xref: /freebsd/sys/vm/uma_core.c (revision fc5ef1ca4f293be626e3d021aed0be976a79fcab)
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 #include "opt_ddb.h"
54 #include "opt_param.h"
55 #include "opt_vm.h"
56 
57 #include <sys/param.h>
58 #include <sys/systm.h>
59 #include <sys/bitset.h>
60 #include <sys/eventhandler.h>
61 #include <sys/kernel.h>
62 #include <sys/types.h>
63 #include <sys/queue.h>
64 #include <sys/malloc.h>
65 #include <sys/ktr.h>
66 #include <sys/lock.h>
67 #include <sys/sysctl.h>
68 #include <sys/mutex.h>
69 #include <sys/proc.h>
70 #include <sys/random.h>
71 #include <sys/rwlock.h>
72 #include <sys/sbuf.h>
73 #include <sys/sched.h>
74 #include <sys/smp.h>
75 #include <sys/taskqueue.h>
76 #include <sys/vmmeter.h>
77 
78 #include <vm/vm.h>
79 #include <vm/vm_object.h>
80 #include <vm/vm_page.h>
81 #include <vm/vm_pageout.h>
82 #include <vm/vm_param.h>
83 #include <vm/vm_map.h>
84 #include <vm/vm_kern.h>
85 #include <vm/vm_extern.h>
86 #include <vm/uma.h>
87 #include <vm/uma_int.h>
88 #include <vm/uma_dbg.h>
89 
90 #include <ddb/ddb.h>
91 
92 #ifdef DEBUG_MEMGUARD
93 #include <vm/memguard.h>
94 #endif
95 
96 /*
97  * This is the zone and keg from which all zones are spawned.  The idea is that
98  * even the zone & keg heads are allocated from the allocator, so we use the
99  * bss section to bootstrap us.
100  */
101 static struct uma_keg masterkeg;
102 static struct uma_zone masterzone_k;
103 static struct uma_zone masterzone_z;
104 static uma_zone_t kegs = &masterzone_k;
105 static uma_zone_t zones = &masterzone_z;
106 
107 /* This is the zone from which all of uma_slab_t's are allocated. */
108 static uma_zone_t slabzone;
109 
110 /*
111  * The initial hash tables come out of this zone so they can be allocated
112  * prior to malloc coming up.
113  */
114 static uma_zone_t hashzone;
115 
116 /* The boot-time adjusted value for cache line alignment. */
117 int uma_align_cache = 64 - 1;
118 
119 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets");
120 
121 /*
122  * Are we allowed to allocate buckets?
123  */
124 static int bucketdisable = 1;
125 
126 /* Linked list of all kegs in the system */
127 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs);
128 
129 /* Linked list of all cache-only zones in the system */
130 static LIST_HEAD(,uma_zone) uma_cachezones =
131     LIST_HEAD_INITIALIZER(uma_cachezones);
132 
133 /* This RW lock protects the keg list */
134 static struct rwlock_padalign __exclusive_cache_line uma_rwlock;
135 
136 /*
137  * Pointer and counter to pool of pages, that is preallocated at
138  * startup to bootstrap UMA.  Early zones continue to use the pool
139  * until it is depleted, so allocations may happen after boot, thus
140  * we need a mutex to protect it.
141  */
142 static char *bootmem;
143 static int boot_pages;
144 static struct mtx uma_boot_pages_mtx;
145 
146 static struct sx uma_drain_lock;
147 
148 /* Is the VM done starting up? */
149 static int booted = 0;
150 #define	UMA_STARTUP	1
151 #define	UMA_STARTUP2	2
152 
153 /*
154  * This is the handle used to schedule events that need to happen
155  * outside of the allocation fast path.
156  */
157 static struct callout uma_callout;
158 #define	UMA_TIMEOUT	20		/* Seconds for callout interval. */
159 
160 /*
161  * This structure is passed as the zone ctor arg so that I don't have to create
162  * a special allocation function just for zones.
163  */
164 struct uma_zctor_args {
165 	const char *name;
166 	size_t size;
167 	uma_ctor ctor;
168 	uma_dtor dtor;
169 	uma_init uminit;
170 	uma_fini fini;
171 	uma_import import;
172 	uma_release release;
173 	void *arg;
174 	uma_keg_t keg;
175 	int align;
176 	uint32_t flags;
177 };
178 
179 struct uma_kctor_args {
180 	uma_zone_t zone;
181 	size_t size;
182 	uma_init uminit;
183 	uma_fini fini;
184 	int align;
185 	uint32_t flags;
186 };
187 
188 struct uma_bucket_zone {
189 	uma_zone_t	ubz_zone;
190 	char		*ubz_name;
191 	int		ubz_entries;	/* Number of items it can hold. */
192 	int		ubz_maxsize;	/* Maximum allocation size per-item. */
193 };
194 
195 /*
196  * Compute the actual number of bucket entries to pack them in power
197  * of two sizes for more efficient space utilization.
198  */
199 #define	BUCKET_SIZE(n)						\
200     (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *))
201 
202 #define	BUCKET_MAX	BUCKET_SIZE(256)
203 
204 struct uma_bucket_zone bucket_zones[] = {
205 	{ NULL, "4 Bucket", BUCKET_SIZE(4), 4096 },
206 	{ NULL, "6 Bucket", BUCKET_SIZE(6), 3072 },
207 	{ NULL, "8 Bucket", BUCKET_SIZE(8), 2048 },
208 	{ NULL, "12 Bucket", BUCKET_SIZE(12), 1536 },
209 	{ NULL, "16 Bucket", BUCKET_SIZE(16), 1024 },
210 	{ NULL, "32 Bucket", BUCKET_SIZE(32), 512 },
211 	{ NULL, "64 Bucket", BUCKET_SIZE(64), 256 },
212 	{ NULL, "128 Bucket", BUCKET_SIZE(128), 128 },
213 	{ NULL, "256 Bucket", BUCKET_SIZE(256), 64 },
214 	{ NULL, NULL, 0}
215 };
216 
217 /*
218  * Flags and enumerations to be passed to internal functions.
219  */
220 enum zfreeskip { SKIP_NONE = 0, SKIP_DTOR, SKIP_FINI };
221 
222 /* Prototypes.. */
223 
224 static void *noobj_alloc(uma_zone_t, vm_size_t, uint8_t *, int);
225 static void *page_alloc(uma_zone_t, vm_size_t, uint8_t *, int);
226 static void *startup_alloc(uma_zone_t, vm_size_t, uint8_t *, int);
227 static void page_free(void *, vm_size_t, uint8_t);
228 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int);
229 static void cache_drain(uma_zone_t);
230 static void bucket_drain(uma_zone_t, uma_bucket_t);
231 static void bucket_cache_drain(uma_zone_t zone);
232 static int keg_ctor(void *, int, void *, int);
233 static void keg_dtor(void *, int, void *);
234 static int zone_ctor(void *, int, void *, int);
235 static void zone_dtor(void *, int, void *);
236 static int zero_init(void *, int, int);
237 static void keg_small_init(uma_keg_t keg);
238 static void keg_large_init(uma_keg_t keg);
239 static void zone_foreach(void (*zfunc)(uma_zone_t));
240 static void zone_timeout(uma_zone_t zone);
241 static int hash_alloc(struct uma_hash *);
242 static int hash_expand(struct uma_hash *, struct uma_hash *);
243 static void hash_free(struct uma_hash *hash);
244 static void uma_timeout(void *);
245 static void uma_startup3(void);
246 static void *zone_alloc_item(uma_zone_t, void *, int);
247 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip);
248 static void bucket_enable(void);
249 static void bucket_init(void);
250 static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int);
251 static void bucket_free(uma_zone_t zone, uma_bucket_t, void *);
252 static void bucket_zone_drain(void);
253 static uma_bucket_t zone_alloc_bucket(uma_zone_t zone, void *, int flags);
254 static uma_slab_t zone_fetch_slab(uma_zone_t zone, uma_keg_t last, int flags);
255 static uma_slab_t zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int flags);
256 static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab);
257 static void slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item);
258 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit,
259     uma_fini fini, int align, uint32_t flags);
260 static int zone_import(uma_zone_t zone, void **bucket, int max, int flags);
261 static void zone_release(uma_zone_t zone, void **bucket, int cnt);
262 static void uma_zero_item(void *item, uma_zone_t zone);
263 
264 void uma_print_zone(uma_zone_t);
265 void uma_print_stats(void);
266 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS);
267 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS);
268 
269 #ifdef INVARIANTS
270 static void uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item);
271 static void uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item);
272 #endif
273 
274 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
275 
276 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLTYPE_INT,
277     0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones");
278 
279 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLTYPE_STRUCT,
280     0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats");
281 
282 static int zone_warnings = 1;
283 SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RWTUN, &zone_warnings, 0,
284     "Warn when UMA zones becomes full");
285 
286 /*
287  * This routine checks to see whether or not it's safe to enable buckets.
288  */
289 static void
290 bucket_enable(void)
291 {
292 	bucketdisable = vm_page_count_min();
293 }
294 
295 /*
296  * Initialize bucket_zones, the array of zones of buckets of various sizes.
297  *
298  * For each zone, calculate the memory required for each bucket, consisting
299  * of the header and an array of pointers.
300  */
301 static void
302 bucket_init(void)
303 {
304 	struct uma_bucket_zone *ubz;
305 	int size;
306 
307 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) {
308 		size = roundup(sizeof(struct uma_bucket), sizeof(void *));
309 		size += sizeof(void *) * ubz->ubz_entries;
310 		ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
311 		    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
312 		    UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET);
313 	}
314 }
315 
316 /*
317  * Given a desired number of entries for a bucket, return the zone from which
318  * to allocate the bucket.
319  */
320 static struct uma_bucket_zone *
321 bucket_zone_lookup(int entries)
322 {
323 	struct uma_bucket_zone *ubz;
324 
325 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
326 		if (ubz->ubz_entries >= entries)
327 			return (ubz);
328 	ubz--;
329 	return (ubz);
330 }
331 
332 static int
333 bucket_select(int size)
334 {
335 	struct uma_bucket_zone *ubz;
336 
337 	ubz = &bucket_zones[0];
338 	if (size > ubz->ubz_maxsize)
339 		return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1);
340 
341 	for (; ubz->ubz_entries != 0; ubz++)
342 		if (ubz->ubz_maxsize < size)
343 			break;
344 	ubz--;
345 	return (ubz->ubz_entries);
346 }
347 
348 static uma_bucket_t
349 bucket_alloc(uma_zone_t zone, void *udata, int flags)
350 {
351 	struct uma_bucket_zone *ubz;
352 	uma_bucket_t bucket;
353 
354 	/*
355 	 * This is to stop us from allocating per cpu buckets while we're
356 	 * running out of vm.boot_pages.  Otherwise, we would exhaust the
357 	 * boot pages.  This also prevents us from allocating buckets in
358 	 * low memory situations.
359 	 */
360 	if (bucketdisable)
361 		return (NULL);
362 	/*
363 	 * To limit bucket recursion we store the original zone flags
364 	 * in a cookie passed via zalloc_arg/zfree_arg.  This allows the
365 	 * NOVM flag to persist even through deep recursions.  We also
366 	 * store ZFLAG_BUCKET once we have recursed attempting to allocate
367 	 * a bucket for a bucket zone so we do not allow infinite bucket
368 	 * recursion.  This cookie will even persist to frees of unused
369 	 * buckets via the allocation path or bucket allocations in the
370 	 * free path.
371 	 */
372 	if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
373 		udata = (void *)(uintptr_t)zone->uz_flags;
374 	else {
375 		if ((uintptr_t)udata & UMA_ZFLAG_BUCKET)
376 			return (NULL);
377 		udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET);
378 	}
379 	if ((uintptr_t)udata & UMA_ZFLAG_CACHEONLY)
380 		flags |= M_NOVM;
381 	ubz = bucket_zone_lookup(zone->uz_count);
382 	if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0)
383 		ubz++;
384 	bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags);
385 	if (bucket) {
386 #ifdef INVARIANTS
387 		bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
388 #endif
389 		bucket->ub_cnt = 0;
390 		bucket->ub_entries = ubz->ubz_entries;
391 	}
392 
393 	return (bucket);
394 }
395 
396 static void
397 bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata)
398 {
399 	struct uma_bucket_zone *ubz;
400 
401 	KASSERT(bucket->ub_cnt == 0,
402 	    ("bucket_free: Freeing a non free bucket."));
403 	if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
404 		udata = (void *)(uintptr_t)zone->uz_flags;
405 	ubz = bucket_zone_lookup(bucket->ub_entries);
406 	uma_zfree_arg(ubz->ubz_zone, bucket, udata);
407 }
408 
409 static void
410 bucket_zone_drain(void)
411 {
412 	struct uma_bucket_zone *ubz;
413 
414 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
415 		zone_drain(ubz->ubz_zone);
416 }
417 
418 static void
419 zone_log_warning(uma_zone_t zone)
420 {
421 	static const struct timeval warninterval = { 300, 0 };
422 
423 	if (!zone_warnings || zone->uz_warning == NULL)
424 		return;
425 
426 	if (ratecheck(&zone->uz_ratecheck, &warninterval))
427 		printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning);
428 }
429 
430 static inline void
431 zone_maxaction(uma_zone_t zone)
432 {
433 
434 	if (zone->uz_maxaction.ta_func != NULL)
435 		taskqueue_enqueue(taskqueue_thread, &zone->uz_maxaction);
436 }
437 
438 static void
439 zone_foreach_keg(uma_zone_t zone, void (*kegfn)(uma_keg_t))
440 {
441 	uma_klink_t klink;
442 
443 	LIST_FOREACH(klink, &zone->uz_kegs, kl_link)
444 		kegfn(klink->kl_keg);
445 }
446 
447 /*
448  * Routine called by timeout which is used to fire off some time interval
449  * based calculations.  (stats, hash size, etc.)
450  *
451  * Arguments:
452  *	arg   Unused
453  *
454  * Returns:
455  *	Nothing
456  */
457 static void
458 uma_timeout(void *unused)
459 {
460 	bucket_enable();
461 	zone_foreach(zone_timeout);
462 
463 	/* Reschedule this event */
464 	callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
465 }
466 
467 /*
468  * Routine to perform timeout driven calculations.  This expands the
469  * hashes and does per cpu statistics aggregation.
470  *
471  *  Returns nothing.
472  */
473 static void
474 keg_timeout(uma_keg_t keg)
475 {
476 
477 	KEG_LOCK(keg);
478 	/*
479 	 * Expand the keg hash table.
480 	 *
481 	 * This is done if the number of slabs is larger than the hash size.
482 	 * What I'm trying to do here is completely reduce collisions.  This
483 	 * may be a little aggressive.  Should I allow for two collisions max?
484 	 */
485 	if (keg->uk_flags & UMA_ZONE_HASH &&
486 	    keg->uk_pages / keg->uk_ppera >= keg->uk_hash.uh_hashsize) {
487 		struct uma_hash newhash;
488 		struct uma_hash oldhash;
489 		int ret;
490 
491 		/*
492 		 * This is so involved because allocating and freeing
493 		 * while the keg lock is held will lead to deadlock.
494 		 * I have to do everything in stages and check for
495 		 * races.
496 		 */
497 		newhash = keg->uk_hash;
498 		KEG_UNLOCK(keg);
499 		ret = hash_alloc(&newhash);
500 		KEG_LOCK(keg);
501 		if (ret) {
502 			if (hash_expand(&keg->uk_hash, &newhash)) {
503 				oldhash = keg->uk_hash;
504 				keg->uk_hash = newhash;
505 			} else
506 				oldhash = newhash;
507 
508 			KEG_UNLOCK(keg);
509 			hash_free(&oldhash);
510 			return;
511 		}
512 	}
513 	KEG_UNLOCK(keg);
514 }
515 
516 static void
517 zone_timeout(uma_zone_t zone)
518 {
519 
520 	zone_foreach_keg(zone, &keg_timeout);
521 }
522 
523 /*
524  * Allocate and zero fill the next sized hash table from the appropriate
525  * backing store.
