xref: /freebsd/sys/vm/uma_core.c (revision 18849b5da0c5eaa88500b457be05b038813b51b1)
1 /*-
2  * Copyright (c) 2002-2005, 2009, 2013 Jeffrey Roberson <jeff@FreeBSD.org>
3  * Copyright (c) 2004, 2005 Bosko Milekic <bmilekic@FreeBSD.org>
4  * Copyright (c) 2004-2006 Robert N. M. Watson
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice unmodified, this list of conditions, and the following
12  *    disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
18  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
19  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
20  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
21  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
22  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
23  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
24  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
25  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
26  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
27  */
28 
29 /*
30  * uma_core.c  Implementation of the Universal Memory allocator
31  *
32  * This allocator is intended to replace the multitude of similar object caches
33  * in the standard FreeBSD kernel.  The intent is to be flexible as well as
34  * efficient.  A primary design goal is to return unused memory to the rest of
35  * the system.  This will make the system as a whole more flexible due to the
36  * ability to move memory to subsystems which most need it instead of leaving
37  * pools of reserved memory unused.
38  *
39  * The basic ideas stem from similar slab/zone based allocators whose algorithms
40  * are well known.
41  *
42  */
43 
44 /*
45  * TODO:
46  *	- Improve memory usage for large allocations
47  *	- Investigate cache size adjustments
48  */
49 
50 #include <sys/cdefs.h>
51 __FBSDID("$FreeBSD$");
52 
53 /* I should really use ktr.. */
54 /*
55 #define UMA_DEBUG 1
56 #define UMA_DEBUG_ALLOC 1
57 #define UMA_DEBUG_ALLOC_1 1
58 */
59 
60 #include "opt_ddb.h"
61 #include "opt_param.h"
62 #include "opt_vm.h"
63 
64 #include <sys/param.h>
65 #include <sys/systm.h>
66 #include <sys/bitset.h>
67 #include <sys/kernel.h>
68 #include <sys/types.h>
69 #include <sys/queue.h>
70 #include <sys/malloc.h>
71 #include <sys/ktr.h>
72 #include <sys/lock.h>
73 #include <sys/sysctl.h>
74 #include <sys/mutex.h>
75 #include <sys/proc.h>
76 #include <sys/random.h>
77 #include <sys/rwlock.h>
78 #include <sys/sbuf.h>
79 #include <sys/sched.h>
80 #include <sys/smp.h>
81 #include <sys/taskqueue.h>
82 #include <sys/vmmeter.h>
83 
84 #include <vm/vm.h>
85 #include <vm/vm_object.h>
86 #include <vm/vm_page.h>
87 #include <vm/vm_pageout.h>
88 #include <vm/vm_param.h>
89 #include <vm/vm_map.h>
90 #include <vm/vm_kern.h>
91 #include <vm/vm_extern.h>
92 #include <vm/uma.h>
93 #include <vm/uma_int.h>
94 #include <vm/uma_dbg.h>
95 
96 #include <ddb/ddb.h>
97 
98 #ifdef DEBUG_MEMGUARD
99 #include <vm/memguard.h>
100 #endif
101 
102 /*
103  * This is the zone and keg from which all zones are spawned.  The idea is that
104  * even the zone & keg heads are allocated from the allocator, so we use the
105  * bss section to bootstrap us.
106  */
107 static struct uma_keg masterkeg;
108 static struct uma_zone masterzone_k;
109 static struct uma_zone masterzone_z;
110 static uma_zone_t kegs = &masterzone_k;
111 static uma_zone_t zones = &masterzone_z;
112 
113 /* This is the zone from which all of uma_slab_t's are allocated. */
114 static uma_zone_t slabzone;
115 
116 /*
117  * The initial hash tables come out of this zone so they can be allocated
118  * prior to malloc coming up.
119  */
120 static uma_zone_t hashzone;
121 
122 /* The boot-time adjusted value for cache line alignment. */
123 int uma_align_cache = 64 - 1;
124 
125 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets");
126 
127 /*
128  * Are we allowed to allocate buckets?
129  */
130 static int bucketdisable = 1;
131 
132 /* Linked list of all kegs in the system */
133 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs);
134 
135 /* Linked list of all cache-only zones in the system */
136 static LIST_HEAD(,uma_zone) uma_cachezones =
137     LIST_HEAD_INITIALIZER(uma_cachezones);
138 
139 /* This RW lock protects the keg list */
140 static struct rwlock_padalign uma_rwlock;
141 
142 /* Linked list of boot time pages */
143 static LIST_HEAD(,uma_slab) uma_boot_pages =
144     LIST_HEAD_INITIALIZER(uma_boot_pages);
145 
146 /* This mutex protects the boot time pages list */
147 static struct mtx_padalign uma_boot_pages_mtx;
148 
149 static struct sx uma_drain_lock;
150 
151 /* Is the VM done starting up? */
152 static int booted = 0;
153 #define	UMA_STARTUP	1
154 #define	UMA_STARTUP2	2
155 
156 /*
157  * This is the handle used to schedule events that need to happen
158  * outside of the allocation fast path.
159  */
160 static struct callout uma_callout;
161 #define	UMA_TIMEOUT	20		/* Seconds for callout interval. */
162 
163 /*
164  * This structure is passed as the zone ctor arg so that I don't have to create
165  * a special allocation function just for zones.
166  */
167 struct uma_zctor_args {
168 	const char *name;
169 	size_t size;
170 	uma_ctor ctor;
171 	uma_dtor dtor;
172 	uma_init uminit;
173 	uma_fini fini;
174 	uma_import import;
175 	uma_release release;
176 	void *arg;
177 	uma_keg_t keg;
178 	int align;
179 	uint32_t flags;
180 };
181 
182 struct uma_kctor_args {
183 	uma_zone_t zone;
184 	size_t size;
185 	uma_init uminit;
186 	uma_fini fini;
187 	int align;
188 	uint32_t flags;
189 };
190 
191 struct uma_bucket_zone {
192 	uma_zone_t	ubz_zone;
193 	char		*ubz_name;
194 	int		ubz_entries;	/* Number of items it can hold. */
195 	int		ubz_maxsize;	/* Maximum allocation size per-item. */
196 };
197 
198 /*
199  * Compute the actual number of bucket entries to pack them in power
200  * of two sizes for more efficient space utilization.
201  */
202 #define	BUCKET_SIZE(n)						\
203     (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *))
204 
205 #define	BUCKET_MAX	BUCKET_SIZE(256)
206 
207 struct uma_bucket_zone bucket_zones[] = {
208 	{ NULL, "4 Bucket", BUCKET_SIZE(4), 4096 },
209 	{ NULL, "6 Bucket", BUCKET_SIZE(6), 3072 },
210 	{ NULL, "8 Bucket", BUCKET_SIZE(8), 2048 },
211 	{ NULL, "12 Bucket", BUCKET_SIZE(12), 1536 },
212 	{ NULL, "16 Bucket", BUCKET_SIZE(16), 1024 },
213 	{ NULL, "32 Bucket", BUCKET_SIZE(32), 512 },
214 	{ NULL, "64 Bucket", BUCKET_SIZE(64), 256 },
215 	{ NULL, "128 Bucket", BUCKET_SIZE(128), 128 },
216 	{ NULL, "256 Bucket", BUCKET_SIZE(256), 64 },
217 	{ NULL, NULL, 0}
218 };
219 
220 /*
221  * Flags and enumerations to be passed to internal functions.
222  */
223 enum zfreeskip { SKIP_NONE = 0, SKIP_DTOR, SKIP_FINI };
224 
225 /* Prototypes.. */
226 
227 static void *noobj_alloc(uma_zone_t, vm_size_t, uint8_t *, int);
228 static void *page_alloc(uma_zone_t, vm_size_t, uint8_t *, int);
229 static void *startup_alloc(uma_zone_t, vm_size_t, uint8_t *, int);
230 static void page_free(void *, vm_size_t, uint8_t);
231 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int);
232 static void cache_drain(uma_zone_t);
233 static void bucket_drain(uma_zone_t, uma_bucket_t);
234 static void bucket_cache_drain(uma_zone_t zone);
235 static int keg_ctor(void *, int, void *, int);
236 static void keg_dtor(void *, int, void *);
237 static int zone_ctor(void *, int, void *, int);
238 static void zone_dtor(void *, int, void *);
239 static int zero_init(void *, int, int);
240 static void keg_small_init(uma_keg_t keg);
241 static void keg_large_init(uma_keg_t keg);
242 static void zone_foreach(void (*zfunc)(uma_zone_t));
243 static void zone_timeout(uma_zone_t zone);
244 static int hash_alloc(struct uma_hash *);
245 static int hash_expand(struct uma_hash *, struct uma_hash *);
246 static void hash_free(struct uma_hash *hash);
247 static void uma_timeout(void *);
248 static void uma_startup3(void);
249 static void *zone_alloc_item(uma_zone_t, void *, int);
250 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip);
251 static void bucket_enable(void);
252 static void bucket_init(void);
253 static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int);
254 static void bucket_free(uma_zone_t zone, uma_bucket_t, void *);
255 static void bucket_zone_drain(void);
256 static uma_bucket_t zone_alloc_bucket(uma_zone_t zone, void *, int flags);
257 static uma_slab_t zone_fetch_slab(uma_zone_t zone, uma_keg_t last, int flags);
258 static uma_slab_t zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int flags);
259 static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab);
260 static void slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item);
261 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit,
262     uma_fini fini, int align, uint32_t flags);
263 static int zone_import(uma_zone_t zone, void **bucket, int max, int flags);
264 static void zone_release(uma_zone_t zone, void **bucket, int cnt);
265 static void uma_zero_item(void *item, uma_zone_t zone);
266 
267 void uma_print_zone(uma_zone_t);
268 void uma_print_stats(void);
269 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS);
270 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS);
271 
272 #ifdef INVARIANTS
273 static void uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item);
274 static void uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item);
275 #endif
276 
277 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
278 
279 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLTYPE_INT,
280     0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones");
281 
282 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLTYPE_STRUCT,
283     0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats");
284 
285 static int zone_warnings = 1;
286 SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RWTUN, &zone_warnings, 0,
287     "Warn when UMA zones becomes full");
288 
289 /*
290  * This routine checks to see whether or not it's safe to enable buckets.
291  */
292 static void
293 bucket_enable(void)
294 {
295 	bucketdisable = vm_page_count_min();
296 }
297 
298 /*
299  * Initialize bucket_zones, the array of zones of buckets of various sizes.
300  *
301  * For each zone, calculate the memory required for each bucket, consisting
302  * of the header and an array of pointers.
303  */
304 static void
305 bucket_init(void)
306 {
307 	struct uma_bucket_zone *ubz;
308 	int size;
309 
310 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) {
311 		size = roundup(sizeof(struct uma_bucket), sizeof(void *));
312 		size += sizeof(void *) * ubz->ubz_entries;
313 		ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
314 		    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
315 		    UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET);
316 	}
317 }
318 
319 /*
320  * Given a desired number of entries for a bucket, return the zone from which
321  * to allocate the bucket.
322  */
323 static struct uma_bucket_zone *
324 bucket_zone_lookup(int entries)
325 {
326 	struct uma_bucket_zone *ubz;
327 
328 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
329 		if (ubz->ubz_entries >= entries)
330 			return (ubz);
331 	ubz--;
332 	return (ubz);
333 }
334 
335 static int
336 bucket_select(int size)
337 {
338 	struct uma_bucket_zone *ubz;
339 
340 	ubz = &bucket_zones[0];
341 	if (size > ubz->ubz_maxsize)
342 		return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1);
343 
344 	for (; ubz->ubz_entries != 0; ubz++)
345 		if (ubz->ubz_maxsize < size)
346 			break;
347 	ubz--;
348 	return (ubz->ubz_entries);
349 }
350 
351 static uma_bucket_t
352 bucket_alloc(uma_zone_t zone, void *udata, int flags)
353 {
354 	struct uma_bucket_zone *ubz;
355 	uma_bucket_t bucket;
356 
357 	/*
358 	 * This is to stop us from allocating per cpu buckets while we're
359 	 * running out of vm.boot_pages.  Otherwise, we would exhaust the
360 	 * boot pages.  This also prevents us from allocating buckets in
361 	 * low memory situations.
362 	 */
363 	if (bucketdisable)
364 		return (NULL);
365 	/*
366 	 * To limit bucket recursion we store the original zone flags
367 	 * in a cookie passed via zalloc_arg/zfree_arg.  This allows the
368 	 * NOVM flag to persist even through deep recursions.  We also
369 	 * store ZFLAG_BUCKET once we have recursed attempting to allocate
370 	 * a bucket for a bucket zone so we do not allow infinite bucket
371 	 * recursion.  This cookie will even persist to frees of unused
372 	 * buckets via the allocation path or bucket allocations in the
373 	 * free path.
374 	 */
375 	if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
376 		udata = (void *)(uintptr_t)zone->uz_flags;
377 	else {
378 		if ((uintptr_t)udata & UMA_ZFLAG_BUCKET)
379 			return (NULL);
380 		udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET);
381 	}
382 	if ((uintptr_t)udata & UMA_ZFLAG_CACHEONLY)
383 		flags |= M_NOVM;
384 	ubz = bucket_zone_lookup(zone->uz_count);
385 	if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0)
386 		ubz++;
387 	bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags);
388 	if (bucket) {
389 #ifdef INVARIANTS
390 		bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
391 #endif
392 		bucket->ub_cnt = 0;
393 		bucket->ub_entries = ubz->ubz_entries;
394 	}
395 
396 	return (bucket);
397 }
398 
399 static void
400 bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata)
401 {
402 	struct uma_bucket_zone *ubz;
403 
404 	KASSERT(bucket->ub_cnt == 0,
405 	    ("bucket_free: Freeing a non free bucket."));
406 	if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
407 		udata = (void *)(uintptr_t)zone->uz_flags;
408 	ubz = bucket_zone_lookup(bucket->ub_entries);
409 	uma_zfree_arg(ubz->ubz_zone, bucket, udata);
410 }
411 
412 static void
413 bucket_zone_drain(void)
414 {
415 	struct uma_bucket_zone *ubz;
416 
417 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
418 		zone_drain(ubz->ubz_zone);
419 }
420 
421 static void
422 zone_log_warning(uma_zone_t zone)
423 {
424 	static const struct timeval warninterval = { 300, 0 };
425 
426 	if (!zone_warnings || zone->uz_warning == NULL)
427 		return;
428 
429 	if (ratecheck(&zone->uz_ratecheck, &warninterval))
430 		printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning);
431 }
432 
433 static inline void
434 zone_maxaction(uma_zone_t zone)
435 {
436 
437 	if (zone->uz_maxaction.ta_func != NULL)
438 		taskqueue_enqueue(taskqueue_thread, &zone->uz_maxaction);
439 }
440 
441 static void
442 zone_foreach_keg(uma_zone_t zone, void (*kegfn)(uma_keg_t))
443 {
444 	uma_klink_t klink;
445 
446 	LIST_FOREACH(klink, &zone->uz_kegs, kl_link)
447 		kegfn(klink->kl_keg);
448 }
449 
450 /*
451  * Routine called by timeout which is used to fire off some time interval
452  * based calculations.  (stats, hash size, etc.)
