < UMA_STARTUP)
1482 			keg->uk_allocf = startup_alloc;
1483 #else
1484 		if (booted < UMA_STARTUP2)
1485 			keg->uk_allocf = startup_alloc;
1486 #endif
1487 	} else if (booted < UMA_STARTUP2 &&
1488 	    (keg->uk_flags & UMA_ZFLAG_INTERNAL))
1489 		keg->uk_allocf = startup_alloc;
1490 
1491 	/*
1492 	 * Initialize keg's lock
1493 	 */
1494 	KEG_LOCK_INIT(keg, (arg->flags & UMA_ZONE_MTXCLASS));
1495 
1496 	/*
1497 	 * If we're putting the slab header in the actual page we need to
1498 	 * figure out where in each page it goes.  This calculates a right
1499 	 * justified offset into the memory on an ALIGN_PTR boundary.
1500 	 */
1501 	if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) {
1502 		u_int totsize;
1503 
1504 		/* Size of the slab struct and free list */
1505 		totsize = sizeof(struct uma_slab);
1506 
1507 		/* Size of the reference counts. */
1508 		if (keg->uk_flags & UMA_ZONE_REFCNT)
1509 			totsize += keg->uk_ipers * sizeof(uint32_t);
1510 
1511 		if (totsize & UMA_ALIGN_PTR)
1512 			totsize = (totsize & ~UMA_ALIGN_PTR) +
1513 			    (UMA_ALIGN_PTR + 1);
1514 		keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - totsize;
1515 
1516 		/*
1517 		 * The only way the following is possible is if with our
1518 		 * UMA_ALIGN_PTR adjustments we are now bigger than
1519 		 * UMA_SLAB_SIZE.  I haven't checked whether this is
1520 		 * mathematically possible for all cases, so we make
1521 		 * sure here anyway.
1522 		 */
1523 		totsize = keg->uk_pgoff + sizeof(struct uma_slab);
1524 		if (keg->uk_flags & UMA_ZONE_REFCNT)
1525 			totsize += keg->uk_ipers * sizeof(uint32_t);
1526 		if (totsize > PAGE_SIZE * keg->uk_ppera) {
1527 			printf("zone %s ipers %d rsize %d size %d\n",
1528 			    zone->uz_name, keg->uk_ipers, keg->uk_rsize,
1529 			    keg->uk_size);
1530 			panic("UMA slab won't fit.");
1531 		}
1532 	}
1533 
1534 	if (keg->uk_flags & UMA_ZONE_HASH)
1535 		hash_alloc(&keg->uk_hash);
1536 
1537 #ifdef UMA_DEBUG
1538 	printf("UMA: %s(%p) size %d(%d) flags %#x ipers %d ppera %d out %d free %d\n",
1539 	    zone->uz_name, zone, keg->uk_size, keg->uk_rsize, keg->uk_flags,
1540 	    keg->uk_ipers, keg->uk_ppera,
1541 	    (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free);
1542 #endif
1543 
1544 	LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
1545 
1546 	mtx_lock(&uma_mtx);
1547 	LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
1548 	mtx_unlock(&uma_mtx);
1549 	return (0);
1550 }
1551 
1552 /*
1553  * Zone header ctor.  This initializes all fields, locks, etc.
1554  *
1555  * Arguments/Returns follow uma_ctor specifications
1556  *	udata  Actually uma_zctor_args
1557  */
1558 static int
1559 zone_ctor(void *mem, int size, void *udata, int flags)
1560 {
1561 	struct uma_zctor_args *arg = udata;
1562 	uma_zone_t zone = mem;
1563 	uma_zone_t z;
1564 	uma_keg_t keg;
1565 
1566 	bzero(zone, size);
1567 	zone->uz_name = arg->name;
1568 	zone->uz_ctor = arg->ctor;
1569 	zone->uz_dtor = arg->dtor;
1570 	zone->uz_slab = zone_fetch_slab;
1571 	zone->uz_init = NULL;
1572 	zone->uz_fini = NULL;
1573 	zone->uz_allocs = 0;
1574 	zone->uz_frees = 0;
1575 	zone->uz_fails = 0;
1576 	zone->uz_sleeps = 0;
1577 	zone->uz_count = 0;
1578 	zone->uz_count_min = 0;
1579 	zone->uz_flags = 0;
1580 	zone->uz_warning = NULL;
1581 	timevalclear(&zone->uz_ratecheck);
1582 	keg = arg->keg;
1583 
1584 	ZONE_LOCK_INIT(zone, (arg->flags & UMA_ZONE_MTXCLASS));
1585 
1586 	/*
1587 	 * This is a pure cache zone, no kegs.
1588 	 */
1589 	if (arg->import) {
1590 		if (arg->flags & UMA_ZONE_VM)
1591 			arg->flags |= UMA_ZFLAG_CACHEONLY;
1592 		zone->uz_flags = arg->flags;
1593 		zone->uz_size = arg->size;
1594 		zone->uz_import = arg->import;
1595 		zone->uz_release = arg->release;
1596 		zone->uz_arg = arg->arg;
1597 		zone->uz_lockptr = &zone->uz_lock;
1598 		mtx_lock(&uma_mtx);
1599 		LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link);
1600 		mtx_unlock(&uma_mtx);
1601 		goto out;
1602 	}
1603 
1604 	/*
1605 	 * Use the regular zone/keg/slab allocator.
1606 	 */
1607 	zone->uz_import = (uma_import)zone_import;
1608 	zone->uz_release = (uma_release)zone_release;
1609 	zone->uz_arg = zone;
1610 
1611 	if (arg->flags & UMA_ZONE_SECONDARY) {
1612 		KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
1613 		zone->uz_init = arg->uminit;
1614 		zone->uz_fini = arg->fini;
1615 		zone->uz_lockptr = &keg->uk_lock;
1616 		zone->uz_flags |= UMA_ZONE_SECONDARY;
1617 		mtx_lock(&uma_mtx);
1618 		ZONE_LOCK(zone);
1619 		LIST_FOREACH(z, &keg->uk_zones, uz_link) {
1620 			if (LIST_NEXT(z, uz_link) == NULL) {
1621 				LIST_INSERT_AFTER(z, zone, uz_link);
1622 				break;
1623 			}
1624 		}
1625 		ZONE_UNLOCK(zone);
1626 		mtx_unlock(&uma_mtx);
1627 	} else if (keg == NULL) {
1628 		if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
1629 		    arg->align, arg->flags)) == NULL)
1630 			return (ENOMEM);
1631 	} else {
1632 		struct uma_kctor_args karg;
1633 		int error;
1634 
1635 		/* We should only be here from uma_startup() */
1636 		karg.size = arg->size;
1637 		karg.uminit = arg->uminit;
1638 		karg.fini = arg->fini;
1639 		karg.align = arg->align;
1640 		karg.flags = arg->flags;
1641 		karg.zone = zone;
1642 		error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
1643 		    flags);
1644 		if (error)
1645 			return (error);
1646 	}
1647 
1648 	/*
1649 	 * Link in the first keg.
1650 	 */
1651 	zone->uz_klink.kl_keg = keg;
1652 	LIST_INSERT_HEAD(&zone->uz_kegs, &zone->uz_klink, kl_link);
1653 	zone->uz_lockptr = &keg->uk_lock;
1654 	zone->uz_size = keg->uk_size;
1655 	zone->uz_flags |= (keg->uk_flags &
1656 	    (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
1657 
1658 	/*
1659 	 * Some internal zones don't have room allocated for the per cpu
1660 	 * caches.  If we're internal, bail out here.
1661 	 */
1662 	if (keg->uk_flags & UMA_ZFLAG_INTERNAL) {
1663 		KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
1664 		    ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
1665 		return (0);
1666 	}
1667 
1668 out:
1669 	if ((arg->flags & UMA_ZONE_MAXBUCKET) == 0)
1670 		zone->uz_count = bucket_select(zone->uz_size);
1671 	else
1672 		zone->uz_count = BUCKET_MAX;
1673 	zone->uz_count_min = zone->uz_count;
1674 
1675 	return (0);
1676 }
1677 
1678 /*
1679  * Keg header dtor.  This frees all data, destroys locks, frees the hash
1680  * table and removes the keg from the global list.
1681  *
1682  * Arguments/Returns follow uma_dtor specifications
1683  *	udata  unused
1684  */
1685 static void
1686 keg_dtor(void *arg, int size, void *udata)
1687 {
1688 	uma_keg_t keg;
1689 
1690 	keg = (uma_keg_t)arg;
1691 	KEG_LOCK(keg);
1692 	if (keg->uk_free != 0) {
1693 		printf("Freed UMA keg (%s) was not empty (%d items). "
1694 		    " Lost %d pages of memory.\n",
1695 		    keg->uk_name ? keg->uk_name : "",
1696 		    keg->uk_free, keg->uk_pages);
1697 	}
1698 	KEG_UNLOCK(keg);
1699 
1700 	hash_free(&keg->uk_hash);
1701 
1702 	KEG_LOCK_FINI(keg);
1703 }
1704 
1705 /*
1706  * Zone header dtor.
1707  *
1708  * Arguments/Returns follow uma_dtor specifications
1709  *	udata  unused
1710  */
1711 static void
1712 zone_dtor(void *arg, int size, void *udata)
1713 {
1714 	uma_klink_t klink;
1715 	uma_zone_t zone;
1716 	uma_keg_t keg;
1717 
1718 	zone = (uma_zone_t)arg;
1719 	keg = zone_first_keg(zone);
1720 
1721 	if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
1722 		cache_drain(zone);
1723 
1724 	mtx_lock(&uma_mtx);
1725 	LIST_REMOVE(zone, uz_link);
1726 	mtx_unlock(&uma_mtx);
1727 	/*
1728 	 * XXX there are some races here where
1729 	 * the zone can be drained but zone lock
1730 	 * released and then refilled before we
1731 	 * remove it... we dont care for now
1732 	 */
1733 	zone_drain_wait(zone, M_WAITOK);
1734 	/*
1735 	 * Unlink all of our kegs.