526  *
527  * Arguments:
528  *	hash  A new hash structure with the old hash size in uh_hashsize
529  *
530  * Returns:
531  *	1 on success and 0 on failure.
532  */
533 static int
534 hash_alloc(struct uma_hash *hash)
535 {
536 	int oldsize;
537 	int alloc;
538 
539 	oldsize = hash->uh_hashsize;
540 
541 	/* We're just going to go to a power of two greater */
542 	if (oldsize)  {
543 		hash->uh_hashsize = oldsize * 2;
544 		alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
545 		hash->uh_slab_hash = (struct slabhead *)malloc(alloc,
546 		    M_UMAHASH, M_NOWAIT);
547 	} else {
548 		alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
549 		hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
550 		    M_WAITOK);
551 		hash->uh_hashsize = UMA_HASH_SIZE_INIT;
552 	}
553 	if (hash->uh_slab_hash) {
554 		bzero(hash->uh_slab_hash, alloc);
555 		hash->uh_hashmask = hash->uh_hashsize - 1;
556 		return (1);
557 	}
558 
559 	return (0);
560 }
561 
562 /*
563  * Expands the hash table for HASH zones.  This is done from zone_timeout
564  * to reduce collisions.  This must not be done in the regular allocation
565  * path, otherwise, we can recurse on the vm while allocating pages.
566  *
567  * Arguments:
568  *	oldhash  The hash you want to expand
569  *	newhash  The hash structure for the new table
570  *
571  * Returns:
572  *	Nothing
573  *
574  * Discussion:
575  */
576 static int
577 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
578 {
579 	uma_slab_t slab;
580 	int hval;
581 	int i;
582 
583 	if (!newhash->uh_slab_hash)
584 		return (0);
585 
586 	if (oldhash->uh_hashsize >= newhash->uh_hashsize)
587 		return (0);
588 
589 	/*
590 	 * I need to investigate hash algorithms for resizing without a
591 	 * full rehash.
592 	 */
593 
594 	for (i = 0; i < oldhash->uh_hashsize; i++)
595 		while (!SLIST_EMPTY(&oldhash->uh_slab_hash[i])) {
596 			slab = SLIST_FIRST(&oldhash->uh_slab_hash[i]);
597 			SLIST_REMOVE_HEAD(&oldhash->uh_slab_hash[i], us_hlink);
598 			hval = UMA_HASH(newhash, slab->us_data);
599 			SLIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
600 			    slab, us_hlink);
601 		}
602 
603 	return (1);
604 }
605 
606 /*
607  * Free the hash bucket to the appropriate backing store.
608  *
609  * Arguments:
610  *	slab_hash  The hash bucket we're freeing
611  *	hashsize   The number of entries in that hash bucket
612  *
613  * Returns:
614  *	Nothing
615  */
616 static void
617 hash_free(struct uma_hash *hash)
618 {
619 	if (hash->uh_slab_hash == NULL)
620 		return;
621 	if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
622 		zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE);
623 	else
624 		free(hash->uh_slab_hash, M_UMAHASH);
625 }
626 
627 /*
628  * Frees all outstanding items in a bucket
629  *
630  * Arguments:
631  *	zone   The zone to free to, must be unlocked.
632  *	bucket The free/alloc bucket with items, cpu queue must be locked.
633  *
634  * Returns:
635  *	Nothing
636  */
637 
638 static void
639 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
640 {
641 	int i;
642 
643 	if (bucket == NULL)
644 		return;
645 
646 	if (zone->uz_fini)
647 		for (i = 0; i < bucket->ub_cnt; i++)
648 			zone->uz_fini(bucket->ub_bucket[i], zone->uz_size);
649 	zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt);
650 	bucket->ub_cnt = 0;
651 }
652 
653 /*
654  * Drains the per cpu caches for a zone.
655  *
656  * NOTE: This may only be called while the zone is being turn down, and not
657  * during normal operation.  This is necessary in order that we do not have
658  * to migrate CPUs to drain the per-CPU caches.
659  *
660  * Arguments:
661  *	zone     The zone to drain, must be unlocked.
662  *
663  * Returns:
664  *	Nothing
665  */
666 static void
667 cache_drain(uma_zone_t zone)
668 {
669 	uma_cache_t cache;
670 	int cpu;
671 
672 	/*
673 	 * XXX: It is safe to not lock the per-CPU caches, because we're
674 	 * tearing down the zone anyway.  I.e., there will be no further use
675 	 * of the caches at this point.
676 	 *
677 	 * XXX: It would good to be able to assert that the zone is being
678 	 * torn down to prevent improper use of cache_drain().
679 	 *
680 	 * XXX: We lock the zone before passing into bucket_cache_drain() as
681 	 * it is used elsewhere.  Should the tear-down path be made special
682 	 * there in some form?
683 	 */
684 	CPU_FOREACH(cpu) {
685 		cache = &zone->uz_cpu[cpu];
686 		bucket_drain(zone, cache->uc_allocbucket);
687 		bucket_drain(zone, cache->uc_freebucket);
688 		if (cache->uc_allocbucket != NULL)
689 			bucket_free(zone, cache->uc_allocbucket, NULL);
690 		if (cache->uc_freebucket != NULL)
691 			bucket_free(zone, cache->uc_freebucket, NULL);
692 		cache->uc_allocbucket = cache->uc_freebucket = NULL;
693 	}
694 	ZONE_LOCK(zone);
695 	bucket_cache_drain(zone);
696 	ZONE_UNLOCK(zone);
697 }
698 
699 static void
700 cache_shrink(uma_zone_t zone)
701 {
702 
703 	if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
704 		return;
705 
706 	ZONE_LOCK(zone);
707 	zone->uz_count = (zone->uz_count_min + zone->uz_count) / 2;
708 	ZONE_UNLOCK(zone);
709 }
710 
711 static void
712 cache_drain_safe_cpu(uma_zone_t zone)
713 {
714 	uma_cache_t cache;
715 	uma_bucket_t b1, b2;
716 
717 	if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
718 		return;
719 
720 	b1 = b2 = NULL;
721 	ZONE_LOCK(zone);
722 	critical_enter();
723 	cache = &zone->uz_cpu[curcpu];
724 	if (cache->uc_allocbucket) {
725 		if (cache->uc_allocbucket->ub_cnt != 0)
726 			LIST_INSERT_HEAD(&zone->uz_buckets,
727 			    cache->uc_allocbucket, ub_link);
728 		else
729 			b1 = cache->uc_allocbucket;
730 		cache->uc_allocbucket = NULL;
731 	}
732 	if (cache->uc_freebucket) {
733 		if (cache->uc_freebucket->ub_cnt != 0)
734 			LIST_INSERT_HEAD(&zone->uz_buckets,
735 			    cache->uc_freebucket, ub_link);
736 		else
737 			b2 = cache->uc_freebucket;
738 		cache->uc_freebucket = NULL;
739 	}
740 	critical_exit();
741 	ZONE_UNLOCK(zone);
742 	if (b1)
743 		bucket_free(zone, b1, NULL);
744 	if (b2)
745 		bucket_free(zone, b2, NULL);
746 }
747 
748 /*
749  * Safely drain per-CPU caches of a zone(s) to alloc bucket.
750  * This is an expensive call because it needs to bind to all CPUs
751  * one by one and enter a critical section on each of them in order
752  * to safely access their cache buckets.
753  * Zone lock must not be held on call this function.
754  */
755 static void
756 cache_drain_safe(uma_zone_t zone)
757 {
758 	int cpu;
759 
760 	/*
761 	 * Polite bucket sizes shrinking was not enouth, shrink aggressively.
762 	 */
763 	if (zone)
764 		cache_shrink(zone);
765 	else
766 		zone_foreach(cache_shrink);
767 
768 	CPU_FOREACH(cpu) {
769 		thread_lock(curthread);
770 		sched_bind(curthread, cpu);
771 		thread_unlock(curthread);
772 
773 		if (zone)
774 			cache_drain_safe_cpu(zone);
775 		else
776 			zone_foreach(cache_drain_safe_cpu);
777 	}
778 	thread_lock(curthread);
779 	sched_unbind(curthread);
780 	thread_unlock(curthread);
781 }
782 
783 /*
784  * Drain the cached buckets from a zone.  Expects a locked zone on entry.
785  */
786 static void
787 bucket_cache_drain(uma_zone_t zone)
788 {
789 	uma_bucket_t bucket;
790 
791 	/*
792 	 * Drain the bucket queues and free the buckets, we just keep two per
793 	 * cpu (alloc/free).
794 	 */
795 	while ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) {
796 		LIST_REMOVE(bucket, ub_link);
797 		ZONE_UNLOCK(zone);
798 		bucket_drain(zone, bucket);
799 		bucket_free(zone, bucket, NULL);
800 		ZONE_LOCK(zone);
801 	}
802 
803 	/*
804 	 * Shrink further bucket sizes.  Price of single zone lock collision
805 	 * is probably lower then price of global cache drain.
806 	 */
807 	if (zone->uz_count > zone->uz_count_min)
808 		zone->uz_count--;
809 }
810 
811 static void
812 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start)
813 {
814 	uint8_t *mem;
815 	int i;
816 	uint8_t flags;
817 
818 	CTR4(KTR_UMA, "keg_free_slab keg %s(%p) slab %p, returning %d bytes",
819 	    keg->uk_name, keg, slab, PAGE_SIZE * keg->uk_ppera);
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 	keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags);
832 }
833 
834 /*
835  * Frees pages from a keg back to the system.  This is done on demand from
836  * the pageout daemon.
837  *
838  * Returns nothing.
839  */
840 static void
841 keg_drain(uma_keg_t keg)
842 {
843 	struct slabhead freeslabs = { 0 };
844 	uma_slab_t slab, tmp;
845 
846 	/*
847 	 * We don't want to take pages from statically allocated kegs at this
848 	 * time
849 	 */
850 	if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL)
851 		return;
852 
853 	CTR3(KTR_UMA, "keg_drain %s(%p) free items: %u",
854 	    keg->uk_name, keg, keg->uk_free);
855 	KEG_LOCK(keg);
856 	if (keg->uk_free == 0)
857 		goto finished;
858 
859 	LIST_FOREACH_SAFE(slab, &keg->uk_free_slab, us_link, tmp) {
860 		/* We have nowhere to free these to. */
861 		if (slab->us_flags & UMA_SLAB_BOOT)
862 			continue;
863 
864 		LIST_REMOVE(slab, us_link);
865 		keg->uk_pages -= keg->uk_ppera;
866 		keg->uk_free -= keg->uk_ipers;
867 
868 		if (keg->uk_flags & UMA_ZONE_HASH)
869 			UMA_HASH_REMOVE(&keg->uk_hash, slab, slab->us_data);
870 
871 		SLIST_INSERT_HEAD(&freeslabs, slab, us_hlink);
872 	}
873 finished:
874 	KEG_UNLOCK(keg);
875 
876 	while ((slab = SLIST_FIRST(&freeslabs)) != NULL) {
877 		SLIST_REMOVE(&freeslabs, slab, uma_slab, us_hlink);
878 		keg_free_slab(keg, slab, keg->uk_ipers);
879 	}
880 }
881 
882 static void
883 zone_drain_wait(uma_zone_t zone, int waitok)
884 {
885 
886 	/*
887 	 * Set draining to interlock with zone_dtor() so we can release our
888 	 * locks as we go.  Only dtor() should do a WAITOK call since it
889 	 * is the only call that knows the structure will still be available
890 	 * when it wakes up.
891 	 */
892 	ZONE_LOCK(zone);
893 	while (zone->uz_flags & UMA_ZFLAG_DRAINING) {
894 		if (waitok == M_NOWAIT)
895 			goto out;
896 		msleep(zone, zone->uz_lockptr, PVM, "zonedrain", 1);
897 	}
898 	zone->uz_flags |= UMA_ZFLAG_DRAINING;
899 	bucket_cache_drain(zone);
900 	ZONE_UNLOCK(zone);
901 	/*
902 	 * The DRAINING flag protects us from being freed while
903 	 * we're running.  Normally the uma_rwlock would protect us but we
904 	 * must be able to release and acquire the right lock for each keg.
905 	 */
906 	zone_foreach_keg(zone, &keg_drain);
907 	ZONE_LOCK(zone);
908 	zone->uz_flags &= ~UMA_ZFLAG_DRAINING;
909 	wakeup(zone);
910 out:
911 	ZONE_UNLOCK(zone);
912 }
913 
914 void
915 zone_drain(uma_zone_t zone)
916 {
917 
918 	zone_drain_wait(zone, M_NOWAIT);
919 }
920 
921 /*
922  * Allocate a new slab for a keg.  This does not insert the slab onto a list.
923  *
924  * Arguments:
925  *	wait  Shall we wait?
926  *
927  * Returns:
928  *	The slab that was allocated or NULL if there is no memory and the
929  *	caller specified M_NOWAIT.
930  */
931 static uma_slab_t
932 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int wait)
933 {
934 	uma_alloc allocf;
935 	uma_slab_t slab;
936 	uint8_t *mem;
937 	uint8_t flags;
938 	int i;
939 
940 	mtx_assert(&keg->uk_lock, MA_OWNED);
941 	slab = NULL;
942 	mem = NULL;
943 
944 	allocf = keg->uk_allocf;
945 	KEG_UNLOCK(keg);
946 
947 	if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
948 		slab = zone_alloc_item(keg->uk_slabzone, NULL, wait);
949 		if (slab == NULL)
950 			goto out;
951 	}
952 
953 	/*
954 	 * This reproduces the old vm_zone behavior of zero filling pages the
955 	 * first time they are added to a zone.
956 	 *
957 	 * Malloced items are zeroed in uma_zalloc.
958 	 */
959 
960 	if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
961 		wait |= M_ZERO;
962 	else
963 		wait &= ~M_ZERO;
964 
965 	if (keg->uk_flags & UMA_ZONE_NODUMP)
966 		wait |= M_NODUMP;
967 
968 	/* zone is passed for legacy reasons. */
969 	mem = allocf(zone, keg->uk_ppera * PAGE_SIZE, &flags, wait);
970 	if (mem == NULL) {
971 		if (keg->uk_flags & UMA_ZONE_OFFPAGE)
972 			zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE);
973 		slab = NULL;
974 		goto out;
975 	}
976 
977 	/* Point the slab into the allocated memory */
978 	if (!(keg->uk_flags & UMA_ZONE_OFFPAGE))
979 		slab = (uma_slab_t )(mem + keg->uk_pgoff);
980 
981 	if (keg->uk_flags & UMA_ZONE_VTOSLAB)
982 		for (i = 0; i < keg->uk_ppera; i++)
983 			vsetslab((vm_offset_t)mem + (i * PAGE_SIZE), slab);
984 
985 	slab->us_keg = keg;
986 	slab->us_data = mem;
987 	slab->us_freecount = keg->uk_ipers;
988 	slab->us_flags = flags;
989 	BIT_FILL(SLAB_SETSIZE, &slab->us_free);
990 #ifdef INVARIANTS
991 	BIT_ZERO(SLAB_SETSIZE, &slab->us_debugfree);
992 #endif
993 
994 	if (keg->uk_init != NULL) {
995 		for (i = 0; i < keg->uk_ipers; i++)
996 			if (keg->uk_init(slab->us_data + (keg->uk_rsize * i),
997 			    keg->uk_size, wait) != 0)
998 				break;
999 		if (i != keg->uk_ipers) {
1000 			keg_free_slab(keg, slab, i);
1001 			slab = NULL;
1002 			goto out;
1003 		}
1004 	}
1005 out:
1006 	KEG_LOCK(keg);
1007 
1008 	CTR3(KTR_UMA, "keg_alloc_slab: allocated slab %p for %s(%p)",
1009 	    slab, keg->uk_name, keg);
1010 
1011 	if (slab != NULL) {
1012 		if (keg->uk_flags & UMA_ZONE_HASH)
1013 			UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
1014 
1015 		keg->uk_pages += keg->uk_ppera;
1016 		keg->uk_free += keg->uk_ipers;
1017 	}
1018 
1019 	return (slab);
1020 }
1021 
1022 /*
1023  * This function is intended to be used early on in place of page_alloc() so
1024  * that we may use the boot time page cache to satisfy allocations before
1025  * the VM is ready.