453  *
454  * Arguments:
455  *	arg   Unused
456  *
457  * Returns:
458  *	Nothing
459  */
460 static void
461 uma_timeout(void *unused)
462 {
463 	bucket_enable();
464 	zone_foreach(zone_timeout);
465 
466 	/* Reschedule this event */
467 	callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
468 }
469 
470 /*
471  * Routine to perform timeout driven calculations.  This expands the
472  * hashes and does per cpu statistics aggregation.
473  *
474  *  Returns nothing.
475  */
476 static void
477 keg_timeout(uma_keg_t keg)
478 {
479 
480 	KEG_LOCK(keg);
481 	/*
482 	 * Expand the keg hash table.
483 	 *
484 	 * This is done if the number of slabs is larger than the hash size.
485 	 * What I'm trying to do here is completely reduce collisions.  This
486 	 * may be a little aggressive.  Should I allow for two collisions max?
487 	 */
488 	if (keg->uk_flags & UMA_ZONE_HASH &&
489 	    keg->uk_pages / keg->uk_ppera >= keg->uk_hash.uh_hashsize) {
490 		struct uma_hash newhash;
491 		struct uma_hash oldhash;
492 		int ret;
493 
494 		/*
495 		 * This is so involved because allocating and freeing
496 		 * while the keg lock is held will lead to deadlock.
497 		 * I have to do everything in stages and check for
498 		 * races.
499 		 */
500 		newhash = keg->uk_hash;
501 		KEG_UNLOCK(keg);
502 		ret = hash_alloc(&newhash);
503 		KEG_LOCK(keg);
504 		if (ret) {
505 			if (hash_expand(&keg->uk_hash, &newhash)) {
506 				oldhash = keg->uk_hash;
507 				keg->uk_hash = newhash;
508 			} else
509 				oldhash = newhash;
510 
511 			KEG_UNLOCK(keg);
512 			hash_free(&oldhash);
513 			return;
514 		}
515 	}
516 	KEG_UNLOCK(keg);
517 }
518 
519 static void
520 zone_timeout(uma_zone_t zone)
521 {
522 
523 	zone_foreach_keg(zone, &keg_timeout);
524 }
525 
526 /*
527  * Allocate and zero fill the next sized hash table from the appropriate
528  * backing store.
529  *
530  * Arguments:
531  *	hash  A new hash structure with the old hash size in uh_hashsize
532  *
533  * Returns:
534  *	1 on success and 0 on failure.
535  */
536 static int
537 hash_alloc(struct uma_hash *hash)
538 {
539 	int oldsize;
540 	int alloc;
541 
542 	oldsize = hash->uh_hashsize;
543 
544 	/* We're just going to go to a power of two greater */
545 	if (oldsize)  {
546 		hash->uh_hashsize = oldsize * 2;
547 		alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
548 		hash->uh_slab_hash = (struct slabhead *)malloc(alloc,
549 		    M_UMAHASH, M_NOWAIT);
550 	} else {
551 		alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
552 		hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
553 		    M_WAITOK);
554 		hash->uh_hashsize = UMA_HASH_SIZE_INIT;
555 	}
556 	if (hash->uh_slab_hash) {
557 		bzero(hash->uh_slab_hash, alloc);
558 		hash->uh_hashmask = hash->uh_hashsize - 1;
559 		return (1);
560 	}
561 
562 	return (0);
563 }
564 
565 /*
566  * Expands the hash table for HASH zones.  This is done from zone_timeout
567  * to reduce collisions.  This must not be done in the regular allocation
568  * path, otherwise, we can recurse on the vm while allocating pages.
569  *
570  * Arguments:
571  *	oldhash  The hash you want to expand
572  *	newhash  The hash structure for the new table
573  *
574  * Returns:
575  *	Nothing
576  *
577  * Discussion:
578  */
579 static int
580 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
581 {
582 	uma_slab_t slab;
583 	int hval;
584 	int i;
585 
586 	if (!newhash->uh_slab_hash)
587 		return (0);
588 
589 	if (oldhash->uh_hashsize >= newhash->uh_hashsize)
590 		return (0);
591 
592 	/*
593 	 * I need to investigate hash algorithms for resizing without a
594 	 * full rehash.
595 	 */
596 
597 	for (i = 0; i < oldhash->uh_hashsize; i++)
598 		while (!SLIST_EMPTY(&oldhash->uh_slab_hash[i])) {
599 			slab = SLIST_FIRST(&oldhash->uh_slab_hash[i]);
600 			SLIST_REMOVE_HEAD(&oldhash->uh_slab_hash[i], us_hlink);
601 			hval = UMA_HASH(newhash, slab->us_data);
602 			SLIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
603 			    slab, us_hlink);
604 		}
605 
606 	return (1);
607 }
608 
609 /*
610  * Free the hash bucket to the appropriate backing store.
611  *
612  * Arguments:
613  *	slab_hash  The hash bucket we're freeing
614  *	hashsize   The number of entries in that hash bucket
615  *
616  * Returns:
617  *	Nothing
618  */
619 static void
620 hash_free(struct uma_hash *hash)
621 {
622 	if (hash->uh_slab_hash == NULL)
623 		return;
624 	if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
625 		zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE);
626 	else
627 		free(hash->uh_slab_hash, M_UMAHASH);
628 }
629 
630 /*
631  * Frees all outstanding items in a bucket
632  *
633  * Arguments:
634  *	zone   The zone to free to, must be unlocked.
635  *	bucket The free/alloc bucket with items, cpu queue must be locked.
636  *
637  * Returns:
638  *	Nothing
639  */
640 
641 static void
642 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
643 {
644 	int i;
645 
646 	if (bucket == NULL)
647 		return;
648 
649 	if (zone->uz_fini)
650 		for (i = 0; i < bucket->ub_cnt; i++)
651 			zone->uz_fini(bucket->ub_bucket[i], zone->uz_size);
652 	zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt);
653 	bucket->ub_cnt = 0;
654 }
655 
656 /*
657  * Drains the per cpu caches for a zone.
658  *
659  * NOTE: This may only be called while the zone is being turn down, and not
660  * during normal operation.  This is necessary in order that we do not have
661  * to migrate CPUs to drain the per-CPU caches.
662  *
663  * Arguments:
664  *	zone     The zone to drain, must be unlocked.
665  *
666  * Returns:
667  *	Nothing
668  */
669 static void
670 cache_drain(uma_zone_t zone)
671 {
672 	uma_cache_t cache;
673 	int cpu;
674 
675 	/*
676 	 * XXX: It is safe to not lock the per-CPU caches, because we're
677 	 * tearing down the zone anyway.  I.e., there will be no further use
678 	 * of the caches at this point.
679 	 *
680 	 * XXX: It would good to be able to assert that the zone is being
681 	 * torn down to prevent improper use of cache_drain().
682 	 *
683 	 * XXX: We lock the zone before passing into bucket_cache_drain() as
684 	 * it is used elsewhere.  Should the tear-down path be made special
685 	 * there in some form?
686 	 */
687 	CPU_FOREACH(cpu) {
688 		cache = &zone->uz_cpu[cpu];
689 		bucket_drain(zone, cache->uc_allocbucket);
690 		bucket_drain(zone, cache->uc_freebucket);
691 		if (cache->uc_allocbucket != NULL)
692 			bucket_free(zone, cache->uc_allocbucket, NULL);
693 		if (cache->uc_freebucket != NULL)
694 			bucket_free(zone, cache->uc_freebucket, NULL);
695 		cache->uc_allocbucket = cache->uc_freebucket = NULL;
696 	}
697 	ZONE_LOCK(zone);
698 	bucket_cache_drain(zone);
699 	ZONE_UNLOCK(zone);
700 }
701 
702 static void
703 cache_shrink(uma_zone_t zone)
704 {
705 
706 	if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
707 		return;
708 
709 	ZONE_LOCK(zone);
710 	zone->uz_count = (zone->uz_count_min + zone->uz_count) / 2;
711 	ZONE_UNLOCK(zone);
712 }
713 
714 static void
715 cache_drain_safe_cpu(uma_zone_t zone)
716 {
717 	uma_cache_t cache;
718 	uma_bucket_t b1, b2;
719 
720 	if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
721 		return;
722 
723 	b1 = b2 = NULL;
724 	ZONE_LOCK(zone);
725 	critical_enter();
726 	cache = &zone->uz_cpu[curcpu];
727 	if (cache->uc_allocbucket) {
728 		if (cache->uc_allocbucket->ub_cnt != 0)
729 			LIST_INSERT_HEAD(&zone->uz_buckets,
730 			    cache->uc_allocbucket, ub_link);
731 		else
732 			b1 = cache->uc_allocbucket;
733 		cache->uc_allocbucket = NULL;
734 	}
735 	if (cache->uc_freebucket) {
736 		if (cache->uc_freebucket->ub_cnt != 0)
737 			LIST_INSERT_HEAD(&zone->uz_buckets,
738 			    cache->uc_freebucket, ub_link);
739 		else
740 			b2 = cache->uc_freebucket;
741 		cache->uc_freebucket = NULL;
742 	}
743 	critical_exit();
744 	ZONE_UNLOCK(zone);
745 	if (b1)
746 		bucket_free(zone, b1, NULL);
747 	if (b2)
748 		bucket_free(zone, b2, NULL);
749 }
750 
751 /*
752  * Safely drain per-CPU caches of a zone(s) to alloc bucket.
753  * This is an expensive call because it needs to bind to all CPUs
754  * one by one and enter a critical section on each of them in order
755  * to safely access their cache buckets.
756  * Zone lock must not be held on call this function.
757  */
758 static void
759 cache_drain_safe(uma_zone_t zone)
760 {
761 	int cpu;
762 
763 	/*
764 	 * Polite bucket sizes shrinking was not enouth, shrink aggressively.
765 	 */
766 	if (zone)
767 		cache_shrink(zone);
768 	else
769 		zone_foreach(cache_shrink);
770 
771 	CPU_FOREACH(cpu) {
772 		thread_lock(curthread);
773 		sched_bind(curthread, cpu);
774 		thread_unlock(curthread);
775 
776 		if (zone)
777 			cache_drain_safe_cpu(zone);
778 		else
779 			zone_foreach(cache_drain_safe_cpu);
780 	}
781 	thread_lock(curthread);
782 	sched_unbind(curthread);
783 	thread_unlock(curthread);
784 }
785 
786 /*
787  * Drain the cached buckets from a zone.  Expects a locked zone on entry.
788  */
789 static void
790 bucket_cache_drain(uma_zone_t zone)
791 {
792 	uma_bucket_t bucket;
793 
794 	/*
795 	 * Drain the bucket queues and free the buckets, we just keep two per
796 	 * cpu (alloc/free).
797 	 */
798 	while ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) {
799 		LIST_REMOVE(bucket, ub_link);
800 		ZONE_UNLOCK(zone);
801 		bucket_drain(zone, bucket);
802 		bucket_free(zone, bucket, NULL);
803 		ZONE_LOCK(zone);
804 	}
805 
806 	/*
807 	 * Shrink further bucket sizes.  Price of single zone lock collision
808 	 * is probably lower then price of global cache drain.
809 	 */
810 	if (zone->uz_count > zone->uz_count_min)
811 		zone->uz_count--;
812 }
813 
814 static void
815 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start)
816 {
817 	uint8_t *mem;
818 	int i;
819 	uint8_t flags;
820 
821 	mem = slab->us_data;
822 	flags = slab->us_flags;
823 	i = start;
824 	if (keg->uk_fini != NULL) {
825 		for (i--; i > -1; i--)
826 			keg->uk_fini(slab->us_data + (keg->uk_rsize * i),
827 			    keg->uk_size);
828 	}
829 	if (keg->uk_flags & UMA_ZONE_OFFPAGE)
830 		zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE);
831 #ifdef UMA_DEBUG
832 	printf("%s: Returning %d bytes.\n", keg->uk_name,
833 	    PAGE_SIZE * keg->uk_ppera);
834 #endif
835 	keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags);
836 }
837 
838 /*
839  * Frees pages from a keg back to the system.  This is done on demand from
840  * the pageout daemon.
841  *
842  * Returns nothing.
843  */
844 static void
845 keg_drain(uma_keg_t keg)
846 {
847 	struct slabhead freeslabs = { 0 };
848 	uma_slab_t slab;
849 	uma_slab_t n;
850 
851 	/*
852 	 * We don't want to take pages from statically allocated kegs at this
853 	 * time
854 	 */
855 	if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL)
856 		return;
857 
858 #ifdef UMA_DEBUG
859 	printf("%s free items: %u\n", keg->uk_name, keg->uk_free);
860 #endif
861 	KEG_LOCK(keg);
862 	if (keg->uk_free == 0)
863 		goto finished;
864 
865 	slab = LIST_FIRST(&keg->uk_free_slab);
866 	while (slab) {
867 		n = LIST_NEXT(slab, us_link);
868 
869 		/* We have no where to free these to */
870 		if (slab->us_flags & UMA_SLAB_BOOT) {
871 			slab = n;
872 			continue;
873 		}
874 
875 		LIST_REMOVE(slab, us_link);
876 		keg->uk_pages -= keg->uk_ppera;
877 		keg->uk_free -= keg->uk_ipers;
878 
879 		if (keg->uk_flags & UMA_ZONE_HASH)
880 			UMA_HASH_REMOVE(&keg->uk_hash, slab, slab->us_data);
881 
882 		SLIST_INSERT_HEAD(&freeslabs, slab, us_hlink);
883 
884 		slab = n;
885 	}
886 finished:
887 	KEG_UNLOCK(keg);
888 
889 	while ((slab = SLIST_FIRST(&freeslabs)) != NULL) {
890 		SLIST_REMOVE(&freeslabs, slab, uma_slab, us_hlink);
891 		keg_free_slab(keg, slab, keg->uk_ipers);
892 	}
893 }
894 
895 static void
896 zone_drain_wait(uma_zone_t zone, int waitok)
897 {
898 
899 	/*
900 	 * Set draining to interlock with zone_dtor() so we can release our
901 	 * locks as we go.  Only dtor() should do a WAITOK call since it
902 	 * is the only call that knows the structure will still be available
903 	 * when it wakes up.
904 	 */
905 	ZONE_LOCK(zone);
906 	while (zone->uz_flags & UMA_ZFLAG_DRAINING) {
907 		if (waitok == M_NOWAIT)
908 			goto out;
909 		msleep(zone, zone->uz_lockptr, PVM, "zonedrain", 1);
910 	}
911 	zone->uz_flags |= UMA_ZFLAG_DRAINING;
912 	bucket_cache_drain(zone);
913 	ZONE_UNLOCK(zone);
914 	/*
915 	 * The DRAINING flag protects us from being freed while
916 	 * we're running.  Normally the uma_rwlock would protect us but we
917 	 * must be able to release and acquire the right lock for each keg.
918 	 */
919 	zone_foreach_keg(zone, &keg_drain);
920 	ZONE_LOCK(zone);
921 	zone->uz_flags &= ~UMA_ZFLAG_DRAINING;
922 	wakeup(zone);
923 out:
924 	ZONE_UNLOCK(zone);
925 }
926 
927 void
928 zone_drain(uma_zone_t zone)
929 {
930 
931 	zone_drain_wait(zone, M_NOWAIT);
932 }
933 
934 /*
935  * Allocate a new slab for a keg.  This does not insert the slab onto a list.
936  *
937  * Arguments:
938  *	wait  Shall we wait?
939  *
940  * Returns:
941  *	The slab that was allocated or NULL if there is no memory and the
942  *	caller specified M_NOWAIT.
