xref: /freebsd/sys/vm/uma_core.c (revision a812392203d7c4c3f0db9d8a0f3391374c49c71f)
1 /*-
2  * Copyright (c) 2002-2005, 2009, 2013 Jeffrey Roberson <jeff@FreeBSD.org>
3  * Copyright (c) 2004, 2005 Bosko Milekic <bmilekic@FreeBSD.org>
4  * Copyright (c) 2004-2006 Robert N. M. Watson
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice unmodified, this list of conditions, and the following
12  *    disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
18  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
19  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
20  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
21  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
22  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
23  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
24  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
25  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
26  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
27  */
28 
29 /*
30  * uma_core.c  Implementation of the Universal Memory allocator
31  *
32  * This allocator is intended to replace the multitude of similar object caches
33  * in the standard FreeBSD kernel.  The intent is to be flexible as well as
34  * effecient.  A primary design goal is to return unused memory to the rest of
35  * the system.  This will make the system as a whole more flexible due to the
36  * ability to move memory to subsystems which most need it instead of leaving
37  * pools of reserved memory unused.
38  *
39  * The basic ideas stem from similar slab/zone based allocators whose algorithms
40  * are well known.
41  *
42  */
43 
44 /*
45  * TODO:
46  *	- Improve memory usage for large allocations
47  *	- Investigate cache size adjustments
48  */
49 
50 #include <sys/cdefs.h>
51 __FBSDID("$FreeBSD$");
52 
53 /* I should really use ktr.. */
54 /*
55 #define UMA_DEBUG 1
56 #define UMA_DEBUG_ALLOC 1
57 #define UMA_DEBUG_ALLOC_1 1
58 */
59 
60 #include "opt_ddb.h"
61 #include "opt_param.h"
62 #include "opt_vm.h"
63 
64 #include <sys/param.h>
65 #include <sys/systm.h>
66 #include <sys/bitset.h>
67 #include <sys/kernel.h>
68 #include <sys/types.h>
69 #include <sys/queue.h>
70 #include <sys/malloc.h>
71 #include <sys/ktr.h>
72 #include <sys/lock.h>
73 #include <sys/sysctl.h>
74 #include <sys/mutex.h>
75 #include <sys/proc.h>
76 #include <sys/random.h>
77 #include <sys/rwlock.h>
78 #include <sys/sbuf.h>
79 #include <sys/sched.h>
80 #include <sys/smp.h>
81 #include <sys/vmmeter.h>
82 
83 #include <vm/vm.h>
84 #include <vm/vm_object.h>
85 #include <vm/vm_page.h>
86 #include <vm/vm_pageout.h>
87 #include <vm/vm_param.h>
88 #include <vm/vm_map.h>
89 #include <vm/vm_kern.h>
90 #include <vm/vm_extern.h>
91 #include <vm/uma.h>
92 #include <vm/uma_int.h>
93 #include <vm/uma_dbg.h>
94 
95 #include <ddb/ddb.h>
96 
97 #ifdef DEBUG_MEMGUARD
98 #include <vm/memguard.h>
99 #endif
100 
101 /*
102  * This is the zone and keg from which all zones are spawned.  The idea is that
103  * even the zone & keg heads are allocated from the allocator, so we use the
104  * bss section to bootstrap us.
105  */
106 static struct uma_keg masterkeg;
107 static struct uma_zone masterzone_k;
108 static struct uma_zone masterzone_z;
109 static uma_zone_t kegs = &masterzone_k;
110 static uma_zone_t zones = &masterzone_z;
111 
112 /* This is the zone from which all of uma_slab_t's are allocated. */
113 static uma_zone_t slabzone;
114 static uma_zone_t slabrefzone;	/* With refcounters (for UMA_ZONE_REFCNT) */
115 
116 /*
117  * The initial hash tables come out of this zone so they can be allocated
118  * prior to malloc coming up.
119  */
120 static uma_zone_t hashzone;
121 
122 /* The boot-time adjusted value for cache line alignment. */
123 int uma_align_cache = 64 - 1;
124 
125 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets");
126 
127 /*
128  * Are we allowed to allocate buckets?
129  */
130 static int bucketdisable = 1;
131 
132 /* Linked list of all kegs in the system */
133 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs);
134 
135 /* Linked list of all cache-only zones in the system */
136 static LIST_HEAD(,uma_zone) uma_cachezones =
137     LIST_HEAD_INITIALIZER(uma_cachezones);
138 
139 /* This RW lock protects the keg list */
140 static struct rwlock_padalign uma_rwlock;
141 
142 /* Linked list of boot time pages */
143 static LIST_HEAD(,uma_slab) uma_boot_pages =
144     LIST_HEAD_INITIALIZER(uma_boot_pages);
145 
146 /* This mutex protects the boot time pages list */
147 static struct mtx_padalign uma_boot_pages_mtx;
148 
149 static struct sx uma_drain_lock;
150 
151 /* Is the VM done starting up? */
152 static int booted = 0;
153 #define	UMA_STARTUP	1
154 #define	UMA_STARTUP2	2
155 
156 /*
157  * Only mbuf clusters use ref zones.  Just provide enough references
158  * to support the one user.  New code should not use the ref facility.
159  */
160 static const u_int uma_max_ipers_ref = PAGE_SIZE / MCLBYTES;
161 
162 /*
163  * This is the handle used to schedule events that need to happen
164  * outside of the allocation fast path.
165  */
166 static struct callout uma_callout;
167 #define	UMA_TIMEOUT	20		/* Seconds for callout interval. */
168 
169 /*
170  * This structure is passed as the zone ctor arg so that I don't have to create
171  * a special allocation function just for zones.
172  */
173 struct uma_zctor_args {
174 	const char *name;
175 	size_t size;
176 	uma_ctor ctor;
177 	uma_dtor dtor;
178 	uma_init uminit;
179 	uma_fini fini;
180 	uma_import import;
181 	uma_release release;
182 	void *arg;
183 	uma_keg_t keg;
184 	int align;
185 	uint32_t flags;
186 };
187 
188 struct uma_kctor_args {
189 	uma_zone_t zone;
190 	size_t size;
191 	uma_init uminit;
192 	uma_fini fini;
193 	int align;
194 	uint32_t flags;
195 };
196 
197 struct uma_bucket_zone {
198 	uma_zone_t	ubz_zone;
199 	char		*ubz_name;
200 	int		ubz_entries;	/* Number of items it can hold. */
201 	int		ubz_maxsize;	/* Maximum allocation size per-item. */
202 };
203 
204 /*
205  * Compute the actual number of bucket entries to pack them in power
206  * of two sizes for more efficient space utilization.
207  */
208 #define	BUCKET_SIZE(n)						\
209     (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *))
210 
211 #define	BUCKET_MAX	BUCKET_SIZE(256)
212 
213 struct uma_bucket_zone bucket_zones[] = {
214 	{ NULL, "4 Bucket", BUCKET_SIZE(4), 4096 },
215 	{ NULL, "6 Bucket", BUCKET_SIZE(6), 3072 },
216 	{ NULL, "8 Bucket", BUCKET_SIZE(8), 2048 },
217 	{ NULL, "12 Bucket", BUCKET_SIZE(12), 1536 },
218 	{ NULL, "16 Bucket", BUCKET_SIZE(16), 1024 },
219 	{ NULL, "32 Bucket", BUCKET_SIZE(32), 512 },
220 	{ NULL, "64 Bucket", BUCKET_SIZE(64), 256 },
221 	{ NULL, "128 Bucket", BUCKET_SIZE(128), 128 },
222 	{ NULL, "256 Bucket", BUCKET_SIZE(256), 64 },
223 	{ NULL, NULL, 0}
224 };
225 
226 /*
227  * Flags and enumerations to be passed to internal functions.
228  */
229 enum zfreeskip { SKIP_NONE = 0, SKIP_DTOR, SKIP_FINI };
230 
231 /* Prototypes.. */
232 
233 static void *noobj_alloc(uma_zone_t, int, uint8_t *, int);
234 static void *page_alloc(uma_zone_t, int, uint8_t *, int);
235 static void *startup_alloc(uma_zone_t, int, uint8_t *, int);
236 static void page_free(void *, int, uint8_t);
237 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int);
238 static void cache_drain(uma_zone_t);
239 static void bucket_drain(uma_zone_t, uma_bucket_t);
240 static void bucket_cache_drain(uma_zone_t zone);
241 static int keg_ctor(void *, int, void *, int);
242 static void keg_dtor(void *, int, void *);
243 static int zone_ctor(void *, int, void *, int);
244 static void zone_dtor(void *, int, void *);
245 static int zero_init(void *, int, int);
246 static void keg_small_init(uma_keg_t keg);
247 static void keg_large_init(uma_keg_t keg);
248 static void zone_foreach(void (*zfunc)(uma_zone_t));
249 static void zone_timeout(uma_zone_t zone);
250 static int hash_alloc(struct uma_hash *);
251 static int hash_expand(struct uma_hash *, struct uma_hash *);
252 static void hash_free(struct uma_hash *hash);
253 static void uma_timeout(void *);
254 static void uma_startup3(void);
255 static void *zone_alloc_item(uma_zone_t, void *, int);
256 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip);
257 static void bucket_enable(void);
258 static void bucket_init(void);
259 static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int);
260 static void bucket_free(uma_zone_t zone, uma_bucket_t, void *);
261 static void bucket_zone_drain(void);
262 static uma_bucket_t zone_alloc_bucket(uma_zone_t zone, void *, int flags);
263 static uma_slab_t zone_fetch_slab(uma_zone_t zone, uma_keg_t last, int flags);
264 static uma_slab_t zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int flags);
265 static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab);
266 static void slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item);
267 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit,
268     uma_fini fini, int align, uint32_t flags);
269 static int zone_import(uma_zone_t zone, void **bucket, int max, int flags);
270 static void zone_release(uma_zone_t zone, void **bucket, int cnt);
271 static void uma_zero_item(void *item, uma_zone_t zone);
272 
273 void uma_print_zone(uma_zone_t);
274 void uma_print_stats(void);
275 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS);
276 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS);
277 
278 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
279 
280 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLTYPE_INT,
281     0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones");
282 
283 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLTYPE_STRUCT,
284     0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats");
285 
286 static int zone_warnings = 1;
287 SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RWTUN, &zone_warnings, 0,
288     "Warn when UMA zones becomes full");
289 
290 /*
291  * This routine checks to see whether or not it's safe to enable buckets.
292  */
293 static void
294 bucket_enable(void)
295 {
296 	bucketdisable = vm_page_count_min();
297 }
298 
299 /*
300  * Initialize bucket_zones, the array of zones of buckets of various sizes.
301  *
302  * For each zone, calculate the memory required for each bucket, consisting
303  * of the header and an array of pointers.
304  */
305 static void
306 bucket_init(void)
307 {
308 	struct uma_bucket_zone *ubz;
309 	int size;
310 	int i;
311 
312 	for (i = 0, ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) {
313 		size = roundup(sizeof(struct uma_bucket), sizeof(void *));
314 		size += sizeof(void *) * ubz->ubz_entries;
315 		ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
316 		    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
317 		    UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET);
318 	}
319 }
320 
321 /*
322  * Given a desired number of entries for a bucket, return the zone from which
323  * to allocate the bucket.
324  */
325 static struct uma_bucket_zone *
326 bucket_zone_lookup(int entries)
327 {
328 	struct uma_bucket_zone *ubz;
329 
330 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
331 		if (ubz->ubz_entries >= entries)
332 			return (ubz);
333 	ubz--;
334 	return (ubz);
335 }
336 
337 static int
338 bucket_select(int size)
339 {
340 	struct uma_bucket_zone *ubz;
341 
342 	ubz = &bucket_zones[0];
343 	if (size > ubz->ubz_maxsize)
344 		return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1);
345 
346 	for (; ubz->ubz_entries != 0; ubz++)
347 		if (ubz->ubz_maxsize < size)
348 			break;
349 	ubz--;
350 	return (ubz->ubz_entries);
351 }
352 
353 static uma_bucket_t
354 bucket_alloc(uma_zone_t zone, void *udata, int flags)
355 {
356 	struct uma_bucket_zone *ubz;
357 	uma_bucket_t bucket;
358 
359 	/*
360 	 * This is to stop us from allocating per cpu buckets while we're
361 	 * running out of vm.boot_pages.  Otherwise, we would exhaust the
362 	 * boot pages.  This also prevents us from allocating buckets in
363 	 * low memory situations.