1026  */
1027 static void *
1028 startup_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *pflag, int wait)
1029 {
1030 	uma_keg_t keg;
1031 	void *mem;
1032 	int pages;
1033 
1034 	keg = zone_first_keg(zone);
1035 	pages = howmany(bytes, PAGE_SIZE);
1036 	KASSERT(pages > 0, ("startup_alloc can't reserve 0 pages\n"));
1037 
1038 	/*
1039 	 * Check our small startup cache to see if it has pages remaining.
1040 	 */
1041 	mtx_lock(&uma_boot_pages_mtx);
1042 	if (pages <= boot_pages) {
1043 		mem = bootmem;
1044 		boot_pages -= pages;
1045 		bootmem += pages * PAGE_SIZE;
1046 		mtx_unlock(&uma_boot_pages_mtx);
1047 		*pflag = UMA_SLAB_BOOT;
1048 		return (mem);
1049 	}
1050 	mtx_unlock(&uma_boot_pages_mtx);
1051 	if (booted < UMA_STARTUP2)
1052 		panic("UMA: Increase vm.boot_pages");
1053 	/*
1054 	 * Now that we've booted reset these users to their real allocator.
1055 	 */
1056 #ifdef UMA_MD_SMALL_ALLOC
1057 	keg->uk_allocf = (keg->uk_ppera > 1) ? page_alloc : uma_small_alloc;
1058 #else
1059 	keg->uk_allocf = page_alloc;
1060 #endif
1061 	return keg->uk_allocf(zone, bytes, pflag, wait);
1062 }
1063 
1064 /*
1065  * Allocates a number of pages from the system
1066  *
1067  * Arguments:
1068  *	bytes  The number of bytes requested
1069  *	wait  Shall we wait?
1070  *
1071  * Returns:
1072  *	A pointer to the alloced memory or possibly
1073  *	NULL if M_NOWAIT is set.
1074  */
1075 static void *
1076 page_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *pflag, int wait)
1077 {
1078 	void *p;	/* Returned page */
1079 
1080 	*pflag = UMA_SLAB_KMEM;
1081 	p = (void *) kmem_malloc(kmem_arena, bytes, wait);
1082 
1083 	return (p);
1084 }
1085 
1086 /*
1087  * Allocates a number of pages from within an object
1088  *
1089  * Arguments:
1090  *	bytes  The number of bytes requested
1091  *	wait   Shall we wait?
1092  *
1093  * Returns:
1094  *	A pointer to the alloced memory or possibly
1095  *	NULL if M_NOWAIT is set.
1096  */
1097 static void *
1098 noobj_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *flags, int wait)
1099 {
1100 	TAILQ_HEAD(, vm_page) alloctail;
1101 	u_long npages;
1102 	vm_offset_t retkva, zkva;
1103 	vm_page_t p, p_next;
1104 	uma_keg_t keg;
1105 
1106 	TAILQ_INIT(&alloctail);
1107 	keg = zone_first_keg(zone);
1108 
1109 	npages = howmany(bytes, PAGE_SIZE);
1110 	while (npages > 0) {
1111 		p = vm_page_alloc(NULL, 0, VM_ALLOC_INTERRUPT |
1112 		    VM_ALLOC_WIRED | VM_ALLOC_NOOBJ |
1113 		    (wait & M_WAITOK) ? VM_ALLOC_WAITOK : VM_ALLOC_NOWAIT);
1114 		if (p != NULL) {
1115 			/*
1116 			 * Since the page does not belong to an object, its
1117 			 * listq is unused.
1118 			 */
1119 			TAILQ_INSERT_TAIL(&alloctail, p, listq);
1120 			npages--;
1121 			continue;
1122 		}
1123 		/*
1124 		 * Page allocation failed, free intermediate pages and
1125 		 * exit.
1126 		 */
1127 		TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1128 			vm_page_unwire(p, PQ_NONE);
1129 			vm_page_free(p);
1130 		}
1131 		return (NULL);
1132 	}
1133 	*flags = UMA_SLAB_PRIV;
1134 	zkva = keg->uk_kva +
1135 	    atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
1136 	retkva = zkva;
1137 	TAILQ_FOREACH(p, &alloctail, listq) {
1138 		pmap_qenter(zkva, &p, 1);
1139 		zkva += PAGE_SIZE;
1140 	}
1141 
1142 	return ((void *)retkva);
1143 }
1144 
1145 /*
1146  * Frees a number of pages to the system
1147  *
1148  * Arguments:
1149  *	mem   A pointer to the memory to be freed
1150  *	size  The size of the memory being freed
1151  *	flags The original p->us_flags field
1152  *
1153  * Returns:
1154  *	Nothing
1155  */
1156 static void
1157 page_free(void *mem, vm_size_t size, uint8_t flags)
1158 {
1159 	struct vmem *vmem;
1160 
1161 	if (flags & UMA_SLAB_KMEM)
1162 		vmem = kmem_arena;
1163 	else if (flags & UMA_SLAB_KERNEL)
1164 		vmem = kernel_arena;
1165 	else
1166 		panic("UMA: page_free used with invalid flags %x", flags);
1167 
1168 	kmem_free(vmem, (vm_offset_t)mem, size);
1169 }
1170 
1171 /*
1172  * Zero fill initializer
1173  *
1174  * Arguments/Returns follow uma_init specifications
1175  */
1176 static int
1177 zero_init(void *mem, int size, int flags)
1178 {
1179 	bzero(mem, size);
1180 	return (0);
1181 }
1182 
1183 /*
1184  * Finish creating a small uma keg.  This calculates ipers, and the keg size.
1185  *
1186  * Arguments
1187  *	keg  The zone we should initialize
1188  *
1189  * Returns
1190  *	Nothing
1191  */
1192 static void
1193 keg_small_init(uma_keg_t keg)
1194 {
1195 	u_int rsize;
1196 	u_int memused;
1197 	u_int wastedspace;
1198 	u_int shsize;
1199 	u_int slabsize;
1200 
1201 	if (keg->uk_flags & UMA_ZONE_PCPU) {
1202 		u_int ncpus = (mp_maxid + 1) ? (mp_maxid + 1) : MAXCPU;
1203 
1204 		slabsize = sizeof(struct pcpu);
1205 		keg->uk_ppera = howmany(ncpus * sizeof(struct pcpu),
1206 		    PAGE_SIZE);
1207 	} else {
1208 		slabsize = UMA_SLAB_SIZE;
1209 		keg->uk_ppera = 1;
1210 	}
1211 
1212 	/*
1213 	 * Calculate the size of each allocation (rsize) according to
1214 	 * alignment.  If the requested size is smaller than we have
1215 	 * allocation bits for we round it up.
1216 	 */
1217 	rsize = keg->uk_size;
1218 	if (rsize < slabsize / SLAB_SETSIZE)
1219 		rsize = slabsize / SLAB_SETSIZE;
1220 	if (rsize & keg->uk_align)
1221 		rsize = (rsize & ~keg->uk_align) + (keg->uk_align + 1);
1222 	keg->uk_rsize = rsize;
1223 
1224 	KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
1225 	    keg->uk_rsize < sizeof(struct pcpu),
1226 	    ("%s: size %u too large", __func__, keg->uk_rsize));
1227 
1228 	if (keg->uk_flags & UMA_ZONE_OFFPAGE)
1229 		shsize = 0;
1230 	else
1231 		shsize = sizeof(struct uma_slab);
1232 
1233 	keg->uk_ipers = (slabsize - shsize) / rsize;
1234 	KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE,
1235 	    ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers));
1236 
1237 	memused = keg->uk_ipers * rsize + shsize;
1238 	wastedspace = slabsize - memused;
1239 
1240 	/*
1241 	 * We can't do OFFPAGE if we're internal or if we've been
1242 	 * asked to not go to the VM for buckets.  If we do this we
1243 	 * may end up going to the VM  for slabs which we do not
1244 	 * want to do if we're UMA_ZFLAG_CACHEONLY as a result
1245 	 * of UMA_ZONE_VM, which clearly forbids it.
1246 	 */
1247 	if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) ||
1248 	    (keg->uk_flags & UMA_ZFLAG_CACHEONLY))
1249 		return;
1250 
1251 	/*
1252 	 * See if using an OFFPAGE slab will limit our waste.  Only do
1253 	 * this if it permits more items per-slab.
1254 	 *
1255 	 * XXX We could try growing slabsize to limit max waste as well.
1256 	 * Historically this was not done because the VM could not
1257 	 * efficiently handle contiguous allocations.
1258 	 */
1259 	if ((wastedspace >= slabsize / UMA_MAX_WASTE) &&
1260 	    (keg->uk_ipers < (slabsize / keg->uk_rsize))) {
1261 		keg->uk_ipers = slabsize / keg->uk_rsize;
1262 		KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE,
1263 		    ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers));
1264 		CTR6(KTR_UMA, "UMA decided we need offpage slab headers for "
1265 		    "keg: %s(%p), calculated wastedspace = %d, "
1266 		    "maximum wasted space allowed = %d, "
1267 		    "calculated ipers = %d, "
1268 		    "new wasted space = %d\n", keg->uk_name, keg, wastedspace,
1269 		    slabsize / UMA_MAX_WASTE, keg->uk_ipers,
1270 		    slabsize - keg->uk_ipers * keg->uk_rsize);
1271 		keg->uk_flags |= UMA_ZONE_OFFPAGE;
1272 	}
1273 
1274 	if ((keg->uk_flags & UMA_ZONE_OFFPAGE) &&
1275 	    (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1276 		keg->uk_flags |= UMA_ZONE_HASH;
1277 }
1278 
1279 /*
1280  * Finish creating a large (> UMA_SLAB_SIZE) uma kegs.  Just give in and do
1281  * OFFPAGE for now.  When I can allow for more dynamic slab sizes this will be
1282  * more complicated.
1283  *
1284  * Arguments
1285  *	keg  The keg we should initialize
1286  *
1287  * Returns
1288  *	Nothing
1289  */
1290 static void
1291 keg_large_init(uma_keg_t keg)
1292 {
1293 	u_int shsize;
1294 
1295 	KASSERT(keg != NULL, ("Keg is null in keg_large_init"));
1296 	KASSERT((keg->uk_flags & UMA_ZFLAG_CACHEONLY) == 0,
1297 	    ("keg_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY keg"));
1298 	KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0,
1299 	    ("%s: Cannot large-init a UMA_ZONE_PCPU keg", __func__));
1300 
1301 	keg->uk_ppera = howmany(keg->uk_size, PAGE_SIZE);
1302 	keg->uk_ipers = 1;
1303 	keg->uk_rsize = keg->uk_size;
1304 
1305 	/* Check whether we have enough space to not do OFFPAGE. */
1306 	if ((keg->uk_flags & UMA_ZONE_OFFPAGE) == 0) {
1307 		shsize = sizeof(struct uma_slab);
1308 		if (shsize & UMA_ALIGN_PTR)
1309 			shsize = (shsize & ~UMA_ALIGN_PTR) +
1310 			    (UMA_ALIGN_PTR + 1);
1311 
1312 		if (PAGE_SIZE * keg->uk_ppera - keg->uk_rsize < shsize) {
1313 			/*
1314 			 * We can't do OFFPAGE if we're internal, in which case
1315 			 * we need an extra page per allocation to contain the
1316 			 * slab header.
1317 			 */
1318 			if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) == 0)
1319 				keg->uk_flags |= UMA_ZONE_OFFPAGE;
1320 			else
1321 				keg->uk_ppera++;
1322 		}
1323 	}
1324 
1325 	if ((keg->uk_flags & UMA_ZONE_OFFPAGE) &&
1326 	    (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1327 		keg->uk_flags |= UMA_ZONE_HASH;
1328 }
1329 
1330 static void
1331 keg_cachespread_init(uma_keg_t keg)
1332 {
1333 	int alignsize;
1334 	int trailer;
1335 	int pages;
1336 	int rsize;
1337 
1338 	KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0,
1339 	    ("%s: Cannot cachespread-init a UMA_ZONE_PCPU keg", __func__));
1340 
1341 	alignsize = keg->uk_align + 1;
1342 	rsize = keg->uk_size;
1343 	/*
1344 	 * We want one item to start on every align boundary in a page.  To
1345 	 * do this we will span pages.  We will also extend the item by the
1346 	 * size of align if it is an even multiple of align.  Otherwise, it
1347 	 * would fall on the same boundary every time.
1348 	 */
1349 	if (rsize & keg->uk_align)
1350 		rsize = (rsize & ~keg->uk_align) + alignsize;
1351 	if ((rsize & alignsize) == 0)
1352 		rsize += alignsize;
1353 	trailer = rsize - keg->uk_size;
1354 	pages = (rsize * (PAGE_SIZE / alignsize)) / PAGE_SIZE;
1355 	pages = MIN(pages, (128 * 1024) / PAGE_SIZE);
1356 	keg->uk_rsize = rsize;
1357 	keg->uk_ppera = pages;
1358 	keg->uk_ipers = ((pages * PAGE_SIZE) + trailer) / rsize;
1359 	keg->uk_flags |= UMA_ZONE_OFFPAGE | UMA_ZONE_VTOSLAB;
1360 	KASSERT(keg->uk_ipers <= SLAB_SETSIZE,
1361 	    ("%s: keg->uk_ipers too high(%d) increase max_ipers", __func__,
1362 	    keg->uk_ipers));
1363 }
1364 
1365 /*
1366  * Keg header ctor.  This initializes all fields, locks, etc.  And inserts
1367  * the keg onto the global keg list.
1368  *
1369  * Arguments/Returns follow uma_ctor specifications
1370  *	udata  Actually uma_kctor_args
1371  */
1372 static int
1373 keg_ctor(void *mem, int size, void *udata, int flags)
1374 {
1375 	struct uma_kctor_args *arg = udata;
1376 	uma_keg_t keg = mem;
1377 	uma_zone_t zone;
1378 
1379 	bzero(keg, size);
1380 	keg->uk_size = arg->size;
1381 	keg->uk_init = arg->uminit;
1382 	keg->uk_fini = arg->fini;
1383 	keg->uk_align = arg->align;
1384 	keg->uk_free = 0;
1385 	keg->uk_reserve = 0;
1386 	keg->uk_pages = 0;
1387 	keg->uk_flags = arg->flags;
1388 	keg->uk_slabzone = NULL;
1389 
1390 	/*
1391 	 * The master zone is passed to us at keg-creation time.
1392 	 */
1393 	zone = arg->zone;
1394 	keg->uk_name = zone->uz_name;
1395 
1396 	if (arg->flags & UMA_ZONE_VM)
1397 		keg->uk_flags |= UMA_ZFLAG_CACHEONLY;
1398 
1399 	if (arg->flags & UMA_ZONE_ZINIT)
1400 		keg->uk_init = zero_init;
1401 
1402 	if (arg->flags & UMA_ZONE_MALLOC)
1403 		keg->uk_flags |= UMA_ZONE_VTOSLAB;
1404 
1405 	if (arg->flags & UMA_ZONE_PCPU)
1406 #ifdef SMP
1407 		keg->uk_flags |= UMA_ZONE_OFFPAGE;
1408 #else
1409 		keg->uk_flags &= ~UMA_ZONE_PCPU;
1410 #endif
1411 
1412 	if (keg->uk_flags & UMA_ZONE_CACHESPREAD) {
1413 		keg_cachespread_init(keg);
1414 	} else {
1415 		if (keg->uk_size > (UMA_SLAB_SIZE - sizeof(struct uma_slab)))
1416 			keg_large_init(keg);
1417 		else
1418 			keg_small_init(keg);
1419 	}
1420 
1421 	if (keg->uk_flags & UMA_ZONE_OFFPAGE)
1422 		keg->uk_slabzone = slabzone;
1423 
1424 	/*
1425 	 * If we haven't booted yet we need allocations to go through the
1426 	 * startup cache until the vm is ready.