1736 	 */
1737 	while ((klink = LIST_FIRST(&zone->uz_kegs)) != NULL) {
1738 		klink->kl_keg = NULL;
1739 		LIST_REMOVE(klink, kl_link);
1740 		if (klink == &zone->uz_klink)
1741 			continue;
1742 		free(klink, M_TEMP);
1743 	}
1744 	/*
1745 	 * We only destroy kegs from non secondary zones.
1746 	 */
1747 	if (keg != NULL && (zone->uz_flags & UMA_ZONE_SECONDARY) == 0)  {
1748 		mtx_lock(&uma_mtx);
1749 		LIST_REMOVE(keg, uk_link);
1750 		mtx_unlock(&uma_mtx);
1751 		zone_free_item(kegs, keg, NULL, SKIP_NONE);
1752 	}
1753 	ZONE_LOCK_FINI(zone);
1754 }
1755 
1756 /*
1757  * Traverses every zone in the system and calls a callback
1758  *
1759  * Arguments:
1760  *	zfunc  A pointer to a function which accepts a zone
1761  *		as an argument.
1762  *
1763  * Returns:
1764  *	Nothing
1765  */
1766 static void
1767 zone_foreach(void (*zfunc)(uma_zone_t))
1768 {
1769 	uma_keg_t keg;
1770 	uma_zone_t zone;
1771 
1772 	mtx_lock(&uma_mtx);
1773 	LIST_FOREACH(keg, &uma_kegs, uk_link) {
1774 		LIST_FOREACH(zone, &keg->uk_zones, uz_link)
1775 			zfunc(zone);
1776 	}
1777 	mtx_unlock(&uma_mtx);
1778 }
1779 
1780 /* Public functions */
1781 /* See uma.h */
1782 void
1783 uma_startup(void *bootmem, int boot_pages)
1784 {
1785 	struct uma_zctor_args args;
1786 	uma_slab_t slab;
1787 	u_int slabsize;
1788 	int i;
1789 
1790 #ifdef UMA_DEBUG
1791 	printf("Creating uma keg headers zone and keg.\n");
1792 #endif
1793 	mtx_init(&uma_mtx, "UMA lock", NULL, MTX_DEF);
1794 
1795 	/* "manually" create the initial zone */
1796 	memset(&args, 0, sizeof(args));
1797 	args.name = "UMA Kegs";
1798 	args.size = sizeof(struct uma_keg);
1799 	args.ctor = keg_ctor;
1800 	args.dtor = keg_dtor;
1801 	args.uminit = zero_init;
1802 	args.fini = NULL;
1803 	args.keg = &masterkeg;
1804 	args.align = 32 - 1;
1805 	args.flags = UMA_ZFLAG_INTERNAL;
1806 	/* The initial zone has no Per cpu queues so it's smaller */
1807 	zone_ctor(kegs, sizeof(struct uma_zone), &args, M_WAITOK);
1808 
1809 #ifdef UMA_DEBUG
1810 	printf("Filling boot free list.\n");
1811 #endif
1812 	for (i = 0; i < boot_pages; i++) {
1813 		slab = (uma_slab_t)((uint8_t *)bootmem + (i * UMA_SLAB_SIZE));
1814 		slab->us_data = (uint8_t *)slab;
1815 		slab->us_flags = UMA_SLAB_BOOT;
1816 		LIST_INSERT_HEAD(&uma_boot_pages, slab, us_link);
1817 	}
1818 	mtx_init(&uma_boot_pages_mtx, "UMA boot pages", NULL, MTX_DEF);
1819 
1820 #ifdef UMA_DEBUG
1821 	printf("Creating uma zone headers zone and keg.\n");
1822 #endif
1823 	args.name = "UMA Zones";
1824 	args.size = sizeof(struct uma_zone) +
1825 	    (sizeof(struct uma_cache) * (mp_maxid + 1));
1826 	args.ctor = zone_ctor;
1827 	args.dtor = zone_dtor;
1828 	args.uminit = zero_init;
1829 	args.fini = NULL;
1830 	args.keg = NULL;
1831 	args.align = 32 - 1;
1832 	args.flags = UMA_ZFLAG_INTERNAL;
1833 	/* The initial zone has no Per cpu queues so it's smaller */
1834 	zone_ctor(zones, sizeof(struct uma_zone), &args, M_WAITOK);
1835 
1836 #ifdef UMA_DEBUG
1837 	printf("Initializing pcpu cache locks.\n");
1838 #endif
1839 #ifdef UMA_DEBUG
1840 	printf("Creating slab and hash zones.\n");
1841 #endif
1842 
1843 	/* Now make a zone for slab headers */
1844 	slabzone = uma_zcreate("UMA Slabs",
1845 				sizeof(struct uma_slab),
1846 				NULL, NULL, NULL, NULL,
1847 				UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1848 
1849 	/*
1850 	 * We also create a zone for the bigger slabs with reference
1851 	 * counts in them, to accomodate UMA_ZONE_REFCNT zones.
1852 	 */
1853 	slabsize = sizeof(struct uma_slab_refcnt);
1854 	slabsize += uma_max_ipers_ref * sizeof(uint32_t);
1855 	slabrefzone = uma_zcreate("UMA RCntSlabs",
1856 				  slabsize,
1857 				  NULL, NULL, NULL, NULL,
1858 				  UMA_ALIGN_PTR,
1859 				  UMA_ZFLAG_INTERNAL);
1860 
1861 	hashzone = uma_zcreate("UMA Hash",
1862 	    sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
1863 	    NULL, NULL, NULL, NULL,
1864 	    UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1865 
1866 	bucket_init();
1867 
1868 	booted = UMA_STARTUP;
1869 
1870 #ifdef UMA_DEBUG
1871 	printf("UMA startup complete.\n");
1872 #endif
1873 }
1874 
1875 /* see uma.h */
1876 void
1877 uma_startup2(void)
1878 {
1879 	booted = UMA_STARTUP2;
1880 	bucket_enable();
1881 #ifdef UMA_DEBUG
1882 	printf("UMA startup2 complete.\n");
1883 #endif
1884 }
1885 
1886 /*
1887  * Initialize our callout handle
1888  *
1889  */
1890 
1891 static void
1892 uma_startup3(void)
1893 {
1894 #ifdef UMA_DEBUG
1895 	printf("Starting callout.\n");
1896 #endif
1897 	callout_init(&uma_callout, CALLOUT_MPSAFE);
1898 	callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
1899 #ifdef UMA_DEBUG
1900 	printf("UMA startup3 complete.\n");
1901 #endif
1902 }
1903 
1904 static uma_keg_t
1905 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
1906 		int align, uint32_t flags)
1907 {
1908 	struct uma_kctor_args args;
1909 
1910 	args.size = size;
1911 	args.uminit = uminit;
1912 	args.fini = fini;
1913 	args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
1914 	args.flags = flags;
1915 	args.zone = zone;
1916 	return (zone_alloc_item(kegs, &args, M_WAITOK));
1917 }
1918 
1919 /* See uma.h */
1920 void
1921 uma_set_align(int align)
1922 {
1923 
1924 	if (align != UMA_ALIGN_CACHE)
1925 		uma_align_cache = align;
1926 }
1927 
1928 /* See uma.h */
1929 uma_zone_t
1930 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
1931 		uma_init uminit, uma_fini fini, int align, uint32_t flags)
1932 
1933 {
1934 	struct uma_zctor_args args;
1935 
1936 	/* This stuff is essential for the zone ctor */
1937 	memset(&args, 0, sizeof(args));
1938 	args.name = name;
1939 	args.size = size;
1940 	args.ctor = ctor;
1941 	args.dtor = dtor;
1942 	args.uminit = uminit;
1943 	args.fini = fini;
1944 	args.align = align;
1945 	args.flags = flags;
1946 	args.keg = NULL;
1947 
1948 	return (zone_alloc_item(zones, &args, M_WAITOK));
1949 }
1950 
1951 /* See uma.h */
1952 uma_zone_t
1953 uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor,
1954 		    uma_init zinit, uma_fini zfini, uma_zone_t master)
1955 {
1956 	struct uma_zctor_args args;
1957 	uma_keg_t keg;
1958 
1959 	keg = zone_first_keg(master);
1960 	memset(&args, 0, sizeof(args));
1961 	args.name = name;
1962 	args.size = keg->uk_size;
1963 	args.ctor = ctor;
1964 	args.dtor = dtor;
1965 	args.uminit = zinit;
1966 	args.fini = zfini;
1967 	args.align = keg->uk_align;
1968 	args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
1969 	args.keg = keg;
1970 
1971 	/* XXX Attaches only one keg of potentially many. */
1972 	return (zone_alloc_item(zones, &args, M_WAITOK));
1973 }
1974 
1975 /* See uma.h */
1976 uma_zone_t
1977 uma_zcache_create(char *name, int size, uma_ctor ctor, uma_dtor dtor,
1978 		    uma_init zinit, uma_fini zfini, uma_import zimport,
1979 		    uma_release zrelease, void *arg, int flags)
1980 {
1981 	struct uma_zctor_args args;
1982 
1983 	memset(&args, 0, sizeof(args));
1984 	args.name = name;
1985 	args.size = size;
1986 	args.ctor = ctor;
1987 	args.dtor = dtor;
1988 	args.uminit = zinit;
1989 	args.fini = zfini;
1990 	args.import = zimport;
1991 	args.release = zrelease;
1992 	args.arg = arg;
1993 	args.align = 0;
1994 	args.flags = flags;
1995 
1996 	return (zone_alloc_item(zones, &args, M_WAITOK));
1997 }
1998 
1999 static void
2000 zone_lock_pair(uma_zone_t a, uma_zone_t b)
2001 {
2002 	if (a < b) {
2003 		ZONE_LOCK(a);
2004 		mtx_lock_flags(b->uz_lockptr, MTX_DUPOK);
2005 	} else {
2006 		ZONE_LOCK(b);
2007 		mtx_lock_flags(a->uz_lockptr, MTX_DUPOK);
2008 	}
2009 }
2010 
2011 static void
2012 zone_unlock_pair(uma_zone_t a, uma_zone_t b)
2013 {
2014 
2015 	ZONE_UNLOCK(a);
2016 	ZONE_UNLOCK(b);
2017 }
2018 
2019 int
2020 uma_zsecond_add(uma_zone_t zone, uma_zone_t master)
2021 {
2022 	uma_klink_t klink;
2023 	uma_klink_t kl;
2024 	int error;
2025 
2026 	error = 0;
2027 	klink = malloc(sizeof(*klink), M_TEMP, M_WAITOK | M_ZERO);
2028 
2029 	zone_lock_pair(zone, master);
2030 	/*
2031 	 * zone must use vtoslab() to resolve objects and must already be
2032 	 * a secondary.