943  */
944 static uma_slab_t
945 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int wait)
946 {
947 	uma_alloc allocf;
948 	uma_slab_t slab;
949 	uint8_t *mem;
950 	uint8_t flags;
951 	int i;
952 
953 	mtx_assert(&keg->uk_lock, MA_OWNED);
954 	slab = NULL;
955 	mem = NULL;
956 
957 #ifdef UMA_DEBUG
958 	printf("alloc_slab:  Allocating a new slab for %s\n", keg->uk_name);
959 #endif
960 	allocf = keg->uk_allocf;
961 	KEG_UNLOCK(keg);
962 
963 	if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
964 		slab = zone_alloc_item(keg->uk_slabzone, NULL, wait);
965 		if (slab == NULL)
966 			goto out;
967 	}
968 
969 	/*
970 	 * This reproduces the old vm_zone behavior of zero filling pages the
971 	 * first time they are added to a zone.
972 	 *
973 	 * Malloced items are zeroed in uma_zalloc.
974 	 */
975 
976 	if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
977 		wait |= M_ZERO;
978 	else
979 		wait &= ~M_ZERO;
980 
981 	if (keg->uk_flags & UMA_ZONE_NODUMP)
982 		wait |= M_NODUMP;
983 
984 	/* zone is passed for legacy reasons. */
985 	mem = allocf(zone, keg->uk_ppera * PAGE_SIZE, &flags, wait);
986 	if (mem == NULL) {
987 		if (keg->uk_flags & UMA_ZONE_OFFPAGE)
988 			zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE);
989 		slab = NULL;
990 		goto out;
991 	}
992 
993 	/* Point the slab into the allocated memory */
994 	if (!(keg->uk_flags & UMA_ZONE_OFFPAGE))
995 		slab = (uma_slab_t )(mem + keg->uk_pgoff);
996 
997 	if (keg->uk_flags & UMA_ZONE_VTOSLAB)
998 		for (i = 0; i < keg->uk_ppera; i++)
999 			vsetslab((vm_offset_t)mem + (i * PAGE_SIZE), slab);
1000 
1001 	slab->us_keg = keg;
1002 	slab->us_data = mem;
1003 	slab->us_freecount = keg->uk_ipers;
1004 	slab->us_flags = flags;
1005 	BIT_FILL(SLAB_SETSIZE, &slab->us_free);
1006 #ifdef INVARIANTS
1007 	BIT_ZERO(SLAB_SETSIZE, &slab->us_debugfree);
1008 #endif
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, vm_size_t 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, vm_size_t 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, vm_size_t 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_NONE);
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, vm_size_t 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_OFFPAGE)
1257 		shsize = 0;
1258 	else
1259 		shsize = sizeof(struct uma_slab);
1260 
1261 	keg->uk_ipers = (keg->uk_slabsize - shsize) / rsize;
1262 	KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE,
1263 	    ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers));
1264 
1265 	memused = keg->uk_ipers * rsize + shsize;
1266 	wastedspace = keg->uk_slabsize - memused;
1267 
1268 	/*
1269 	 * We can't do OFFPAGE if we're internal or if we've been
1270 	 * asked to not go to the VM for buckets.  If we do this we
1271 	 * may end up going to the VM  for slabs which we do not
1272 	 * want to do if we're UMA_ZFLAG_CACHEONLY as a result
1273 	 * of UMA_ZONE_VM, which clearly forbids it.
1274 	 */
1275 	if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) ||
1276 	    (keg->uk_flags & UMA_ZFLAG_CACHEONLY))
1277 		return;
1278 
1279 	/*
1280 	 * See if using an OFFPAGE slab will limit our waste.  Only do
1281 	 * this if it permits more items per-slab.
1282 	 *
1283 	 * XXX We could try growing slabsize to limit max waste as well.
1284 	 * Historically this was not done because the VM could not
1285 	 * efficiently handle contiguous allocations.
1286 	 */
1287 	if ((wastedspace >= keg->uk_slabsize / UMA_MAX_WASTE) &&
1288 	    (keg->uk_ipers < (keg->uk_slabsize / keg->uk_rsize))) {
1289 		keg->uk_ipers = keg->uk_slabsize / keg->uk_rsize;
1290 		KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE,
1291 		    ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers));
1292 #ifdef UMA_DEBUG
1293 		printf("UMA decided we need offpage slab headers for "
1294 		    "keg: %s, calculated wastedspace = %d, "
1295 		    "maximum wasted space allowed = %d, "
1296 		    "calculated ipers = %d, "
1297 		    "new wasted space = %d\n", keg->uk_name, wastedspace,
1298 		    keg->uk_slabsize / UMA_MAX_WASTE, keg->uk_ipers,
1299 		    keg->uk_slabsize - keg->uk_ipers * keg->uk_rsize);
1300 #endif
1301 		keg->uk_flags |= UMA_ZONE_OFFPAGE;
1302 	}
1303 
1304 	if ((keg->uk_flags & UMA_ZONE_OFFPAGE) &&
1305 	    (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1306 		keg->uk_flags |= UMA_ZONE_HASH;
1307 }
1308 
1309 /*
1310  * Finish creating a large (> UMA_SLAB_SIZE) uma kegs.  Just give in and do
1311  * OFFPAGE for now.  When I can allow for more dynamic slab sizes this will be
1312  * more complicated.
1313  *
1314  * Arguments
1315  *	keg  The keg we should initialize
1316  *
1317  * Returns
1318  *	Nothing
1319  */
1320 static void
1321 keg_large_init(uma_keg_t keg)
1322 {
1323 	u_int shsize;
1324 
1325 	KASSERT(keg != NULL, ("Keg is null in keg_large_init"));
1326 	KASSERT((keg->uk_flags & UMA_ZFLAG_CACHEONLY) == 0,
1327 	    ("keg_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY keg"));
1328 	KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0,
1329 	    ("%s: Cannot large-init a UMA_ZONE_PCPU keg", __func__));
1330 
1331 	keg->uk_ppera = howmany(keg->uk_size, PAGE_SIZE);
1332 	keg->uk_slabsize = keg->uk_ppera * PAGE_SIZE;
1333 	keg->uk_ipers = 1;
1334 	keg->uk_rsize = keg->uk_size;
1335 
1336 	/* We can't do OFFPAGE if we're internal, bail out here. */
1337 	if (keg->uk_flags & UMA_ZFLAG_INTERNAL)
1338 		return;
1339 
1340 	/* Check whether we have enough space to not do OFFPAGE. */
1341 	if ((keg->uk_flags & UMA_ZONE_OFFPAGE) == 0) {
1342 		shsize = sizeof(struct uma_slab);
1343 		if (shsize & UMA_ALIGN_PTR)
1344 			shsize = (shsize & ~UMA_ALIGN_PTR) +
1345 			    (UMA_ALIGN_PTR + 1);
1346 
1347 		if ((PAGE_SIZE * keg->uk_ppera) - keg->uk_rsize < shsize)
1348 			keg->uk_flags |= UMA_ZONE_OFFPAGE;
1349 	}
1350 
1351 	if ((keg->uk_flags & UMA_ZONE_OFFPAGE) &&
1352 	    (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1353 		keg->uk_flags |= UMA_ZONE_HASH;
1354 }
1355 
1356 static void
1357 keg_cachespread_init(uma_keg_t keg)
1358 {
1359 	int alignsize;
1360 	int trailer;
1361 	int pages;
1362 	int rsize;
1363 
1364 	KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0,
1365 	    ("%s: Cannot cachespread-init a UMA_ZONE_PCPU keg", __func__));
1366 
1367 	alignsize = keg->uk_align + 1;
1368 	rsize = keg->uk_size;
1369 	/*
1370 	 * We want one item to start on every align boundary in a page.  To
1371 	 * do this we will span pages.  We will also extend the item by the
1372 	 * size of align if it is an even multiple of align.  Otherwise, it
1373 	 * would fall on the same boundary every time.
1374 	 */
1375 	if (rsize & keg->uk_align)
1376 		rsize = (rsize & ~keg->uk_align) + alignsize;
1377 	if ((rsize & alignsize) == 0)
1378 		rsize += alignsize;
1379 	trailer = rsize - keg->uk_size;
1380 	pages = (rsize * (PAGE_SIZE / alignsize)) / PAGE_SIZE;
1381 	pages = MIN(pages, (128 * 1024) / PAGE_SIZE);
1382 	keg->uk_rsize = rsize;
1383 	keg->uk_ppera = pages;
1384 	keg->uk_slabsize = UMA_SLAB_SIZE;
1385 	keg->uk_ipers = ((pages * PAGE_SIZE) + trailer) / rsize;
1386 	keg->uk_flags |= UMA_ZONE_OFFPAGE | UMA_ZONE_VTOSLAB;
1387 	KASSERT(keg->uk_ipers <= SLAB_SETSIZE,
1388 	    ("%s: keg->uk_ipers too high(%d) increase max_ipers", __func__,
1389 	    keg->uk_ipers));
1390 }
1391 
1392 /*
1393  * Keg header ctor.  This initializes all fields, locks, etc.  And inserts
1394  * the keg onto the global keg list.
1395  *
1396  * Arguments/Returns follow uma_ctor specifications
1397  *	udata  Actually uma_kctor_args
1398  */
1399 static int
1400 keg_ctor(void *mem, int size, void *udata, int flags)
1401 {
1402 	struct uma_kctor_args *arg = udata;
1403 	uma_keg_t keg = mem;
1404 	uma_zone_t zone;
1405 
1406 	bzero(keg, size);
1407 	keg->uk_size = arg->size;
1408 	keg->uk_init = arg->uminit;
1409 	keg->uk_fini = arg->fini;
1410 	keg->uk_align = arg->align;
1411 	keg->uk_free = 0;
1412 	keg->uk_reserve = 0;
1413 	keg->uk_pages = 0;
1414 	keg->uk_flags = arg->flags;
1415 	keg->uk_allocf = page_alloc;
1416 	keg->uk_freef = page_free;
1417 	keg->uk_slabzone = NULL;
1418 
1419 	/*
1420 	 * The master zone is passed to us at keg-creation time.
1421 	 */
1422 	zone = arg->zone;
1423 	keg->uk_name = zone->uz_name;
1424 
1425 	if (arg->flags & UMA_ZONE_VM)
1426 		keg->uk_flags |= UMA_ZFLAG_CACHEONLY;
1427 
1428 	if (arg->flags & UMA_ZONE_ZINIT)
1429 		keg->uk_init = zero_init;
1430 
1431 	if (arg->flags & UMA_ZONE_MALLOC)
1432 		keg->uk_flags |= UMA_ZONE_VTOSLAB;
1433 
1434 	if (arg->flags & UMA_ZONE_PCPU)
1435 #ifdef SMP
1436 		keg->uk_flags |= UMA_ZONE_OFFPAGE;
1437 #else
1438 		keg->uk_flags &= ~UMA_ZONE_PCPU;
1439 #endif
1440 
1441 	if (keg->uk_flags & UMA_ZONE_CACHESPREAD) {
1442 		keg_cachespread_init(keg);
1443 	} else {
1444 		if (keg->uk_size > (UMA_SLAB_SIZE - sizeof(struct uma_slab)))
1445 			keg_large_init(keg);
1446 		else
1447 			keg_small_init(keg);
1448 	}
1449 
1450 	if (keg->uk_flags & UMA_ZONE_OFFPAGE)
1451 		keg->uk_slabzone = slabzone;
1452 
1453 	/*
1454 	 * If we haven't booted yet we need allocations to go through the
1455 	 * startup cache until the vm is ready.
1456 	 */
1457 	if (keg->uk_ppera == 1) {
1458 #ifdef UMA_MD_SMALL_ALLOC
1459 		keg->uk_allocf = uma_small_alloc;
1460 		keg->uk_freef = uma_small_free;
1461 
1462 		if (booted < UMA_STARTUP)
1463 			keg->uk_allocf = startup_alloc;
1464 #else
1465 		if (booted < UMA_STARTUP2)
1466 			keg->uk_allocf = startup_alloc;
1467 #endif
1468 	} else if (booted < UMA_STARTUP2 &&
1469 	    (keg->uk_flags & UMA_ZFLAG_INTERNAL))
1470 		keg->uk_allocf = startup_alloc;
1471 
1472 	/*
1473 	 * Initialize keg's lock
1474 	 */
1475 	KEG_LOCK_INIT(keg, (arg->flags & UMA_ZONE_MTXCLASS));
1476 
1477 	/*
1478 	 * If we're putting the slab header in the actual page we need to
1479 	 * figure out where in each page it goes.  This calculates a right
1480 	 * justified offset into the memory on an ALIGN_PTR boundary.
1481 	 */
1482 	if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) {
1483 		u_int totsize;
1484 
1485 		/* Size of the slab struct and free list */
1486 		totsize = sizeof(struct uma_slab);
1487 
1488 		if (totsize & UMA_ALIGN_PTR)
1489 			totsize = (totsize & ~UMA_ALIGN_PTR) +
1490 			    (UMA_ALIGN_PTR + 1);
1491 		keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - totsize;
1492 
1493 		/*
1494 		 * The only way the following is possible is if with our
1495 		 * UMA_ALIGN_PTR adjustments we are now bigger than
1496 		 * UMA_SLAB_SIZE.  I haven't checked whether this is
1497 		 * mathematically possible for all cases, so we make
1498 		 * sure here anyway.
1499 		 */
1500 		totsize = keg->uk_pgoff + sizeof(struct uma_slab);
1501 		if (totsize > PAGE_SIZE * keg->uk_ppera) {
1502 			printf("zone %s ipers %d rsize %d size %d\n",
1503 			    zone->uz_name, keg->uk_ipers, keg->uk_rsize,
1504 			    keg->uk_size);
1505 			panic("UMA slab won't fit.");
1506 		}
1507 	}
1508 
1509 	if (keg->uk_flags & UMA_ZONE_HASH)
1510 		hash_alloc(&keg->uk_hash);
1511 
1512 #ifdef UMA_DEBUG
1513 	printf("UMA: %s(%p) size %d(%d) flags %#x ipers %d ppera %d out %d free %d\n",
1514 	    zone->uz_name, zone, keg->uk_size, keg->uk_rsize, keg->uk_flags,
1515 	    keg->uk_ipers, keg->uk_ppera,
1516 	    (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free);
1517 #endif
1518 
1519 	LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
1520 
1521 	rw_wlock(&uma_rwlock);
1522 	LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
1523 	rw_wunlock(&uma_rwlock);
1524 	return (0);
1525 }
1526 
1527 /*
1528  * Zone header ctor.  This initializes all fields, locks, etc.
1529  *
1530  * Arguments/Returns follow uma_ctor specifications
1531  *	udata  Actually uma_zctor_args
1532  */
1533 static int
1534 zone_ctor(void *mem, int size, void *udata, int flags)
1535 {
1536 	struct uma_zctor_args *arg = udata;
1537 	uma_zone_t zone = mem;
1538 	uma_zone_t z;
1539 	uma_keg_t keg;
1540 
1541 	bzero(zone, size);
1542 	zone->uz_name = arg->name;
1543 	zone->uz_ctor = arg->ctor;
1544 	zone->uz_dtor = arg->dtor;
1545 	zone->uz_slab = zone_fetch_slab;
1546 	zone->uz_init = NULL;
1547 	zone->uz_fini = NULL;
1548 	zone->uz_allocs = 0;
1549 	zone->uz_frees = 0;
1550 	zone->uz_fails = 0;
1551 	zone->uz_sleeps = 0;
1552 	zone->uz_count = 0;
1553 	zone->uz_count_min = 0;
1554 	zone->uz_flags = 0;
1555 	zone->uz_warning = NULL;
1556 	timevalclear(&zone->uz_ratecheck);
1557 	keg = arg->keg;
1558 
1559 	ZONE_LOCK_INIT(zone, (arg->flags & UMA_ZONE_MTXCLASS));
1560 
1561 	/*
1562 	 * This is a pure cache zone, no kegs.
