xref: /freebsd/sys/vm/uma_core.c (revision ee7b0571c2c18bdec848ed2044223cc88db29bd8)
1 /*-
2  * Copyright (c) 2002-2005, 2009, 2013 Jeffrey Roberson <jeff@FreeBSD.org>
3  * Copyright (c) 2004, 2005 Bosko Milekic <bmilekic@FreeBSD.org>
4  * Copyright (c) 2004-2006 Robert N. M. Watson
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice unmodified, this list of conditions, and the following
12  *    disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
18  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
19  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
20  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
21  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
22  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
23  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
24  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
25  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
26  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
27  */
28 
29 /*
30  * uma_core.c  Implementation of the Universal Memory allocator
31  *
32  * This allocator is intended to replace the multitude of similar object caches
33  * in the standard FreeBSD kernel.  The intent is to be flexible as well as
34  * effecient.  A primary design goal is to return unused memory to the rest of
35  * the system.  This will make the system as a whole more flexible due to the
36  * ability to move memory to subsystems which most need it instead of leaving
37  * pools of reserved memory unused.
38  *
39  * The basic ideas stem from similar slab/zone based allocators whose algorithms
40  * are well known.
41  *
42  */
43 
44 /*
45  * TODO:
46  *	- Improve memory usage for large allocations
47  *	- Investigate cache size adjustments
48  */
49 
50 #include <sys/cdefs.h>
51 __FBSDID("$FreeBSD$");
52 
53 /* I should really use ktr.. */
54 /*
55 #define UMA_DEBUG 1
56 #define UMA_DEBUG_ALLOC 1
57 #define UMA_DEBUG_ALLOC_1 1
58 */
59 
60 #include "opt_ddb.h"
61 #include "opt_param.h"
62 #include "opt_vm.h"
63 
64 #include <sys/param.h>
65 #include <sys/systm.h>
66 #include <sys/bitset.h>
67 #include <sys/kernel.h>
68 #include <sys/types.h>
69 #include <sys/queue.h>
70 #include <sys/malloc.h>
71 #include <sys/ktr.h>
72 #include <sys/lock.h>
73 #include <sys/sysctl.h>
74 #include <sys/mutex.h>
75 #include <sys/proc.h>
76 #include <sys/rwlock.h>
77 #include <sys/sbuf.h>
78 #include <sys/sched.h>
79 #include <sys/smp.h>
80 #include <sys/vmmeter.h>
81 
82 #include <vm/vm.h>
83 #include <vm/vm_object.h>
84 #include <vm/vm_page.h>
85 #include <vm/vm_pageout.h>
86 #include <vm/vm_param.h>
87 #include <vm/vm_map.h>
88 #include <vm/vm_kern.h>
89 #include <vm/vm_extern.h>
90 #include <vm/uma.h>
91 #include <vm/uma_int.h>
92 #include <vm/uma_dbg.h>
93 
94 #include <ddb/ddb.h>
95 
96 #ifdef DEBUG_MEMGUARD
97 #include <vm/memguard.h>
98 #endif
99 
100 /*
101  * This is the zone and keg from which all zones are spawned.  The idea is that
102  * even the zone & keg heads are allocated from the allocator, so we use the
103  * bss section to bootstrap us.
104  */
105 static struct uma_keg masterkeg;
106 static struct uma_zone masterzone_k;
107 static struct uma_zone masterzone_z;
108 static uma_zone_t kegs = &masterzone_k;
109 static uma_zone_t zones = &masterzone_z;
110 
111 /* This is the zone from which all of uma_slab_t's are allocated. */
112 static uma_zone_t slabzone;
113 static uma_zone_t slabrefzone;	/* With refcounters (for UMA_ZONE_REFCNT) */
114 
115 /*
116  * The initial hash tables come out of this zone so they can be allocated
117  * prior to malloc coming up.
118  */
119 static uma_zone_t hashzone;
120 
121 /* The boot-time adjusted value for cache line alignment. */
122 int uma_align_cache = 64 - 1;
123 
124 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets");
125 
126 /*
127  * Are we allowed to allocate buckets?
128  */
129 static int bucketdisable = 1;
130 
131 /* Linked list of all kegs in the system */
132 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs);
133 
134 /* Linked list of all cache-only zones in the system */
135 static LIST_HEAD(,uma_zone) uma_cachezones =
136     LIST_HEAD_INITIALIZER(uma_cachezones);
137 
138 /* This mutex protects the keg list */
139 static struct mtx_padalign uma_mtx;
140 
141 /* Linked list of boot time pages */
142 static LIST_HEAD(,uma_slab) uma_boot_pages =
143     LIST_HEAD_INITIALIZER(uma_boot_pages);
144 
145 /* This mutex protects the boot time pages list */
146 static struct mtx_padalign uma_boot_pages_mtx;
147 
148 /* Is the VM done starting up? */
149 static int booted = 0;
150 #define	UMA_STARTUP	1
151 #define	UMA_STARTUP2	2
152 
153 /*
154  * Only mbuf clusters use ref zones.  Just provide enough references
155  * to support the one user.  New code should not use the ref facility.
156  */
157 static const u_int uma_max_ipers_ref = PAGE_SIZE / MCLBYTES;
158 
159 /*
160  * This is the handle used to schedule events that need to happen
161  * outside of the allocation fast path.
162  */
163 static struct callout uma_callout;
164 #define	UMA_TIMEOUT	20		/* Seconds for callout interval. */
165 
166 /*
167  * This structure is passed as the zone ctor arg so that I don't have to create
168  * a special allocation function just for zones.
169  */
170 struct uma_zctor_args {
171 	const char *name;
172 	size_t size;
173 	uma_ctor ctor;
174 	uma_dtor dtor;
175 	uma_init uminit;
176 	uma_fini fini;
177 	uma_import import;
178 	uma_release release;
179 	void *arg;
180 	uma_keg_t keg;
181 	int align;
182 	uint32_t flags;
183 };
184 
185 struct uma_kctor_args {
186 	uma_zone_t zone;
187 	size_t size;
188 	uma_init uminit;
189 	uma_fini fini;
190 	int align;
191 	uint32_t flags;
192 };
193 
194 struct uma_bucket_zone {
195 	uma_zone_t	ubz_zone;
196 	char		*ubz_name;
197 	int		ubz_entries;	/* Number of items it can hold. */
198 	int		ubz_maxsize;	/* Maximum allocation size per-item. */
199 };
200 
201 /*
202  * Compute the actual number of bucket entries to pack them in power
203  * of two sizes for more efficient space utilization.
204  */
205 #define	BUCKET_SIZE(n)						\
206     (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *))
207 
208 #define	BUCKET_MAX	BUCKET_SIZE(256)
209 
210 struct uma_bucket_zone bucket_zones[] = {
211 	{ NULL, "4 Bucket", BUCKET_SIZE(4), 4096 },
212 	{ NULL, "6 Bucket", BUCKET_SIZE(6), 3072 },
213 	{ NULL, "8 Bucket", BUCKET_SIZE(8), 2048 },
214 	{ NULL, "12 Bucket", BUCKET_SIZE(12), 1536 },
215 	{ NULL, "16 Bucket", BUCKET_SIZE(16), 1024 },
216 	{ NULL, "32 Bucket", BUCKET_SIZE(32), 512 },
217 	{ NULL, "64 Bucket", BUCKET_SIZE(64), 256 },
218 	{ NULL, "128 Bucket", BUCKET_SIZE(128), 128 },
219 	{ NULL, "256 Bucket", BUCKET_SIZE(256), 64 },
220 	{ NULL, NULL, 0}
221 };
222 
223 /*
224  * Flags and enumerations to be passed to internal functions.
225  */
226 enum zfreeskip { SKIP_NONE = 0, SKIP_DTOR, SKIP_FINI };
227 
228 /* Prototypes.. */
229 
230 static void *noobj_alloc(uma_zone_t, int, uint8_t *, int);
231 static void *page_alloc(uma_zone_t, int, uint8_t *, int);
232 static void *startup_alloc(uma_zone_t, int, uint8_t *, int);
233 static void page_free(void *, int, uint8_t);
234 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int);
235 static void cache_drain(uma_zone_t);
236 static void bucket_drain(uma_zone_t, uma_bucket_t);
237 static void bucket_cache_drain(uma_zone_t zone);
238 static int keg_ctor(void *, int, void *, int);
239 static void keg_dtor(void *, int, void *);
240 static int zone_ctor(void *, int, void *, int);
241 static void zone_dtor(void *, int, void *);
242 static int zero_init(void *, int, int);
243 static void keg_small_init(uma_keg_t keg);
244 static void keg_large_init(uma_keg_t keg);
245 static void zone_foreach(void (*zfunc)(uma_zone_t));
246 static void zone_timeout(uma_zone_t zone);
247 static int hash_alloc(struct uma_hash *);
248 static int hash_expand(struct uma_hash *, struct uma_hash *);
249 static void hash_free(struct uma_hash *hash);
250 static void uma_timeout(void *);
251 static void uma_startup3(void);
252 static void *zone_alloc_item(uma_zone_t, void *, int);
253 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip);
254 static void bucket_enable(void);
255 static void bucket_init(void);
256 static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int);
257 static void bucket_free(uma_zone_t zone, uma_bucket_t, void *);
258 static void bucket_zone_drain(void);
259 static uma_bucket_t zone_alloc_bucket(uma_zone_t zone, void *, int flags);
260 static uma_slab_t zone_fetch_slab(uma_zone_t zone, uma_keg_t last, int flags);
261 static uma_slab_t zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int flags);
262 static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab);
263 static void slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item);
264 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit,
265     uma_fini fini, int align, uint32_t flags);
266 static int zone_import(uma_zone_t zone, void **bucket, int max, int flags);
267 static void zone_release(uma_zone_t zone, void **bucket, int cnt);
268 static void uma_zero_item(void *item, uma_zone_t zone);
269 
270 void uma_print_zone(uma_zone_t);
271 void uma_print_stats(void);
272 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS);
273 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS);
274 
275 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
276 
277 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLTYPE_INT,
278     0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones");
279 
280 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLTYPE_STRUCT,
281     0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats");
282 
283 static int zone_warnings = 1;
284 SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RWTUN, &zone_warnings, 0,
285     "Warn when UMA zones becomes full");
286 
287 /*
288  * This routine checks to see whether or not it's safe to enable buckets.
289  */
290 static void
291 bucket_enable(void)
292 {
293 	bucketdisable = vm_page_count_min();
294 }
295 
296 /*
297  * Initialize bucket_zones, the array of zones of buckets of various sizes.
298  *
299  * For each zone, calculate the memory required for each bucket, consisting
300  * of the header and an array of pointers.
301  */
302 static void
303 bucket_init(void)
304 {
305 	struct uma_bucket_zone *ubz;
306 	int size;
307 	int i;
308 
309 	for (i = 0, ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) {
310 		size = roundup(sizeof(struct uma_bucket), sizeof(void *));
311 		size += sizeof(void *) * ubz->ubz_entries;
312 		ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
313 		    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
314 		    UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET);
315 	}
316 }
317 
318 /*
319  * Given a desired number of entries for a bucket, return the zone from which
320  * to allocate the bucket.
321  */
322 static struct uma_bucket_zone *
323 bucket_zone_lookup(int entries)
324 {
325 	struct uma_bucket_zone *ubz;
326 
327 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
328 		if (ubz->ubz_entries >= entries)
329 			return (ubz);
330 	ubz--;
331 	return (ubz);
332 }
333 
334 static int
335 bucket_select(int size)
336 {
337 	struct uma_bucket_zone *ubz;
338 
339 	ubz = &bucket_zones[0];
340 	if (size > ubz->ubz_maxsize)
341 		return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1);
342 
343 	for (; ubz->ubz_entries != 0; ubz++)
344 		if (ubz->ubz_maxsize < size)
345 			break;
346 	ubz--;
347 	return (ubz->ubz_entries);
348 }
349 
350 static uma_bucket_t
351 bucket_alloc(uma_zone_t zone, void *udata, int flags)
352 {
353 	struct uma_bucket_zone *ubz;
354 	uma_bucket_t bucket;
355 
356 	/*
357 	 * This is to stop us from allocating per cpu buckets while we're
358 	 * running out of vm.boot_pages.  Otherwise, we would exhaust the
359 	 * boot pages.  This also prevents us from allocating buckets in
360 	 * low memory situations.
361 	 */
362 	if (bucketdisable)
363 		return (NULL);
364 	/*
365 	 * To limit bucket recursion we store the original zone flags
366 	 * in a cookie passed via zalloc_arg/zfree_arg.  This allows the
367 	 * NOVM flag to persist even through deep recursions.  We also
368 	 * store ZFLAG_BUCKET once we have recursed attempting to allocate
369 	 * a bucket for a bucket zone so we do not allow infinite bucket
370 	 * recursion.  This cookie will even persist to frees of unused
371 	 * buckets via the allocation path or bucket allocations in the
372 	 * free path.
373 	 */
374 	if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
375 		udata = (void *)(uintptr_t)zone->uz_flags;
376 	else {
377 		if ((uintptr_t)udata & UMA_ZFLAG_BUCKET)
378 			return (NULL);
379 		udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET);
380 	}
381 	if ((uintptr_t)udata & UMA_ZFLAG_CACHEONLY)
382 		flags |= M_NOVM;
383 	ubz = bucket_zone_lookup(zone->uz_count);
384 	if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0)
385 		ubz++;
386 	bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags);
387 	if (bucket) {
388 #ifdef INVARIANTS
389 		bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
390 #endif
391 		bucket->ub_cnt = 0;
392 		bucket->ub_entries = ubz->ubz_entries;
393 	}
394 
395 	return (bucket);
396 }
397 
398 static void
399 bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata)
400 {
401 	struct uma_bucket_zone *ubz;
402 
403 	KASSERT(bucket->ub_cnt == 0,
404 	    ("bucket_free: Freeing a non free bucket."));
405 	if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
406 		udata = (void *)(uintptr_t)zone->uz_flags;
407 	ubz = bucket_zone_lookup(bucket->ub_entries);
408 	uma_zfree_arg(ubz->ubz_zone, bucket, udata);
409 }
410 
411 static void
412 bucket_zone_drain(void)
413 {
414 	struct uma_bucket_zone *ubz;
415 
416 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
417 		zone_drain(ubz->ubz_zone);
418 }
419 
420 static void
421 zone_log_warning(uma_zone_t zone)
422 {
423 	static const struct timeval warninterval = { 300, 0 };
424 
425 	if (!zone_warnings || zone->uz_warning == NULL)
426 		return;
427 
428 	if (ratecheck(&zone->uz_ratecheck, &warninterval))
429 		printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning);
430 }
431 
432 static void
433 zone_foreach_keg(uma_zone_t zone, void (*kegfn)(uma_keg_t))
434 {
435 	uma_klink_t klink;
436 
437 	LIST_FOREACH(klink, &zone->uz_kegs, kl_link)
438 		kegfn(klink->kl_keg);
439 }
440 
441 /*
442  * Routine called by timeout which is used to fire off some time interval
443  * based calculations.  (stats, hash size, etc.)
444  *
445  * Arguments:
446  *	arg   Unused
447  *
448  * Returns:
449  *	Nothing
450  */
451 static void
452 uma_timeout(void *unused)
453 {
454 	bucket_enable();
455 	zone_foreach(zone_timeout);
456 
457 	/* Reschedule this event */
458 	callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
459 }
460 
461 /*
462  * Routine to perform timeout driven calculations.  This expands the
463  * hashes and does per cpu statistics aggregation.
464  *
465  *  Returns nothing.
466  */
467 static void
468 keg_timeout(uma_keg_t keg)
469 {
470 
471 	KEG_LOCK(keg);
472 	/*
473 	 * Expand the keg hash table.
474 	 *
475 	 * This is done if the number of slabs is larger than the hash size.
476 	 * What I'm trying to do here is completely reduce collisions.  This
477 	 * may be a little aggressive.  Should I allow for two collisions max?