1427 	 */
1428 	if (booted < UMA_STARTUP2)
1429 		keg->uk_allocf = startup_alloc;
1430 #ifdef UMA_MD_SMALL_ALLOC
1431 	else if (keg->uk_ppera == 1)
1432 		keg->uk_allocf = uma_small_alloc;
1433 #endif
1434 	else
1435 		keg->uk_allocf = page_alloc;
1436 #ifdef UMA_MD_SMALL_ALLOC
1437 	if (keg->uk_ppera == 1)
1438 		keg->uk_freef = uma_small_free;
1439 	else
1440 #endif
1441 		keg->uk_freef = page_free;
1442 
1443 	/*
1444 	 * Initialize keg's lock
1445 	 */
1446 	KEG_LOCK_INIT(keg, (arg->flags & UMA_ZONE_MTXCLASS));
1447 
1448 	/*
1449 	 * If we're putting the slab header in the actual page we need to
1450 	 * figure out where in each page it goes.  This calculates a right
1451 	 * justified offset into the memory on an ALIGN_PTR boundary.
1452 	 */
1453 	if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) {
1454 		u_int totsize;
1455 
1456 		/* Size of the slab struct and free list */
1457 		totsize = sizeof(struct uma_slab);
1458 
1459 		if (totsize & UMA_ALIGN_PTR)
1460 			totsize = (totsize & ~UMA_ALIGN_PTR) +
1461 			    (UMA_ALIGN_PTR + 1);
1462 		keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - totsize;
1463 
1464 		/*
1465 		 * The only way the following is possible is if with our
1466 		 * UMA_ALIGN_PTR adjustments we are now bigger than
1467 		 * UMA_SLAB_SIZE.  I haven't checked whether this is
1468 		 * mathematically possible for all cases, so we make
1469 		 * sure here anyway.
1470 		 */
1471 		totsize = keg->uk_pgoff + sizeof(struct uma_slab);
1472 		if (totsize > PAGE_SIZE * keg->uk_ppera) {
1473 			printf("zone %s ipers %d rsize %d size %d\n",
1474 			    zone->uz_name, keg->uk_ipers, keg->uk_rsize,
1475 			    keg->uk_size);
1476 			panic("UMA slab won't fit.");
1477 		}
1478 	}
1479 
1480 	if (keg->uk_flags & UMA_ZONE_HASH)
1481 		hash_alloc(&keg->uk_hash);
1482 
1483 	CTR5(KTR_UMA, "keg_ctor %p zone %s(%p) out %d free %d\n",
1484 	    keg, zone->uz_name, zone,
1485 	    (keg->uk_pages / keg->uk_ppera) * keg->uk_ipers - keg->uk_free,
1486 	    keg->uk_free);
1487 
1488 	LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
1489 
1490 	rw_wlock(&uma_rwlock);
1491 	LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
1492 	rw_wunlock(&uma_rwlock);
1493 	return (0);
1494 }
1495 
1496 /*
1497  * Zone header ctor.  This initializes all fields, locks, etc.
1498  *
1499  * Arguments/Returns follow uma_ctor specifications
1500  *	udata  Actually uma_zctor_args
1501  */
1502 static int
1503 zone_ctor(void *mem, int size, void *udata, int flags)
1504 {
1505 	struct uma_zctor_args *arg = udata;
1506 	uma_zone_t zone = mem;
1507 	uma_zone_t z;
1508 	uma_keg_t keg;
1509 
1510 	bzero(zone, size);
1511 	zone->uz_name = arg->name;
1512 	zone->uz_ctor = arg->ctor;
1513 	zone->uz_dtor = arg->dtor;
1514 	zone->uz_slab = zone_fetch_slab;
1515 	zone->uz_init = NULL;
1516 	zone->uz_fini = NULL;
1517 	zone->uz_allocs = 0;
1518 	zone->uz_frees = 0;
1519 	zone->uz_fails = 0;
1520 	zone->uz_sleeps = 0;
1521 	zone->uz_count = 0;
1522 	zone->uz_count_min = 0;
1523 	zone->uz_flags = 0;
1524 	zone->uz_warning = NULL;
1525 	timevalclear(&zone->uz_ratecheck);
1526 	keg = arg->keg;
1527 
1528 	ZONE_LOCK_INIT(zone, (arg->flags & UMA_ZONE_MTXCLASS));
1529 
1530 	/*
1531 	 * This is a pure cache zone, no kegs.
1532 	 */
1533 	if (arg->import) {
1534 		if (arg->flags & UMA_ZONE_VM)
1535 			arg->flags |= UMA_ZFLAG_CACHEONLY;
1536 		zone->uz_flags = arg->flags;
1537 		zone->uz_size = arg->size;
1538 		zone->uz_import = arg->import;
1539 		zone->uz_release = arg->release;
1540 		zone->uz_arg = arg->arg;
1541 		zone->uz_lockptr = &zone->uz_lock;
1542 		rw_wlock(&uma_rwlock);
1543 		LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link);
1544 		rw_wunlock(&uma_rwlock);
1545 		goto out;
1546 	}
1547 
1548 	/*
1549 	 * Use the regular zone/keg/slab allocator.
1550 	 */
1551 	zone->uz_import = (uma_import)zone_import;
1552 	zone->uz_release = (uma_release)zone_release;
1553 	zone->uz_arg = zone;
1554 
1555 	if (arg->flags & UMA_ZONE_SECONDARY) {
1556 		KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
1557 		zone->uz_init = arg->uminit;
1558 		zone->uz_fini = arg->fini;
1559 		zone->uz_lockptr = &keg->uk_lock;
1560 		zone->uz_flags |= UMA_ZONE_SECONDARY;
1561 		rw_wlock(&uma_rwlock);
1562 		ZONE_LOCK(zone);
1563 		LIST_FOREACH(z, &keg->uk_zones, uz_link) {
1564 			if (LIST_NEXT(z, uz_link) == NULL) {
1565 				LIST_INSERT_AFTER(z, zone, uz_link);
1566 				break;
1567 			}
1568 		}
1569 		ZONE_UNLOCK(zone);
1570 		rw_wunlock(&uma_rwlock);
1571 	} else if (keg == NULL) {
1572 		if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
1573 		    arg->align, arg->flags)) == NULL)
1574 			return (ENOMEM);
1575 	} else {
1576 		struct uma_kctor_args karg;
1577 		int error;
1578 
1579 		/* We should only be here from uma_startup() */
1580 		karg.size = arg->size;
1581 		karg.uminit = arg->uminit;
1582 		karg.fini = arg->fini;
1583 		karg.align = arg->align;
1584 		karg.flags = arg->flags;
1585 		karg.zone = zone;
1586 		error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
1587 		    flags);
1588 		if (error)
1589 			return (error);
1590 	}
1591 
1592 	/*
1593 	 * Link in the first keg.
1594 	 */
1595 	zone->uz_klink.kl_keg = keg;
1596 	LIST_INSERT_HEAD(&zone->uz_kegs, &zone->uz_klink, kl_link);
1597 	zone->uz_lockptr = &keg->uk_lock;
1598 	zone->uz_size = keg->uk_size;
1599 	zone->uz_flags |= (keg->uk_flags &
1600 	    (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
1601 
1602 	/*
1603 	 * Some internal zones don't have room allocated for the per cpu
1604 	 * caches.  If we're internal, bail out here.
1605 	 */
1606 	if (keg->uk_flags & UMA_ZFLAG_INTERNAL) {
1607 		KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
1608 		    ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
1609 		return (0);
1610 	}
1611 
1612 out:
1613 	if ((arg->flags & UMA_ZONE_MAXBUCKET) == 0)
1614 		zone->uz_count = bucket_select(zone->uz_size);
1615 	else
1616 		zone->uz_count = BUCKET_MAX;
1617 	zone->uz_count_min = zone->uz_count;
1618 
1619 	return (0);
1620 }
1621 
1622 /*
1623  * Keg header dtor.  This frees all data, destroys locks, frees the hash
1624  * table and removes the keg from the global list.
1625  *
1626  * Arguments/Returns follow uma_dtor specifications
1627  *	udata  unused
1628  */
1629 static void
1630 keg_dtor(void *arg, int size, void *udata)
1631 {
1632 	uma_keg_t keg;
1633 
1634 	keg = (uma_keg_t)arg;
1635 	KEG_LOCK(keg);
1636 	if (keg->uk_free != 0) {
1637 		printf("Freed UMA keg (%s) was not empty (%d items). "
1638 		    " Lost %d pages of memory.\n",
1639 		    keg->uk_name ? keg->uk_name : "",
1640 		    keg->uk_free, keg->uk_pages);
1641 	}
1642 	KEG_UNLOCK(keg);
1643 
1644 	hash_free(&keg->uk_hash);
1645 
1646 	KEG_LOCK_FINI(keg);
1647 }
1648 
1649 /*
1650  * Zone header dtor.
1651  *
1652  * Arguments/Returns follow uma_dtor specifications
1653  *	udata  unused
1654  */
1655 static void
1656 zone_dtor(void *arg, int size, void *udata)
1657 {
1658 	uma_klink_t klink;
1659 	uma_zone_t zone;
1660 	uma_keg_t keg;
1661 
1662 	zone = (uma_zone_t)arg;
1663 	keg = zone_first_keg(zone);
1664 
1665 	if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
1666 		cache_drain(zone);
1667 
1668 	rw_wlock(&uma_rwlock);
1669 	LIST_REMOVE(zone, uz_link);
1670 	rw_wunlock(&uma_rwlock);
1671 	/*
1672 	 * XXX there are some races here where
1673 	 * the zone can be drained but zone lock
1674 	 * released and then refilled before we
1675 	 * remove it... we dont care for now
1676 	 */
1677 	zone_drain_wait(zone, M_WAITOK);
1678 	/*
1679 	 * Unlink all of our kegs.
1680 	 */
1681 	while ((klink = LIST_FIRST(&zone->uz_kegs)) != NULL) {
1682 		klink->kl_keg = NULL;
1683 		LIST_REMOVE(klink, kl_link);
1684 		if (klink == &zone->uz_klink)
1685 			continue;
1686 		free(klink, M_TEMP);
1687 	}
1688 	/*
1689 	 * We only destroy kegs from non secondary zones.
1690 	 */
1691 	if (keg != NULL && (zone->uz_flags & UMA_ZONE_SECONDARY) == 0)  {
1692 		rw_wlock(&uma_rwlock);
1693 		LIST_REMOVE(keg, uk_link);
1694 		rw_wunlock(&uma_rwlock);
1695 		zone_free_item(kegs, keg, NULL, SKIP_NONE);
1696 	}
1697 	ZONE_LOCK_FINI(zone);
1698 }
1699 
1700 /*
1701  * Traverses every zone in the system and calls a callback
1702  *
1703  * Arguments:
1704  *	zfunc  A pointer to a function which accepts a zone
1705  *		as an argument.
1706  *
1707  * Returns:
1708  *	Nothing
1709  */
1710 static void
1711 zone_foreach(void (*zfunc)(uma_zone_t))
1712 {
1713 	uma_keg_t keg;
1714 	uma_zone_t zone;
1715 
1716 	rw_rlock(&uma_rwlock);
1717 	LIST_FOREACH(keg, &uma_kegs, uk_link) {
1718 		LIST_FOREACH(zone, &keg->uk_zones, uz_link)
1719 			zfunc(zone);
1720 	}
1721 	rw_runlock(&uma_rwlock);
1722 }
1723 
1724 /* Public functions */
1725 /* See uma.h */
1726 void
1727 uma_startup(void *mem, int npages)
1728 {
1729 	struct uma_zctor_args args;
1730 
1731 	rw_init(&uma_rwlock, "UMA lock");
1732 
1733 	/* "manually" create the initial zone */
1734 	memset(&args, 0, sizeof(args));
1735 	args.name = "UMA Kegs";
1736 	args.size = sizeof(struct uma_keg);
1737 	args.ctor = keg_ctor;
1738 	args.dtor = keg_dtor;
1739 	args.uminit = zero_init;
1740 	args.fini = NULL;
1741 	args.keg = &masterkeg;
1742 	args.align = 32 - 1;
1743 	args.flags = UMA_ZFLAG_INTERNAL;
1744 	/* The initial zone has no Per cpu queues so it's smaller */
1745 	zone_ctor(kegs, sizeof(struct uma_zone), &args, M_WAITOK);
1746 
1747 	mtx_init(&uma_boot_pages_mtx, "UMA boot pages", NULL, MTX_DEF);
1748 	bootmem = mem;
1749 	boot_pages = npages;
1750 
1751 	args.name = "UMA Zones";
1752 	args.size = sizeof(struct uma_zone) +
1753 	    (sizeof(struct uma_cache) * (mp_maxid + 1));
1754 	args.ctor = zone_ctor;
1755 	args.dtor = zone_dtor;
1756 	args.uminit = zero_init;
1757 	args.fini = NULL;
1758 	args.keg = NULL;
1759 	args.align = 32 - 1;
1760 	args.flags = UMA_ZFLAG_INTERNAL;
1761 	/* The initial zone has no Per cpu queues so it's smaller */
1762 	zone_ctor(zones, sizeof(struct uma_zone), &args, M_WAITOK);
1763 
1764 	/* Now make a zone for slab headers */
1765 	slabzone = uma_zcreate("UMA Slabs",
1766 				sizeof(struct uma_slab),
1767 				NULL, NULL, NULL, NULL,
1768 				UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1769 
1770 	hashzone = uma_zcreate("UMA Hash",
1771 	    sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
1772 	    NULL, NULL, NULL, NULL,
1773 	    UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1774 
1775 	bucket_init();
1776 
1777 	booted = UMA_STARTUP;
1778 }
1779 
1780 /* see uma.h */
1781 void
1782 uma_startup2(void)
1783 {
1784 	booted = UMA_STARTUP2;
1785 	bucket_enable();
1786 	sx_init(&uma_drain_lock, "umadrain");
1787 }
1788 
1789 /*
1790  * Initialize our callout handle
1791  *
1792  */
1793 
1794 static void
1795 uma_startup3(void)
1796 {
1797 
1798 	callout_init(&uma_callout, 1);
1799 	callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
1800 }
1801 
1802 static uma_keg_t
1803 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
1804 		int align, uint32_t flags)
1805 {
1806 	struct uma_kctor_args args;
1807 
1808 	args.size = size;
1809 	args.uminit = uminit;
1810 	args.fini = fini;
1811 	args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
1812 	args.flags = flags;
1813 	args.zone = zone;
1814 	return (zone_alloc_item(kegs, &args, M_WAITOK));
1815 }
1816 
1817 /* See uma.h */
1818 void
1819 uma_set_align(int align)
1820 {
1821 
1822 	if (align != UMA_ALIGN_CACHE)
1823 		uma_align_cache = align;
1824 }
1825 
1826 /* See uma.h */
1827 uma_zone_t
1828 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
1829 		uma_init uminit, uma_fini fini, int align, uint32_t flags)
1830 
1831 {
1832 	struct uma_zctor_args args;
1833 	uma_zone_t res;
1834 	bool locked;
1835 
1836 	KASSERT(powerof2(align + 1), ("invalid zone alignment %d for \"%s\"",
1837 	    align, name));
1838 
1839 	/* This stuff is essential for the zone ctor */
1840 	memset(&args, 0, sizeof(args));
1841 	args.name = name;
1842 	args.size = size;
1843 	args.ctor = ctor;
1844 	args.dtor = dtor;
1845 	args.uminit = uminit;
1846 	args.fini = fini;
1847 #ifdef  INVARIANTS
1848 	/*
1849 	 * If a zone is being created with an empty constructor and
1850 	 * destructor, pass UMA constructor/destructor which checks for
1851 	 * memory use after free.