2033 	 */
2034 	if ((zone->uz_flags & (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY))
2035 	    != (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) {
2036 		error = EINVAL;
2037 		goto out;
2038 	}
2039 	/*
2040 	 * The new master must also use vtoslab().
2041 	 */
2042 	if ((zone->uz_flags & UMA_ZONE_VTOSLAB) != UMA_ZONE_VTOSLAB) {
2043 		error = EINVAL;
2044 		goto out;
2045 	}
2046 	/*
2047 	 * Both must either be refcnt, or not be refcnt.
2048 	 */
2049 	if ((zone->uz_flags & UMA_ZONE_REFCNT) !=
2050 	    (master->uz_flags & UMA_ZONE_REFCNT)) {
2051 		error = EINVAL;
2052 		goto out;
2053 	}
2054 	/*
2055 	 * The underlying object must be the same size.  rsize
2056 	 * may be different.
2057 	 */
2058 	if (master->uz_size != zone->uz_size) {
2059 		error = E2BIG;
2060 		goto out;
2061 	}
2062 	/*
2063 	 * Put it at the end of the list.
2064 	 */
2065 	klink->kl_keg = zone_first_keg(master);
2066 	LIST_FOREACH(kl, &zone->uz_kegs, kl_link) {
2067 		if (LIST_NEXT(kl, kl_link) == NULL) {
2068 			LIST_INSERT_AFTER(kl, klink, kl_link);
2069 			break;
2070 		}
2071 	}
2072 	klink = NULL;
2073 	zone->uz_flags |= UMA_ZFLAG_MULTI;
2074 	zone->uz_slab = zone_fetch_slab_multi;
2075 
2076 out:
2077 	zone_unlock_pair(zone, master);
2078 	if (klink != NULL)
2079 		free(klink, M_TEMP);
2080 
2081 	return (error);
2082 }
2083 
2084 
2085 /* See uma.h */
2086 void
2087 uma_zdestroy(uma_zone_t zone)
2088 {
2089 
2090 	zone_free_item(zones, zone, NULL, SKIP_NONE);
2091 }
2092 
2093 /* See uma.h */
2094 void *
2095 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
2096 {
2097 	void *item;
2098 	uma_cache_t cache;
2099 	uma_bucket_t bucket;
2100 	int lockfail;
2101 	int cpu;
2102 
2103 	/* This is the fast path allocation */
2104 #ifdef UMA_DEBUG_ALLOC_1
2105 	printf("Allocating one item from %s(%p)\n", zone->uz_name, zone);
2106 #endif
2107 	CTR3(KTR_UMA, "uma_zalloc_arg thread %x zone %s flags %d", curthread,
2108 	    zone->uz_name, flags);
2109 
2110 	if (flags & M_WAITOK) {
2111 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
2112 		    "uma_zalloc_arg: zone \"%s\"", zone->uz_name);
2113 	}
2114 #ifdef DEBUG_MEMGUARD
2115 	if (memguard_cmp_zone(zone)) {
2116 		item = memguard_alloc(zone->uz_size, flags);
2117 		if (item != NULL) {
2118 			/*
2119 			 * Avoid conflict with the use-after-free
2120 			 * protecting infrastructure from INVARIANTS.
2121 			 */
2122 			if (zone->uz_init != NULL &&
2123 			    zone->uz_init != mtrash_init &&
2124 			    zone->uz_init(item, zone->uz_size, flags) != 0)
2125 				return (NULL);
2126 			if (zone->uz_ctor != NULL &&
2127 			    zone->uz_ctor != mtrash_ctor &&
2128 			    zone->uz_ctor(item, zone->uz_size, udata,
2129 			    flags) != 0) {
2130 			    	zone->uz_fini(item, zone->uz_size);
2131 				return (NULL);
2132 			}
2133 			return (item);
2134 		}
2135 		/* This is unfortunate but should not be fatal. */
2136 	}
2137 #endif
2138 	/*
2139 	 * If possible, allocate from the per-CPU cache.  There are two
2140 	 * requirements for safe access to the per-CPU cache: (1) the thread
2141 	 * accessing the cache must not be preempted or yield during access,
2142 	 * and (2) the thread must not migrate CPUs without switching which
2143 	 * cache it accesses.  We rely on a critical section to prevent
2144 	 * preemption and migration.  We release the critical section in
2145 	 * order to acquire the zone mutex if we are unable to allocate from
2146 	 * the current cache; when we re-acquire the critical section, we
2147 	 * must detect and handle migration if it has occurred.
2148 	 */
2149 	critical_enter();
2150 	cpu = curcpu;
2151 	cache = &zone->uz_cpu[cpu];
2152 
2153 zalloc_start:
2154 	bucket = cache->uc_allocbucket;
2155 	if (bucket != NULL && bucket->ub_cnt > 0) {
2156 		bucket->ub_cnt--;
2157 		item = bucket->ub_bucket[bucket->ub_cnt];
2158 #ifdef INVARIANTS
2159 		bucket->ub_bucket[bucket->ub_cnt] = NULL;
2160 #endif
2161 		KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled."));
2162 		cache->uc_allocs++;
2163 		critical_exit();
2164 		if (zone->uz_ctor != NULL &&
2165 		    zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2166 			atomic_add_long(&zone->uz_fails, 1);
2167 			zone_free_item(zone, item, udata, SKIP_DTOR);
2168 			return (NULL);
2169 		}
2170 #ifdef INVARIANTS
2171 		uma_dbg_alloc(zone, NULL, item);
2172 #endif
2173 		if (flags & M_ZERO)
2174 			uma_zero_item(item, zone);
2175 		return (item);
2176 	}
2177 
2178 	/*
2179 	 * We have run out of items in our alloc bucket.
2180 	 * See if we can switch with our free bucket.
2181 	 */
2182 	bucket = cache->uc_freebucket;
2183 	if (bucket != NULL && bucket->ub_cnt > 0) {
2184 #ifdef UMA_DEBUG_ALLOC
2185 		printf("uma_zalloc: Swapping empty with alloc.\n");
2186 #endif
2187 		cache->uc_freebucket = cache->uc_allocbucket;
2188 		cache->uc_allocbucket = bucket;
2189 		goto zalloc_start;
2190 	}
2191 
2192 	/*
2193 	 * Discard any empty allocation bucket while we hold no locks.
2194 	 */
2195 	bucket = cache->uc_allocbucket;
2196 	cache->uc_allocbucket = NULL;
2197 	critical_exit();
2198 	if (bucket != NULL)
2199 		bucket_free(zone, bucket, udata);
2200 
2201 	/* Short-circuit for zones without buckets and low memory. */
2202 	if (zone->uz_count == 0 || bucketdisable)
2203 		goto zalloc_item;
2204 
2205 	/*
2206 	 * Attempt to retrieve the item from the per-CPU cache has failed, so
2207 	 * we must go back to the zone.  This requires the zone lock, so we
2208 	 * must drop the critical section, then re-acquire it when we go back
2209 	 * to the cache.  Since the critical section is released, we may be
2210 	 * preempted or migrate.  As such, make sure not to maintain any
2211 	 * thread-local state specific to the cache from prior to releasing
2212 	 * the critical section.
2213 	 */
2214 	lockfail = 0;
2215 	if (ZONE_TRYLOCK(zone) == 0) {
2216 		/* Record contention to size the buckets. */
2217 		ZONE_LOCK(zone);
2218 		lockfail = 1;
2219 	}
2220 	critical_enter();
2221 	cpu = curcpu;
2222 	cache = &zone->uz_cpu[cpu];
2223 
2224 	/*
2225 	 * Since we have locked the zone we may as well send back our stats.
2226 	 */
2227 	atomic_add_long(&zone->uz_allocs, cache->uc_allocs);
2228 	atomic_add_long(&zone->uz_frees, cache->uc_frees);
2229 	cache->uc_allocs = 0;
2230 	cache->uc_frees = 0;
2231 
2232 	/* See if we lost the race to fill the cache. */
2233 	if (cache->uc_allocbucket != NULL) {
2234 		ZONE_UNLOCK(zone);
2235 		goto zalloc_start;
2236 	}
2237 
2238 	/*
2239 	 * Check the zone's cache of buckets.