478 	 */
479 	if (keg->uk_flags & UMA_ZONE_HASH &&
480 	    keg->uk_pages / keg->uk_ppera >= keg->uk_hash.uh_hashsize) {
481 		struct uma_hash newhash;
482 		struct uma_hash oldhash;
483 		int ret;
484 
485 		/*
486 		 * This is so involved because allocating and freeing
487 		 * while the keg lock is held will lead to deadlock.
488 		 * I have to do everything in stages and check for
489 		 * races.
490 		 */
491 		newhash = keg->uk_hash;
492 		KEG_UNLOCK(keg);
493 		ret = hash_alloc(&newhash);
494 		KEG_LOCK(keg);
495 		if (ret) {
496 			if (hash_expand(&keg->uk_hash, &newhash)) {
497 				oldhash = keg->uk_hash;
498 				keg->uk_hash = newhash;
499 			} else
500 				oldhash = newhash;
501 
502 			KEG_UNLOCK(keg);
503 			hash_free(&oldhash);
504 			return;
505 		}
506 	}
507 	KEG_UNLOCK(keg);
508 }
509 
510 static void
511 zone_timeout(uma_zone_t zone)
512 {
513 
514 	zone_foreach_keg(zone, &keg_timeout);
515 }
516 
517 /*
518  * Allocate and zero fill the next sized hash table from the appropriate
519  * backing store.
520  *
521  * Arguments:
522  *	hash  A new hash structure with the old hash size in uh_hashsize
523  *
524  * Returns:
525  *	1 on sucess and 0 on failure.
526  */
527 static int
528 hash_alloc(struct uma_hash *hash)
529 {
530 	int oldsize;
531 	int alloc;
532 
533 	oldsize = hash->uh_hashsize;
534 
535 	/* We're just going to go to a power of two greater */
536 	if (oldsize)  {
537 		hash->uh_hashsize = oldsize * 2;
538 		alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
539 		hash->uh_slab_hash = (struct slabhead *)malloc(alloc,
540 		    M_UMAHASH, M_NOWAIT);
541 	} else {
542 		alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
543 		hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
544 		    M_WAITOK);
545 		hash->uh_hashsize = UMA_HASH_SIZE_INIT;
546 	}
547 	if (hash->uh_slab_hash) {
548 		bzero(hash->uh_slab_hash, alloc);
549 		hash->uh_hashmask = hash->uh_hashsize - 1;
550 		return (1);
551 	}
552 
553 	return (0);
554 }
555 
556 /*
557  * Expands the hash table for HASH zones.  This is done from zone_timeout
558  * to reduce collisions.  This must not be done in the regular allocation
559  * path, otherwise, we can recurse on the vm while allocating pages.
560  *
561  * Arguments:
562  *	oldhash  The hash you want to expand
563  *	newhash  The hash structure for the new table
564  *
565  * Returns:
566  *	Nothing
567  *
568  * Discussion:
569  */
570 static int
571 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
572 {
573 	uma_slab_t slab;
574 	int hval;
575 	int i;
576 
577 	if (!newhash->uh_slab_hash)
578 		return (0);
579 
580 	if (oldhash->uh_hashsize >= newhash->uh_hashsize)
581 		return (0);
582 
583 	/*
584 	 * I need to investigate hash algorithms for resizing without a
585 	 * full rehash.
586 	 */
587 
588 	for (i = 0; i < oldhash->uh_hashsize; i++)
589 		while (!SLIST_EMPTY(&oldhash->uh_slab_hash[i])) {
590 			slab = SLIST_FIRST(&oldhash->uh_slab_hash[i]);
591 			SLIST_REMOVE_HEAD(&oldhash->uh_slab_hash[i], us_hlink);
592 			hval = UMA_HASH(newhash, slab->us_data);
593 			SLIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
594 			    slab, us_hlink);
595 		}
596 
597 	return (1);
598 }
599 
600 /*
601  * Free the hash bucket to the appropriate backing store.
602  *
603  * Arguments:
604  *	slab_hash  The hash bucket we're freeing
605  *	hashsize   The number of entries in that hash bucket
606  *
607  * Returns:
608  *	Nothing
609  */
610 static void
611 hash_free(struct uma_hash *hash)
612 {
613 	if (hash->uh_slab_hash == NULL)
614 		return;
615 	if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
616 		zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE);
617 	else
618 		free(hash->uh_slab_hash, M_UMAHASH);
619 }
620 
621 /*
622  * Frees all outstanding items in a bucket
623  *
624  * Arguments:
625  *	zone   The zone to free to, must be unlocked.
626  *	bucket The free/alloc bucket with items, cpu queue must be locked.
627  *
628  * Returns:
629  *	Nothing
630  */
631 
632 static void
633 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
634 {
635 	int i;
636 
637 	if (bucket == NULL)
638 		return;
639 
640 	if (zone->uz_fini)
641 		for (i = 0; i < bucket->ub_cnt; i++)
642 			zone->uz_fini(bucket->ub_bucket[i], zone->uz_size);
643 	zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt);
644 	bucket->ub_cnt = 0;
645 }
646 
647 /*
648  * Drains the per cpu caches for a zone.
649  *
650  * NOTE: This may only be called while the zone is being turn down, and not
651  * during normal operation.  This is necessary in order that we do not have
652  * to migrate CPUs to drain the per-CPU caches.
653  *
654  * Arguments:
655  *	zone     The zone to drain, must be unlocked.
656  *
657  * Returns:
658  *	Nothing
659  */
660 static void
661 cache_drain(uma_zone_t zone)
662 {
663 	uma_cache_t cache;
664 	int cpu;
665 
666 	/*
667 	 * XXX: It is safe to not lock the per-CPU caches, because we're
668 	 * tearing down the zone anyway.  I.e., there will be no further use
669 	 * of the caches at this point.
670 	 *
671 	 * XXX: It would good to be able to assert that the zone is being
672 	 * torn down to prevent improper use of cache_drain().
673 	 *
674 	 * XXX: We lock the zone before passing into bucket_cache_drain() as
675 	 * it is used elsewhere.  Should the tear-down path be made special
676 	 * there in some form?
677 	 */
678 	CPU_FOREACH(cpu) {
679 		cache = &zone->uz_cpu[cpu];
680 		bucket_drain(zone, cache->uc_allocbucket);
681 		bucket_drain(zone, cache->uc_freebucket);
682 		if (cache->uc_allocbucket != NULL)
683 			bucket_free(zone, cache->uc_allocbucket, NULL);
684 		if (cache->uc_freebucket != NULL)
685 			bucket_free(zone, cache->uc_freebucket, NULL);
686 		cache->uc_allocbucket = cache->uc_freebucket = NULL;
687 	}
688 	ZONE_LOCK(zone);
689 	bucket_cache_drain(zone);
690 	ZONE_UNLOCK(zone);
691 }
692 
693 static void
694 cache_shrink(uma_zone_t zone)
695 {
696 
697 	if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
698 		return;
699 
700 	ZONE_LOCK(zone);
701 	zone->uz_count = (zone->uz_count_min + zone->uz_count) / 2;
702 	ZONE_UNLOCK(zone);
703 }
704 
705 static void
706 cache_drain_safe_cpu(uma_zone_t zone)
707 {
708 	uma_cache_t cache;
709 	uma_bucket_t b1, b2;
710 
711 	if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
712 		return;
713 
714 	b1 = b2 = NULL;
715 	ZONE_LOCK(zone);
716 	critical_enter();
717 	cache = &zone->uz_cpu[curcpu];
718 	if (cache->uc_allocbucket) {
719 		if (cache->uc_allocbucket->ub_cnt != 0)
720 			LIST_INSERT_HEAD(&zone->uz_buckets,
721 			    cache->uc_allocbucket, ub_link);
722 		else
723 			b1 = cache->uc_allocbucket;
724 		cache->uc_allocbucket = NULL;
725 	}
726 	if (cache->uc_freebucket) {
727 		if (cache->uc_freebucket->ub_cnt != 0)
728 			LIST_INSERT_HEAD(&zone->uz_buckets,
729 			    cache->uc_freebucket, ub_link);
730 		else
731 			b2 = cache->uc_freebucket;
732 		cache->uc_freebucket = NULL;
733 	}
734 	critical_exit();
735 	ZONE_UNLOCK(zone);
736 	if (b1)
737 		bucket_free(zone, b1, NULL);
738 	if (b2)
739 		bucket_free(zone, b2, NULL);
740 }
741 
742 /*
743  * Safely drain per-CPU caches of a zone(s) to alloc bucket.
744  * This is an expensive call because it needs to bind to all CPUs
745  * one by one and enter a critical section on each of them in order
746  * to safely access their cache buckets.
747  * Zone lock must not be held on call this function.
748  */
749 static void
750 cache_drain_safe(uma_zone_t zone)
751 {
752 	int cpu;
753 
754 	/*
755 	 * Polite bucket sizes shrinking was not enouth, shrink aggressively.
756 	 */
757 	if (zone)
758 		cache_shrink(zone);
759 	else
760 		zone_foreach(cache_shrink);
761 
762 	CPU_FOREACH(cpu) {
763 		thread_lock(curthread);
764 		sched_bind(curthread, cpu);
765 		thread_unlock(curthread);
766 
767 		if (zone)
768 			cache_drain_safe_cpu(zone);
769 		else
770 			zone_foreach(cache_drain_safe_cpu);
771 	}
772 	thread_lock(curthread);
773 	sched_unbind(curthread);
774 	thread_unlock(curthread);
775 }
776 
777 /*
778  * Drain the cached buckets from a zone.  Expects a locked zone on entry.
779  */
780 static void
781 bucket_cache_drain(uma_zone_t zone)
782 {
783 	uma_bucket_t bucket;
784 
785 	/*
786 	 * Drain the bucket queues and free the buckets, we just keep two per
787 	 * cpu (alloc/free).
788 	 */
789 	while ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) {
790 		LIST_REMOVE(bucket, ub_link);
791 		ZONE_UNLOCK(zone);
792 		bucket_drain(zone, bucket);
793 		bucket_free(zone, bucket, NULL);
794 		ZONE_LOCK(zone);
795 	}
796 
797 	/*
798 	 * Shrink further bucket sizes.  Price of single zone lock collision
799 	 * is probably lower then price of global cache drain.
800 	 */
801 	if (zone->uz_count > zone->uz_count_min)
802 		zone->uz_count--;
803 }
804 
805 static void
806 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start)
807 {
808 	uint8_t *mem;
809 	int i;
810 	uint8_t flags;
811 
812 	mem = slab->us_data;
813 	flags = slab->us_flags;
814 	i = start;
815 	if (keg->uk_fini != NULL) {
816 		for (i--; i > -1; i--)
817 			keg->uk_fini(slab->us_data + (keg->uk_rsize * i),
818 			    keg->uk_size);
819 	}
820 	if (keg->uk_flags & UMA_ZONE_OFFPAGE)
821 		zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE);
822 #ifdef UMA_DEBUG
823 	printf("%s: Returning %d bytes.\n", keg->uk_name,
824 	    PAGE_SIZE * keg->uk_ppera);
825 #endif
826 	keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags);
827 }
828 
829 /*
830  * Frees pages from a keg back to the system.  This is done on demand from
831  * the pageout daemon.
832  *
833  * Returns nothing.
834  */
835 static void
836 keg_drain(uma_keg_t keg)
837 {
838 	struct slabhead freeslabs = { 0 };
839 	uma_slab_t slab;
840 	uma_slab_t n;
841 
842 	/*
843 	 * We don't want to take pages from statically allocated kegs at this
844 	 * time
845 	 */
846 	if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL)
847 		return;
848 
849 #ifdef UMA_DEBUG
850 	printf("%s free items: %u\n", keg->uk_name, keg->uk_free);
851 #endif
852 	KEG_LOCK(keg);
853 	if (keg->uk_free == 0)
854 		goto finished;
855 
856 	slab = LIST_FIRST(&keg->uk_free_slab);
857 	while (slab) {
858 		n = LIST_NEXT(slab, us_link);
859 
860 		/* We have no where to free these to */
861 		if (slab->us_flags & UMA_SLAB_BOOT) {
862 			slab = n;
863 			continue;
864 		}
865 
866 		LIST_REMOVE(slab, us_link);
867 		keg->uk_pages -= keg->uk_ppera;
868 		keg->uk_free -= keg->uk_ipers;
869 
870 		if (keg->uk_flags & UMA_ZONE_HASH)
871 			UMA_HASH_REMOVE(&keg->uk_hash, slab, slab->us_data);
872 
873 		SLIST_INSERT_HEAD(&freeslabs, slab, us_hlink);
874 
875 		slab = n;
876 	}
877 finished:
878 	KEG_UNLOCK(keg);
879 
880 	while ((slab = SLIST_FIRST(&freeslabs)) != NULL) {
881 		SLIST_REMOVE(&freeslabs, slab, uma_slab, us_hlink);
882 		keg_free_slab(keg, slab, keg->uk_ipers);
883 	}
884 }
885 
886 static void
887 zone_drain_wait(uma_zone_t zone, int waitok)
888 {
889 
890 	/*
891 	 * Set draining to interlock with zone_dtor() so we can release our
892 	 * locks as we go.  Only dtor() should do a WAITOK call since it
893 	 * is the only call that knows the structure will still be available
894 	 * when it wakes up.
895 	 */
896 	ZONE_LOCK(zone);
897 	while (zone->uz_flags & UMA_ZFLAG_DRAINING) {
898 		if (waitok == M_NOWAIT)
899 			goto out;
900 		mtx_unlock(&uma_mtx);
901 		msleep(zone, zone->uz_lockptr, PVM, "zonedrain", 1);
902 		mtx_lock(&uma_mtx);
903 	}
904 	zone->uz_flags |= UMA_ZFLAG_DRAINING;
905 	bucket_cache_drain(zone);
906 	ZONE_UNLOCK(zone);
907 	/*
908 	 * The DRAINING flag protects us from being freed while
909 	 * we're running.  Normally the uma_mtx would protect us but we
910 	 * must be able to release and acquire the right lock for each keg.
911 	 */
912 	zone_foreach_keg(zone, &keg_drain);
913 	ZONE_LOCK(zone);
914 	zone->uz_flags &= ~UMA_ZFLAG_DRAINING;
915 	wakeup(zone);
916 out:
917 	ZONE_UNLOCK(zone);
918 }
919 
920 void
921 zone_drain(uma_zone_t zone)
922 {
923 
924 	zone_drain_wait(zone, M_NOWAIT);
925 }
926 
927 /*
928  * Allocate a new slab for a keg.  This does not insert the slab onto a list.
929  *
930  * Arguments:
931  *	wait  Shall we wait?
932  *
933  * Returns:
934  *	The slab that was allocated or NULL if there is no memory and the
935  *	caller specified M_NOWAIT.
936  */
937 static uma_slab_t
938 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int wait)
939 {
940 	uma_slabrefcnt_t slabref;
941 	uma_alloc allocf;
942 	uma_slab_t slab;
943 	uint8_t *mem;
944 	uint8_t flags;
945 	int i;
946 
947 	mtx_assert(&keg->uk_lock, MA_OWNED);
948 	slab = NULL;
949 	mem = NULL;
950 
951 #ifdef UMA_DEBUG
952 	printf("alloc_slab:  Allocating a new slab for %s\n", keg->uk_name);
953 #endif
954 	allocf = keg->uk_allocf;
955 	KEG_UNLOCK(keg);
956 
957 	if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
958 		slab = zone_alloc_item(keg->uk_slabzone, NULL, wait);
959 		if (slab == NULL)
960 			goto out;
961 	}
962 
963 	/*
964 	 * This reproduces the old vm_zone behavior of zero filling pages the
965 	 * first time they are added to a zone.
966 	 *
967 	 * Malloced items are zeroed in uma_zalloc.
