xref: /freebsd/sys/vm/uma_core.c (revision 732a02b4e77866604a120a275c082bb6221bd2ff)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2002-2019 Jeffrey Roberson <jeff@FreeBSD.org>
5  * Copyright (c) 2004, 2005 Bosko Milekic <bmilekic@FreeBSD.org>
6  * Copyright (c) 2004-2006 Robert N. M. Watson
7  * All rights reserved.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice unmodified, this list of conditions, and the following
14  *    disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
20  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
21  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
22  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
23  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
24  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
25  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
26  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
27  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
28  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
29  */
30 
31 /*
32  * uma_core.c  Implementation of the Universal Memory allocator
33  *
34  * This allocator is intended to replace the multitude of similar object caches
35  * in the standard FreeBSD kernel.  The intent is to be flexible as well as
36  * efficient.  A primary design goal is to return unused memory to the rest of
37  * the system.  This will make the system as a whole more flexible due to the
38  * ability to move memory to subsystems which most need it instead of leaving
39  * pools of reserved memory unused.
40  *
41  * The basic ideas stem from similar slab/zone based allocators whose algorithms
42  * are well known.
43  *
44  */
45 
46 /*
47  * TODO:
48  *	- Improve memory usage for large allocations
49  *	- Investigate cache size adjustments
50  */
51 
52 #include <sys/cdefs.h>
53 __FBSDID("$FreeBSD$");
54 
55 #include "opt_ddb.h"
56 #include "opt_param.h"
57 #include "opt_vm.h"
58 
59 #include <sys/param.h>
60 #include <sys/systm.h>
61 #include <sys/bitset.h>
62 #include <sys/domainset.h>
63 #include <sys/eventhandler.h>
64 #include <sys/kernel.h>
65 #include <sys/types.h>
66 #include <sys/limits.h>
67 #include <sys/queue.h>
68 #include <sys/malloc.h>
69 #include <sys/ktr.h>
70 #include <sys/lock.h>
71 #include <sys/sysctl.h>
72 #include <sys/mutex.h>
73 #include <sys/proc.h>
74 #include <sys/random.h>
75 #include <sys/rwlock.h>
76 #include <sys/sbuf.h>
77 #include <sys/sched.h>
78 #include <sys/sleepqueue.h>
79 #include <sys/smp.h>
80 #include <sys/smr.h>
81 #include <sys/taskqueue.h>
82 #include <sys/vmmeter.h>
83 
84 #include <vm/vm.h>
85 #include <vm/vm_domainset.h>
86 #include <vm/vm_object.h>
87 #include <vm/vm_page.h>
88 #include <vm/vm_pageout.h>
89 #include <vm/vm_param.h>
90 #include <vm/vm_phys.h>
91 #include <vm/vm_pagequeue.h>
92 #include <vm/vm_map.h>
93 #include <vm/vm_kern.h>
94 #include <vm/vm_extern.h>
95 #include <vm/uma.h>
96 #include <vm/uma_int.h>
97 #include <vm/uma_dbg.h>
98 
99 #include <ddb/ddb.h>
100 
101 #ifdef DEBUG_MEMGUARD
102 #include <vm/memguard.h>
103 #endif
104 
105 #include <machine/md_var.h>
106 
107 #ifdef INVARIANTS
108 #define	UMA_ALWAYS_CTORDTOR	1
109 #else
110 #define	UMA_ALWAYS_CTORDTOR	0
111 #endif
112 
113 /*
114  * This is the zone and keg from which all zones are spawned.
115  */
116 static uma_zone_t kegs;
117 static uma_zone_t zones;
118 
119 /*
120  * On INVARIANTS builds, the slab contains a second bitset of the same size,
121  * "dbg_bits", which is laid out immediately after us_free.
122  */
123 #ifdef INVARIANTS
124 #define	SLAB_BITSETS	2
125 #else
126 #define	SLAB_BITSETS	1
127 #endif
128 
129 /*
130  * These are the two zones from which all offpage uma_slab_ts are allocated.
131  *
132  * One zone is for slab headers that can represent a larger number of items,
133  * making the slabs themselves more efficient, and the other zone is for
134  * headers that are smaller and represent fewer items, making the headers more
135  * efficient.
136  */
137 #define	SLABZONE_SIZE(setsize)					\
138     (sizeof(struct uma_hash_slab) + BITSET_SIZE(setsize) * SLAB_BITSETS)
139 #define	SLABZONE0_SETSIZE	(PAGE_SIZE / 16)
140 #define	SLABZONE1_SETSIZE	SLAB_MAX_SETSIZE
141 #define	SLABZONE0_SIZE	SLABZONE_SIZE(SLABZONE0_SETSIZE)
142 #define	SLABZONE1_SIZE	SLABZONE_SIZE(SLABZONE1_SETSIZE)
143 static uma_zone_t slabzones[2];
144 
145 /*
146  * The initial hash tables come out of this zone so they can be allocated
147  * prior to malloc coming up.
148  */
149 static uma_zone_t hashzone;
150 
151 /* The boot-time adjusted value for cache line alignment. */
152 int uma_align_cache = 64 - 1;
153 
154 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets");
155 static MALLOC_DEFINE(M_UMA, "UMA", "UMA Misc");
156 
157 /*
158  * Are we allowed to allocate buckets?
159  */
160 static int bucketdisable = 1;
161 
162 /* Linked list of all kegs in the system */
163 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs);
164 
165 /* Linked list of all cache-only zones in the system */
166 static LIST_HEAD(,uma_zone) uma_cachezones =
167     LIST_HEAD_INITIALIZER(uma_cachezones);
168 
169 /* This RW lock protects the keg list */
170 static struct rwlock_padalign __exclusive_cache_line uma_rwlock;
171 
172 /*
173  * First available virual address for boot time allocations.
174  */
175 static vm_offset_t bootstart;
176 static vm_offset_t bootmem;
177 
178 static struct sx uma_reclaim_lock;
179 
180 /*
181  * kmem soft limit, initialized by uma_set_limit().  Ensure that early
182  * allocations don't trigger a wakeup of the reclaim thread.
183  */
184 unsigned long uma_kmem_limit = LONG_MAX;
185 SYSCTL_ULONG(_vm, OID_AUTO, uma_kmem_limit, CTLFLAG_RD, &uma_kmem_limit, 0,
186     "UMA kernel memory soft limit");
187 unsigned long uma_kmem_total;
188 SYSCTL_ULONG(_vm, OID_AUTO, uma_kmem_total, CTLFLAG_RD, &uma_kmem_total, 0,
189     "UMA kernel memory usage");
190 
191 /* Is the VM done starting up? */
192 static enum {
193 	BOOT_COLD,
194 	BOOT_KVA,
195 	BOOT_RUNNING,
196 	BOOT_SHUTDOWN,
197 } booted = BOOT_COLD;
198 
199 /*
200  * This is the handle used to schedule events that need to happen
201  * outside of the allocation fast path.
202  */
203 static struct callout uma_callout;
204 #define	UMA_TIMEOUT	20		/* Seconds for callout interval. */
205 
206 /*
207  * This structure is passed as the zone ctor arg so that I don't have to create
208  * a special allocation function just for zones.
209  */
210 struct uma_zctor_args {
211 	const char *name;
212 	size_t size;
213 	uma_ctor ctor;
214 	uma_dtor dtor;
215 	uma_init uminit;
216 	uma_fini fini;
217 	uma_import import;
218 	uma_release release;
219 	void *arg;
220 	uma_keg_t keg;
221 	int align;
222 	uint32_t flags;
223 };
224 
225 struct uma_kctor_args {
226 	uma_zone_t zone;
227 	size_t size;
228 	uma_init uminit;
229 	uma_fini fini;
230 	int align;
231 	uint32_t flags;
232 };
233 
234 struct uma_bucket_zone {
235 	uma_zone_t	ubz_zone;
236 	const char	*ubz_name;
237 	int		ubz_entries;	/* Number of items it can hold. */
238 	int		ubz_maxsize;	/* Maximum allocation size per-item. */
239 };
240 
241 /*
242  * Compute the actual number of bucket entries to pack them in power
243  * of two sizes for more efficient space utilization.
244  */
245 #define	BUCKET_SIZE(n)						\
246     (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *))
247 
248 #define	BUCKET_MAX	BUCKET_SIZE(256)
249 #define	BUCKET_MIN	2
250 
251 struct uma_bucket_zone bucket_zones[] = {
252 	/* Literal bucket sizes. */
253 	{ NULL, "2 Bucket", 2, 4096 },
254 	{ NULL, "4 Bucket", 4, 3072 },
255 	{ NULL, "8 Bucket", 8, 2048 },
256 	{ NULL, "16 Bucket", 16, 1024 },
257 	/* Rounded down power of 2 sizes for efficiency. */
258 	{ NULL, "32 Bucket", BUCKET_SIZE(32), 512 },
259 	{ NULL, "64 Bucket", BUCKET_SIZE(64), 256 },
260 	{ NULL, "128 Bucket", BUCKET_SIZE(128), 128 },
261 	{ NULL, "256 Bucket", BUCKET_SIZE(256), 64 },
262 	{ NULL, NULL, 0}
263 };
264 
265 /*
266  * Flags and enumerations to be passed to internal functions.
267  */
268 enum zfreeskip {
269 	SKIP_NONE =	0,
270 	SKIP_CNT =	0x00000001,
271 	SKIP_DTOR =	0x00010000,
272 	SKIP_FINI =	0x00020000,
273 };
274 
275 /* Prototypes.. */
276 
277 void	uma_startup1(vm_offset_t);
278 void	uma_startup2(void);
279 
280 static void *noobj_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
281 static void *page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
282 static void *pcpu_page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
283 static void *startup_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
284 static void *contig_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
285 static void page_free(void *, vm_size_t, uint8_t);
286 static void pcpu_page_free(void *, vm_size_t, uint8_t);
287 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int, int, int);
288 static void cache_drain(uma_zone_t);
289 static void bucket_drain(uma_zone_t, uma_bucket_t);
290 static void bucket_cache_reclaim(uma_zone_t zone, bool);
291 static int keg_ctor(void *, int, void *, int);
292 static void keg_dtor(void *, int, void *);
293 static int zone_ctor(void *, int, void *, int);
294 static void zone_dtor(void *, int, void *);
295 static inline void item_dtor(uma_zone_t zone, void *item, int size,
296     void *udata, enum zfreeskip skip);
297 static int zero_init(void *, int, int);
298 static void zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata,
299     int itemdomain, bool ws);
300 static void zone_foreach(void (*zfunc)(uma_zone_t, void *), void *);
301 static void zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *), void *);
302 static void zone_timeout(uma_zone_t zone, void *);
303 static int hash_alloc(struct uma_hash *, u_int);
304 static int hash_expand(struct uma_hash *, struct uma_hash *);
305 static void hash_free(struct uma_hash *hash);
306 static void uma_timeout(void *);
307 static void uma_startup3(void);
308 static void uma_shutdown(void);
309 static void *zone_alloc_item(uma_zone_t, void *, int, int);
310 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip);
311 static int zone_alloc_limit(uma_zone_t zone, int count, int flags);
312 static void zone_free_limit(uma_zone_t zone, int count);
313 static void bucket_enable(void);
314 static void bucket_init(void);
315 static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int);
316 static void bucket_free(uma_zone_t zone, uma_bucket_t, void *);
317 static void bucket_zone_drain(void);
318 static uma_bucket_t zone_alloc_bucket(uma_zone_t, void *, int, int);
319 static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab);
320 static void slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item);
321 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit,
322     uma_fini fini, int align, uint32_t flags);
323 static int zone_import(void *, void **, int, int, int);
324 static void zone_release(void *, void **, int);
325 static bool cache_alloc(uma_zone_t, uma_cache_t, void *, int);
326 static bool cache_free(uma_zone_t, uma_cache_t, void *, void *, int);
327 
328 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS);
329 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS);
330 static int sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS);
331 static int sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS);
332 static int sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS);
333 static int sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS);
334 static int sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS);
335 
336 static uint64_t uma_zone_get_allocs(uma_zone_t zone);
337 
338 static SYSCTL_NODE(_vm, OID_AUTO, debug, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
339     "Memory allocation debugging");
340 
341 #ifdef INVARIANTS
342 static uint64_t uma_keg_get_allocs(uma_keg_t zone);
343 static inline struct noslabbits *slab_dbg_bits(uma_slab_t slab, uma_keg_t keg);
344 
345 static bool uma_dbg_kskip(uma_keg_t keg, void *mem);
346 static bool uma_dbg_zskip(uma_zone_t zone, void *mem);
347 static void uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item);
348 static void uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item);
349 
350 static u_int dbg_divisor = 1;
351 SYSCTL_UINT(_vm_debug, OID_AUTO, divisor,
352     CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &dbg_divisor, 0,
353     "Debug & thrash every this item in memory allocator");
354 
355 static counter_u64_t uma_dbg_cnt = EARLY_COUNTER;
356 static counter_u64_t uma_skip_cnt = EARLY_COUNTER;
357 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, trashed, CTLFLAG_RD,
358     &uma_dbg_cnt, "memory items debugged");
359 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, skipped, CTLFLAG_RD,
360     &uma_skip_cnt, "memory items skipped, not debugged");
361 #endif
362 
363 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
364 
365 SYSCTL_NODE(_vm, OID_AUTO, uma, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
366     "Universal Memory Allocator");
367 
368 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLFLAG_MPSAFE|CTLTYPE_INT,
369     0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones");
370 
371 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLFLAG_MPSAFE|CTLTYPE_STRUCT,
372     0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats");
373 
374 static int zone_warnings = 1;
375 SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RWTUN, &zone_warnings, 0,
376     "Warn when UMA zones becomes full");
377 
378 static int multipage_slabs = 1;
379 TUNABLE_INT("vm.debug.uma_multipage_slabs", &multipage_slabs);
380 SYSCTL_INT(_vm_debug, OID_AUTO, uma_multipage_slabs,
381     CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &multipage_slabs, 0,
382     "UMA may choose larger slab sizes for better efficiency");
383 
384 /*
385  * Select the slab zone for an offpage slab with the given maximum item count.
386  */
387 static inline uma_zone_t
388 slabzone(int ipers)
389 {
390 
391 	return (slabzones[ipers > SLABZONE0_SETSIZE]);
392 }
393 
394 /*
395  * This routine checks to see whether or not it's safe to enable buckets.
396  */
397 static void
398 bucket_enable(void)
399 {
400 
401 	KASSERT(booted >= BOOT_KVA, ("Bucket enable before init"));
402 	bucketdisable = vm_page_count_min();
403 }
404 
405 /*
406  * Initialize bucket_zones, the array of zones of buckets of various sizes.
407  *
408  * For each zone, calculate the memory required for each bucket, consisting
409  * of the header and an array of pointers.
410  */
411 static void
412 bucket_init(void)
413 {
414 	struct uma_bucket_zone *ubz;
415 	int size;
416 
417 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) {
418 		size = roundup(sizeof(struct uma_bucket), sizeof(void *));
419 		size += sizeof(void *) * ubz->ubz_entries;
420 		ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
421 		    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
422 		    UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET |
423 		    UMA_ZONE_FIRSTTOUCH);
424 	}
425 }
426 
427 /*
428  * Given a desired number of entries for a bucket, return the zone from which
429  * to allocate the bucket.
430  */
431 static struct uma_bucket_zone *
432 bucket_zone_lookup(int entries)
433 {
434 	struct uma_bucket_zone *ubz;
435 
436 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
437 		if (ubz->ubz_entries >= entries)
438 			return (ubz);
439 	ubz--;
440 	return (ubz);
441 }
442 
443 static struct uma_bucket_zone *
444 bucket_zone_max(uma_zone_t zone, int nitems)
445 {
446 	struct uma_bucket_zone *ubz;
447 	int bpcpu;
448 
449 	bpcpu = 2;
450 	if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
451 		/* Count the cross-domain bucket. */
452 		bpcpu++;
453 
454 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
455 		if (ubz->ubz_entries * bpcpu * mp_ncpus > nitems)
456 			break;
457 	if (ubz == &bucket_zones[0])
458 		ubz = NULL;
459 	else
460 		ubz--;
461 	return (ubz);
462 }
463 
464 static int
465 bucket_select(int size)
466 {
467 	struct uma_bucket_zone *ubz;
468 
469 	ubz = &bucket_zones[0];
470 	if (size > ubz->ubz_maxsize)
471 		return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1);
472 
473 	for (; ubz->ubz_entries != 0; ubz++)
474 		if (ubz->ubz_maxsize < size)
475 			break;
476 	ubz--;
477 	return (ubz->ubz_entries);
478 }
479 
480 static uma_bucket_t
481 bucket_alloc(uma_zone_t zone, void *udata, int flags)
482 {
483 	struct uma_bucket_zone *ubz;
484 	uma_bucket_t bucket;
485 
486 	/*
487 	 * Don't allocate buckets early in boot.
488 	 */
489 	if (__predict_false(booted < BOOT_KVA))
490 		return (NULL);
491 
492 	/*
493 	 * To limit bucket recursion we store the original zone flags
494 	 * in a cookie passed via zalloc_arg/zfree_arg.  This allows the
495 	 * NOVM flag to persist even through deep recursions.  We also
496 	 * store ZFLAG_BUCKET once we have recursed attempting to allocate
497 	 * a bucket for a bucket zone so we do not allow infinite bucket
498 	 * recursion.  This cookie will even persist to frees of unused
499 	 * buckets via the allocation path or bucket allocations in the
500 	 * free path.
501 	 */
502 	if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
503 		udata = (void *)(uintptr_t)zone->uz_flags;
504 	else {
505 		if ((uintptr_t)udata & UMA_ZFLAG_BUCKET)
506 			return (NULL);
507 		udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET);
508 	}
509 	if (((uintptr_t)udata & UMA_ZONE_VM) != 0)
510 		flags |= M_NOVM;
511 	ubz = bucket_zone_lookup(zone->uz_bucket_size);
512 	if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0)
513 		ubz++;
514 	bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags);
515 	if (bucket) {
516 #ifdef INVARIANTS
517 		bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
518 #endif
519 		bucket->ub_cnt = 0;
520 		bucket->ub_entries = ubz->ubz_entries;
521 		bucket->ub_seq = SMR_SEQ_INVALID;
522 		CTR3(KTR_UMA, "bucket_alloc: zone %s(%p) allocated bucket %p",
523 		    zone->uz_name, zone, bucket);
524 	}
525 
526 	return (bucket);
527 }
528 
529 static void
530 bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata)
531 {
532 	struct uma_bucket_zone *ubz;
533 
534 	if (bucket->ub_cnt != 0)
535 		bucket_drain(zone, bucket);
536 
537 	KASSERT(bucket->ub_cnt == 0,
538 	    ("bucket_free: Freeing a non free bucket."));
539 	KASSERT(bucket->ub_seq == SMR_SEQ_INVALID,
540 	    ("bucket_free: Freeing an SMR bucket."));
541 	if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
542 		udata = (void *)(uintptr_t)zone->uz_flags;
543 	ubz = bucket_zone_lookup(bucket->ub_entries);
544 	uma_zfree_arg(ubz->ubz_zone, bucket, udata);
545 }
546 
547 static void
548 bucket_zone_drain(void)
549 {
550 	struct uma_bucket_zone *ubz;
551 
552 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
553 		uma_zone_reclaim(ubz->ubz_zone, UMA_RECLAIM_DRAIN);
554 }
555 
556 /*
557  * Acquire the domain lock and record contention.
558  */
559 static uma_zone_domain_t
560 zone_domain_lock(uma_zone_t zone, int domain)
561 {
562 	uma_zone_domain_t zdom;
563 	bool lockfail;
564 
565 	zdom = ZDOM_GET(zone, domain);
566 	lockfail = false;
567 	if (ZDOM_OWNED(zdom))
568 		lockfail = true;
569 	ZDOM_LOCK(zdom);
570 	/* This is unsynchronized.  The counter does not need to be precise. */
571 	if (lockfail && zone->uz_bucket_size < zone->uz_bucket_size_max)
572 		zone->uz_bucket_size++;
573 	return (zdom);
574 }
575 
576 /*
577  * Search for the domain with the least cached items and return it if it
578  * is out of balance with the preferred domain.
579  */
580 static __noinline int
581 zone_domain_lowest(uma_zone_t zone, int pref)
582 {
583 	long least, nitems, prefitems;
584 	int domain;
585 	int i;
586 
587 	prefitems = least = LONG_MAX;
588 	domain = 0;
589 	for (i = 0; i < vm_ndomains; i++) {
590 		nitems = ZDOM_GET(zone, i)->uzd_nitems;
591 		if (nitems < least) {
592 			domain = i;
593 			least = nitems;
594 		}
595 		if (domain == pref)
596 			prefitems = nitems;
597 	}
598 	if (prefitems < least * 2)
599 		return (pref);
600 
601 	return (domain);
602 }
603 
604 /*
605  * Search for the domain with the most cached items and return it or the
606  * preferred domain if it has enough to proceed.