1563 	 */
1564 	if (arg->import) {
1565 		if (arg->flags & UMA_ZONE_VM)
1566 			arg->flags |= UMA_ZFLAG_CACHEONLY;
1567 		zone->uz_flags = arg->flags;
1568 		zone->uz_size = arg->size;
1569 		zone->uz_import = arg->import;
1570 		zone->uz_release = arg->release;
1571 		zone->uz_arg = arg->arg;
1572 		zone->uz_lockptr = &zone->uz_lock;
1573 		rw_wlock(&uma_rwlock);
1574 		LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link);
1575 		rw_wunlock(&uma_rwlock);
1576 		goto out;
1577 	}
1578 
1579 	/*
1580 	 * Use the regular zone/keg/slab allocator.
1581 	 */
1582 	zone->uz_import = (uma_import)zone_import;
1583 	zone->uz_release = (uma_release)zone_release;
1584 	zone->uz_arg = zone;
1585 
1586 	if (arg->flags & UMA_ZONE_SECONDARY) {
1587 		KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
1588 		zone->uz_init = arg->uminit;
1589 		zone->uz_fini = arg->fini;
1590 		zone->uz_lockptr = &keg->uk_lock;
1591 		zone->uz_flags |= UMA_ZONE_SECONDARY;
1592 		rw_wlock(&uma_rwlock);
1593 		ZONE_LOCK(zone);
1594 		LIST_FOREACH(z, &keg->uk_zones, uz_link) {
1595 			if (LIST_NEXT(z, uz_link) == NULL) {
1596 				LIST_INSERT_AFTER(z, zone, uz_link);
1597 				break;
1598 			}
1599 		}
1600 		ZONE_UNLOCK(zone);
1601 		rw_wunlock(&uma_rwlock);
1602 	} else if (keg == NULL) {
1603 		if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
1604 		    arg->align, arg->flags)) == NULL)
1605 			return (ENOMEM);
1606 	} else {
1607 		struct uma_kctor_args karg;
1608 		int error;
1609 
1610 		/* We should only be here from uma_startup() */
1611 		karg.size = arg->size;
1612 		karg.uminit = arg->uminit;
1613 		karg.fini = arg->fini;
1614 		karg.align = arg->align;
1615 		karg.flags = arg->flags;
1616 		karg.zone = zone;
1617 		error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
1618 		    flags);
1619 		if (error)
1620 			return (error);
1621 	}
1622 
1623 	/*
1624 	 * Link in the first keg.
1625 	 */
1626 	zone->uz_klink.kl_keg = keg;
1627 	LIST_INSERT_HEAD(&zone->uz_kegs, &zone->uz_klink, kl_link);
1628 	zone->uz_lockptr = &keg->uk_lock;
1629 	zone->uz_size = keg->uk_size;
1630 	zone->uz_flags |= (keg->uk_flags &
1631 	    (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
1632 
1633 	/*
1634 	 * Some internal zones don't have room allocated for the per cpu
1635 	 * caches.  If we're internal, bail out here.
1636 	 */
1637 	if (keg->uk_flags & UMA_ZFLAG_INTERNAL) {
1638 		KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
1639 		    ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
1640 		return (0);
1641 	}
1642 
1643 out:
1644 	if ((arg->flags & UMA_ZONE_MAXBUCKET) == 0)
1645 		zone->uz_count = bucket_select(zone->uz_size);
1646 	else
1647 		zone->uz_count = BUCKET_MAX;
1648 	zone->uz_count_min = zone->uz_count;
1649 
1650 	return (0);
1651 }
1652 
1653 /*
1654  * Keg header dtor.  This frees all data, destroys locks, frees the hash
1655  * table and removes the keg from the global list.
1656  *
1657  * Arguments/Returns follow uma_dtor specifications
1658  *	udata  unused
1659  */
1660 static void
1661 keg_dtor(void *arg, int size, void *udata)
1662 {
1663 	uma_keg_t keg;
1664 
1665 	keg = (uma_keg_t)arg;
1666 	KEG_LOCK(keg);
1667 	if (keg->uk_free != 0) {
1668 		printf("Freed UMA keg (%s) was not empty (%d items). "
1669 		    " Lost %d pages of memory.\n",
1670 		    keg->uk_name ? keg->uk_name : "",
1671 		    keg->uk_free, keg->uk_pages);
1672 	}
1673 	KEG_UNLOCK(keg);
1674 
1675 	hash_free(&keg->uk_hash);
1676 
1677 	KEG_LOCK_FINI(keg);
1678 }
1679 
1680 /*
1681  * Zone header dtor.
1682  *
1683  * Arguments/Returns follow uma_dtor specifications
1684  *	udata  unused
1685  */
1686 static void
1687 zone_dtor(void *arg, int size, void *udata)
1688 {
1689 	uma_klink_t klink;
1690 	uma_zone_t zone;
1691 	uma_keg_t keg;
1692 
1693 	zone = (uma_zone_t)arg;
1694 	keg = zone_first_keg(zone);
1695 
1696 	if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
1697 		cache_drain(zone);
1698 
1699 	rw_wlock(&uma_rwlock);
1700 	LIST_REMOVE(zone, uz_link);
1701 	rw_wunlock(&uma_rwlock);
1702 	/*
1703 	 * XXX there are some races here where
1704 	 * the zone can be drained but zone lock
1705 	 * released and then refilled before we
1706 	 * remove it... we dont care for now
1707 	 */
1708 	zone_drain_wait(zone, M_WAITOK);
1709 	/*
1710 	 * Unlink all of our kegs.
1711 	 */
1712 	while ((klink = LIST_FIRST(&zone->uz_kegs)) != NULL) {
1713 		klink->kl_keg = NULL;
1714 		LIST_REMOVE(klink, kl_link);
1715 		if (klink == &zone->uz_klink)
1716 			continue;
1717 		free(klink, M_TEMP);
1718 	}
1719 	/*
1720 	 * We only destroy kegs from non secondary zones.
1721 	 */
1722 	if (keg != NULL && (zone->uz_flags & UMA_ZONE_SECONDARY) == 0)  {
1723 		rw_wlock(&uma_rwlock);
1724 		LIST_REMOVE(keg, uk_link);
1725 		rw_wunlock(&uma_rwlock);
1726 		zone_free_item(kegs, keg, NULL, SKIP_NONE);
1727 	}
1728 	ZONE_LOCK_FINI(zone);
1729 }
1730 
1731 /*
1732  * Traverses every zone in the system and calls a callback
1733  *
1734  * Arguments:
1735  *	zfunc  A pointer to a function which accepts a zone
1736  *		as an argument.
1737  *
1738  * Returns:
1739  *	Nothing
1740  */
1741 static void
1742 zone_foreach(void (*zfunc)(uma_zone_t))
1743 {
1744 	uma_keg_t keg;
1745 	uma_zone_t zone;
1746 
1747 	rw_rlock(&uma_rwlock);
1748 	LIST_FOREACH(keg, &uma_kegs, uk_link) {
1749 		LIST_FOREACH(zone, &keg->uk_zones, uz_link)
1750 			zfunc(zone);
1751 	}
1752 	rw_runlock(&uma_rwlock);
1753 }
1754 
1755 /* Public functions */
1756 /* See uma.h */
1757 void
1758 uma_startup(void *bootmem, int boot_pages)
1759 {
1760 	struct uma_zctor_args args;
1761 	uma_slab_t slab;
1762 	int i;
1763 
1764 #ifdef UMA_DEBUG
1765 	printf("Creating uma keg headers zone and keg.\n");
1766 #endif
1767 	rw_init(&uma_rwlock, "UMA lock");
1768 
1769 	/* "manually" create the initial zone */
1770 	memset(&args, 0, sizeof(args));
1771 	args.name = "UMA Kegs";
1772 	args.size = sizeof(struct uma_keg);
1773 	args.ctor = keg_ctor;
1774 	args.dtor = keg_dtor;
1775 	args.uminit = zero_init;
1776 	args.fini = NULL;
1777 	args.keg = &masterkeg;
1778 	args.align = 32 - 1;
1779 	args.flags = UMA_ZFLAG_INTERNAL;
1780 	/* The initial zone has no Per cpu queues so it's smaller */
1781 	zone_ctor(kegs, sizeof(struct uma_zone), &args, M_WAITOK);
1782 
1783 #ifdef UMA_DEBUG
1784 	printf("Filling boot free list.\n");
1785 #endif
1786 	for (i = 0; i < boot_pages; i++) {
1787 		slab = (uma_slab_t)((uint8_t *)bootmem + (i * UMA_SLAB_SIZE));
1788 		slab->us_data = (uint8_t *)slab;
1789 		slab->us_flags = UMA_SLAB_BOOT;
1790 		LIST_INSERT_HEAD(&uma_boot_pages, slab, us_link);
1791 	}
1792 	mtx_init(&uma_boot_pages_mtx, "UMA boot pages", NULL, MTX_DEF);
1793 
1794 #ifdef UMA_DEBUG
1795 	printf("Creating uma zone headers zone and keg.\n");
1796 #endif
1797 	args.name = "UMA Zones";
1798 	args.size = sizeof(struct uma_zone) +
1799 	    (sizeof(struct uma_cache) * (mp_maxid + 1));
1800 	args.ctor = zone_ctor;
1801 	args.dtor = zone_dtor;
1802 	args.uminit = zero_init;
1803 	args.fini = NULL;
1804 	args.keg = NULL;
1805 	args.align = 32 - 1;
1806 	args.flags = UMA_ZFLAG_INTERNAL;
1807 	/* The initial zone has no Per cpu queues so it's smaller */
1808 	zone_ctor(zones, sizeof(struct uma_zone), &args, M_WAITOK);
1809 
1810 #ifdef UMA_DEBUG
1811 	printf("Creating slab and hash zones.\n");
1812 #endif
1813 
1814 	/* Now make a zone for slab headers */
1815 	slabzone = uma_zcreate("UMA Slabs",
1816 				sizeof(struct uma_slab),
1817 				NULL, NULL, NULL, NULL,
1818 				UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1819 
1820 	hashzone = uma_zcreate("UMA Hash",
1821 	    sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
1822 	    NULL, NULL, NULL, NULL,
1823 	    UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1824 
1825 	bucket_init();
1826 
1827 	booted = UMA_STARTUP;
1828 
1829 #ifdef UMA_DEBUG
1830 	printf("UMA startup complete.\n");
1831 #endif
1832 }
1833 
1834 /* see uma.h */
1835 void
1836 uma_startup2(void)
1837 {
1838 	booted = UMA_STARTUP2;
1839 	bucket_enable();
1840 	sx_init(&uma_drain_lock, "umadrain");
1841 #ifdef UMA_DEBUG
1842 	printf("UMA startup2 complete.\n");
1843 #endif
1844 }
1845 
1846 /*
1847  * Initialize our callout handle
1848  *
1849  */
1850 
1851 static void
1852 uma_startup3(void)
1853 {
1854 #ifdef UMA_DEBUG
1855 	printf("Starting callout.\n");
1856 #endif
1857 	callout_init(&uma_callout, 1);
1858 	callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
1859 #ifdef UMA_DEBUG
1860 	printf("UMA startup3 complete.\n");
1861 #endif
1862 }
1863 
1864 static uma_keg_t
1865 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
1866 		int align, uint32_t flags)
1867 {
1868 	struct uma_kctor_args args;
1869 
1870 	args.size = size;
1871 	args.uminit = uminit;
1872 	args.fini = fini;
1873 	args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
1874 	args.flags = flags;
1875 	args.zone = zone;
1876 	return (zone_alloc_item(kegs, &args, M_WAITOK));
1877 }
1878 
1879 /* See uma.h */
1880 void
1881 uma_set_align(int align)
1882 {
1883 
1884 	if (align != UMA_ALIGN_CACHE)
1885 		uma_align_cache = align;
1886 }
1887 
1888 /* See uma.h */
1889 uma_zone_t
1890 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
1891 		uma_init uminit, uma_fini fini, int align, uint32_t flags)
1892 
1893 {
1894 	struct uma_zctor_args args;
1895 	uma_zone_t res;
1896 	bool locked;
1897 
1898 	/* This stuff is essential for the zone ctor */
1899 	memset(&args, 0, sizeof(args));
1900 	args.name = name;
1901 	args.size = size;
1902 	args.ctor = ctor;
1903 	args.dtor = dtor;
1904 	args.uminit = uminit;
1905 	args.fini = fini;
1906 #ifdef  INVARIANTS
1907 	/*
1908 	 * If a zone is being created with an empty constructor and
1909 	 * destructor, pass UMA constructor/destructor which checks for
1910 	 * memory use after free.
1911 	 */
1912 	if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOFREE))) &&
1913 	    ctor == NULL && dtor == NULL && uminit == NULL && fini == NULL) {
1914 		args.ctor = trash_ctor;
1915 		args.dtor = trash_dtor;
1916 		args.uminit = trash_init;
1917 		args.fini = trash_fini;
1918 	}
1919 #endif
1920 	args.align = align;
1921 	args.flags = flags;
1922 	args.keg = NULL;
1923 
1924 	if (booted < UMA_STARTUP2) {
1925 		locked = false;
1926 	} else {
1927 		sx_slock(&uma_drain_lock);
1928 		locked = true;
1929 	}
1930 	res = zone_alloc_item(zones, &args, M_WAITOK);
1931 	if (locked)
1932 		sx_sunlock(&uma_drain_lock);
1933 	return (res);
1934 }
1935 
1936 /* See uma.h */
1937 uma_zone_t
1938 uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor,
1939 		    uma_init zinit, uma_fini zfini, uma_zone_t master)
1940 {
1941 	struct uma_zctor_args args;
1942 	uma_keg_t keg;
1943 	uma_zone_t res;
1944 	bool locked;
1945 
1946 	keg = zone_first_keg(master);
1947 	memset(&args, 0, sizeof(args));
1948 	args.name = name;
1949 	args.size = keg->uk_size;
1950 	args.ctor = ctor;
1951 	args.dtor = dtor;
1952 	args.uminit = zinit;
1953 	args.fini = zfini;
1954 	args.align = keg->uk_align;
1955 	args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
1956 	args.keg = keg;
1957 
1958 	if (booted < UMA_STARTUP2) {
1959 		locked = false;
1960 	} else {
1961 		sx_slock(&uma_drain_lock);
1962 		locked = true;
1963 	}
1964 	/* XXX Attaches only one keg of potentially many. */
1965 	res = zone_alloc_item(zones, &args, M_WAITOK);
1966 	if (locked)
1967 		sx_sunlock(&uma_drain_lock);
1968 	return (res);
1969 }
1970 
1971 /* See uma.h */
1972 uma_zone_t
1973 uma_zcache_create(char *name, int size, uma_ctor ctor, uma_dtor dtor,
1974 		    uma_init zinit, uma_fini zfini, uma_import zimport,
1975 		    uma_release zrelease, void *arg, int flags)
1976 {
1977 	struct uma_zctor_args args;
1978 
1979 	memset(&args, 0, sizeof(args));
1980 	args.name = name;
1981 	args.size = size;
1982 	args.ctor = ctor;
1983 	args.dtor = dtor;
1984 	args.uminit = zinit;
1985 	args.fini = zfini;
1986 	args.import = zimport;
1987 	args.release = zrelease;
1988 	args.arg = arg;
1989 	args.align = 0;
1990 	args.flags = flags;
1991 
1992 	return (zone_alloc_item(zones, &args, M_WAITOK));
1993 }
1994 
1995 static void
1996 zone_lock_pair(uma_zone_t a, uma_zone_t b)
1997 {
1998 	if (a < b) {
1999 		ZONE_LOCK(a);
2000 		mtx_lock_flags(b->uz_lockptr, MTX_DUPOK);
2001 	} else {
2002 		ZONE_LOCK(b);
2003 		mtx_lock_flags(a->uz_lockptr, MTX_DUPOK);
2004 	}
2005 }
2006 
2007 static void
2008 zone_unlock_pair(uma_zone_t a, uma_zone_t b)
2009 {
2010 
2011 	ZONE_UNLOCK(a);
2012 	ZONE_UNLOCK(b);
2013 }
2014 
2015 int
2016 uma_zsecond_add(uma_zone_t zone, uma_zone_t master)
2017 {
2018 	uma_klink_t klink;
2019 	uma_klink_t kl;
2020 	int error;
2021 
2022 	error = 0;
2023 	klink = malloc(sizeof(*klink), M_TEMP, M_WAITOK | M_ZERO);
2024 
2025 	zone_lock_pair(zone, master);
2026 	/*
2027 	 * zone must use vtoslab() to resolve objects and must already be
2028 	 * a secondary.