968 	 */
969 
970 	if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
971 		wait |= M_ZERO;
972 	else
973 		wait &= ~M_ZERO;
974 
975 	if (keg->uk_flags & UMA_ZONE_NODUMP)
976 		wait |= M_NODUMP;
977 
978 	/* zone is passed for legacy reasons. */
979 	mem = allocf(zone, keg->uk_ppera * PAGE_SIZE, &flags, wait);
980 	if (mem == NULL) {
981 		if (keg->uk_flags & UMA_ZONE_OFFPAGE)
982 			zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE);
983 		slab = NULL;
984 		goto out;
985 	}
986 
987 	/* Point the slab into the allocated memory */
988 	if (!(keg->uk_flags & UMA_ZONE_OFFPAGE))
989 		slab = (uma_slab_t )(mem + keg->uk_pgoff);
990 
991 	if (keg->uk_flags & UMA_ZONE_VTOSLAB)
992 		for (i = 0; i < keg->uk_ppera; i++)
993 			vsetslab((vm_offset_t)mem + (i * PAGE_SIZE), slab);
994 
995 	slab->us_keg = keg;
996 	slab->us_data = mem;
997 	slab->us_freecount = keg->uk_ipers;
998 	slab->us_flags = flags;
999 	BIT_FILL(SLAB_SETSIZE, &slab->us_free);
1000 #ifdef INVARIANTS
1001 	BIT_ZERO(SLAB_SETSIZE, &slab->us_debugfree);
1002 #endif
1003 	if (keg->uk_flags & UMA_ZONE_REFCNT) {
1004 		slabref = (uma_slabrefcnt_t)slab;
1005 		for (i = 0; i < keg->uk_ipers; i++)
1006 			slabref->us_refcnt[i] = 0;
1007 	}
1008 
1009 	if (keg->uk_init != NULL) {
1010 		for (i = 0; i < keg->uk_ipers; i++)
1011 			if (keg->uk_init(slab->us_data + (keg->uk_rsize * i),
1012 			    keg->uk_size, wait) != 0)
1013 				break;
1014 		if (i != keg->uk_ipers) {
1015 			keg_free_slab(keg, slab, i);
1016 			slab = NULL;
1017 			goto out;
1018 		}
1019 	}
1020 out:
1021 	KEG_LOCK(keg);
1022 
1023 	if (slab != NULL) {
1024 		if (keg->uk_flags & UMA_ZONE_HASH)
1025 			UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
1026 
1027 		keg->uk_pages += keg->uk_ppera;
1028 		keg->uk_free += keg->uk_ipers;
1029 	}
1030 
1031 	return (slab);
1032 }
1033 
1034 /*
1035  * This function is intended to be used early on in place of page_alloc() so
1036  * that we may use the boot time page cache to satisfy allocations before
1037  * the VM is ready.
1038  */
1039 static void *
1040 startup_alloc(uma_zone_t zone, int bytes, uint8_t *pflag, int wait)
1041 {
1042 	uma_keg_t keg;
1043 	uma_slab_t tmps;
1044 	int pages, check_pages;
1045 
1046 	keg = zone_first_keg(zone);
1047 	pages = howmany(bytes, PAGE_SIZE);
1048 	check_pages = pages - 1;
1049 	KASSERT(pages > 0, ("startup_alloc can't reserve 0 pages\n"));
1050 
1051 	/*
1052 	 * Check our small startup cache to see if it has pages remaining.
1053 	 */
1054 	mtx_lock(&uma_boot_pages_mtx);
1055 
1056 	/* First check if we have enough room. */
1057 	tmps = LIST_FIRST(&uma_boot_pages);
1058 	while (tmps != NULL && check_pages-- > 0)
1059 		tmps = LIST_NEXT(tmps, us_link);
1060 	if (tmps != NULL) {
1061 		/*
1062 		 * It's ok to lose tmps references.  The last one will
1063 		 * have tmps->us_data pointing to the start address of
1064 		 * "pages" contiguous pages of memory.
1065 		 */
1066 		while (pages-- > 0) {
1067 			tmps = LIST_FIRST(&uma_boot_pages);
1068 			LIST_REMOVE(tmps, us_link);
1069 		}
1070 		mtx_unlock(&uma_boot_pages_mtx);
1071 		*pflag = tmps->us_flags;
1072 		return (tmps->us_data);
1073 	}
1074 	mtx_unlock(&uma_boot_pages_mtx);
1075 	if (booted < UMA_STARTUP2)
1076 		panic("UMA: Increase vm.boot_pages");
1077 	/*
1078 	 * Now that we've booted reset these users to their real allocator.
1079 	 */
1080 #ifdef UMA_MD_SMALL_ALLOC
1081 	keg->uk_allocf = (keg->uk_ppera > 1) ? page_alloc : uma_small_alloc;
1082 #else
1083 	keg->uk_allocf = page_alloc;
1084 #endif
1085 	return keg->uk_allocf(zone, bytes, pflag, wait);
1086 }
1087 
1088 /*
1089  * Allocates a number of pages from the system
1090  *
1091  * Arguments:
1092  *	bytes  The number of bytes requested
1093  *	wait  Shall we wait?
1094  *
1095  * Returns:
1096  *	A pointer to the alloced memory or possibly
1097  *	NULL if M_NOWAIT is set.
1098  */
1099 static void *
1100 page_alloc(uma_zone_t zone, int bytes, uint8_t *pflag, int wait)
1101 {
1102 	void *p;	/* Returned page */
1103 
1104 	*pflag = UMA_SLAB_KMEM;
1105 	p = (void *) kmem_malloc(kmem_arena, bytes, wait);
1106 
1107 	return (p);
1108 }
1109 
1110 /*
1111  * Allocates a number of pages from within an object
1112  *
1113  * Arguments:
1114  *	bytes  The number of bytes requested
1115  *	wait   Shall we wait?
1116  *
1117  * Returns:
1118  *	A pointer to the alloced memory or possibly
1119  *	NULL if M_NOWAIT is set.
1120  */
1121 static void *
1122 noobj_alloc(uma_zone_t zone, int bytes, uint8_t *flags, int wait)
1123 {
1124 	TAILQ_HEAD(, vm_page) alloctail;
1125 	u_long npages;
1126 	vm_offset_t retkva, zkva;
1127 	vm_page_t p, p_next;
1128 	uma_keg_t keg;
1129 
1130 	TAILQ_INIT(&alloctail);
1131 	keg = zone_first_keg(zone);
1132 
1133 	npages = howmany(bytes, PAGE_SIZE);
1134 	while (npages > 0) {
1135 		p = vm_page_alloc(NULL, 0, VM_ALLOC_INTERRUPT |
1136 		    VM_ALLOC_WIRED | VM_ALLOC_NOOBJ);
1137 		if (p != NULL) {
1138 			/*
1139 			 * Since the page does not belong to an object, its
1140 			 * listq is unused.
1141 			 */
1142 			TAILQ_INSERT_TAIL(&alloctail, p, listq);
1143 			npages--;
1144 			continue;
1145 		}
1146 		if (wait & M_WAITOK) {
1147 			VM_WAIT;
1148 			continue;
1149 		}
1150 
1151 		/*
1152 		 * Page allocation failed, free intermediate pages and
1153 		 * exit.
1154 		 */
1155 		TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1156 			vm_page_unwire(p, PQ_INACTIVE);
1157 			vm_page_free(p);
1158 		}
1159 		return (NULL);
1160 	}
1161 	*flags = UMA_SLAB_PRIV;
1162 	zkva = keg->uk_kva +
1163 	    atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
1164 	retkva = zkva;
1165 	TAILQ_FOREACH(p, &alloctail, listq) {
1166 		pmap_qenter(zkva, &p, 1);
1167 		zkva += PAGE_SIZE;
1168 	}
1169 
1170 	return ((void *)retkva);
1171 }
1172 
1173 /*
1174  * Frees a number of pages to the system
1175  *
1176  * Arguments:
1177  *	mem   A pointer to the memory to be freed
1178  *	size  The size of the memory being freed
1179  *	flags The original p->us_flags field
1180  *
1181  * Returns:
1182  *	Nothing
1183  */
1184 static void
1185 page_free(void *mem, int size, uint8_t flags)
1186 {
1187 	struct vmem *vmem;
1188 
1189 	if (flags & UMA_SLAB_KMEM)
1190 		vmem = kmem_arena;
1191 	else if (flags & UMA_SLAB_KERNEL)
1192 		vmem = kernel_arena;
1193 	else
1194 		panic("UMA: page_free used with invalid flags %d", flags);
1195 
1196 	kmem_free(vmem, (vm_offset_t)mem, size);
1197 }
1198 
1199 /*
1200  * Zero fill initializer
1201  *
1202  * Arguments/Returns follow uma_init specifications
1203  */
1204 static int
1205 zero_init(void *mem, int size, int flags)
1206 {
1207 	bzero(mem, size);
1208 	return (0);
1209 }
1210 
1211 /*
1212  * Finish creating a small uma keg.  This calculates ipers, and the keg size.
1213  *
1214  * Arguments
1215  *	keg  The zone we should initialize
1216  *
1217  * Returns
1218  *	Nothing
1219  */
1220 static void
1221 keg_small_init(uma_keg_t keg)
1222 {
1223 	u_int rsize;
1224 	u_int memused;
1225 	u_int wastedspace;
1226 	u_int shsize;
1227 
1228 	if (keg->uk_flags & UMA_ZONE_PCPU) {
1229 		u_int ncpus = mp_ncpus ? mp_ncpus : MAXCPU;
1230 
1231 		keg->uk_slabsize = sizeof(struct pcpu);
1232 		keg->uk_ppera = howmany(ncpus * sizeof(struct pcpu),
1233 		    PAGE_SIZE);
1234 	} else {
1235 		keg->uk_slabsize = UMA_SLAB_SIZE;
1236 		keg->uk_ppera = 1;
1237 	}
1238 
1239 	/*
1240 	 * Calculate the size of each allocation (rsize) according to
1241 	 * alignment.  If the requested size is smaller than we have
1242 	 * allocation bits for we round it up.
1243 	 */
1244 	rsize = keg->uk_size;
1245 	if (rsize < keg->uk_slabsize / SLAB_SETSIZE)
1246 		rsize = keg->uk_slabsize / SLAB_SETSIZE;
1247 	if (rsize & keg->uk_align)
1248 		rsize = (rsize & ~keg->uk_align) + (keg->uk_align + 1);
1249 	keg->uk_rsize = rsize;
1250 
1251 	KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
1252 	    keg->uk_rsize < sizeof(struct pcpu),
1253 	    ("%s: size %u too large", __func__, keg->uk_rsize));
1254 
1255 	if (keg->uk_flags & UMA_ZONE_REFCNT)
1256 		rsize += sizeof(uint32_t);
1257 
1258 	if (keg->uk_flags & UMA_ZONE_OFFPAGE)
1259 		shsize = 0;
1260 	else
1261 		shsize = sizeof(struct uma_slab);
1262 
1263 	keg->uk_ipers = (keg->uk_slabsize - shsize) / rsize;
1264 	KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE,
1265 	    ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers));
1266 
1267 	memused = keg->uk_ipers * rsize + shsize;
1268 	wastedspace = keg->uk_slabsize - memused;
1269 
1270 	/*
1271 	 * We can't do OFFPAGE if we're internal or if we've been
1272 	 * asked to not go to the VM for buckets.  If we do this we
1273 	 * may end up going to the VM  for slabs which we do not
1274 	 * want to do if we're UMA_ZFLAG_CACHEONLY as a result
1275 	 * of UMA_ZONE_VM, which clearly forbids it.
1276 	 */
1277 	if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) ||
1278 	    (keg->uk_flags & UMA_ZFLAG_CACHEONLY))
1279 		return;
1280 
1281 	/*
1282 	 * See if using an OFFPAGE slab will limit our waste.  Only do
1283 	 * this if it permits more items per-slab.
1284 	 *
1285 	 * XXX We could try growing slabsize to limit max waste as well.
1286 	 * Historically this was not done because the VM could not
1287 	 * efficiently handle contiguous allocations.
1288 	 */
1289 	if ((wastedspace >= keg->uk_slabsize / UMA_MAX_WASTE) &&
1290 	    (keg->uk_ipers < (keg->uk_slabsize / keg->uk_rsize))) {
1291 		keg->uk_ipers = keg->uk_slabsize / keg->uk_rsize;
1292 		KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE,
1293 		    ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers));
1294 #ifdef UMA_DEBUG
1295 		printf("UMA decided we need offpage slab headers for "
1296 		    "keg: %s, calculated wastedspace = %d, "
1297 		    "maximum wasted space allowed = %d, "
1298 		    "calculated ipers = %d, "
1299 		    "new wasted space = %d\n", keg->uk_name, wastedspace,
1300 		    keg->uk_slabsize / UMA_MAX_WASTE, keg->uk_ipers,
1301 		    keg->uk_slabsize - keg->uk_ipers * keg->uk_rsize);
1302 #endif
1303 		keg->uk_flags |= UMA_ZONE_OFFPAGE;
1304 	}
1305 
1306 	if ((keg->uk_flags & UMA_ZONE_OFFPAGE) &&
1307 	    (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1308 		keg->uk_flags |= UMA_ZONE_HASH;
1309 }
1310 
1311 /*
1312  * Finish creating a large (> UMA_SLAB_SIZE) uma kegs.  Just give in and do
1313  * OFFPAGE for now.  When I can allow for more dynamic slab sizes this will be
1314  * more complicated.
1315  *
1316  * Arguments
1317  *	keg  The keg we should initialize
1318  *
1319  * Returns
1320  *	Nothing
1321  */
1322 static void
1323 keg_large_init(uma_keg_t keg)
1324 {
1325 	u_int shsize;
1326 
1327 	KASSERT(keg != NULL, ("Keg is null in keg_large_init"));
1328 	KASSERT((keg->uk_flags & UMA_ZFLAG_CACHEONLY) == 0,
1329 	    ("keg_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY keg"));
1330 	KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0,
1331 	    ("%s: Cannot large-init a UMA_ZONE_PCPU keg", __func__));
1332 
1333 	keg->uk_ppera = howmany(keg->uk_size, PAGE_SIZE);
1334 	keg->uk_slabsize = keg->uk_ppera * PAGE_SIZE;
1335 	keg->uk_ipers = 1;
1336 	keg->uk_rsize = keg->uk_size;
1337 
1338 	/* We can't do OFFPAGE if we're internal, bail out here. */
1339 	if (keg->uk_flags & UMA_ZFLAG_INTERNAL)
1340 		return;
1341 
1342 	/* Check whether we have enough space to not do OFFPAGE. */
1343 	if ((keg->uk_flags & UMA_ZONE_OFFPAGE) == 0) {
1344 		shsize = sizeof(struct uma_slab);
1345 		if (keg->uk_flags & UMA_ZONE_REFCNT)
1346 			shsize += keg->uk_ipers * sizeof(uint32_t);
1347 		if (shsize & UMA_ALIGN_PTR)
1348 			shsize = (shsize & ~UMA_ALIGN_PTR) +
1349 			    (UMA_ALIGN_PTR + 1);
1350 
1351 		if ((PAGE_SIZE * keg->uk_ppera) - keg->uk_rsize < shsize)
1352 			keg->uk_flags |= UMA_ZONE_OFFPAGE;
1353 	}
1354 
1355 	if ((keg->uk_flags & UMA_ZONE_OFFPAGE) &&
1356 	    (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0)
1357 		keg->uk_flags |= UMA_ZONE_HASH;
1358 }
1359 
1360 static void
1361 keg_cachespread_init(uma_keg_t keg)
1362 {
1363 	int alignsize;
1364 	int trailer;
1365 	int pages;
1366 	int rsize;
1367 
1368 	KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0,
1369 	    ("%s: Cannot cachespread-init a UMA_ZONE_PCPU keg", __func__));
1370 
1371 	alignsize = keg->uk_align + 1;
1372 	rsize = keg->uk_size;
1373 	/*
1374 	 * We want one item to start on every align boundary in a page.  To
1375 	 * do this we will span pages.  We will also extend the item by the
1376 	 * size of align if it is an even multiple of align.  Otherwise, it
1377 	 * would fall on the same boundary every time.