607  */
608 static __noinline int
609 zone_domain_highest(uma_zone_t zone, int pref)
610 {
611 	long most, nitems;
612 	int domain;
613 	int i;
614 
615 	if (ZDOM_GET(zone, pref)->uzd_nitems > BUCKET_MAX)
616 		return (pref);
617 
618 	most = 0;
619 	domain = 0;
620 	for (i = 0; i < vm_ndomains; i++) {
621 		nitems = ZDOM_GET(zone, i)->uzd_nitems;
622 		if (nitems > most) {
623 			domain = i;
624 			most = nitems;
625 		}
626 	}
627 
628 	return (domain);
629 }
630 
631 /*
632  * Safely subtract cnt from imax.
633  */
634 static void
635 zone_domain_imax_sub(uma_zone_domain_t zdom, int cnt)
636 {
637 	long new;
638 	long old;
639 
640 	old = zdom->uzd_imax;
641 	do {
642 		if (old <= cnt)
643 			new = 0;
644 		else
645 			new = old - cnt;
646 	} while (atomic_fcmpset_long(&zdom->uzd_imax, &old, new) == 0);
647 }
648 
649 /*
650  * Set the maximum imax value.
651  */
652 static void
653 zone_domain_imax_set(uma_zone_domain_t zdom, int nitems)
654 {
655 	long old;
656 
657 	old = zdom->uzd_imax;
658 	do {
659 		if (old >= nitems)
660 			break;
661 	} while (atomic_fcmpset_long(&zdom->uzd_imax, &old, nitems) == 0);
662 }
663 
664 /*
665  * Attempt to satisfy an allocation by retrieving a full bucket from one of the
666  * zone's caches.  If a bucket is found the zone is not locked on return.
667  */
668 static uma_bucket_t
669 zone_fetch_bucket(uma_zone_t zone, uma_zone_domain_t zdom, bool reclaim)
670 {
671 	uma_bucket_t bucket;
672 	int i;
673 	bool dtor = false;
674 
675 	ZDOM_LOCK_ASSERT(zdom);
676 
677 	if ((bucket = STAILQ_FIRST(&zdom->uzd_buckets)) == NULL)
678 		return (NULL);
679 
680 	/* SMR Buckets can not be re-used until readers expire. */
681 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
682 	    bucket->ub_seq != SMR_SEQ_INVALID) {
683 		if (!smr_poll(zone->uz_smr, bucket->ub_seq, false))
684 			return (NULL);
685 		bucket->ub_seq = SMR_SEQ_INVALID;
686 		dtor = (zone->uz_dtor != NULL) || UMA_ALWAYS_CTORDTOR;
687 		if (STAILQ_NEXT(bucket, ub_link) != NULL)
688 			zdom->uzd_seq = STAILQ_NEXT(bucket, ub_link)->ub_seq;
689 	}
690 	MPASS(zdom->uzd_nitems >= bucket->ub_cnt);
691 	STAILQ_REMOVE_HEAD(&zdom->uzd_buckets, ub_link);
692 	zdom->uzd_nitems -= bucket->ub_cnt;
693 
694 	/*
695 	 * Shift the bounds of the current WSS interval to avoid
696 	 * perturbing the estimate.
697 	 */
698 	if (reclaim) {
699 		zdom->uzd_imin -= lmin(zdom->uzd_imin, bucket->ub_cnt);
700 		zone_domain_imax_sub(zdom, bucket->ub_cnt);
701 	} else if (zdom->uzd_imin > zdom->uzd_nitems)
702 		zdom->uzd_imin = zdom->uzd_nitems;
703 
704 	ZDOM_UNLOCK(zdom);
705 	if (dtor)
706 		for (i = 0; i < bucket->ub_cnt; i++)
707 			item_dtor(zone, bucket->ub_bucket[i], zone->uz_size,
708 			    NULL, SKIP_NONE);
709 
710 	return (bucket);
711 }
712 
713 /*
714  * Insert a full bucket into the specified cache.  The "ws" parameter indicates
715  * whether the bucket's contents should be counted as part of the zone's working
716  * set.  The bucket may be freed if it exceeds the bucket limit.
717  */
718 static void
719 zone_put_bucket(uma_zone_t zone, int domain, uma_bucket_t bucket, void *udata,
720     const bool ws)
721 {
722 	uma_zone_domain_t zdom;
723 
724 	/* We don't cache empty buckets.  This can happen after a reclaim. */
725 	if (bucket->ub_cnt == 0)
726 		goto out;
727 	zdom = zone_domain_lock(zone, domain);
728 
729 	/*
730 	 * Conditionally set the maximum number of items.
731 	 */
732 	zdom->uzd_nitems += bucket->ub_cnt;
733 	if (__predict_true(zdom->uzd_nitems < zone->uz_bucket_max)) {
734 		if (ws)
735 			zone_domain_imax_set(zdom, zdom->uzd_nitems);
736 		if (STAILQ_EMPTY(&zdom->uzd_buckets))
737 			zdom->uzd_seq = bucket->ub_seq;
738 		STAILQ_INSERT_TAIL(&zdom->uzd_buckets, bucket, ub_link);
739 		ZDOM_UNLOCK(zdom);
740 		return;
741 	}
742 	zdom->uzd_nitems -= bucket->ub_cnt;
743 	ZDOM_UNLOCK(zdom);
744 out:
745 	bucket_free(zone, bucket, udata);
746 }
747 
748 /* Pops an item out of a per-cpu cache bucket. */
749 static inline void *
750 cache_bucket_pop(uma_cache_t cache, uma_cache_bucket_t bucket)
751 {
752 	void *item;
753 
754 	CRITICAL_ASSERT(curthread);
755 
756 	bucket->ucb_cnt--;
757 	item = bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt];
758 #ifdef INVARIANTS
759 	bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = NULL;
760 	KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled."));
761 #endif
762 	cache->uc_allocs++;
763 
764 	return (item);
765 }
766 
767 /* Pushes an item into a per-cpu cache bucket. */
768 static inline void
769 cache_bucket_push(uma_cache_t cache, uma_cache_bucket_t bucket, void *item)
770 {
771 
772 	CRITICAL_ASSERT(curthread);
773 	KASSERT(bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] == NULL,
774 	    ("uma_zfree: Freeing to non free bucket index."));
775 
776 	bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = item;
777 	bucket->ucb_cnt++;
778 	cache->uc_frees++;
779 }
780 
781 /*
782  * Unload a UMA bucket from a per-cpu cache.
783  */
784 static inline uma_bucket_t
785 cache_bucket_unload(uma_cache_bucket_t bucket)
786 {
787 	uma_bucket_t b;
788 
789 	b = bucket->ucb_bucket;
790 	if (b != NULL) {
791 		MPASS(b->ub_entries == bucket->ucb_entries);
792 		b->ub_cnt = bucket->ucb_cnt;
793 		bucket->ucb_bucket = NULL;
794 		bucket->ucb_entries = bucket->ucb_cnt = 0;
795 	}
796 
797 	return (b);
798 }
799 
800 static inline uma_bucket_t
801 cache_bucket_unload_alloc(uma_cache_t cache)
802 {
803 
804 	return (cache_bucket_unload(&cache->uc_allocbucket));
805 }
806 
807 static inline uma_bucket_t
808 cache_bucket_unload_free(uma_cache_t cache)
809 {
810 
811 	return (cache_bucket_unload(&cache->uc_freebucket));
812 }
813 
814 static inline uma_bucket_t
815 cache_bucket_unload_cross(uma_cache_t cache)
816 {
817 
818 	return (cache_bucket_unload(&cache->uc_crossbucket));
819 }
820 
821 /*
822  * Load a bucket into a per-cpu cache bucket.
823  */
824 static inline void
825 cache_bucket_load(uma_cache_bucket_t bucket, uma_bucket_t b)
826 {
827 
828 	CRITICAL_ASSERT(curthread);
829 	MPASS(bucket->ucb_bucket == NULL);
830 	MPASS(b->ub_seq == SMR_SEQ_INVALID);
831 
832 	bucket->ucb_bucket = b;
833 	bucket->ucb_cnt = b->ub_cnt;
834 	bucket->ucb_entries = b->ub_entries;
835 }
836 
837 static inline void
838 cache_bucket_load_alloc(uma_cache_t cache, uma_bucket_t b)
839 {
840 
841 	cache_bucket_load(&cache->uc_allocbucket, b);
842 }
843 
844 static inline void
845 cache_bucket_load_free(uma_cache_t cache, uma_bucket_t b)
846 {
847 
848 	cache_bucket_load(&cache->uc_freebucket, b);
849 }
850 
851 #ifdef NUMA
852 static inline void
853 cache_bucket_load_cross(uma_cache_t cache, uma_bucket_t b)
854 {
855 
856 	cache_bucket_load(&cache->uc_crossbucket, b);
857 }
858 #endif
859 
860 /*
861  * Copy and preserve ucb_spare.
862  */
863 static inline void
864 cache_bucket_copy(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
865 {
866 
867 	b1->ucb_bucket = b2->ucb_bucket;
868 	b1->ucb_entries = b2->ucb_entries;
869 	b1->ucb_cnt = b2->ucb_cnt;
870 }
871 
872 /*
873  * Swap two cache buckets.
874  */
875 static inline void
876 cache_bucket_swap(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
877 {
878 	struct uma_cache_bucket b3;
879 
880 	CRITICAL_ASSERT(curthread);
881 
882 	cache_bucket_copy(&b3, b1);
883 	cache_bucket_copy(b1, b2);
884 	cache_bucket_copy(b2, &b3);
885 }
886 
887 /*
888  * Attempt to fetch a bucket from a zone on behalf of the current cpu cache.
889  */
890 static uma_bucket_t
891 cache_fetch_bucket(uma_zone_t zone, uma_cache_t cache, int domain)
892 {
893 	uma_zone_domain_t zdom;
894 	uma_bucket_t bucket;
895 
896 	/*
897 	 * Avoid the lock if possible.
898 	 */
899 	zdom = ZDOM_GET(zone, domain);
900 	if (zdom->uzd_nitems == 0)
901 		return (NULL);
902 
903 	if ((cache_uz_flags(cache) & UMA_ZONE_SMR) != 0 &&
904 	    !smr_poll(zone->uz_smr, zdom->uzd_seq, false))
905 		return (NULL);
906 
907 	/*
908 	 * Check the zone's cache of buckets.
909 	 */
910 	zdom = zone_domain_lock(zone, domain);
911 	if ((bucket = zone_fetch_bucket(zone, zdom, false)) != NULL) {
912 		KASSERT(bucket->ub_cnt != 0,
913 		    ("cache_fetch_bucket: Returning an empty bucket."));
914 		return (bucket);
915 	}
916 	ZDOM_UNLOCK(zdom);
917 
918 	return (NULL);
919 }
920 
921 static void
922 zone_log_warning(uma_zone_t zone)
923 {
924 	static const struct timeval warninterval = { 300, 0 };
925 
926 	if (!zone_warnings || zone->uz_warning == NULL)
927 		return;
928 
929 	if (ratecheck(&zone->uz_ratecheck, &warninterval))
930 		printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning);
931 }
932 
933 static inline void
934 zone_maxaction(uma_zone_t zone)
935 {
936 
937 	if (zone->uz_maxaction.ta_func != NULL)
938 		taskqueue_enqueue(taskqueue_thread, &zone->uz_maxaction);
939 }
940 
941 /*
942  * Routine called by timeout which is used to fire off some time interval
943  * based calculations.  (stats, hash size, etc.)
944  *
945  * Arguments:
946  *	arg   Unused
947  *
948  * Returns:
949  *	Nothing
950  */
951 static void
952 uma_timeout(void *unused)
953 {
954 	bucket_enable();
955 	zone_foreach(zone_timeout, NULL);
956 
957 	/* Reschedule this event */
958 	callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
959 }
960 
961 /*
962  * Update the working set size estimate for the zone's bucket cache.
963  * The constants chosen here are somewhat arbitrary.  With an update period of
964  * 20s (UMA_TIMEOUT), this estimate is dominated by zone activity over the
965  * last 100s.
966  */
967 static void
968 zone_domain_update_wss(uma_zone_domain_t zdom)
969 {
970 	long wss;
971 
972 	ZDOM_LOCK(zdom);
973 	MPASS(zdom->uzd_imax >= zdom->uzd_imin);
974 	wss = zdom->uzd_imax - zdom->uzd_imin;
975 	zdom->uzd_imax = zdom->uzd_imin = zdom->uzd_nitems;
976 	zdom->uzd_wss = (4 * wss + zdom->uzd_wss) / 5;
977 	ZDOM_UNLOCK(zdom);
978 }
979 
980 /*
981  * Routine to perform timeout driven calculations.  This expands the
982  * hashes and does per cpu statistics aggregation.
983  *
984  *  Returns nothing.
985  */
986 static void
987 zone_timeout(uma_zone_t zone, void *unused)
988 {
989 	uma_keg_t keg;
990 	u_int slabs, pages;
991 
992 	if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
993 		goto update_wss;
994 
995 	keg = zone->uz_keg;
996 
997 	/*
998 	 * Hash zones are non-numa by definition so the first domain
999 	 * is the only one present.
1000 	 */
1001 	KEG_LOCK(keg, 0);
1002 	pages = keg->uk_domain[0].ud_pages;
1003 
1004 	/*
1005 	 * Expand the keg hash table.
1006 	 *
1007 	 * This is done if the number of slabs is larger than the hash size.
1008 	 * What I'm trying to do here is completely reduce collisions.  This
1009 	 * may be a little aggressive.  Should I allow for two collisions max?
1010 	 */
1011 	if ((slabs = pages / keg->uk_ppera) > keg->uk_hash.uh_hashsize) {
1012 		struct uma_hash newhash;
1013 		struct uma_hash oldhash;
1014 		int ret;
1015 
1016 		/*
1017 		 * This is so involved because allocating and freeing
1018 		 * while the keg lock is held will lead to deadlock.
1019 		 * I have to do everything in stages and check for
1020 		 * races.
1021 		 */
1022 		KEG_UNLOCK(keg, 0);
1023 		ret = hash_alloc(&newhash, 1 << fls(slabs));
1024 		KEG_LOCK(keg, 0);
1025 		if (ret) {
1026 			if (hash_expand(&keg->uk_hash, &newhash)) {
1027 				oldhash = keg->uk_hash;
1028 				keg->uk_hash = newhash;
1029 			} else
1030 				oldhash = newhash;
1031 
1032 			KEG_UNLOCK(keg, 0);
1033 			hash_free(&oldhash);
1034 			goto update_wss;
1035 		}
1036 	}
1037 	KEG_UNLOCK(keg, 0);
1038 
1039 update_wss:
1040 	for (int i = 0; i < vm_ndomains; i++)
1041 		zone_domain_update_wss(ZDOM_GET(zone, i));
1042 }
1043 
1044 /*
1045  * Allocate and zero fill the next sized hash table from the appropriate
1046  * backing store.
1047  *
1048  * Arguments:
1049  *	hash  A new hash structure with the old hash size in uh_hashsize
1050  *
1051  * Returns:
1052  *	1 on success and 0 on failure.
1053  */
1054 static int
1055 hash_alloc(struct uma_hash *hash, u_int size)
1056 {
1057 	size_t alloc;
1058 
1059 	KASSERT(powerof2(size), ("hash size must be power of 2"));
1060 	if (size > UMA_HASH_SIZE_INIT)  {
1061 		hash->uh_hashsize = size;
1062 		alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
1063 		hash->uh_slab_hash = malloc(alloc, M_UMAHASH, M_NOWAIT);
1064 	} else {
1065 		alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
1066 		hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
1067 		    UMA_ANYDOMAIN, M_WAITOK);
1068 		hash->uh_hashsize = UMA_HASH_SIZE_INIT;
1069 	}
1070 	if (hash->uh_slab_hash) {
1071 		bzero(hash->uh_slab_hash, alloc);
1072 		hash->uh_hashmask = hash->uh_hashsize - 1;
1073 		return (1);
1074 	}
1075 
1076 	return (0);
1077 }
1078 
1079 /*
1080  * Expands the hash table for HASH zones.  This is done from zone_timeout
1081  * to reduce collisions.  This must not be done in the regular allocation
1082  * path, otherwise, we can recurse on the vm while allocating pages.
1083  *
1084  * Arguments:
1085  *	oldhash  The hash you want to expand
1086  *	newhash  The hash structure for the new table
1087  *
1088  * Returns:
1089  *	Nothing
1090  *
1091  * Discussion:
1092  */
1093 static int
1094 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
1095 {
1096 	uma_hash_slab_t slab;
1097 	u_int hval;
1098 	u_int idx;
1099 
1100 	if (!newhash->uh_slab_hash)
1101 		return (0);
1102 
1103 	if (oldhash->uh_hashsize >= newhash->uh_hashsize)
1104 		return (0);
1105 
1106 	/*
1107 	 * I need to investigate hash algorithms for resizing without a
1108 	 * full rehash.
1109 	 */
1110 
1111 	for (idx = 0; idx < oldhash->uh_hashsize; idx++)
1112 		while (!LIST_EMPTY(&oldhash->uh_slab_hash[idx])) {
1113 			slab = LIST_FIRST(&oldhash->uh_slab_hash[idx]);
1114 			LIST_REMOVE(slab, uhs_hlink);
1115 			hval = UMA_HASH(newhash, slab->uhs_data);
1116 			LIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
1117 			    slab, uhs_hlink);
1118 		}
1119 
1120 	return (1);
1121 }
1122 
1123 /*
1124  * Free the hash bucket to the appropriate backing store.
1125  *
1126  * Arguments:
1127  *	slab_hash  The hash bucket we're freeing
1128  *	hashsize   The number of entries in that hash bucket
1129  *
1130  * Returns:
1131  *	Nothing
1132  */
1133 static void
1134 hash_free(struct uma_hash *hash)
1135 {
1136 	if (hash->uh_slab_hash == NULL)
1137 		return;
1138 	if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
1139 		zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE);
1140 	else
1141 		free(hash->uh_slab_hash, M_UMAHASH);
1142 }
1143 
1144 /*
1145  * Frees all outstanding items in a bucket
1146  *
1147  * Arguments:
1148  *	zone   The zone to free to, must be unlocked.
1149  *	bucket The free/alloc bucket with items.
1150  *
1151  * Returns:
1152  *	Nothing
1153  */
1154 static void
1155 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
1156 {
1157 	int i;
1158 
1159 	if (bucket->ub_cnt == 0)
1160 		return;
1161 
1162 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
1163 	    bucket->ub_seq != SMR_SEQ_INVALID) {
1164 		smr_wait(zone->uz_smr, bucket->ub_seq);
1165 		bucket->ub_seq = SMR_SEQ_INVALID;
1166 		for (i = 0; i < bucket->ub_cnt; i++)
1167 			item_dtor(zone, bucket->ub_bucket[i],
1168 			    zone->uz_size, NULL, SKIP_NONE);
1169 	}
1170 	if (zone->uz_fini)
1171 		for (i = 0; i < bucket->ub_cnt; i++)
1172 			zone->uz_fini(bucket->ub_bucket[i], zone->uz_size);
1173 	zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt);
1174 	if (zone->uz_max_items > 0)
1175 		zone_free_limit(zone, bucket->ub_cnt);
1176 #ifdef INVARIANTS
1177 	bzero(bucket->ub_bucket, sizeof(void *) * bucket->ub_cnt);
1178 #endif
1179 	bucket->ub_cnt = 0;
1180 }
1181 
1182 /*
1183  * Drains the per cpu caches for a zone.
1184  *
1185  * NOTE: This may only be called while the zone is being torn down, and not
1186  * during normal operation.  This is necessary in order that we do not have
1187  * to migrate CPUs to drain the per-CPU caches.
1188  *
1189  * Arguments:
1190  *	zone     The zone to drain, must be unlocked.
1191  *
1192  * Returns:
1193  *	Nothing
1194  */
1195 static void
1196 cache_drain(uma_zone_t zone)
1197 {
1198 	uma_cache_t cache;
1199 	uma_bucket_t bucket;
1200 	smr_seq_t seq;
1201 	int cpu;
1202 
1203 	/*
1204 	 * XXX: It is safe to not lock the per-CPU caches, because we're
1205 	 * tearing down the zone anyway.  I.e., there will be no further use
1206 	 * of the caches at this point.
1207 	 *
1208 	 * XXX: It would good to be able to assert that the zone is being
1209 	 * torn down to prevent improper use of cache_drain().