2240 	 */
2241 	if ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) {
2242 		KASSERT(bucket->ub_cnt != 0,
2243 		    ("uma_zalloc_arg: Returning an empty bucket."));
2244 
2245 		LIST_REMOVE(bucket, ub_link);
2246 		cache->uc_allocbucket = bucket;
2247 		ZONE_UNLOCK(zone);
2248 		goto zalloc_start;
2249 	}
2250 	/* We are no longer associated with this CPU. */
2251 	critical_exit();
2252 
2253 	/*
2254 	 * We bump the uz count when the cache size is insufficient to
2255 	 * handle the working set.
2256 	 */
2257 	if (lockfail && zone->uz_count < BUCKET_MAX)
2258 		zone->uz_count++;
2259 	ZONE_UNLOCK(zone);
2260 
2261 	/*
2262 	 * Now lets just fill a bucket and put it on the free list.  If that
2263 	 * works we'll restart the allocation from the begining and it
2264 	 * will use the just filled bucket.
2265 	 */
2266 	bucket = zone_alloc_bucket(zone, udata, flags);
2267 	if (bucket != NULL) {
2268 		ZONE_LOCK(zone);
2269 		critical_enter();
2270 		cpu = curcpu;
2271 		cache = &zone->uz_cpu[cpu];
2272 		/*
2273 		 * See if we lost the race or were migrated.  Cache the
2274 		 * initialized bucket to make this less likely or claim
2275 		 * the memory directly.
2276 		 */
2277 		if (cache->uc_allocbucket == NULL)
2278 			cache->uc_allocbucket = bucket;
2279 		else
2280 			LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link);
2281 		ZONE_UNLOCK(zone);
2282 		goto zalloc_start;
2283 	}
2284 
2285 	/*
2286 	 * We may not be able to get a bucket so return an actual item.
2287 	 */
2288 #ifdef UMA_DEBUG
2289 	printf("uma_zalloc_arg: Bucketzone returned NULL\n");
2290 #endif
2291 
2292 zalloc_item:
2293 	item = zone_alloc_item(zone, udata, flags);
2294 
2295 	return (item);
2296 }
2297 
2298 static uma_slab_t
2299 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int flags)
2300 {
2301 	uma_slab_t slab;
2302 	int reserve;
2303 
2304 	mtx_assert(&keg->uk_lock, MA_OWNED);
2305 	slab = NULL;
2306 	reserve = 0;
2307 	if ((flags & M_USE_RESERVE) == 0)
2308 		reserve = keg->uk_reserve;
2309 
2310 	for (;;) {
2311 		/*
2312 		 * Find a slab with some space.  Prefer slabs that are partially
2313 		 * used over those that are totally full.  This helps to reduce
2314 		 * fragmentation.
2315 		 */
2316 		if (keg->uk_free > reserve) {
2317 			if (!LIST_EMPTY(&keg->uk_part_slab)) {
2318 				slab = LIST_FIRST(&keg->uk_part_slab);
2319 			} else {
2320 				slab = LIST_FIRST(&keg->uk_free_slab);
2321 				LIST_REMOVE(slab, us_link);
2322 				LIST_INSERT_HEAD(&keg->uk_part_slab, slab,
2323 				    us_link);
2324 			}
2325 			MPASS(slab->us_keg == keg);
2326 			return (slab);
2327 		}
2328 
2329 		/*
2330 		 * M_NOVM means don't ask at all!
2331 		 */
2332 		if (flags & M_NOVM)
2333 			break;
2334 
2335 		if (keg->uk_maxpages && keg->uk_pages >= keg->uk_maxpages) {
2336 			keg->uk_flags |= UMA_ZFLAG_FULL;
2337 			/*
2338 			 * If this is not a multi-zone, set the FULL bit.
2339 			 * Otherwise slab_multi() takes care of it.
2340 			 */
2341 			if ((zone->uz_flags & UMA_ZFLAG_MULTI) == 0) {
2342 				zone->uz_flags |= UMA_ZFLAG_FULL;
2343 				zone_log_warning(zone);
2344 			}
2345 			if (flags & M_NOWAIT)
2346 				break;
2347 			zone->uz_sleeps++;
2348 			msleep(keg, &keg->uk_lock, PVM, "keglimit", 0);
2349 			continue;
2350 		}
2351 		slab = keg_alloc_slab(keg, zone, flags);
2352 		/*
2353 		 * If we got a slab here it's safe to mark it partially used
2354 		 * and return.  We assume that the caller is going to remove
2355 		 * at least one item.
2356 		 */
2357 		if (slab) {
2358 			MPASS(slab->us_keg == keg);
2359 			LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2360 			return (slab);
2361 		}
2362 		/*
2363 		 * We might not have been able to get a slab but another cpu
2364 		 * could have while we were unlocked.  Check again before we
2365 		 * fail.
2366 		 */
2367 		flags |= M_NOVM;
2368 	}
2369 	return (slab);
2370 }
2371 
2372 static uma_slab_t
2373 zone_fetch_slab(uma_zone_t zone, uma_keg_t keg, int flags)
2374 {
2375 	uma_slab_t slab;
2376 
2377 	if (keg == NULL) {
2378 		keg = zone_first_keg(zone);
2379 		KEG_LOCK(keg);
2380 	}
2381 
2382 	for (;;) {
2383 		slab = keg_fetch_slab(keg, zone, flags);
2384 		if (slab)
2385 			return (slab);
2386 		if (flags & (M_NOWAIT | M_NOVM))
2387 			break;
2388 	}
2389 	KEG_UNLOCK(keg);
2390 	return (NULL);
2391 }
2392 
2393 /*
2394  * uma_zone_fetch_slab_multi:  Fetches a slab from one available keg.  Returns
2395  * with the keg locked.  On NULL no lock is held.
2396  *
2397  * The last pointer is used to seed the search.  It is not required.
2398  */
2399 static uma_slab_t
2400 zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int rflags)
2401 {
2402 	uma_klink_t klink;
2403 	uma_slab_t slab;
2404 	uma_keg_t keg;
2405 	int flags;
2406 	int empty;
2407 	int full;
2408 
2409 	/*
2410 	 * Don't wait on the first pass.  This will skip limit tests
2411 	 * as well.  We don't want to block if we can find a provider
2412 	 * without blocking.
2413 	 */
2414 	flags = (rflags & ~M_WAITOK) | M_NOWAIT;
2415 	/*
2416 	 * Use the last slab allocated as a hint for where to start
2417 	 * the search.
2418 	 */
2419 	if (last != NULL) {
2420 		slab = keg_fetch_slab(last, zone, flags);
2421 		if (slab)
2422 			return (slab);
2423 		KEG_UNLOCK(last);
2424 	}
2425 	/*
2426 	 * Loop until we have a slab incase of transient failures
2427 	 * while M_WAITOK is specified.  I'm not sure this is 100%
2428 	 * required but we've done it for so long now.
2429 	 */
2430 	for (;;) {
2431 		empty = 0;
2432 		full = 0;
2433 		/*
2434 		 * Search the available kegs for slabs.  Be careful to hold the
2435 		 * correct lock while calling into the keg layer.
2436 		 */
2437 		LIST_FOREACH(klink, &zone->uz_kegs, kl_link) {
2438 			keg = klink->kl_keg;
2439 			KEG_LOCK(keg);
2440 			if ((keg->uk_flags & UMA_ZFLAG_FULL) == 0) {
2441 				slab = keg_fetch_slab(keg, zone, flags);
2442 				if (slab)
2443 					return (slab);
2444 			}
2445 			if (keg->uk_flags & UMA_ZFLAG_FULL)
2446 				full++;
2447 			else
2448 				empty++;
2449 			KEG_UNLOCK(keg);
2450 		}
2451 		if (rflags & (M_NOWAIT | M_NOVM))
2452 			break;
2453 		flags = rflags;
2454 		/*
2455 		 * All kegs are full.  XXX We can't atomically check all kegs
2456 		 * and sleep so just sleep for a short period and retry.