1852 	 */
1853 	if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOFREE))) &&
1854 	    ctor == NULL && dtor == NULL && uminit == NULL && fini == NULL) {
1855 		args.ctor = trash_ctor;
1856 		args.dtor = trash_dtor;
1857 		args.uminit = trash_init;
1858 		args.fini = trash_fini;
1859 	}
1860 #endif
1861 	args.align = align;
1862 	args.flags = flags;
1863 	args.keg = NULL;
1864 
1865 	if (booted < UMA_STARTUP2) {
1866 		locked = false;
1867 	} else {
1868 		sx_slock(&uma_drain_lock);
1869 		locked = true;
1870 	}
1871 	res = zone_alloc_item(zones, &args, M_WAITOK);
1872 	if (locked)
1873 		sx_sunlock(&uma_drain_lock);
1874 	return (res);
1875 }
1876 
1877 /* See uma.h */
1878 uma_zone_t
1879 uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor,
1880 		    uma_init zinit, uma_fini zfini, uma_zone_t master)
1881 {
1882 	struct uma_zctor_args args;
1883 	uma_keg_t keg;
1884 	uma_zone_t res;
1885 	bool locked;
1886 
1887 	keg = zone_first_keg(master);
1888 	memset(&args, 0, sizeof(args));
1889 	args.name = name;
1890 	args.size = keg->uk_size;
1891 	args.ctor = ctor;
1892 	args.dtor = dtor;
1893 	args.uminit = zinit;
1894 	args.fini = zfini;
1895 	args.align = keg->uk_align;
1896 	args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
1897 	args.keg = keg;
1898 
1899 	if (booted < UMA_STARTUP2) {
1900 		locked = false;
1901 	} else {
1902 		sx_slock(&uma_drain_lock);
1903 		locked = true;
1904 	}
1905 	/* XXX Attaches only one keg of potentially many. */
1906 	res = zone_alloc_item(zones, &args, M_WAITOK);
1907 	if (locked)
1908 		sx_sunlock(&uma_drain_lock);
1909 	return (res);
1910 }
1911 
1912 /* See uma.h */
1913 uma_zone_t
1914 uma_zcache_create(char *name, int size, uma_ctor ctor, uma_dtor dtor,
1915 		    uma_init zinit, uma_fini zfini, uma_import zimport,
1916 		    uma_release zrelease, void *arg, int flags)
1917 {
1918 	struct uma_zctor_args args;
1919 
1920 	memset(&args, 0, sizeof(args));
1921 	args.name = name;
1922 	args.size = size;
1923 	args.ctor = ctor;
1924 	args.dtor = dtor;
1925 	args.uminit = zinit;
1926 	args.fini = zfini;
1927 	args.import = zimport;
1928 	args.release = zrelease;
1929 	args.arg = arg;
1930 	args.align = 0;
1931 	args.flags = flags;
1932 
1933 	return (zone_alloc_item(zones, &args, M_WAITOK));
1934 }
1935 
1936 static void
1937 zone_lock_pair(uma_zone_t a, uma_zone_t b)
1938 {
1939 	if (a < b) {
1940 		ZONE_LOCK(a);
1941 		mtx_lock_flags(b->uz_lockptr, MTX_DUPOK);
1942 	} else {
1943 		ZONE_LOCK(b);
1944 		mtx_lock_flags(a->uz_lockptr, MTX_DUPOK);
1945 	}
1946 }
1947 
1948 static void
1949 zone_unlock_pair(uma_zone_t a, uma_zone_t b)
1950 {
1951 
1952 	ZONE_UNLOCK(a);
1953 	ZONE_UNLOCK(b);
1954 }
1955 
1956 int
1957 uma_zsecond_add(uma_zone_t zone, uma_zone_t master)
1958 {
1959 	uma_klink_t klink;
1960 	uma_klink_t kl;
1961 	int error;
1962 
1963 	error = 0;
1964 	klink = malloc(sizeof(*klink), M_TEMP, M_WAITOK | M_ZERO);
1965 
1966 	zone_lock_pair(zone, master);
1967 	/*
1968 	 * zone must use vtoslab() to resolve objects and must already be
1969 	 * a secondary.
1970 	 */
1971 	if ((zone->uz_flags & (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY))
1972 	    != (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) {
1973 		error = EINVAL;
1974 		goto out;
1975 	}
1976 	/*
1977 	 * The new master must also use vtoslab().
1978 	 */
1979 	if ((zone->uz_flags & UMA_ZONE_VTOSLAB) != UMA_ZONE_VTOSLAB) {
1980 		error = EINVAL;
1981 		goto out;
1982 	}
1983 
1984 	/*
1985 	 * The underlying object must be the same size.  rsize
1986 	 * may be different.
1987 	 */
1988 	if (master->uz_size != zone->uz_size) {
1989 		error = E2BIG;
1990 		goto out;
1991 	}
1992 	/*
1993 	 * Put it at the end of the list.
1994 	 */
1995 	klink->kl_keg = zone_first_keg(master);
1996 	LIST_FOREACH(kl, &zone->uz_kegs, kl_link) {
1997 		if (LIST_NEXT(kl, kl_link) == NULL) {
1998 			LIST_INSERT_AFTER(kl, klink, kl_link);
1999 			break;
2000 		}
2001 	}
2002 	klink = NULL;
2003 	zone->uz_flags |= UMA_ZFLAG_MULTI;
2004 	zone->uz_slab = zone_fetch_slab_multi;
2005 
2006 out:
2007 	zone_unlock_pair(zone, master);
2008 	if (klink != NULL)
2009 		free(klink, M_TEMP);
2010 
2011 	return (error);
2012 }
2013 
2014 
2015 /* See uma.h */
2016 void
2017 uma_zdestroy(uma_zone_t zone)
2018 {
2019 
2020 	sx_slock(&uma_drain_lock);
2021 	zone_free_item(zones, zone, NULL, SKIP_NONE);
2022 	sx_sunlock(&uma_drain_lock);
2023 }
2024 
2025 void
2026 uma_zwait(uma_zone_t zone)
2027 {
2028 	void *item;
2029 
2030 	item = uma_zalloc_arg(zone, NULL, M_WAITOK);
2031 	uma_zfree(zone, item);
2032 }
2033 
2034 /* See uma.h */
2035 void *
2036 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
2037 {
2038 	void *item;
2039 	uma_cache_t cache;
2040 	uma_bucket_t bucket;
2041 	int lockfail;
2042 	int cpu;
2043 
2044 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
2045 	random_harvest_fast_uma(&zone, sizeof(zone), 1, RANDOM_UMA);
2046 
2047 	/* This is the fast path allocation */
2048 	CTR4(KTR_UMA, "uma_zalloc_arg thread %x zone %s(%p) flags %d",
2049 	    curthread, zone->uz_name, zone, flags);
2050 
2051 	if (flags & M_WAITOK) {
2052 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
2053 		    "uma_zalloc_arg: zone \"%s\"", zone->uz_name);
2054 	}
2055 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
2056 	    ("uma_zalloc_arg: called with spinlock or critical section held"));
2057 
2058 #ifdef DEBUG_MEMGUARD
2059 	if (memguard_cmp_zone(zone)) {
2060 		item = memguard_alloc(zone->uz_size, flags);
2061 		if (item != NULL) {
2062 			if (zone->uz_init != NULL &&
2063 			    zone->uz_init(item, zone->uz_size, flags) != 0)
2064 				return (NULL);
2065 			if (zone->uz_ctor != NULL &&
2066 			    zone->uz_ctor(item, zone->uz_size, udata,
2067 			    flags) != 0) {
2068 			    	zone->uz_fini(item, zone->uz_size);
2069 				return (NULL);
2070 			}
2071 			return (item);
2072 		}
2073 		/* This is unfortunate but should not be fatal. */
2074 	}
2075 #endif
2076 	/*
2077 	 * If possible, allocate from the per-CPU cache.  There are two
2078 	 * requirements for safe access to the per-CPU cache: (1) the thread
2079 	 * accessing the cache must not be preempted or yield during access,
2080 	 * and (2) the thread must not migrate CPUs without switching which
2081 	 * cache it accesses.  We rely on a critical section to prevent
2082 	 * preemption and migration.  We release the critical section in
2083 	 * order to acquire the zone mutex if we are unable to allocate from
2084 	 * the current cache; when we re-acquire the critical section, we
2085 	 * must detect and handle migration if it has occurred.
2086 	 */
2087 	critical_enter();
2088 	cpu = curcpu;
2089 	cache = &zone->uz_cpu[cpu];
2090 
2091 zalloc_start:
2092 	bucket = cache->uc_allocbucket;
2093 	if (bucket != NULL && bucket->ub_cnt > 0) {
2094 		bucket->ub_cnt--;
2095 		item = bucket->ub_bucket[bucket->ub_cnt];
2096 #ifdef INVARIANTS
2097 		bucket->ub_bucket[bucket->ub_cnt] = NULL;
2098 #endif
2099 		KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled."));
2100 		cache->uc_allocs++;
2101 		critical_exit();
2102 		if (zone->uz_ctor != NULL &&
2103 		    zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2104 			atomic_add_long(&zone->uz_fails, 1);
2105 			zone_free_item(zone, item, udata, SKIP_DTOR);
2106 			return (NULL);
2107 		}
2108 #ifdef INVARIANTS
2109 		uma_dbg_alloc(zone, NULL, item);
2110 #endif
2111 		if (flags & M_ZERO)
2112 			uma_zero_item(item, zone);
2113 		return (item);
2114 	}
2115 
2116 	/*
2117 	 * We have run out of items in our alloc bucket.
2118 	 * See if we can switch with our free bucket.
2119 	 */
2120 	bucket = cache->uc_freebucket;
2121 	if (bucket != NULL && bucket->ub_cnt > 0) {
2122 		CTR2(KTR_UMA,
2123 		    "uma_zalloc: zone %s(%p) swapping empty with alloc",
2124 		    zone->uz_name, zone);
2125 		cache->uc_freebucket = cache->uc_allocbucket;
2126 		cache->uc_allocbucket = bucket;
2127 		goto zalloc_start;
2128 	}
2129 
2130 	/*
2131 	 * Discard any empty allocation bucket while we hold no locks.
2132 	 */
2133 	bucket = cache->uc_allocbucket;
2134 	cache->uc_allocbucket = NULL;
2135 	critical_exit();
2136 	if (bucket != NULL)
2137 		bucket_free(zone, bucket, udata);
2138 
2139 	/* Short-circuit for zones without buckets and low memory. */
2140 	if (zone->uz_count == 0 || bucketdisable)
2141 		goto zalloc_item;
2142 
2143 	/*
2144 	 * Attempt to retrieve the item from the per-CPU cache has failed, so
2145 	 * we must go back to the zone.  This requires the zone lock, so we
2146 	 * must drop the critical section, then re-acquire it when we go back
2147 	 * to the cache.  Since the critical section is released, we may be
2148 	 * preempted or migrate.  As such, make sure not to maintain any
2149 	 * thread-local state specific to the cache from prior to releasing
2150 	 * the critical section.
2151 	 */
2152 	lockfail = 0;
2153 	if (ZONE_TRYLOCK(zone) == 0) {
2154 		/* Record contention to size the buckets. */
2155 		ZONE_LOCK(zone);
2156 		lockfail = 1;
2157 	}
2158 	critical_enter();
2159 	cpu = curcpu;
2160 	cache = &zone->uz_cpu[cpu];
2161 
2162 	/*
2163 	 * Since we have locked the zone we may as well send back our stats.
2164 	 */
2165 	atomic_add_long(&zone->uz_allocs, cache->uc_allocs);
2166 	atomic_add_long(&zone->uz_frees, cache->uc_frees);
2167 	cache->uc_allocs = 0;
2168 	cache->uc_frees = 0;
2169 
2170 	/* See if we lost the race to fill the cache. */
2171 	if (cache->uc_allocbucket != NULL) {
2172 		ZONE_UNLOCK(zone);
2173 		goto zalloc_start;
2174 	}
2175 
2176 	/*
2177 	 * Check the zone's cache of buckets.
2178 	 */
2179 	if ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) {
2180 		KASSERT(bucket->ub_cnt != 0,
2181 		    ("uma_zalloc_arg: Returning an empty bucket."));
2182 
2183 		LIST_REMOVE(bucket, ub_link);
2184 		cache->uc_allocbucket = bucket;
2185 		ZONE_UNLOCK(zone);
2186 		goto zalloc_start;
2187 	}
2188 	/* We are no longer associated with this CPU. */
2189 	critical_exit();
2190 
2191 	/*
2192 	 * We bump the uz count when the cache size is insufficient to
2193 	 * handle the working set.
2194 	 */
2195 	if (lockfail && zone->uz_count < BUCKET_MAX)
2196 		zone->uz_count++;
2197 	ZONE_UNLOCK(zone);
2198 
2199 	/*
2200 	 * Now lets just fill a bucket and put it on the free list.  If that
2201 	 * works we'll restart the allocation from the beginning and it
2202 	 * will use the just filled bucket.
2203 	 */
2204 	bucket = zone_alloc_bucket(zone, udata, flags);
2205 	CTR3(KTR_UMA, "uma_zalloc: zone %s(%p) bucket zone returned %p",
2206 	    zone->uz_name, zone, bucket);
2207 	if (bucket != NULL) {
2208 		ZONE_LOCK(zone);
2209 		critical_enter();
2210 		cpu = curcpu;
2211 		cache = &zone->uz_cpu[cpu];
2212 		/*
2213 		 * See if we lost the race or were migrated.  Cache the
2214 		 * initialized bucket to make this less likely or claim
2215 		 * the memory directly.
2216 		 */
2217 		if (cache->uc_allocbucket == NULL)
2218 			cache->uc_allocbucket = bucket;
2219 		else
2220 			LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link);
2221 		ZONE_UNLOCK(zone);
2222 		goto zalloc_start;
2223 	}
2224 
2225 	/*
2226 	 * We may not be able to get a bucket so return an actual item.
2227 	 */
2228 zalloc_item:
2229 	item = zone_alloc_item(zone, udata, flags);
2230 
2231 	return (item);
2232 }
2233 
2234 static uma_slab_t
2235 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int flags)
2236 {
2237 	uma_slab_t slab;
2238 	int reserve;
2239 
2240 	mtx_assert(&keg->uk_lock, MA_OWNED);
2241 	slab = NULL;
2242 	reserve = 0;
2243 	if ((flags & M_USE_RESERVE) == 0)
2244 		reserve = keg->uk_reserve;
2245 
2246 	for (;;) {
2247 		/*
2248 		 * Find a slab with some space.  Prefer slabs that are partially
2249 		 * used over those that are totally full.  This helps to reduce
2250 		 * fragmentation.
2251 		 */
2252 		if (keg->uk_free > reserve) {
2253 			if (!LIST_EMPTY(&keg->uk_part_slab)) {
2254 				slab = LIST_FIRST(&keg->uk_part_slab);
2255 			} else {
2256 				slab = LIST_FIRST(&keg->uk_free_slab);
2257 				LIST_REMOVE(slab, us_link);
2258 				LIST_INSERT_HEAD(&keg->uk_part_slab, slab,
2259 				    us_link);
2260 			}
2261 			MPASS(slab->us_keg == keg);
2262 			return (slab);
2263 		}
2264 
2265 		/*
2266 		 * M_NOVM means don't ask at all!
2267 		 */
2268 		if (flags & M_NOVM)
2269 			break;
2270 
2271 		if (keg->uk_maxpages && keg->uk_pages >= keg->uk_maxpages) {
2272 			keg->uk_flags |= UMA_ZFLAG_FULL;
2273 			/*
2274 			 * If this is not a multi-zone, set the FULL bit.
2275 			 * Otherwise slab_multi() takes care of it.