1210 	 */
1211 	seq = SMR_SEQ_INVALID;
1212 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
1213 		seq = smr_advance(zone->uz_smr);
1214 	CPU_FOREACH(cpu) {
1215 		cache = &zone->uz_cpu[cpu];
1216 		bucket = cache_bucket_unload_alloc(cache);
1217 		if (bucket != NULL)
1218 			bucket_free(zone, bucket, NULL);
1219 		bucket = cache_bucket_unload_free(cache);
1220 		if (bucket != NULL) {
1221 			bucket->ub_seq = seq;
1222 			bucket_free(zone, bucket, NULL);
1223 		}
1224 		bucket = cache_bucket_unload_cross(cache);
1225 		if (bucket != NULL) {
1226 			bucket->ub_seq = seq;
1227 			bucket_free(zone, bucket, NULL);
1228 		}
1229 	}
1230 	bucket_cache_reclaim(zone, true);
1231 }
1232 
1233 static void
1234 cache_shrink(uma_zone_t zone, void *unused)
1235 {
1236 
1237 	if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1238 		return;
1239 
1240 	zone->uz_bucket_size =
1241 	    (zone->uz_bucket_size_min + zone->uz_bucket_size) / 2;
1242 }
1243 
1244 static void
1245 cache_drain_safe_cpu(uma_zone_t zone, void *unused)
1246 {
1247 	uma_cache_t cache;
1248 	uma_bucket_t b1, b2, b3;
1249 	int domain;
1250 
1251 	if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1252 		return;
1253 
1254 	b1 = b2 = b3 = NULL;
1255 	critical_enter();
1256 	cache = &zone->uz_cpu[curcpu];
1257 	domain = PCPU_GET(domain);
1258 	b1 = cache_bucket_unload_alloc(cache);
1259 
1260 	/*
1261 	 * Don't flush SMR zone buckets.  This leaves the zone without a
1262 	 * bucket and forces every free to synchronize().
1263 	 */
1264 	if ((zone->uz_flags & UMA_ZONE_SMR) == 0) {
1265 		b2 = cache_bucket_unload_free(cache);
1266 		b3 = cache_bucket_unload_cross(cache);
1267 	}
1268 	critical_exit();
1269 
1270 	if (b1 != NULL)
1271 		zone_free_bucket(zone, b1, NULL, domain, false);
1272 	if (b2 != NULL)
1273 		zone_free_bucket(zone, b2, NULL, domain, false);
1274 	if (b3 != NULL) {
1275 		/* Adjust the domain so it goes to zone_free_cross. */
1276 		domain = (domain + 1) % vm_ndomains;
1277 		zone_free_bucket(zone, b3, NULL, domain, false);
1278 	}
1279 }
1280 
1281 /*
1282  * Safely drain per-CPU caches of a zone(s) to alloc bucket.
1283  * This is an expensive call because it needs to bind to all CPUs
1284  * one by one and enter a critical section on each of them in order
1285  * to safely access their cache buckets.
1286  * Zone lock must not be held on call this function.
1287  */
1288 static void
1289 pcpu_cache_drain_safe(uma_zone_t zone)
1290 {
1291 	int cpu;
1292 
1293 	/*
1294 	 * Polite bucket sizes shrinking was not enough, shrink aggressively.
1295 	 */
1296 	if (zone)
1297 		cache_shrink(zone, NULL);
1298 	else
1299 		zone_foreach(cache_shrink, NULL);
1300 
1301 	CPU_FOREACH(cpu) {
1302 		thread_lock(curthread);
1303 		sched_bind(curthread, cpu);
1304 		thread_unlock(curthread);
1305 
1306 		if (zone)
1307 			cache_drain_safe_cpu(zone, NULL);
1308 		else
1309 			zone_foreach(cache_drain_safe_cpu, NULL);
1310 	}
1311 	thread_lock(curthread);
1312 	sched_unbind(curthread);
1313 	thread_unlock(curthread);
1314 }
1315 
1316 /*
1317  * Reclaim cached buckets from a zone.  All buckets are reclaimed if the caller
1318  * requested a drain, otherwise the per-domain caches are trimmed to either
1319  * estimated working set size.
1320  */
1321 static void
1322 bucket_cache_reclaim(uma_zone_t zone, bool drain)
1323 {
1324 	uma_zone_domain_t zdom;
1325 	uma_bucket_t bucket;
1326 	long target;
1327 	int i;
1328 
1329 	/*
1330 	 * Shrink the zone bucket size to ensure that the per-CPU caches
1331 	 * don't grow too large.
1332 	 */
1333 	if (zone->uz_bucket_size > zone->uz_bucket_size_min)
1334 		zone->uz_bucket_size--;
1335 
1336 	for (i = 0; i < vm_ndomains; i++) {
1337 		/*
1338 		 * The cross bucket is partially filled and not part of
1339 		 * the item count.  Reclaim it individually here.
1340 		 */
1341 		zdom = ZDOM_GET(zone, i);
1342 		if ((zone->uz_flags & UMA_ZONE_SMR) == 0 || drain) {
1343 			ZONE_CROSS_LOCK(zone);
1344 			bucket = zdom->uzd_cross;
1345 			zdom->uzd_cross = NULL;
1346 			ZONE_CROSS_UNLOCK(zone);
1347 			if (bucket != NULL)
1348 				bucket_free(zone, bucket, NULL);
1349 		}
1350 
1351 		/*
1352 		 * If we were asked to drain the zone, we are done only once
1353 		 * this bucket cache is empty.  Otherwise, we reclaim items in
1354 		 * excess of the zone's estimated working set size.  If the
1355 		 * difference nitems - imin is larger than the WSS estimate,
1356 		 * then the estimate will grow at the end of this interval and
1357 		 * we ignore the historical average.
1358 		 */
1359 		ZDOM_LOCK(zdom);
1360 		target = drain ? 0 : lmax(zdom->uzd_wss, zdom->uzd_nitems -
1361 		    zdom->uzd_imin);
1362 		while (zdom->uzd_nitems > target) {
1363 			bucket = zone_fetch_bucket(zone, zdom, true);
1364 			if (bucket == NULL)
1365 				break;
1366 			bucket_free(zone, bucket, NULL);
1367 			ZDOM_LOCK(zdom);
1368 		}
1369 		ZDOM_UNLOCK(zdom);
1370 	}
1371 }
1372 
1373 static void
1374 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start)
1375 {
1376 	uint8_t *mem;
1377 	int i;
1378 	uint8_t flags;
1379 
1380 	CTR4(KTR_UMA, "keg_free_slab keg %s(%p) slab %p, returning %d bytes",
1381 	    keg->uk_name, keg, slab, PAGE_SIZE * keg->uk_ppera);
1382 
1383 	mem = slab_data(slab, keg);
1384 	flags = slab->us_flags;
1385 	i = start;
1386 	if (keg->uk_fini != NULL) {
1387 		for (i--; i > -1; i--)
1388 #ifdef INVARIANTS
1389 		/*
1390 		 * trash_fini implies that dtor was trash_dtor. trash_fini
1391 		 * would check that memory hasn't been modified since free,
1392 		 * which executed trash_dtor.
1393 		 * That's why we need to run uma_dbg_kskip() check here,
1394 		 * albeit we don't make skip check for other init/fini
1395 		 * invocations.
1396 		 */
1397 		if (!uma_dbg_kskip(keg, slab_item(slab, keg, i)) ||
1398 		    keg->uk_fini != trash_fini)
1399 #endif
1400 			keg->uk_fini(slab_item(slab, keg, i), keg->uk_size);
1401 	}
1402 	if (keg->uk_flags & UMA_ZFLAG_OFFPAGE)
1403 		zone_free_item(slabzone(keg->uk_ipers), slab_tohashslab(slab),
1404 		    NULL, SKIP_NONE);
1405 	keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags);
1406 	uma_total_dec(PAGE_SIZE * keg->uk_ppera);
1407 }
1408 
1409 /*
1410  * Frees pages from a keg back to the system.  This is done on demand from
1411  * the pageout daemon.
1412  *
1413  * Returns nothing.
1414  */
1415 static void
1416 keg_drain(uma_keg_t keg)
1417 {
1418 	struct slabhead freeslabs;
1419 	uma_domain_t dom;
1420 	uma_slab_t slab, tmp;
1421 	int i, n;
1422 
1423 	if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL)
1424 		return;
1425 
1426 	for (i = 0; i < vm_ndomains; i++) {
1427 		CTR4(KTR_UMA, "keg_drain %s(%p) domain %d free items: %u",
1428 		    keg->uk_name, keg, i, dom->ud_free_items);
1429 		dom = &keg->uk_domain[i];
1430 		LIST_INIT(&freeslabs);
1431 
1432 		KEG_LOCK(keg, i);
1433 		if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0) {
1434 			LIST_FOREACH(slab, &dom->ud_free_slab, us_link)
1435 				UMA_HASH_REMOVE(&keg->uk_hash, slab);
1436 		}
1437 		n = dom->ud_free_slabs;
1438 		LIST_SWAP(&freeslabs, &dom->ud_free_slab, uma_slab, us_link);
1439 		dom->ud_free_slabs = 0;
1440 		dom->ud_free_items -= n * keg->uk_ipers;
1441 		dom->ud_pages -= n * keg->uk_ppera;
1442 		KEG_UNLOCK(keg, i);
1443 
1444 		LIST_FOREACH_SAFE(slab, &freeslabs, us_link, tmp)
1445 			keg_free_slab(keg, slab, keg->uk_ipers);
1446 	}
1447 }
1448 
1449 static void
1450 zone_reclaim(uma_zone_t zone, int waitok, bool drain)
1451 {
1452 
1453 	/*
1454 	 * Set draining to interlock with zone_dtor() so we can release our
1455 	 * locks as we go.  Only dtor() should do a WAITOK call since it
1456 	 * is the only call that knows the structure will still be available
1457 	 * when it wakes up.
1458 	 */
1459 	ZONE_LOCK(zone);
1460 	while (zone->uz_flags & UMA_ZFLAG_RECLAIMING) {
1461 		if (waitok == M_NOWAIT)
1462 			goto out;
1463 		msleep(zone, &ZDOM_GET(zone, 0)->uzd_lock, PVM, "zonedrain",
1464 		    1);
1465 	}
1466 	zone->uz_flags |= UMA_ZFLAG_RECLAIMING;
1467 	ZONE_UNLOCK(zone);
1468 	bucket_cache_reclaim(zone, drain);
1469 
1470 	/*
1471 	 * The DRAINING flag protects us from being freed while
1472 	 * we're running.  Normally the uma_rwlock would protect us but we
1473 	 * must be able to release and acquire the right lock for each keg.
1474 	 */
1475 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0)
1476 		keg_drain(zone->uz_keg);
1477 	ZONE_LOCK(zone);
1478 	zone->uz_flags &= ~UMA_ZFLAG_RECLAIMING;
1479 	wakeup(zone);
1480 out:
1481 	ZONE_UNLOCK(zone);
1482 }
1483 
1484 static void
1485 zone_drain(uma_zone_t zone, void *unused)
1486 {
1487 
1488 	zone_reclaim(zone, M_NOWAIT, true);
1489 }
1490 
1491 static void
1492 zone_trim(uma_zone_t zone, void *unused)
1493 {
1494 
1495 	zone_reclaim(zone, M_NOWAIT, false);
1496 }
1497 
1498 /*
1499  * Allocate a new slab for a keg and inserts it into the partial slab list.
1500  * The keg should be unlocked on entry.  If the allocation succeeds it will
1501  * be locked on return.
1502  *
1503  * Arguments:
1504  *	flags   Wait flags for the item initialization routine
1505  *	aflags  Wait flags for the slab allocation
1506  *
1507  * Returns:
1508  *	The slab that was allocated or NULL if there is no memory and the
1509  *	caller specified M_NOWAIT.
1510  */
1511 static uma_slab_t
1512 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int domain, int flags,
1513     int aflags)
1514 {
1515 	uma_domain_t dom;
1516 	uma_alloc allocf;
1517 	uma_slab_t slab;
1518 	unsigned long size;
1519 	uint8_t *mem;
1520 	uint8_t sflags;
1521 	int i;
1522 
1523 	KASSERT(domain >= 0 && domain < vm_ndomains,
1524 	    ("keg_alloc_slab: domain %d out of range", domain));
1525 
1526 	allocf = keg->uk_allocf;
1527 	slab = NULL;
1528 	mem = NULL;
1529 	if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) {
1530 		uma_hash_slab_t hslab;
1531 		hslab = zone_alloc_item(slabzone(keg->uk_ipers), NULL,
1532 		    domain, aflags);
1533 		if (hslab == NULL)
1534 			goto fail;
1535 		slab = &hslab->uhs_slab;
1536 	}
1537 
1538 	/*
1539 	 * This reproduces the old vm_zone behavior of zero filling pages the
1540 	 * first time they are added to a zone.
1541 	 *
1542 	 * Malloced items are zeroed in uma_zalloc.
1543 	 */
1544 
1545 	if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
1546 		aflags |= M_ZERO;
1547 	else
1548 		aflags &= ~M_ZERO;
1549 
1550 	if (keg->uk_flags & UMA_ZONE_NODUMP)
1551 		aflags |= M_NODUMP;
1552 
1553 	/* zone is passed for legacy reasons. */
1554 	size = keg->uk_ppera * PAGE_SIZE;
1555 	mem = allocf(zone, size, domain, &sflags, aflags);
1556 	if (mem == NULL) {
1557 		if (keg->uk_flags & UMA_ZFLAG_OFFPAGE)
1558 			zone_free_item(slabzone(keg->uk_ipers),
1559 			    slab_tohashslab(slab), NULL, SKIP_NONE);
1560 		goto fail;
1561 	}
1562 	uma_total_inc(size);
1563 
1564 	/* For HASH zones all pages go to the same uma_domain. */
1565 	if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
1566 		domain = 0;
1567 
1568 	/* Point the slab into the allocated memory */
1569 	if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE))
1570 		slab = (uma_slab_t )(mem + keg->uk_pgoff);
1571 	else
1572 		slab_tohashslab(slab)->uhs_data = mem;
1573 
1574 	if (keg->uk_flags & UMA_ZFLAG_VTOSLAB)
1575 		for (i = 0; i < keg->uk_ppera; i++)
1576 			vsetzoneslab((vm_offset_t)mem + (i * PAGE_SIZE),
1577 			    zone, slab);
1578 
1579 	slab->us_freecount = keg->uk_ipers;
1580 	slab->us_flags = sflags;
1581 	slab->us_domain = domain;
1582 
1583 	BIT_FILL(keg->uk_ipers, &slab->us_free);
1584 #ifdef INVARIANTS
1585 	BIT_ZERO(keg->uk_ipers, slab_dbg_bits(slab, keg));
1586 #endif
1587 
1588 	if (keg->uk_init != NULL) {
1589 		for (i = 0; i < keg->uk_ipers; i++)
1590 			if (keg->uk_init(slab_item(slab, keg, i),
1591 			    keg->uk_size, flags) != 0)
1592 				break;
1593 		if (i != keg->uk_ipers) {
1594 			keg_free_slab(keg, slab, i);
1595 			goto fail;
1596 		}
1597 	}
1598 	KEG_LOCK(keg, domain);
1599 
1600 	CTR3(KTR_UMA, "keg_alloc_slab: allocated slab %p for %s(%p)",
1601 	    slab, keg->uk_name, keg);
1602 
1603 	if (keg->uk_flags & UMA_ZFLAG_HASH)
1604 		UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
1605 
1606 	/*
1607 	 * If we got a slab here it's safe to mark it partially used
1608 	 * and return.  We assume that the caller is going to remove
1609 	 * at least one item.
1610 	 */
1611 	dom = &keg->uk_domain[domain];
1612 	LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
1613 	dom->ud_pages += keg->uk_ppera;
1614 	dom->ud_free_items += keg->uk_ipers;
1615 
1616 	return (slab);
1617 
1618 fail:
1619 	return (NULL);
1620 }
1621 
1622 /*
1623  * This function is intended to be used early on in place of page_alloc() so
1624  * that we may use the boot time page cache to satisfy allocations before
1625  * the VM is ready.
1626  */
1627 static void *
1628 startup_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1629     int wait)
1630 {
1631 	vm_paddr_t pa;
1632 	vm_page_t m;
1633 	void *mem;
1634 	int pages;
1635 	int i;
1636 
1637 	pages = howmany(bytes, PAGE_SIZE);
1638 	KASSERT(pages > 0, ("%s can't reserve 0 pages", __func__));
1639 
1640 	*pflag = UMA_SLAB_BOOT;
1641 	m = vm_page_alloc_contig_domain(NULL, 0, domain,
1642 	    malloc2vm_flags(wait) | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED, pages,
1643 	    (vm_paddr_t)0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT);
1644 	if (m == NULL)
1645 		return (NULL);
1646 
1647 	pa = VM_PAGE_TO_PHYS(m);
1648 	for (i = 0; i < pages; i++, pa += PAGE_SIZE) {
1649 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
1650     defined(__riscv) || defined(__powerpc64__)
1651 		if ((wait & M_NODUMP) == 0)
1652 			dump_add_page(pa);
1653 #endif
1654 	}
1655 	/* Allocate KVA and indirectly advance bootmem. */
1656 	mem = (void *)pmap_map(&bootmem, m->phys_addr,
1657 	    m->phys_addr + (pages * PAGE_SIZE), VM_PROT_READ | VM_PROT_WRITE);
1658         if ((wait & M_ZERO) != 0)
1659                 bzero(mem, pages * PAGE_SIZE);
1660 
1661         return (mem);
1662 }
1663 
1664 static void
1665 startup_free(void *mem, vm_size_t bytes)
1666 {
1667 	vm_offset_t va;
1668 	vm_page_t m;
1669 
1670 	va = (vm_offset_t)mem;
1671 	m = PHYS_TO_VM_PAGE(pmap_kextract(va));
1672 	pmap_remove(kernel_pmap, va, va + bytes);
1673 	for (; bytes != 0; bytes -= PAGE_SIZE, m++) {
1674 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
1675     defined(__riscv) || defined(__powerpc64__)
1676 		dump_drop_page(VM_PAGE_TO_PHYS(m));
1677 #endif
1678 		vm_page_unwire_noq(m);
1679 		vm_page_free(m);
1680 	}
1681 }
1682 
1683 /*
1684  * Allocates a number of pages from the system
1685  *
1686  * Arguments:
1687  *	bytes  The number of bytes requested
1688  *	wait  Shall we wait?
1689  *
1690  * Returns:
1691  *	A pointer to the alloced memory or possibly
1692  *	NULL if M_NOWAIT is set.
1693  */
1694 static void *
1695 page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1696     int wait)
1697 {
1698 	void *p;	/* Returned page */
1699 
1700 	*pflag = UMA_SLAB_KERNEL;
1701 	p = (void *)kmem_malloc_domainset(DOMAINSET_FIXED(domain), bytes, wait);
1702 
1703 	return (p);
1704 }
1705 
1706 static void *
1707 pcpu_page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1708     int wait)
1709 {
1710 	struct pglist alloctail;
1711 	vm_offset_t addr, zkva;
1712 	int cpu, flags;
1713 	vm_page_t p, p_next;
1714 #ifdef NUMA
1715 	struct pcpu *pc;
1716 #endif
1717 
1718 	MPASS(bytes == (mp_maxid + 1) * PAGE_SIZE);
1719 
1720 	TAILQ_INIT(&alloctail);
1721 	flags = VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | VM_ALLOC_NOOBJ |
1722 	    malloc2vm_flags(wait);
1723 	*pflag = UMA_SLAB_KERNEL;
1724 	for (cpu = 0; cpu <= mp_maxid; cpu++) {
1725 		if (CPU_ABSENT(cpu)) {
1726 			p = vm_page_alloc(NULL, 0, flags);
1727 		} else {
1728 #ifndef NUMA
1729 			p = vm_page_alloc(NULL, 0, flags);
1730 #else
1731 			pc = pcpu_find(cpu);
1732 			if (__predict_false(VM_DOMAIN_EMPTY(pc->pc_domain)))
1733 				p = NULL;
1734 			else
1735 				p = vm_page_alloc_domain(NULL, 0,
1736 				    pc->pc_domain, flags);
1737 			if (__predict_false(p == NULL))
1738 				p = vm_page_alloc(NULL, 0, flags);
1739 #endif
1740 		}
1741 		if (__predict_false(p == NULL))
1742 			goto fail;
1743 		TAILQ_INSERT_TAIL(&alloctail, p, listq);
1744 	}
1745 	if ((addr = kva_alloc(bytes)) == 0)
1746 		goto fail;
1747 	zkva = addr;
1748 	TAILQ_FOREACH(p, &alloctail, listq) {
1749 		pmap_qenter(zkva, &p, 1);
1750 		zkva += PAGE_SIZE;
1751 	}
1752 	return ((void*)addr);
1753 fail:
1754 	TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1755 		vm_page_unwire_noq(p);
1756 		vm_page_free(p);
1757 	}
1758 	return (NULL);
1759 }
1760 
1761 /*
1762  * Allocates a number of pages from within an object
1763  *
1764  * Arguments:
1765  *	bytes  The number of bytes requested
1766  *	wait   Shall we wait?