364 	 */
365 	if (bucketdisable)
366 		return (NULL);
367 	/*
368 	 * To limit bucket recursion we store the original zone flags
369 	 * in a cookie passed via zalloc_arg/zfree_arg.  This allows the
370 	 * NOVM flag to persist even through deep recursions.  We also
371 	 * store ZFLAG_BUCKET once we have recursed attempting to allocate
372 	 * a bucket for a bucket zone so we do not allow infinite bucket
373 	 * recursion.  This cookie will even persist to frees of unused
374 	 * buckets via the allocation path or bucket allocations in the
375 	 * free path.
376 	 */
377 	if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
378 		udata = (void *)(uintptr_t)zone->uz_flags;
379 	else {
380 		if ((uintptr_t)udata & UMA_ZFLAG_BUCKET)
381 			return (NULL);
382 		udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET);
383 	}
384 	if ((uintptr_t)udata & UMA_ZFLAG_CACHEONLY)
385 		flags |= M_NOVM;
386 	ubz = bucket_zone_lookup(zone->uz_count);
387 	if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0)
388 		ubz++;
389 	bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags);
390 	if (bucket) {
391 #ifdef INVARIANTS
392 		bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
393 #endif
394 		bucket->ub_cnt = 0;
395 		bucket->ub_entries = ubz->ubz_entries;
396 	}
397 
398 	return (bucket);
399 }
400 
401 static void
402 bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata)
403 {
404 	struct uma_bucket_zone *ubz;
405 
406 	KASSERT(bucket->ub_cnt == 0,
407 	    ("bucket_free: Freeing a non free bucket."));
408 	if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
409 		udata = (void *)(uintptr_t)zone->uz_flags;
410 	ubz = bucket_zone_lookup(bucket->ub_entries);
411 	uma_zfree_arg(ubz->ubz_zone, bucket, udata);
412 }
413 
414 static void
415 bucket_zone_drain(void)
416 {
417 	struct uma_bucket_zone *ubz;
418 
419 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
420 		zone_drain(ubz->ubz_zone);
421 }
422 
423 static void
424 zone_log_warning(uma_zone_t zone)
425 {
426 	static const struct timeval warninterval = { 300, 0 };
427 
428 	if (!zone_warnings || zone->uz_warning == NULL)
429 		return;
430 
431 	if (ratecheck(&zone->uz_ratecheck, &warninterval))
432 		printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning);
433 }
434 
435 static void
436 zone_foreach_keg(uma_zone_t zone, void (*kegfn)(uma_keg_t))
437 {
438 	uma_klink_t klink;
439 
440 	LIST_FOREACH(klink, &zone->uz_kegs, kl_link)
441 		kegfn(klink->kl_keg);
442 }
443 
444 /*
445  * Routine called by timeout which is used to fire off some time interval
446  * based calculations.  (stats, hash size, etc.)
447  *
448  * Arguments:
449  *	arg   Unused
450  *
451  * Returns:
452  *	Nothing
453  */
454 static void
455 uma_timeout(void *unused)
456 {
457 	bucket_enable();
458 	zone_foreach(zone_timeout);
459 
460 	/* Reschedule this event */
461 	callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
462 }
463 
464 /*
465  * Routine to perform timeout driven calculations.  This expands the
466  * hashes and does per cpu statistics aggregation.
467  *
468  *  Returns nothing.
469  */
470 static void
471 keg_timeout(uma_keg_t keg)
472 {
473 
474 	KEG_LOCK(keg);
475 	/*
476 	 * Expand the keg hash table.
477 	 *
478 	 * This is done if the number of slabs is larger than the hash size.
479 	 * What I'm trying to do here is completely reduce collisions.  This
480 	 * may be a little aggressive.  Should I allow for two collisions max?
481 	 */
482 	if (keg->uk_flags & UMA_ZONE_HASH &&
483 	    keg->uk_pages / keg->uk_ppera >= keg->uk_hash.uh_hashsize) {
484 		struct uma_hash newhash;
485 		struct uma_hash oldhash;
486 		int ret;
487 
488 		/*
489 		 * This is so involved because allocating and freeing
490 		 * while the keg lock is held will lead to deadlock.
491 		 * I have to do everything in stages and check for
492 		 * races.
493 		 */
494 		newhash = keg->uk_hash;
495 		KEG_UNLOCK(keg);
496 		ret = hash_alloc(&newhash);
497 		KEG_LOCK(keg);
498 		if (ret) {
499 			if (hash_expand(&keg->uk_hash, &newhash)) {
500 				oldhash = keg->uk_hash;
501 				keg->uk_hash = newhash;
502 			} else
503 				oldhash = newhash;
504 
505 			KEG_UNLOCK(keg);
506 			hash_free(&oldhash);
507 			return;
508 		}
509 	}
510 	KEG_UNLOCK(keg);
511 }
512 
513 static void
514 zone_timeout(uma_zone_t zone)
515 {
516 
517 	zone_foreach_keg(zone, &keg_timeout);
518 }
519 
520 /*
521  * Allocate and zero fill the next sized hash table from the appropriate
522  * backing store.
523  *
524  * Arguments:
525  *	hash  A new hash structure with the old hash size in uh_hashsize
526  *
527  * Returns:
528  *	1 on sucess and 0 on failure.
529  */
530 static int
531 hash_alloc(struct uma_hash *hash)
532 {
533 	int oldsize;
534 	int alloc;
535 
536 	oldsize = hash->uh_hashsize;
537 
538 	/* We're just going to go to a power of two greater */
539 	if (oldsize)  {
540 		hash->uh_hashsize = oldsize * 2;
541 		alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
542 		hash->uh_slab_hash = (struct slabhead *)malloc(alloc,
543 		    M_UMAHASH, M_NOWAIT);
544 	} else {
545 		alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
546 		hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
547 		    M_WAITOK);
548 		hash->uh_hashsize = UMA_HASH_SIZE_INIT;
549 	}
550 	if (hash->uh_slab_hash) {
551 		bzero(hash->uh_slab_hash, alloc);
552 		hash->uh_hashmask = hash->uh_hashsize - 1;
553 		return (1);
554 	}
555 
556 	return (0);
557 }
558 
559 /*
560  * Expands the hash table for HASH zones.  This is done from zone_timeout
561  * to reduce collisions.  This must not be done in the regular allocation
562  * path, otherwise, we can recurse on the vm while allocating pages.
563  *
564  * Arguments:
565  *	oldhash  The hash you want to expand
566  *	newhash  The hash structure for the new table
567  *
568  * Returns:
569  *	Nothing
570  *
571  * Discussion:
572  */
573 static int
574 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
575 {
576 	uma_slab_t slab;
577 	int hval;
578 	int i;
579 
580 	if (!newhash->uh_slab_hash)
581 		return (0);
582 
583 	if (oldhash->uh_hashsize >= newhash->uh_hashsize)
584 		return (0);
585 
586 	/*
587 	 * I need to investigate hash algorithms for resizing without a
588 	 * full rehash.
589 	 */
590 
591 	for (i = 0; i < oldhash->uh_hashsize; i++)
592 		while (!SLIST_EMPTY(&oldhash->uh_slab_hash[i])) {
593 			slab = SLIST_FIRST(&oldhash->uh_slab_hash[i]);
594 			SLIST_REMOVE_HEAD(&oldhash->uh_slab_hash[i], us_hlink);
595 			hval = UMA_HASH(newhash, slab->us_data);
596 			SLIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
597 			    slab, us_hlink);
598 		}
599 
600 	return (1);
601 }
602 
603 /*
604  * Free the hash bucket to the appropriate backing store.
605  *
606  * Arguments:
607  *	slab_hash  The hash bucket we're freeing
608  *	hashsize   The number of entries in that hash bucket
609  *
610  * Returns:
611  *	Nothing
612  */
613 static void
614 hash_free(struct uma_hash *hash)
615 {
616 	if (hash->uh_slab_hash == NULL)
617 		return;
618 	if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
619 		zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE);
620 	else
621 		free(hash->uh_slab_hash, M_UMAHASH);
622 }
623 
624 /*
625  * Frees all outstanding items in a bucket
626  *
627  * Arguments:
628  *	zone   The zone to free to, must be unlocked.
629  *	bucket The free/alloc bucket with items, cpu queue must be locked.
630  *
631  * Returns:
632  *	Nothing
633  */
634 
635 static void
636 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
637 {
638 	int i;
639 
640 	if (bucket == NULL)
641 		return;
642 
643 	if (zone->uz_fini)
644 		for (i = 0; i < bucket->ub_cnt; i++)
645 			zone->uz_fini(bucket->ub_bucket[i], zone->uz_size);
646 	zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt);
647 	bucket->ub_cnt = 0;
648 }
649 
650 /*
651  * Drains the per cpu caches for a zone.
652  *
653  * NOTE: This may only be called while the zone is being turn down, and not
654  * during normal operation.  This is necessary in order that we do not have
655  * to migrate CPUs to drain the per-CPU caches.
656  *
657  * Arguments:
658  *	zone     The zone to drain, must be unlocked.
659  *
660  * Returns:
661  *	Nothing
662  */
663 static void
664 cache_drain(uma_zone_t zone)
665 {
666 	uma_cache_t cache;
667 	int cpu;
668 
669 	/*
670 	 * XXX: It is safe to not lock the per-CPU caches, because we're
671 	 * tearing down the zone anyway.  I.e., there will be no further use
672 	 * of the caches at this point.
673 	 *
674 	 * XXX: It would good to be able to assert that the zone is being
675 	 * torn down to prevent improper use of cache_drain().
676 	 *
677 	 * XXX: We lock the zone before passing into bucket_cache_drain() as
678 	 * it is used elsewhere.  Should the tear-down path be made special
679 	 * there in some form?
680 	 */
681 	CPU_FOREACH(cpu) {
682 		cache = &zone->uz_cpu[cpu];
683 		bucket_drain(zone, cache->uc_allocbucket);
684 		bucket_drain(zone, cache->uc_freebucket);
685 		if (cache->uc_allocbucket != NULL)
686 			bucket_free(zone, cache->uc_allocbucket, NULL);
687 		if (cache->uc_freebucket != NULL)
688 			bucket_free(zone, cache->uc_freebucket, NULL);
689 		cache->uc_allocbucket = cache->uc_freebucket = NULL;
690 	}
691 	ZONE_LOCK(zone);
692 	bucket_cache_drain(zone);
693 	ZONE_UNLOCK(zone);
694 }
695 
696 static void
697 cache_shrink(uma_zone_t zone)
698 {
699 
700 	if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
701 		return;
702 
703 	ZONE_LOCK(zone);
704 	zone->uz_count = (zone->uz_count_min + zone->uz_count) / 2;
705 	ZONE_UNLOCK(zone);
706 }
707 
708 static void
709 cache_drain_safe_cpu(uma_zone_t zone)
710 {
711 	uma_cache_t cache;
712 	uma_bucket_t b1, b2;
713 
714 	if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
715 		return;
716 
717 	b1 = b2 = NULL;
718 	ZONE_LOCK(zone);
719 	critical_enter();
720 	cache = &zone->uz_cpu[curcpu];
721 	if (cache->uc_allocbucket) {
722 		if (cache->uc_allocbucket->ub_cnt != 0)
723 			LIST_INSERT_HEAD(&zone->uz_buckets,
724 			    cache->uc_allocbucket, ub_link);
725 		else
726 			b1 = cache->uc_allocbucket;
727 		cache->uc_allocbucket = NULL;
728 	}
729 	if (cache->uc_freebucket) {
730 		if (cache->uc_freebucket->ub_cnt != 0)
731 			LIST_INSERT_HEAD(&zone->uz_buckets,
732 			    cache->uc_freebucket, ub_link);
733 		else
734 			b2 = cache->uc_freebucket;
735 		cache->uc_freebucket = NULL;
736 	}
737 	critical_exit();
738 	ZONE_UNLOCK(zone);
739 	if (b1)
740 		bucket_free(zone, b1, NULL);
741 	if (b2)
742 		bucket_free(zone, b2, NULL);
743 }
744 
745 /*
746  * Safely drain per-CPU caches of a zone(s) to alloc bucket.