2276 			 */
2277 			if ((zone->uz_flags & UMA_ZFLAG_MULTI) == 0) {
2278 				zone->uz_flags |= UMA_ZFLAG_FULL;
2279 				zone_log_warning(zone);
2280 				zone_maxaction(zone);
2281 			}
2282 			if (flags & M_NOWAIT)
2283 				break;
2284 			zone->uz_sleeps++;
2285 			msleep(keg, &keg->uk_lock, PVM, "keglimit", 0);
2286 			continue;
2287 		}
2288 		slab = keg_alloc_slab(keg, zone, flags);
2289 		/*
2290 		 * If we got a slab here it's safe to mark it partially used
2291 		 * and return.  We assume that the caller is going to remove
2292 		 * at least one item.
2293 		 */
2294 		if (slab) {
2295 			MPASS(slab->us_keg == keg);
2296 			LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2297 			return (slab);
2298 		}
2299 		/*
2300 		 * We might not have been able to get a slab but another cpu
2301 		 * could have while we were unlocked.  Check again before we
2302 		 * fail.
2303 		 */
2304 		flags |= M_NOVM;
2305 	}
2306 	return (slab);
2307 }
2308 
2309 static uma_slab_t
2310 zone_fetch_slab(uma_zone_t zone, uma_keg_t keg, int flags)
2311 {
2312 	uma_slab_t slab;
2313 
2314 	if (keg == NULL) {
2315 		keg = zone_first_keg(zone);
2316 		KEG_LOCK(keg);
2317 	}
2318 
2319 	for (;;) {
2320 		slab = keg_fetch_slab(keg, zone, flags);
2321 		if (slab)
2322 			return (slab);
2323 		if (flags & (M_NOWAIT | M_NOVM))
2324 			break;
2325 	}
2326 	KEG_UNLOCK(keg);
2327 	return (NULL);
2328 }
2329 
2330 /*
2331  * uma_zone_fetch_slab_multi:  Fetches a slab from one available keg.  Returns
2332  * with the keg locked.  On NULL no lock is held.
2333  *
2334  * The last pointer is used to seed the search.  It is not required.
2335  */
2336 static uma_slab_t
2337 zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int rflags)
2338 {
2339 	uma_klink_t klink;
2340 	uma_slab_t slab;
2341 	uma_keg_t keg;
2342 	int flags;
2343 	int empty;
2344 	int full;
2345 
2346 	/*
2347 	 * Don't wait on the first pass.  This will skip limit tests
2348 	 * as well.  We don't want to block if we can find a provider
2349 	 * without blocking.
2350 	 */
2351 	flags = (rflags & ~M_WAITOK) | M_NOWAIT;
2352 	/*
2353 	 * Use the last slab allocated as a hint for where to start
2354 	 * the search.
2355 	 */
2356 	if (last != NULL) {
2357 		slab = keg_fetch_slab(last, zone, flags);
2358 		if (slab)
2359 			return (slab);
2360 		KEG_UNLOCK(last);
2361 	}
2362 	/*
2363 	 * Loop until we have a slab incase of transient failures
2364 	 * while M_WAITOK is specified.  I'm not sure this is 100%
2365 	 * required but we've done it for so long now.
2366 	 */
2367 	for (;;) {
2368 		empty = 0;
2369 		full = 0;
2370 		/*
2371 		 * Search the available kegs for slabs.  Be careful to hold the
2372 		 * correct lock while calling into the keg layer.
2373 		 */
2374 		LIST_FOREACH(klink, &zone->uz_kegs, kl_link) {
2375 			keg = klink->kl_keg;
2376 			KEG_LOCK(keg);
2377 			if ((keg->uk_flags & UMA_ZFLAG_FULL) == 0) {
2378 				slab = keg_fetch_slab(keg, zone, flags);
2379 				if (slab)
2380 					return (slab);
2381 			}
2382 			if (keg->uk_flags & UMA_ZFLAG_FULL)
2383 				full++;
2384 			else
2385 				empty++;
2386 			KEG_UNLOCK(keg);
2387 		}
2388 		if (rflags & (M_NOWAIT | M_NOVM))
2389 			break;
2390 		flags = rflags;
2391 		/*
2392 		 * All kegs are full.  XXX We can't atomically check all kegs
2393 		 * and sleep so just sleep for a short period and retry.
2394 		 */
2395 		if (full && !empty) {
2396 			ZONE_LOCK(zone);
2397 			zone->uz_flags |= UMA_ZFLAG_FULL;
2398 			zone->uz_sleeps++;
2399 			zone_log_warning(zone);
2400 			zone_maxaction(zone);
2401 			msleep(zone, zone->uz_lockptr, PVM,
2402 			    "zonelimit", hz/100);
2403 			zone->uz_flags &= ~UMA_ZFLAG_FULL;
2404 			ZONE_UNLOCK(zone);
2405 			continue;
2406 		}
2407 	}
2408 	return (NULL);
2409 }
2410 
2411 static void *
2412 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
2413 {
2414 	void *item;
2415 	uint8_t freei;
2416 
2417 	MPASS(keg == slab->us_keg);
2418 	mtx_assert(&keg->uk_lock, MA_OWNED);
2419 
2420 	freei = BIT_FFS(SLAB_SETSIZE, &slab->us_free) - 1;
2421 	BIT_CLR(SLAB_SETSIZE, freei, &slab->us_free);
2422 	item = slab->us_data + (keg->uk_rsize * freei);
2423 	slab->us_freecount--;
2424 	keg->uk_free--;
2425 
2426 	/* Move this slab to the full list */
2427 	if (slab->us_freecount == 0) {
2428 		LIST_REMOVE(slab, us_link);
2429 		LIST_INSERT_HEAD(&keg->uk_full_slab, slab, us_link);
2430 	}
2431 
2432 	return (item);
2433 }
2434 
2435 static int
2436 zone_import(uma_zone_t zone, void **bucket, int max, int flags)
2437 {
2438 	uma_slab_t slab;
2439 	uma_keg_t keg;
2440 	int i;
2441 
2442 	slab = NULL;
2443 	keg = NULL;
2444 	/* Try to keep the buckets totally full */
2445 	for (i = 0; i < max; ) {
2446 		if ((slab = zone->uz_slab(zone, keg, flags)) == NULL)
2447 			break;
2448 		keg = slab->us_keg;
2449 		while (slab->us_freecount && i < max) {
2450 			bucket[i++] = slab_alloc_item(keg, slab);
2451 			if (keg->uk_free <= keg->uk_reserve)
2452 				break;
2453 		}
2454 		/* Don't grab more than one slab at a time. */
2455 		flags &= ~M_WAITOK;
2456 		flags |= M_NOWAIT;
2457 	}
2458 	if (slab != NULL)
2459 		KEG_UNLOCK(keg);
2460 
2461 	return i;
2462 }
2463 
2464 static uma_bucket_t
2465 zone_alloc_bucket(uma_zone_t zone, void *udata, int flags)
2466 {
2467 	uma_bucket_t bucket;
2468 	int max;
2469 
2470 	/* Don't wait for buckets, preserve caller's NOVM setting. */
2471 	bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
2472 	if (bucket == NULL)
2473 		return (NULL);
2474 
2475 	max = MIN(bucket->ub_entries, zone->uz_count);
2476 	bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
2477 	    max, flags);
2478 
2479 	/*
2480 	 * Initialize the memory if necessary.
2481 	 */
2482 	if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
2483 		int i;
2484 
2485 		for (i = 0; i < bucket->ub_cnt; i++)
2486 			if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size,
2487 			    flags) != 0)
2488 				break;
2489 		/*
2490 		 * If we couldn't initialize the whole bucket, put the
2491 		 * rest back onto the freelist.
2492 		 */
2493 		if (i != bucket->ub_cnt) {
2494 			zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
2495 			    bucket->ub_cnt - i);
2496 #ifdef INVARIANTS
2497 			bzero(&bucket->ub_bucket[i],
2498 			    sizeof(void *) * (bucket->ub_cnt - i));
2499 #endif
2500 			bucket->ub_cnt = i;
2501 		}
2502 	}
2503 
2504 	if (bucket->ub_cnt == 0) {
2505 		bucket_free(zone, bucket, udata);
2506 		atomic_add_long(&zone->uz_fails, 1);
2507 		return (NULL);
2508 	}
2509 
2510 	return (bucket);
2511 }
2512 
2513 /*
2514  * Allocates a single item from a zone.
2515  *
2516  * Arguments
2517  *	zone   The zone to alloc for.
2518  *	udata  The data to be passed to the constructor.
2519  *	flags  M_WAITOK, M_NOWAIT, M_ZERO.
2520  *
2521  * Returns
2522  *	NULL if there is no memory and M_NOWAIT is set
2523  *	An item if successful
2524  */
2525 
2526 static void *
2527 zone_alloc_item(uma_zone_t zone, void *udata, int flags)
2528 {
2529 	void *item;
2530 
2531 	item = NULL;
2532 
2533 	if (zone->uz_import(zone->uz_arg, &item, 1, flags) != 1)
2534 		goto fail;
2535 	atomic_add_long(&zone->uz_allocs, 1);
2536 
2537 	/*
2538 	 * We have to call both the zone's init (not the keg's init)
2539 	 * and the zone's ctor.  This is because the item is going from
2540 	 * a keg slab directly to the user, and the user is expecting it
2541 	 * to be both zone-init'd as well as zone-ctor'd.
2542 	 */
2543 	if (zone->uz_init != NULL) {
2544 		if (zone->uz_init(item, zone->uz_size, flags) != 0) {
2545 			zone_free_item(zone, item, udata, SKIP_FINI);
2546 			goto fail;
2547 		}
2548 	}
2549 	if (zone->uz_ctor != NULL) {
2550 		if (zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2551 			zone_free_item(zone, item, udata, SKIP_DTOR);
2552 			goto fail;
2553 		}
2554 	}
2555 #ifdef INVARIANTS
2556 	uma_dbg_alloc(zone, NULL, item);
2557 #endif
2558 	if (flags & M_ZERO)
2559 		uma_zero_item(item, zone);
2560 
2561 	CTR3(KTR_UMA, "zone_alloc_item item %p from %s(%p)", item,
2562 	    zone->uz_name, zone);
2563 
2564 	return (item);
2565 
2566 fail:
2567 	CTR2(KTR_UMA, "zone_alloc_item failed from %s(%p)",
2568 	    zone->uz_name, zone);
2569 	atomic_add_long(&zone->uz_fails, 1);
2570 	return (NULL);
2571 }
2572 
2573 /* See uma.h */
2574 void
2575 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
2576 {
2577 	uma_cache_t cache;
2578 	uma_bucket_t bucket;
2579 	int lockfail;
2580 	int cpu;
2581 
2582 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
2583 	random_harvest_fast_uma(&zone, sizeof(zone), 1, RANDOM_UMA);
2584 
2585 	CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread,
2586 	    zone->uz_name);
2587 
2588 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
2589 	    ("uma_zfree_arg: called with spinlock or critical section held"));
2590 
2591         /* uma_zfree(..., NULL) does nothing, to match free(9). */
2592         if (item == NULL)
2593                 return;
2594 #ifdef DEBUG_MEMGUARD
2595 	if (is_memguard_addr(item)) {
2596 		if (zone->uz_dtor != NULL)
2597 			zone->uz_dtor(item, zone->uz_size, udata);
2598 		if (zone->uz_fini != NULL)
2599 			zone->uz_fini(item, zone->uz_size);
2600 		memguard_free(item);
2601 		return;
2602 	}
2603 #endif
2604 #ifdef INVARIANTS
2605 	if (zone->uz_flags & UMA_ZONE_MALLOC)
2606 		uma_dbg_free(zone, udata, item);
2607 	else
2608 		uma_dbg_free(zone, NULL, item);
2609 #endif
2610 	if (zone->uz_dtor != NULL)
2611 		zone->uz_dtor(item, zone->uz_size, udata);
2612 
2613 	/*
2614 	 * The race here is acceptable.  If we miss it we'll just have to wait
2615 	 * a little longer for the limits to be reset.
2616 	 */
2617 	if (zone->uz_flags & UMA_ZFLAG_FULL)
2618 		goto zfree_item;
2619 
2620 	/*
2621 	 * If possible, free to the per-CPU cache.  There are two
2622 	 * requirements for safe access to the per-CPU cache: (1) the thread
2623 	 * accessing the cache must not be preempted or yield during access,
2624 	 * and (2) the thread must not migrate CPUs without switching which
2625 	 * cache it accesses.  We rely on a critical section to prevent
2626 	 * preemption and migration.  We release the critical section in
2627 	 * order to acquire the zone mutex if we are unable to free to the
2628 	 * current cache; when we re-acquire the critical section, we must
2629 	 * detect and handle migration if it has occurred.
2630 	 */
2631 zfree_restart:
2632 	critical_enter();
2633 	cpu = curcpu;
2634 	cache = &zone->uz_cpu[cpu];
2635 
2636 zfree_start:
2637 	/*
2638 	 * Try to free into the allocbucket first to give LIFO ordering
2639 	 * for cache-hot datastructures.  Spill over into the freebucket
2640 	 * if necessary.  Alloc will swap them if one runs dry.
2641 	 */
2642 	bucket = cache->uc_allocbucket;
2643 	if (bucket == NULL || bucket->ub_cnt >= bucket->ub_entries)
2644 		bucket = cache->uc_freebucket;
2645 	if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) {
2646 		KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL,
2647 		    ("uma_zfree: Freeing to non free bucket index."));
2648 		bucket->ub_bucket[bucket->ub_cnt] = item;
2649 		bucket->ub_cnt++;
2650 		cache->uc_frees++;
2651 		critical_exit();
2652 		return;
2653 	}
2654 
2655 	/*
2656 	 * We must go back the zone, which requires acquiring the zone lock,
2657 	 * which in turn means we must release and re-acquire the critical
2658 	 * section.  Since the critical section is released, we may be
2659 	 * preempted or migrate.  As such, make sure not to maintain any
2660 	 * thread-local state specific to the cache from prior to releasing
2661 	 * the critical section.
2662 	 */
2663 	critical_exit();
2664 	if (zone->uz_count == 0 || bucketdisable)
2665 		goto zfree_item;
2666 
2667 	lockfail = 0;
2668 	if (ZONE_TRYLOCK(zone) == 0) {
2669 		/* Record contention to size the buckets. */
2670 		ZONE_LOCK(zone);
2671 		lockfail = 1;
2672 	}
2673 	critical_enter();
2674 	cpu = curcpu;
2675 	cache = &zone->uz_cpu[cpu];
2676 
2677 	/*
2678 	 * Since we have locked the zone we may as well send back our stats.
2679 	 */
2680 	atomic_add_long(&zone->uz_allocs, cache->uc_allocs);
2681 	atomic_add_long(&zone->uz_frees, cache->uc_frees);
2682 	cache->uc_allocs = 0;
2683 	cache->uc_frees = 0;
2684 
2685 	bucket = cache->uc_freebucket;
2686 	if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) {
2687 		ZONE_UNLOCK(zone);
2688 		goto zfree_start;
2689 	}
2690 	cache->uc_freebucket = NULL;
2691 	/* We are no longer associated with this CPU. */
2692 	critical_exit();
2693 
2694 	/* Can we throw this on the zone full list? */
2695 	if (bucket != NULL) {
2696 		CTR3(KTR_UMA,
2697 		    "uma_zfree: zone %s(%p) putting bucket %p on free list",
2698 		    zone->uz_name, zone, bucket);
2699 		/* ub_cnt is pointing to the last free item */
2700 		KASSERT(bucket->ub_cnt != 0,
2701 		    ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n"));
2702 		LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link);
2703 	}
2704 
2705 	/*
2706 	 * We bump the uz count when the cache size is insufficient to
2707 	 * handle the working set.
2708 	 */
2709 	if (lockfail && zone->uz_count < BUCKET_MAX)
2710 		zone->uz_count++;
2711 	ZONE_UNLOCK(zone);
2712 
2713 	bucket = bucket_alloc(zone, udata, M_NOWAIT);
2714 	CTR3(KTR_UMA, "uma_zfree: zone %s(%p) allocated bucket %p",
2715 	    zone->uz_name, zone, bucket);
2716 	if (bucket) {
2717 		critical_enter();
2718 		cpu = curcpu;
2719 		cache = &zone->uz_cpu[cpu];
2720 		if (cache->uc_freebucket == NULL) {
2721 			cache->uc_freebucket = bucket;
2722 			goto zfree_start;
2723 		}
2724 		/*
2725 		 * We lost the race, start over.  We have to drop our
2726 		 * critical section to free the bucket.