1378 	 */
1379 	if (rsize & keg->uk_align)
1380 		rsize = (rsize & ~keg->uk_align) + alignsize;
1381 	if ((rsize & alignsize) == 0)
1382 		rsize += alignsize;
1383 	trailer = rsize - keg->uk_size;
1384 	pages = (rsize * (PAGE_SIZE / alignsize)) / PAGE_SIZE;
1385 	pages = MIN(pages, (128 * 1024) / PAGE_SIZE);
1386 	keg->uk_rsize = rsize;
1387 	keg->uk_ppera = pages;
1388 	keg->uk_slabsize = UMA_SLAB_SIZE;
1389 	keg->uk_ipers = ((pages * PAGE_SIZE) + trailer) / rsize;
1390 	keg->uk_flags |= UMA_ZONE_OFFPAGE | UMA_ZONE_VTOSLAB;
1391 	KASSERT(keg->uk_ipers <= SLAB_SETSIZE,
1392 	    ("%s: keg->uk_ipers too high(%d) increase max_ipers", __func__,
1393 	    keg->uk_ipers));
1394 }
1395 
1396 /*
1397  * Keg header ctor.  This initializes all fields, locks, etc.  And inserts
1398  * the keg onto the global keg list.
1399  *
1400  * Arguments/Returns follow uma_ctor specifications
1401  *	udata  Actually uma_kctor_args
1402  */
1403 static int
1404 keg_ctor(void *mem, int size, void *udata, int flags)
1405 {
1406 	struct uma_kctor_args *arg = udata;
1407 	uma_keg_t keg = mem;
1408 	uma_zone_t zone;
1409 
1410 	bzero(keg, size);
1411 	keg->uk_size = arg->size;
1412 	keg->uk_init = arg->uminit;
1413 	keg->uk_fini = arg->fini;
1414 	keg->uk_align = arg->align;
1415 	keg->uk_free = 0;
1416 	keg->uk_reserve = 0;
1417 	keg->uk_pages = 0;
1418 	keg->uk_flags = arg->flags;
1419 	keg->uk_allocf = page_alloc;
1420 	keg->uk_freef = page_free;
1421 	keg->uk_slabzone = NULL;
1422 
1423 	/*
1424 	 * The master zone is passed to us at keg-creation time.
1425 	 */
1426 	zone = arg->zone;
1427 	keg->uk_name = zone->uz_name;
1428 
1429 	if (arg->flags & UMA_ZONE_VM)
1430 		keg->uk_flags |= UMA_ZFLAG_CACHEONLY;
1431 
1432 	if (arg->flags & UMA_ZONE_ZINIT)
1433 		keg->uk_init = zero_init;
1434 
1435 	if (arg->flags & UMA_ZONE_REFCNT || arg->flags & UMA_ZONE_MALLOC)
1436 		keg->uk_flags |= UMA_ZONE_VTOSLAB;
1437 
1438 	if (arg->flags & UMA_ZONE_PCPU)
1439 #ifdef SMP
1440 		keg->uk_flags |= UMA_ZONE_OFFPAGE;
1441 #else
1442 		keg->uk_flags &= ~UMA_ZONE_PCPU;
1443 #endif
1444 
1445 	if (keg->uk_flags & UMA_ZONE_CACHESPREAD) {
1446 		keg_cachespread_init(keg);
1447 	} else if (keg->uk_flags & UMA_ZONE_REFCNT) {
1448 		if (keg->uk_size >
1449 		    (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt) -
1450 		    sizeof(uint32_t)))
1451 			keg_large_init(keg);
1452 		else
1453 			keg_small_init(keg);
1454 	} else {
1455 		if (keg->uk_size > (UMA_SLAB_SIZE - sizeof(struct uma_slab)))
1456 			keg_large_init(keg);
1457 		else
1458 			keg_small_init(keg);
1459 	}
1460 
1461 	if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
1462 		if (keg->uk_flags & UMA_ZONE_REFCNT) {
1463 			if (keg->uk_ipers > uma_max_ipers_ref)
1464 				panic("Too many ref items per zone: %d > %d\n",
1465 				    keg->uk_ipers, uma_max_ipers_ref);
1466 			keg->uk_slabzone = slabrefzone;
1467 		} else
1468 			keg->uk_slabzone = slabzone;
1469 	}
1470 
1471 	/*
1472 	 * If we haven't booted yet we need allocations to go through the
1473 	 * startup cache until the vm is ready.
1474 	 */
1475 	if (keg->uk_ppera == 1) {
1476 #ifdef UMA_MD_SMALL_ALLOC
1477 		keg->uk_allocf = uma_small_alloc;
1478 		keg->uk_freef = uma_small_free;
1479 
1480 		if (booted < UMA_STARTUP)
1481 			keg->uk_allocf = startup_alloc;
1482 #else
1483 		if (booted < UMA_STARTUP2)
1484 			keg->uk_allocf = startup_alloc;
1485 #endif
1486 	} else if (booted < UMA_STARTUP2 &&
1487 	    (keg->uk_flags & UMA_ZFLAG_INTERNAL))
1488 		keg->uk_allocf = startup_alloc;
1489 
1490 	/*
1491 	 * Initialize keg's lock
1492 	 */
1493 	KEG_LOCK_INIT(keg, (arg->flags & UMA_ZONE_MTXCLASS));
1494 
1495 	/*
1496 	 * If we're putting the slab header in the actual page we need to
1497 	 * figure out where in each page it goes.  This calculates a right
1498 	 * justified offset into the memory on an ALIGN_PTR boundary.
1499 	 */
1500 	if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) {
1501 		u_int totsize;
1502 
1503 		/* Size of the slab struct and free list */
1504 		totsize = sizeof(struct uma_slab);
1505 
1506 		/* Size of the reference counts. */
1507 		if (keg->uk_flags & UMA_ZONE_REFCNT)
1508 			totsize += keg->uk_ipers * sizeof(uint32_t);
1509 
1510 		if (totsize & UMA_ALIGN_PTR)
1511 			totsize = (totsize & ~UMA_ALIGN_PTR) +
1512 			    (UMA_ALIGN_PTR + 1);
1513 		keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - totsize;
1514 
1515 		/*
1516 		 * The only way the following is possible is if with our
1517 		 * UMA_ALIGN_PTR adjustments we are now bigger than
1518 		 * UMA_SLAB_SIZE.  I haven't checked whether this is
1519 		 * mathematically possible for all cases, so we make
1520 		 * sure here anyway.
1521 		 */
1522 		totsize = keg->uk_pgoff + sizeof(struct uma_slab);
1523 		if (keg->uk_flags & UMA_ZONE_REFCNT)
1524 			totsize += keg->uk_ipers * sizeof(uint32_t);
1525 		if (totsize > PAGE_SIZE * keg->uk_ppera) {
1526 			printf("zone %s ipers %d rsize %d size %d\n",
1527 			    zone->uz_name, keg->uk_ipers, keg->uk_rsize,
1528 			    keg->uk_size);
1529 			panic("UMA slab won't fit.");
1530 		}
1531 	}
1532 
1533 	if (keg->uk_flags & UMA_ZONE_HASH)
1534 		hash_alloc(&keg->uk_hash);
1535 
1536 #ifdef UMA_DEBUG
1537 	printf("UMA: %s(%p) size %d(%d) flags %#x ipers %d ppera %d out %d free %d\n",
1538 	    zone->uz_name, zone, keg->uk_size, keg->uk_rsize, keg->uk_flags,
1539 	    keg->uk_ipers, keg->uk_ppera,
1540 	    (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free);
1541 #endif
1542 
1543 	LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
1544 
1545 	mtx_lock(&uma_mtx);
1546 	LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
1547 	mtx_unlock(&uma_mtx);
1548 	return (0);
1549 }
1550 
1551 /*
1552  * Zone header ctor.  This initializes all fields, locks, etc.
1553  *
1554  * Arguments/Returns follow uma_ctor specifications
1555  *	udata  Actually uma_zctor_args
1556  */
1557 static int
1558 zone_ctor(void *mem, int size, void *udata, int flags)
1559 {
1560 	struct uma_zctor_args *arg = udata;
1561 	uma_zone_t zone = mem;
1562 	uma_zone_t z;
1563 	uma_keg_t keg;
1564 
1565 	bzero(zone, size);
1566 	zone->uz_name = arg->name;
1567 	zone->uz_ctor = arg->ctor;
1568 	zone->uz_dtor = arg->dtor;
1569 	zone->uz_slab = zone_fetch_slab;
1570 	zone->uz_init = NULL;
1571 	zone->uz_fini = NULL;
1572 	zone->uz_allocs = 0;
1573 	zone->uz_frees = 0;
1574 	zone->uz_fails = 0;
1575 	zone->uz_sleeps = 0;
1576 	zone->uz_count = 0;
1577 	zone->uz_count_min = 0;
1578 	zone->uz_flags = 0;
1579 	zone->uz_warning = NULL;
1580 	timevalclear(&zone->uz_ratecheck);
1581 	keg = arg->keg;
1582 
1583 	ZONE_LOCK_INIT(zone, (arg->flags & UMA_ZONE_MTXCLASS));
1584 
1585 	/*
1586 	 * This is a pure cache zone, no kegs.
1587 	 */
1588 	if (arg->import) {
1589 		if (arg->flags & UMA_ZONE_VM)
1590 			arg->flags |= UMA_ZFLAG_CACHEONLY;
1591 		zone->uz_flags = arg->flags;
1592 		zone->uz_size = arg->size;
1593 		zone->uz_import = arg->import;
1594 		zone->uz_release = arg->release;
1595 		zone->uz_arg = arg->arg;
1596 		zone->uz_lockptr = &zone->uz_lock;
1597 		mtx_lock(&uma_mtx);
1598 		LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link);
1599 		mtx_unlock(&uma_mtx);
1600 		goto out;
1601 	}
1602 
1603 	/*
1604 	 * Use the regular zone/keg/slab allocator.
1605 	 */
1606 	zone->uz_import = (uma_import)zone_import;
1607 	zone->uz_release = (uma_release)zone_release;
1608 	zone->uz_arg = zone;
1609 
1610 	if (arg->flags & UMA_ZONE_SECONDARY) {
1611 		KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
1612 		zone->uz_init = arg->uminit;
1613 		zone->uz_fini = arg->fini;
1614 		zone->uz_lockptr = &keg->uk_lock;
1615 		zone->uz_flags |= UMA_ZONE_SECONDARY;
1616 		mtx_lock(&uma_mtx);
1617 		ZONE_LOCK(zone);
1618 		LIST_FOREACH(z, &keg->uk_zones, uz_link) {
1619 			if (LIST_NEXT(z, uz_link) == NULL) {
1620 				LIST_INSERT_AFTER(z, zone, uz_link);
1621 				break;
1622 			}
1623 		}
1624 		ZONE_UNLOCK(zone);
1625 		mtx_unlock(&uma_mtx);
1626 	} else if (keg == NULL) {
1627 		if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
1628 		    arg->align, arg->flags)) == NULL)
1629 			return (ENOMEM);
1630 	} else {
1631 		struct uma_kctor_args karg;
1632 		int error;
1633 
1634 		/* We should only be here from uma_startup() */
1635 		karg.size = arg->size;
1636 		karg.uminit = arg->uminit;
1637 		karg.fini = arg->fini;
1638 		karg.align = arg->align;
1639 		karg.flags = arg->flags;
1640 		karg.zone = zone;
1641 		error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
1642 		    flags);
1643 		if (error)
1644 			return (error);
1645 	}
1646 
1647 	/*
1648 	 * Link in the first keg.
1649 	 */
1650 	zone->uz_klink.kl_keg = keg;
1651 	LIST_INSERT_HEAD(&zone->uz_kegs, &zone->uz_klink, kl_link);
1652 	zone->uz_lockptr = &keg->uk_lock;
1653 	zone->uz_size = keg->uk_size;
1654 	zone->uz_flags |= (keg->uk_flags &
1655 	    (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
1656 
1657 	/*
1658 	 * Some internal zones don't have room allocated for the per cpu
1659 	 * caches.  If we're internal, bail out here.
1660 	 */
1661 	if (keg->uk_flags & UMA_ZFLAG_INTERNAL) {
1662 		KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
1663 		    ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
1664 		return (0);
1665 	}
1666 
1667 out:
1668 	if ((arg->flags & UMA_ZONE_MAXBUCKET) == 0)
1669 		zone->uz_count = bucket_select(zone->uz_size);
1670 	else
1671 		zone->uz_count = BUCKET_MAX;
1672 	zone->uz_count_min = zone->uz_count;
1673 
1674 	return (0);
1675 }
1676 
1677 /*
1678  * Keg header dtor.  This frees all data, destroys locks, frees the hash
1679  * table and removes the keg from the global list.
1680  *
1681  * Arguments/Returns follow uma_dtor specifications
1682  *	udata  unused
1683  */
1684 static void
1685 keg_dtor(void *arg, int size, void *udata)
1686 {
1687 	uma_keg_t keg;
1688 
1689 	keg = (uma_keg_t)arg;
1690 	KEG_LOCK(keg);
1691 	if (keg->uk_free != 0) {
1692 		printf("Freed UMA keg (%s) was not empty (%d items). "
1693 		    " Lost %d pages of memory.\n",
1694 		    keg->uk_name ? keg->uk_name : "",
1695 		    keg->uk_free, keg->uk_pages);
1696 	}
1697 	KEG_UNLOCK(keg);
1698 
1699 	hash_free(&keg->uk_hash);
1700 
1701 	KEG_LOCK_FINI(keg);
1702 }
1703 
1704 /*
1705  * Zone header dtor.
1706  *
1707  * Arguments/Returns follow uma_dtor specifications
1708  *	udata  unused
1709  */
1710 static void
1711 zone_dtor(void *arg, int size, void *udata)
1712 {
1713 	uma_klink_t klink;
1714 	uma_zone_t zone;
1715 	uma_keg_t keg;
1716 
1717 	zone = (uma_zone_t)arg;
1718 	keg = zone_first_keg(zone);
1719 
1720 	if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
1721 		cache_drain(zone);
1722 
1723 	mtx_lock(&uma_mtx);
1724 	LIST_REMOVE(zone, uz_link);
1725 	mtx_unlock(&uma_mtx);
1726 	/*
1727 	 * XXX there are some races here where
1728 	 * the zone can be drained but zone lock
1729 	 * released and then refilled before we
1730 	 * remove it... we dont care for now
1731 	 */
1732 	zone_drain_wait(zone, M_WAITOK);
1733 	/*
1734 	 * Unlink all of our kegs.
1735 	 */
1736 	while ((klink = LIST_FIRST(&zone->uz_kegs)) != NULL) {
1737 		klink->kl_keg = NULL;
1738 		LIST_REMOVE(klink, kl_link);
1739 		if (klink == &zone->uz_klink)
1740 			continue;
1741 		free(klink, M_TEMP);
1742 	}
1743 	/*
1744 	 * We only destroy kegs from non secondary zones.
1745 	 */
1746 	if (keg != NULL && (zone->uz_flags & UMA_ZONE_SECONDARY) == 0)  {
1747 		mtx_lock(&uma_mtx);
1748 		LIST_REMOVE(keg, uk_link);
1749 		mtx_unlock(&uma_mtx);
1750 		zone_free_item(kegs, keg, NULL, SKIP_NONE);
1751 	}
1752 	ZONE_LOCK_FINI(zone);
1753 }
1754 
1755 /*
1756  * Traverses every zone in the system and calls a callback
1757  *
1758  * Arguments:
1759  *	zfunc  A pointer to a function which accepts a zone
1760  *		as an argument.