747  * This is an expensive call because it needs to bind to all CPUs
748  * one by one and enter a critical section on each of them in order
749  * to safely access their cache buckets.
750  * Zone lock must not be held on call this function.
751  */
752 static void
753 cache_drain_safe(uma_zone_t zone)
754 {
755 	int cpu;
756 
757 	/*
758 	 * Polite bucket sizes shrinking was not enouth, shrink aggressively.
759 	 */
760 	if (zone)
761 		cache_shrink(zone);
762 	else
763 		zone_foreach(cache_shrink);
764 
765 	CPU_FOREACH(cpu) {
766 		thread_lock(curthread);
767 		sched_bind(curthread, cpu);
768 		thread_unlock(curthread);
769 
770 		if (zone)
771 			cache_drain_safe_cpu(zone);
772 		else
773 			zone_foreach(cache_drain_safe_cpu);
774 	}
775 	thread_lock(curthread);
776 	sched_unbind(curthread);
777 	thread_unlock(curthread);
778 }
779 
780 /*
781  * Drain the cached buckets from a zone.  Expects a locked zone on entry.
782  */
783 static void
784 bucket_cache_drain(uma_zone_t zone)
785 {
786 	uma_bucket_t bucket;
787 
788 	/*
789 	 * Drain the bucket queues and free the buckets, we just keep two per
790 	 * cpu (alloc/free).
791 	 */
792 	while ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) {
793 		LIST_REMOVE(bucket, ub_link);
794 		ZONE_UNLOCK(zone);
795 		bucket_drain(zone, bucket);
796 		bucket_free(zone, bucket, NULL);
797 		ZONE_LOCK(zone);
798 	}
799 
800 	/*
801 	 * Shrink further bucket sizes.  Price of single zone lock collision
802 	 * is probably lower then price of global cache drain.
803 	 */
804 	if (zone->uz_count > zone->uz_count_min)
805 		zone->uz_count--;
806 }
807 
808 static void
809 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start)
810 {
811 	uint8_t *mem;
812 	int i;
813 	uint8_t flags;
814 
815 	mem = slab->us_data;
816 	flags = slab->us_flags;
817 	i = start;
818 	if (keg->uk_fini != NULL) {
819 		for (i--; i > -1; i--)
820 			keg->uk_fini(slab->us_data + (keg->uk_rsize * i),
821 			    keg->uk_size);
822 	}
823 	if (keg->uk_flags & UMA_ZONE_OFFPAGE)
824 		zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE);
825 #ifdef UMA_DEBUG
826 	printf("%s: Returning %d bytes.\n", keg->uk_name,
827 	    PAGE_SIZE * keg->uk_ppera);
828 #endif
829 	keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags);
830 }
831 
832 /*
833  * Frees pages from a keg back to the system.  This is done on demand from
834  * the pageout daemon.
835  *
836  * Returns nothing.
837  */
838 static void
839 keg_drain(uma_keg_t keg)
840 {
841 	struct slabhead freeslabs = { 0 };
842 	uma_slab_t slab;
843 	uma_slab_t n;
844 
845 	/*
846 	 * We don't want to take pages from statically allocated kegs at this
847 	 * time
848 	 */
849 	if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL)
850 		return;
851 
852 #ifdef UMA_DEBUG
853 	printf("%s free items: %u\n", keg->uk_name, keg->uk_free);
854 #endif
855 	KEG_LOCK(keg);
856 	if (keg->uk_free == 0)
857 		goto finished;
858 
859 	slab = LIST_FIRST(&keg->uk_free_slab);
860 	while (slab) {
861 		n = LIST_NEXT(slab, us_link);
862 
863 		/* We have no where to free these to */
864 		if (slab->us_flags & UMA_SLAB_BOOT) {
865 			slab = n;
866 			continue;
867 		}
868 
869 		LIST_REMOVE(slab, us_link);
870 		keg->uk_pages -= keg->uk_ppera;
871 		keg->uk_free -= keg->uk_ipers;
872 
873 		if (keg->uk_flags & UMA_ZONE_HASH)
874 			UMA_HASH_REMOVE(&keg->uk_hash, slab, slab->us_data);
875 
876 		SLIST_INSERT_HEAD(&freeslabs, slab, us_hlink);
877 
878 		slab = n;
879 	}
880 finished:
881 	KEG_UNLOCK(keg);
882 
883 	while ((slab = SLIST_FIRST(&freeslabs)) != NULL) {
884 		SLIST_REMOVE(&freeslabs, slab, uma_slab, us_hlink);
885 		keg_free_slab(keg, slab, keg->uk_ipers);
886 	}
887 }
888 
889 static void
890 zone_drain_wait(uma_zone_t zone, int waitok)
891 {
892 
893 	/*
894 	 * Set draining to interlock with zone_dtor() so we can release our
895 	 * locks as we go.  Only dtor() should do a WAITOK call since it
896 	 * is the only call that knows the structure will still be available
897 	 * when it wakes up.
898 	 */
899 	ZONE_LOCK(zone);
900 	while (zone->uz_flags & UMA_ZFLAG_DRAINING) {
901 		if (waitok == M_NOWAIT)
902 			goto out;
903 		msleep(zone, zone->uz_lockptr, PVM, "zonedrain", 1);
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_rwlock 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 	rw_wlock(&uma_rwlock);
1547 	LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
1548 	rw_wunlock(&uma_rwlock);
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 		rw_wlock(&uma_rwlock);
1599 		LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link);
1600 		rw_wunlock(&uma_rwlock);
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 		rw_wlock(&uma_rwlock);
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 		rw_wunlock(&uma_rwlock);
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 	rw_wlock(&uma_rwlock);
1725 	LIST_REMOVE(zone, uz_link);
1726 	rw_wunlock(&uma_rwlock);
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 		rw_wlock(&uma_rwlock);
1749 		LIST_REMOVE(keg, uk_link);
1750 		rw_wunlock(&uma_rwlock);
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 	rw_rlock(&uma_rwlock);
1773 	LIST_FOREACH(keg, &uma_kegs, uk_link) {
1774 		LIST_FOREACH(zone, &keg->uk_zones, uz_link)
1775 			zfunc(zone);
1776 	}
1777 	rw_runlock(&uma_rwlock);
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 	rw_init(&uma_rwlock, "UMA lock");
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("Creating slab and hash zones.\n");
1838 #endif
1839 
1840 	/* Now make a zone for slab headers */
1841 	slabzone = uma_zcreate("UMA Slabs",
1842 				sizeof(struct uma_slab),
1843 				NULL, NULL, NULL, NULL,
1844 				UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1845 
1846 	/*
1847 	 * We also create a zone for the bigger slabs with reference
1848 	 * counts in them, to accomodate UMA_ZONE_REFCNT zones.
1849 	 */
1850 	slabsize = sizeof(struct uma_slab_refcnt);
1851 	slabsize += uma_max_ipers_ref * sizeof(uint32_t);
1852 	slabrefzone = uma_zcreate("UMA RCntSlabs",
1853 				  slabsize,
1854 				  NULL, NULL, NULL, NULL,
1855 				  UMA_ALIGN_PTR,
1856 				  UMA_ZFLAG_INTERNAL);
1857 
1858 	hashzone = uma_zcreate("UMA Hash",
1859 	    sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
1860 	    NULL, NULL, NULL, NULL,
1861 	    UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1862 
1863 	bucket_init();
1864 
1865 	booted = UMA_STARTUP;
1866 
1867 #ifdef UMA_DEBUG
1868 	printf("UMA startup complete.\n");
1869 #endif
1870 }
1871 
1872 /* see uma.h */
1873 void
1874 uma_startup2(void)
1875 {
1876 	booted = UMA_STARTUP2;
1877 	bucket_enable();
1878 	sx_init(&uma_drain_lock, "umadrain");
1879 #ifdef UMA_DEBUG
1880 	printf("UMA startup2 complete.\n");
1881 #endif
1882 }
1883 
1884 /*
1885  * Initialize our callout handle
1886  *
1887  */
1888 
1889 static void
1890 uma_startup3(void)
1891 {
1892 #ifdef UMA_DEBUG
1893 	printf("Starting callout.\n");
1894 #endif
1895 	callout_init(&uma_callout, CALLOUT_MPSAFE);
1896 	callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
1897 #ifdef UMA_DEBUG
1898 	printf("UMA startup3 complete.\n");
1899 #endif
1900 }
1901 
1902 static uma_keg_t
1903 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
1904 		int align, uint32_t flags)
1905 {
1906 	struct uma_kctor_args args;
1907 
1908 	args.size = size;
1909 	args.uminit = uminit;
1910 	args.fini = fini;
1911 	args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
1912 	args.flags = flags;
1913 	args.zone = zone;
1914 	return (zone_alloc_item(kegs, &args, M_WAITOK));
1915 }
1916 
1917 /* See uma.h */
1918 void
1919 uma_set_align(int align)
1920 {
1921 
1922 	if (align != UMA_ALIGN_CACHE)
1923 		uma_align_cache = align;
1924 }
1925 
1926 /* See uma.h */
1927 uma_zone_t
1928 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
1929 		uma_init uminit, uma_fini fini, int align, uint32_t flags)
1930 
1931 {
1932 	struct uma_zctor_args args;
1933 	uma_zone_t res;
1934 	bool locked;
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 	if (booted < UMA_STARTUP2) {
1949 		locked = false;
1950 	} else {
1951 		sx_slock(&uma_drain_lock);
1952 		locked = true;
1953 	}
1954 	res = zone_alloc_item(zones, &args, M_WAITOK);
1955 	if (locked)
1956 		sx_sunlock(&uma_drain_lock);
1957 	return (res);
1958 }
1959 
1960 /* See uma.h */
1961 uma_zone_t
1962 uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor,
1963 		    uma_init zinit, uma_fini zfini, uma_zone_t master)
1964 {
1965 	struct uma_zctor_args args;
1966 	uma_keg_t keg;
1967 	uma_zone_t res;
1968 	bool locked;
1969 
1970 	keg = zone_first_keg(master);
1971 	memset(&args, 0, sizeof(args));
1972 	args.name = name;
1973 	args.size = keg->uk_size;
1974 	args.ctor = ctor;
1975 	args.dtor = dtor;
1976 	args.uminit = zinit;
1977 	args.fini = zfini;
1978 	args.align = keg->uk_align;
1979 	args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
1980 	args.keg = keg;
1981 
1982 	if (booted < UMA_STARTUP2) {
1983 		locked = false;
1984 	} else {
1985 		sx_slock(&uma_drain_lock);
1986 		locked = true;
1987 	}
1988 	/* XXX Attaches only one keg of potentially many. */
1989 	res = zone_alloc_item(zones, &args, M_WAITOK);
1990 	if (locked)
1991 		sx_sunlock(&uma_drain_lock);
1992 	return (res);
1993 }
1994 
1995 /* See uma.h */
1996 uma_zone_t
1997 uma_zcache_create(char *name, int size, uma_ctor ctor, uma_dtor dtor,
1998 		    uma_init zinit, uma_fini zfini, uma_import zimport,
1999 		    uma_release zrelease, void *arg, int flags)
2000 {
2001 	struct uma_zctor_args args;
2002 
2003 	memset(&args, 0, sizeof(args));
2004 	args.name = name;
2005 	args.size = size;
2006 	args.ctor = ctor;
2007 	args.dtor = dtor;
2008 	args.uminit = zinit;
2009 	args.fini = zfini;
2010 	args.import = zimport;
2011 	args.release = zrelease;
2012 	args.arg = arg;
2013 	args.align = 0;
2014 	args.flags = flags;
2015 
2016 	return (zone_alloc_item(zones, &args, M_WAITOK));
2017 }
2018 
2019 static void
2020 zone_lock_pair(uma_zone_t a, uma_zone_t b)
2021 {
2022 	if (a < b) {
2023 		ZONE_LOCK(a);
2024 		mtx_lock_flags(b->uz_lockptr, MTX_DUPOK);
2025 	} else {
2026 		ZONE_LOCK(b);
2027 		mtx_lock_flags(a->uz_lockptr, MTX_DUPOK);
2028 	}
2029 }
2030 
2031 static void
2032 zone_unlock_pair(uma_zone_t a, uma_zone_t b)
2033 {
2034 
2035 	ZONE_UNLOCK(a);
2036 	ZONE_UNLOCK(b);
2037 }
2038 
2039 int
2040 uma_zsecond_add(uma_zone_t zone, uma_zone_t master)
2041 {
2042 	uma_klink_t klink;
2043 	uma_klink_t kl;
2044 	int error;
2045 
2046 	error = 0;
2047 	klink = malloc(sizeof(*klink), M_TEMP, M_WAITOK | M_ZERO);
2048 
2049 	zone_lock_pair(zone, master);
2050 	/*
2051 	 * zone must use vtoslab() to resolve objects and must already be
2052 	 * a secondary.