2727 		 */
2728 		critical_exit();
2729 		bucket_free(zone, bucket, udata);
2730 		goto zfree_restart;
2731 	}
2732 
2733 	/*
2734 	 * If nothing else caught this, we'll just do an internal free.
2735 	 */
2736 zfree_item:
2737 	zone_free_item(zone, item, udata, SKIP_DTOR);
2738 
2739 	return;
2740 }
2741 
2742 static void
2743 slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item)
2744 {
2745 	uint8_t freei;
2746 
2747 	mtx_assert(&keg->uk_lock, MA_OWNED);
2748 	MPASS(keg == slab->us_keg);
2749 
2750 	/* Do we need to remove from any lists? */
2751 	if (slab->us_freecount+1 == keg->uk_ipers) {
2752 		LIST_REMOVE(slab, us_link);
2753 		LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
2754 	} else if (slab->us_freecount == 0) {
2755 		LIST_REMOVE(slab, us_link);
2756 		LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2757 	}
2758 
2759 	/* Slab management. */
2760 	freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
2761 	BIT_SET(SLAB_SETSIZE, freei, &slab->us_free);
2762 	slab->us_freecount++;
2763 
2764 	/* Keg statistics. */
2765 	keg->uk_free++;
2766 }
2767 
2768 static void
2769 zone_release(uma_zone_t zone, void **bucket, int cnt)
2770 {
2771 	void *item;
2772 	uma_slab_t slab;
2773 	uma_keg_t keg;
2774 	uint8_t *mem;
2775 	int clearfull;
2776 	int i;
2777 
2778 	clearfull = 0;
2779 	keg = zone_first_keg(zone);
2780 	KEG_LOCK(keg);
2781 	for (i = 0; i < cnt; i++) {
2782 		item = bucket[i];
2783 		if (!(zone->uz_flags & UMA_ZONE_VTOSLAB)) {
2784 			mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
2785 			if (zone->uz_flags & UMA_ZONE_HASH) {
2786 				slab = hash_sfind(&keg->uk_hash, mem);
2787 			} else {
2788 				mem += keg->uk_pgoff;
2789 				slab = (uma_slab_t)mem;
2790 			}
2791 		} else {
2792 			slab = vtoslab((vm_offset_t)item);
2793 			if (slab->us_keg != keg) {
2794 				KEG_UNLOCK(keg);
2795 				keg = slab->us_keg;
2796 				KEG_LOCK(keg);
2797 			}
2798 		}
2799 		slab_free_item(keg, slab, item);
2800 		if (keg->uk_flags & UMA_ZFLAG_FULL) {
2801 			if (keg->uk_pages < keg->uk_maxpages) {
2802 				keg->uk_flags &= ~UMA_ZFLAG_FULL;
2803 				clearfull = 1;
2804 			}
2805 
2806 			/*
2807 			 * We can handle one more allocation. Since we're
2808 			 * clearing ZFLAG_FULL, wake up all procs blocked
2809 			 * on pages. This should be uncommon, so keeping this
2810 			 * simple for now (rather than adding count of blocked
2811 			 * threads etc).
2812 			 */
2813 			wakeup(keg);
2814 		}
2815 	}
2816 	KEG_UNLOCK(keg);
2817 	if (clearfull) {
2818 		ZONE_LOCK(zone);
2819 		zone->uz_flags &= ~UMA_ZFLAG_FULL;
2820 		wakeup(zone);
2821 		ZONE_UNLOCK(zone);
2822 	}
2823 
2824 }
2825 
2826 /*
2827  * Frees a single item to any zone.
2828  *
2829  * Arguments:
2830  *	zone   The zone to free to
2831  *	item   The item we're freeing
2832  *	udata  User supplied data for the dtor
2833  *	skip   Skip dtors and finis
2834  */
2835 static void
2836 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
2837 {
2838 
2839 #ifdef INVARIANTS
2840 	if (skip == SKIP_NONE) {
2841 		if (zone->uz_flags & UMA_ZONE_MALLOC)
2842 			uma_dbg_free(zone, udata, item);
2843 		else
2844 			uma_dbg_free(zone, NULL, item);
2845 	}
2846 #endif
2847 	if (skip < SKIP_DTOR && zone->uz_dtor)
2848 		zone->uz_dtor(item, zone->uz_size, udata);
2849 
2850 	if (skip < SKIP_FINI && zone->uz_fini)
2851 		zone->uz_fini(item, zone->uz_size);
2852 
2853 	atomic_add_long(&zone->uz_frees, 1);
2854 	zone->uz_release(zone->uz_arg, &item, 1);
2855 }
2856 
2857 /* See uma.h */
2858 int
2859 uma_zone_set_max(uma_zone_t zone, int nitems)
2860 {
2861 	uma_keg_t keg;
2862 
2863 	keg = zone_first_keg(zone);
2864 	if (keg == NULL)
2865 		return (0);
2866 	KEG_LOCK(keg);
2867 	keg->uk_maxpages = (nitems / keg->uk_ipers) * keg->uk_ppera;
2868 	if (keg->uk_maxpages * keg->uk_ipers < nitems)
2869 		keg->uk_maxpages += keg->uk_ppera;
2870 	nitems = (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers;
2871 	KEG_UNLOCK(keg);
2872 
2873 	return (nitems);
2874 }
2875 
2876 /* See uma.h */
2877 int
2878 uma_zone_get_max(uma_zone_t zone)
2879 {
2880 	int nitems;
2881 	uma_keg_t keg;
2882 
2883 	keg = zone_first_keg(zone);
2884 	if (keg == NULL)
2885 		return (0);
2886 	KEG_LOCK(keg);
2887 	nitems = (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers;
2888 	KEG_UNLOCK(keg);
2889 
2890 	return (nitems);
2891 }
2892 
2893 /* See uma.h */
2894 void
2895 uma_zone_set_warning(uma_zone_t zone, const char *warning)
2896 {
2897 
2898 	ZONE_LOCK(zone);
2899 	zone->uz_warning = warning;
2900 	ZONE_UNLOCK(zone);
2901 }
2902 
2903 /* See uma.h */
2904 void
2905 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction)
2906 {
2907 
2908 	ZONE_LOCK(zone);
2909 	TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone);
2910 	ZONE_UNLOCK(zone);
2911 }
2912 
2913 /* See uma.h */
2914 int
2915 uma_zone_get_cur(uma_zone_t zone)
2916 {
2917 	int64_t nitems;
2918 	u_int i;
2919 
2920 	ZONE_LOCK(zone);
2921 	nitems = zone->uz_allocs - zone->uz_frees;
2922 	CPU_FOREACH(i) {
2923 		/*
2924 		 * See the comment in sysctl_vm_zone_stats() regarding the
2925 		 * safety of accessing the per-cpu caches. With the zone lock
2926 		 * held, it is safe, but can potentially result in stale data.
2927 		 */
2928 		nitems += zone->uz_cpu[i].uc_allocs -
2929 		    zone->uz_cpu[i].uc_frees;
2930 	}
2931 	ZONE_UNLOCK(zone);
2932 
2933 	return (nitems < 0 ? 0 : nitems);
2934 }
2935 
2936 /* See uma.h */
2937 void
2938 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
2939 {
2940 	uma_keg_t keg;
2941 
2942 	keg = zone_first_keg(zone);
2943 	KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type"));
2944 	KEG_LOCK(keg);
2945 	KASSERT(keg->uk_pages == 0,
2946 	    ("uma_zone_set_init on non-empty keg"));
2947 	keg->uk_init = uminit;
2948 	KEG_UNLOCK(keg);
2949 }
2950 
2951 /* See uma.h */
2952 void
2953 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
2954 {
2955 	uma_keg_t keg;
2956 
2957 	keg = zone_first_keg(zone);
2958 	KASSERT(keg != NULL, ("uma_zone_set_fini: Invalid zone type"));
2959 	KEG_LOCK(keg);
2960 	KASSERT(keg->uk_pages == 0,
2961 	    ("uma_zone_set_fini on non-empty keg"));
2962 	keg->uk_fini = fini;
2963 	KEG_UNLOCK(keg);
2964 }
2965 
2966 /* See uma.h */
2967 void
2968 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
2969 {
2970 
2971 	ZONE_LOCK(zone);
2972 	KASSERT(zone_first_keg(zone)->uk_pages == 0,
2973 	    ("uma_zone_set_zinit on non-empty keg"));
2974 	zone->uz_init = zinit;
2975 	ZONE_UNLOCK(zone);
2976 }
2977 
2978 /* See uma.h */
2979 void
2980 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
2981 {
2982 
2983 	ZONE_LOCK(zone);
2984 	KASSERT(zone_first_keg(zone)->uk_pages == 0,
2985 	    ("uma_zone_set_zfini on non-empty keg"));
2986 	zone->uz_fini = zfini;
2987 	ZONE_UNLOCK(zone);
2988 }
2989 
2990 /* See uma.h */
2991 /* XXX uk_freef is not actually used with the zone locked */
2992 void
2993 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
2994 {
2995 	uma_keg_t keg;
2996 
2997 	keg = zone_first_keg(zone);
2998 	KASSERT(keg != NULL, ("uma_zone_set_freef: Invalid zone type"));
2999 	KEG_LOCK(keg);
3000 	keg->uk_freef = freef;
3001 	KEG_UNLOCK(keg);
3002 }
3003 
3004 /* See uma.h */
3005 /* XXX uk_allocf is not actually used with the zone locked */
3006 void
3007 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
3008 {
3009 	uma_keg_t keg;
3010 
3011 	keg = zone_first_keg(zone);
3012 	KEG_LOCK(keg);
3013 	keg->uk_allocf = allocf;
3014 	KEG_UNLOCK(keg);
3015 }
3016 
3017 /* See uma.h */
3018 void
3019 uma_zone_reserve(uma_zone_t zone, int items)
3020 {
3021 	uma_keg_t keg;
3022 
3023 	keg = zone_first_keg(zone);
3024 	if (keg == NULL)
3025 		return;
3026 	KEG_LOCK(keg);
3027 	keg->uk_reserve = items;
3028 	KEG_UNLOCK(keg);
3029 
3030 	return;
3031 }
3032 
3033 /* See uma.h */
3034 int
3035 uma_zone_reserve_kva(uma_zone_t zone, int count)
3036 {
3037 	uma_keg_t keg;
3038 	vm_offset_t kva;
3039 	u_int pages;
3040 
3041 	keg = zone_first_keg(zone);
3042 	if (keg == NULL)
3043 		return (0);
3044 	pages = count / keg->uk_ipers;
3045 
3046 	if (pages * keg->uk_ipers < count)
3047 		pages++;
3048 	pages *= keg->uk_ppera;
3049 
3050 #ifdef UMA_MD_SMALL_ALLOC
3051 	if (keg->uk_ppera > 1) {
3052 #else
3053 	if (1) {
3054 #endif
3055 		kva = kva_alloc((vm_size_t)pages * PAGE_SIZE);
3056 		if (kva == 0)
3057 			return (0);
3058 	} else
3059 		kva = 0;
3060 	KEG_LOCK(keg);
3061 	keg->uk_kva = kva;
3062 	keg->uk_offset = 0;
3063 	keg->uk_maxpages = pages;
3064 #ifdef UMA_MD_SMALL_ALLOC
3065 	keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
3066 #else
3067 	keg->uk_allocf = noobj_alloc;
3068 #endif
3069 	keg->uk_flags |= UMA_ZONE_NOFREE;
3070 	KEG_UNLOCK(keg);
3071 
3072 	return (1);
3073 }
3074 
3075 /* See uma.h */
3076 void
3077 uma_prealloc(uma_zone_t zone, int items)
3078 {
3079 	int slabs;
3080 	uma_slab_t slab;
3081 	uma_keg_t keg;
3082 
3083 	keg = zone_first_keg(zone);
3084 	if (keg == NULL)
3085 		return;
3086 	KEG_LOCK(keg);
3087 	slabs = items / keg->uk_ipers;
3088 	if (slabs * keg->uk_ipers < items)
3089 		slabs++;
3090 	while (slabs > 0) {
3091 		slab = keg_alloc_slab(keg, zone, M_WAITOK);
3092 		if (slab == NULL)
3093 			break;
3094 		MPASS(slab->us_keg == keg);
3095 		LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
3096 		slabs--;
3097 	}
3098 	KEG_UNLOCK(keg);
3099 }
3100 
3101 /* See uma.h */
3102 static void
3103 uma_reclaim_locked(bool kmem_danger)
3104 {
3105 
3106 	CTR0(KTR_UMA, "UMA: vm asked us to release pages!");
3107 	sx_assert(&uma_drain_lock, SA_XLOCKED);
3108 	bucket_enable();
3109 	zone_foreach(zone_drain);
3110 	if (vm_page_count_min() || kmem_danger) {
3111 		cache_drain_safe(NULL);
3112 		zone_foreach(zone_drain);
3113 	}
3114 	/*
3115 	 * Some slabs may have been freed but this zone will be visited early
3116 	 * we visit again so that we can free pages that are empty once other
3117 	 * zones are drained.  We have to do the same for buckets.