1761  *
1762  * Returns:
1763  *	Nothing
1764  */
1765 static void
1766 zone_foreach(void (*zfunc)(uma_zone_t))
1767 {
1768 	uma_keg_t keg;
1769 	uma_zone_t zone;
1770 
1771 	mtx_lock(&uma_mtx);
1772 	LIST_FOREACH(keg, &uma_kegs, uk_link) {
1773 		LIST_FOREACH(zone, &keg->uk_zones, uz_link)
1774 			zfunc(zone);
1775 	}
1776 	mtx_unlock(&uma_mtx);
1777 }
1778 
1779 /* Public functions */
1780 /* See uma.h */
1781 void
1782 uma_startup(void *bootmem, int boot_pages)
1783 {
1784 	struct uma_zctor_args args;
1785 	uma_slab_t slab;
1786 	u_int slabsize;
1787 	int i;
1788 
1789 #ifdef UMA_DEBUG
1790 	printf("Creating uma keg headers zone and keg.\n");
1791 #endif
1792 	mtx_init(&uma_mtx, "UMA lock", NULL, MTX_DEF);
1793 
1794 	/* "manually" create the initial zone */
1795 	memset(&args, 0, sizeof(args));
1796 	args.name = "UMA Kegs";
1797 	args.size = sizeof(struct uma_keg);
1798 	args.ctor = keg_ctor;
1799 	args.dtor = keg_dtor;
1800 	args.uminit = zero_init;
1801 	args.fini = NULL;
1802 	args.keg = &masterkeg;
1803 	args.align = 32 - 1;
1804 	args.flags = UMA_ZFLAG_INTERNAL;
1805 	/* The initial zone has no Per cpu queues so it's smaller */
1806 	zone_ctor(kegs, sizeof(struct uma_zone), &args, M_WAITOK);
1807 
1808 #ifdef UMA_DEBUG
1809 	printf("Filling boot free list.\n");
1810 #endif
1811 	for (i = 0; i < boot_pages; i++) {
1812 		slab = (uma_slab_t)((uint8_t *)bootmem + (i * UMA_SLAB_SIZE));
1813 		slab->us_data = (uint8_t *)slab;
1814 		slab->us_flags = UMA_SLAB_BOOT;
1815 		LIST_INSERT_HEAD(&uma_boot_pages, slab, us_link);
1816 	}
1817 	mtx_init(&uma_boot_pages_mtx, "UMA boot pages", NULL, MTX_DEF);
1818 
1819 #ifdef UMA_DEBUG
1820 	printf("Creating uma zone headers zone and keg.\n");
1821 #endif
1822 	args.name = "UMA Zones";
1823 	args.size = sizeof(struct uma_zone) +
1824 	    (sizeof(struct uma_cache) * (mp_maxid + 1));
1825 	args.ctor = zone_ctor;
1826 	args.dtor = zone_dtor;
1827 	args.uminit = zero_init;
1828 	args.fini = NULL;
1829 	args.keg = NULL;
1830 	args.align = 32 - 1;
1831 	args.flags = UMA_ZFLAG_INTERNAL;
1832 	/* The initial zone has no Per cpu queues so it's smaller */
1833 	zone_ctor(zones, sizeof(struct uma_zone), &args, M_WAITOK);
1834 
1835 #ifdef UMA_DEBUG
1836 	printf("Initializing pcpu cache locks.\n");
1837 #endif
1838 #ifdef UMA_DEBUG
1839 	printf("Creating slab and hash zones.\n");
1840 #endif
1841 
1842 	/* Now make a zone for slab headers */
1843 	slabzone = uma_zcreate("UMA Slabs",
1844 				sizeof(struct uma_slab),
1845 				NULL, NULL, NULL, NULL,
1846 				UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1847 
1848 	/*
1849 	 * We also create a zone for the bigger slabs with reference
1850 	 * counts in them, to accomodate UMA_ZONE_REFCNT zones.
1851 	 */
1852 	slabsize = sizeof(struct uma_slab_refcnt);
1853 	slabsize += uma_max_ipers_ref * sizeof(uint32_t);
1854 	slabrefzone = uma_zcreate("UMA RCntSlabs",
1855 				  slabsize,
1856 				  NULL, NULL, NULL, NULL,
1857 				  UMA_ALIGN_PTR,
1858 				  UMA_ZFLAG_INTERNAL);
1859 
1860 	hashzone = uma_zcreate("UMA Hash",
1861 	    sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
1862 	    NULL, NULL, NULL, NULL,
1863 	    UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1864 
1865 	bucket_init();
1866 
1867 	booted = UMA_STARTUP;
1868 
1869 #ifdef UMA_DEBUG
1870 	printf("UMA startup complete.\n");
1871 #endif
1872 }
1873 
1874 /* see uma.h */
1875 void
1876 uma_startup2(void)
1877 {
1878 	booted = UMA_STARTUP2;
1879 	bucket_enable();
1880 #ifdef UMA_DEBUG
1881 	printf("UMA startup2 complete.\n");
1882 #endif
1883 }
1884 
1885 /*
1886  * Initialize our callout handle
1887  *
1888  */
1889 
1890 static void
1891 uma_startup3(void)
1892 {
1893 #ifdef UMA_DEBUG
1894 	printf("Starting callout.\n");
1895 #endif
1896 	callout_init(&uma_callout, CALLOUT_MPSAFE);
1897 	callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
1898 #ifdef UMA_DEBUG
1899 	printf("UMA startup3 complete.\n");
1900 #endif
1901 }
1902 
1903 static uma_keg_t
1904 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
1905 		int align, uint32_t flags)
1906 {
1907 	struct uma_kctor_args args;
1908 
1909 	args.size = size;
1910 	args.uminit = uminit;
1911 	args.fini = fini;
1912 	args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
1913 	args.flags = flags;
1914 	args.zone = zone;
1915 	return (zone_alloc_item(kegs, &args, M_WAITOK));
1916 }
1917 
1918 /* See uma.h */
1919 void
1920 uma_set_align(int align)
1921 {
1922 
1923 	if (align != UMA_ALIGN_CACHE)
1924 		uma_align_cache = align;
1925 }
1926 
1927 /* See uma.h */
1928 uma_zone_t
1929 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
1930 		uma_init uminit, uma_fini fini, int align, uint32_t flags)
1931 
1932 {
1933 	struct uma_zctor_args args;
1934 
1935 	/* This stuff is essential for the zone ctor */
1936 	memset(&args, 0, sizeof(args));
1937 	args.name = name;
1938 	args.size = size;
1939 	args.ctor = ctor;
1940 	args.dtor = dtor;
1941 	args.uminit = uminit;
1942 	args.fini = fini;
1943 	args.align = align;
1944 	args.flags = flags;
1945 	args.keg = NULL;
1946 
1947 	return (zone_alloc_item(zones, &args, M_WAITOK));
1948 }
1949 
1950 /* See uma.h */
1951 uma_zone_t
1952 uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor,
1953 		    uma_init zinit, uma_fini zfini, uma_zone_t master)
1954 {
1955 	struct uma_zctor_args args;
1956 	uma_keg_t keg;
1957 
1958 	keg = zone_first_keg(master);
1959 	memset(&args, 0, sizeof(args));
1960 	args.name = name;
1961 	args.size = keg->uk_size;
1962 	args.ctor = ctor;
1963 	args.dtor = dtor;
1964 	args.uminit = zinit;
1965 	args.fini = zfini;
1966 	args.align = keg->uk_align;
1967 	args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
1968 	args.keg = keg;
1969 
1970 	/* XXX Attaches only one keg of potentially many. */
1971 	return (zone_alloc_item(zones, &args, M_WAITOK));
1972 }
1973 
1974 /* See uma.h */
1975 uma_zone_t
1976 uma_zcache_create(char *name, int size, uma_ctor ctor, uma_dtor dtor,
1977 		    uma_init zinit, uma_fini zfini, uma_import zimport,
1978 		    uma_release zrelease, void *arg, int flags)
1979 {
1980 	struct uma_zctor_args args;
1981 
1982 	memset(&args, 0, sizeof(args));
1983 	args.name = name;
1984 	args.size = size;
1985 	args.ctor = ctor;
1986 	args.dtor = dtor;
1987 	args.uminit = zinit;
1988 	args.fini = zfini;
1989 	args.import = zimport;
1990 	args.release = zrelease;
1991 	args.arg = arg;
1992 	args.align = 0;
1993 	args.flags = flags;
1994 
1995 	return (zone_alloc_item(zones, &args, M_WAITOK));
1996 }
1997 
1998 static void
1999 zone_lock_pair(uma_zone_t a, uma_zone_t b)
2000 {
2001 	if (a < b) {
2002 		ZONE_LOCK(a);
2003 		mtx_lock_flags(b->uz_lockptr, MTX_DUPOK);
2004 	} else {
2005 		ZONE_LOCK(b);
2006 		mtx_lock_flags(a->uz_lockptr, MTX_DUPOK);
2007 	}
2008 }
2009 
2010 static void
2011 zone_unlock_pair(uma_zone_t a, uma_zone_t b)
2012 {
2013 
2014 	ZONE_UNLOCK(a);
2015 	ZONE_UNLOCK(b);
2016 }
2017 
2018 int
2019 uma_zsecond_add(uma_zone_t zone, uma_zone_t master)
2020 {
2021 	uma_klink_t klink;
2022 	uma_klink_t kl;
2023 	int error;
2024 
2025 	error = 0;
2026 	klink = malloc(sizeof(*klink), M_TEMP, M_WAITOK | M_ZERO);
2027 
2028 	zone_lock_pair(zone, master);
2029 	/*
2030 	 * zone must use vtoslab() to resolve objects and must already be
2031 	 * a secondary.
2032 	 */
2033 	if ((zone->uz_flags & (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY))
2034 	    != (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) {
2035 		error = EINVAL;
2036 		goto out;
2037 	}
2038 	/*
2039 	 * The new master must also use vtoslab().
2040 	 */
2041 	if ((zone->uz_flags & UMA_ZONE_VTOSLAB) != UMA_ZONE_VTOSLAB) {
2042 		error = EINVAL;
2043 		goto out;
2044 	}
2045 	/*
2046 	 * Both must either be refcnt, or not be refcnt.
2047 	 */
2048 	if ((zone->uz_flags & UMA_ZONE_REFCNT) !=
2049 	    (master->uz_flags & UMA_ZONE_REFCNT)) {
2050 		error = EINVAL;
2051 		goto out;
2052 	}
2053 	/*
2054 	 * The underlying object must be the same size.  rsize
2055 	 * may be different.
2056 	 */
2057 	if (master->uz_size != zone->uz_size) {
2058 		error = E2BIG;
2059 		goto out;
2060 	}
2061 	/*
2062 	 * Put it at the end of the list.
2063 	 */
2064 	klink->kl_keg = zone_first_keg(master);
2065 	LIST_FOREACH(kl, &zone->uz_kegs, kl_link) {
2066 		if (LIST_NEXT(kl, kl_link) == NULL) {
2067 			LIST_INSERT_AFTER(kl, klink, kl_link);
2068 			break;
2069 		}
2070 	}
2071 	klink = NULL;
2072 	zone->uz_flags |= UMA_ZFLAG_MULTI;
2073 	zone->uz_slab = zone_fetch_slab_multi;
2074 
2075 out:
2076 	zone_unlock_pair(zone, master);
2077 	if (klink != NULL)
2078 		free(klink, M_TEMP);
2079 
2080 	return (error);
2081 }
2082 
2083 
2084 /* See uma.h */
2085 void
2086 uma_zdestroy(uma_zone_t zone)
2087 {
2088 
2089 	zone_free_item(zones, zone, NULL, SKIP_NONE);
2090 }
2091 
2092 /* See uma.h */
2093 void *
2094 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
2095 {
2096 	void *item;
2097 	uma_cache_t cache;
2098 	uma_bucket_t bucket;
2099 	int lockfail;
2100 	int cpu;
2101 
2102 	/* This is the fast path allocation */
2103 #ifdef UMA_DEBUG_ALLOC_1
2104 	printf("Allocating one item from %s(%p)\n", zone->uz_name, zone);
2105 #endif
2106 	CTR3(KTR_UMA, "uma_zalloc_arg thread %x zone %s flags %d", curthread,
2107 	    zone->uz_name, flags);
2108 
2109 	if (flags & M_WAITOK) {
2110 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
2111 		    "uma_zalloc_arg: zone \"%s\"", zone->uz_name);
2112 	}
2113 #ifdef DEBUG_MEMGUARD
2114 	if (memguard_cmp_zone(zone)) {
2115 		item = memguard_alloc(zone->uz_size, flags);
2116 		if (item != NULL) {
2117 			/*
2118 			 * Avoid conflict with the use-after-free
2119 			 * protecting infrastructure from INVARIANTS.
2120 			 */
2121 			if (zone->uz_init != NULL &&
2122 			    zone->uz_init != mtrash_init &&
2123 			    zone->uz_init(item, zone->uz_size, flags) != 0)
2124 				return (NULL);
2125 			if (zone->uz_ctor != NULL &&
2126 			    zone->uz_ctor != mtrash_ctor &&
2127 			    zone->uz_ctor(item, zone->uz_size, udata,
2128 			    flags) != 0) {
2129 			    	zone->uz_fini(item, zone->uz_size);
2130 				return (NULL);
2131 			}
2132 			return (item);
2133 		}
2134 		/* This is unfortunate but should not be fatal. */
2135 	}
2136 #endif
2137 	/*
2138 	 * If possible, allocate from the per-CPU cache.  There are two
2139 	 * requirements for safe access to the per-CPU cache: (1) the thread
2140 	 * accessing the cache must not be preempted or yield during access,
2141 	 * and (2) the thread must not migrate CPUs without switching which
2142 	 * cache it accesses.  We rely on a critical section to prevent
2143 	 * preemption and migration.  We release the critical section in
2144 	 * order to acquire the zone mutex if we are unable to allocate from
2145 	 * the current cache; when we re-acquire the critical section, we
2146 	 * must detect and handle migration if it has occurred.
2147 	 */
2148 	critical_enter();
2149 	cpu = curcpu;
2150 	cache = &zone->uz_cpu[cpu];
2151 
2152 zalloc_start:
2153 	bucket = cache->uc_allocbucket;
2154 	if (bucket != NULL && bucket->ub_cnt > 0) {
2155 		bucket->ub_cnt--;
2156 		item = bucket->ub_bucket[bucket->ub_cnt];
2157 #ifdef INVARIANTS
2158 		bucket->ub_bucket[bucket->ub_cnt] = NULL;
2159 #endif
2160 		KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled."));
2161 		cache->uc_allocs++;
2162 		critical_exit();
2163 		if (zone->uz_ctor != NULL &&
2164 		    zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2165 			atomic_add_long(&zone->uz_fails, 1);
2166 			zone_free_item(zone, item, udata, SKIP_DTOR);
2167 			return (NULL);
2168 		}
2169 #ifdef INVARIANTS
2170 		uma_dbg_alloc(zone, NULL, item);
2171 #endif
2172 		if (flags & M_ZERO)
2173 			uma_zero_item(item, zone);
2174 		return (item);
2175 	}
2176 
2177 	/*
2178 	 * We have run out of items in our alloc bucket.
2179 	 * See if we can switch with our free bucket.
2180 	 */
2181 	bucket = cache->uc_freebucket;
2182 	if (bucket != NULL && bucket->ub_cnt > 0) {
2183 #ifdef UMA_DEBUG_ALLOC
2184 		printf("uma_zalloc: Swapping empty with alloc.\n");
2185 #endif
2186 		cache->uc_freebucket = cache->uc_allocbucket;
2187 		cache->uc_allocbucket = bucket;
2188 		goto zalloc_start;
2189 	}
2190 
2191 	/*
2192 	 * Discard any empty allocation bucket while we hold no locks.
2193 	 */
2194 	bucket = cache->uc_allocbucket;
2195 	cache->uc_allocbucket = NULL;
2196 	critical_exit();
2197 	if (bucket != NULL)
2198 		bucket_free(zone, bucket, udata);
2199 
2200 	/* Short-circuit for zones without buckets and low memory. */
2201 	if (zone->uz_count == 0 || bucketdisable)
2202 		goto zalloc_item;
2203 
2204 	/*
2205 	 * Attempt to retrieve the item from the per-CPU cache has failed, so
2206 	 * we must go back to the zone.  This requires the zone lock, so we
2207 	 * must drop the critical section, then re-acquire it when we go back
2208 	 * to the cache.  Since the critical section is released, we may be
2209 	 * preempted or migrate.  As such, make sure not to maintain any
2210 	 * thread-local state specific to the cache from prior to releasing
2211 	 * the critical section.
2212 	 */
2213 	lockfail = 0;
2214 	if (ZONE_TRYLOCK(zone) == 0) {
2215 		/* Record contention to size the buckets. */
2216 		ZONE_LOCK(zone);
2217 		lockfail = 1;
2218 	}
2219 	critical_enter();
2220 	cpu = curcpu;
2221 	cache = &zone->uz_cpu[cpu];
2222 
2223 	/*
2224 	 * Since we have locked the zone we may as well send back our stats.