1767  *
1768  * Returns:
1769  *	A pointer to the alloced memory or possibly
1770  *	NULL if M_NOWAIT is set.
1771  */
1772 static void *
1773 noobj_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags,
1774     int wait)
1775 {
1776 	TAILQ_HEAD(, vm_page) alloctail;
1777 	u_long npages;
1778 	vm_offset_t retkva, zkva;
1779 	vm_page_t p, p_next;
1780 	uma_keg_t keg;
1781 
1782 	TAILQ_INIT(&alloctail);
1783 	keg = zone->uz_keg;
1784 
1785 	npages = howmany(bytes, PAGE_SIZE);
1786 	while (npages > 0) {
1787 		p = vm_page_alloc_domain(NULL, 0, domain, VM_ALLOC_INTERRUPT |
1788 		    VM_ALLOC_WIRED | VM_ALLOC_NOOBJ |
1789 		    ((wait & M_WAITOK) != 0 ? VM_ALLOC_WAITOK :
1790 		    VM_ALLOC_NOWAIT));
1791 		if (p != NULL) {
1792 			/*
1793 			 * Since the page does not belong to an object, its
1794 			 * listq is unused.
1795 			 */
1796 			TAILQ_INSERT_TAIL(&alloctail, p, listq);
1797 			npages--;
1798 			continue;
1799 		}
1800 		/*
1801 		 * Page allocation failed, free intermediate pages and
1802 		 * exit.
1803 		 */
1804 		TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1805 			vm_page_unwire_noq(p);
1806 			vm_page_free(p);
1807 		}
1808 		return (NULL);
1809 	}
1810 	*flags = UMA_SLAB_PRIV;
1811 	zkva = keg->uk_kva +
1812 	    atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
1813 	retkva = zkva;
1814 	TAILQ_FOREACH(p, &alloctail, listq) {
1815 		pmap_qenter(zkva, &p, 1);
1816 		zkva += PAGE_SIZE;
1817 	}
1818 
1819 	return ((void *)retkva);
1820 }
1821 
1822 /*
1823  * Allocate physically contiguous pages.
1824  */
1825 static void *
1826 contig_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1827     int wait)
1828 {
1829 
1830 	*pflag = UMA_SLAB_KERNEL;
1831 	return ((void *)kmem_alloc_contig_domainset(DOMAINSET_FIXED(domain),
1832 	    bytes, wait, 0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT));
1833 }
1834 
1835 /*
1836  * Frees a number of pages to the system
1837  *
1838  * Arguments:
1839  *	mem   A pointer to the memory to be freed
1840  *	size  The size of the memory being freed
1841  *	flags The original p->us_flags field
1842  *
1843  * Returns:
1844  *	Nothing
1845  */
1846 static void
1847 page_free(void *mem, vm_size_t size, uint8_t flags)
1848 {
1849 
1850 	if ((flags & UMA_SLAB_BOOT) != 0) {
1851 		startup_free(mem, size);
1852 		return;
1853 	}
1854 
1855 	KASSERT((flags & UMA_SLAB_KERNEL) != 0,
1856 	    ("UMA: page_free used with invalid flags %x", flags));
1857 
1858 	kmem_free((vm_offset_t)mem, size);
1859 }
1860 
1861 /*
1862  * Frees pcpu zone allocations
1863  *
1864  * Arguments:
1865  *	mem   A pointer to the memory to be freed
1866  *	size  The size of the memory being freed
1867  *	flags The original p->us_flags field
1868  *
1869  * Returns:
1870  *	Nothing
1871  */
1872 static void
1873 pcpu_page_free(void *mem, vm_size_t size, uint8_t flags)
1874 {
1875 	vm_offset_t sva, curva;
1876 	vm_paddr_t paddr;
1877 	vm_page_t m;
1878 
1879 	MPASS(size == (mp_maxid+1)*PAGE_SIZE);
1880 
1881 	if ((flags & UMA_SLAB_BOOT) != 0) {
1882 		startup_free(mem, size);
1883 		return;
1884 	}
1885 
1886 	sva = (vm_offset_t)mem;
1887 	for (curva = sva; curva < sva + size; curva += PAGE_SIZE) {
1888 		paddr = pmap_kextract(curva);
1889 		m = PHYS_TO_VM_PAGE(paddr);
1890 		vm_page_unwire_noq(m);
1891 		vm_page_free(m);
1892 	}
1893 	pmap_qremove(sva, size >> PAGE_SHIFT);
1894 	kva_free(sva, size);
1895 }
1896 
1897 
1898 /*
1899  * Zero fill initializer
1900  *
1901  * Arguments/Returns follow uma_init specifications
1902  */
1903 static int
1904 zero_init(void *mem, int size, int flags)
1905 {
1906 	bzero(mem, size);
1907 	return (0);
1908 }
1909 
1910 #ifdef INVARIANTS
1911 static struct noslabbits *
1912 slab_dbg_bits(uma_slab_t slab, uma_keg_t keg)
1913 {
1914 
1915 	return ((void *)((char *)&slab->us_free + BITSET_SIZE(keg->uk_ipers)));
1916 }
1917 #endif
1918 
1919 /*
1920  * Actual size of embedded struct slab (!OFFPAGE).
1921  */
1922 static size_t
1923 slab_sizeof(int nitems)
1924 {
1925 	size_t s;
1926 
1927 	s = sizeof(struct uma_slab) + BITSET_SIZE(nitems) * SLAB_BITSETS;
1928 	return (roundup(s, UMA_ALIGN_PTR + 1));
1929 }
1930 
1931 #define	UMA_FIXPT_SHIFT	31
1932 #define	UMA_FRAC_FIXPT(n, d)						\
1933 	((uint32_t)(((uint64_t)(n) << UMA_FIXPT_SHIFT) / (d)))
1934 #define	UMA_FIXPT_PCT(f)						\
1935 	((u_int)(((uint64_t)100 * (f)) >> UMA_FIXPT_SHIFT))
1936 #define	UMA_PCT_FIXPT(pct)	UMA_FRAC_FIXPT((pct), 100)
1937 #define	UMA_MIN_EFF	UMA_PCT_FIXPT(100 - UMA_MAX_WASTE)
1938 
1939 /*
1940  * Compute the number of items that will fit in a slab.  If hdr is true, the
1941  * item count may be limited to provide space in the slab for an inline slab
1942  * header.  Otherwise, all slab space will be provided for item storage.
1943  */
1944 static u_int
1945 slab_ipers_hdr(u_int size, u_int rsize, u_int slabsize, bool hdr)
1946 {
1947 	u_int ipers;
1948 	u_int padpi;
1949 
1950 	/* The padding between items is not needed after the last item. */
1951 	padpi = rsize - size;
1952 
1953 	if (hdr) {
1954 		/*
1955 		 * Start with the maximum item count and remove items until
1956 		 * the slab header first alongside the allocatable memory.
1957 		 */
1958 		for (ipers = MIN(SLAB_MAX_SETSIZE,
1959 		    (slabsize + padpi - slab_sizeof(1)) / rsize);
1960 		    ipers > 0 &&
1961 		    ipers * rsize - padpi + slab_sizeof(ipers) > slabsize;
1962 		    ipers--)
1963 			continue;
1964 	} else {
1965 		ipers = MIN((slabsize + padpi) / rsize, SLAB_MAX_SETSIZE);
1966 	}
1967 
1968 	return (ipers);
1969 }
1970 
1971 struct keg_layout_result {
1972 	u_int format;
1973 	u_int slabsize;
1974 	u_int ipers;
1975 	u_int eff;
1976 };
1977 
1978 static void
1979 keg_layout_one(uma_keg_t keg, u_int rsize, u_int slabsize, u_int fmt,
1980     struct keg_layout_result *kl)
1981 {
1982 	u_int total;
1983 
1984 	kl->format = fmt;
1985 	kl->slabsize = slabsize;
1986 
1987 	/* Handle INTERNAL as inline with an extra page. */
1988 	if ((fmt & UMA_ZFLAG_INTERNAL) != 0) {
1989 		kl->format &= ~UMA_ZFLAG_INTERNAL;
1990 		kl->slabsize += PAGE_SIZE;
1991 	}
1992 
1993 	kl->ipers = slab_ipers_hdr(keg->uk_size, rsize, kl->slabsize,
1994 	    (fmt & UMA_ZFLAG_OFFPAGE) == 0);
1995 
1996 	/* Account for memory used by an offpage slab header. */
1997 	total = kl->slabsize;
1998 	if ((fmt & UMA_ZFLAG_OFFPAGE) != 0)
1999 		total += slabzone(kl->ipers)->uz_keg->uk_rsize;
2000 
2001 	kl->eff = UMA_FRAC_FIXPT(kl->ipers * rsize, total);
2002 }
2003 
2004 /*
2005  * Determine the format of a uma keg.  This determines where the slab header
2006  * will be placed (inline or offpage) and calculates ipers, rsize, and ppera.
2007  *
2008  * Arguments
2009  *	keg  The zone we should initialize
2010  *
2011  * Returns
2012  *	Nothing
2013  */
2014 static void
2015 keg_layout(uma_keg_t keg)
2016 {
2017 	struct keg_layout_result kl = {}, kl_tmp;
2018 	u_int fmts[2];
2019 	u_int alignsize;
2020 	u_int nfmt;
2021 	u_int pages;
2022 	u_int rsize;
2023 	u_int slabsize;
2024 	u_int i, j;
2025 
2026 	KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
2027 	    (keg->uk_size <= UMA_PCPU_ALLOC_SIZE &&
2028 	     (keg->uk_flags & UMA_ZONE_CACHESPREAD) == 0),
2029 	    ("%s: cannot configure for PCPU: keg=%s, size=%u, flags=0x%b",
2030 	     __func__, keg->uk_name, keg->uk_size, keg->uk_flags,
2031 	     PRINT_UMA_ZFLAGS));
2032 	KASSERT((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) == 0 ||
2033 	    (keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0,
2034 	    ("%s: incompatible flags 0x%b", __func__, keg->uk_flags,
2035 	     PRINT_UMA_ZFLAGS));
2036 
2037 	alignsize = keg->uk_align + 1;
2038 
2039 	/*
2040 	 * Calculate the size of each allocation (rsize) according to
2041 	 * alignment.  If the requested size is smaller than we have
2042 	 * allocation bits for we round it up.
2043 	 */
2044 	rsize = MAX(keg->uk_size, UMA_SMALLEST_UNIT);
2045 	rsize = roundup2(rsize, alignsize);
2046 
2047 	if ((keg->uk_flags & UMA_ZONE_CACHESPREAD) != 0) {
2048 		/*
2049 		 * We want one item to start on every align boundary in a page.
2050 		 * To do this we will span pages.  We will also extend the item
2051 		 * by the size of align if it is an even multiple of align.
2052 		 * Otherwise, it would fall on the same boundary every time.
2053 		 */
2054 		if ((rsize & alignsize) == 0)
2055 			rsize += alignsize;
2056 		slabsize = rsize * (PAGE_SIZE / alignsize);
2057 		slabsize = MIN(slabsize, rsize * SLAB_MAX_SETSIZE);
2058 		slabsize = MIN(slabsize, UMA_CACHESPREAD_MAX_SIZE);
2059 		slabsize = round_page(slabsize);
2060 	} else {
2061 		/*
2062 		 * Start with a slab size of as many pages as it takes to
2063 		 * represent a single item.  We will try to fit as many
2064 		 * additional items into the slab as possible.
2065 		 */
2066 		slabsize = round_page(keg->uk_size);
2067 	}
2068 
2069 	/* Build a list of all of the available formats for this keg. */
2070 	nfmt = 0;
2071 
2072 	/* Evaluate an inline slab layout. */
2073 	if ((keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0)
2074 		fmts[nfmt++] = 0;
2075 
2076 	/* TODO: vm_page-embedded slab. */
2077 
2078 	/*
2079 	 * We can't do OFFPAGE if we're internal or if we've been
2080 	 * asked to not go to the VM for buckets.  If we do this we
2081 	 * may end up going to the VM for slabs which we do not want
2082 	 * to do if we're UMA_ZONE_VM, which clearly forbids it.
2083 	 * In those cases, evaluate a pseudo-format called INTERNAL
2084 	 * which has an inline slab header and one extra page to
2085 	 * guarantee that it fits.
2086 	 *
2087 	 * Otherwise, see if using an OFFPAGE slab will improve our
2088 	 * efficiency.
2089 	 */
2090 	if ((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) != 0)
2091 		fmts[nfmt++] = UMA_ZFLAG_INTERNAL;
2092 	else
2093 		fmts[nfmt++] = UMA_ZFLAG_OFFPAGE;
2094 
2095 	/*
2096 	 * Choose a slab size and format which satisfy the minimum efficiency.
2097 	 * Prefer the smallest slab size that meets the constraints.
2098 	 *
2099 	 * Start with a minimum slab size, to accommodate CACHESPREAD.  Then,
2100 	 * for small items (up to PAGE_SIZE), the iteration increment is one
2101 	 * page; and for large items, the increment is one item.
2102 	 */
2103 	i = (slabsize + rsize - keg->uk_size) / MAX(PAGE_SIZE, rsize);
2104 	KASSERT(i >= 1, ("keg %s(%p) flags=0x%b slabsize=%u, rsize=%u, i=%u",
2105 	    keg->uk_name, keg, keg->uk_flags, PRINT_UMA_ZFLAGS, slabsize,
2106 	    rsize, i));
2107 	for ( ; ; i++) {
2108 		slabsize = (rsize <= PAGE_SIZE) ? ptoa(i) :
2109 		    round_page(rsize * (i - 1) + keg->uk_size);
2110 
2111 		for (j = 0; j < nfmt; j++) {
2112 			/* Only if we have no viable format yet. */
2113 			if ((fmts[j] & UMA_ZFLAG_INTERNAL) != 0 &&
2114 			    kl.ipers > 0)
2115 				continue;
2116 
2117 			keg_layout_one(keg, rsize, slabsize, fmts[j], &kl_tmp);
2118 			if (kl_tmp.eff <= kl.eff)
2119 				continue;
2120 
2121 			kl = kl_tmp;
2122 
2123 			CTR6(KTR_UMA, "keg %s layout: format %#x "
2124 			    "(ipers %u * rsize %u) / slabsize %#x = %u%% eff",
2125 			    keg->uk_name, kl.format, kl.ipers, rsize,
2126 			    kl.slabsize, UMA_FIXPT_PCT(kl.eff));
2127 
2128 			/* Stop when we reach the minimum efficiency. */
2129 			if (kl.eff >= UMA_MIN_EFF)
2130 				break;
2131 		}
2132 
2133 		if (kl.eff >= UMA_MIN_EFF || !multipage_slabs ||
2134 		    slabsize >= SLAB_MAX_SETSIZE * rsize ||
2135 		    (keg->uk_flags & (UMA_ZONE_PCPU | UMA_ZONE_CONTIG)) != 0)
2136 			break;
2137 	}
2138 
2139 	pages = atop(kl.slabsize);
2140 	if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
2141 		pages *= mp_maxid + 1;
2142 
2143 	keg->uk_rsize = rsize;
2144 	keg->uk_ipers = kl.ipers;
2145 	keg->uk_ppera = pages;
2146 	keg->uk_flags |= kl.format;
2147 
2148 	/*
2149 	 * How do we find the slab header if it is offpage or if not all item
2150 	 * start addresses are in the same page?  We could solve the latter
2151 	 * case with vaddr alignment, but we don't.
2152 	 */
2153 	if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0 ||
2154 	    (keg->uk_ipers - 1) * rsize >= PAGE_SIZE) {
2155 		if ((keg->uk_flags & UMA_ZONE_NOTPAGE) != 0)
2156 			keg->uk_flags |= UMA_ZFLAG_HASH;
2157 		else
2158 			keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2159 	}
2160 
2161 	CTR6(KTR_UMA, "%s: keg=%s, flags=%#x, rsize=%u, ipers=%u, ppera=%u",
2162 	    __func__, keg->uk_name, keg->uk_flags, rsize, keg->uk_ipers,
2163 	    pages);
2164 	KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_MAX_SETSIZE,
2165 	    ("%s: keg=%s, flags=0x%b, rsize=%u, ipers=%u, ppera=%u", __func__,
2166 	     keg->uk_name, keg->uk_flags, PRINT_UMA_ZFLAGS, rsize,
2167 	     keg->uk_ipers, pages));
2168 }
2169 
2170 /*
2171  * Keg header ctor.  This initializes all fields, locks, etc.  And inserts
2172  * the keg onto the global keg list.
2173  *
2174  * Arguments/Returns follow uma_ctor specifications
2175  *	udata  Actually uma_kctor_args
2176  */
2177 static int
2178 keg_ctor(void *mem, int size, void *udata, int flags)
2179 {
2180 	struct uma_kctor_args *arg = udata;
2181 	uma_keg_t keg = mem;
2182 	uma_zone_t zone;
2183 	int i;
2184 
2185 	bzero(keg, size);
2186 	keg->uk_size = arg->size;
2187 	keg->uk_init = arg->uminit;
2188 	keg->uk_fini = arg->fini;
2189 	keg->uk_align = arg->align;
2190 	keg->uk_reserve = 0;
2191 	keg->uk_flags = arg->flags;
2192 
2193 	/*
2194 	 * We use a global round-robin policy by default.  Zones with
2195 	 * UMA_ZONE_FIRSTTOUCH set will use first-touch instead, in which
2196 	 * case the iterator is never run.
2197 	 */
2198 	keg->uk_dr.dr_policy = DOMAINSET_RR();
2199 	keg->uk_dr.dr_iter = 0;
2200 
2201 	/*
2202 	 * The master zone is passed to us at keg-creation time.
2203 	 */
2204 	zone = arg->zone;
2205 	keg->uk_name = zone->uz_name;
2206 
2207 	if (arg->flags & UMA_ZONE_ZINIT)
2208 		keg->uk_init = zero_init;
2209 
2210 	if (arg->flags & UMA_ZONE_MALLOC)
2211 		keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2212 
2213 #ifndef SMP
2214 	keg->uk_flags &= ~UMA_ZONE_PCPU;
2215 #endif
2216 
2217 	keg_layout(keg);
2218 
2219 	/*
2220 	 * Use a first-touch NUMA policy for kegs that pmap_extract() will
2221 	 * work on.  Use round-robin for everything else.
2222 	 *
2223 	 * Zones may override the default by specifying either.
2224 	 */
2225 #ifdef NUMA
2226 	if ((keg->uk_flags &
2227 	    (UMA_ZONE_ROUNDROBIN | UMA_ZFLAG_CACHE | UMA_ZONE_NOTPAGE)) == 0)
2228 		keg->uk_flags |= UMA_ZONE_FIRSTTOUCH;
2229 	else if ((keg->uk_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2230 		keg->uk_flags |= UMA_ZONE_ROUNDROBIN;
2231 #endif
2232 
2233 	/*
2234 	 * If we haven't booted yet we need allocations to go through the
2235 	 * startup cache until the vm is ready.
2236 	 */
2237 #ifdef UMA_MD_SMALL_ALLOC
2238 	if (keg->uk_ppera == 1)
2239 		keg->uk_allocf = uma_small_alloc;
2240 	else
2241 #endif
2242 	if (booted < BOOT_KVA)
2243 		keg->uk_allocf = startup_alloc;
2244 	else if (keg->uk_flags & UMA_ZONE_PCPU)
2245 		keg->uk_allocf = pcpu_page_alloc;
2246 	else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 && keg->uk_ppera > 1)
2247 		keg->uk_allocf = contig_alloc;
2248 	else
2249 		keg->uk_allocf = page_alloc;
2250 #ifdef UMA_MD_SMALL_ALLOC
2251 	if (keg->uk_ppera == 1)
2252 		keg->uk_freef = uma_small_free;
2253 	else
2254 #endif
2255 	if (keg->uk_flags & UMA_ZONE_PCPU)
2256 		keg->uk_freef = pcpu_page_free;
2257 	else
2258 		keg->uk_freef = page_free;
2259 
2260 	/*
2261 	 * Initialize keg's locks.
2262 	 */
2263 	for (i = 0; i < vm_ndomains; i++)
2264 		KEG_LOCK_INIT(keg, i, (arg->flags & UMA_ZONE_MTXCLASS));
2265 
2266 	/*
2267 	 * If we're putting the slab header in the actual page we need to
2268 	 * figure out where in each page it goes.  See slab_sizeof
2269 	 * definition.
2270 	 */
2271 	if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE)) {
2272 		size_t shsize;
2273 
2274 		shsize = slab_sizeof(keg->uk_ipers);
2275 		keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - shsize;
2276 		/*
2277 		 * The only way the following is possible is if with our
2278 		 * UMA_ALIGN_PTR adjustments we are now bigger than
2279 		 * UMA_SLAB_SIZE.  I haven't checked whether this is
2280 		 * mathematically possible for all cases, so we make
2281 		 * sure here anyway.