2029 	 */
2030 	if ((zone->uz_flags & (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY))
2031 	    != (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) {
2032 		error = EINVAL;
2033 		goto out;
2034 	}
2035 	/*
2036 	 * The new master must also use vtoslab().
2037 	 */
2038 	if ((zone->uz_flags & UMA_ZONE_VTOSLAB) != UMA_ZONE_VTOSLAB) {
2039 		error = EINVAL;
2040 		goto out;
2041 	}
2042 
2043 	/*
2044 	 * The underlying object must be the same size.  rsize
2045 	 * may be different.
2046 	 */
2047 	if (master->uz_size != zone->uz_size) {
2048 		error = E2BIG;
2049 		goto out;
2050 	}
2051 	/*
2052 	 * Put it at the end of the list.
2053 	 */
2054 	klink->kl_keg = zone_first_keg(master);
2055 	LIST_FOREACH(kl, &zone->uz_kegs, kl_link) {
2056 		if (LIST_NEXT(kl, kl_link) == NULL) {
2057 			LIST_INSERT_AFTER(kl, klink, kl_link);
2058 			break;
2059 		}
2060 	}
2061 	klink = NULL;
2062 	zone->uz_flags |= UMA_ZFLAG_MULTI;
2063 	zone->uz_slab = zone_fetch_slab_multi;
2064 
2065 out:
2066 	zone_unlock_pair(zone, master);
2067 	if (klink != NULL)
2068 		free(klink, M_TEMP);
2069 
2070 	return (error);
2071 }
2072 
2073 
2074 /* See uma.h */
2075 void
2076 uma_zdestroy(uma_zone_t zone)
2077 {
2078 
2079 	sx_slock(&uma_drain_lock);
2080 	zone_free_item(zones, zone, NULL, SKIP_NONE);
2081 	sx_sunlock(&uma_drain_lock);
2082 }
2083 
2084 /* See uma.h */
2085 void *
2086 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
2087 {
2088 	void *item;
2089 	uma_cache_t cache;
2090 	uma_bucket_t bucket;
2091 	int lockfail;
2092 	int cpu;
2093 
2094 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
2095 	random_harvest_fast_uma(&zone, sizeof(zone), 1, RANDOM_UMA);
2096 
2097 	/* This is the fast path allocation */
2098 #ifdef UMA_DEBUG_ALLOC_1
2099 	printf("Allocating one item from %s(%p)\n", zone->uz_name, zone);
2100 #endif
2101 	CTR3(KTR_UMA, "uma_zalloc_arg thread %x zone %s flags %d", curthread,
2102 	    zone->uz_name, flags);
2103 
2104 	if (flags & M_WAITOK) {
2105 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
2106 		    "uma_zalloc_arg: zone \"%s\"", zone->uz_name);
2107 	}
2108 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
2109 	    ("uma_zalloc_arg: called with spinlock or critical section held"));
2110 
2111 #ifdef DEBUG_MEMGUARD
2112 	if (memguard_cmp_zone(zone)) {
2113 		item = memguard_alloc(zone->uz_size, flags);
2114 		if (item != NULL) {
2115 			/*
2116 			 * Avoid conflict with the use-after-free
2117 			 * protecting infrastructure from INVARIANTS.
2118 			 */
2119 			if (zone->uz_init != NULL &&
2120 			    zone->uz_init != mtrash_init &&
2121 			    zone->uz_init(item, zone->uz_size, flags) != 0)
2122 				return (NULL);
2123 			if (zone->uz_ctor != NULL &&
2124 			    zone->uz_ctor != mtrash_ctor &&
2125 			    zone->uz_ctor(item, zone->uz_size, udata,
2126 			    flags) != 0) {
2127 			    	zone->uz_fini(item, zone->uz_size);
2128 				return (NULL);
2129 			}
2130 			return (item);
2131 		}
2132 		/* This is unfortunate but should not be fatal. */
2133 	}
2134 #endif
2135 	/*
2136 	 * If possible, allocate from the per-CPU cache.  There are two
2137 	 * requirements for safe access to the per-CPU cache: (1) the thread
2138 	 * accessing the cache must not be preempted or yield during access,
2139 	 * and (2) the thread must not migrate CPUs without switching which
2140 	 * cache it accesses.  We rely on a critical section to prevent
2141 	 * preemption and migration.  We release the critical section in
2142 	 * order to acquire the zone mutex if we are unable to allocate from
2143 	 * the current cache; when we re-acquire the critical section, we
2144 	 * must detect and handle migration if it has occurred.
2145 	 */
2146 	critical_enter();
2147 	cpu = curcpu;
2148 	cache = &zone->uz_cpu[cpu];
2149 
2150 zalloc_start:
2151 	bucket = cache->uc_allocbucket;
2152 	if (bucket != NULL && bucket->ub_cnt > 0) {
2153 		bucket->ub_cnt--;
2154 		item = bucket->ub_bucket[bucket->ub_cnt];
2155 #ifdef INVARIANTS
2156 		bucket->ub_bucket[bucket->ub_cnt] = NULL;
2157 #endif
2158 		KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled."));
2159 		cache->uc_allocs++;
2160 		critical_exit();
2161 		if (zone->uz_ctor != NULL &&
2162 		    zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2163 			atomic_add_long(&zone->uz_fails, 1);
2164 			zone_free_item(zone, item, udata, SKIP_DTOR);
2165 			return (NULL);
2166 		}
2167 #ifdef INVARIANTS
2168 		uma_dbg_alloc(zone, NULL, item);
2169 #endif
2170 		if (flags & M_ZERO)
2171 			uma_zero_item(item, zone);
2172 		return (item);
2173 	}
2174 
2175 	/*
2176 	 * We have run out of items in our alloc bucket.
2177 	 * See if we can switch with our free bucket.
2178 	 */
2179 	bucket = cache->uc_freebucket;
2180 	if (bucket != NULL && bucket->ub_cnt > 0) {
2181 #ifdef UMA_DEBUG_ALLOC
2182 		printf("uma_zalloc: Swapping empty with alloc.\n");
2183 #endif
2184 		cache->uc_freebucket = cache->uc_allocbucket;
2185 		cache->uc_allocbucket = bucket;
2186 		goto zalloc_start;
2187 	}
2188 
2189 	/*
2190 	 * Discard any empty allocation bucket while we hold no locks.
2191 	 */
2192 	bucket = cache->uc_allocbucket;
2193 	cache->uc_allocbucket = NULL;
2194 	critical_exit();
2195 	if (bucket != NULL)
2196 		bucket_free(zone, bucket, udata);
2197 
2198 	/* Short-circuit for zones without buckets and low memory. */
2199 	if (zone->uz_count == 0 || bucketdisable)
2200 		goto zalloc_item;
2201 
2202 	/*
2203 	 * Attempt to retrieve the item from the per-CPU cache has failed, so
2204 	 * we must go back to the zone.  This requires the zone lock, so we
2205 	 * must drop the critical section, then re-acquire it when we go back
2206 	 * to the cache.  Since the critical section is released, we may be
2207 	 * preempted or migrate.  As such, make sure not to maintain any
2208 	 * thread-local state specific to the cache from prior to releasing
2209 	 * the critical section.
2210 	 */
2211 	lockfail = 0;
2212 	if (ZONE_TRYLOCK(zone) == 0) {
2213 		/* Record contention to size the buckets. */
2214 		ZONE_LOCK(zone);
2215 		lockfail = 1;
2216 	}
2217 	critical_enter();
2218 	cpu = curcpu;
2219 	cache = &zone->uz_cpu[cpu];
2220 
2221 	/*
2222 	 * Since we have locked the zone we may as well send back our stats.
2223 	 */
2224 	atomic_add_long(&zone->uz_allocs, cache->uc_allocs);
2225 	atomic_add_long(&zone->uz_frees, cache->uc_frees);
2226 	cache->uc_allocs = 0;
2227 	cache->uc_frees = 0;
2228 
2229 	/* See if we lost the race to fill the cache. */
2230 	if (cache->uc_allocbucket != NULL) {
2231 		ZONE_UNLOCK(zone);
2232 		goto zalloc_start;
2233 	}
2234 
2235 	/*
2236 	 * Check the zone's cache of buckets.
2237 	 */
2238 	if ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) {
2239 		KASSERT(bucket->ub_cnt != 0,
2240 		    ("uma_zalloc_arg: Returning an empty bucket."));
2241 
2242 		LIST_REMOVE(bucket, ub_link);
2243 		cache->uc_allocbucket = bucket;
2244 		ZONE_UNLOCK(zone);
2245 		goto zalloc_start;
2246 	}
2247 	/* We are no longer associated with this CPU. */
2248 	critical_exit();
2249 
2250 	/*
2251 	 * We bump the uz count when the cache size is insufficient to
2252 	 * handle the working set.
2253 	 */
2254 	if (lockfail && zone->uz_count < BUCKET_MAX)
2255 		zone->uz_count++;
2256 	ZONE_UNLOCK(zone);
2257 
2258 	/*
2259 	 * Now lets just fill a bucket and put it on the free list.  If that
2260 	 * works we'll restart the allocation from the beginning and it
2261 	 * will use the just filled bucket.
2262 	 */
2263 	bucket = zone_alloc_bucket(zone, udata, flags);
2264 	if (bucket != NULL) {
2265 		ZONE_LOCK(zone);
2266 		critical_enter();
2267 		cpu = curcpu;
2268 		cache = &zone->uz_cpu[cpu];
2269 		/*
2270 		 * See if we lost the race or were migrated.  Cache the
2271 		 * initialized bucket to make this less likely or claim
2272 		 * the memory directly.
2273 		 */
2274 		if (cache->uc_allocbucket == NULL)
2275 			cache->uc_allocbucket = bucket;
2276 		else
2277 			LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link);
2278 		ZONE_UNLOCK(zone);
2279 		goto zalloc_start;
2280 	}
2281 
2282 	/*
2283 	 * We may not be able to get a bucket so return an actual item.
2284 	 */
2285 #ifdef UMA_DEBUG
2286 	printf("uma_zalloc_arg: Bucketzone returned NULL\n");
2287 #endif
2288 
2289 zalloc_item:
2290 	item = zone_alloc_item(zone, udata, flags);
2291 
2292 	return (item);
2293 }
2294 
2295 static uma_slab_t
2296 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int flags)
2297 {
2298 	uma_slab_t slab;
2299 	int reserve;
2300 
2301 	mtx_assert(&keg->uk_lock, MA_OWNED);
2302 	slab = NULL;
2303 	reserve = 0;
2304 	if ((flags & M_USE_RESERVE) == 0)
2305 		reserve = keg->uk_reserve;
2306 
2307 	for (;;) {
2308 		/*
2309 		 * Find a slab with some space.  Prefer slabs that are partially
2310 		 * used over those that are totally full.  This helps to reduce
2311 		 * fragmentation.
2312 		 */
2313 		if (keg->uk_free > reserve) {
2314 			if (!LIST_EMPTY(&keg->uk_part_slab)) {
2315 				slab = LIST_FIRST(&keg->uk_part_slab);
2316 			} else {
2317 				slab = LIST_FIRST(&keg->uk_free_slab);
2318 				LIST_REMOVE(slab, us_link);
2319 				LIST_INSERT_HEAD(&keg->uk_part_slab, slab,
2320 				    us_link);
2321 			}
2322 			MPASS(slab->us_keg == keg);
2323 			return (slab);
2324 		}
2325 
2326 		/*
2327 		 * M_NOVM means don't ask at all!
2328 		 */
2329 		if (flags & M_NOVM)
2330 			break;
2331 
2332 		if (keg->uk_maxpages && keg->uk_pages >= keg->uk_maxpages) {
2333 			keg->uk_flags |= UMA_ZFLAG_FULL;
2334 			/*
2335 			 * If this is not a multi-zone, set the FULL bit.
2336 			 * Otherwise slab_multi() takes care of it.
2337 			 */
2338 			if ((zone->uz_flags & UMA_ZFLAG_MULTI) == 0) {
2339 				zone->uz_flags |= UMA_ZFLAG_FULL;
2340 				zone_log_warning(zone);
2341 				zone_maxaction(zone);
2342 			}
2343 			if (flags & M_NOWAIT)
2344 				break;
2345 			zone->uz_sleeps++;
2346 			msleep(keg, &keg->uk_lock, PVM, "keglimit", 0);
2347 			continue;
2348 		}
2349 		slab = keg_alloc_slab(keg, zone, flags);
2350 		/*
2351 		 * If we got a slab here it's safe to mark it partially used
2352 		 * and return.  We assume that the caller is going to remove
2353 		 * at least one item.
2354 		 */
2355 		if (slab) {
2356 			MPASS(slab->us_keg == keg);
2357 			LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2358 			return (slab);
2359 		}
2360 		/*
2361 		 * We might not have been able to get a slab but another cpu
2362 		 * could have while we were unlocked.  Check again before we
2363 		 * fail.
2364 		 */
2365 		flags |= M_NOVM;
2366 	}
2367 	return (slab);
2368 }
2369 
2370 static uma_slab_t
2371 zone_fetch_slab(uma_zone_t zone, uma_keg_t keg, int flags)
2372 {
2373 	uma_slab_t slab;
2374 
2375 	if (keg == NULL) {
2376 		keg = zone_first_keg(zone);
2377 		KEG_LOCK(keg);
2378 	}
2379 
2380 	for (;;) {
2381 		slab = keg_fetch_slab(keg, zone, flags);
2382 		if (slab)
2383 			return (slab);
2384 		if (flags & (M_NOWAIT | M_NOVM))
2385 			break;
2386 	}
2387 	KEG_UNLOCK(keg);
2388 	return (NULL);
2389 }
2390 
2391 /*
2392  * uma_zone_fetch_slab_multi:  Fetches a slab from one available keg.  Returns
2393  * with the keg locked.  On NULL no lock is held.