2053 	 */
2054 	if ((zone->uz_flags & (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY))
2055 	    != (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) {
2056 		error = EINVAL;
2057 		goto out;
2058 	}
2059 	/*
2060 	 * The new master must also use vtoslab().
2061 	 */
2062 	if ((zone->uz_flags & UMA_ZONE_VTOSLAB) != UMA_ZONE_VTOSLAB) {
2063 		error = EINVAL;
2064 		goto out;
2065 	}
2066 	/*
2067 	 * Both must either be refcnt, or not be refcnt.
2068 	 */
2069 	if ((zone->uz_flags & UMA_ZONE_REFCNT) !=
2070 	    (master->uz_flags & UMA_ZONE_REFCNT)) {
2071 		error = EINVAL;
2072 		goto out;
2073 	}
2074 	/*
2075 	 * The underlying object must be the same size.  rsize
2076 	 * may be different.
2077 	 */
2078 	if (master->uz_size != zone->uz_size) {
2079 		error = E2BIG;
2080 		goto out;
2081 	}
2082 	/*
2083 	 * Put it at the end of the list.
2084 	 */
2085 	klink->kl_keg = zone_first_keg(master);
2086 	LIST_FOREACH(kl, &zone->uz_kegs, kl_link) {
2087 		if (LIST_NEXT(kl, kl_link) == NULL) {
2088 			LIST_INSERT_AFTER(kl, klink, kl_link);
2089 			break;
2090 		}
2091 	}
2092 	klink = NULL;
2093 	zone->uz_flags |= UMA_ZFLAG_MULTI;
2094 	zone->uz_slab = zone_fetch_slab_multi;
2095 
2096 out:
2097 	zone_unlock_pair(zone, master);
2098 	if (klink != NULL)
2099 		free(klink, M_TEMP);
2100 
2101 	return (error);
2102 }
2103 
2104 
2105 /* See uma.h */
2106 void
2107 uma_zdestroy(uma_zone_t zone)
2108 {
2109 
2110 	sx_slock(&uma_drain_lock);
2111 	zone_free_item(zones, zone, NULL, SKIP_NONE);
2112 	sx_sunlock(&uma_drain_lock);
2113 }
2114 
2115 /* See uma.h */
2116 void *
2117 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
2118 {
2119 	void *item;
2120 	uma_cache_t cache;
2121 	uma_bucket_t bucket;
2122 	int lockfail;
2123 	int cpu;
2124 
2125 #if 0
2126 	/* XXX: FIX!! Do not enable this in CURRENT!! MarkM */
2127 	/* The entropy here is desirable, but the harvesting is expensive */
2128 	random_harvest(&(zone->uz_name), sizeof(void *), 1, RANDOM_UMA_ALLOC);
2129 #endif
2130 
2131 	/* This is the fast path allocation */
2132 #ifdef UMA_DEBUG_ALLOC_1
2133 	printf("Allocating one item from %s(%p)\n", zone->uz_name, zone);
2134 #endif
2135 	CTR3(KTR_UMA, "uma_zalloc_arg thread %x zone %s flags %d", curthread,
2136 	    zone->uz_name, flags);
2137 
2138 	if (flags & M_WAITOK) {
2139 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
2140 		    "uma_zalloc_arg: zone \"%s\"", zone->uz_name);
2141 	}
2142 #ifdef DEBUG_MEMGUARD
2143 	if (memguard_cmp_zone(zone)) {
2144 		item = memguard_alloc(zone->uz_size, flags);
2145 		if (item != NULL) {
2146 			/*
2147 			 * Avoid conflict with the use-after-free
2148 			 * protecting infrastructure from INVARIANTS.
2149 			 */
2150 			if (zone->uz_init != NULL &&
2151 			    zone->uz_init != mtrash_init &&
2152 			    zone->uz_init(item, zone->uz_size, flags) != 0)
2153 				return (NULL);
2154 			if (zone->uz_ctor != NULL &&
2155 			    zone->uz_ctor != mtrash_ctor &&
2156 			    zone->uz_ctor(item, zone->uz_size, udata,
2157 			    flags) != 0) {
2158 			    	zone->uz_fini(item, zone->uz_size);
2159 				return (NULL);
2160 			}
2161 #if 0
2162 			/* XXX: FIX!! Do not enable this in CURRENT!! MarkM */
2163 			/* The entropy here is desirable, but the harvesting is expensive */
2164 			random_harvest(&item, sizeof(void *), 1, RANDOM_UMA_ALLOC);
2165 #endif
2166 			return (item);
2167 		}
2168 		/* This is unfortunate but should not be fatal. */
2169 	}
2170 #endif
2171 	/*
2172 	 * If possible, allocate from the per-CPU cache.  There are two
2173 	 * requirements for safe access to the per-CPU cache: (1) the thread
2174 	 * accessing the cache must not be preempted or yield during access,
2175 	 * and (2) the thread must not migrate CPUs without switching which
2176 	 * cache it accesses.  We rely on a critical section to prevent
2177 	 * preemption and migration.  We release the critical section in
2178 	 * order to acquire the zone mutex if we are unable to allocate from
2179 	 * the current cache; when we re-acquire the critical section, we
2180 	 * must detect and handle migration if it has occurred.
2181 	 */
2182 	critical_enter();
2183 	cpu = curcpu;
2184 	cache = &zone->uz_cpu[cpu];
2185 
2186 zalloc_start:
2187 	bucket = cache->uc_allocbucket;
2188 	if (bucket != NULL && bucket->ub_cnt > 0) {
2189 		bucket->ub_cnt--;
2190 		item = bucket->ub_bucket[bucket->ub_cnt];
2191 #ifdef INVARIANTS
2192 		bucket->ub_bucket[bucket->ub_cnt] = NULL;
2193 #endif
2194 		KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled."));
2195 		cache->uc_allocs++;
2196 		critical_exit();
2197 		if (zone->uz_ctor != NULL &&
2198 		    zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2199 			atomic_add_long(&zone->uz_fails, 1);
2200 			zone_free_item(zone, item, udata, SKIP_DTOR);
2201 			return (NULL);
2202 		}
2203 #ifdef INVARIANTS
2204 		uma_dbg_alloc(zone, NULL, item);
2205 #endif
2206 		if (flags & M_ZERO)
2207 			uma_zero_item(item, zone);
2208 #if 0
2209 		/* XXX: FIX!! Do not enable this in CURRENT!! MarkM */
2210 		/* The entropy here is desirable, but the harvesting is expensive */
2211 		random_harvest(&item, sizeof(void *), 1, RANDOM_UMA_ALLOC);
2212 #endif
2213 		return (item);
2214 	}
2215 
2216 	/*
2217 	 * We have run out of items in our alloc bucket.
2218 	 * See if we can switch with our free bucket.
2219 	 */
2220 	bucket = cache->uc_freebucket;
2221 	if (bucket != NULL && bucket->ub_cnt > 0) {
2222 #ifdef UMA_DEBUG_ALLOC
2223 		printf("uma_zalloc: Swapping empty with alloc.\n");
2224 #endif
2225 		cache->uc_freebucket = cache->uc_allocbucket;
2226 		cache->uc_allocbucket = bucket;
2227 		goto zalloc_start;
2228 	}
2229 
2230 	/*
2231 	 * Discard any empty allocation bucket while we hold no locks.
2232 	 */
2233 	bucket = cache->uc_allocbucket;
2234 	cache->uc_allocbucket = NULL;
2235 	critical_exit();
2236 	if (bucket != NULL)
2237 		bucket_free(zone, bucket, udata);
2238 
2239 	/* Short-circuit for zones without buckets and low memory. */
2240 	if (zone->uz_count == 0 || bucketdisable)
2241 		goto zalloc_item;
2242 
2243 	/*
2244 	 * Attempt to retrieve the item from the per-CPU cache has failed, so
2245 	 * we must go back to the zone.  This requires the zone lock, so we
2246 	 * must drop the critical section, then re-acquire it when we go back
2247 	 * to the cache.  Since the critical section is released, we may be
2248 	 * preempted or migrate.  As such, make sure not to maintain any
2249 	 * thread-local state specific to the cache from prior to releasing
2250 	 * the critical section.
2251 	 */
2252 	lockfail = 0;
2253 	if (ZONE_TRYLOCK(zone) == 0) {
2254 		/* Record contention to size the buckets. */
2255 		ZONE_LOCK(zone);
2256 		lockfail = 1;
2257 	}
2258 	critical_enter();
2259 	cpu = curcpu;
2260 	cache = &zone->uz_cpu[cpu];
2261 
2262 	/*
2263 	 * Since we have locked the zone we may as well send back our stats.
2264 	 */
2265 	atomic_add_long(&zone->uz_allocs, cache->uc_allocs);
2266 	atomic_add_long(&zone->uz_frees, cache->uc_frees);
2267 	cache->uc_allocs = 0;
2268 	cache->uc_frees = 0;
2269 
2270 	/* See if we lost the race to fill the cache. */
2271 	if (cache->uc_allocbucket != NULL) {
2272 		ZONE_UNLOCK(zone);
2273 		goto zalloc_start;
2274 	}
2275 
2276 	/*
2277 	 * Check the zone's cache of buckets.
2278 	 */
2279 	if ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) {
2280 		KASSERT(bucket->ub_cnt != 0,
2281 		    ("uma_zalloc_arg: Returning an empty bucket."));
2282 
2283 		LIST_REMOVE(bucket, ub_link);
2284 		cache->uc_allocbucket = bucket;
2285 		ZONE_UNLOCK(zone);
2286 		goto zalloc_start;
2287 	}
2288 	/* We are no longer associated with this CPU. */
2289 	critical_exit();
2290 
2291 	/*
2292 	 * We bump the uz count when the cache size is insufficient to
2293 	 * handle the working set.
2294 	 */
2295 	if (lockfail && zone->uz_count < BUCKET_MAX)
2296 		zone->uz_count++;
2297 	ZONE_UNLOCK(zone);
2298 
2299 	/*
2300 	 * Now lets just fill a bucket and put it on the free list.  If that
2301 	 * works we'll restart the allocation from the begining and it
2302 	 * will use the just filled bucket.