2225 	 */
2226 	atomic_add_long(&zone->uz_allocs, cache->uc_allocs);
2227 	atomic_add_long(&zone->uz_frees, cache->uc_frees);
2228 	cache->uc_allocs = 0;
2229 	cache->uc_frees = 0;
2230 
2231 	/* See if we lost the race to fill the cache. */
2232 	if (cache->uc_allocbucket != NULL) {
2233 		ZONE_UNLOCK(zone);
2234 		goto zalloc_start;
2235 	}
2236 
2237 	/*
2238 	 * Check the zone's cache of buckets.
2239 	 */
2240 	if ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) {
2241 		KASSERT(bucket->ub_cnt != 0,
2242 		    ("uma_zalloc_arg: Returning an empty bucket."));
2243 
2244 		LIST_REMOVE(bucket, ub_link);
2245 		cache->uc_allocbucket = bucket;
2246 		ZONE_UNLOCK(zone);
2247 		goto zalloc_start;
2248 	}
2249 	/* We are no longer associated with this CPU. */
2250 	critical_exit();
2251 
2252 	/*
2253 	 * We bump the uz count when the cache size is insufficient to
2254 	 * handle the working set.
2255 	 */
2256 	if (lockfail && zone->uz_count < BUCKET_MAX)
2257 		zone->uz_count++;
2258 	ZONE_UNLOCK(zone);
2259 
2260 	/*
2261 	 * Now lets just fill a bucket and put it on the free list.  If that
2262 	 * works we'll restart the allocation from the begining and it
2263 	 * will use the just filled bucket.
2264 	 */
2265 	bucket = zone_alloc_bucket(zone, udata, flags);
2266 	if (bucket != NULL) {
2267 		ZONE_LOCK(zone);
2268 		critical_enter();
2269 		cpu = curcpu;
2270 		cache = &zone->uz_cpu[cpu];
2271 		/*
2272 		 * See if we lost the race or were migrated.  Cache the
2273 		 * initialized bucket to make this less likely or claim
2274 		 * the memory directly.
2275 		 */
2276 		if (cache->uc_allocbucket == NULL)
2277 			cache->uc_allocbucket = bucket;
2278 		else
2279 			LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link);
2280 		ZONE_UNLOCK(zone);
2281 		goto zalloc_start;
2282 	}
2283 
2284 	/*
2285 	 * We may not be able to get a bucket so return an actual item.
2286 	 */
2287 #ifdef UMA_DEBUG
2288 	printf("uma_zalloc_arg: Bucketzone returned NULL\n");
2289 #endif
2290 
2291 zalloc_item:
2292 	item = zone_alloc_item(zone, udata, flags);
2293 
2294 	return (item);
2295 }
2296 
2297 static uma_slab_t
2298 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int flags)
2299 {
2300 	uma_slab_t slab;
2301 	int reserve;
2302 
2303 	mtx_assert(&keg->uk_lock, MA_OWNED);
2304 	slab = NULL;
2305 	reserve = 0;
2306 	if ((flags & M_USE_RESERVE) == 0)
2307 		reserve = keg->uk_reserve;
2308 
2309 	for (;;) {
2310 		/*
2311 		 * Find a slab with some space.  Prefer slabs that are partially
2312 		 * used over those that are totally full.  This helps to reduce
2313 		 * fragmentation.
2314 		 */
2315 		if (keg->uk_free > reserve) {
2316 			if (!LIST_EMPTY(&keg->uk_part_slab)) {
2317 				slab = LIST_FIRST(&keg->uk_part_slab);
2318 			} else {
2319 				slab = LIST_FIRST(&keg->uk_free_slab);
2320 				LIST_REMOVE(slab, us_link);
2321 				LIST_INSERT_HEAD(&keg->uk_part_slab, slab,
2322 				    us_link);
2323 			}
2324 			MPASS(slab->us_keg == keg);
2325 			return (slab);
2326 		}
2327 
2328 		/*
2329 		 * M_NOVM means don't ask at all!
2330 		 */
2331 		if (flags & M_NOVM)
2332 			break;
2333 
2334 		if (keg->uk_maxpages && keg->uk_pages >= keg->uk_maxpages) {
2335 			keg->uk_flags |= UMA_ZFLAG_FULL;
2336 			/*
2337 			 * If this is not a multi-zone, set the FULL bit.
2338 			 * Otherwise slab_multi() takes care of it.
2339 			 */
2340 			if ((zone->uz_flags & UMA_ZFLAG_MULTI) == 0) {
2341 				zone->uz_flags |= UMA_ZFLAG_FULL;
2342 				zone_log_warning(zone);
2343 			}
2344 			if (flags & M_NOWAIT)
2345 				break;
2346 			zone->uz_sleeps++;
2347 			msleep(keg, &keg->uk_lock, PVM, "keglimit", 0);
2348 			continue;
2349 		}
2350 		slab = keg_alloc_slab(keg, zone, flags);
2351 		/*
2352 		 * If we got a slab here it's safe to mark it partially used
2353 		 * and return.  We assume that the caller is going to remove
2354 		 * at least one item.
2355 		 */
2356 		if (slab) {
2357 			MPASS(slab->us_keg == keg);
2358 			LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2359 			return (slab);
2360 		}
2361 		/*
2362 		 * We might not have been able to get a slab but another cpu
2363 		 * could have while we were unlocked.  Check again before we
2364 		 * fail.
2365 		 */
2366 		flags |= M_NOVM;
2367 	}
2368 	return (slab);
2369 }
2370 
2371 static uma_slab_t
2372 zone_fetch_slab(uma_zone_t zone, uma_keg_t keg, int flags)
2373 {
2374 	uma_slab_t slab;
2375 
2376 	if (keg == NULL) {
2377 		keg = zone_first_keg(zone);
2378 		KEG_LOCK(keg);
2379 	}
2380 
2381 	for (;;) {
2382 		slab = keg_fetch_slab(keg, zone, flags);
2383 		if (slab)
2384 			return (slab);
2385 		if (flags & (M_NOWAIT | M_NOVM))
2386 			break;
2387 	}
2388 	KEG_UNLOCK(keg);
2389 	return (NULL);
2390 }
2391 
2392 /*
2393  * uma_zone_fetch_slab_multi:  Fetches a slab from one available keg.  Returns
2394  * with the keg locked.  On NULL no lock is held.
2395  *
2396  * The last pointer is used to seed the search.  It is not required.
2397  */
2398 static uma_slab_t
2399 zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int rflags)
2400 {
2401 	uma_klink_t klink;
2402 	uma_slab_t slab;
2403 	uma_keg_t keg;
2404 	int flags;
2405 	int empty;
2406 	int full;
2407 
2408 	/*
2409 	 * Don't wait on the first pass.  This will skip limit tests
2410 	 * as well.  We don't want to block if we can find a provider
2411 	 * without blocking.
2412 	 */
2413 	flags = (rflags & ~M_WAITOK) | M_NOWAIT;
2414 	/*
2415 	 * Use the last slab allocated as a hint for where to start
2416 	 * the search.
2417 	 */
2418 	if (last != NULL) {
2419 		slab = keg_fetch_slab(last, zone, flags);
2420 		if (slab)
2421 			return (slab);
2422 		KEG_UNLOCK(last);
2423 	}
2424 	/*
2425 	 * Loop until we have a slab incase of transient failures
2426 	 * while M_WAITOK is specified.  I'm not sure this is 100%
2427 	 * required but we've done it for so long now.
2428 	 */
2429 	for (;;) {
2430 		empty = 0;
2431 		full = 0;
2432 		/*
2433 		 * Search the available kegs for slabs.  Be careful to hold the
2434 		 * correct lock while calling into the keg layer.
2435 		 */
2436 		LIST_FOREACH(klink, &zone->uz_kegs, kl_link) {
2437 			keg = klink->kl_keg;
2438 			KEG_LOCK(keg);
2439 			if ((keg->uk_flags & UMA_ZFLAG_FULL) == 0) {
2440 				slab = keg_fetch_slab(keg, zone, flags);
2441 				if (slab)
2442 					return (slab);
2443 			}
2444 			if (keg->uk_flags & UMA_ZFLAG_FULL)
2445 				full++;
2446 			else
2447 				empty++;
2448 			KEG_UNLOCK(keg);
2449 		}
2450 		if (rflags & (M_NOWAIT | M_NOVM))
2451 			break;
2452 		flags = rflags;
2453 		/*
2454 		 * All kegs are full.  XXX We can't atomically check all kegs
2455 		 * and sleep so just sleep for a short period and retry.
2456 		 */
2457 		if (full && !empty) {
2458 			ZONE_LOCK(zone);
2459 			zone->uz_flags |= UMA_ZFLAG_FULL;
2460 			zone->uz_sleeps++;
2461 			zone_log_warning(zone);
2462 			msleep(zone, zone->uz_lockptr, PVM,
2463 			    "zonelimit", hz/100);
2464 			zone->uz_flags &= ~UMA_ZFLAG_FULL;
2465 			ZONE_UNLOCK(zone);
2466 			continue;
2467 		}
2468 	}
2469 	return (NULL);
2470 }
2471 
2472 static void *
2473 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
2474 {
2475 	void *item;
2476 	uint8_t freei;
2477 
2478 	MPASS(keg == slab->us_keg);
2479 	mtx_assert(&keg->uk_lock, MA_OWNED);
2480 
2481 	freei = BIT_FFS(SLAB_SETSIZE, &slab->us_free) - 1;
2482 	BIT_CLR(SLAB_SETSIZE, freei, &slab->us_free);
2483 	item = slab->us_data + (keg->uk_rsize * freei);
2484 	slab->us_freecount--;
2485 	keg->uk_free--;
2486 
2487 	/* Move this slab to the full list */
2488 	if (slab->us_freecount == 0) {
2489 		LIST_REMOVE(slab, us_link);
2490 		LIST_INSERT_HEAD(&keg->uk_full_slab, slab, us_link);
2491 	}
2492 
2493 	return (item);
2494 }
2495 
2496 static int
2497 zone_import(uma_zone_t zone, void **bucket, int max, int flags)
2498 {
2499 	uma_slab_t slab;
2500 	uma_keg_t keg;
2501 	int i;
2502 
2503 	slab = NULL;
2504 	keg = NULL;
2505 	/* Try to keep the buckets totally full */
2506 	for (i = 0; i < max; ) {
2507 		if ((slab = zone->uz_slab(zone, keg, flags)) == NULL)
2508 			break;
2509 		keg = slab->us_keg;
2510 		while (slab->us_freecount && i < max) {
2511 			bucket[i++] = slab_alloc_item(keg, slab);
2512 			if (keg->uk_free <= keg->uk_reserve)
2513 				break;
2514 		}
2515 		/* Don't grab more than one slab at a time. */
2516 		flags &= ~M_WAITOK;
2517 		flags |= M_NOWAIT;
2518 	}
2519 	if (slab != NULL)
2520 		KEG_UNLOCK(keg);
2521 
2522 	return i;
2523 }
2524 
2525 static uma_bucket_t
2526 zone_alloc_bucket(uma_zone_t zone, void *udata, int flags)
2527 {
2528 	uma_bucket_t bucket;
2529 	int max;
2530 
2531 	/* Don't wait for buckets, preserve caller's NOVM setting. */
2532 	bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
2533 	if (bucket == NULL)
2534 		return (NULL);
2535 
2536 	max = MIN(bucket->ub_entries, zone->uz_count);
2537 	bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
2538 	    max, flags);
2539 
2540 	/*
2541 	 * Initialize the memory if necessary.
2542 	 */
2543 	if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
2544 		int i;
2545 
2546 		for (i = 0; i < bucket->ub_cnt; i++)
2547 			if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size,
2548 			    flags) != 0)
2549 				break;
2550 		/*
2551 		 * If we couldn't initialize the whole bucket, put the
2552 		 * rest back onto the freelist.
2553 		 */
2554 		if (i != bucket->ub_cnt) {
2555 			zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
2556 			    bucket->ub_cnt - i);
2557 #ifdef INVARIANTS
2558 			bzero(&bucket->ub_bucket[i],
2559 			    sizeof(void *) * (bucket->ub_cnt - i));
2560 #endif
2561 			bucket->ub_cnt = i;
2562 		}
2563 	}
2564 
2565 	if (bucket->ub_cnt == 0) {
2566 		bucket_free(zone, bucket, udata);
2567 		atomic_add_long(&zone->uz_fails, 1);
2568 		return (NULL);
2569 	}
2570 
2571 	return (bucket);
2572 }
2573 
2574 /*
2575  * Allocates a single item from a zone.
2576  *
2577  * Arguments
2578  *	zone   The zone to alloc for.
2579  *	udata  The data to be passed to the constructor.
2580  *	flags  M_WAITOK, M_NOWAIT, M_ZERO.
2581  *
2582  * Returns
2583  *	NULL if there is no memory and M_NOWAIT is set
2584  *	An item if successful
2585  */
2586 
2587 static void *
2588 zone_alloc_item(uma_zone_t zone, void *udata, int flags)
2589 {
2590 	void *item;
2591 
2592 	item = NULL;
2593 
2594 #ifdef UMA_DEBUG_ALLOC
2595 	printf("INTERNAL: Allocating one item from %s(%p)\n", zone->uz_name, zone);
2596 #endif
2597 	if (zone->uz_import(zone->uz_arg, &item, 1, flags) != 1)
2598 		goto fail;
2599 	atomic_add_long(&zone->uz_allocs, 1);
2600 
2601 	/*
2602 	 * We have to call both the zone's init (not the keg's init)
2603 	 * and the zone's ctor.  This is because the item is going from
2604 	 * a keg slab directly to the user, and the user is expecting it
2605 	 * to be both zone-init'd as well as zone-ctor'd.
2606 	 */
2607 	if (zone->uz_init != NULL) {
2608 		if (zone->uz_init(item, zone->uz_size, flags) != 0) {
2609 			zone_free_item(zone, item, udata, SKIP_FINI);
2610 			goto fail;
2611 		}
2612 	}
2613 	if (zone->uz_ctor != NULL) {
2614 		if (zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) {
2615 			zone_free_item(zone, item, udata, SKIP_DTOR);
2616 			goto fail;
2617 		}
2618 	}
2619 #ifdef INVARIANTS
2620 	uma_dbg_alloc(zone, NULL, item);
2621 #endif
2622 	if (flags & M_ZERO)
2623 		uma_zero_item(item, zone);
2624 
2625 	return (item);
2626 
2627 fail:
2628 	atomic_add_long(&zone->uz_fails, 1);
2629 	return (NULL);
2630 }
2631 
2632 /* See uma.h */
2633 void
2634 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
2635 {
2636 	uma_cache_t cache;
2637 	uma_bucket_t bucket;
2638 	int lockfail;
2639 	int cpu;
2640 
2641 #ifdef UMA_DEBUG_ALLOC_1
2642 	printf("Freeing item %p to %s(%p)\n", item, zone->uz_name, zone);
2643 #endif
2644 	CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread,
2645 	    zone->uz_name);
2646 
2647         /* uma_zfree(..., NULL) does nothing, to match free(9). */
2648         if (item == NULL)
2649                 return;
2650 #ifdef DEBUG_MEMGUARD
2651 	if (is_memguard_addr(item)) {
2652 		if (zone->uz_dtor != NULL && zone->uz_dtor != mtrash_dtor)
2653 			zone->uz_dtor(item, zone->uz_size, udata);
2654 		if (zone->uz_fini != NULL && zone->uz_fini != mtrash_fini)
2655 			zone->uz_fini(item, zone->uz_size);
2656 		memguard_free(item);
2657 		return;
2658 	}
2659 #endif
2660 #ifdef INVARIANTS
2661 	if (zone->uz_flags & UMA_ZONE_MALLOC)
2662 		uma_dbg_free(zone, udata, item);
2663 	else
2664 		uma_dbg_free(zone, NULL, item);
2665 #endif
2666 	if (zone->uz_dtor != NULL)
2667 		zone->uz_dtor(item, zone->uz_size, udata);
2668 
2669 	/*
2670 	 * The race here is acceptable.  If we miss it we'll just have to wait
2671 	 * a little longer for the limits to be reset.