2394  *
2395  * The last pointer is used to seed the search.  It is not required.
2396  */
2397 static uma_slab_t
2398 zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int rflags)
2399 {
2400 	uma_klink_t klink;
2401 	uma_slab_t slab;
2402 	uma_keg_t keg;
2403 	int flags;
2404 	int empty;
2405 	int full;
2406 
2407 	/*
2408 	 * Don't wait on the first pass.  This will skip limit tests
2409 	 * as well.  We don't want to block if we can find a provider
2410 	 * without blocking.
2411 	 */
2412 	flags = (rflags & ~M_WAITOK) | M_NOWAIT;
2413 	/*
2414 	 * Use the last slab allocated as a hint for where to start
2415 	 * the search.
2416 	 */
2417 	if (last != NULL) {
2418 		slab = keg_fetch_slab(last, zone, flags);
2419 		if (slab)
2420 			return (slab);
2421 		KEG_UNLOCK(last);
2422 	}
2423 	/*
2424 	 * Loop until we have a slab incase of transient failures
2425 	 * while M_WAITOK is specified.  I'm not sure this is 100%
2426 	 * required but we've done it for so long now.
2427 	 */
2428 	for (;;) {
2429 		empty = 0;
2430 		full = 0;
2431 		/*
2432 		 * Search the available kegs for slabs.  Be careful to hold the
2433 		 * correct lock while calling into the keg layer.
2434 		 */
2435 		LIST_FOREACH(klink, &zone->uz_kegs, kl_link) {
2436 			keg = klink->kl_keg;
2437 			KEG_LOCK(keg);
2438 			if ((keg->uk_flags & UMA_ZFLAG_FULL) == 0) {
2439 				slab = keg_fetch_slab(keg, zone, flags);
2440 				if (slab)
2441 					return (slab);
2442 			}
2443 			if (keg->uk_flags & UMA_ZFLAG_FULL)
2444 				full++;
2445 			else
2446 				empty++;
2447 			KEG_UNLOCK(keg);
2448 		}
2449 		if (rflags & (M_NOWAIT | M_NOVM))
2450 			break;
2451 		flags = rflags;
2452 		/*
2453 		 * All kegs are full.  XXX We can't atomically check all kegs
2454 		 * and sleep so just sleep for a short period and retry.
2455 		 */
2456 		if (full && !empty) {
2457 			ZONE_LOCK(zone);
2458 			zone->uz_flags |= UMA_ZFLAG_FULL;
2459 			zone->uz_sleeps++;
2460 			zone_log_warning(zone);
2461 			zone_maxaction(zone);
2462 			msleep(zone, zone->uz_lockptr, PVM,
2463 			    "zonelimit", hz/100);
2464 			zone->uz_flags &= ~UMA_ZFLAG_FULL;
2465 			ZONE_UNLOCK(zone);
2466 			continue;
2467 		}
2468 	}
2469 	return (NULL);
2470 }
2471 
2472 static void *
2473 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
2474 {
2475 	void *item;
2476 	uint8_t freei;
2477 
2478 	MPASS(keg == slab->us_keg);
2479 	mtx_assert(&keg->uk_lock, MA_OWNED);
2480 
2481 	freei = BIT_FFS(SLAB_SETSIZE, &slab->us_free) - 1;
2482 	BIT_CLR(SLAB_SETSIZE, freei, &slab->us_free);
2483 	item = slab->us_data + (keg->uk_rsize * freei);
2484 	slab->us_freecount--;
2485 	keg->uk_free--;
2486 
2487 	/* Move this slab to the full list */
2488 	if (slab->us_freecount == 0) {
2489 		LIST_REMOVE(slab, us_link);
2490 		LIST_INSERT_HEAD(&keg->uk_full_slab, slab, us_link);
2491 	}
2492 
2493 	return (item);
2494 }
2495 
2496 static int
2497 zone_import(uma_zone_t zone, void **bucket, int max, int flags)
2498 {
2499 	uma_slab_t slab;
2500 	uma_keg_t keg;
2501 	int i;
2502 
2503 	slab = NULL;
2504 	keg = NULL;
2505 	/* Try to keep the buckets totally full */
2506 	for (i = 0; i < max; ) {
2507 		if ((slab = zone->uz_slab(zone, keg, flags)) == NULL)
2508 			break;
2509 		keg = slab->us_keg;
2510 		while (slab->us_freecount && i < max) {
2511 			bucket[i++] = slab_alloc_item(keg, slab);
2512 			if (keg->uk_free <= keg->uk_reserve)
2513 				break;
2514 		}
2515 		/* Don't grab more than one slab at a time. */
2516 		flags &= ~M_WAITOK;
2517 		flags |= M_NOWAIT;
2518 	}
2519 	if (slab != NULL)
2520 		KEG_UNLOCK(keg);
2521 
2522 	return i;
2523 }
2524 
2525 static uma_bucket_t
2526 zone_alloc_bucket(uma_zone_t zone, void *udata, int flags)
2527 {
2528 	uma_bucket_t bucket;
2529 	int max;
2530 
2531 	/* Don't wait for buckets, preserve caller's NOVM setting. */
2532 	bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
2533 	if (bucket == NULL)
2534 		return (NULL);
2535 
2536 	max = MIN(bucket->ub_entries, zone->uz_count);
2537 	bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
2538 	    max, flags);
2539 
2540 	/*
2541 	 * Initialize the memory if necessary.
2542 	 */
2543 	if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
2544 		int i;
2545 
2546 		for (i = 0; i < bucket->ub_cnt; i++)
2547 			if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size,
2548 			    flags) != 0)
2549 				break;
2550 		/*
2551 		 * If we couldn't initialize the whole bucket, put the
2552 		 * rest back onto the freelist.
2553 		 */
2554 		if (i != bucket->ub_cnt) {
2555 			zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
2556 			    bucket->ub_cnt - i);
2557 #ifdef INVARIANTS
2558 			bzero(&bucket->ub_bucket[i],
2559 			    sizeof(void *) * (bucket->ub_cnt - i));
2560 #endif
2561 			bucket->ub_cnt = i;
2562 		}
2563 	}
2564 
2565 	if (bucket->ub_cnt == 0) {
2566 		bucket_free(zone, bucket, udata);
2567 		atomic_add_long(&zone->uz_fails, 1);
2568 		return (NULL);
2569 	}
2570 
2571 	return (bucket);
2572 }
2573 
2574 /*
2575  * Allocates a single item from a zone.
2576  *
2577  * Arguments
2578  *	zone   The zone to alloc for.
2579  *	udata  The data to be passed to the constructor.
2580  *	flags  M_WAITOK, M_NOWAIT, M_ZERO.
2581  *
2582  * Returns
2583  *	NULL if there is no memory and M_NOWAIT is set
2584  *	An item if successful
2585  */
2586 
2587 static void *
2588 zone_alloc_item(uma_zone_t zone, void *udata, int flags)
2589 {
2590 	void *item;
2591 
2592 	item = NULL;
2593 
2594 #ifdef UMA_DEBUG_ALLOC
2595 	printf("INTERNAL: Allocating one item from %s(%p)\n", zone->uz_name, zone);
2596 #endif
2597 	if (zone->uz_import(zone->uz_arg, &item, 1, flags) != 1)
2598 		goto fail;
2599 	atomic_add_long(&zone->uz_allocs, 1);
2600 
2601 	/*
2602 	 * We have to call both the zone's init (not the keg's init)
2603 	 * and the zone's ctor.  This is because the item is going from
2604 	 * a keg slab directly to the user, and the user is expecting it
2605 	 * to be both zone-init'd as well as zone-ctor'd.
2606 	 */
2607 	if (zone->uz_init != NULL) {
2608 		if (zone->uz_init(item, zone->uz_size, flags) != 0) {
2609 			zone_free_item(zone, item, udata, SKIP_FINI);
2610 			goto fail;
2611 		}
2612 	}
2613 	if (zone->uz_ctor != NULL) {
2614 		if (zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2615 			zone_free_item(zone, item, udata, SKIP_DTOR);
2616 			goto fail;
2617 		}
2618 	}
2619 #ifdef INVARIANTS
2620 	uma_dbg_alloc(zone, NULL, item);
2621 #endif
2622 	if (flags & M_ZERO)
2623 		uma_zero_item(item, zone);
2624 
2625 	return (item);
2626 
2627 fail:
2628 	atomic_add_long(&zone->uz_fails, 1);
2629 	return (NULL);
2630 }
2631 
2632 /* See uma.h */
2633 void
2634 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
2635 {
2636 	uma_cache_t cache;
2637 	uma_bucket_t bucket;
2638 	int lockfail;
2639 	int cpu;
2640 
2641 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
2642 	random_harvest_fast_uma(&zone, sizeof(zone), 1, RANDOM_UMA);
2643 
2644 #ifdef UMA_DEBUG_ALLOC_1
2645 	printf("Freeing item %p to %s(%p)\n", item, zone->uz_name, zone);
2646 #endif
2647 	CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread,
2648 	    zone->uz_name);
2649 
2650 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
2651 	    ("uma_zfree_arg: called with spinlock or critical section held"));
2652 
2653         /* uma_zfree(..., NULL) does nothing, to match free(9). */
2654         if (item == NULL)
2655                 return;
2656 #ifdef DEBUG_MEMGUARD
2657 	if (is_memguard_addr(item)) {
2658 		if (zone->uz_dtor != NULL && zone->uz_dtor != mtrash_dtor)
2659 			zone->uz_dtor(item, zone->uz_size, udata);
2660 		if (zone->uz_fini != NULL && zone->uz_fini != mtrash_fini)
2661 			zone->uz_fini(item, zone->uz_size);
2662 		memguard_free(item);
2663 		return;
2664 	}
2665 #endif
2666 #ifdef INVARIANTS
2667 	if (zone->uz_flags & UMA_ZONE_MALLOC)
2668 		uma_dbg_free(zone, udata, item);
2669 	else
2670 		uma_dbg_free(zone, NULL, item);
2671 #endif
2672 	if (zone->uz_dtor != NULL)
2673 		zone->uz_dtor(item, zone->uz_size, udata);
2674 
2675 	/*
2676 	 * The race here is acceptable.  If we miss it we'll just have to wait
2677 	 * a little longer for the limits to be reset.
2678 	 */
2679 	if (zone->uz_flags & UMA_ZFLAG_FULL)
2680 		goto zfree_item;
2681 
2682 	/*
2683 	 * If possible, free to the per-CPU cache.  There are two
2684 	 * requirements for safe access to the per-CPU cache: (1) the thread
2685 	 * accessing the cache must not be preempted or yield during access,
2686 	 * and (2) the thread must not migrate CPUs without switching which
2687 	 * cache it accesses.  We rely on a critical section to prevent
2688 	 * preemption and migration.  We release the critical section in
2689 	 * order to acquire the zone mutex if we are unable to free to the
2690 	 * current cache; when we re-acquire the critical section, we must
2691 	 * detect and handle migration if it has occurred.
2692 	 */
2693 zfree_restart:
2694 	critical_enter();
2695 	cpu = curcpu;
2696 	cache = &zone->uz_cpu[cpu];
2697 
2698 zfree_start:
2699 	/*
2700 	 * Try to free into the allocbucket first to give LIFO ordering
2701 	 * for cache-hot datastructures.  Spill over into the freebucket
2702 	 * if necessary.  Alloc will swap them if one runs dry.
2703 	 */
2704 	bucket = cache->uc_allocbucket;
2705 	if (bucket == NULL || bucket->ub_cnt >= bucket->ub_entries)
2706 		bucket = cache->uc_freebucket;
2707 	if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) {
2708 		KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL,
2709 		    ("uma_zfree: Freeing to non free bucket index."));
2710 		bucket->ub_bucket[bucket->ub_cnt] = item;
2711 		bucket->ub_cnt++;
2712 		cache->uc_frees++;
2713 		critical_exit();
2714 		return;
2715 	}
2716 
2717 	/*
2718 	 * We must go back the zone, which requires acquiring the zone lock,
2719 	 * which in turn means we must release and re-acquire the critical
2720 	 * section.  Since the critical section is released, we may be
2721 	 * preempted or migrate.  As such, make sure not to maintain any
2722 	 * thread-local state specific to the cache from prior to releasing
2723 	 * the critical section.
2724 	 */
2725 	critical_exit();
2726 	if (zone->uz_count == 0 || bucketdisable)
2727 		goto zfree_item;
2728 
2729 	lockfail = 0;
2730 	if (ZONE_TRYLOCK(zone) == 0) {
2731 		/* Record contention to size the buckets. */
2732 		ZONE_LOCK(zone);
2733 		lockfail = 1;
2734 	}
2735 	critical_enter();
2736 	cpu = curcpu;
2737 	cache = &zone->uz_cpu[cpu];
2738 
2739 	/*
2740 	 * Since we have locked the zone we may as well send back our stats.
2741 	 */
2742 	atomic_add_long(&zone->uz_allocs, cache->uc_allocs);
2743 	atomic_add_long(&zone->uz_frees, cache->uc_frees);
2744 	cache->uc_allocs = 0;
2745 	cache->uc_frees = 0;
2746 
2747 	bucket = cache->uc_freebucket;
2748 	if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) {
2749 		ZONE_UNLOCK(zone);
2750 		goto zfree_start;
2751 	}
2752 	cache->uc_freebucket = NULL;
2753 
2754 	/* Can we throw this on the zone full list? */
2755 	if (bucket != NULL) {
2756 #ifdef UMA_DEBUG_ALLOC
2757 		printf("uma_zfree: Putting old bucket on the free list.\n");
2758 #endif
2759 		/* ub_cnt is pointing to the last free item */
2760 		KASSERT(bucket->ub_cnt != 0,
2761 		    ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n"));
2762 		LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link);
2763 	}
2764 
2765 	/* We are no longer associated with this CPU. */
2766 	critical_exit();
2767 
2768 	/*
2769 	 * We bump the uz count when the cache size is insufficient to
2770 	 * handle the working set.
2771 	 */
2772 	if (lockfail && zone->uz_count < BUCKET_MAX)
2773 		zone->uz_count++;
2774 	ZONE_UNLOCK(zone);
2775 
2776 #ifdef UMA_DEBUG_ALLOC
2777 	printf("uma_zfree: Allocating new free bucket.\n");
2778 #endif
2779 	bucket = bucket_alloc(zone, udata, M_NOWAIT);
2780 	if (bucket) {
2781 		critical_enter();
2782 		cpu = curcpu;
2783 		cache = &zone->uz_cpu[cpu];
2784 		if (cache->uc_freebucket == NULL) {
2785 			cache->uc_freebucket = bucket;
2786 			goto zfree_start;
2787 		}
2788 		/*
2789 		 * We lost the race, start over.  We have to drop our
2790 		 * critical section to free the bucket.
2791 		 */
2792 		critical_exit();
2793 		bucket_free(zone, bucket, udata);
2794 		goto zfree_restart;
2795 	}
2796 
2797 	/*
2798 	 * If nothing else caught this, we'll just do an internal free.