2282 		 */
2283 		KASSERT(keg->uk_pgoff + shsize <= PAGE_SIZE * keg->uk_ppera,
2284 		    ("zone %s ipers %d rsize %d size %d slab won't fit",
2285 		    zone->uz_name, keg->uk_ipers, keg->uk_rsize, keg->uk_size));
2286 	}
2287 
2288 	if (keg->uk_flags & UMA_ZFLAG_HASH)
2289 		hash_alloc(&keg->uk_hash, 0);
2290 
2291 	CTR3(KTR_UMA, "keg_ctor %p zone %s(%p)", keg, zone->uz_name, zone);
2292 
2293 	LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
2294 
2295 	rw_wlock(&uma_rwlock);
2296 	LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
2297 	rw_wunlock(&uma_rwlock);
2298 	return (0);
2299 }
2300 
2301 static void
2302 zone_kva_available(uma_zone_t zone, void *unused)
2303 {
2304 	uma_keg_t keg;
2305 
2306 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
2307 		return;
2308 	KEG_GET(zone, keg);
2309 
2310 	if (keg->uk_allocf == startup_alloc) {
2311 		/* Switch to the real allocator. */
2312 		if (keg->uk_flags & UMA_ZONE_PCPU)
2313 			keg->uk_allocf = pcpu_page_alloc;
2314 		else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 &&
2315 		    keg->uk_ppera > 1)
2316 			keg->uk_allocf = contig_alloc;
2317 		else
2318 			keg->uk_allocf = page_alloc;
2319 	}
2320 }
2321 
2322 static void
2323 zone_alloc_counters(uma_zone_t zone, void *unused)
2324 {
2325 
2326 	zone->uz_allocs = counter_u64_alloc(M_WAITOK);
2327 	zone->uz_frees = counter_u64_alloc(M_WAITOK);
2328 	zone->uz_fails = counter_u64_alloc(M_WAITOK);
2329 	zone->uz_xdomain = counter_u64_alloc(M_WAITOK);
2330 }
2331 
2332 static void
2333 zone_alloc_sysctl(uma_zone_t zone, void *unused)
2334 {
2335 	uma_zone_domain_t zdom;
2336 	uma_domain_t dom;
2337 	uma_keg_t keg;
2338 	struct sysctl_oid *oid, *domainoid;
2339 	int domains, i, cnt;
2340 	static const char *nokeg = "cache zone";
2341 	char *c;
2342 
2343 	/*
2344 	 * Make a sysctl safe copy of the zone name by removing
2345 	 * any special characters and handling dups by appending
2346 	 * an index.
2347 	 */
2348 	if (zone->uz_namecnt != 0) {
2349 		/* Count the number of decimal digits and '_' separator. */
2350 		for (i = 1, cnt = zone->uz_namecnt; cnt != 0; i++)
2351 			cnt /= 10;
2352 		zone->uz_ctlname = malloc(strlen(zone->uz_name) + i + 1,
2353 		    M_UMA, M_WAITOK);
2354 		sprintf(zone->uz_ctlname, "%s_%d", zone->uz_name,
2355 		    zone->uz_namecnt);
2356 	} else
2357 		zone->uz_ctlname = strdup(zone->uz_name, M_UMA);
2358 	for (c = zone->uz_ctlname; *c != '\0'; c++)
2359 		if (strchr("./\\ -", *c) != NULL)
2360 			*c = '_';
2361 
2362 	/*
2363 	 * Basic parameters at the root.
2364 	 */
2365 	zone->uz_oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_vm_uma),
2366 	    OID_AUTO, zone->uz_ctlname, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2367 	oid = zone->uz_oid;
2368 	SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2369 	    "size", CTLFLAG_RD, &zone->uz_size, 0, "Allocation size");
2370 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2371 	    "flags", CTLFLAG_RD | CTLTYPE_STRING | CTLFLAG_MPSAFE,
2372 	    zone, 0, sysctl_handle_uma_zone_flags, "A",
2373 	    "Allocator configuration flags");
2374 	SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2375 	    "bucket_size", CTLFLAG_RD, &zone->uz_bucket_size, 0,
2376 	    "Desired per-cpu cache size");
2377 	SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2378 	    "bucket_size_max", CTLFLAG_RD, &zone->uz_bucket_size_max, 0,
2379 	    "Maximum allowed per-cpu cache size");
2380 
2381 	/*
2382 	 * keg if present.
2383 	 */
2384 	if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
2385 		domains = vm_ndomains;
2386 	else
2387 		domains = 1;
2388 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2389 	    "keg", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2390 	keg = zone->uz_keg;
2391 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0) {
2392 		SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2393 		    "name", CTLFLAG_RD, keg->uk_name, "Keg name");
2394 		SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2395 		    "rsize", CTLFLAG_RD, &keg->uk_rsize, 0,
2396 		    "Real object size with alignment");
2397 		SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2398 		    "ppera", CTLFLAG_RD, &keg->uk_ppera, 0,
2399 		    "pages per-slab allocation");
2400 		SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2401 		    "ipers", CTLFLAG_RD, &keg->uk_ipers, 0,
2402 		    "items available per-slab");
2403 		SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2404 		    "align", CTLFLAG_RD, &keg->uk_align, 0,
2405 		    "item alignment mask");
2406 		SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2407 		    "efficiency", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2408 		    keg, 0, sysctl_handle_uma_slab_efficiency, "I",
2409 		    "Slab utilization (100 - internal fragmentation %)");
2410 		domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(oid),
2411 		    OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2412 		for (i = 0; i < domains; i++) {
2413 			dom = &keg->uk_domain[i];
2414 			oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2415 			    OID_AUTO, VM_DOMAIN(i)->vmd_name,
2416 			    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2417 			SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2418 			    "pages", CTLFLAG_RD, &dom->ud_pages, 0,
2419 			    "Total pages currently allocated from VM");
2420 			SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2421 			    "free_items", CTLFLAG_RD, &dom->ud_free_items, 0,
2422 			    "items free in the slab layer");
2423 		}
2424 	} else
2425 		SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2426 		    "name", CTLFLAG_RD, nokeg, "Keg name");
2427 
2428 	/*
2429 	 * Information about zone limits.
2430 	 */
2431 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2432 	    "limit", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2433 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2434 	    "items", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2435 	    zone, 0, sysctl_handle_uma_zone_items, "QU",
2436 	    "current number of allocated items if limit is set");
2437 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2438 	    "max_items", CTLFLAG_RD, &zone->uz_max_items, 0,
2439 	    "Maximum number of cached items");
2440 	SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2441 	    "sleepers", CTLFLAG_RD, &zone->uz_sleepers, 0,
2442 	    "Number of threads sleeping at limit");
2443 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2444 	    "sleeps", CTLFLAG_RD, &zone->uz_sleeps, 0,
2445 	    "Total zone limit sleeps");
2446 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2447 	    "bucket_max", CTLFLAG_RD, &zone->uz_bucket_max, 0,
2448 	    "Maximum number of items in each domain's bucket cache");
2449 
2450 	/*
2451 	 * Per-domain zone information.
2452 	 */
2453 	domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid),
2454 	    OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2455 	for (i = 0; i < domains; i++) {
2456 		zdom = ZDOM_GET(zone, i);
2457 		oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2458 		    OID_AUTO, VM_DOMAIN(i)->vmd_name,
2459 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2460 		SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2461 		    "nitems", CTLFLAG_RD, &zdom->uzd_nitems,
2462 		    "number of items in this domain");
2463 		SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2464 		    "imax", CTLFLAG_RD, &zdom->uzd_imax,
2465 		    "maximum item count in this period");
2466 		SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2467 		    "imin", CTLFLAG_RD, &zdom->uzd_imin,
2468 		    "minimum item count in this period");
2469 		SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2470 		    "wss", CTLFLAG_RD, &zdom->uzd_wss,
2471 		    "Working set size");
2472 	}
2473 
2474 	/*
2475 	 * General statistics.
2476 	 */
2477 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2478 	    "stats", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2479 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2480 	    "current", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2481 	    zone, 1, sysctl_handle_uma_zone_cur, "I",
2482 	    "Current number of allocated items");
2483 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2484 	    "allocs", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2485 	    zone, 0, sysctl_handle_uma_zone_allocs, "QU",
2486 	    "Total allocation calls");
2487 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2488 	    "frees", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2489 	    zone, 0, sysctl_handle_uma_zone_frees, "QU",
2490 	    "Total free calls");
2491 	SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2492 	    "fails", CTLFLAG_RD, &zone->uz_fails,
2493 	    "Number of allocation failures");
2494 	SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2495 	    "xdomain", CTLFLAG_RD, &zone->uz_xdomain,
2496 	    "Free calls from the wrong domain");
2497 }
2498 
2499 struct uma_zone_count {
2500 	const char	*name;
2501 	int		count;
2502 };
2503 
2504 static void
2505 zone_count(uma_zone_t zone, void *arg)
2506 {
2507 	struct uma_zone_count *cnt;
2508 
2509 	cnt = arg;
2510 	/*
2511 	 * Some zones are rapidly created with identical names and
2512 	 * destroyed out of order.  This can lead to gaps in the count.
2513 	 * Use one greater than the maximum observed for this name.
2514 	 */
2515 	if (strcmp(zone->uz_name, cnt->name) == 0)
2516 		cnt->count = MAX(cnt->count,
2517 		    zone->uz_namecnt + 1);
2518 }
2519 
2520 static void
2521 zone_update_caches(uma_zone_t zone)
2522 {
2523 	int i;
2524 
2525 	for (i = 0; i <= mp_maxid; i++) {
2526 		cache_set_uz_size(&zone->uz_cpu[i], zone->uz_size);
2527 		cache_set_uz_flags(&zone->uz_cpu[i], zone->uz_flags);
2528 	}
2529 }
2530 
2531 /*
2532  * Zone header ctor.  This initializes all fields, locks, etc.
2533  *
2534  * Arguments/Returns follow uma_ctor specifications
2535  *	udata  Actually uma_zctor_args
2536  */
2537 static int
2538 zone_ctor(void *mem, int size, void *udata, int flags)
2539 {
2540 	struct uma_zone_count cnt;
2541 	struct uma_zctor_args *arg = udata;
2542 	uma_zone_domain_t zdom;
2543 	uma_zone_t zone = mem;
2544 	uma_zone_t z;
2545 	uma_keg_t keg;
2546 	int i;
2547 
2548 	bzero(zone, size);
2549 	zone->uz_name = arg->name;
2550 	zone->uz_ctor = arg->ctor;
2551 	zone->uz_dtor = arg->dtor;
2552 	zone->uz_init = NULL;
2553 	zone->uz_fini = NULL;
2554 	zone->uz_sleeps = 0;
2555 	zone->uz_bucket_size = 0;
2556 	zone->uz_bucket_size_min = 0;
2557 	zone->uz_bucket_size_max = BUCKET_MAX;
2558 	zone->uz_flags = (arg->flags & UMA_ZONE_SMR);
2559 	zone->uz_warning = NULL;
2560 	/* The domain structures follow the cpu structures. */
2561 	zone->uz_bucket_max = ULONG_MAX;
2562 	timevalclear(&zone->uz_ratecheck);
2563 
2564 	/* Count the number of duplicate names. */
2565 	cnt.name = arg->name;
2566 	cnt.count = 0;
2567 	zone_foreach(zone_count, &cnt);
2568 	zone->uz_namecnt = cnt.count;
2569 	ZONE_CROSS_LOCK_INIT(zone);
2570 
2571 	for (i = 0; i < vm_ndomains; i++) {
2572 		zdom = ZDOM_GET(zone, i);
2573 		ZDOM_LOCK_INIT(zone, zdom, (arg->flags & UMA_ZONE_MTXCLASS));
2574 		STAILQ_INIT(&zdom->uzd_buckets);
2575 	}
2576 
2577 #ifdef INVARIANTS
2578 	if (arg->uminit == trash_init && arg->fini == trash_fini)
2579 		zone->uz_flags |= UMA_ZFLAG_TRASH | UMA_ZFLAG_CTORDTOR;
2580 #endif
2581 
2582 	/*
2583 	 * This is a pure cache zone, no kegs.
2584 	 */
2585 	if (arg->import) {
2586 		KASSERT((arg->flags & UMA_ZFLAG_CACHE) != 0,
2587 		    ("zone_ctor: Import specified for non-cache zone."));
2588 		zone->uz_flags = arg->flags;
2589 		zone->uz_size = arg->size;
2590 		zone->uz_import = arg->import;
2591 		zone->uz_release = arg->release;
2592 		zone->uz_arg = arg->arg;
2593 #ifdef NUMA
2594 		/*
2595 		 * Cache zones are round-robin unless a policy is
2596 		 * specified because they may have incompatible
2597 		 * constraints.
2598 		 */
2599 		if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2600 			zone->uz_flags |= UMA_ZONE_ROUNDROBIN;
2601 #endif
2602 		rw_wlock(&uma_rwlock);
2603 		LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link);
2604 		rw_wunlock(&uma_rwlock);
2605 		goto out;
2606 	}
2607 
2608 	/*
2609 	 * Use the regular zone/keg/slab allocator.
2610 	 */
2611 	zone->uz_import = zone_import;
2612 	zone->uz_release = zone_release;
2613 	zone->uz_arg = zone;
2614 	keg = arg->keg;
2615 
2616 	if (arg->flags & UMA_ZONE_SECONDARY) {
2617 		KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
2618 		    ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
2619 		KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
2620 		zone->uz_init = arg->uminit;
2621 		zone->uz_fini = arg->fini;
2622 		zone->uz_flags |= UMA_ZONE_SECONDARY;
2623 		rw_wlock(&uma_rwlock);
2624 		ZONE_LOCK(zone);
2625 		LIST_FOREACH(z, &keg->uk_zones, uz_link) {
2626 			if (LIST_NEXT(z, uz_link) == NULL) {
2627 				LIST_INSERT_AFTER(z, zone, uz_link);
2628 				break;
2629 			}
2630 		}
2631 		ZONE_UNLOCK(zone);
2632 		rw_wunlock(&uma_rwlock);
2633 	} else if (keg == NULL) {
2634 		if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
2635 		    arg->align, arg->flags)) == NULL)
2636 			return (ENOMEM);
2637 	} else {
2638 		struct uma_kctor_args karg;
2639 		int error;
2640 
2641 		/* We should only be here from uma_startup() */
2642 		karg.size = arg->size;
2643 		karg.uminit = arg->uminit;
2644 		karg.fini = arg->fini;
2645 		karg.align = arg->align;
2646 		karg.flags = (arg->flags & ~UMA_ZONE_SMR);
2647 		karg.zone = zone;
2648 		error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
2649 		    flags);
2650 		if (error)
2651 			return (error);
2652 	}
2653 
2654 	/* Inherit properties from the keg. */
2655 	zone->uz_keg = keg;
2656 	zone->uz_size = keg->uk_size;
2657 	zone->uz_flags |= (keg->uk_flags &
2658 	    (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
2659 
2660 out:
2661 	if (__predict_true(booted >= BOOT_RUNNING)) {
2662 		zone_alloc_counters(zone, NULL);
2663 		zone_alloc_sysctl(zone, NULL);
2664 	} else {
2665 		zone->uz_allocs = EARLY_COUNTER;
2666 		zone->uz_frees = EARLY_COUNTER;
2667 		zone->uz_fails = EARLY_COUNTER;
2668 	}
2669 
2670 	/* Caller requests a private SMR context. */
2671 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
2672 		zone->uz_smr = smr_create(zone->uz_name, 0, 0);
2673 
2674 	KASSERT((arg->flags & (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET)) !=
2675 	    (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET),
2676 	    ("Invalid zone flag combination"));
2677 	if (arg->flags & UMA_ZFLAG_INTERNAL)
2678 		zone->uz_bucket_size_max = zone->uz_bucket_size = 0;
2679 	if ((arg->flags & UMA_ZONE_MAXBUCKET) != 0)
2680 		zone->uz_bucket_size = BUCKET_MAX;
2681 	else if ((arg->flags & UMA_ZONE_MINBUCKET) != 0)
2682 		zone->uz_bucket_size_max = zone->uz_bucket_size = BUCKET_MIN;
2683 	else if ((arg->flags & UMA_ZONE_NOBUCKET) != 0)
2684 		zone->uz_bucket_size = 0;
2685 	else
2686 		zone->uz_bucket_size = bucket_select(zone->uz_size);
2687 	zone->uz_bucket_size_min = zone->uz_bucket_size;
2688 	if (zone->uz_dtor != NULL || zone->uz_ctor != NULL)
2689 		zone->uz_flags |= UMA_ZFLAG_CTORDTOR;
2690 	zone_update_caches(zone);
2691 
2692 	return (0);
2693 }
2694 
2695 /*
2696  * Keg header dtor.  This frees all data, destroys locks, frees the hash
2697  * table and removes the keg from the global list.
2698  *
2699  * Arguments/Returns follow uma_dtor specifications
2700  *	udata  unused
2701  */
2702 static void
2703 keg_dtor(void *arg, int size, void *udata)
2704 {
2705 	uma_keg_t keg;
2706 	uint32_t free, pages;
2707 	int i;
2708 
2709 	keg = (uma_keg_t)arg;
2710 	free = pages = 0;
2711 	for (i = 0; i < vm_ndomains; i++) {
2712 		free += keg->uk_domain[i].ud_free_items;
2713 		pages += keg->uk_domain[i].ud_pages;
2714 		KEG_LOCK_FINI(keg, i);
2715 	}
2716 	if (pages != 0)
2717 		printf("Freed UMA keg (%s) was not empty (%u items). "
2718 		    " Lost %u pages of memory.\n",
2719 		    keg->uk_name ? keg->uk_name : "",
2720 		    pages / keg->uk_ppera * keg->uk_ipers - free, pages);
2721 
2722 	hash_free(&keg->uk_hash);
2723 }
2724 
2725 /*
2726  * Zone header dtor.
2727  *
2728  * Arguments/Returns follow uma_dtor specifications
2729  *	udata  unused
2730  */
2731 static void
2732 zone_dtor(void *arg, int size, void *udata)
2733 {
2734 	uma_zone_t zone;
2735 	uma_keg_t keg;
2736 	int i;
2737 
2738 	zone = (uma_zone_t)arg;
2739 
2740 	sysctl_remove_oid(zone->uz_oid, 1, 1);
2741 
2742 	if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
2743 		cache_drain(zone);
2744 
2745 	rw_wlock(&uma_rwlock);
2746 	LIST_REMOVE(zone, uz_link);
2747 	rw_wunlock(&uma_rwlock);
2748 	zone_reclaim(zone, M_WAITOK, true);
2749 
2750 	/*
2751 	 * We only destroy kegs from non secondary/non cache zones.
2752 	 */
2753 	if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) {
2754 		keg = zone->uz_keg;
2755 		rw_wlock(&uma_rwlock);
2756 		LIST_REMOVE(keg, uk_link);
2757 		rw_wunlock(&uma_rwlock);
2758 		zone_free_item(kegs, keg, NULL, SKIP_NONE);
2759 	}
2760 	counter_u64_free(zone->uz_allocs);
2761 	counter_u64_free(zone->uz_frees);
2762 	counter_u64_free(zone->uz_fails);
2763 	counter_u64_free(zone->uz_xdomain);
2764 	free(zone->uz_ctlname, M_UMA);
2765 	for (i = 0; i < vm_ndomains; i++)
2766 		ZDOM_LOCK_FINI(ZDOM_GET(zone, i));
2767 	ZONE_CROSS_LOCK_FINI(zone);
2768 }
2769 
2770 static void
2771 zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *arg), void *arg)
2772 {
2773 	uma_keg_t keg;
2774 	uma_zone_t zone;
2775 
2776 	LIST_FOREACH(keg, &uma_kegs, uk_link) {
2777 		LIST_FOREACH(zone, &keg->uk_zones, uz_link)
2778 			zfunc(zone, arg);
2779 	}
2780 	LIST_FOREACH(zone, &uma_cachezones, uz_link)
2781 		zfunc(zone, arg);
2782 }
2783 
2784 /*
2785  * Traverses every zone in the system and calls a callback
2786  *
2787  * Arguments:
2788  *	zfunc  A pointer to a function which accepts a zone
2789  *		as an argument.
2790  *
2791  * Returns:
2792  *	Nothing
2793  */
2794 static void
2795 zone_foreach(void (*zfunc)(uma_zone_t, void *arg), void *arg)
2796 {
2797 
2798 	rw_rlock(&uma_rwlock);
2799 	zone_foreach_unlocked(zfunc, arg);
2800 	rw_runlock(&uma_rwlock);
2801 }
2802 
2803 /*
2804  * Initialize the kernel memory allocator.  This is done after pages can be
2805  * allocated but before general KVA is available.