2303 	 */
2304 	bucket = zone_alloc_bucket(zone, udata, flags);
2305 	if (bucket != NULL) {
2306 		ZONE_LOCK(zone);
2307 		critical_enter();
2308 		cpu = curcpu;
2309 		cache = &zone->uz_cpu[cpu];
2310 		/*
2311 		 * See if we lost the race or were migrated.  Cache the
2312 		 * initialized bucket to make this less likely or claim
2313 		 * the memory directly.
2314 		 */
2315 		if (cache->uc_allocbucket == NULL)
2316 			cache->uc_allocbucket = bucket;
2317 		else
2318 			LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link);
2319 		ZONE_UNLOCK(zone);
2320 		goto zalloc_start;
2321 	}
2322 
2323 	/*
2324 	 * We may not be able to get a bucket so return an actual item.
2325 	 */
2326 #ifdef UMA_DEBUG
2327 	printf("uma_zalloc_arg: Bucketzone returned NULL\n");
2328 #endif
2329 
2330 zalloc_item:
2331 	item = zone_alloc_item(zone, udata, flags);
2332 
2333 #if 0
2334 	/* XXX: FIX!! Do not enable this in CURRENT!! MarkM */
2335 	/* The entropy here is desirable, but the harvesting is expensive */
2336 	random_harvest(&item, sizeof(void *), 1, RANDOM_UMA_ALLOC);
2337 #endif
2338 	return (item);
2339 }
2340 
2341 static uma_slab_t
2342 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int flags)
2343 {
2344 	uma_slab_t slab;
2345 	int reserve;
2346 
2347 	mtx_assert(&keg->uk_lock, MA_OWNED);
2348 	slab = NULL;
2349 	reserve = 0;
2350 	if ((flags & M_USE_RESERVE) == 0)
2351 		reserve = keg->uk_reserve;
2352 
2353 	for (;;) {
2354 		/*
2355 		 * Find a slab with some space.  Prefer slabs that are partially
2356 		 * used over those that are totally full.  This helps to reduce
2357 		 * fragmentation.
2358 		 */
2359 		if (keg->uk_free > reserve) {
2360 			if (!LIST_EMPTY(&keg->uk_part_slab)) {
2361 				slab = LIST_FIRST(&keg->uk_part_slab);
2362 			} else {
2363 				slab = LIST_FIRST(&keg->uk_free_slab);
2364 				LIST_REMOVE(slab, us_link);
2365 				LIST_INSERT_HEAD(&keg->uk_part_slab, slab,
2366 				    us_link);
2367 			}
2368 			MPASS(slab->us_keg == keg);
2369 			return (slab);
2370 		}
2371 
2372 		/*
2373 		 * M_NOVM means don't ask at all!
2374 		 */
2375 		if (flags & M_NOVM)
2376 			break;
2377 
2378 		if (keg->uk_maxpages && keg->uk_pages >= keg->uk_maxpages) {
2379 			keg->uk_flags |= UMA_ZFLAG_FULL;
2380 			/*
2381 			 * If this is not a multi-zone, set the FULL bit.
2382 			 * Otherwise slab_multi() takes care of it.
2383 			 */
2384 			if ((zone->uz_flags & UMA_ZFLAG_MULTI) == 0) {
2385 				zone->uz_flags |= UMA_ZFLAG_FULL;
2386 				zone_log_warning(zone);
2387 			}
2388 			if (flags & M_NOWAIT)
2389 				break;
2390 			zone->uz_sleeps++;
2391 			msleep(keg, &keg->uk_lock, PVM, "keglimit", 0);
2392 			continue;
2393 		}
2394 		slab = keg_alloc_slab(keg, zone, flags);
2395 		/*
2396 		 * If we got a slab here it's safe to mark it partially used
2397 		 * and return.  We assume that the caller is going to remove
2398 		 * at least one item.
2399 		 */
2400 		if (slab) {
2401 			MPASS(slab->us_keg == keg);
2402 			LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2403 			return (slab);
2404 		}
2405 		/*
2406 		 * We might not have been able to get a slab but another cpu
2407 		 * could have while we were unlocked.  Check again before we
2408 		 * fail.
2409 		 */
2410 		flags |= M_NOVM;
2411 	}
2412 	return (slab);
2413 }
2414 
2415 static uma_slab_t
2416 zone_fetch_slab(uma_zone_t zone, uma_keg_t keg, int flags)
2417 {
2418 	uma_slab_t slab;
2419 
2420 	if (keg == NULL) {
2421 		keg = zone_first_keg(zone);
2422 		KEG_LOCK(keg);
2423 	}
2424 
2425 	for (;;) {
2426 		slab = keg_fetch_slab(keg, zone, flags);
2427 		if (slab)
2428 			return (slab);
2429 		if (flags & (M_NOWAIT | M_NOVM))
2430 			break;
2431 	}
2432 	KEG_UNLOCK(keg);
2433 	return (NULL);
2434 }
2435 
2436 /*
2437  * uma_zone_fetch_slab_multi:  Fetches a slab from one available keg.  Returns
2438  * with the keg locked.  On NULL no lock is held.
2439  *
2440  * The last pointer is used to seed the search.  It is not required.
2441  */
2442 static uma_slab_t
2443 zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int rflags)
2444 {
2445 	uma_klink_t klink;
2446 	uma_slab_t slab;
2447 	uma_keg_t keg;
2448 	int flags;
2449 	int empty;
2450 	int full;
2451 
2452 	/*
2453 	 * Don't wait on the first pass.  This will skip limit tests
2454 	 * as well.  We don't want to block if we can find a provider
2455 	 * without blocking.
2456 	 */
2457 	flags = (rflags & ~M_WAITOK) | M_NOWAIT;
2458 	/*
2459 	 * Use the last slab allocated as a hint for where to start
2460 	 * the search.
2461 	 */
2462 	if (last != NULL) {
2463 		slab = keg_fetch_slab(last, zone, flags);
2464 		if (slab)
2465 			return (slab);
2466 		KEG_UNLOCK(last);
2467 	}
2468 	/*
2469 	 * Loop until we have a slab incase of transient failures
2470 	 * while M_WAITOK is specified.  I'm not sure this is 100%
2471 	 * required but we've done it for so long now.
2472 	 */
2473 	for (;;) {
2474 		empty = 0;
2475 		full = 0;
2476 		/*
2477 		 * Search the available kegs for slabs.  Be careful to hold the
2478 		 * correct lock while calling into the keg layer.
2479 		 */
2480 		LIST_FOREACH(klink, &zone->uz_kegs, kl_link) {
2481 			keg = klink->kl_keg;
2482 			KEG_LOCK(keg);
2483 			if ((keg->uk_flags & UMA_ZFLAG_FULL) == 0) {
2484 				slab = keg_fetch_slab(keg, zone, flags);
2485 				if (slab)
2486 					return (slab);
2487 			}
2488 			if (keg->uk_flags & UMA_ZFLAG_FULL)
2489 				full++;
2490 			else
2491 				empty++;
2492 			KEG_UNLOCK(keg);
2493 		}
2494 		if (rflags & (M_NOWAIT | M_NOVM))
2495 			break;
2496 		flags = rflags;
2497 		/*
2498 		 * All kegs are full.  XXX We can't atomically check all kegs
2499 		 * and sleep so just sleep for a short period and retry.
2500 		 */
2501 		if (full && !empty) {
2502 			ZONE_LOCK(zone);
2503 			zone->uz_flags |= UMA_ZFLAG_FULL;
2504 			zone->uz_sleeps++;
2505 			zone_log_warning(zone);
2506 			msleep(zone, zone->uz_lockptr, PVM,
2507 			    "zonelimit", hz/100);
2508 			zone->uz_flags &= ~UMA_ZFLAG_FULL;
2509 			ZONE_UNLOCK(zone);
2510 			continue;
2511 		}
2512 	}
2513 	return (NULL);
2514 }
2515 
2516 static void *
2517 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
2518 {
2519 	void *item;
2520 	uint8_t freei;
2521 
2522 	MPASS(keg == slab->us_keg);
2523 	mtx_assert(&keg->uk_lock, MA_OWNED);
2524 
2525 	freei = BIT_FFS(SLAB_SETSIZE, &slab->us_free) - 1;
2526 	BIT_CLR(SLAB_SETSIZE, freei, &slab->us_free);
2527 	item = slab->us_data + (keg->uk_rsize * freei);
2528 	slab->us_freecount--;
2529 	keg->uk_free--;
2530 
2531 	/* Move this slab to the full list */
2532 	if (slab->us_freecount == 0) {
2533 		LIST_REMOVE(slab, us_link);
2534 		LIST_INSERT_HEAD(&keg->uk_full_slab, slab, us_link);
2535 	}
2536 
2537 	return (item);
2538 }
2539 
2540 static int
2541 zone_import(uma_zone_t zone, void **bucket, int max, int flags)
2542 {
2543 	uma_slab_t slab;
2544 	uma_keg_t keg;
2545 	int i;
2546 
2547 	slab = NULL;
2548 	keg = NULL;
2549 	/* Try to keep the buckets totally full */
2550 	for (i = 0; i < max; ) {
2551 		if ((slab = zone->uz_slab(zone, keg, flags)) == NULL)
2552 			break;
2553 		keg = slab->us_keg;
2554 		while (slab->us_freecount && i < max) {
2555 			bucket[i++] = slab_alloc_item(keg, slab);
2556 			if (keg->uk_free <= keg->uk_reserve)
2557 				break;
2558 		}
2559 		/* Don't grab more than one slab at a time. */
2560 		flags &= ~M_WAITOK;
2561 		flags |= M_NOWAIT;
2562 	}
2563 	if (slab != NULL)
2564 		KEG_UNLOCK(keg);
2565 
2566 	return i;
2567 }
2568 
2569 static uma_bucket_t
2570 zone_alloc_bucket(uma_zone_t zone, void *udata, int flags)
2571 {
2572 	uma_bucket_t bucket;
2573 	int max;
2574 
2575 	/* Don't wait for buckets, preserve caller's NOVM setting. */
2576 	bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
2577 	if (bucket == NULL)
2578 		return (NULL);
2579 
2580 	max = MIN(bucket->ub_entries, zone->uz_count);
2581 	bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
2582 	    max, flags);
2583 
2584 	/*
2585 	 * Initialize the memory if necessary.
2586 	 */
2587 	if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
2588 		int i;
2589 
2590 		for (i = 0; i < bucket->ub_cnt; i++)
2591 			if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size,
2592 			    flags) != 0)
2593 				break;
2594 		/*
2595 		 * If we couldn't initialize the whole bucket, put the
2596 		 * rest back onto the freelist.
2597 		 */
2598 		if (i != bucket->ub_cnt) {
2599 			zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
2600 			    bucket->ub_cnt - i);
2601 #ifdef INVARIANTS
2602 			bzero(&bucket->ub_bucket[i],
2603 			    sizeof(void *) * (bucket->ub_cnt - i));
2604 #endif
2605 			bucket->ub_cnt = i;
2606 		}
2607 	}
2608 
2609 	if (bucket->ub_cnt == 0) {
2610 		bucket_free(zone, bucket, udata);
2611 		atomic_add_long(&zone->uz_fails, 1);
2612 		return (NULL);
2613 	}
2614 
2615 	return (bucket);
2616 }
2617 
2618 /*
2619  * Allocates a single item from a zone.
2620  *
2621  * Arguments
2622  *	zone   The zone to alloc for.
2623  *	udata  The data to be passed to the constructor.
2624  *	flags  M_WAITOK, M_NOWAIT, M_ZERO.