2799 	 */
2800 zfree_item:
2801 	zone_free_item(zone, item, udata, SKIP_DTOR);
2802 
2803 	return;
2804 }
2805 
2806 static void
2807 slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item)
2808 {
2809 	uint8_t freei;
2810 
2811 	mtx_assert(&keg->uk_lock, MA_OWNED);
2812 	MPASS(keg == slab->us_keg);
2813 
2814 	/* Do we need to remove from any lists? */
2815 	if (slab->us_freecount+1 == keg->uk_ipers) {
2816 		LIST_REMOVE(slab, us_link);
2817 		LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
2818 	} else if (slab->us_freecount == 0) {
2819 		LIST_REMOVE(slab, us_link);
2820 		LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2821 	}
2822 
2823 	/* Slab management. */
2824 	freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
2825 	BIT_SET(SLAB_SETSIZE, freei, &slab->us_free);
2826 	slab->us_freecount++;
2827 
2828 	/* Keg statistics. */
2829 	keg->uk_free++;
2830 }
2831 
2832 static void
2833 zone_release(uma_zone_t zone, void **bucket, int cnt)
2834 {
2835 	void *item;
2836 	uma_slab_t slab;
2837 	uma_keg_t keg;
2838 	uint8_t *mem;
2839 	int clearfull;
2840 	int i;
2841 
2842 	clearfull = 0;
2843 	keg = zone_first_keg(zone);
2844 	KEG_LOCK(keg);
2845 	for (i = 0; i < cnt; i++) {
2846 		item = bucket[i];
2847 		if (!(zone->uz_flags & UMA_ZONE_VTOSLAB)) {
2848 			mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
2849 			if (zone->uz_flags & UMA_ZONE_HASH) {
2850 				slab = hash_sfind(&keg->uk_hash, mem);
2851 			} else {
2852 				mem += keg->uk_pgoff;
2853 				slab = (uma_slab_t)mem;
2854 			}
2855 		} else {
2856 			slab = vtoslab((vm_offset_t)item);
2857 			if (slab->us_keg != keg) {
2858 				KEG_UNLOCK(keg);
2859 				keg = slab->us_keg;
2860 				KEG_LOCK(keg);
2861 			}
2862 		}
2863 		slab_free_item(keg, slab, item);
2864 		if (keg->uk_flags & UMA_ZFLAG_FULL) {
2865 			if (keg->uk_pages < keg->uk_maxpages) {
2866 				keg->uk_flags &= ~UMA_ZFLAG_FULL;
2867 				clearfull = 1;
2868 			}
2869 
2870 			/*
2871 			 * We can handle one more allocation. Since we're
2872 			 * clearing ZFLAG_FULL, wake up all procs blocked
2873 			 * on pages. This should be uncommon, so keeping this
2874 			 * simple for now (rather than adding count of blocked
2875 			 * threads etc).
2876 			 */
2877 			wakeup(keg);
2878 		}
2879 	}
2880 	KEG_UNLOCK(keg);
2881 	if (clearfull) {
2882 		ZONE_LOCK(zone);
2883 		zone->uz_flags &= ~UMA_ZFLAG_FULL;
2884 		wakeup(zone);
2885 		ZONE_UNLOCK(zone);
2886 	}
2887 
2888 }
2889 
2890 /*
2891  * Frees a single item to any zone.
2892  *
2893  * Arguments:
2894  *	zone   The zone to free to
2895  *	item   The item we're freeing
2896  *	udata  User supplied data for the dtor
2897  *	skip   Skip dtors and finis
2898  */
2899 static void
2900 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
2901 {
2902 
2903 #ifdef INVARIANTS
2904 	if (skip == SKIP_NONE) {
2905 		if (zone->uz_flags & UMA_ZONE_MALLOC)
2906 			uma_dbg_free(zone, udata, item);
2907 		else
2908 			uma_dbg_free(zone, NULL, item);
2909 	}
2910 #endif
2911 	if (skip < SKIP_DTOR && zone->uz_dtor)
2912 		zone->uz_dtor(item, zone->uz_size, udata);
2913 
2914 	if (skip < SKIP_FINI && zone->uz_fini)
2915 		zone->uz_fini(item, zone->uz_size);
2916 
2917 	atomic_add_long(&zone->uz_frees, 1);
2918 	zone->uz_release(zone->uz_arg, &item, 1);
2919 }
2920 
2921 /* See uma.h */
2922 int
2923 uma_zone_set_max(uma_zone_t zone, int nitems)
2924 {
2925 	uma_keg_t keg;
2926 
2927 	keg = zone_first_keg(zone);
2928 	if (keg == NULL)
2929 		return (0);
2930 	KEG_LOCK(keg);
2931 	keg->uk_maxpages = (nitems / keg->uk_ipers) * keg->uk_ppera;
2932 	if (keg->uk_maxpages * keg->uk_ipers < nitems)
2933 		keg->uk_maxpages += keg->uk_ppera;
2934 	nitems = keg->uk_maxpages * keg->uk_ipers;
2935 	KEG_UNLOCK(keg);
2936 
2937 	return (nitems);
2938 }
2939 
2940 /* See uma.h */
2941 int
2942 uma_zone_get_max(uma_zone_t zone)
2943 {
2944 	int nitems;
2945 	uma_keg_t keg;
2946 
2947 	keg = zone_first_keg(zone);
2948 	if (keg == NULL)
2949 		return (0);
2950 	KEG_LOCK(keg);
2951 	nitems = keg->uk_maxpages * keg->uk_ipers;
2952 	KEG_UNLOCK(keg);
2953 
2954 	return (nitems);
2955 }
2956 
2957 /* See uma.h */
2958 void
2959 uma_zone_set_warning(uma_zone_t zone, const char *warning)
2960 {
2961 
2962 	ZONE_LOCK(zone);
2963 	zone->uz_warning = warning;
2964 	ZONE_UNLOCK(zone);
2965 }
2966 
2967 /* See uma.h */
2968 void
2969 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction)
2970 {
2971 
2972 	ZONE_LOCK(zone);
2973 	TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone);
2974 	ZONE_UNLOCK(zone);
2975 }
2976 
2977 /* See uma.h */
2978 int
2979 uma_zone_get_cur(uma_zone_t zone)
2980 {
2981 	int64_t nitems;
2982 	u_int i;
2983 
2984 	ZONE_LOCK(zone);
2985 	nitems = zone->uz_allocs - zone->uz_frees;
2986 	CPU_FOREACH(i) {
2987 		/*
2988 		 * See the comment in sysctl_vm_zone_stats() regarding the
2989 		 * safety of accessing the per-cpu caches. With the zone lock
2990 		 * held, it is safe, but can potentially result in stale data.
2991 		 */
2992 		nitems += zone->uz_cpu[i].uc_allocs -
2993 		    zone->uz_cpu[i].uc_frees;
2994 	}
2995 	ZONE_UNLOCK(zone);
2996 
2997 	return (nitems < 0 ? 0 : nitems);
2998 }
2999 
3000 /* See uma.h */
3001 void
3002 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
3003 {
3004 	uma_keg_t keg;
3005 
3006 	keg = zone_first_keg(zone);
3007 	KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type"));
3008 	KEG_LOCK(keg);
3009 	KASSERT(keg->uk_pages == 0,
3010 	    ("uma_zone_set_init on non-empty keg"));
3011 	keg->uk_init = uminit;
3012 	KEG_UNLOCK(keg);
3013 }
3014 
3015 /* See uma.h */
3016 void
3017 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
3018 {
3019 	uma_keg_t keg;
3020 
3021 	keg = zone_first_keg(zone);
3022 	KASSERT(keg != NULL, ("uma_zone_set_fini: Invalid zone type"));
3023 	KEG_LOCK(keg);
3024 	KASSERT(keg->uk_pages == 0,
3025 	    ("uma_zone_set_fini on non-empty keg"));
3026 	keg->uk_fini = fini;
3027 	KEG_UNLOCK(keg);
3028 }
3029 
3030 /* See uma.h */
3031 void
3032 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
3033 {
3034 
3035 	ZONE_LOCK(zone);
3036 	KASSERT(zone_first_keg(zone)->uk_pages == 0,
3037 	    ("uma_zone_set_zinit on non-empty keg"));
3038 	zone->uz_init = zinit;
3039 	ZONE_UNLOCK(zone);
3040 }
3041 
3042 /* See uma.h */
3043 void
3044 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
3045 {
3046 
3047 	ZONE_LOCK(zone);
3048 	KASSERT(zone_first_keg(zone)->uk_pages == 0,
3049 	    ("uma_zone_set_zfini on non-empty keg"));
3050 	zone->uz_fini = zfini;
3051 	ZONE_UNLOCK(zone);
3052 }
3053 
3054 /* See uma.h */
3055 /* XXX uk_freef is not actually used with the zone locked */
3056 void
3057 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
3058 {
3059 	uma_keg_t keg;
3060 
3061 	keg = zone_first_keg(zone);
3062 	KASSERT(keg != NULL, ("uma_zone_set_freef: Invalid zone type"));
3063 	KEG_LOCK(keg);
3064 	keg->uk_freef = freef;
3065 	KEG_UNLOCK(keg);
3066 }
3067 
3068 /* See uma.h */
3069 /* XXX uk_allocf is not actually used with the zone locked */
3070 void
3071 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
3072 {
3073 	uma_keg_t keg;
3074 
3075 	keg = zone_first_keg(zone);
3076 	KEG_LOCK(keg);
3077 	keg->uk_allocf = allocf;
3078 	KEG_UNLOCK(keg);
3079 }
3080 
3081 /* See uma.h */
3082 void
3083 uma_zone_reserve(uma_zone_t zone, int items)
3084 {
3085 	uma_keg_t keg;
3086 
3087 	keg = zone_first_keg(zone);
3088 	if (keg == NULL)
3089 		return;
3090 	KEG_LOCK(keg);
3091 	keg->uk_reserve = items;
3092 	KEG_UNLOCK(keg);
3093 
3094 	return;
3095 }
3096 
3097 /* See uma.h */
3098 int
3099 uma_zone_reserve_kva(uma_zone_t zone, int count)
3100 {
3101 	uma_keg_t keg;
3102 	vm_offset_t kva;
3103 	u_int pages;
3104 
3105 	keg = zone_first_keg(zone);
3106 	if (keg == NULL)
3107 		return (0);
3108 	pages = count / keg->uk_ipers;
3109 
3110 	if (pages * keg->uk_ipers < count)
3111 		pages++;
3112 
3113 #ifdef UMA_MD_SMALL_ALLOC
3114 	if (keg->uk_ppera > 1) {
3115 #else
3116 	if (1) {
3117 #endif
3118 		kva = kva_alloc((vm_size_t)pages * UMA_SLAB_SIZE);
3119 		if (kva == 0)
3120 			return (0);
3121 	} else
3122 		kva = 0;
3123 	KEG_LOCK(keg);
3124 	keg->uk_kva = kva;
3125 	keg->uk_offset = 0;
3126 	keg->uk_maxpages = pages;
3127 #ifdef UMA_MD_SMALL_ALLOC
3128 	keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
3129 #else
3130 	keg->uk_allocf = noobj_alloc;
3131 #endif
3132 	keg->uk_flags |= UMA_ZONE_NOFREE;
3133 	KEG_UNLOCK(keg);
3134 
3135 	return (1);
3136 }
3137 
3138 /* See uma.h */
3139 void
3140 uma_prealloc(uma_zone_t zone, int items)
3141 {
3142 	int slabs;
3143 	uma_slab_t slab;
3144 	uma_keg_t keg;
3145 
3146 	keg = zone_first_keg(zone);
3147 	if (keg == NULL)
3148 		return;
3149 	KEG_LOCK(keg);
3150 	slabs = items / keg->uk_ipers;
3151 	if (slabs * keg->uk_ipers < items)
3152 		slabs++;
3153 	while (slabs > 0) {
3154 		slab = keg_alloc_slab(keg, zone, M_WAITOK);
3155 		if (slab == NULL)
3156 			break;
3157 		MPASS(slab->us_keg == keg);
3158 		LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
3159 		slabs--;
3160 	}
3161 	KEG_UNLOCK(keg);
3162 }
3163 
3164 /* See uma.h */
3165 static void
3166 uma_reclaim_locked(bool kmem_danger)
3167 {
3168 
3169 #ifdef UMA_DEBUG
3170 	printf("UMA: vm asked us to release pages!\n");
3171 #endif
3172 	sx_assert(&uma_drain_lock, SA_XLOCKED);
3173 	bucket_enable();
3174 	zone_foreach(zone_drain);
3175 	if (vm_page_count_min() || kmem_danger) {
3176 		cache_drain_safe(NULL);
3177 		zone_foreach(zone_drain);
3178 	}
3179 	/*
3180 	 * Some slabs may have been freed but this zone will be visited early
3181 	 * we visit again so that we can free pages that are empty once other
3182 	 * zones are drained.  We have to do the same for buckets.
3183 	 */
3184 	zone_drain(slabzone);
3185 	bucket_zone_drain();
3186 }
3187 
3188 void
3189 uma_reclaim(void)
3190 {
3191 
3192 	sx_xlock(&uma_drain_lock);
3193 	uma_reclaim_locked(false);
3194 	sx_xunlock(&uma_drain_lock);
3195 }
3196 
3197 static int uma_reclaim_needed;
3198 
3199 void
3200 uma_reclaim_wakeup(void)
3201 {
3202 
3203 	uma_reclaim_needed = 1;
3204 	wakeup(&uma_reclaim_needed);
3205 }
3206 
3207 void
3208 uma_reclaim_worker(void *arg __unused)
3209 {
3210 
3211 	sx_xlock(&uma_drain_lock);
3212 	for (;;) {
3213 		sx_sleep(&uma_reclaim_needed, &uma_drain_lock, PVM,
3214 		    "umarcl", 0);
3215 		if (uma_reclaim_needed) {
3216 			uma_reclaim_needed = 0;
3217 			uma_reclaim_locked(true);
3218 		}
3219 	}
3220 }
3221 
3222 /* See uma.h */
3223 int
3224 uma_zone_exhausted(uma_zone_t zone)
3225 {
3226 	int full;
3227 
3228 	ZONE_LOCK(zone);
3229 	full = (zone->uz_flags & UMA_ZFLAG_FULL);
3230 	ZONE_UNLOCK(zone);
3231 	return (full);
3232 }
3233 
3234 int
3235 uma_zone_exhausted_nolock(uma_zone_t zone)
3236 {
3237 	return (zone->uz_flags & UMA_ZFLAG_FULL);
3238 }
3239 
3240 void *
3241 uma_large_malloc(vm_size_t size, int wait)
3242 {
3243 	void *mem;
3244 	uma_slab_t slab;
3245 	uint8_t flags;
3246 
3247 	slab = zone_alloc_item(slabzone, NULL, wait);
3248 	if (slab == NULL)
3249 		return (NULL);
3250 	mem = page_alloc(NULL, size, &flags, wait);
3251 	if (mem) {
3252 		vsetslab((vm_offset_t)mem, slab);
3253 		slab->us_data = mem;
3254 		slab->us_flags = flags | UMA_SLAB_MALLOC;
3255 		slab->us_size = size;
3256 	} else {
3257 		zone_free_item(slabzone, slab, NULL, SKIP_NONE);
3258 	}
3259 
3260 	return (mem);
3261 }
3262 
3263 void
3264 uma_large_free(uma_slab_t slab)
3265 {
3266 
3267 	page_free(slab->us_data, slab->us_size, slab->us_flags);
3268 	zone_free_item(slabzone, slab, NULL, SKIP_NONE);
3269 }
3270 
3271 static void
3272 uma_zero_item(void *item, uma_zone_t zone)
3273 {
3274 
3275 	if (zone->uz_flags & UMA_ZONE_PCPU) {
3276 		for (int i = 0; i < mp_ncpus; i++)
3277 			bzero(zpcpu_get_cpu(item, i), zone->uz_size);
3278 	} else
3279 		bzero(item, zone->uz_size);
3280 }
3281 
3282 void
3283 uma_print_stats(void)
3284 {
3285 	zone_foreach(uma_print_zone);
3286 }
3287 
3288 static void
3289 slab_print(uma_slab_t slab)
3290 {
3291 	printf("slab: keg %p, data %p, freecount %d\n",
3292 		slab->us_keg, slab->us_data, slab->us_freecount);
3293 }
3294 
3295 static void
3296 cache_print(uma_cache_t cache)
3297 {
3298 	printf("alloc: %p(%d), free: %p(%d)\n",
3299 		cache->uc_allocbucket,
3300 		cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0,
3301 		cache->uc_freebucket,
3302 		cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0);
3303 }
3304 
3305 static void
3306 uma_print_keg(uma_keg_t keg)
3307 {
3308 	uma_slab_t slab;
3309 
3310 	printf("keg: %s(%p) size %d(%d) flags %#x ipers %d ppera %d "
3311 	    "out %d free %d limit %d\n",
3312 	    keg->uk_name, keg, keg->uk_size, keg->uk_rsize, keg->uk_flags,
3313 	    keg->uk_ipers, keg->uk_ppera,
3314 	    (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free,
3315 	    (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers);
3316 	printf("Part slabs:\n");
3317 	LIST_FOREACH(slab, &keg->uk_part_slab, us_link)
3318 		slab_print(slab);
3319 	printf("Free slabs:\n");
3320 	LIST_FOREACH(slab, &keg->uk_free_slab, us_link)
3321 		slab_print(slab);
3322 	printf("Full slabs:\n");
3323 	LIST_FOREACH(slab, &keg->uk_full_slab, us_link)
3324 		slab_print(slab);
3325 }
3326 
3327 void
3328 uma_print_zone(uma_zone_t zone)
3329 {
3330 	uma_cache_t cache;
3331 	uma_klink_t kl;
3332 	int i;
3333 
3334 	printf("zone: %s(%p) size %d flags %#x\n",
3335 	    zone->uz_name, zone, zone->uz_size, zone->uz_flags);
3336 	LIST_FOREACH(kl, &zone->uz_kegs, kl_link)
3337 		uma_print_keg(kl->kl_keg);
3338 	CPU_FOREACH(i) {
3339 		cache = &zone->uz_cpu[i];
3340 		printf("CPU %d Cache:\n", i);
3341 		cache_print(cache);
3342 	}
3343 }
3344 
3345 #ifdef DDB
3346 /*
3347  * Generate statistics across both the zone and its per-cpu cache's.  Return
3348  * desired statistics if the pointer is non-NULL for that statistic.