2457 		 */
2458 		if (full && !empty) {
2459 			ZONE_LOCK(zone);
2460 			zone->uz_flags |= UMA_ZFLAG_FULL;
2461 			zone->uz_sleeps++;
2462 			zone_log_warning(zone);
2463 			msleep(zone, zone->uz_lockptr, PVM,
2464 			    "zonelimit", hz/100);
2465 			zone->uz_flags &= ~UMA_ZFLAG_FULL;
2466 			ZONE_UNLOCK(zone);
2467 			continue;
2468 		}
2469 	}
2470 	return (NULL);
2471 }
2472 
2473 static void *
2474 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
2475 {
2476 	void *item;
2477 	uint8_t freei;
2478 
2479 	MPASS(keg == slab->us_keg);
2480 	mtx_assert(&keg->uk_lock, MA_OWNED);
2481 
2482 	freei = BIT_FFS(SLAB_SETSIZE, &slab->us_free) - 1;
2483 	BIT_CLR(SLAB_SETSIZE, freei, &slab->us_free);
2484 	item = slab->us_data + (keg->uk_rsize * freei);
2485 	slab->us_freecount--;
2486 	keg->uk_free--;
2487 
2488 	/* Move this slab to the full list */
2489 	if (slab->us_freecount == 0) {
2490 		LIST_REMOVE(slab, us_link);
2491 		LIST_INSERT_HEAD(&keg->uk_full_slab, slab, us_link);
2492 	}
2493 
2494 	return (item);
2495 }
2496 
2497 static int
2498 zone_import(uma_zone_t zone, void **bucket, int max, int flags)
2499 {
2500 	uma_slab_t slab;
2501 	uma_keg_t keg;
2502 	int i;
2503 
2504 	slab = NULL;
2505 	keg = NULL;
2506 	/* Try to keep the buckets totally full */
2507 	for (i = 0; i < max; ) {
2508 		if ((slab = zone->uz_slab(zone, keg, flags)) == NULL)
2509 			break;
2510 		keg = slab->us_keg;
2511 		while (slab->us_freecount && i < max) {
2512 			bucket[i++] = slab_alloc_item(keg, slab);
2513 			if (keg->uk_free <= keg->uk_reserve)
2514 				break;
2515 		}
2516 		/* Don't grab more than one slab at a time. */
2517 		flags &= ~M_WAITOK;
2518 		flags |= M_NOWAIT;
2519 	}
2520 	if (slab != NULL)
2521 		KEG_UNLOCK(keg);
2522 
2523 	return i;
2524 }
2525 
2526 static uma_bucket_t
2527 zone_alloc_bucket(uma_zone_t zone, void *udata, int flags)
2528 {
2529 	uma_bucket_t bucket;
2530 	int max;
2531 
2532 	/* Don't wait for buckets, preserve caller's NOVM setting. */
2533 	bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
2534 	if (bucket == NULL)
2535 		return (NULL);
2536 
2537 	max = MIN(bucket->ub_entries, zone->uz_count);
2538 	bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
2539 	    max, flags);
2540 
2541 	/*
2542 	 * Initialize the memory if necessary.
2543 	 */
2544 	if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
2545 		int i;
2546 
2547 		for (i = 0; i < bucket->ub_cnt; i++)
2548 			if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size,
2549 			    flags) != 0)
2550 				break;
2551 		/*
2552 		 * If we couldn't initialize the whole bucket, put the
2553 		 * rest back onto the freelist.
2554 		 */
2555 		if (i != bucket->ub_cnt) {
2556 			zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
2557 			    bucket->ub_cnt - i);
2558 #ifdef INVARIANTS
2559 			bzero(&bucket->ub_bucket[i],
2560 			    sizeof(void *) * (bucket->ub_cnt - i));
2561 #endif
2562 			bucket->ub_cnt = i;
2563 		}
2564 	}
2565 
2566 	if (bucket->ub_cnt == 0) {
2567 		bucket_free(zone, bucket, udata);
2568 		atomic_add_long(&zone->uz_fails, 1);
2569 		return (NULL);
2570 	}
2571 
2572 	return (bucket);
2573 }
2574 
2575 /*
2576  * Allocates a single item from a zone.
2577  *
2578  * Arguments
2579  *	zone   The zone to alloc for.
2580  *	udata  The data to be passed to the constructor.
2581  *	flags  M_WAITOK, M_NOWAIT, M_ZERO.
2582  *
2583  * Returns
2584  *	NULL if there is no memory and M_NOWAIT is set
2585  *	An item if successful
2586  */
2587 
2588 static void *
2589 zone_alloc_item(uma_zone_t zone, void *udata, int flags)
2590 {
2591 	void *item;
2592 
2593 	item = NULL;
2594 
2595 #ifdef UMA_DEBUG_ALLOC
2596 	printf("INTERNAL: Allocating one item from %s(%p)\n", zone->uz_name, zone);
2597 #endif
2598 	if (zone->uz_import(zone->uz_arg, &item, 1, flags) != 1)
2599 		goto fail;
2600 	atomic_add_long(&zone->uz_allocs, 1);
2601 
2602 	/*
2603 	 * We have to call both the zone's init (not the keg's init)
2604 	 * and the zone's ctor.  This is because the item is going from
2605 	 * a keg slab directly to the user, and the user is expecting it
2606 	 * to be both zone-init'd as well as zone-ctor'd.
2607 	 */
2608 	if (zone->uz_init != NULL) {
2609 		if (zone->uz_init(item, zone->uz_size, flags) != 0) {
2610 			zone_free_item(zone, item, udata, SKIP_FINI);
2611 			goto fail;
2612 		}
2613 	}
2614 	if (zone->uz_ctor != NULL) {
2615 		if (zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2616 			zone_free_item(zone, item, udata, SKIP_DTOR);
2617 			goto fail;
2618 		}
2619 	}
2620 #ifdef INVARIANTS
2621 	uma_dbg_alloc(zone, NULL, item);
2622 #endif
2623 	if (flags & M_ZERO)
2624 		uma_zero_item(item, zone);
2625 
2626 	return (item);
2627 
2628 fail:
2629 	atomic_add_long(&zone->uz_fails, 1);
2630 	return (NULL);
2631 }
2632 
2633 /* See uma.h */
2634 void
2635 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
2636 {
2637 	uma_cache_t cache;
2638 	uma_bucket_t bucket;
2639 	int lockfail;
2640 	int cpu;
2641 
2642 #ifdef UMA_DEBUG_ALLOC_1
2643 	printf("Freeing item %p to %s(%p)\n", item, zone->uz_name, zone);
2644 #endif
2645 	CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread,
2646 	    zone->uz_name);
2647 
2648         /* uma_zfree(..., NULL) does nothing, to match free(9). */
2649         if (item == NULL)
2650                 return;
2651 #ifdef DEBUG_MEMGUARD
2652 	if (is_memguard_addr(item)) {
2653 		if (zone->uz_dtor != NULL && zone->uz_dtor != mtrash_dtor)
2654 			zone->uz_dtor(item, zone->uz_size, udata);
2655 		if (zone->uz_fini != NULL && zone->uz_fini != mtrash_fini)
2656 			zone->uz_fini(item, zone->uz_size);
2657 		memguard_free(item);
2658 		return;
2659 	}
2660 #endif
2661 #ifdef INVARIANTS
2662 	if (zone->uz_flags & UMA_ZONE_MALLOC)
2663 		uma_dbg_free(zone, udata, item);
2664 	else
2665 		uma_dbg_free(zone, NULL, item);
2666 #endif
2667 	if (zone->uz_dtor != NULL)
2668 		zone->uz_dtor(item, zone->uz_size, udata);
2669 
2670 	/*
2671 	 * The race here is acceptable.  If we miss it we'll just have to wait
2672 	 * a little longer for the limits to be reset.
2673 	 */
2674 	if (zone->uz_flags & UMA_ZFLAG_FULL)
2675 		goto zfree_item;
2676 
2677 	/*
2678 	 * If possible, free to the per-CPU cache.  There are two
2679 	 * requirements for safe access to the per-CPU cache: (1) the thread
2680 	 * accessing the cache must not be preempted or yield during access,
2681 	 * and (2) the thread must not migrate CPUs without switching which
2682 	 * cache it accesses.  We rely on a critical section to prevent
2683 	 * preemption and migration.  We release the critical section in
2684 	 * order to acquire the zone mutex if we are unable to free to the
2685 	 * current cache; when we re-acquire the critical section, we must
2686 	 * detect and handle migration if it has occurred.
2687 	 */
2688 zfree_restart:
2689 	critical_enter();
2690 	cpu = curcpu;
2691 	cache = &zone->uz_cpu[cpu];
2692 
2693 zfree_start:
2694 	/*
2695 	 * Try to free into the allocbucket first to give LIFO ordering
2696 	 * for cache-hot datastructures.  Spill over into the freebucket
2697 	 * if necessary.  Alloc will swap them if one runs dry.
2698 	 */
2699 	bucket = cache->uc_allocbucket;
2700 	if (bucket == NULL || bucket->ub_cnt >= bucket->ub_entries)
2701 		bucket = cache->uc_freebucket;
2702 	if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) {
2703 		KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL,
2704 		    ("uma_zfree: Freeing to non free bucket index."));
2705 		bucket->ub_bucket[bucket->ub_cnt] = item;
2706 		bucket->ub_cnt++;
2707 		cache->uc_frees++;
2708 		critical_exit();
2709 		return;
2710 	}
2711 
2712 	/*
2713 	 * We must go back the zone, which requires acquiring the zone lock,
2714 	 * which in turn means we must release and re-acquire the critical
2715 	 * section.  Since the critical section is released, we may be
2716 	 * preempted or migrate.  As such, make sure not to maintain any
2717 	 * thread-local state specific to the cache from prior to releasing
2718 	 * the critical section.
2719 	 */
2720 	critical_exit();
2721 	if (zone->uz_count == 0 || bucketdisable)
2722 		goto zfree_item;
2723 
2724 	lockfail = 0;
2725 	if (ZONE_TRYLOCK(zone) == 0) {
2726 		/* Record contention to size the buckets. */
2727 		ZONE_LOCK(zone);
2728 		lockfail = 1;
2729 	}
2730 	critical_enter();
2731 	cpu = curcpu;
2732 	cache = &zone->uz_cpu[cpu];
2733 
2734 	/*
2735 	 * Since we have locked the zone we may as well send back our stats.
2736 	 */
2737 	atomic_add_long(&zone->uz_allocs, cache->uc_allocs);
2738 	atomic_add_long(&zone->uz_frees, cache->uc_frees);
2739 	cache->uc_allocs = 0;
2740 	cache->uc_frees = 0;
2741 
2742 	bucket = cache->uc_freebucket;
2743 	if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) {
2744 		ZONE_UNLOCK(zone);
2745 		goto zfree_start;
2746 	}
2747 	cache->uc_freebucket = NULL;
2748 
2749 	/* Can we throw this on the zone full list? */
2750 	if (bucket != NULL) {
2751 #ifdef UMA_DEBUG_ALLOC
2752 		printf("uma_zfree: Putting old bucket on the free list.\n");
2753 #endif
2754 		/* ub_cnt is pointing to the last free item */
2755 		KASSERT(bucket->ub_cnt != 0,
2756 		    ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n"));
2757 		LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link);
2758 	}
2759 
2760 	/* We are no longer associated with this CPU. */
2761 	critical_exit();
2762 
2763 	/*
2764 	 * We bump the uz count when the cache size is insufficient to
2765 	 * handle the working set.