2672 	 */
2673 	if (zone->uz_flags & UMA_ZFLAG_FULL)
2674 		goto zfree_item;
2675 
2676 	/*
2677 	 * If possible, free to the per-CPU cache.  There are two
2678 	 * requirements for safe access to the per-CPU cache: (1) the thread
2679 	 * accessing the cache must not be preempted or yield during access,
2680 	 * and (2) the thread must not migrate CPUs without switching which
2681 	 * cache it accesses.  We rely on a critical section to prevent
2682 	 * preemption and migration.  We release the critical section in
2683 	 * order to acquire the zone mutex if we are unable to free to the
2684 	 * current cache; when we re-acquire the critical section, we must
2685 	 * detect and handle migration if it has occurred.
2686 	 */
2687 zfree_restart:
2688 	critical_enter();
2689 	cpu = curcpu;
2690 	cache = &zone->uz_cpu[cpu];
2691 
2692 zfree_start:
2693 	/*
2694 	 * Try to free into the allocbucket first to give LIFO ordering
2695 	 * for cache-hot datastructures.  Spill over into the freebucket
2696 	 * if necessary.  Alloc will swap them if one runs dry.
2697 	 */
2698 	bucket = cache->uc_allocbucket;
2699 	if (bucket == NULL || bucket->ub_cnt >= bucket->ub_entries)
2700 		bucket = cache->uc_freebucket;
2701 	if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) {
2702 		KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL,
2703 		    ("uma_zfree: Freeing to non free bucket index."));
2704 		bucket->ub_bucket[bucket->ub_cnt] = item;
2705 		bucket->ub_cnt++;
2706 		cache->uc_frees++;
2707 		critical_exit();
2708 		return;
2709 	}
2710 
2711 	/*
2712 	 * We must go back the zone, which requires acquiring the zone lock,
2713 	 * which in turn means we must release and re-acquire the critical
2714 	 * section.  Since the critical section is released, we may be
2715 	 * preempted or migrate.  As such, make sure not to maintain any
2716 	 * thread-local state specific to the cache from prior to releasing
2717 	 * the critical section.
2718 	 */
2719 	critical_exit();
2720 	if (zone->uz_count == 0 || bucketdisable)
2721 		goto zfree_item;
2722 
2723 	lockfail = 0;
2724 	if (ZONE_TRYLOCK(zone) == 0) {
2725 		/* Record contention to size the buckets. */
2726 		ZONE_LOCK(zone);
2727 		lockfail = 1;
2728 	}
2729 	critical_enter();
2730 	cpu = curcpu;
2731 	cache = &zone->uz_cpu[cpu];
2732 
2733 	/*
2734 	 * Since we have locked the zone we may as well send back our stats.
2735 	 */
2736 	atomic_add_long(&zone->uz_allocs, cache->uc_allocs);
2737 	atomic_add_long(&zone->uz_frees, cache->uc_frees);
2738 	cache->uc_allocs = 0;
2739 	cache->uc_frees = 0;
2740 
2741 	bucket = cache->uc_freebucket;
2742 	if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) {
2743 		ZONE_UNLOCK(zone);
2744 		goto zfree_start;
2745 	}
2746 	cache->uc_freebucket = NULL;
2747 
2748 	/* Can we throw this on the zone full list? */
2749 	if (bucket != NULL) {
2750 #ifdef UMA_DEBUG_ALLOC
2751 		printf("uma_zfree: Putting old bucket on the free list.\n");
2752 #endif
2753 		/* ub_cnt is pointing to the last free item */
2754 		KASSERT(bucket->ub_cnt != 0,
2755 		    ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n"));
2756 		LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link);
2757 	}
2758 
2759 	/* We are no longer associated with this CPU. */
2760 	critical_exit();
2761 
2762 	/*
2763 	 * We bump the uz count when the cache size is insufficient to
2764 	 * handle the working set.
2765 	 */
2766 	if (lockfail && zone->uz_count < BUCKET_MAX)
2767 		zone->uz_count++;
2768 	ZONE_UNLOCK(zone);
2769 
2770 #ifdef UMA_DEBUG_ALLOC
2771 	printf("uma_zfree: Allocating new free bucket.\n");
2772 #endif
2773 	bucket = bucket_alloc(zone, udata, M_NOWAIT);
2774 	if (bucket) {
2775 		critical_enter();
2776 		cpu = curcpu;
2777 		cache = &zone->uz_cpu[cpu];
2778 		if (cache->uc_freebucket == NULL) {
2779 			cache->uc_freebucket = bucket;
2780 			goto zfree_start;
2781 		}
2782 		/*
2783 		 * We lost the race, start over.  We have to drop our
2784 		 * critical section to free the bucket.
2785 		 */
2786 		critical_exit();
2787 		bucket_free(zone, bucket, udata);
2788 		goto zfree_restart;
2789 	}
2790 
2791 	/*
2792 	 * If nothing else caught this, we'll just do an internal free.
2793 	 */
2794 zfree_item:
2795 	zone_free_item(zone, item, udata, SKIP_DTOR);
2796 
2797 	return;
2798 }
2799 
2800 static void
2801 slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item)
2802 {
2803 	uint8_t freei;
2804 
2805 	mtx_assert(&keg->uk_lock, MA_OWNED);
2806 	MPASS(keg == slab->us_keg);
2807 
2808 	/* Do we need to remove from any lists? */
2809 	if (slab->us_freecount+1 == keg->uk_ipers) {
2810 		LIST_REMOVE(slab, us_link);
2811 		LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
2812 	} else if (slab->us_freecount == 0) {
2813 		LIST_REMOVE(slab, us_link);
2814 		LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2815 	}
2816 
2817 	/* Slab management. */
2818 	freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
2819 	BIT_SET(SLAB_SETSIZE, freei, &slab->us_free);
2820 	slab->us_freecount++;
2821 
2822 	/* Keg statistics. */
2823 	keg->uk_free++;
2824 }
2825 
2826 static void
2827 zone_release(uma_zone_t zone, void **bucket, int cnt)
2828 {
2829 	void *item;
2830 	uma_slab_t slab;
2831 	uma_keg_t keg;
2832 	uint8_t *mem;
2833 	int clearfull;
2834 	int i;
2835 
2836 	clearfull = 0;
2837 	keg = zone_first_keg(zone);
2838 	KEG_LOCK(keg);
2839 	for (i = 0; i < cnt; i++) {
2840 		item = bucket[i];
2841 		if (!(zone->uz_flags & UMA_ZONE_VTOSLAB)) {
2842 			mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
2843 			if (zone->uz_flags & UMA_ZONE_HASH) {
2844 				slab = hash_sfind(&keg->uk_hash, mem);
2845 			} else {
2846 				mem += keg->uk_pgoff;
2847 				slab = (uma_slab_t)mem;
2848 			}
2849 		} else {
2850 			slab = vtoslab((vm_offset_t)item);
2851 			if (slab->us_keg != keg) {
2852 				KEG_UNLOCK(keg);
2853 				keg = slab->us_keg;
2854 				KEG_LOCK(keg);
2855 			}
2856 		}
2857 		slab_free_item(keg, slab, item);
2858 		if (keg->uk_flags & UMA_ZFLAG_FULL) {
2859 			if (keg->uk_pages < keg->uk_maxpages) {
2860 				keg->uk_flags &= ~UMA_ZFLAG_FULL;
2861 				clearfull = 1;
2862 			}
2863 
2864 			/*
2865 			 * We can handle one more allocation. Since we're
2866 			 * clearing ZFLAG_FULL, wake up all procs blocked
2867 			 * on pages. This should be uncommon, so keeping this
2868 			 * simple for now (rather than adding count of blocked
2869 			 * threads etc).
2870 			 */
2871 			wakeup(keg);
2872 		}
2873 	}
2874 	KEG_UNLOCK(keg);
2875 	if (clearfull) {
2876 		ZONE_LOCK(zone);
2877 		zone->uz_flags &= ~UMA_ZFLAG_FULL;
2878 		wakeup(zone);
2879 		ZONE_UNLOCK(zone);
2880 	}
2881 
2882 }
2883 
2884 /*
2885  * Frees a single item to any zone.
2886  *
2887  * Arguments:
2888  *	zone   The zone to free to
2889  *	item   The item we're freeing
2890  *	udata  User supplied data for the dtor
2891  *	skip   Skip dtors and finis
2892  */
2893 static void
2894 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
2895 {
2896 
2897 #ifdef INVARIANTS
2898 	if (skip == SKIP_NONE) {
2899 		if (zone->uz_flags & UMA_ZONE_MALLOC)
2900 			uma_dbg_free(zone, udata, item);
2901 		else
2902 			uma_dbg_free(zone, NULL, item);
2903 	}
2904 #endif
2905 	if (skip < SKIP_DTOR && zone->uz_dtor)
2906 		zone->uz_dtor(item, zone->uz_size, udata);
2907 
2908 	if (skip < SKIP_FINI && zone->uz_fini)
2909 		zone->uz_fini(item, zone->uz_size);
2910 
2911 	atomic_add_long(&zone->uz_frees, 1);
2912 	zone->uz_release(zone->uz_arg, &item, 1);
2913 }
2914 
2915 /* See uma.h */
2916 int
2917 uma_zone_set_max(uma_zone_t zone, int nitems)
2918 {
2919 	uma_keg_t keg;
2920 
2921 	keg = zone_first_keg(zone);
2922 	if (keg == NULL)
2923 		return (0);
2924 	KEG_LOCK(keg);
2925 	keg->uk_maxpages = (nitems / keg->uk_ipers) * keg->uk_ppera;
2926 	if (keg->uk_maxpages * keg->uk_ipers < nitems)
2927 		keg->uk_maxpages += keg->uk_ppera;
2928 	nitems = keg->uk_maxpages * keg->uk_ipers;
2929 	KEG_UNLOCK(keg);
2930 
2931 	return (nitems);
2932 }
2933 
2934 /* See uma.h */
2935 int
2936 uma_zone_get_max(uma_zone_t zone)
2937 {
2938 	int nitems;
2939 	uma_keg_t keg;
2940 
2941 	keg = zone_first_keg(zone);
2942 	if (keg == NULL)
2943 		return (0);
2944 	KEG_LOCK(keg);
2945 	nitems = keg->uk_maxpages * keg->uk_ipers;
2946 	KEG_UNLOCK(keg);
2947 
2948 	return (nitems);
2949 }
2950 
2951 /* See uma.h */
2952 void
2953 uma_zone_set_warning(uma_zone_t zone, const char *warning)
2954 {
2955 
2956 	ZONE_LOCK(zone);
2957 	zone->uz_warning = warning;
2958 	ZONE_UNLOCK(zone);
2959 }
2960 
2961 /* See uma.h */
2962 int
2963 uma_zone_get_cur(uma_zone_t zone)
2964 {
2965 	int64_t nitems;
2966 	u_int i;
2967 
2968 	ZONE_LOCK(zone);
2969 	nitems = zone->uz_allocs - zone->uz_frees;
2970 	CPU_FOREACH(i) {
2971 		/*
2972 		 * See the comment in sysctl_vm_zone_stats() regarding the
2973 		 * safety of accessing the per-cpu caches. With the zone lock
2974 		 * held, it is safe, but can potentially result in stale data.
2975 		 */
2976 		nitems += zone->uz_cpu[i].uc_allocs -
2977 		    zone->uz_cpu[i].uc_frees;
2978 	}
2979 	ZONE_UNLOCK(zone);
2980 
2981 	return (nitems < 0 ? 0 : nitems);
2982 }
2983 
2984 /* See uma.h */
2985 void
2986 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
2987 {
2988 	uma_keg_t keg;
2989 
2990 	keg = zone_first_keg(zone);
2991 	KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type"));
2992 	KEG_LOCK(keg);
2993 	KASSERT(keg->uk_pages == 0,
2994 	    ("uma_zone_set_init on non-empty keg"));
2995 	keg->uk_init = uminit;
2996 	KEG_UNLOCK(keg);
2997 }
2998 
2999 /* See uma.h */
3000 void
3001 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
3002 {
3003 	uma_keg_t keg;
3004 
3005 	keg = zone_first_keg(zone);
3006 	KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type"));
3007 	KEG_LOCK(keg);
3008 	KASSERT(keg->uk_pages == 0,
3009 	    ("uma_zone_set_fini on non-empty keg"));
3010 	keg->uk_fini = fini;
3011 	KEG_UNLOCK(keg);
3012 }
3013 
3014 /* See uma.h */
3015 void
3016 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
3017 {
3018 
3019 	ZONE_LOCK(zone);
3020 	KASSERT(zone_first_keg(zone)->uk_pages == 0,
3021 	    ("uma_zone_set_zinit on non-empty keg"));
3022 	zone->uz_init = zinit;
3023 	ZONE_UNLOCK(zone);
3024 }
3025 
3026 /* See uma.h */
3027 void
3028 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
3029 {
3030 
3031 	ZONE_LOCK(zone);
3032 	KASSERT(zone_first_keg(zone)->uk_pages == 0,
3033 	    ("uma_zone_set_zfini on non-empty keg"));
3034 	zone->uz_fini = zfini;
3035 	ZONE_UNLOCK(zone);
3036 }
3037 
3038 /* See uma.h */
3039 /* XXX uk_freef is not actually used with the zone locked */
3040 void
3041 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
3042 {
3043 	uma_keg_t keg;
3044 
3045 	keg = zone_first_keg(zone);
3046 	KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type"));
3047 	KEG_LOCK(keg);
3048 	keg->uk_freef = freef;
3049 	KEG_UNLOCK(keg);
3050 }
3051 
3052 /* See uma.h */
3053 /* XXX uk_allocf is not actually used with the zone locked */
3054 void
3055 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
3056 {
3057 	uma_keg_t keg;
3058 
3059 	keg = zone_first_keg(zone);
3060 	KEG_LOCK(keg);
3061 	keg->uk_allocf = allocf;
3062 	KEG_UNLOCK(keg);
3063 }
3064 
3065 /* See uma.h */
3066 void
3067 uma_zone_reserve(uma_zone_t zone, int items)
3068 {
3069 	uma_keg_t keg;
3070 
3071 	keg = zone_first_keg(zone);
3072 	if (keg == NULL)
3073 		return;
3074 	KEG_LOCK(keg);
3075 	keg->uk_reserve = items;
3076 	KEG_UNLOCK(keg);
3077 
3078 	return;
3079 }
3080 
3081 /* See uma.h */
3082 int
3083 uma_zone_reserve_kva(uma_zone_t zone, int count)
3084 {
3085 	uma_keg_t keg;
3086 	vm_offset_t kva;
3087 	int pages;
3088 
3089 	keg = zone_first_keg(zone);
3090 	if (keg == NULL)
3091 		return (0);
3092 	pages = count / keg->uk_ipers;
3093 
3094 	if (pages * keg->uk_ipers < count)
3095 		pages++;
3096 
3097 #ifdef UMA_MD_SMALL_ALLOC
3098 	if (keg->uk_ppera > 1) {
3099 #else
3100 	if (1) {
3101 #endif
3102 		kva = kva_alloc(pages * UMA_SLAB_SIZE);
3103 		if (kva == 0)
3104 			return (0);
3105 	} else
3106 		kva = 0;
3107 	KEG_LOCK(keg);
3108 	keg->uk_kva = kva;
3109 	keg->uk_offset = 0;
3110 	keg->uk_maxpages = pages;
3111 #ifdef UMA_MD_SMALL_ALLOC
3112 	keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
3113 #else
3114 	keg->uk_allocf = noobj_alloc;
3115 #endif
3116 	keg->uk_flags |= UMA_ZONE_NOFREE;
3117 	KEG_UNLOCK(keg);
3118 
3119 	return (1);
3120 }
3121 
3122 /* See uma.h */
3123 void
3124 uma_prealloc(uma_zone_t zone, int items)
3125 {
3126 	int slabs;
3127 	uma_slab_t slab;
3128 	uma_keg_t keg;
3129 
3130 	keg = zone_first_keg(zone);
3131 	if (keg == NULL)
3132 		return;
3133 	KEG_LOCK(keg);
3134 	slabs = items / keg->uk_ipers;
3135 	if (slabs * keg->uk_ipers < items)
3136 		slabs++;
3137 	while (slabs > 0) {
3138 		slab = keg_alloc_slab(keg, zone, M_WAITOK);
3139 		if (slab == NULL)
3140 			break;
3141 		MPASS(slab->us_keg == keg);
3142 		LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
3143 		slabs--;
3144 	}
3145 	KEG_UNLOCK(keg);
3146 }
3147 
3148 /* See uma.h */
3149 uint32_t *
3150 uma_find_refcnt(uma_zone_t zone, void *item)
3151 {
3152 	uma_slabrefcnt_t slabref;
3153 	uma_slab_t slab;
3154 	uma_keg_t keg;
3155 	uint32_t *refcnt;
3156 	int idx;
3157 
3158 	slab = vtoslab((vm_offset_t)item & (~UMA_SLAB_MASK));
3159 	slabref = (uma_slabrefcnt_t)slab;
3160 	keg = slab->us_keg;
3161 	KASSERT(keg->uk_flags & UMA_ZONE_REFCNT,
3162 	    ("uma_find_refcnt(): zone possibly not UMA_ZONE_REFCNT"));
3163 	idx = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize;
3164 	refcnt = &slabref->us_refcnt[idx];
3165 	return refcnt;
3166 }
3167 
3168 /* See uma.h */
3169 void
3170 uma_reclaim(void)
3171 {
3172 #ifdef UMA_DEBUG
3173 	printf("UMA: vm asked us to release pages!\n");
3174 #endif
3175 	bucket_enable();
3176 	zone_foreach(zone_drain);
3177 	if (vm_page_count_min()) {
3178 		cache_drain_safe(NULL);
3179 		zone_foreach(zone_drain);
3180 	}
3181 	/*
3182 	 * Some slabs may have been freed but this zone will be visited early
3183 	 * we visit again so that we can free pages that are empty once other
3184 	 * zones are drained.  We have to do the same for buckets.