2625  *
2626  * Returns
2627  *	NULL if there is no memory and M_NOWAIT is set
2628  *	An item if successful
2629  */
2630 
2631 static void *
2632 zone_alloc_item(uma_zone_t zone, void *udata, int flags)
2633 {
2634 	void *item;
2635 
2636 	item = NULL;
2637 
2638 #ifdef UMA_DEBUG_ALLOC
2639 	printf("INTERNAL: Allocating one item from %s(%p)\n", zone->uz_name, zone);
2640 #endif
2641 	if (zone->uz_import(zone->uz_arg, &item, 1, flags) != 1)
2642 		goto fail;
2643 	atomic_add_long(&zone->uz_allocs, 1);
2644 
2645 	/*
2646 	 * We have to call both the zone's init (not the keg's init)
2647 	 * and the zone's ctor.  This is because the item is going from
2648 	 * a keg slab directly to the user, and the user is expecting it
2649 	 * to be both zone-init'd as well as zone-ctor'd.
2650 	 */
2651 	if (zone->uz_init != NULL) {
2652 		if (zone->uz_init(item, zone->uz_size, flags) != 0) {
2653 			zone_free_item(zone, item, udata, SKIP_FINI);
2654 			goto fail;
2655 		}
2656 	}
2657 	if (zone->uz_ctor != NULL) {
2658 		if (zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2659 			zone_free_item(zone, item, udata, SKIP_DTOR);
2660 			goto fail;
2661 		}
2662 	}
2663 #ifdef INVARIANTS
2664 	uma_dbg_alloc(zone, NULL, item);
2665 #endif
2666 	if (flags & M_ZERO)
2667 		uma_zero_item(item, zone);
2668 
2669 	return (item);
2670 
2671 fail:
2672 	atomic_add_long(&zone->uz_fails, 1);
2673 	return (NULL);
2674 }
2675 
2676 /* See uma.h */
2677 void
2678 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
2679 {
2680 	uma_cache_t cache;
2681 	uma_bucket_t bucket;
2682 	int lockfail;
2683 	int cpu;
2684 
2685 #if 0
2686 	/* XXX: FIX!! Do not enable this in CURRENT!! MarkM */
2687 	/* The entropy here is desirable, but the harvesting is expensive */
2688 	struct entropy {
2689 		const void *uz_name;
2690 		const void *item;
2691 	} entropy;
2692 
2693 	entropy.uz_name = zone->uz_name;
2694 	entropy.item = item;
2695 	random_harvest(&entropy, sizeof(struct entropy), 2, RANDOM_UMA_ALLOC);
2696 #endif
2697 
2698 #ifdef UMA_DEBUG_ALLOC_1
2699 	printf("Freeing item %p to %s(%p)\n", item, zone->uz_name, zone);
2700 #endif
2701 	CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread,
2702 	    zone->uz_name);
2703 
2704         /* uma_zfree(..., NULL) does nothing, to match free(9). */
2705         if (item == NULL)
2706                 return;
2707 #ifdef DEBUG_MEMGUARD
2708 	if (is_memguard_addr(item)) {
2709 		if (zone->uz_dtor != NULL && zone->uz_dtor != mtrash_dtor)
2710 			zone->uz_dtor(item, zone->uz_size, udata);
2711 		if (zone->uz_fini != NULL && zone->uz_fini != mtrash_fini)
2712 			zone->uz_fini(item, zone->uz_size);
2713 		memguard_free(item);
2714 		return;
2715 	}
2716 #endif
2717 #ifdef INVARIANTS
2718 	if (zone->uz_flags & UMA_ZONE_MALLOC)
2719 		uma_dbg_free(zone, udata, item);
2720 	else
2721 		uma_dbg_free(zone, NULL, item);
2722 #endif
2723 	if (zone->uz_dtor != NULL)
2724 		zone->uz_dtor(item, zone->uz_size, udata);
2725 
2726 	/*
2727 	 * The race here is acceptable.  If we miss it we'll just have to wait
2728 	 * a little longer for the limits to be reset.
2729 	 */
2730 	if (zone->uz_flags & UMA_ZFLAG_FULL)
2731 		goto zfree_item;
2732 
2733 	/*
2734 	 * If possible, free to the per-CPU cache.  There are two
2735 	 * requirements for safe access to the per-CPU cache: (1) the thread
2736 	 * accessing the cache must not be preempted or yield during access,
2737 	 * and (2) the thread must not migrate CPUs without switching which
2738 	 * cache it accesses.  We rely on a critical section to prevent
2739 	 * preemption and migration.  We release the critical section in
2740 	 * order to acquire the zone mutex if we are unable to free to the
2741 	 * current cache; when we re-acquire the critical section, we must
2742 	 * detect and handle migration if it has occurred.
2743 	 */
2744 zfree_restart:
2745 	critical_enter();
2746 	cpu = curcpu;
2747 	cache = &zone->uz_cpu[cpu];
2748 
2749 zfree_start:
2750 	/*
2751 	 * Try to free into the allocbucket first to give LIFO ordering
2752 	 * for cache-hot datastructures.  Spill over into the freebucket
2753 	 * if necessary.  Alloc will swap them if one runs dry.
2754 	 */
2755 	bucket = cache->uc_allocbucket;
2756 	if (bucket == NULL || bucket->ub_cnt >= bucket->ub_entries)
2757 		bucket = cache->uc_freebucket;
2758 	if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) {
2759 		KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL,
2760 		    ("uma_zfree: Freeing to non free bucket index."));
2761 		bucket->ub_bucket[bucket->ub_cnt] = item;
2762 		bucket->ub_cnt++;
2763 		cache->uc_frees++;
2764 		critical_exit();
2765 		return;
2766 	}
2767 
2768 	/*
2769 	 * We must go back the zone, which requires acquiring the zone lock,
2770 	 * which in turn means we must release and re-acquire the critical
2771 	 * section.  Since the critical section is released, we may be
2772 	 * preempted or migrate.  As such, make sure not to maintain any
2773 	 * thread-local state specific to the cache from prior to releasing
2774 	 * the critical section.
2775 	 */
2776 	critical_exit();
2777 	if (zone->uz_count == 0 || bucketdisable)
2778 		goto zfree_item;
2779 
2780 	lockfail = 0;
2781 	if (ZONE_TRYLOCK(zone) == 0) {
2782 		/* Record contention to size the buckets. */
2783 		ZONE_LOCK(zone);
2784 		lockfail = 1;
2785 	}
2786 	critical_enter();
2787 	cpu = curcpu;
2788 	cache = &zone->uz_cpu[cpu];
2789 
2790 	/*
2791 	 * Since we have locked the zone we may as well send back our stats.
2792 	 */
2793 	atomic_add_long(&zone->uz_allocs, cache->uc_allocs);
2794 	atomic_add_long(&zone->uz_frees, cache->uc_frees);
2795 	cache->uc_allocs = 0;
2796 	cache->uc_frees = 0;
2797 
2798 	bucket = cache->uc_freebucket;
2799 	if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) {
2800 		ZONE_UNLOCK(zone);
2801 		goto zfree_start;
2802 	}
2803 	cache->uc_freebucket = NULL;
2804 
2805 	/* Can we throw this on the zone full list? */
2806 	if (bucket != NULL) {
2807 #ifdef UMA_DEBUG_ALLOC
2808 		printf("uma_zfree: Putting old bucket on the free list.\n");
2809 #endif
2810 		/* ub_cnt is pointing to the last free item */
2811 		KASSERT(bucket->ub_cnt != 0,
2812 		    ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n"));
2813 		LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link);
2814 	}
2815 
2816 	/* We are no longer associated with this CPU. */
2817 	critical_exit();
2818 
2819 	/*
2820 	 * We bump the uz count when the cache size is insufficient to
2821 	 * handle the working set.
2822 	 */
2823 	if (lockfail && zone->uz_count < BUCKET_MAX)
2824 		zone->uz_count++;
2825 	ZONE_UNLOCK(zone);
2826 
2827 #ifdef UMA_DEBUG_ALLOC
2828 	printf("uma_zfree: Allocating new free bucket.\n");
2829 #endif
2830 	bucket = bucket_alloc(zone, udata, M_NOWAIT);
2831 	if (bucket) {
2832 		critical_enter();
2833 		cpu = curcpu;
2834 		cache = &zone->uz_cpu[cpu];
2835 		if (cache->uc_freebucket == NULL) {
2836 			cache->uc_freebucket = bucket;
2837 			goto zfree_start;
2838 		}
2839 		/*
2840 		 * We lost the race, start over.  We have to drop our
2841 		 * critical section to free the bucket.
2842 		 */
2843 		critical_exit();
2844 		bucket_free(zone, bucket, udata);
2845 		goto zfree_restart;
2846 	}
2847 
2848 	/*
2849 	 * If nothing else caught this, we'll just do an internal free.
2850 	 */
2851 zfree_item:
2852 	zone_free_item(zone, item, udata, SKIP_DTOR);
2853 
2854 	return;
2855 }
2856 
2857 static void
2858 slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item)
2859 {
2860 	uint8_t freei;
2861 
2862 	mtx_assert(&keg->uk_lock, MA_OWNED);
2863 	MPASS(keg == slab->us_keg);
2864 
2865 	/* Do we need to remove from any lists? */
2866 	if (slab->us_freecount+1 == keg->uk_ipers) {
2867 		LIST_REMOVE(slab, us_link);
2868 		LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
2869 	} else if (slab->us_freecount == 0) {
2870 		LIST_REMOVE(slab, us_link);
2871 		LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2872 	}
2873 
2874 	/* Slab management. */
2875 	freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
2876 	BIT_SET(SLAB_SETSIZE, freei, &slab->us_free);
2877 	slab->us_freecount++;
2878 
2879 	/* Keg statistics. */
2880 	keg->uk_free++;
2881 }
2882 
2883 static void
2884 zone_release(uma_zone_t zone, void **bucket, int cnt)
2885 {
2886 	void *item;
2887 	uma_slab_t slab;
2888 	uma_keg_t keg;
2889 	uint8_t *mem;
2890 	int clearfull;
2891 	int i;
2892 
2893 	clearfull = 0;
2894 	keg = zone_first_keg(zone);
2895 	KEG_LOCK(keg);
2896 	for (i = 0; i < cnt; i++) {
2897 		item = bucket[i];
2898 		if (!(zone->uz_flags & UMA_ZONE_VTOSLAB)) {
2899 			mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
2900 			if (zone->uz_flags & UMA_ZONE_HASH) {
2901 				slab = hash_sfind(&keg->uk_hash, mem);
2902 			} else {
2903 				mem += keg->uk_pgoff;
2904 				slab = (uma_slab_t)mem;
2905 			}
2906 		} else {
2907 			slab = vtoslab((vm_offset_t)item);
2908 			if (slab->us_keg != keg) {
2909 				KEG_UNLOCK(keg);
2910 				keg = slab->us_keg;
2911 				KEG_LOCK(keg);
2912 			}
2913 		}
2914 		slab_free_item(keg, slab, item);
2915 		if (keg->uk_flags & UMA_ZFLAG_FULL) {
2916 			if (keg->uk_pages < keg->uk_maxpages) {
2917 				keg->uk_flags &= ~UMA_ZFLAG_FULL;
2918 				clearfull = 1;
2919 			}
2920 
2921 			/*
2922 			 * We can handle one more allocation. Since we're
2923 			 * clearing ZFLAG_FULL, wake up all procs blocked
2924 			 * on pages. This should be uncommon, so keeping this
2925 			 * simple for now (rather than adding count of blocked
2926 			 * threads etc).
2927 			 */
2928 			wakeup(keg);
2929 		}
2930 	}
2931 	KEG_UNLOCK(keg);
2932 	if (clearfull) {
2933 		ZONE_LOCK(zone);
2934 		zone->uz_flags &= ~UMA_ZFLAG_FULL;
2935 		wakeup(zone);
2936 		ZONE_UNLOCK(zone);
2937 	}
2938 
2939 }
2940 
2941 /*
2942  * Frees a single item to any zone.