2766 	 */
2767 	if (lockfail && zone->uz_count < BUCKET_MAX)
2768 		zone->uz_count++;
2769 	ZONE_UNLOCK(zone);
2770 
2771 #ifdef UMA_DEBUG_ALLOC
2772 	printf("uma_zfree: Allocating new free bucket.\n");
2773 #endif
2774 	bucket = bucket_alloc(zone, udata, M_NOWAIT);
2775 	if (bucket) {
2776 		critical_enter();
2777 		cpu = curcpu;
2778 		cache = &zone->uz_cpu[cpu];
2779 		if (cache->uc_freebucket == NULL) {
2780 			cache->uc_freebucket = bucket;
2781 			goto zfree_start;
2782 		}
2783 		/*
2784 		 * We lost the race, start over.  We have to drop our
2785 		 * critical section to free the bucket.
2786 		 */
2787 		critical_exit();
2788 		bucket_free(zone, bucket, udata);
2789 		goto zfree_restart;
2790 	}
2791 
2792 	/*
2793 	 * If nothing else caught this, we'll just do an internal free.
2794 	 */
2795 zfree_item:
2796 	zone_free_item(zone, item, udata, SKIP_DTOR);
2797 
2798 	return;
2799 }
2800 
2801 static void
2802 slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item)
2803 {
2804 	uint8_t freei;
2805 
2806 	mtx_assert(&keg->uk_lock, MA_OWNED);
2807 	MPASS(keg == slab->us_keg);
2808 
2809 	/* Do we need to remove from any lists? */
2810 	if (slab->us_freecount+1 == keg->uk_ipers) {
2811 		LIST_REMOVE(slab, us_link);
2812 		LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
2813 	} else if (slab->us_freecount == 0) {
2814 		LIST_REMOVE(slab, us_link);
2815 		LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2816 	}
2817 
2818 	/* Slab management. */
2819 	freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
2820 	BIT_SET(SLAB_SETSIZE, freei, &slab->us_free);
2821 	slab->us_freecount++;
2822 
2823 	/* Keg statistics. */
2824 	keg->uk_free++;
2825 }
2826 
2827 static void
2828 zone_release(uma_zone_t zone, void **bucket, int cnt)
2829 {
2830 	void *item;
2831 	uma_slab_t slab;
2832 	uma_keg_t keg;
2833 	uint8_t *mem;
2834 	int clearfull;
2835 	int i;
2836 
2837 	clearfull = 0;
2838 	keg = zone_first_keg(zone);
2839 	KEG_LOCK(keg);
2840 	for (i = 0; i < cnt; i++) {
2841 		item = bucket[i];
2842 		if (!(zone->uz_flags & UMA_ZONE_VTOSLAB)) {
2843 			mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
2844 			if (zone->uz_flags & UMA_ZONE_HASH) {
2845 				slab = hash_sfind(&keg->uk_hash, mem);
2846 			} else {
2847 				mem += keg->uk_pgoff;
2848 				slab = (uma_slab_t)mem;
2849 			}
2850 		} else {
2851 			slab = vtoslab((vm_offset_t)item);
2852 			if (slab->us_keg != keg) {
2853 				KEG_UNLOCK(keg);
2854 				keg = slab->us_keg;
2855 				KEG_LOCK(keg);
2856 			}
2857 		}
2858 		slab_free_item(keg, slab, item);
2859 		if (keg->uk_flags & UMA_ZFLAG_FULL) {
2860 			if (keg->uk_pages < keg->uk_maxpages) {
2861 				keg->uk_flags &= ~UMA_ZFLAG_FULL;
2862 				clearfull = 1;
2863 			}
2864 
2865 			/*
2866 			 * We can handle one more allocation. Since we're
2867 			 * clearing ZFLAG_FULL, wake up all procs blocked
2868 			 * on pages. This should be uncommon, so keeping this
2869 			 * simple for now (rather than adding count of blocked
2870 			 * threads etc).
2871 			 */
2872 			wakeup(keg);
2873 		}
2874 	}
2875 	KEG_UNLOCK(keg);
2876 	if (clearfull) {
2877 		ZONE_LOCK(zone);
2878 		zone->uz_flags &= ~UMA_ZFLAG_FULL;
2879 		wakeup(zone);
2880 		ZONE_UNLOCK(zone);
2881 	}
2882 
2883 }
2884 
2885 /*
2886  * Frees a single item to any zone.
2887  *
2888  * Arguments:
2889  *	zone   The zone to free to
2890  *	item   The item we're freeing
2891  *	udata  User supplied data for the dtor
2892  *	skip   Skip dtors and finis
2893  */
2894 static void
2895 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
2896 {
2897 
2898 #ifdef INVARIANTS
2899 	if (skip == SKIP_NONE) {
2900 		if (zone->uz_flags & UMA_ZONE_MALLOC)
2901 			uma_dbg_free(zone, udata, item);
2902 		else
2903 			uma_dbg_free(zone, NULL, item);
2904 	}
2905 #endif
2906 	if (skip < SKIP_DTOR && zone->uz_dtor)
2907 		zone->uz_dtor(item, zone->uz_size, udata);
2908 
2909 	if (skip < SKIP_FINI && zone->uz_fini)
2910 		zone->uz_fini(item, zone->uz_size);
2911 
2912 	atomic_add_long(&zone->uz_frees, 1);
2913 	zone->uz_release(zone->uz_arg, &item, 1);
2914 }
2915 
2916 /* See uma.h */
2917 int
2918 uma_zone_set_max(uma_zone_t zone, int nitems)
2919 {
2920 	uma_keg_t keg;
2921 
2922 	keg = zone_first_keg(zone);
2923 	if (keg == NULL)
2924 		return (0);
2925 	KEG_LOCK(keg);
2926 	keg->uk_maxpages = (nitems / keg->uk_ipers) * keg->uk_ppera;
2927 	if (keg->uk_maxpages * keg->uk_ipers < nitems)
2928 		keg->uk_maxpages += keg->uk_ppera;
2929 	nitems = keg->uk_maxpages * keg->uk_ipers;
2930 	KEG_UNLOCK(keg);
2931 
2932 	return (nitems);
2933 }
2934 
2935 /* See uma.h */
2936 int
2937 uma_zone_get_max(uma_zone_t zone)
2938 {
2939 	int nitems;
2940 	uma_keg_t keg;
2941 
2942 	keg = zone_first_keg(zone);
2943 	if (keg == NULL)
2944 		return (0);
2945 	KEG_LOCK(keg);
2946 	nitems = keg->uk_maxpages * keg->uk_ipers;
2947 	KEG_UNLOCK(keg);
2948 
2949 	return (nitems);
2950 }
2951 
2952 /* See uma.h */
2953 void
2954 uma_zone_set_warning(uma_zone_t zone, const char *warning)
2955 {
2956 
2957 	ZONE_LOCK(zone);
2958 	zone->uz_warning = warning;
2959 	ZONE_UNLOCK(zone);
2960 }
2961 
2962 /* See uma.h */
2963 int
2964 uma_zone_get_cur(uma_zone_t zone)
2965 {
2966 	int64_t nitems;
2967 	u_int i;
2968 
2969 	ZONE_LOCK(zone);
2970 	nitems = zone->uz_allocs - zone->uz_frees;
2971 	CPU_FOREACH(i) {
2972 		/*
2973 		 * See the comment in sysctl_vm_zone_stats() regarding the
2974 		 * safety of accessing the per-cpu caches. With the zone lock
2975 		 * held, it is safe, but can potentially result in stale data.