3118 	 */
3119 	zone_drain(slabzone);
3120 	bucket_zone_drain();
3121 }
3122 
3123 void
3124 uma_reclaim(void)
3125 {
3126 
3127 	sx_xlock(&uma_drain_lock);
3128 	uma_reclaim_locked(false);
3129 	sx_xunlock(&uma_drain_lock);
3130 }
3131 
3132 static int uma_reclaim_needed;
3133 
3134 void
3135 uma_reclaim_wakeup(void)
3136 {
3137 
3138 	uma_reclaim_needed = 1;
3139 	wakeup(&uma_reclaim_needed);
3140 }
3141 
3142 void
3143 uma_reclaim_worker(void *arg __unused)
3144 {
3145 
3146 	sx_xlock(&uma_drain_lock);
3147 	for (;;) {
3148 		sx_sleep(&uma_reclaim_needed, &uma_drain_lock, PVM,
3149 		    "umarcl", 0);
3150 		if (uma_reclaim_needed) {
3151 			uma_reclaim_needed = 0;
3152 			sx_xunlock(&uma_drain_lock);
3153 			EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM);
3154 			sx_xlock(&uma_drain_lock);
3155 			uma_reclaim_locked(true);
3156 		}
3157 	}
3158 }
3159 
3160 /* See uma.h */
3161 int
3162 uma_zone_exhausted(uma_zone_t zone)
3163 {
3164 	int full;
3165 
3166 	ZONE_LOCK(zone);
3167 	full = (zone->uz_flags & UMA_ZFLAG_FULL);
3168 	ZONE_UNLOCK(zone);
3169 	return (full);
3170 }
3171 
3172 int
3173 uma_zone_exhausted_nolock(uma_zone_t zone)
3174 {
3175 	return (zone->uz_flags & UMA_ZFLAG_FULL);
3176 }
3177 
3178 void *
3179 uma_large_malloc(vm_size_t size, int wait)
3180 {
3181 	void *mem;
3182 	uma_slab_t slab;
3183 	uint8_t flags;
3184 
3185 	slab = zone_alloc_item(slabzone, NULL, wait);
3186 	if (slab == NULL)
3187 		return (NULL);
3188 	mem = page_alloc(NULL, size, &flags, wait);
3189 	if (mem) {
3190 		vsetslab((vm_offset_t)mem, slab);
3191 		slab->us_data = mem;
3192 		slab->us_flags = flags | UMA_SLAB_MALLOC;
3193 		slab->us_size = size;
3194 	} else {
3195 		zone_free_item(slabzone, slab, NULL, SKIP_NONE);
3196 	}
3197 
3198 	return (mem);
3199 }
3200 
3201 void
3202 uma_large_free(uma_slab_t slab)
3203 {
3204 
3205 	page_free(slab->us_data, slab->us_size, slab->us_flags);
3206 	zone_free_item(slabzone, slab, NULL, SKIP_NONE);
3207 }
3208 
3209 static void
3210 uma_zero_item(void *item, uma_zone_t zone)
3211 {
3212 	int i;
3213 
3214 	if (zone->uz_flags & UMA_ZONE_PCPU) {
3215 		CPU_FOREACH(i)
3216 			bzero(zpcpu_get_cpu(item, i), zone->uz_size);
3217 	} else
3218 		bzero(item, zone->uz_size);
3219 }
3220 
3221 void
3222 uma_print_stats(void)
3223 {
3224 	zone_foreach(uma_print_zone);
3225 }
3226 
3227 static void
3228 slab_print(uma_slab_t slab)
3229 {
3230 	printf("slab: keg %p, data %p, freecount %d\n",
3231 		slab->us_keg, slab->us_data, slab->us_freecount);
3232 }
3233 
3234 static void
3235 cache_print(uma_cache_t cache)
3236 {
3237 	printf("alloc: %p(%d), free: %p(%d)\n",
3238 		cache->uc_allocbucket,
3239 		cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0,
3240 		cache->uc_freebucket,
3241 		cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0);
3242 }
3243 
3244 static void
3245 uma_print_keg(uma_keg_t keg)
3246 {
3247 	uma_slab_t slab;
3248 
3249 	printf("keg: %s(%p) size %d(%d) flags %#x ipers %d ppera %d "
3250 	    "out %d free %d limit %d\n",
3251 	    keg->uk_name, keg, keg->uk_size, keg->uk_rsize, keg->uk_flags,
3252 	    keg->uk_ipers, keg->uk_ppera,
3253 	    (keg->uk_pages / keg->uk_ppera) * keg->uk_ipers - keg->uk_free,
3254 	    keg->uk_free, (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers);
3255 	printf("Part slabs:\n");
3256 	LIST_FOREACH(slab, &keg->uk_part_slab, us_link)
3257 		slab_print(slab);
3258 	printf("Free slabs:\n");
3259 	LIST_FOREACH(slab, &keg->uk_free_slab, us_link)
3260 		slab_print(slab);
3261 	printf("Full slabs:\n");
3262 	LIST_FOREACH(slab, &keg->uk_full_slab, us_link)
3263 		slab_print(slab);
3264 }
3265 
3266 void
3267 uma_print_zone(uma_zone_t zone)
3268 {
3269 	uma_cache_t cache;
3270 	uma_klink_t kl;
3271 	int i;
3272 
3273 	printf("zone: %s(%p) size %d flags %#x\n",
3274 	    zone->uz_name, zone, zone->uz_size, zone->uz_flags);
3275 	LIST_FOREACH(kl, &zone->uz_kegs, kl_link)
3276 		uma_print_keg(kl->kl_keg);
3277 	CPU_FOREACH(i) {
3278 		cache = &zone->uz_cpu[i];
3279 		printf("CPU %d Cache:\n", i);
3280 		cache_print(cache);
3281 	}
3282 }
3283 
3284 #ifdef DDB
3285 /*
3286  * Generate statistics across both the zone and its per-cpu cache's.  Return
3287  * desired statistics if the pointer is non-NULL for that statistic.
3288  *
3289  * Note: does not update the zone statistics, as it can't safely clear the
3290  * per-CPU cache statistic.
3291  *
3292  * XXXRW: Following the uc_allocbucket and uc_freebucket pointers here isn't
3293  * safe from off-CPU; we should modify the caches to track this information
3294  * directly so that we don't have to.
3295  */
3296 static void
3297 uma_zone_sumstat(uma_zone_t z, int *cachefreep, uint64_t *allocsp,
3298     uint64_t *freesp, uint64_t *sleepsp)
3299 {
3300 	uma_cache_t cache;
3301 	uint64_t allocs, frees, sleeps;
3302 	int cachefree, cpu;
3303 
3304 	allocs = frees = sleeps = 0;
3305 	cachefree = 0;
3306 	CPU_FOREACH(cpu) {
3307 		cache = &z->uz_cpu[cpu];
3308 		if (cache->uc_allocbucket != NULL)
3309 			cachefree += cache->uc_allocbucket->ub_cnt;
3310 		if (cache->uc_freebucket != NULL)
3311 			cachefree += cache->uc_freebucket->ub_cnt;
3312 		allocs += cache->uc_allocs;
3313 		frees += cache->uc_frees;
3314 	}
3315 	allocs += z->uz_allocs;
3316 	frees += z->uz_frees;
3317 	sleeps += z->uz_sleeps;
3318 	if (cachefreep != NULL)
3319 		*cachefreep = cachefree;
3320 	if (allocsp != NULL)
3321 		*allocsp = allocs;
3322 	if (freesp != NULL)
3323 		*freesp = frees;
3324 	if (sleepsp != NULL)
3325 		*sleepsp = sleeps;
3326 }
3327 #endif /* DDB */
3328 
3329 static int
3330 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
3331 {
3332 	uma_keg_t kz;
3333 	uma_zone_t z;
3334 	int count;
3335 
3336 	count = 0;
3337 	rw_rlock(&uma_rwlock);
3338 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3339 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
3340 			count++;
3341 	}
3342 	rw_runlock(&uma_rwlock);
3343 	return (sysctl_handle_int(oidp, &count, 0, req));
3344 }
3345 
3346 static int
3347 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
3348 {
3349 	struct uma_stream_header ush;
3350 	struct uma_type_header uth;
3351 	struct uma_percpu_stat ups;
3352 	uma_bucket_t bucket;
3353 	struct sbuf sbuf;
3354 	uma_cache_t cache;
3355 	uma_klink_t kl;
3356 	uma_keg_t kz;
3357 	uma_zone_t z;
3358 	uma_keg_t k;
3359 	int count, error, i;
3360 
3361 	error = sysctl_wire_old_buffer(req, 0);
3362 	if (error != 0)
3363 		return (error);
3364 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
3365 	sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL);
3366 
3367 	count = 0;
3368 	rw_rlock(&uma_rwlock);
3369 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3370 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
3371 			count++;
3372 	}
3373 
3374 	/*
3375 	 * Insert stream header.
3376 	 */
3377 	bzero(&ush, sizeof(ush));
3378 	ush.ush_version = UMA_STREAM_VERSION;
3379 	ush.ush_maxcpus = (mp_maxid + 1);
3380 	ush.ush_count = count;
3381 	(void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
3382 
3383 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3384 		LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3385 			bzero(&uth, sizeof(uth));
3386 			ZONE_LOCK(z);
3387 			strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
3388 			uth.uth_align = kz->uk_align;
3389 			uth.uth_size = kz->uk_size;
3390 			uth.uth_rsize = kz->uk_rsize;
3391 			LIST_FOREACH(kl, &z->uz_kegs, kl_link) {
3392 				k = kl->kl_keg;
3393 				uth.uth_maxpages += k->uk_maxpages;
3394 				uth.uth_pages += k->uk_pages;
3395 				uth.uth_keg_free += k->uk_free;
3396 				uth.uth_limit = (k->uk_maxpages / k->uk_ppera)
3397 				    * k->uk_ipers;
3398 			}
3399 
3400 			/*
3401 			 * A zone is secondary is it is not the first entry
3402 			 * on the keg's zone list.
3403 			 */
3404 			if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
3405 			    (LIST_FIRST(&kz->uk_zones) != z))
3406 				uth.uth_zone_flags = UTH_ZONE_SECONDARY;
3407 
3408 			LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3409 				uth.uth_zone_free += bucket->ub_cnt;
3410 			uth.uth_allocs = z->uz_allocs;
3411 			uth.uth_frees = z->uz_frees;
3412 			uth.uth_fails = z->uz_fails;
3413 			uth.uth_sleeps = z->uz_sleeps;
3414 			(void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
3415 			/*
3416 			 * While it is not normally safe to access the cache
3417 			 * bucket pointers while not on the CPU that owns the
3418 			 * cache, we only allow the pointers to be exchanged
3419 			 * without the zone lock held, not invalidated, so
3420 			 * accept the possible race associated with bucket
3421 			 * exchange during monitoring.
3422 			 */
3423 			for (i = 0; i < (mp_maxid + 1); i++) {
3424 				bzero(&ups, sizeof(ups));
3425 				if (kz->uk_flags & UMA_ZFLAG_INTERNAL)
3426 					goto skip;
3427 				if (CPU_ABSENT(i))
3428 					goto skip;
3429 				cache = &z->uz_cpu[i];
3430 				if (cache->uc_allocbucket != NULL)
3431 					ups.ups_cache_free +=
3432 					    cache->uc_allocbucket->ub_cnt;
3433 				if (cache->uc_freebucket != NULL)
3434 					ups.ups_cache_free +=
3435 					    cache->uc_freebucket->ub_cnt;
3436 				ups.ups_allocs = cache->uc_allocs;
3437 				ups.ups_frees = cache->uc_frees;
3438 skip:
3439 				(void)sbuf_bcat(&sbuf, &ups, sizeof(ups));
3440 			}
3441 			ZONE_UNLOCK(z);
3442 		}
3443 	}
3444 	rw_runlock(&uma_rwlock);
3445 	error = sbuf_finish(&sbuf);
3446 	sbuf_delete(&sbuf);
3447 	return (error);
3448 }
3449 
3450 int
3451 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS)
3452 {
3453 	uma_zone_t zone = *(uma_zone_t *)arg1;
3454 	int error, max;
3455 
3456 	max = uma_zone_get_max(zone);
3457 	error = sysctl_handle_int(oidp, &max, 0, req);
3458 	if (error || !req->newptr)
3459 		return (error);
3460 
3461 	uma_zone_set_max(zone, max);
3462 
3463 	return (0);
3464 }
3465 
3466 int
3467 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS)
3468 {
3469 	uma_zone_t zone = *(uma_zone_t *)arg1;
3470 	int cur;
3471 
3472 	cur = uma_zone_get_cur(zone);
3473 	return (sysctl_handle_int(oidp, &cur, 0, req));
3474 }
3475 
3476 #ifdef INVARIANTS
3477 static uma_slab_t
3478 uma_dbg_getslab(uma_zone_t zone, void *item)
3479 {
3480 	uma_slab_t slab;
3481 	uma_keg_t keg;
3482 	uint8_t *mem;
3483 
3484 	mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
3485 	if (zone->uz_flags & UMA_ZONE_VTOSLAB) {
3486 		slab = vtoslab((vm_offset_t)mem);
3487 	} else {
3488 		/*
3489 		 * It is safe to return the slab here even though the
3490 		 * zone is unlocked because the item's allocation state
3491 		 * essentially holds a reference.
3492 		 */
3493 		ZONE_LOCK(zone);
3494 		keg = LIST_FIRST(&zone->uz_kegs)->kl_keg;
3495 		if (keg->uk_flags & UMA_ZONE_HASH)
3496 			slab = hash_sfind(&keg->uk_hash, mem);
3497 		else
3498 			slab = (uma_slab_t)(mem + keg->uk_pgoff);
3499 		ZONE_UNLOCK(zone);
3500 	}
3501 
3502 	return (slab);
3503 }
3504 
3505 /*
3506  * Set up the slab's freei data such that uma_dbg_free can function.
3507  *
3508  */
3509 static void
3510 uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item)
3511 {
3512 	uma_keg_t keg;
3513 	int freei;
3514 
3515 	if (zone_first_keg(zone) == NULL)
3516 		return;
3517 	if (slab == NULL) {
3518 		slab = uma_dbg_getslab(zone, item);
3519 		if (slab == NULL)
3520 			panic("uma: item %p did not belong to zone %s\n",
3521 			    item, zone->uz_name);
3522 	}
3523 	keg = slab->us_keg;
3524 	freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
3525 
3526 	if (BIT_ISSET(SLAB_SETSIZE, freei, &slab->us_debugfree))
3527 		panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)\n",
3528 		    item, zone, zone->uz_name, slab, freei);
3529 	BIT_SET_ATOMIC(SLAB_SETSIZE, freei, &slab->us_debugfree);
3530 
3531 	return;
3532 }
3533 
3534 /*
3535  * Verifies freed addresses.  Checks for alignment, valid slab membership
3536  * and duplicate frees.
3537  *
3538  */
3539 static void
3540 uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item)
3541 {
3542 	uma_keg_t keg;
3543 	int freei;
3544 
3545 	if (zone_first_keg(zone) == NULL)
3546 		return;
3547 	if (slab == NULL) {
3548 		slab = uma_dbg_getslab(zone, item);
3549 		if (slab == NULL)
3550 			panic("uma: Freed item %p did not belong to zone %s\n",
3551 			    item, zone->uz_name);
3552 	}
3553 	keg = slab->us_keg;
3554 	freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
3555 
3556 	if (freei >= keg->uk_ipers)
3557 		panic("Invalid free of %p from zone %p(%s) slab %p(%d)\n",
3558 		    item, zone, zone->uz_name, slab, freei);
3559 
3560 	if (((freei * keg->uk_rsize) + slab->us_data) != item)
3561 		panic("Unaligned free of %p from zone %p(%s) slab %p(%d)\n",
3562 		    item, zone, zone->uz_name, slab, freei);
3563 
3564 	if (!BIT_ISSET(SLAB_SETSIZE, freei, &slab->us_debugfree))
3565 		panic("Duplicate free of %p from zone %p(%s) slab %p(%d)\n",
3566 		    item, zone, zone->uz_name, slab, freei);
3567 
3568 	BIT_CLR_ATOMIC(SLAB_SETSIZE, freei, &slab->us_debugfree);
3569 }
3570 #endif /* INVARIANTS */
3571 
3572 #ifdef DDB
3573 DB_SHOW_COMMAND(uma, db_show_uma)
3574 {
3575 	uint64_t allocs, frees, sleeps;
3576 	uma_bucket_t bucket;
3577 	uma_keg_t kz;
3578 	uma_zone_t z;
3579 	int cachefree;
3580 
3581 	db_printf("%18s %8s %8s %8s %12s %8s %8s\n", "Zone", "Size", "Used",
3582 	    "Free", "Requests", "Sleeps", "Bucket");
3583 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3584 		LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3585 			if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
3586 				allocs = z->uz_allocs;
3587 				frees = z->uz_frees;
3588 				sleeps = z->uz_sleeps;
3589 				cachefree = 0;
3590 			} else
3591 				uma_zone_sumstat(z, &cachefree, &allocs,
3592 				    &frees, &sleeps);
3593 			if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
3594 			    (LIST_FIRST(&kz->uk_zones) != z)))
3595 				cachefree += kz->uk_free;
3596 			LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3597 				cachefree += bucket->ub_cnt;
3598 			db_printf("%18s %8ju %8jd %8d %12ju %8ju %8u\n",
3599 			    z->uz_name, (uintmax_t)kz->uk_size,
3600 			    (intmax_t)(allocs - frees), cachefree,
3601 			    (uintmax_t)allocs, sleeps, z->uz_count);
3602 			if (db_pager_quit)
3603 				return;
3604 		}
3605 	}
3606 }
3607 
3608 DB_SHOW_COMMAND(umacache, db_show_umacache)
3609 {
3610 	uint64_t allocs, frees;
3611 	uma_bucket_t bucket;
3612 	uma_zone_t z;
3613 	int cachefree;
3614 
3615 	db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
3616 	    "Requests", "Bucket");
3617 	LIST_FOREACH(z, &uma_cachezones, uz_link) {
3618 		uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL);
3619 		LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3620 			cachefree += bucket->ub_cnt;
3621 		db_printf("%18s %8ju %8jd %8d %12ju %8u\n",
3622 		    z->uz_name, (uintmax_t)z->uz_size,
3623 		    (intmax_t)(allocs - frees), cachefree,
3624 		    (uintmax_t)allocs, z->uz_count);
3625 		if (db_pager_quit)
3626 			return;
3627 	}
3628 }
3629 #endif	/* DDB */
3630