2806  */
2807 void
2808 uma_startup1(vm_offset_t virtual_avail)
2809 {
2810 	struct uma_zctor_args args;
2811 	size_t ksize, zsize, size;
2812 	uma_keg_t masterkeg;
2813 	uintptr_t m;
2814 	uint8_t pflag;
2815 
2816 	bootstart = bootmem = virtual_avail;
2817 
2818 	rw_init(&uma_rwlock, "UMA lock");
2819 	sx_init(&uma_reclaim_lock, "umareclaim");
2820 
2821 	ksize = sizeof(struct uma_keg) +
2822 	    (sizeof(struct uma_domain) * vm_ndomains);
2823 	ksize = roundup(ksize, UMA_SUPER_ALIGN);
2824 	zsize = sizeof(struct uma_zone) +
2825 	    (sizeof(struct uma_cache) * (mp_maxid + 1)) +
2826 	    (sizeof(struct uma_zone_domain) * vm_ndomains);
2827 	zsize = roundup(zsize, UMA_SUPER_ALIGN);
2828 
2829 	/* Allocate the zone of zones, zone of kegs, and zone of zones keg. */
2830 	size = (zsize * 2) + ksize;
2831 	m = (uintptr_t)startup_alloc(NULL, size, 0, &pflag, M_NOWAIT | M_ZERO);
2832 	zones = (uma_zone_t)m;
2833 	m += zsize;
2834 	kegs = (uma_zone_t)m;
2835 	m += zsize;
2836 	masterkeg = (uma_keg_t)m;
2837 
2838 	/* "manually" create the initial zone */
2839 	memset(&args, 0, sizeof(args));
2840 	args.name = "UMA Kegs";
2841 	args.size = ksize;
2842 	args.ctor = keg_ctor;
2843 	args.dtor = keg_dtor;
2844 	args.uminit = zero_init;
2845 	args.fini = NULL;
2846 	args.keg = masterkeg;
2847 	args.align = UMA_SUPER_ALIGN - 1;
2848 	args.flags = UMA_ZFLAG_INTERNAL;
2849 	zone_ctor(kegs, zsize, &args, M_WAITOK);
2850 
2851 	args.name = "UMA Zones";
2852 	args.size = zsize;
2853 	args.ctor = zone_ctor;
2854 	args.dtor = zone_dtor;
2855 	args.uminit = zero_init;
2856 	args.fini = NULL;
2857 	args.keg = NULL;
2858 	args.align = UMA_SUPER_ALIGN - 1;
2859 	args.flags = UMA_ZFLAG_INTERNAL;
2860 	zone_ctor(zones, zsize, &args, M_WAITOK);
2861 
2862 	/* Now make zones for slab headers */
2863 	slabzones[0] = uma_zcreate("UMA Slabs 0", SLABZONE0_SIZE,
2864 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2865 	slabzones[1] = uma_zcreate("UMA Slabs 1", SLABZONE1_SIZE,
2866 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2867 
2868 	hashzone = uma_zcreate("UMA Hash",
2869 	    sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
2870 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2871 
2872 	bucket_init();
2873 	smr_init();
2874 }
2875 
2876 #ifndef UMA_MD_SMALL_ALLOC
2877 extern void vm_radix_reserve_kva(void);
2878 #endif
2879 
2880 /*
2881  * Advertise the availability of normal kva allocations and switch to
2882  * the default back-end allocator.  Marks the KVA we consumed on startup
2883  * as used in the map.
2884  */
2885 void
2886 uma_startup2(void)
2887 {
2888 
2889 	if (bootstart != bootmem) {
2890 		vm_map_lock(kernel_map);
2891 		(void)vm_map_insert(kernel_map, NULL, 0, bootstart, bootmem,
2892 		    VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT);
2893 		vm_map_unlock(kernel_map);
2894 	}
2895 
2896 #ifndef UMA_MD_SMALL_ALLOC
2897 	/* Set up radix zone to use noobj_alloc. */
2898 	vm_radix_reserve_kva();
2899 #endif
2900 
2901 	booted = BOOT_KVA;
2902 	zone_foreach_unlocked(zone_kva_available, NULL);
2903 	bucket_enable();
2904 }
2905 
2906 /*
2907  * Finish our initialization steps.
2908  */
2909 static void
2910 uma_startup3(void)
2911 {
2912 
2913 #ifdef INVARIANTS
2914 	TUNABLE_INT_FETCH("vm.debug.divisor", &dbg_divisor);
2915 	uma_dbg_cnt = counter_u64_alloc(M_WAITOK);
2916 	uma_skip_cnt = counter_u64_alloc(M_WAITOK);
2917 #endif
2918 	zone_foreach_unlocked(zone_alloc_counters, NULL);
2919 	zone_foreach_unlocked(zone_alloc_sysctl, NULL);
2920 	callout_init(&uma_callout, 1);
2921 	callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
2922 	booted = BOOT_RUNNING;
2923 
2924 	EVENTHANDLER_REGISTER(shutdown_post_sync, uma_shutdown, NULL,
2925 	    EVENTHANDLER_PRI_FIRST);
2926 }
2927 
2928 static void
2929 uma_shutdown(void)
2930 {
2931 
2932 	booted = BOOT_SHUTDOWN;
2933 }
2934 
2935 static uma_keg_t
2936 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
2937 		int align, uint32_t flags)
2938 {
2939 	struct uma_kctor_args args;
2940 
2941 	args.size = size;
2942 	args.uminit = uminit;
2943 	args.fini = fini;
2944 	args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
2945 	args.flags = flags;
2946 	args.zone = zone;
2947 	return (zone_alloc_item(kegs, &args, UMA_ANYDOMAIN, M_WAITOK));
2948 }
2949 
2950 /* Public functions */
2951 /* See uma.h */
2952 void
2953 uma_set_align(int align)
2954 {
2955 
2956 	if (align != UMA_ALIGN_CACHE)
2957 		uma_align_cache = align;
2958 }
2959 
2960 /* See uma.h */
2961 uma_zone_t
2962 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
2963 		uma_init uminit, uma_fini fini, int align, uint32_t flags)
2964 
2965 {
2966 	struct uma_zctor_args args;
2967 	uma_zone_t res;
2968 
2969 	KASSERT(powerof2(align + 1), ("invalid zone alignment %d for \"%s\"",
2970 	    align, name));
2971 
2972 	/* This stuff is essential for the zone ctor */
2973 	memset(&args, 0, sizeof(args));
2974 	args.name = name;
2975 	args.size = size;
2976 	args.ctor = ctor;
2977 	args.dtor = dtor;
2978 	args.uminit = uminit;
2979 	args.fini = fini;
2980 #ifdef  INVARIANTS
2981 	/*
2982 	 * Inject procedures which check for memory use after free if we are
2983 	 * allowed to scramble the memory while it is not allocated.  This
2984 	 * requires that: UMA is actually able to access the memory, no init
2985 	 * or fini procedures, no dependency on the initial value of the
2986 	 * memory, and no (legitimate) use of the memory after free.  Note,
2987 	 * the ctor and dtor do not need to be empty.
2988 	 */
2989 	if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOTOUCH |
2990 	    UMA_ZONE_NOFREE))) && uminit == NULL && fini == NULL) {
2991 		args.uminit = trash_init;
2992 		args.fini = trash_fini;
2993 	}
2994 #endif
2995 	args.align = align;
2996 	args.flags = flags;
2997 	args.keg = NULL;
2998 
2999 	sx_slock(&uma_reclaim_lock);
3000 	res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3001 	sx_sunlock(&uma_reclaim_lock);
3002 
3003 	return (res);
3004 }
3005 
3006 /* See uma.h */
3007 uma_zone_t
3008 uma_zsecond_create(const char *name, uma_ctor ctor, uma_dtor dtor,
3009     uma_init zinit, uma_fini zfini, uma_zone_t master)
3010 {
3011 	struct uma_zctor_args args;
3012 	uma_keg_t keg;
3013 	uma_zone_t res;
3014 
3015 	keg = master->uz_keg;
3016 	memset(&args, 0, sizeof(args));
3017 	args.name = name;
3018 	args.size = keg->uk_size;
3019 	args.ctor = ctor;
3020 	args.dtor = dtor;
3021 	args.uminit = zinit;
3022 	args.fini = zfini;
3023 	args.align = keg->uk_align;
3024 	args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
3025 	args.keg = keg;
3026 
3027 	sx_slock(&uma_reclaim_lock);
3028 	res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3029 	sx_sunlock(&uma_reclaim_lock);
3030 
3031 	return (res);
3032 }
3033 
3034 /* See uma.h */
3035 uma_zone_t
3036 uma_zcache_create(const char *name, int size, uma_ctor ctor, uma_dtor dtor,
3037     uma_init zinit, uma_fini zfini, uma_import zimport, uma_release zrelease,
3038     void *arg, int flags)
3039 {
3040 	struct uma_zctor_args args;
3041 
3042 	memset(&args, 0, sizeof(args));
3043 	args.name = name;
3044 	args.size = size;
3045 	args.ctor = ctor;
3046 	args.dtor = dtor;
3047 	args.uminit = zinit;
3048 	args.fini = zfini;
3049 	args.import = zimport;
3050 	args.release = zrelease;
3051 	args.arg = arg;
3052 	args.align = 0;
3053 	args.flags = flags | UMA_ZFLAG_CACHE;
3054 
3055 	return (zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK));
3056 }
3057 
3058 /* See uma.h */
3059 void
3060 uma_zdestroy(uma_zone_t zone)
3061 {
3062 
3063 	/*
3064 	 * Large slabs are expensive to reclaim, so don't bother doing
3065 	 * unnecessary work if we're shutting down.
3066 	 */
3067 	if (booted == BOOT_SHUTDOWN &&
3068 	    zone->uz_fini == NULL && zone->uz_release == zone_release)
3069 		return;
3070 	sx_slock(&uma_reclaim_lock);
3071 	zone_free_item(zones, zone, NULL, SKIP_NONE);
3072 	sx_sunlock(&uma_reclaim_lock);
3073 }
3074 
3075 void
3076 uma_zwait(uma_zone_t zone)
3077 {
3078 
3079 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
3080 		uma_zfree_smr(zone, uma_zalloc_smr(zone, M_WAITOK));
3081 	else if ((zone->uz_flags & UMA_ZONE_PCPU) != 0)
3082 		uma_zfree_pcpu(zone, uma_zalloc_pcpu(zone, M_WAITOK));
3083 	else
3084 		uma_zfree(zone, uma_zalloc(zone, M_WAITOK));
3085 }
3086 
3087 void *
3088 uma_zalloc_pcpu_arg(uma_zone_t zone, void *udata, int flags)
3089 {
3090 	void *item, *pcpu_item;
3091 #ifdef SMP
3092 	int i;
3093 
3094 	MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3095 #endif
3096 	item = uma_zalloc_arg(zone, udata, flags & ~M_ZERO);
3097 	if (item == NULL)
3098 		return (NULL);
3099 	pcpu_item = zpcpu_base_to_offset(item);
3100 	if (flags & M_ZERO) {
3101 #ifdef SMP
3102 		for (i = 0; i <= mp_maxid; i++)
3103 			bzero(zpcpu_get_cpu(pcpu_item, i), zone->uz_size);
3104 #else
3105 		bzero(item, zone->uz_size);
3106 #endif
3107 	}
3108 	return (pcpu_item);
3109 }
3110 
3111 /*
3112  * A stub while both regular and pcpu cases are identical.
3113  */
3114 void
3115 uma_zfree_pcpu_arg(uma_zone_t zone, void *pcpu_item, void *udata)
3116 {
3117 	void *item;
3118 
3119 #ifdef SMP
3120 	MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3121 #endif
3122 	item = zpcpu_offset_to_base(pcpu_item);
3123 	uma_zfree_arg(zone, item, udata);
3124 }
3125 
3126 static inline void *
3127 item_ctor(uma_zone_t zone, int uz_flags, int size, void *udata, int flags,
3128     void *item)
3129 {
3130 #ifdef INVARIANTS
3131 	bool skipdbg;
3132 
3133 	skipdbg = uma_dbg_zskip(zone, item);
3134 	if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3135 	    zone->uz_ctor != trash_ctor)
3136 		trash_ctor(item, size, udata, flags);
3137 #endif
3138 	/* Check flags before loading ctor pointer. */
3139 	if (__predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0) &&
3140 	    __predict_false(zone->uz_ctor != NULL) &&
3141 	    zone->uz_ctor(item, size, udata, flags) != 0) {
3142 		counter_u64_add(zone->uz_fails, 1);
3143 		zone_free_item(zone, item, udata, SKIP_DTOR | SKIP_CNT);
3144 		return (NULL);
3145 	}
3146 #ifdef INVARIANTS
3147 	if (!skipdbg)
3148 		uma_dbg_alloc(zone, NULL, item);
3149 #endif
3150 	if (__predict_false(flags & M_ZERO))
3151 		return (memset(item, 0, size));
3152 
3153 	return (item);
3154 }
3155 
3156 static inline void
3157 item_dtor(uma_zone_t zone, void *item, int size, void *udata,
3158     enum zfreeskip skip)
3159 {
3160 #ifdef INVARIANTS
3161 	bool skipdbg;
3162 
3163 	skipdbg = uma_dbg_zskip(zone, item);
3164 	if (skip == SKIP_NONE && !skipdbg) {
3165 		if ((zone->uz_flags & UMA_ZONE_MALLOC) != 0)
3166 			uma_dbg_free(zone, udata, item);
3167 		else
3168 			uma_dbg_free(zone, NULL, item);
3169 	}
3170 #endif
3171 	if (__predict_true(skip < SKIP_DTOR)) {
3172 		if (zone->uz_dtor != NULL)
3173 			zone->uz_dtor(item, size, udata);
3174 #ifdef INVARIANTS
3175 		if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3176 		    zone->uz_dtor != trash_dtor)
3177 			trash_dtor(item, size, udata);
3178 #endif
3179 	}
3180 }
3181 
3182 #if defined(INVARIANTS) || defined(DEBUG_MEMGUARD) || defined(WITNESS)
3183 #define	UMA_ZALLOC_DEBUG
3184 static int
3185 uma_zalloc_debug(uma_zone_t zone, void **itemp, void *udata, int flags)
3186 {
3187 	int error;
3188 
3189 	error = 0;
3190 #ifdef WITNESS
3191 	if (flags & M_WAITOK) {
3192 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
3193 		    "uma_zalloc_debug: zone \"%s\"", zone->uz_name);
3194 	}
3195 #endif
3196 
3197 #ifdef INVARIANTS
3198 	KASSERT((flags & M_EXEC) == 0,
3199 	    ("uma_zalloc_debug: called with M_EXEC"));
3200 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3201 	    ("uma_zalloc_debug: called within spinlock or critical section"));
3202 	KASSERT((zone->uz_flags & UMA_ZONE_PCPU) == 0 || (flags & M_ZERO) == 0,
3203 	    ("uma_zalloc_debug: allocating from a pcpu zone with M_ZERO"));
3204 #endif
3205 
3206 #ifdef DEBUG_MEMGUARD
3207 	if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && memguard_cmp_zone(zone)) {
3208 		void *item;
3209 		item = memguard_alloc(zone->uz_size, flags);
3210 		if (item != NULL) {
3211 			error = EJUSTRETURN;
3212 			if (zone->uz_init != NULL &&
3213 			    zone->uz_init(item, zone->uz_size, flags) != 0) {
3214 				*itemp = NULL;
3215 				return (error);
3216 			}
3217 			if (zone->uz_ctor != NULL &&
3218 			    zone->uz_ctor(item, zone->uz_size, udata,
3219 			    flags) != 0) {
3220 				counter_u64_add(zone->uz_fails, 1);
3221 			    	zone->uz_fini(item, zone->uz_size);
3222 				*itemp = NULL;
3223 				return (error);
3224 			}
3225 			*itemp = item;
3226 			return (error);
3227 		}
3228 		/* This is unfortunate but should not be fatal. */
3229 	}
3230 #endif
3231 	return (error);
3232 }
3233 
3234 static int
3235 uma_zfree_debug(uma_zone_t zone, void *item, void *udata)
3236 {
3237 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3238 	    ("uma_zfree_debug: called with spinlock or critical section held"));
3239 
3240 #ifdef DEBUG_MEMGUARD
3241 	if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && is_memguard_addr(item)) {
3242 		if (zone->uz_dtor != NULL)
3243 			zone->uz_dtor(item, zone->uz_size, udata);
3244 		if (zone->uz_fini != NULL)
3245 			zone->uz_fini(item, zone->uz_size);
3246 		memguard_free(item);
3247 		return (EJUSTRETURN);
3248 	}
3249 #endif
3250 	return (0);
3251 }
3252 #endif
3253 
3254 static inline void *
3255 cache_alloc_item(uma_zone_t zone, uma_cache_t cache, uma_cache_bucket_t bucket,
3256     void *udata, int flags)
3257 {
3258 	void *item;
3259 	int size, uz_flags;
3260 
3261 	item = cache_bucket_pop(cache, bucket);
3262 	size = cache_uz_size(cache);
3263 	uz_flags = cache_uz_flags(cache);
3264 	critical_exit();
3265 	return (item_ctor(zone, uz_flags, size, udata, flags, item));
3266 }
3267 
3268 static __noinline void *
3269 cache_alloc_retry(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3270 {
3271 	uma_cache_bucket_t bucket;
3272 	int domain;
3273 
3274 	while (cache_alloc(zone, cache, udata, flags)) {
3275 		cache = &zone->uz_cpu[curcpu];
3276 		bucket = &cache->uc_allocbucket;
3277 		if (__predict_false(bucket->ucb_cnt == 0))
3278 			continue;
3279 		return (cache_alloc_item(zone, cache, bucket, udata, flags));
3280 	}
3281 	critical_exit();
3282 
3283 	/*
3284 	 * We can not get a bucket so try to return a single item.
3285 	 */
3286 	if (zone->uz_flags & UMA_ZONE_FIRSTTOUCH)
3287 		domain = PCPU_GET(domain);
3288 	else
3289 		domain = UMA_ANYDOMAIN;
3290 	return (zone_alloc_item(zone, udata, domain, flags));
3291 }
3292 
3293 /* See uma.h */
3294 void *
3295 uma_zalloc_smr(uma_zone_t zone, int flags)
3296 {
3297 	uma_cache_bucket_t bucket;
3298 	uma_cache_t cache;
3299 
3300 #ifdef UMA_ZALLOC_DEBUG
3301 	void *item;
3302 
3303 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
3304 	    ("uma_zalloc_arg: called with non-SMR zone.\n"));
3305 	if (uma_zalloc_debug(zone, &item, NULL, flags) == EJUSTRETURN)
3306 		return (item);
3307 #endif
3308 
3309 	critical_enter();
3310 	cache = &zone->uz_cpu[curcpu];
3311 	bucket = &cache->uc_allocbucket;
3312 	if (__predict_false(bucket->ucb_cnt == 0))
3313 		return (cache_alloc_retry(zone, cache, NULL, flags));
3314 	return (cache_alloc_item(zone, cache, bucket, NULL, flags));
3315 }
3316 
3317 /* See uma.h */
3318 void *
3319 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
3320 {
3321 	uma_cache_bucket_t bucket;
3322 	uma_cache_t cache;
3323 
3324 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3325 	random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3326 
3327 	/* This is the fast path allocation */
3328 	CTR3(KTR_UMA, "uma_zalloc_arg zone %s(%p) flags %d", zone->uz_name,
3329 	    zone, flags);
3330 
3331 #ifdef UMA_ZALLOC_DEBUG
3332 	void *item;
3333 
3334 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3335 	    ("uma_zalloc_arg: called with SMR zone.\n"));
3336 	if (uma_zalloc_debug(zone, &item, udata, flags) == EJUSTRETURN)
3337 		return (item);
3338 #endif
3339 
3340 	/*
3341 	 * If possible, allocate from the per-CPU cache.  There are two
3342 	 * requirements for safe access to the per-CPU cache: (1) the thread
3343 	 * accessing the cache must not be preempted or yield during access,
3344 	 * and (2) the thread must not migrate CPUs without switching which
3345 	 * cache it accesses.  We rely on a critical section to prevent
3346 	 * preemption and migration.  We release the critical section in
3347 	 * order to acquire the zone mutex if we are unable to allocate from
3348 	 * the current cache; when we re-acquire the critical section, we
3349 	 * must detect and handle migration if it has occurred.
3350 	 */
3351 	critical_enter();
3352 	cache = &zone->uz_cpu[curcpu];
3353 	bucket = &cache->uc_allocbucket;
3354 	if (__predict_false(bucket->ucb_cnt == 0))
3355 		return (cache_alloc_retry(zone, cache, udata, flags));
3356 	return (cache_alloc_item(zone, cache, bucket, udata, flags));
3357 }
3358 
3359 /*
3360  * Replenish an alloc bucket and possibly restore an old one.  Called in
3361  * a critical section.  Returns in a critical section.