2976 		 */
2977 		nitems += zone->uz_cpu[i].uc_allocs -
2978 		    zone->uz_cpu[i].uc_frees;
2979 	}
2980 	ZONE_UNLOCK(zone);
2981 
2982 	return (nitems < 0 ? 0 : nitems);
2983 }
2984 
2985 /* See uma.h */
2986 void
2987 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
2988 {
2989 	uma_keg_t keg;
2990 
2991 	keg = zone_first_keg(zone);
2992 	KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type"));
2993 	KEG_LOCK(keg);
2994 	KASSERT(keg->uk_pages == 0,
2995 	    ("uma_zone_set_init on non-empty keg"));
2996 	keg->uk_init = uminit;
2997 	KEG_UNLOCK(keg);
2998 }
2999 
3000 /* See uma.h */
3001 void
3002 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
3003 {
3004 	uma_keg_t keg;
3005 
3006 	keg = zone_first_keg(zone);
3007 	KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type"));
3008 	KEG_LOCK(keg);
3009 	KASSERT(keg->uk_pages == 0,
3010 	    ("uma_zone_set_fini on non-empty keg"));
3011 	keg->uk_fini = fini;
3012 	KEG_UNLOCK(keg);
3013 }
3014 
3015 /* See uma.h */
3016 void
3017 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
3018 {
3019 
3020 	ZONE_LOCK(zone);
3021 	KASSERT(zone_first_keg(zone)->uk_pages == 0,
3022 	    ("uma_zone_set_zinit on non-empty keg"));
3023 	zone->uz_init = zinit;
3024 	ZONE_UNLOCK(zone);
3025 }
3026 
3027 /* See uma.h */
3028 void
3029 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
3030 {
3031 
3032 	ZONE_LOCK(zone);
3033 	KASSERT(zone_first_keg(zone)->uk_pages == 0,
3034 	    ("uma_zone_set_zfini on non-empty keg"));
3035 	zone->uz_fini = zfini;
3036 	ZONE_UNLOCK(zone);
3037 }
3038 
3039 /* See uma.h */
3040 /* XXX uk_freef is not actually used with the zone locked */
3041 void
3042 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
3043 {
3044 	uma_keg_t keg;
3045 
3046 	keg = zone_first_keg(zone);
3047 	KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type"));
3048 	KEG_LOCK(keg);
3049 	keg->uk_freef = freef;
3050 	KEG_UNLOCK(keg);
3051 }
3052 
3053 /* See uma.h */
3054 /* XXX uk_allocf is not actually used with the zone locked */
3055 void
3056 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
3057 {
3058 	uma_keg_t keg;
3059 
3060 	keg = zone_first_keg(zone);
3061 	KEG_LOCK(keg);
3062 	keg->uk_allocf = allocf;
3063 	KEG_UNLOCK(keg);
3064 }
3065 
3066 /* See uma.h */
3067 void
3068 uma_zone_reserve(uma_zone_t zone, int items)
3069 {
3070 	uma_keg_t keg;
3071 
3072 	keg = zone_first_keg(zone);
3073 	if (keg == NULL)
3074 		return;
3075 	KEG_LOCK(keg);
3076 	keg->uk_reserve = items;
3077 	KEG_UNLOCK(keg);
3078 
3079 	return;
3080 }
3081 
3082 /* See uma.h */
3083 int
3084 uma_zone_reserve_kva(uma_zone_t zone, int count)
3085 {
3086 	uma_keg_t keg;
3087 	vm_offset_t kva;
3088 	int pages;
3089 
3090 	keg = zone_first_keg(zone);
3091 	if (keg == NULL)
3092 		return (0);
3093 	pages = count / keg->uk_ipers;
3094 
3095 	if (pages * keg->uk_ipers < count)
3096 		pages++;
3097 
3098 #ifdef UMA_MD_SMALL_ALLOC
3099 	if (keg->uk_ppera > 1) {
3100 #else
3101 	if (1) {
3102 #endif
3103 		kva = kva_alloc(pages * UMA_SLAB_SIZE);
3104 		if (kva == 0)
3105 			return (0);
3106 	} else
3107 		kva = 0;
3108 	KEG_LOCK(keg);
3109 	keg->uk_kva = kva;
3110 	keg->uk_offset = 0;
3111 	keg->uk_maxpages = pages;
3112 #ifdef UMA_MD_SMALL_ALLOC
3113 	keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
3114 #else
3115 	keg->uk_allocf = noobj_alloc;
3116 #endif
3117 	keg->uk_flags |= UMA_ZONE_NOFREE;
3118 	KEG_UNLOCK(keg);
3119 
3120 	return (1);
3121 }
3122 
3123 /* See uma.h */
3124 void
3125 uma_prealloc(uma_zone_t zone, int items)
3126 {
3127 	int slabs;
3128 	uma_slab_t slab;
3129 	uma_keg_t keg;
3130 
3131 	keg = zone_first_keg(zone);
3132 	if (keg == NULL)
3133 		return;
3134 	KEG_LOCK(keg);
3135 	slabs = items / keg->uk_ipers;
3136 	if (slabs * keg->uk_ipers < items)
3137 		slabs++;
3138 	while (slabs > 0) {
3139 		slab = keg_alloc_slab(keg, zone, M_WAITOK);
3140 		if (slab == NULL)
3141 			break;
3142 		MPASS(slab->us_keg == keg);
3143 		LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
3144 		slabs--;
3145 	}
3146 	KEG_UNLOCK(keg);
3147 }
3148 
3149 /* See uma.h */
3150 uint32_t *
3151 uma_find_refcnt(uma_zone_t zone, void *item)
3152 {
3153 	uma_slabrefcnt_t slabref;
3154 	uma_slab_t slab;
3155 	uma_keg_t keg;
3156 	uint32_t *refcnt;
3157 	int idx;
3158 
3159 	slab = vtoslab((vm_offset_t)item & (~UMA_SLAB_MASK));
3160 	slabref = (uma_slabrefcnt_t)slab;
3161 	keg = slab->us_keg;
3162 	KASSERT(keg->uk_flags & UMA_ZONE_REFCNT,
3163 	    ("uma_find_refcnt(): zone possibly not UMA_ZONE_REFCNT"));
3164 	idx = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
3165 	refcnt = &slabref->us_refcnt[idx];
3166 	return refcnt;
3167 }
3168 
3169 /* See uma.h */
3170 void
3171 uma_reclaim(void)
3172 {
3173 #ifdef UMA_DEBUG
3174 	printf("UMA: vm asked us to release pages!\n");
3175 #endif
3176 	bucket_enable();
3177 	zone_foreach(zone_drain);
3178 	if (vm_page_count_min()) {
3179 		cache_drain_safe(NULL);
3180 		zone_foreach(zone_drain);
3181 	}
3182 	/*
3183 	 * Some slabs may have been freed but this zone will be visited early
3184 	 * we visit again so that we can free pages that are empty once other
3185 	 * zones are drained.  We have to do the same for buckets.
3186 	 */
3187 	zone_drain(slabzone);
3188 	zone_drain(slabrefzone);
3189 	bucket_zone_drain();
3190 }
3191 
3192 /* See uma.h */
3193 int
3194 uma_zone_exhausted(uma_zone_t zone)
3195 {
3196 	int full;
3197 
3198 	ZONE_LOCK(zone);
3199 	full = (zone->uz_flags & UMA_ZFLAG_FULL);
3200 	ZONE_UNLOCK(zone);
3201 	return (full);
3202 }
3203 
3204 int
3205 uma_zone_exhausted_nolock(uma_zone_t zone)
3206 {
3207 	return (zone->uz_flags & UMA_ZFLAG_FULL);
3208 }
3209 
3210 void *
3211 uma_large_malloc(int size, int wait)
3212 {
3213 	void *mem;
3214 	uma_slab_t slab;
3215 	uint8_t flags;
3216 
3217 	slab = zone_alloc_item(slabzone, NULL, wait);
3218 	if (slab == NULL)
3219 		return (NULL);
3220 	mem = page_alloc(NULL, size, &flags, wait);
3221 	if (mem) {
3222 		vsetslab((vm_offset_t)mem, slab);
3223 		slab->us_data = mem;
3224 		slab->us_flags = flags | UMA_SLAB_MALLOC;
3225 		slab->us_size = size;
3226 	} else {
3227 		zone_free_item(slabzone, slab, NULL, SKIP_NONE);
3228 	}
3229 
3230 	return (mem);
3231 }
3232 
3233 void
3234 uma_large_free(uma_slab_t slab)
3235 {
3236 
3237 	page_free(slab->us_data, slab->us_size, slab->us_flags);
3238 	zone_free_item(slabzone, slab, NULL, SKIP_NONE);
3239 }
3240 
3241 static void
3242 uma_zero_item(void *item, uma_zone_t zone)
3243 {
3244 
3245 	if (zone->uz_flags & UMA_ZONE_PCPU) {
3246 		for (int i = 0; i < mp_ncpus; i++)
3247 			bzero(zpcpu_get_cpu(item, i), zone->uz_size);
3248 	} else
3249 		bzero(item, zone->uz_size);
3250 }
3251 
3252 void
3253 uma_print_stats(void)
3254 {
3255 	zone_foreach(uma_print_zone);
3256 }
3257 
3258 static void
3259 slab_print(uma_slab_t slab)
3260 {
3261 	printf("slab: keg %p, data %p, freecount %d\n",
3262 		slab->us_keg, slab->us_data, slab->us_freecount);
3263 }
3264 
3265 static void
3266 cache_print(uma_cache_t cache)
3267 {
3268 	printf("alloc: %p(%d), free: %p(%d)\n",
3269 		cache->uc_allocbucket,
3270 		cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0,
3271 		cache->uc_freebucket,
3272 		cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0);
3273 }
3274 
3275 static void
3276 uma_print_keg(uma_keg_t keg)
3277 {
3278 	uma_slab_t slab;
3279 
3280 	printf("keg: %s(%p) size %d(%d) flags %#x ipers %d ppera %d "
3281 	    "out %d free %d limit %d\n",
3282 	    keg->uk_name, keg, keg->uk_size, keg->uk_rsize, keg->uk_flags,
3283 	    keg->uk_ipers, keg->uk_ppera,
3284 	    (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free,
3285 	    (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers);
3286 	printf("Part slabs:\n");
3287 	LIST_FOREACH(slab, &keg->uk_part_slab, us_link)
3288 		slab_print(slab);
3289 	printf("Free slabs:\n");
3290 	LIST_FOREACH(slab, &keg->uk_free_slab, us_link)
3291 		slab_print(slab);
3292 	printf("Full slabs:\n");
3293 	LIST_FOREACH(slab, &keg->uk_full_slab, us_link)
3294 		slab_print(slab);
3295 }
3296 
3297 void
3298 uma_print_zone(uma_zone_t zone)
3299 {
3300 	uma_cache_t cache;
3301 	uma_klink_t kl;
3302 	int i;
3303 
3304 	printf("zone: %s(%p) size %d flags %#x\n",
3305 	    zone->uz_name, zone, zone->uz_size, zone->uz_flags);
3306 	LIST_FOREACH(kl, &zone->uz_kegs, kl_link)
3307 		uma_print_keg(kl->kl_keg);
3308 	CPU_FOREACH(i) {
3309 		cache = &zone->uz_cpu[i];
3310 		printf("CPU %d Cache:\n", i);
3311 		cache_print(cache);
3312 	}
3313 }
3314 
3315 #ifdef DDB
3316 /*
3317  * Generate statistics across both the zone and its per-cpu cache's.  Return
3318  * desired statistics if the pointer is non-NULL for that statistic.