2943  *
2944  * Arguments:
2945  *	zone   The zone to free to
2946  *	item   The item we're freeing
2947  *	udata  User supplied data for the dtor
2948  *	skip   Skip dtors and finis
2949  */
2950 static void
2951 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
2952 {
2953 
2954 #ifdef INVARIANTS
2955 	if (skip == SKIP_NONE) {
2956 		if (zone->uz_flags & UMA_ZONE_MALLOC)
2957 			uma_dbg_free(zone, udata, item);
2958 		else
2959 			uma_dbg_free(zone, NULL, item);
2960 	}
2961 #endif
2962 	if (skip < SKIP_DTOR && zone->uz_dtor)
2963 		zone->uz_dtor(item, zone->uz_size, udata);
2964 
2965 	if (skip < SKIP_FINI && zone->uz_fini)
2966 		zone->uz_fini(item, zone->uz_size);
2967 
2968 	atomic_add_long(&zone->uz_frees, 1);
2969 	zone->uz_release(zone->uz_arg, &item, 1);
2970 }
2971 
2972 /* See uma.h */
2973 int
2974 uma_zone_set_max(uma_zone_t zone, int nitems)
2975 {
2976 	uma_keg_t keg;
2977 
2978 	keg = zone_first_keg(zone);
2979 	if (keg == NULL)
2980 		return (0);
2981 	KEG_LOCK(keg);
2982 	keg->uk_maxpages = (nitems / keg->uk_ipers) * keg->uk_ppera;
2983 	if (keg->uk_maxpages * keg->uk_ipers < nitems)
2984 		keg->uk_maxpages += keg->uk_ppera;
2985 	nitems = keg->uk_maxpages * keg->uk_ipers;
2986 	KEG_UNLOCK(keg);
2987 
2988 	return (nitems);
2989 }
2990 
2991 /* See uma.h */
2992 int
2993 uma_zone_get_max(uma_zone_t zone)
2994 {
2995 	int nitems;
2996 	uma_keg_t keg;
2997 
2998 	keg = zone_first_keg(zone);
2999 	if (keg == NULL)
3000 		return (0);
3001 	KEG_LOCK(keg);
3002 	nitems = keg->uk_maxpages * keg->uk_ipers;
3003 	KEG_UNLOCK(keg);
3004 
3005 	return (nitems);
3006 }
3007 
3008 /* See uma.h */
3009 void
3010 uma_zone_set_warning(uma_zone_t zone, const char *warning)
3011 {
3012 
3013 	ZONE_LOCK(zone);
3014 	zone->uz_warning = warning;
3015 	ZONE_UNLOCK(zone);
3016 }
3017 
3018 /* See uma.h */
3019 int
3020 uma_zone_get_cur(uma_zone_t zone)
3021 {
3022 	int64_t nitems;
3023 	u_int i;
3024 
3025 	ZONE_LOCK(zone);
3026 	nitems = zone->uz_allocs - zone->uz_frees;
3027 	CPU_FOREACH(i) {
3028 		/*
3029 		 * See the comment in sysctl_vm_zone_stats() regarding the
3030 		 * safety of accessing the per-cpu caches. With the zone lock
3031 		 * held, it is safe, but can potentially result in stale data.
3032 		 */
3033 		nitems += zone->uz_cpu[i].uc_allocs -
3034 		    zone->uz_cpu[i].uc_frees;
3035 	}
3036 	ZONE_UNLOCK(zone);
3037 
3038 	return (nitems < 0 ? 0 : nitems);
3039 }
3040 
3041 /* See uma.h */
3042 void
3043 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
3044 {
3045 	uma_keg_t keg;
3046 
3047 	keg = zone_first_keg(zone);
3048 	KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type"));
3049 	KEG_LOCK(keg);
3050 	KASSERT(keg->uk_pages == 0,
3051 	    ("uma_zone_set_init on non-empty keg"));
3052 	keg->uk_init = uminit;
3053 	KEG_UNLOCK(keg);
3054 }
3055 
3056 /* See uma.h */
3057 void
3058 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
3059 {
3060 	uma_keg_t keg;
3061 
3062 	keg = zone_first_keg(zone);
3063 	KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type"));
3064 	KEG_LOCK(keg);
3065 	KASSERT(keg->uk_pages == 0,
3066 	    ("uma_zone_set_fini on non-empty keg"));
3067 	keg->uk_fini = fini;
3068 	KEG_UNLOCK(keg);
3069 }
3070 
3071 /* See uma.h */
3072 void
3073 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
3074 {
3075 
3076 	ZONE_LOCK(zone);
3077 	KASSERT(zone_first_keg(zone)->uk_pages == 0,
3078 	    ("uma_zone_set_zinit on non-empty keg"));
3079 	zone->uz_init = zinit;
3080 	ZONE_UNLOCK(zone);
3081 }
3082 
3083 /* See uma.h */
3084 void
3085 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
3086 {
3087 
3088 	ZONE_LOCK(zone);
3089 	KASSERT(zone_first_keg(zone)->uk_pages == 0,
3090 	    ("uma_zone_set_zfini on non-empty keg"));
3091 	zone->uz_fini = zfini;
3092 	ZONE_UNLOCK(zone);
3093 }
3094 
3095 /* See uma.h */
3096 /* XXX uk_freef is not actually used with the zone locked */
3097 void
3098 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
3099 {
3100 	uma_keg_t keg;
3101 
3102 	keg = zone_first_keg(zone);
3103 	KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type"));
3104 	KEG_LOCK(keg);
3105 	keg->uk_freef = freef;
3106 	KEG_UNLOCK(keg);
3107 }
3108 
3109 /* See uma.h */
3110 /* XXX uk_allocf is not actually used with the zone locked */
3111 void
3112 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
3113 {
3114 	uma_keg_t keg;
3115 
3116 	keg = zone_first_keg(zone);
3117 	KEG_LOCK(keg);
3118 	keg->uk_allocf = allocf;
3119 	KEG_UNLOCK(keg);
3120 }
3121 
3122 /* See uma.h */
3123 void
3124 uma_zone_reserve(uma_zone_t zone, int items)
3125 {
3126 	uma_keg_t keg;
3127 
3128 	keg = zone_first_keg(zone);
3129 	if (keg == NULL)
3130 		return;
3131 	KEG_LOCK(keg);
3132 	keg->uk_reserve = items;
3133 	KEG_UNLOCK(keg);
3134 
3135 	return;
3136 }
3137 
3138 /* See uma.h */
3139 int
3140 uma_zone_reserve_kva(uma_zone_t zone, int count)
3141 {
3142 	uma_keg_t keg;
3143 	vm_offset_t kva;
3144 	int pages;
3145 
3146 	keg = zone_first_keg(zone);
3147 	if (keg == NULL)
3148 		return (0);
3149 	pages = count / keg->uk_ipers;
3150 
3151 	if (pages * keg->uk_ipers < count)
3152 		pages++;
3153 
3154 #ifdef UMA_MD_SMALL_ALLOC
3155 	if (keg->uk_ppera > 1) {
3156 #else
3157 	if (1) {
3158 #endif
3159 		kva = kva_alloc(pages * UMA_SLAB_SIZE);
3160 		if (kva == 0)
3161 			return (0);
3162 	} else
3163 		kva = 0;
3164 	KEG_LOCK(keg);
3165 	keg->uk_kva = kva;
3166 	keg->uk_offset = 0;
3167 	keg->uk_maxpages = pages;
3168 #ifdef UMA_MD_SMALL_ALLOC
3169 	keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
3170 #else
3171 	keg->uk_allocf = noobj_alloc;
3172 #endif
3173 	keg->uk_flags |= UMA_ZONE_NOFREE;
3174 	KEG_UNLOCK(keg);
3175 
3176 	return (1);
3177 }
3178 
3179 /* See uma.h */
3180 void
3181 uma_prealloc(uma_zone_t zone, int items)
3182 {
3183 	int slabs;
3184 	uma_slab_t slab;
3185 	uma_keg_t keg;
3186 
3187 	keg = zone_first_keg(zone);
3188 	if (keg == NULL)
3189 		return;
3190 	KEG_LOCK(keg);
3191 	slabs = items / keg->uk_ipers;
3192 	if (slabs * keg->uk_ipers < items)
3193 		slabs++;
3194 	while (slabs > 0) {
3195 		slab = keg_alloc_slab(keg, zone, M_WAITOK);
3196 		if (slab == NULL)
3197 			break;
3198 		MPASS(slab->us_keg == keg);
3199 		LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
3200 		slabs--;
3201 	}
3202 	KEG_UNLOCK(keg);
3203 }
3204 
3205 /* See uma.h */
3206 uint32_t *
3207 uma_find_refcnt(uma_zone_t zone, void *item)
3208 {
3209 	uma_slabrefcnt_t slabref;
3210 	uma_slab_t slab;
3211 	uma_keg_t keg;
3212 	uint32_t *refcnt;
3213 	int idx;
3214 
3215 	slab = vtoslab((vm_offset_t)item & (~UMA_SLAB_MASK));
3216 	slabref = (uma_slabrefcnt_t)slab;
3217 	keg = slab->us_keg;
3218 	KASSERT(keg->uk_flags & UMA_ZONE_REFCNT,
3219 	    ("uma_find_refcnt(): zone possibly not UMA_ZONE_REFCNT"));
3220 	idx = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
3221 	refcnt = &slabref->us_refcnt[idx];
3222 	return refcnt;
3223 }
3224 
3225 /* See uma.h */
3226 void
3227 uma_reclaim(void)
3228 {
3229 #ifdef UMA_DEBUG
3230 	printf("UMA: vm asked us to release pages!\n");
3231 #endif
3232 	sx_xlock(&uma_drain_lock);
3233 	bucket_enable();
3234 	zone_foreach(zone_drain);
3235 	if (vm_page_count_min()) {
3236 		cache_drain_safe(NULL);
3237 		zone_foreach(zone_drain);
3238 	}
3239 	/*
3240 	 * Some slabs may have been freed but this zone will be visited early
3241 	 * we visit again so that we can free pages that are empty once other
3242 	 * zones are drained.  We have to do the same for buckets.