3185 	 */
3186 	zone_drain(slabzone);
3187 	zone_drain(slabrefzone);
3188 	bucket_zone_drain();
3189 }
3190 
3191 /* See uma.h */
3192 int
3193 uma_zone_exhausted(uma_zone_t zone)
3194 {
3195 	int full;
3196 
3197 	ZONE_LOCK(zone);
3198 	full = (zone->uz_flags & UMA_ZFLAG_FULL);
3199 	ZONE_UNLOCK(zone);
3200 	return (full);
3201 }
3202 
3203 int
3204 uma_zone_exhausted_nolock(uma_zone_t zone)
3205 {
3206 	return (zone->uz_flags & UMA_ZFLAG_FULL);
3207 }
3208 
3209 void *
3210 uma_large_malloc(int size, int wait)
3211 {
3212 	void *mem;
3213 	uma_slab_t slab;
3214 	uint8_t flags;
3215 
3216 	slab = zone_alloc_item(slabzone, NULL, wait);
3217 	if (slab == NULL)
3218 		return (NULL);
3219 	mem = page_alloc(NULL, size, &flags, wait);
3220 	if (mem) {
3221 		vsetslab((vm_offset_t)mem, slab);
3222 		slab->us_data = mem;
3223 		slab->us_flags = flags | UMA_SLAB_MALLOC;
3224 		slab->us_size = size;
3225 	} else {
3226 		zone_free_item(slabzone, slab, NULL, SKIP_NONE);
3227 	}
3228 
3229 	return (mem);
3230 }
3231 
3232 void
3233 uma_large_free(uma_slab_t slab)
3234 {
3235 
3236 	page_free(slab->us_data, slab->us_size, slab->us_flags);
3237 	zone_free_item(slabzone, slab, NULL, SKIP_NONE);
3238 }
3239 
3240 static void
3241 uma_zero_item(void *item, uma_zone_t zone)
3242 {
3243 
3244 	if (zone->uz_flags & UMA_ZONE_PCPU) {
3245 		for (int i = 0; i < mp_ncpus; i++)
3246 			bzero(zpcpu_get_cpu(item, i), zone->uz_size);
3247 	} else
3248 		bzero(item, zone->uz_size);
3249 }
3250 
3251 void
3252 uma_print_stats(void)
3253 {
3254 	zone_foreach(uma_print_zone);
3255 }
3256 
3257 static void
3258 slab_print(uma_slab_t slab)
3259 {
3260 	printf("slab: keg %p, data %p, freecount %d\n",
3261 		slab->us_keg, slab->us_data, slab->us_freecount);
3262 }
3263 
3264 static void
3265 cache_print(uma_cache_t cache)
3266 {
3267 	printf("alloc: %p(%d), free: %p(%d)\n",
3268 		cache->uc_allocbucket,
3269 		cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0,
3270 		cache->uc_freebucket,
3271 		cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0);
3272 }
3273 
3274 static void
3275 uma_print_keg(uma_keg_t keg)
3276 {
3277 	uma_slab_t slab;
3278 
3279 	printf("keg: %s(%p) size %d(%d) flags %#x ipers %d ppera %d "
3280 	    "out %d free %d limit %d\n",
3281 	    keg->uk_name, keg, keg->uk_size, keg->uk_rsize, keg->uk_flags,
3282 	    keg->uk_ipers, keg->uk_ppera,
3283 	    (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free,
3284 	    (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers);
3285 	printf("Part slabs:\n");
3286 	LIST_FOREACH(slab, &keg->uk_part_slab, us_link)
3287 		slab_print(slab);
3288 	printf("Free slabs:\n");
3289 	LIST_FOREACH(slab, &keg->uk_free_slab, us_link)
3290 		slab_print(slab);
3291 	printf("Full slabs:\n");
3292 	LIST_FOREACH(slab, &keg->uk_full_slab, us_link)
3293 		slab_print(slab);
3294 }
3295 
3296 void
3297 uma_print_zone(uma_zone_t zone)
3298 {
3299 	uma_cache_t cache;
3300 	uma_klink_t kl;
3301 	int i;
3302 
3303 	printf("zone: %s(%p) size %d flags %#x\n",
3304 	    zone->uz_name, zone, zone->uz_size, zone->uz_flags);
3305 	LIST_FOREACH(kl, &zone->uz_kegs, kl_link)
3306 		uma_print_keg(kl->kl_keg);
3307 	CPU_FOREACH(i) {
3308 		cache = &zone->uz_cpu[i];
3309 		printf("CPU %d Cache:\n", i);
3310 		cache_print(cache);
3311 	}
3312 }
3313 
3314 #ifdef DDB
3315 /*
3316  * Generate statistics across both the zone and its per-cpu cache's.  Return
3317  * desired statistics if the pointer is non-NULL for that statistic.
3318  *
3319  * Note: does not update the zone statistics, as it can't safely clear the
3320  * per-CPU cache statistic.
3321  *
3322  * XXXRW: Following the uc_allocbucket and uc_freebucket pointers here isn't
3323  * safe from off-CPU; we should modify the caches to track this information
3324  * directly so that we don't have to.
3325  */
3326 static void
3327 uma_zone_sumstat(uma_zone_t z, int *cachefreep, uint64_t *allocsp,
3328     uint64_t *freesp, uint64_t *sleepsp)
3329 {
3330 	uma_cache_t cache;
3331 	uint64_t allocs, frees, sleeps;
3332 	int cachefree, cpu;
3333 
3334 	allocs = frees = sleeps = 0;
3335 	cachefree = 0;
3336 	CPU_FOREACH(cpu) {
3337 		cache = &z->uz_cpu[cpu];
3338 		if (cache->uc_allocbucket != NULL)
3339 			cachefree += cache->uc_allocbucket->ub_cnt;
3340 		if (cache->uc_freebucket != NULL)
3341 			cachefree += cache->uc_freebucket->ub_cnt;
3342 		allocs += cache->uc_allocs;
3343 		frees += cache->uc_frees;
3344 	}
3345 	allocs += z->uz_allocs;
3346 	frees += z->uz_frees;
3347 	sleeps += z->uz_sleeps;
3348 	if (cachefreep != NULL)
3349 		*cachefreep = cachefree;
3350 	if (allocsp != NULL)
3351 		*allocsp = allocs;
3352 	if (freesp != NULL)
3353 		*freesp = frees;
3354 	if (sleepsp != NULL)
3355 		*sleepsp = sleeps;
3356 }
3357 #endif /* DDB */
3358 
3359 static int
3360 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
3361 {
3362 	uma_keg_t kz;
3363 	uma_zone_t z;
3364 	int count;
3365 
3366 	count = 0;
3367 	mtx_lock(&uma_mtx);
3368 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3369 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
3370 			count++;
3371 	}
3372 	mtx_unlock(&uma_mtx);
3373 	return (sysctl_handle_int(oidp, &count, 0, req));
3374 }
3375 
3376 static int
3377 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
3378 {
3379 	struct uma_stream_header ush;
3380 	struct uma_type_header uth;
3381 	struct uma_percpu_stat ups;
3382 	uma_bucket_t bucket;
3383 	struct sbuf sbuf;
3384 	uma_cache_t cache;
3385 	uma_klink_t kl;
3386 	uma_keg_t kz;
3387 	uma_zone_t z;
3388 	uma_keg_t k;
3389 	int count, error, i;
3390 
3391 	error = sysctl_wire_old_buffer(req, 0);
3392 	if (error != 0)
3393 		return (error);
3394 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
3395 
3396 	count = 0;
3397 	mtx_lock(&uma_mtx);
3398 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3399 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
3400 			count++;
3401 	}
3402 
3403 	/*
3404 	 * Insert stream header.
3405 	 */
3406 	bzero(&ush, sizeof(ush));
3407 	ush.ush_version = UMA_STREAM_VERSION;
3408 	ush.ush_maxcpus = (mp_maxid + 1);
3409 	ush.ush_count = count;
3410 	(void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
3411 
3412 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3413 		LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3414 			bzero(&uth, sizeof(uth));
3415 			ZONE_LOCK(z);
3416 			strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
3417 			uth.uth_align = kz->uk_align;
3418 			uth.uth_size = kz->uk_size;
3419 			uth.uth_rsize = kz->uk_rsize;
3420 			LIST_FOREACH(kl, &z->uz_kegs, kl_link) {
3421 				k = kl->kl_keg;
3422 				uth.uth_maxpages += k->uk_maxpages;
3423 				uth.uth_pages += k->uk_pages;
3424 				uth.uth_keg_free += k->uk_free;
3425 				uth.uth_limit = (k->uk_maxpages / k->uk_ppera)
3426 				    * k->uk_ipers;
3427 			}
3428 
3429 			/*
3430 			 * A zone is secondary is it is not the first entry
3431 			 * on the keg's zone list.
3432 			 */
3433 			if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
3434 			    (LIST_FIRST(&kz->uk_zones) != z))
3435 				uth.uth_zone_flags = UTH_ZONE_SECONDARY;
3436 
3437 			LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3438 				uth.uth_zone_free += bucket->ub_cnt;
3439 			uth.uth_allocs = z->uz_allocs;
3440 			uth.uth_frees = z->uz_frees;
3441 			uth.uth_fails = z->uz_fails;
3442 			uth.uth_sleeps = z->uz_sleeps;
3443 			(void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
3444 			/*
3445 			 * While it is not normally safe to access the cache
3446 			 * bucket pointers while not on the CPU that owns the
3447 			 * cache, we only allow the pointers to be exchanged
3448 			 * without the zone lock held, not invalidated, so
3449 			 * accept the possible race associated with bucket
3450 			 * exchange during monitoring.
3451 			 */
3452 			for (i = 0; i < (mp_maxid + 1); i++) {
3453 				bzero(&ups, sizeof(ups));
3454 				if (kz->uk_flags & UMA_ZFLAG_INTERNAL)
3455 					goto skip;
3456 				if (CPU_ABSENT(i))
3457 					goto skip;
3458 				cache = &z->uz_cpu[i];
3459 				if (cache->uc_allocbucket != NULL)
3460 					ups.ups_cache_free +=
3461 					    cache->uc_allocbucket->ub_cnt;
3462 				if (cache->uc_freebucket != NULL)
3463 					ups.ups_cache_free +=
3464 					    cache->uc_freebucket->ub_cnt;
3465 				ups.ups_allocs = cache->uc_allocs;
3466 				ups.ups_frees = cache->uc_frees;
3467 skip:
3468 				(void)sbuf_bcat(&sbuf, &ups, sizeof(ups));
3469 			}
3470 			ZONE_UNLOCK(z);
3471 		}
3472 	}
3473 	mtx_unlock(&uma_mtx);
3474 	error = sbuf_finish(&sbuf);
3475 	sbuf_delete(&sbuf);
3476 	return (error);
3477 }
3478 
3479 int
3480 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS)
3481 {
3482 	uma_zone_t zone = *(uma_zone_t *)arg1;
3483 	int error, max, old;
3484 
3485 	old = max = uma_zone_get_max(zone);
3486 	error = sysctl_handle_int(oidp, &max, 0, req);
3487 	if (error || !req->newptr)
3488 		return (error);
3489 
3490 	if (max < old)
3491 		return (EINVAL);
3492 
3493 	uma_zone_set_max(zone, max);
3494 
3495 	return (0);
3496 }
3497 
3498 int
3499 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS)
3500 {
3501 	uma_zone_t zone = *(uma_zone_t *)arg1;
3502 	int cur;
3503 
3504 	cur = uma_zone_get_cur(zone);
3505 	return (sysctl_handle_int(oidp, &cur, 0, req));
3506 }
3507 
3508 #ifdef DDB
3509 DB_SHOW_COMMAND(uma, db_show_uma)
3510 {
3511 	uint64_t allocs, frees, sleeps;
3512 	uma_bucket_t bucket;
3513 	uma_keg_t kz;
3514 	uma_zone_t z;
3515 	int cachefree;
3516 
3517 	db_printf("%18s %8s %8s %8s %12s %8s %8s\n", "Zone", "Size", "Used",
3518 	    "Free", "Requests", "Sleeps", "Bucket");
3519 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
3520 		LIST_FOREACH(z, &kz->uk_zones, uz_link) {
3521 			if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
3522 				allocs = z->uz_allocs;
3523 				frees = z->uz_frees;
3524 				sleeps = z->uz_sleeps;
3525 				cachefree = 0;
3526 			} else
3527 				uma_zone_sumstat(z, &cachefree, &allocs,
3528 				    &frees, &sleeps);
3529 			if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
3530 			    (LIST_FIRST(&kz->uk_zones) != z)))
3531 				cachefree += kz->uk_free;
3532 			LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3533 				cachefree += bucket->ub_cnt;
3534 			db_printf("%18s %8ju %8jd %8d %12ju %8ju %8u\n",
3535 			    z->uz_name, (uintmax_t)kz->uk_size,
3536 			    (intmax_t)(allocs - frees), cachefree,
3537 			    (uintmax_t)allocs, sleeps, z->uz_count);
3538 			if (db_pager_quit)
3539 				return;
3540 		}
3541 	}
3542 }
3543 
3544 DB_SHOW_COMMAND(umacache, db_show_umacache)
3545 {
3546 	uint64_t allocs, frees;
3547 	uma_bucket_t bucket;
3548 	uma_zone_t z;
3549 	int cachefree;
3550 
3551 	db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
3552 	    "Requests", "Bucket");
3553 	LIST_FOREACH(z, &uma_cachezones, uz_link) {
3554 		uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL);
3555 		LIST_FOREACH(bucket, &z->uz_buckets, ub_link)
3556 			cachefree += bucket->ub_cnt;
3557 		db_printf("%18s %8ju %8jd %8d %12ju %8u\n",
3558 		    z->uz_name, (uintmax_t)z->uz_size,
3559 		    (intmax_t)(allocs - frees), cachefree,
3560 		    (uintmax_t)allocs, z->uz_count);
3561 		if (db_pager_quit)
3562 			return;
3563 	}
3564 }
3565 #endif
3566