3319  *
3320  * Note: does not update the zone statistics, as it can't safely clear the
3321  * per-CPU cache statistic.
3322  *
3323  * XXXRW: Following the uc_allocbucket and uc_freebucket pointers here isn't
3324  * safe from off-CPU; we should modify the caches to track this information
3325  * directly so that we don't have to.
3326  */
3327 static void
3328 uma_zone_sumstat(uma_zone_t z, int *cachefreep, uint64_t *allocsp,
3329     uint64_t *freesp, uint64_t *sleepsp)
3330 {
3331 	uma_cache_t cache;
3332 	uint64_t allocs, frees, sleeps;
3333 	int cachefree, cpu;
3334 
3335 	allocs = frees = sleeps = 0;
3336 	cachefree = 0;
3337 	CPU_FOREACH(cpu) {
3338 		cache = &z->uz_cpu[cpu];
3339 		if (cache->uc_allocbucket != NULL)
3340 			cachefree += cache->uc_allocbucket->ub_cnt;
3341 		if (cache->uc_freebucket != NULL)
3342 			cachefree += cache->uc_freebucket->ub_cnt;
3343 		allocs += cache->uc_allocs;
3344 		frees += cache->uc_frees;
3345 	}
3346 	allocs += z->uz_allocs;
3347 	frees += z->uz_frees;
3348 	sleeps += z->uz_sleeps;
3349 	if (cachefreep != NULL)
3350 		*cachefreep = cachefree;
3351 	if (allocsp != NULL)
3352 		*allocsp = allocs;
3353 	if (freesp != NULL)
3354 		*freesp = frees;
3355 	if (sleepsp != NULL)
3356 		*sleepsp = sleeps;
3357 }
3358 #endif /* DDB */
3359 
3360 static int
3361 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
3362 {
3363 	uma_keg_t kz;
3364 	uma_zone_t z;
3365 	int count;
3366 
3367 	count = 0;
3368 	mtx_lock(&uma_mtx);
3369 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3370 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
3371 			count++;
3372 	}
3373 	mtx_unlock(&uma_mtx);
3374 	return (sysctl_handle_int(oidp, &count, 0, req));
3375 }
3376 
3377 static int
3378 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
3379 {
3380 	struct uma_stream_header ush;
3381 	struct uma_type_header uth;
3382 	struct uma_percpu_stat ups;
3383 	uma_bucket_t bucket;
3384 	struct sbuf sbuf;
3385 	uma_cache_t cache;
3386 	uma_klink_t kl;
3387 	uma_keg_t kz;
3388 	uma_zone_t z;
3389 	uma_keg_t k;
3390 	int count, error, i;
3391 
3392 	error = sysctl_wire_old_buffer(req, 0);
3393 	if (error != 0)
3394 		return (error);
3395 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
3396 
3397 	count = 0;
3398 	mtx_lock(&uma_mtx);
3399 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3400 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
3401 			count++;
3402 	}
3403 
3404 	/*
3405 	 * Insert stream header.
3406 	 */
3407 	bzero(&ush, sizeof(ush));
3408 	ush.ush_version = UMA_STREAM_VERSION;
3409 	ush.ush_maxcpus = (mp_maxid + 1);
3410 	ush.ush_count = count;
3411 	(void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
3412 
3413 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3414 		LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3415 			bzero(&uth, sizeof(uth));
3416 			ZONE_LOCK(z);
3417 			strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
3418 			uth.uth_align = kz->uk_align;
3419 			uth.uth_size = kz->uk_size;
3420 			uth.uth_rsize = kz->uk_rsize;
3421 			LIST_FOREACH(kl, &z->uz_kegs, kl_link) {
3422 				k = kl->kl_keg;
3423 				uth.uth_maxpages += k->uk_maxpages;
3424 				uth.uth_pages += k->uk_pages;
3425 				uth.uth_keg_free += k->uk_free;
3426 				uth.uth_limit = (k->uk_maxpages / k->uk_ppera)
3427 				    * k->uk_ipers;
3428 			}
3429 
3430 			/*
3431 			 * A zone is secondary is it is not the first entry
3432 			 * on the keg's zone list.
3433 			 */
3434 			if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
3435 			    (LIST_FIRST(&kz->uk_zones) != z))
3436 				uth.uth_zone_flags = UTH_ZONE_SECONDARY;
3437 
3438 			LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3439 				uth.uth_zone_free += bucket->ub_cnt;
3440 			uth.uth_allocs = z->uz_allocs;
3441 			uth.uth_frees = z->uz_frees;
3442 			uth.uth_fails = z->uz_fails;
3443 			uth.uth_sleeps = z->uz_sleeps;
3444 			(void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
3445 			/*
3446 			 * While it is not normally safe to access the cache
3447 			 * bucket pointers while not on the CPU that owns the
3448 			 * cache, we only allow the pointers to be exchanged
3449 			 * without the zone lock held, not invalidated, so
3450 			 * accept the possible race associated with bucket
3451 			 * exchange during monitoring.
3452 			 */
3453 			for (i = 0; i < (mp_maxid + 1); i++) {
3454 				bzero(&ups, sizeof(ups));
3455 				if (kz->uk_flags & UMA_ZFLAG_INTERNAL)
3456 					goto skip;
3457 				if (CPU_ABSENT(i))
3458 					goto skip;
3459 				cache = &z->uz_cpu[i];
3460 				if (cache->uc_allocbucket != NULL)
3461 					ups.ups_cache_free +=
3462 					    cache->uc_allocbucket->ub_cnt;
3463 				if (cache->uc_freebucket != NULL)
3464 					ups.ups_cache_free +=
3465 					    cache->uc_freebucket->ub_cnt;
3466 				ups.ups_allocs = cache->uc_allocs;
3467 				ups.ups_frees = cache->uc_frees;
3468 skip:
3469 				(void)sbuf_bcat(&sbuf, &ups, sizeof(ups));
3470 			}
3471 			ZONE_UNLOCK(z);
3472 		}
3473 	}
3474 	mtx_unlock(&uma_mtx);
3475 	error = sbuf_finish(&sbuf);
3476 	sbuf_delete(&sbuf);
3477 	return (error);
3478 }
3479 
3480 int
3481 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS)
3482 {
3483 	uma_zone_t zone = *(uma_zone_t *)arg1;
3484 	int error, max, old;
3485 
3486 	old = max = uma_zone_get_max(zone);
3487 	error = sysctl_handle_int(oidp, &max, 0, req);
3488 	if (error || !req->newptr)
3489 		return (error);
3490 
3491 	if (max < old)
3492 		return (EINVAL);
3493 
3494 	uma_zone_set_max(zone, max);
3495 
3496 	return (0);
3497 }
3498 
3499 int
3500 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS)
3501 {
3502 	uma_zone_t zone = *(uma_zone_t *)arg1;
3503 	int cur;
3504 
3505 	cur = uma_zone_get_cur(zone);
3506 	return (sysctl_handle_int(oidp, &cur, 0, req));
3507 }
3508 
3509 #ifdef DDB
3510 DB_SHOW_COMMAND(uma, db_show_uma)
3511 {
3512 	uint64_t allocs, frees, sleeps;
3513 	uma_bucket_t bucket;
3514 	uma_keg_t kz;
3515 	uma_zone_t z;
3516 	int cachefree;
3517 
3518 	db_printf("%18s %8s %8s %8s %12s %8s %8s\n", "Zone", "Size", "Used",
3519 	    "Free", "Requests", "Sleeps", "Bucket");
3520 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3521 		LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3522 			if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
3523 				allocs = z->uz_allocs;
3524 				frees = z->uz_frees;
3525 				sleeps = z->uz_sleeps;
3526 				cachefree = 0;
3527 			} else
3528 				uma_zone_sumstat(z, &cachefree, &allocs,
3529 				    &frees, &sleeps);
3530 			if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
3531 			    (LIST_FIRST(&kz->uk_zones) != z)))
3532 				cachefree += kz->uk_free;
3533 			LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3534 				cachefree += bucket->ub_cnt;
3535 			db_printf("%18s %8ju %8jd %8d %12ju %8ju %8u\n",
3536 			    z->uz_name, (uintmax_t)kz->uk_size,
3537 			    (intmax_t)(allocs - frees), cachefree,
3538 			    (uintmax_t)allocs, sleeps, z->uz_count);
3539 			if (db_pager_quit)
3540 				return;
3541 		}
3542 	}
3543 }
3544 
3545 DB_SHOW_COMMAND(umacache, db_show_umacache)
3546 {
3547 	uint64_t allocs, frees;
3548 	uma_bucket_t bucket;
3549 	uma_zone_t z;
3550 	int cachefree;
3551 
3552 	db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
3553 	    "Requests", "Bucket");
3554 	LIST_FOREACH(z, &uma_cachezones, uz_link) {
3555 		uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL);
3556 		LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3557 			cachefree += bucket->ub_cnt;
3558 		db_printf("%18s %8ju %8jd %8d %12ju %8u\n",
3559 		    z->uz_name, (uintmax_t)z->uz_size,
3560 		    (intmax_t)(allocs - frees), cachefree,
3561 		    (uintmax_t)allocs, z->uz_count);
3562 		if (db_pager_quit)
3563 			return;
3564 	}
3565 }
3566 #endif
3567