3362  *
3363  * A false return value indicates an allocation failure.
3364  * A true return value indicates success and the caller should retry.
3365  */
3366 static __noinline bool
3367 cache_alloc(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3368 {
3369 	uma_bucket_t bucket;
3370 	int domain;
3371 	bool new;
3372 
3373 	CRITICAL_ASSERT(curthread);
3374 
3375 	/*
3376 	 * If we have run out of items in our alloc bucket see
3377 	 * if we can switch with the free bucket.
3378 	 *
3379 	 * SMR Zones can't re-use the free bucket until the sequence has
3380 	 * expired.
3381 	 */
3382 	if ((cache_uz_flags(cache) & UMA_ZONE_SMR) == 0 &&
3383 	    cache->uc_freebucket.ucb_cnt != 0) {
3384 		cache_bucket_swap(&cache->uc_freebucket,
3385 		    &cache->uc_allocbucket);
3386 		return (true);
3387 	}
3388 
3389 	/*
3390 	 * Discard any empty allocation bucket while we hold no locks.
3391 	 */
3392 	bucket = cache_bucket_unload_alloc(cache);
3393 	critical_exit();
3394 
3395 	if (bucket != NULL) {
3396 		KASSERT(bucket->ub_cnt == 0,
3397 		    ("cache_alloc: Entered with non-empty alloc bucket."));
3398 		bucket_free(zone, bucket, udata);
3399 	}
3400 
3401 	/* Short-circuit for zones without buckets and low memory. */
3402 	if (zone->uz_bucket_size == 0 || bucketdisable) {
3403 		critical_enter();
3404 		return (false);
3405 	}
3406 
3407 	/*
3408 	 * Attempt to retrieve the item from the per-CPU cache has failed, so
3409 	 * we must go back to the zone.  This requires the zdom lock, so we
3410 	 * must drop the critical section, then re-acquire it when we go back
3411 	 * to the cache.  Since the critical section is released, we may be
3412 	 * preempted or migrate.  As such, make sure not to maintain any
3413 	 * thread-local state specific to the cache from prior to releasing
3414 	 * the critical section.
3415 	 */
3416 	domain = PCPU_GET(domain);
3417 	if ((cache_uz_flags(cache) & UMA_ZONE_ROUNDROBIN) != 0)
3418 		domain = zone_domain_highest(zone, domain);
3419 	bucket = cache_fetch_bucket(zone, cache, domain);
3420 	if (bucket == NULL) {
3421 		bucket = zone_alloc_bucket(zone, udata, domain, flags);
3422 		new = true;
3423 	} else
3424 		new = false;
3425 
3426 	CTR3(KTR_UMA, "uma_zalloc: zone %s(%p) bucket zone returned %p",
3427 	    zone->uz_name, zone, bucket);
3428 	if (bucket == NULL) {
3429 		critical_enter();
3430 		return (false);
3431 	}
3432 
3433 	/*
3434 	 * See if we lost the race or were migrated.  Cache the
3435 	 * initialized bucket to make this less likely or claim
3436 	 * the memory directly.
3437 	 */
3438 	critical_enter();
3439 	cache = &zone->uz_cpu[curcpu];
3440 	if (cache->uc_allocbucket.ucb_bucket == NULL &&
3441 	    ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) == 0 ||
3442 	    domain == PCPU_GET(domain))) {
3443 		if (new)
3444 			atomic_add_long(&ZDOM_GET(zone, domain)->uzd_imax,
3445 			    bucket->ub_cnt);
3446 		cache_bucket_load_alloc(cache, bucket);
3447 		return (true);
3448 	}
3449 
3450 	/*
3451 	 * We lost the race, release this bucket and start over.
3452 	 */
3453 	critical_exit();
3454 	zone_put_bucket(zone, domain, bucket, udata, false);
3455 	critical_enter();
3456 
3457 	return (true);
3458 }
3459 
3460 void *
3461 uma_zalloc_domain(uma_zone_t zone, void *udata, int domain, int flags)
3462 {
3463 
3464 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3465 	random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3466 
3467 	/* This is the fast path allocation */
3468 	CTR4(KTR_UMA, "uma_zalloc_domain zone %s(%p) domain %d flags %d",
3469 	    zone->uz_name, zone, domain, flags);
3470 
3471 	if (flags & M_WAITOK) {
3472 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
3473 		    "uma_zalloc_domain: zone \"%s\"", zone->uz_name);
3474 	}
3475 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3476 	    ("uma_zalloc_domain: called with spinlock or critical section held"));
3477 
3478 	return (zone_alloc_item(zone, udata, domain, flags));
3479 }
3480 
3481 /*
3482  * Find a slab with some space.  Prefer slabs that are partially used over those
3483  * that are totally full.  This helps to reduce fragmentation.
3484  *
3485  * If 'rr' is 1, search all domains starting from 'domain'.  Otherwise check
3486  * only 'domain'.
3487  */
3488 static uma_slab_t
3489 keg_first_slab(uma_keg_t keg, int domain, bool rr)
3490 {
3491 	uma_domain_t dom;
3492 	uma_slab_t slab;
3493 	int start;
3494 
3495 	KASSERT(domain >= 0 && domain < vm_ndomains,
3496 	    ("keg_first_slab: domain %d out of range", domain));
3497 	KEG_LOCK_ASSERT(keg, domain);
3498 
3499 	slab = NULL;
3500 	start = domain;
3501 	do {
3502 		dom = &keg->uk_domain[domain];
3503 		if ((slab = LIST_FIRST(&dom->ud_part_slab)) != NULL)
3504 			return (slab);
3505 		if ((slab = LIST_FIRST(&dom->ud_free_slab)) != NULL) {
3506 			LIST_REMOVE(slab, us_link);
3507 			dom->ud_free_slabs--;
3508 			LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
3509 			return (slab);
3510 		}
3511 		if (rr)
3512 			domain = (domain + 1) % vm_ndomains;
3513 	} while (domain != start);
3514 
3515 	return (NULL);
3516 }
3517 
3518 /*
3519  * Fetch an existing slab from a free or partial list.  Returns with the
3520  * keg domain lock held if a slab was found or unlocked if not.
3521  */
3522 static uma_slab_t
3523 keg_fetch_free_slab(uma_keg_t keg, int domain, bool rr, int flags)
3524 {
3525 	uma_slab_t slab;
3526 	uint32_t reserve;
3527 
3528 	/* HASH has a single free list. */
3529 	if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
3530 		domain = 0;
3531 
3532 	KEG_LOCK(keg, domain);
3533 	reserve = (flags & M_USE_RESERVE) != 0 ? 0 : keg->uk_reserve;
3534 	if (keg->uk_domain[domain].ud_free_items <= reserve ||
3535 	    (slab = keg_first_slab(keg, domain, rr)) == NULL) {
3536 		KEG_UNLOCK(keg, domain);
3537 		return (NULL);
3538 	}
3539 	return (slab);
3540 }
3541 
3542 static uma_slab_t
3543 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int rdomain, const int flags)
3544 {
3545 	struct vm_domainset_iter di;
3546 	uma_slab_t slab;
3547 	int aflags, domain;
3548 	bool rr;
3549 
3550 restart:
3551 	/*
3552 	 * Use the keg's policy if upper layers haven't already specified a
3553 	 * domain (as happens with first-touch zones).
3554 	 *
3555 	 * To avoid races we run the iterator with the keg lock held, but that
3556 	 * means that we cannot allow the vm_domainset layer to sleep.  Thus,
3557 	 * clear M_WAITOK and handle low memory conditions locally.
3558 	 */
3559 	rr = rdomain == UMA_ANYDOMAIN;
3560 	if (rr) {
3561 		aflags = (flags & ~M_WAITOK) | M_NOWAIT;
3562 		vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
3563 		    &aflags);
3564 	} else {
3565 		aflags = flags;
3566 		domain = rdomain;
3567 	}
3568 
3569 	for (;;) {
3570 		slab = keg_fetch_free_slab(keg, domain, rr, flags);
3571 		if (slab != NULL)
3572 			return (slab);
3573 
3574 		/*
3575 		 * M_NOVM means don't ask at all!
3576 		 */
3577 		if (flags & M_NOVM)
3578 			break;
3579 
3580 		slab = keg_alloc_slab(keg, zone, domain, flags, aflags);
3581 		if (slab != NULL)
3582 			return (slab);
3583 		if (!rr && (flags & M_WAITOK) == 0)
3584 			break;
3585 		if (rr && vm_domainset_iter_policy(&di, &domain) != 0) {
3586 			if ((flags & M_WAITOK) != 0) {
3587 				vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask);
3588 				goto restart;
3589 			}
3590 			break;
3591 		}
3592 	}
3593 
3594 	/*
3595 	 * We might not have been able to get a slab but another cpu
3596 	 * could have while we were unlocked.  Check again before we
3597 	 * fail.
3598 	 */
3599 	if ((slab = keg_fetch_free_slab(keg, domain, rr, flags)) != NULL)
3600 		return (slab);
3601 
3602 	return (NULL);
3603 }
3604 
3605 static void *
3606 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
3607 {
3608 	uma_domain_t dom;
3609 	void *item;
3610 	int freei;
3611 
3612 	KEG_LOCK_ASSERT(keg, slab->us_domain);
3613 
3614 	dom = &keg->uk_domain[slab->us_domain];
3615 	freei = BIT_FFS(keg->uk_ipers, &slab->us_free) - 1;
3616 	BIT_CLR(keg->uk_ipers, freei, &slab->us_free);
3617 	item = slab_item(slab, keg, freei);
3618 	slab->us_freecount--;
3619 	dom->ud_free_items--;
3620 
3621 	/*
3622 	 * Move this slab to the full list.  It must be on the partial list, so
3623 	 * we do not need to update the free slab count.  In particular,
3624 	 * keg_fetch_slab() always returns slabs on the partial list.
3625 	 */
3626 	if (slab->us_freecount == 0) {
3627 		LIST_REMOVE(slab, us_link);
3628 		LIST_INSERT_HEAD(&dom->ud_full_slab, slab, us_link);
3629 	}
3630 
3631 	return (item);
3632 }
3633 
3634 static int
3635 zone_import(void *arg, void **bucket, int max, int domain, int flags)
3636 {
3637 	uma_domain_t dom;
3638 	uma_zone_t zone;
3639 	uma_slab_t slab;
3640 	uma_keg_t keg;
3641 #ifdef NUMA
3642 	int stripe;
3643 #endif
3644 	int i;
3645 
3646 	zone = arg;
3647 	slab = NULL;
3648 	keg = zone->uz_keg;
3649 	/* Try to keep the buckets totally full */
3650 	for (i = 0; i < max; ) {
3651 		if ((slab = keg_fetch_slab(keg, zone, domain, flags)) == NULL)
3652 			break;
3653 #ifdef NUMA
3654 		stripe = howmany(max, vm_ndomains);
3655 #endif
3656 		dom = &keg->uk_domain[slab->us_domain];
3657 		while (slab->us_freecount && i < max) {
3658 			bucket[i++] = slab_alloc_item(keg, slab);
3659 			if (dom->ud_free_items <= keg->uk_reserve)
3660 				break;
3661 #ifdef NUMA
3662 			/*
3663 			 * If the zone is striped we pick a new slab for every
3664 			 * N allocations.  Eliminating this conditional will
3665 			 * instead pick a new domain for each bucket rather
3666 			 * than stripe within each bucket.  The current option
3667 			 * produces more fragmentation and requires more cpu
3668 			 * time but yields better distribution.
3669 			 */
3670 			if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0 &&
3671 			    vm_ndomains > 1 && --stripe == 0)
3672 				break;
3673 #endif
3674 		}
3675 		KEG_UNLOCK(keg, slab->us_domain);
3676 		/* Don't block if we allocated any successfully. */
3677 		flags &= ~M_WAITOK;
3678 		flags |= M_NOWAIT;
3679 	}
3680 
3681 	return i;
3682 }
3683 
3684 static int
3685 zone_alloc_limit_hard(uma_zone_t zone, int count, int flags)
3686 {
3687 	uint64_t old, new, total, max;
3688 
3689 	/*
3690 	 * The hard case.  We're going to sleep because there were existing
3691 	 * sleepers or because we ran out of items.  This routine enforces
3692 	 * fairness by keeping fifo order.
3693 	 *
3694 	 * First release our ill gotten gains and make some noise.
3695 	 */
3696 	for (;;) {
3697 		zone_free_limit(zone, count);
3698 		zone_log_warning(zone);
3699 		zone_maxaction(zone);
3700 		if (flags & M_NOWAIT)
3701 			return (0);
3702 
3703 		/*
3704 		 * We need to allocate an item or set ourself as a sleeper
3705 		 * while the sleepq lock is held to avoid wakeup races.  This
3706 		 * is essentially a home rolled semaphore.
3707 		 */
3708 		sleepq_lock(&zone->uz_max_items);
3709 		old = zone->uz_items;
3710 		do {
3711 			MPASS(UZ_ITEMS_SLEEPERS(old) < UZ_ITEMS_SLEEPERS_MAX);
3712 			/* Cache the max since we will evaluate twice. */
3713 			max = zone->uz_max_items;
3714 			if (UZ_ITEMS_SLEEPERS(old) != 0 ||
3715 			    UZ_ITEMS_COUNT(old) >= max)
3716 				new = old + UZ_ITEMS_SLEEPER;
3717 			else
3718 				new = old + MIN(count, max - old);
3719 		} while (atomic_fcmpset_64(&zone->uz_items, &old, new) == 0);
3720 
3721 		/* We may have successfully allocated under the sleepq lock. */
3722 		if (UZ_ITEMS_SLEEPERS(new) == 0) {
3723 			sleepq_release(&zone->uz_max_items);
3724 			return (new - old);
3725 		}
3726 
3727 		/*
3728 		 * This is in a different cacheline from uz_items so that we
3729 		 * don't constantly invalidate the fastpath cacheline when we
3730 		 * adjust item counts.  This could be limited to toggling on
3731 		 * transitions.
3732 		 */
3733 		atomic_add_32(&zone->uz_sleepers, 1);
3734 		atomic_add_64(&zone->uz_sleeps, 1);
3735 
3736 		/*
3737 		 * We have added ourselves as a sleeper.  The sleepq lock
3738 		 * protects us from wakeup races.  Sleep now and then retry.
3739 		 */
3740 		sleepq_add(&zone->uz_max_items, NULL, "zonelimit", 0, 0);
3741 		sleepq_wait(&zone->uz_max_items, PVM);
3742 
3743 		/*
3744 		 * After wakeup, remove ourselves as a sleeper and try
3745 		 * again.  We no longer have the sleepq lock for protection.
3746 		 *
3747 		 * Subract ourselves as a sleeper while attempting to add
3748 		 * our count.
3749 		 */
3750 		atomic_subtract_32(&zone->uz_sleepers, 1);
3751 		old = atomic_fetchadd_64(&zone->uz_items,
3752 		    -(UZ_ITEMS_SLEEPER - count));
3753 		/* We're no longer a sleeper. */
3754 		old -= UZ_ITEMS_SLEEPER;
3755 
3756 		/*
3757 		 * If we're still at the limit, restart.  Notably do not
3758 		 * block on other sleepers.  Cache the max value to protect
3759 		 * against changes via sysctl.
3760 		 */
3761 		total = UZ_ITEMS_COUNT(old);
3762 		max = zone->uz_max_items;
3763 		if (total >= max)
3764 			continue;
3765 		/* Truncate if necessary, otherwise wake other sleepers. */
3766 		if (total + count > max) {
3767 			zone_free_limit(zone, total + count - max);
3768 			count = max - total;
3769 		} else if (total + count < max && UZ_ITEMS_SLEEPERS(old) != 0)
3770 			wakeup_one(&zone->uz_max_items);
3771 
3772 		return (count);
3773 	}
3774 }
3775 
3776 /*
3777  * Allocate 'count' items from our max_items limit.  Returns the number
3778  * available.  If M_NOWAIT is not specified it will sleep until at least
3779  * one item can be allocated.
3780  */
3781 static int
3782 zone_alloc_limit(uma_zone_t zone, int count, int flags)
3783 {
3784 	uint64_t old;
3785 	uint64_t max;
3786 
3787 	max = zone->uz_max_items;
3788 	MPASS(max > 0);
3789 
3790 	/*
3791 	 * We expect normal allocations to succeed with a simple
3792 	 * fetchadd.
3793 	 */
3794 	old = atomic_fetchadd_64(&zone->uz_items, count);
3795 	if (__predict_true(old + count <= max))
3796 		return (count);
3797 
3798 	/*
3799 	 * If we had some items and no sleepers just return the
3800 	 * truncated value.  We have to release the excess space
3801 	 * though because that may wake sleepers who weren't woken
3802 	 * because we were temporarily over the limit.
3803 	 */
3804 	if (old < max) {
3805 		zone_free_limit(zone, (old + count) - max);
3806 		return (max - old);
3807 	}
3808 	return (zone_alloc_limit_hard(zone, count, flags));
3809 }
3810 
3811 /*
3812  * Free a number of items back to the limit.
3813  */
3814 static void
3815 zone_free_limit(uma_zone_t zone, int count)
3816 {
3817 	uint64_t old;
3818 
3819 	MPASS(count > 0);
3820 
3821 	/*
3822 	 * In the common case we either have no sleepers or
3823 	 * are still over the limit and can just return.
3824 	 */
3825 	old = atomic_fetchadd_64(&zone->uz_items, -count);
3826 	if (__predict_true(UZ_ITEMS_SLEEPERS(old) == 0 ||
3827 	   UZ_ITEMS_COUNT(old) - count >= zone->uz_max_items))
3828 		return;
3829 
3830 	/*
3831 	 * Moderate the rate of wakeups.  Sleepers will continue
3832 	 * to generate wakeups if necessary.
3833 	 */
3834 	wakeup_one(&zone->uz_max_items);
3835 }
3836 
3837 static uma_bucket_t
3838 zone_alloc_bucket(uma_zone_t zone, void *udata, int domain, int flags)
3839 {
3840 	uma_bucket_t bucket;
3841 	int maxbucket, cnt;
3842 
3843 	CTR3(KTR_UMA, "zone_alloc_bucket zone %s(%p) domain %d", zone->uz_name,
3844 	    zone, domain);
3845 
3846 	/* Avoid allocs targeting empty domains. */
3847 	if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
3848 		domain = UMA_ANYDOMAIN;
3849 	if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
3850 		domain = UMA_ANYDOMAIN;
3851 
3852 	if (zone->uz_max_items > 0)
3853 		maxbucket = zone_alloc_limit(zone, zone->uz_bucket_size,
3854 		    M_NOWAIT);
3855 	else
3856 		maxbucket = zone->uz_bucket_size;
3857 	if (maxbucket == 0)
3858 		return (false);
3859 
3860 	/* Don't wait for buckets, preserve caller's NOVM setting. */
3861 	bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
3862 	if (bucket == NULL) {
3863 		cnt = 0;
3864 		goto out;
3865 	}
3866 
3867 	bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
3868 	    MIN(maxbucket, bucket->ub_entries), domain, flags);
3869 
3870 	/*
3871 	 * Initialize the memory if necessary.
3872 	 */
3873 	if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
3874 		int i;
3875 
3876 		for (i = 0; i < bucket->ub_cnt; i++)
3877 			if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size,
3878 			    flags) != 0)
3879 				break;
3880 		/*
3881 		 * If we couldn't initialize the whole bucket, put the
3882 		 * rest back onto the freelist.
3883 		 */
3884 		if (i != bucket->ub_cnt) {
3885 			zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
3886 			    bucket->ub_cnt - i);
3887 #ifdef INVARIANTS
3888 			bzero(&bucket->ub_bucket[i],
3889 			    sizeof(void *) * (bucket->ub_cnt - i));
3890 #endif
3891 			bucket->ub_cnt = i;
3892 		}
3893 	}
3894 
3895 	cnt = bucket->ub_cnt;
3896 	if (bucket->ub_cnt == 0) {
3897 		bucket_free(zone, bucket, udata);
3898 		counter_u64_add(zone->uz_fails, 1);
3899 		bucket = NULL;
3900 	}
3901 out:
3902 	if (zone->uz_max_items > 0 && cnt < maxbucket)
3903 		zone_free_limit(zone, maxbucket - cnt);
3904 
3905 	return (bucket);
3906 }
3907 
3908 /*
3909  * Allocates a single item from a zone.
3910  *
3911  * Arguments
3912  *	zone   The zone to alloc for.
3913  *	udata  The data to be passed to the constructor.
3914  *	domain The domain to allocate from or UMA_ANYDOMAIN.
3915  *	flags  M_WAITOK, M_NOWAIT, M_ZERO.