3349  *
3350  * Note: does not update the zone statistics, as it can't safely clear the
3351  * per-CPU cache statistic.
3352  *
3353  * XXXRW: Following the uc_allocbucket and uc_freebucket pointers here isn't
3354  * safe from off-CPU; we should modify the caches to track this information
3355  * directly so that we don't have to.
3356  */
3357 static void
3358 uma_zone_sumstat(uma_zone_t z, int *cachefreep, uint64_t *allocsp,
3359     uint64_t *freesp, uint64_t *sleepsp)
3360 {
3361 	uma_cache_t cache;
3362 	uint64_t allocs, frees, sleeps;
3363 	int cachefree, cpu;
3364 
3365 	allocs = frees = sleeps = 0;
3366 	cachefree = 0;
3367 	CPU_FOREACH(cpu) {
3368 		cache = &z->uz_cpu[cpu];
3369 		if (cache->uc_allocbucket != NULL)
3370 			cachefree += cache->uc_allocbucket->ub_cnt;
3371 		if (cache->uc_freebucket != NULL)
3372 			cachefree += cache->uc_freebucket->ub_cnt;
3373 		allocs += cache->uc_allocs;
3374 		frees += cache->uc_frees;
3375 	}
3376 	allocs += z->uz_allocs;
3377 	frees += z->uz_frees;
3378 	sleeps += z->uz_sleeps;
3379 	if (cachefreep != NULL)
3380 		*cachefreep = cachefree;
3381 	if (allocsp != NULL)
3382 		*allocsp = allocs;
3383 	if (freesp != NULL)
3384 		*freesp = frees;
3385 	if (sleepsp != NULL)
3386 		*sleepsp = sleeps;
3387 }
3388 #endif /* DDB */
3389 
3390 static int
3391 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
3392 {
3393 	uma_keg_t kz;
3394 	uma_zone_t z;
3395 	int count;
3396 
3397 	count = 0;
3398 	rw_rlock(&uma_rwlock);
3399 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3400 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
3401 			count++;
3402 	}
3403 	rw_runlock(&uma_rwlock);
3404 	return (sysctl_handle_int(oidp, &count, 0, req));
3405 }
3406 
3407 static int
3408 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
3409 {
3410 	struct uma_stream_header ush;
3411 	struct uma_type_header uth;
3412 	struct uma_percpu_stat ups;
3413 	uma_bucket_t bucket;
3414 	struct sbuf sbuf;
3415 	uma_cache_t cache;
3416 	uma_klink_t kl;
3417 	uma_keg_t kz;
3418 	uma_zone_t z;
3419 	uma_keg_t k;
3420 	int count, error, i;
3421 
3422 	error = sysctl_wire_old_buffer(req, 0);
3423 	if (error != 0)
3424 		return (error);
3425 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
3426 	sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL);
3427 
3428 	count = 0;
3429 	rw_rlock(&uma_rwlock);
3430 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3431 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
3432 			count++;
3433 	}
3434 
3435 	/*
3436 	 * Insert stream header.
3437 	 */
3438 	bzero(&ush, sizeof(ush));
3439 	ush.ush_version = UMA_STREAM_VERSION;
3440 	ush.ush_maxcpus = (mp_maxid + 1);
3441 	ush.ush_count = count;
3442 	(void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
3443 
3444 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3445 		LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3446 			bzero(&uth, sizeof(uth));
3447 			ZONE_LOCK(z);
3448 			strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
3449 			uth.uth_align = kz->uk_align;
3450 			uth.uth_size = kz->uk_size;
3451 			uth.uth_rsize = kz->uk_rsize;
3452 			LIST_FOREACH(kl, &z->uz_kegs, kl_link) {
3453 				k = kl->kl_keg;
3454 				uth.uth_maxpages += k->uk_maxpages;
3455 				uth.uth_pages += k->uk_pages;
3456 				uth.uth_keg_free += k->uk_free;
3457 				uth.uth_limit = (k->uk_maxpages / k->uk_ppera)
3458 				    * k->uk_ipers;
3459 			}
3460 
3461 			/*
3462 			 * A zone is secondary is it is not the first entry
3463 			 * on the keg's zone list.
3464 			 */
3465 			if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
3466 			    (LIST_FIRST(&kz->uk_zones) != z))
3467 				uth.uth_zone_flags = UTH_ZONE_SECONDARY;
3468 
3469 			LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3470 				uth.uth_zone_free += bucket->ub_cnt;
3471 			uth.uth_allocs = z->uz_allocs;
3472 			uth.uth_frees = z->uz_frees;
3473 			uth.uth_fails = z->uz_fails;
3474 			uth.uth_sleeps = z->uz_sleeps;
3475 			(void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
3476 			/*
3477 			 * While it is not normally safe to access the cache
3478 			 * bucket pointers while not on the CPU that owns the
3479 			 * cache, we only allow the pointers to be exchanged
3480 			 * without the zone lock held, not invalidated, so
3481 			 * accept the possible race associated with bucket
3482 			 * exchange during monitoring.
3483 			 */
3484 			for (i = 0; i < (mp_maxid + 1); i++) {
3485 				bzero(&ups, sizeof(ups));
3486 				if (kz->uk_flags & UMA_ZFLAG_INTERNAL)
3487 					goto skip;
3488 				if (CPU_ABSENT(i))
3489 					goto skip;
3490 				cache = &z->uz_cpu[i];
3491 				if (cache->uc_allocbucket != NULL)
3492 					ups.ups_cache_free +=
3493 					    cache->uc_allocbucket->ub_cnt;
3494 				if (cache->uc_freebucket != NULL)
3495 					ups.ups_cache_free +=
3496 					    cache->uc_freebucket->ub_cnt;
3497 				ups.ups_allocs = cache->uc_allocs;
3498 				ups.ups_frees = cache->uc_frees;
3499 skip:
3500 				(void)sbuf_bcat(&sbuf, &ups, sizeof(ups));
3501 			}
3502 			ZONE_UNLOCK(z);
3503 		}
3504 	}
3505 	rw_runlock(&uma_rwlock);
3506 	error = sbuf_finish(&sbuf);
3507 	sbuf_delete(&sbuf);
3508 	return (error);
3509 }
3510 
3511 int
3512 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS)
3513 {
3514 	uma_zone_t zone = *(uma_zone_t *)arg1;
3515 	int error, max;
3516 
3517 	max = uma_zone_get_max(zone);
3518 	error = sysctl_handle_int(oidp, &max, 0, req);
3519 	if (error || !req->newptr)
3520 		return (error);
3521 
3522 	uma_zone_set_max(zone, max);
3523 
3524 	return (0);
3525 }
3526 
3527 int
3528 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS)
3529 {
3530 	uma_zone_t zone = *(uma_zone_t *)arg1;
3531 	int cur;
3532 
3533 	cur = uma_zone_get_cur(zone);
3534 	return (sysctl_handle_int(oidp, &cur, 0, req));
3535 }
3536 
3537 #ifdef INVARIANTS
3538 static uma_slab_t
3539 uma_dbg_getslab(uma_zone_t zone, void *item)
3540 {
3541 	uma_slab_t slab;
3542 	uma_keg_t keg;
3543 	uint8_t *mem;
3544 
3545 	mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
3546 	if (zone->uz_flags & UMA_ZONE_VTOSLAB) {
3547 		slab = vtoslab((vm_offset_t)mem);
3548 	} else {
3549 		/*
3550 		 * It is safe to return the slab here even though the
3551 		 * zone is unlocked because the item's allocation state
3552 		 * essentially holds a reference.
3553 		 */
3554 		ZONE_LOCK(zone);
3555 		keg = LIST_FIRST(&zone->uz_kegs)->kl_keg;
3556 		if (keg->uk_flags & UMA_ZONE_HASH)
3557 			slab = hash_sfind(&keg->uk_hash, mem);
3558 		else
3559 			slab = (uma_slab_t)(mem + keg->uk_pgoff);
3560 		ZONE_UNLOCK(zone);
3561 	}
3562 
3563 	return (slab);
3564 }
3565 
3566 /*
3567  * Set up the slab's freei data such that uma_dbg_free can function.
3568  *
3569  */
3570 static void
3571 uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item)
3572 {
3573 	uma_keg_t keg;
3574 	int freei;
3575 
3576 	if (zone_first_keg(zone) == NULL)
3577 		return;
3578 	if (slab == NULL) {
3579 		slab = uma_dbg_getslab(zone, item);
3580 		if (slab == NULL)
3581 			panic("uma: item %p did not belong to zone %s\n",
3582 			    item, zone->uz_name);
3583 	}
3584 	keg = slab->us_keg;
3585 	freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
3586 
3587 	if (BIT_ISSET(SLAB_SETSIZE, freei, &slab->us_debugfree))
3588 		panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)\n",
3589 		    item, zone, zone->uz_name, slab, freei);
3590 	BIT_SET_ATOMIC(SLAB_SETSIZE, freei, &slab->us_debugfree);
3591 
3592 	return;
3593 }
3594 
3595 /*
3596  * Verifies freed addresses.  Checks for alignment, valid slab membership
3597  * and duplicate frees.
3598  *
3599  */
3600 static void
3601 uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item)
3602 {
3603 	uma_keg_t keg;
3604 	int freei;
3605 
3606 	if (zone_first_keg(zone) == NULL)
3607 		return;
3608 	if (slab == NULL) {
3609 		slab = uma_dbg_getslab(zone, item);
3610 		if (slab == NULL)
3611 			panic("uma: Freed item %p did not belong to zone %s\n",
3612 			    item, zone->uz_name);
3613 	}
3614 	keg = slab->us_keg;
3615 	freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
3616 
3617 	if (freei >= keg->uk_ipers)
3618 		panic("Invalid free of %p from zone %p(%s) slab %p(%d)\n",
3619 		    item, zone, zone->uz_name, slab, freei);
3620 
3621 	if (((freei * keg->uk_rsize) + slab->us_data) != item)
3622 		panic("Unaligned free of %p from zone %p(%s) slab %p(%d)\n",
3623 		    item, zone, zone->uz_name, slab, freei);
3624 
3625 	if (!BIT_ISSET(SLAB_SETSIZE, freei, &slab->us_debugfree))
3626 		panic("Duplicate free of %p from zone %p(%s) slab %p(%d)\n",
3627 		    item, zone, zone->uz_name, slab, freei);
3628 
3629 	BIT_CLR_ATOMIC(SLAB_SETSIZE, freei, &slab->us_debugfree);
3630 }
3631 #endif /* INVARIANTS */
3632 
3633 #ifdef DDB
3634 DB_SHOW_COMMAND(uma, db_show_uma)
3635 {
3636 	uint64_t allocs, frees, sleeps;
3637 	uma_bucket_t bucket;
3638 	uma_keg_t kz;
3639 	uma_zone_t z;
3640 	int cachefree;
3641 
3642 	db_printf("%18s %8s %8s %8s %12s %8s %8s\n", "Zone", "Size", "Used",
3643 	    "Free", "Requests", "Sleeps", "Bucket");
3644 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3645 		LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3646 			if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
3647 				allocs = z->uz_allocs;
3648 				frees = z->uz_frees;
3649 				sleeps = z->uz_sleeps;
3650 				cachefree = 0;
3651 			} else
3652 				uma_zone_sumstat(z, &cachefree, &allocs,
3653 				    &frees, &sleeps);
3654 			if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
3655 			    (LIST_FIRST(&kz->uk_zones) != z)))
3656 				cachefree += kz->uk_free;
3657 			LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3658 				cachefree += bucket->ub_cnt;
3659 			db_printf("%18s %8ju %8jd %8d %12ju %8ju %8u\n",
3660 			    z->uz_name, (uintmax_t)kz->uk_size,
3661 			    (intmax_t)(allocs - frees), cachefree,
3662 			    (uintmax_t)allocs, sleeps, z->uz_count);
3663 			if (db_pager_quit)
3664 				return;
3665 		}
3666 	}
3667 }
3668 
3669 DB_SHOW_COMMAND(umacache, db_show_umacache)
3670 {
3671 	uint64_t allocs, frees;
3672 	uma_bucket_t bucket;
3673 	uma_zone_t z;
3674 	int cachefree;
3675 
3676 	db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
3677 	    "Requests", "Bucket");
3678 	LIST_FOREACH(z, &uma_cachezones, uz_link) {
3679 		uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL);
3680 		LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3681 			cachefree += bucket->ub_cnt;
3682 		db_printf("%18s %8ju %8jd %8d %12ju %8u\n",
3683 		    z->uz_name, (uintmax_t)z->uz_size,
3684 		    (intmax_t)(allocs - frees), cachefree,
3685 		    (uintmax_t)allocs, z->uz_count);
3686 		if (db_pager_quit)
3687 			return;
3688 	}
3689 }
3690 #endif	/* DDB */
3691