3243 	 */
3244 	zone_drain(slabzone);
3245 	zone_drain(slabrefzone);
3246 	bucket_zone_drain();
3247 	sx_xunlock(&uma_drain_lock);
3248 }
3249 
3250 /* See uma.h */
3251 int
3252 uma_zone_exhausted(uma_zone_t zone)
3253 {
3254 	int full;
3255 
3256 	ZONE_LOCK(zone);
3257 	full = (zone->uz_flags & UMA_ZFLAG_FULL);
3258 	ZONE_UNLOCK(zone);
3259 	return (full);
3260 }
3261 
3262 int
3263 uma_zone_exhausted_nolock(uma_zone_t zone)
3264 {
3265 	return (zone->uz_flags & UMA_ZFLAG_FULL);
3266 }
3267 
3268 void *
3269 uma_large_malloc(int size, int wait)
3270 {
3271 	void *mem;
3272 	uma_slab_t slab;
3273 	uint8_t flags;
3274 
3275 	slab = zone_alloc_item(slabzone, NULL, wait);
3276 	if (slab == NULL)
3277 		return (NULL);
3278 	mem = page_alloc(NULL, size, &flags, wait);
3279 	if (mem) {
3280 		vsetslab((vm_offset_t)mem, slab);
3281 		slab->us_data = mem;
3282 		slab->us_flags = flags | UMA_SLAB_MALLOC;
3283 		slab->us_size = size;
3284 	} else {
3285 		zone_free_item(slabzone, slab, NULL, SKIP_NONE);
3286 	}
3287 
3288 	return (mem);
3289 }
3290 
3291 void
3292 uma_large_free(uma_slab_t slab)
3293 {
3294 
3295 	page_free(slab->us_data, slab->us_size, slab->us_flags);
3296 	zone_free_item(slabzone, slab, NULL, SKIP_NONE);
3297 }
3298 
3299 static void
3300 uma_zero_item(void *item, uma_zone_t zone)
3301 {
3302 
3303 	if (zone->uz_flags & UMA_ZONE_PCPU) {
3304 		for (int i = 0; i < mp_ncpus; i++)
3305 			bzero(zpcpu_get_cpu(item, i), zone->uz_size);
3306 	} else
3307 		bzero(item, zone->uz_size);
3308 }
3309 
3310 void
3311 uma_print_stats(void)
3312 {
3313 	zone_foreach(uma_print_zone);
3314 }
3315 
3316 static void
3317 slab_print(uma_slab_t slab)
3318 {
3319 	printf("slab: keg %p, data %p, freecount %d\n",
3320 		slab->us_keg, slab->us_data, slab->us_freecount);
3321 }
3322 
3323 static void
3324 cache_print(uma_cache_t cache)
3325 {
3326 	printf("alloc: %p(%d), free: %p(%d)\n",
3327 		cache->uc_allocbucket,
3328 		cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0,
3329 		cache->uc_freebucket,
3330 		cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0);
3331 }
3332 
3333 static void
3334 uma_print_keg(uma_keg_t keg)
3335 {
3336 	uma_slab_t slab;
3337 
3338 	printf("keg: %s(%p) size %d(%d) flags %#x ipers %d ppera %d "
3339 	    "out %d free %d limit %d\n",
3340 	    keg->uk_name, keg, keg->uk_size, keg->uk_rsize, keg->uk_flags,
3341 	    keg->uk_ipers, keg->uk_ppera,
3342 	    (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free,
3343 	    (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers);
3344 	printf("Part slabs:\n");
3345 	LIST_FOREACH(slab, &keg->uk_part_slab, us_link)
3346 		slab_print(slab);
3347 	printf("Free slabs:\n");
3348 	LIST_FOREACH(slab, &keg->uk_free_slab, us_link)
3349 		slab_print(slab);
3350 	printf("Full slabs:\n");
3351 	LIST_FOREACH(slab, &keg->uk_full_slab, us_link)
3352 		slab_print(slab);
3353 }
3354 
3355 void
3356 uma_print_zone(uma_zone_t zone)
3357 {
3358 	uma_cache_t cache;
3359 	uma_klink_t kl;
3360 	int i;
3361 
3362 	printf("zone: %s(%p) size %d flags %#x\n",
3363 	    zone->uz_name, zone, zone->uz_size, zone->uz_flags);
3364 	LIST_FOREACH(kl, &zone->uz_kegs, kl_link)
3365 		uma_print_keg(kl->kl_keg);
3366 	CPU_FOREACH(i) {
3367 		cache = &zone->uz_cpu[i];
3368 		printf("CPU %d Cache:\n", i);
3369 		cache_print(cache);
3370 	}
3371 }
3372 
3373 #ifdef DDB
3374 /*
3375  * Generate statistics across both the zone and its per-cpu cache's.  Return
3376  * desired statistics if the pointer is non-NULL for that statistic.
3377  *
3378  * Note: does not update the zone statistics, as it can't safely clear the
3379  * per-CPU cache statistic.
3380  *
3381  * XXXRW: Following the uc_allocbucket and uc_freebucket pointers here isn't
3382  * safe from off-CPU; we should modify the caches to track this information
3383  * directly so that we don't have to.
3384  */
3385 static void
3386 uma_zone_sumstat(uma_zone_t z, int *cachefreep, uint64_t *allocsp,
3387     uint64_t *freesp, uint64_t *sleepsp)
3388 {
3389 	uma_cache_t cache;
3390 	uint64_t allocs, frees, sleeps;
3391 	int cachefree, cpu;
3392 
3393 	allocs = frees = sleeps = 0;
3394 	cachefree = 0;
3395 	CPU_FOREACH(cpu) {
3396 		cache = &z->uz_cpu[cpu];
3397 		if (cache->uc_allocbucket != NULL)
3398 			cachefree += cache->uc_allocbucket->ub_cnt;
3399 		if (cache->uc_freebucket != NULL)
3400 			cachefree += cache->uc_freebucket->ub_cnt;
3401 		allocs += cache->uc_allocs;
3402 		frees += cache->uc_frees;
3403 	}
3404 	allocs += z->uz_allocs;
3405 	frees += z->uz_frees;
3406 	sleeps += z->uz_sleeps;
3407 	if (cachefreep != NULL)
3408 		*cachefreep = cachefree;
3409 	if (allocsp != NULL)
3410 		*allocsp = allocs;
3411 	if (freesp != NULL)
3412 		*freesp = frees;
3413 	if (sleepsp != NULL)
3414 		*sleepsp = sleeps;
3415 }
3416 #endif /* DDB */
3417 
3418 static int
3419 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
3420 {
3421 	uma_keg_t kz;
3422 	uma_zone_t z;
3423 	int count;
3424 
3425 	count = 0;
3426 	rw_rlock(&uma_rwlock);
3427 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3428 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
3429 			count++;
3430 	}
3431 	rw_runlock(&uma_rwlock);
3432 	return (sysctl_handle_int(oidp, &count, 0, req));
3433 }
3434 
3435 static int
3436 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
3437 {
3438 	struct uma_stream_header ush;
3439 	struct uma_type_header uth;
3440 	struct uma_percpu_stat ups;
3441 	uma_bucket_t bucket;
3442 	struct sbuf sbuf;
3443 	uma_cache_t cache;
3444 	uma_klink_t kl;
3445 	uma_keg_t kz;
3446 	uma_zone_t z;
3447 	uma_keg_t k;
3448 	int count, error, i;
3449 
3450 	error = sysctl_wire_old_buffer(req, 0);
3451 	if (error != 0)
3452 		return (error);
3453 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
3454 
3455 	count = 0;
3456 	rw_rlock(&uma_rwlock);
3457 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3458 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
3459 			count++;
3460 	}
3461 
3462 	/*
3463 	 * Insert stream header.
3464 	 */
3465 	bzero(&ush, sizeof(ush));
3466 	ush.ush_version = UMA_STREAM_VERSION;
3467 	ush.ush_maxcpus = (mp_maxid + 1);
3468 	ush.ush_count = count;
3469 	(void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
3470 
3471 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3472 		LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3473 			bzero(&uth, sizeof(uth));
3474 			ZONE_LOCK(z);
3475 			strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
3476 			uth.uth_align = kz->uk_align;
3477 			uth.uth_size = kz->uk_size;
3478 			uth.uth_rsize = kz->uk_rsize;
3479 			LIST_FOREACH(kl, &z->uz_kegs, kl_link) {
3480 				k = kl->kl_keg;
3481 				uth.uth_maxpages += k->uk_maxpages;
3482 				uth.uth_pages += k->uk_pages;
3483 				uth.uth_keg_free += k->uk_free;
3484 				uth.uth_limit = (k->uk_maxpages / k->uk_ppera)
3485 				    * k->uk_ipers;
3486 			}
3487 
3488 			/*
3489 			 * A zone is secondary is it is not the first entry
3490 			 * on the keg's zone list.
3491 			 */
3492 			if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
3493 			    (LIST_FIRST(&kz->uk_zones) != z))
3494 				uth.uth_zone_flags = UTH_ZONE_SECONDARY;
3495 
3496 			LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3497 				uth.uth_zone_free += bucket->ub_cnt;
3498 			uth.uth_allocs = z->uz_allocs;
3499 			uth.uth_frees = z->uz_frees;
3500 			uth.uth_fails = z->uz_fails;
3501 			uth.uth_sleeps = z->uz_sleeps;
3502 			(void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
3503 			/*
3504 			 * While it is not normally safe to access the cache
3505 			 * bucket pointers while not on the CPU that owns the
3506 			 * cache, we only allow the pointers to be exchanged
3507 			 * without the zone lock held, not invalidated, so
3508 			 * accept the possible race associated with bucket
3509 			 * exchange during monitoring.
3510 			 */
3511 			for (i = 0; i < (mp_maxid + 1); i++) {
3512 				bzero(&ups, sizeof(ups));
3513 				if (kz->uk_flags & UMA_ZFLAG_INTERNAL)
3514 					goto skip;
3515 				if (CPU_ABSENT(i))
3516 					goto skip;
3517 				cache = &z->uz_cpu[i];
3518 				if (cache->uc_allocbucket != NULL)
3519 					ups.ups_cache_free +=
3520 					    cache->uc_allocbucket->ub_cnt;
3521 				if (cache->uc_freebucket != NULL)
3522 					ups.ups_cache_free +=
3523 					    cache->uc_freebucket->ub_cnt;
3524 				ups.ups_allocs = cache->uc_allocs;
3525 				ups.ups_frees = cache->uc_frees;
3526 skip:
3527 				(void)sbuf_bcat(&sbuf, &ups, sizeof(ups));
3528 			}
3529 			ZONE_UNLOCK(z);
3530 		}
3531 	}
3532 	rw_runlock(&uma_rwlock);
3533 	error = sbuf_finish(&sbuf);
3534 	sbuf_delete(&sbuf);
3535 	return (error);
3536 }
3537 
3538 int
3539 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS)
3540 {
3541 	uma_zone_t zone = *(uma_zone_t *)arg1;
3542 	int error, max, old;
3543 
3544 	old = max = uma_zone_get_max(zone);
3545 	error = sysctl_handle_int(oidp, &max, 0, req);
3546 	if (error || !req->newptr)
3547 		return (error);
3548 
3549 	if (max < old)
3550 		return (EINVAL);
3551 
3552 	uma_zone_set_max(zone, max);
3553 
3554 	return (0);
3555 }
3556 
3557 int
3558 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS)
3559 {
3560 	uma_zone_t zone = *(uma_zone_t *)arg1;
3561 	int cur;
3562 
3563 	cur = uma_zone_get_cur(zone);
3564 	return (sysctl_handle_int(oidp, &cur, 0, req));
3565 }
3566 
3567 #ifdef DDB
3568 DB_SHOW_COMMAND(uma, db_show_uma)
3569 {
3570 	uint64_t allocs, frees, sleeps;
3571 	uma_bucket_t bucket;
3572 	uma_keg_t kz;
3573 	uma_zone_t z;
3574 	int cachefree;
3575 
3576 	db_printf("%18s %8s %8s %8s %12s %8s %8s\n", "Zone", "Size", "Used",
3577 	    "Free", "Requests", "Sleeps", "Bucket");
3578 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3579 		LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3580 			if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
3581 				allocs = z->uz_allocs;
3582 				frees = z->uz_frees;
3583 				sleeps = z->uz_sleeps;
3584 				cachefree = 0;
3585 			} else
3586 				uma_zone_sumstat(z, &cachefree, &allocs,
3587 				    &frees, &sleeps);
3588 			if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
3589 			    (LIST_FIRST(&kz->uk_zones) != z)))
3590 				cachefree += kz->uk_free;
3591 			LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3592 				cachefree += bucket->ub_cnt;
3593 			db_printf("%18s %8ju %8jd %8d %12ju %8ju %8u\n",
3594 			    z->uz_name, (uintmax_t)kz->uk_size,
3595 			    (intmax_t)(allocs - frees), cachefree,
3596 			    (uintmax_t)allocs, sleeps, z->uz_count);
3597 			if (db_pager_quit)
3598 				return;
3599 		}
3600 	}
3601 }
3602 
3603 DB_SHOW_COMMAND(umacache, db_show_umacache)
3604 {
3605 	uint64_t allocs, frees;
3606 	uma_bucket_t bucket;
3607 	uma_zone_t z;
3608 	int cachefree;
3609 
3610 	db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
3611 	    "Requests", "Bucket");
3612 	LIST_FOREACH(z, &uma_cachezones, uz_link) {
3613 		uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL);
3614 		LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3615 			cachefree += bucket->ub_cnt;
3616 		db_printf("%18s %8ju %8jd %8d %12ju %8u\n",
3617 		    z->uz_name, (uintmax_t)z->uz_size,
3618 		    (intmax_t)(allocs - frees), cachefree,
3619 		    (uintmax_t)allocs, z->uz_count);
3620 		if (db_pager_quit)
3621 			return;
3622 	}
3623 }
3624 #endif
3625