3916  *
3917  * Returns
3918  *	NULL if there is no memory and M_NOWAIT is set
3919  *	An item if successful
3920  */
3921 
3922 static void *
3923 zone_alloc_item(uma_zone_t zone, void *udata, int domain, int flags)
3924 {
3925 	void *item;
3926 
3927 	if (zone->uz_max_items > 0 && zone_alloc_limit(zone, 1, flags) == 0)
3928 		return (NULL);
3929 
3930 	/* Avoid allocs targeting empty domains. */
3931 	if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
3932 		domain = UMA_ANYDOMAIN;
3933 
3934 	if (zone->uz_import(zone->uz_arg, &item, 1, domain, flags) != 1)
3935 		goto fail_cnt;
3936 
3937 	/*
3938 	 * We have to call both the zone's init (not the keg's init)
3939 	 * and the zone's ctor.  This is because the item is going from
3940 	 * a keg slab directly to the user, and the user is expecting it
3941 	 * to be both zone-init'd as well as zone-ctor'd.
3942 	 */
3943 	if (zone->uz_init != NULL) {
3944 		if (zone->uz_init(item, zone->uz_size, flags) != 0) {
3945 			zone_free_item(zone, item, udata, SKIP_FINI | SKIP_CNT);
3946 			goto fail_cnt;
3947 		}
3948 	}
3949 	item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata, flags,
3950 	    item);
3951 	if (item == NULL)
3952 		goto fail;
3953 
3954 	counter_u64_add(zone->uz_allocs, 1);
3955 	CTR3(KTR_UMA, "zone_alloc_item item %p from %s(%p)", item,
3956 	    zone->uz_name, zone);
3957 
3958 	return (item);
3959 
3960 fail_cnt:
3961 	counter_u64_add(zone->uz_fails, 1);
3962 fail:
3963 	if (zone->uz_max_items > 0)
3964 		zone_free_limit(zone, 1);
3965 	CTR2(KTR_UMA, "zone_alloc_item failed from %s(%p)",
3966 	    zone->uz_name, zone);
3967 
3968 	return (NULL);
3969 }
3970 
3971 /* See uma.h */
3972 void
3973 uma_zfree_smr(uma_zone_t zone, void *item)
3974 {
3975 	uma_cache_t cache;
3976 	uma_cache_bucket_t bucket;
3977 	int itemdomain, uz_flags;
3978 
3979 #ifdef UMA_ZALLOC_DEBUG
3980 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
3981 	    ("uma_zfree_smr: called with non-SMR zone.\n"));
3982 	KASSERT(item != NULL, ("uma_zfree_smr: Called with NULL pointer."));
3983 	SMR_ASSERT_NOT_ENTERED(zone->uz_smr);
3984 	if (uma_zfree_debug(zone, item, NULL) == EJUSTRETURN)
3985 		return;
3986 #endif
3987 	cache = &zone->uz_cpu[curcpu];
3988 	uz_flags = cache_uz_flags(cache);
3989 	itemdomain = 0;
3990 #ifdef NUMA
3991 	if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
3992 		itemdomain = _vm_phys_domain(pmap_kextract((vm_offset_t)item));
3993 #endif
3994 	critical_enter();
3995 	do {
3996 		cache = &zone->uz_cpu[curcpu];
3997 		/* SMR Zones must free to the free bucket. */
3998 		bucket = &cache->uc_freebucket;
3999 #ifdef NUMA
4000 		if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4001 		    PCPU_GET(domain) != itemdomain) {
4002 			bucket = &cache->uc_crossbucket;
4003 		}
4004 #endif
4005 		if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4006 			cache_bucket_push(cache, bucket, item);
4007 			critical_exit();
4008 			return;
4009 		}
4010 	} while (cache_free(zone, cache, NULL, item, itemdomain));
4011 	critical_exit();
4012 
4013 	/*
4014 	 * If nothing else caught this, we'll just do an internal free.
4015 	 */
4016 	zone_free_item(zone, item, NULL, SKIP_NONE);
4017 }
4018 
4019 /* See uma.h */
4020 void
4021 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
4022 {
4023 	uma_cache_t cache;
4024 	uma_cache_bucket_t bucket;
4025 	int itemdomain, uz_flags;
4026 
4027 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
4028 	random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
4029 
4030 	CTR2(KTR_UMA, "uma_zfree_arg zone %s(%p)", zone->uz_name, zone);
4031 
4032 #ifdef UMA_ZALLOC_DEBUG
4033 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
4034 	    ("uma_zfree_arg: called with SMR zone.\n"));
4035 	if (uma_zfree_debug(zone, item, udata) == EJUSTRETURN)
4036 		return;
4037 #endif
4038         /* uma_zfree(..., NULL) does nothing, to match free(9). */
4039         if (item == NULL)
4040                 return;
4041 
4042 	/*
4043 	 * We are accessing the per-cpu cache without a critical section to
4044 	 * fetch size and flags.  This is acceptable, if we are preempted we
4045 	 * will simply read another cpu's line.
4046 	 */
4047 	cache = &zone->uz_cpu[curcpu];
4048 	uz_flags = cache_uz_flags(cache);
4049 	if (UMA_ALWAYS_CTORDTOR ||
4050 	    __predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0))
4051 		item_dtor(zone, item, cache_uz_size(cache), udata, SKIP_NONE);
4052 
4053 	/*
4054 	 * The race here is acceptable.  If we miss it we'll just have to wait
4055 	 * a little longer for the limits to be reset.
4056 	 */
4057 	if (__predict_false(uz_flags & UMA_ZFLAG_LIMIT)) {
4058 		if (zone->uz_sleepers > 0)
4059 			goto zfree_item;
4060 	}
4061 
4062 	/*
4063 	 * If possible, free to the per-CPU cache.  There are two
4064 	 * requirements for safe access to the per-CPU cache: (1) the thread
4065 	 * accessing the cache must not be preempted or yield during access,
4066 	 * and (2) the thread must not migrate CPUs without switching which
4067 	 * cache it accesses.  We rely on a critical section to prevent
4068 	 * preemption and migration.  We release the critical section in
4069 	 * order to acquire the zone mutex if we are unable to free to the
4070 	 * current cache; when we re-acquire the critical section, we must
4071 	 * detect and handle migration if it has occurred.
4072 	 */
4073 	itemdomain = 0;
4074 #ifdef NUMA
4075 	if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4076 		itemdomain = _vm_phys_domain(pmap_kextract((vm_offset_t)item));
4077 #endif
4078 	critical_enter();
4079 	do {
4080 		cache = &zone->uz_cpu[curcpu];
4081 		/*
4082 		 * Try to free into the allocbucket first to give LIFO
4083 		 * ordering for cache-hot datastructures.  Spill over
4084 		 * into the freebucket if necessary.  Alloc will swap
4085 		 * them if one runs dry.
4086 		 */
4087 		bucket = &cache->uc_allocbucket;
4088 #ifdef NUMA
4089 		if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4090 		    PCPU_GET(domain) != itemdomain) {
4091 			bucket = &cache->uc_crossbucket;
4092 		} else
4093 #endif
4094 		if (bucket->ucb_cnt == bucket->ucb_entries &&
4095 		   cache->uc_freebucket.ucb_cnt <
4096 		   cache->uc_freebucket.ucb_entries)
4097 			cache_bucket_swap(&cache->uc_freebucket,
4098 			    &cache->uc_allocbucket);
4099 		if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4100 			cache_bucket_push(cache, bucket, item);
4101 			critical_exit();
4102 			return;
4103 		}
4104 	} while (cache_free(zone, cache, udata, item, itemdomain));
4105 	critical_exit();
4106 
4107 	/*
4108 	 * If nothing else caught this, we'll just do an internal free.
4109 	 */
4110 zfree_item:
4111 	zone_free_item(zone, item, udata, SKIP_DTOR);
4112 }
4113 
4114 #ifdef NUMA
4115 /*
4116  * sort crossdomain free buckets to domain correct buckets and cache
4117  * them.
4118  */
4119 static void
4120 zone_free_cross(uma_zone_t zone, uma_bucket_t bucket, void *udata)
4121 {
4122 	struct uma_bucketlist fullbuckets;
4123 	uma_zone_domain_t zdom;
4124 	uma_bucket_t b;
4125 	smr_seq_t seq;
4126 	void *item;
4127 	int domain;
4128 
4129 	CTR3(KTR_UMA,
4130 	    "uma_zfree: zone %s(%p) draining cross bucket %p",
4131 	    zone->uz_name, zone, bucket);
4132 
4133 	/*
4134 	 * It is possible for buckets to arrive here out of order so we fetch
4135 	 * the current smr seq rather than accepting the bucket's.
4136 	 */
4137 	seq = SMR_SEQ_INVALID;
4138 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
4139 		seq = smr_advance(zone->uz_smr);
4140 
4141 	/*
4142 	 * To avoid having ndomain * ndomain buckets for sorting we have a
4143 	 * lock on the current crossfree bucket.  A full matrix with
4144 	 * per-domain locking could be used if necessary.
4145 	 */
4146 	STAILQ_INIT(&fullbuckets);
4147 	ZONE_CROSS_LOCK(zone);
4148 	while (bucket->ub_cnt > 0) {
4149 		item = bucket->ub_bucket[bucket->ub_cnt - 1];
4150 		domain = _vm_phys_domain(pmap_kextract((vm_offset_t)item));
4151 		zdom = ZDOM_GET(zone, domain);
4152 		if (zdom->uzd_cross == NULL) {
4153 			zdom->uzd_cross = bucket_alloc(zone, udata, M_NOWAIT);
4154 			if (zdom->uzd_cross == NULL)
4155 				break;
4156 		}
4157 		b = zdom->uzd_cross;
4158 		b->ub_bucket[b->ub_cnt++] = item;
4159 		b->ub_seq = seq;
4160 		if (b->ub_cnt == b->ub_entries) {
4161 			STAILQ_INSERT_HEAD(&fullbuckets, b, ub_link);
4162 			zdom->uzd_cross = NULL;
4163 		}
4164 		bucket->ub_cnt--;
4165 	}
4166 	ZONE_CROSS_UNLOCK(zone);
4167 	if (bucket->ub_cnt == 0)
4168 		bucket->ub_seq = SMR_SEQ_INVALID;
4169 	bucket_free(zone, bucket, udata);
4170 
4171 	while ((b = STAILQ_FIRST(&fullbuckets)) != NULL) {
4172 		STAILQ_REMOVE_HEAD(&fullbuckets, ub_link);
4173 		domain = _vm_phys_domain(pmap_kextract(
4174 		    (vm_offset_t)b->ub_bucket[0]));
4175 		zone_put_bucket(zone, domain, b, udata, true);
4176 	}
4177 }
4178 #endif
4179 
4180 static void
4181 zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata,
4182     int itemdomain, bool ws)
4183 {
4184 
4185 #ifdef NUMA
4186 	/*
4187 	 * Buckets coming from the wrong domain will be entirely for the
4188 	 * only other domain on two domain systems.  In this case we can
4189 	 * simply cache them.  Otherwise we need to sort them back to
4190 	 * correct domains.
4191 	 */
4192 	if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4193 	    vm_ndomains > 2 && PCPU_GET(domain) != itemdomain) {
4194 		zone_free_cross(zone, bucket, udata);
4195 		return;
4196 	}
4197 #endif
4198 
4199 	/*
4200 	 * Attempt to save the bucket in the zone's domain bucket cache.
4201 	 */
4202 	CTR3(KTR_UMA,
4203 	    "uma_zfree: zone %s(%p) putting bucket %p on free list",
4204 	    zone->uz_name, zone, bucket);
4205 	/* ub_cnt is pointing to the last free item */
4206 	if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
4207 		itemdomain = zone_domain_lowest(zone, itemdomain);
4208 	zone_put_bucket(zone, itemdomain, bucket, udata, ws);
4209 }
4210 
4211 /*
4212  * Populate a free or cross bucket for the current cpu cache.  Free any
4213  * existing full bucket either to the zone cache or back to the slab layer.
4214  *
4215  * Enters and returns in a critical section.  false return indicates that
4216  * we can not satisfy this free in the cache layer.  true indicates that
4217  * the caller should retry.
4218  */
4219 static __noinline bool
4220 cache_free(uma_zone_t zone, uma_cache_t cache, void *udata, void *item,
4221     int itemdomain)
4222 {
4223 	uma_cache_bucket_t cbucket;
4224 	uma_bucket_t newbucket, bucket;
4225 
4226 	CRITICAL_ASSERT(curthread);
4227 
4228 	if (zone->uz_bucket_size == 0)
4229 		return false;
4230 
4231 	cache = &zone->uz_cpu[curcpu];
4232 	newbucket = NULL;
4233 
4234 	/*
4235 	 * FIRSTTOUCH domains need to free to the correct zdom.  When
4236 	 * enabled this is the zdom of the item.   The bucket is the
4237 	 * cross bucket if the current domain and itemdomain do not match.
4238 	 */
4239 	cbucket = &cache->uc_freebucket;
4240 #ifdef NUMA
4241 	if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4242 		if (PCPU_GET(domain) != itemdomain) {
4243 			cbucket = &cache->uc_crossbucket;
4244 			if (cbucket->ucb_cnt != 0)
4245 				counter_u64_add(zone->uz_xdomain,
4246 				    cbucket->ucb_cnt);
4247 		}
4248 	}
4249 #endif
4250 	bucket = cache_bucket_unload(cbucket);
4251 	KASSERT(bucket == NULL || bucket->ub_cnt == bucket->ub_entries,
4252 	    ("cache_free: Entered with non-full free bucket."));
4253 
4254 	/* We are no longer associated with this CPU. */
4255 	critical_exit();
4256 
4257 	/*
4258 	 * Don't let SMR zones operate without a free bucket.  Force
4259 	 * a synchronize and re-use this one.  We will only degrade
4260 	 * to a synchronize every bucket_size items rather than every
4261 	 * item if we fail to allocate a bucket.
4262 	 */
4263 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0) {
4264 		if (bucket != NULL)
4265 			bucket->ub_seq = smr_advance(zone->uz_smr);
4266 		newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4267 		if (newbucket == NULL && bucket != NULL) {
4268 			bucket_drain(zone, bucket);
4269 			newbucket = bucket;
4270 			bucket = NULL;
4271 		}
4272 	} else if (!bucketdisable)
4273 		newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4274 
4275 	if (bucket != NULL)
4276 		zone_free_bucket(zone, bucket, udata, itemdomain, true);
4277 
4278 	critical_enter();
4279 	if ((bucket = newbucket) == NULL)
4280 		return (false);
4281 	cache = &zone->uz_cpu[curcpu];
4282 #ifdef NUMA
4283 	/*
4284 	 * Check to see if we should be populating the cross bucket.  If it
4285 	 * is already populated we will fall through and attempt to populate
4286 	 * the free bucket.
4287 	 */
4288 	if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4289 		if (PCPU_GET(domain) != itemdomain &&
4290 		    cache->uc_crossbucket.ucb_bucket == NULL) {
4291 			cache_bucket_load_cross(cache, bucket);
4292 			return (true);
4293 		}
4294 	}
4295 #endif
4296 	/*
4297 	 * We may have lost the race to fill the bucket or switched CPUs.
4298 	 */
4299 	if (cache->uc_freebucket.ucb_bucket != NULL) {
4300 		critical_exit();
4301 		bucket_free(zone, bucket, udata);
4302 		critical_enter();
4303 	} else
4304 		cache_bucket_load_free(cache, bucket);
4305 
4306 	return (true);
4307 }
4308 
4309 void
4310 uma_zfree_domain(uma_zone_t zone, void *item, void *udata)
4311 {
4312 
4313 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
4314 	random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
4315 
4316 	CTR2(KTR_UMA, "uma_zfree_domain zone %s(%p)", zone->uz_name, zone);
4317 
4318 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
4319 	    ("uma_zfree_domain: called with spinlock or critical section held"));
4320 
4321         /* uma_zfree(..., NULL) does nothing, to match free(9). */
4322         if (item == NULL)
4323                 return;
4324 	zone_free_item(zone, item, udata, SKIP_NONE);
4325 }
4326 
4327 static void
4328 slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item)
4329 {
4330 	uma_keg_t keg;
4331 	uma_domain_t dom;
4332 	int freei;
4333 
4334 	keg = zone->uz_keg;
4335 	KEG_LOCK_ASSERT(keg, slab->us_domain);
4336 
4337 	/* Do we need to remove from any lists? */
4338 	dom = &keg->uk_domain[slab->us_domain];
4339 	if (slab->us_freecount + 1 == keg->uk_ipers) {
4340 		LIST_REMOVE(slab, us_link);
4341 		LIST_INSERT_HEAD(&dom->ud_free_slab, slab, us_link);
4342 		dom->ud_free_slabs++;
4343 	} else if (slab->us_freecount == 0) {
4344 		LIST_REMOVE(slab, us_link);
4345 		LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
4346 	}
4347 
4348 	/* Slab management. */
4349 	freei = slab_item_index(slab, keg, item);
4350 	BIT_SET(keg->uk_ipers, freei, &slab->us_free);
4351 	slab->us_freecount++;
4352 
4353 	/* Keg statistics. */
4354 	dom->ud_free_items++;
4355 }
4356 
4357 static void
4358 zone_release(void *arg, void **bucket, int cnt)
4359 {
4360 	struct mtx *lock;
4361 	uma_zone_t zone;
4362 	uma_slab_t slab;
4363 	uma_keg_t keg;
4364 	uint8_t *mem;
4365 	void *item;
4366 	int i;
4367 
4368 	zone = arg;
4369 	keg = zone->uz_keg;
4370 	lock = NULL;
4371 	if (__predict_false((zone->uz_flags & UMA_ZFLAG_HASH) != 0))
4372 		lock = KEG_LOCK(keg, 0);
4373 	for (i = 0; i < cnt; i++) {
4374 		item = bucket[i];
4375 		if (__predict_true((zone->uz_flags & UMA_ZFLAG_VTOSLAB) != 0)) {
4376 			slab = vtoslab((vm_offset_t)item);
4377 		} else {
4378 			mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
4379 			if ((zone->uz_flags & UMA_ZFLAG_HASH) != 0)
4380 				slab = hash_sfind(&keg->uk_hash, mem);
4381 			else
4382 				slab = (uma_slab_t)(mem + keg->uk_pgoff);
4383 		}
4384 		if (lock != KEG_LOCKPTR(keg, slab->us_domain)) {
4385 			if (lock != NULL)
4386 				mtx_unlock(lock);
4387 			lock = KEG_LOCK(keg, slab->us_domain);
4388 		}
4389 		slab_free_item(zone, slab, item);
4390 	}
4391 	if (lock != NULL)
4392 		mtx_unlock(lock);
4393 }
4394 
4395 /*
4396  * Frees a single item to any zone.
4397  *
4398  * Arguments:
4399  *	zone   The zone to free to
4400  *	item   The item we're freeing
4401  *	udata  User supplied data for the dtor
4402  *	skip   Skip dtors and finis
4403  */
4404 static __noinline void
4405 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
4406 {
4407 
4408 	/*
4409 	 * If a free is sent directly to an SMR zone we have to
4410 	 * synchronize immediately because the item can instantly
4411 	 * be reallocated. This should only happen in degenerate
4412 	 * cases when no memory is available for per-cpu caches.
4413 	 */
4414 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0 && skip == SKIP_NONE)
4415 		smr_synchronize(zone->uz_smr);
4416 
4417 	item_dtor(zone, item, zone->uz_size, udata, skip);
4418 
4419 	if (skip < SKIP_FINI && zone->uz_fini)
4420 		zone->uz_fini(item, zone->uz_size);
4421 
4422 	zone->uz_release(zone->uz_arg, &item, 1);
4423 
4424 	if (skip & SKIP_CNT)
4425 		return;
4426 
4427 	counter_u64_add(zone->uz_frees, 1);
4428 
4429 	if (zone->uz_max_items > 0)
4430 		zone_free_limit(zone, 1);
4431 }
4432 
4433 /* See uma.h */
4434 int
4435 uma_zone_set_max(uma_zone_t zone, int nitems)
4436 {
4437 	struct uma_bucket_zone *ubz;
4438 	int count;
4439 
4440 	/*
4441 	 * XXX This can misbehave if the zone has any allocations with
4442 	 * no limit and a limit is imposed.  There is currently no
4443 	 * way to clear a limit.
4444 	 */
4445 	ZONE_LOCK(zone);
4446 	ubz = bucket_zone_max(zone, nitems);
4447 	count = ubz != NULL ? ubz->ubz_entries : 0;
4448 	zone->uz_bucket_size_max = zone->uz_bucket_size = count;
4449 	if (zone->uz_bucket_size_min > zone->uz_bucket_size_max)
4450 		zone->uz_bucket_size_min = zone->uz_bucket_size_max;
4451 	zone->uz_max_items = nitems;
4452 	zone->uz_flags |= UMA_ZFLAG_LIMIT;
4453 	zone_update_caches(zone);
4454 	/* We may need to wake waiters. */
4455 	wakeup(&zone->uz_max_items);
4456 	ZONE_UNLOCK(zone);
4457 
4458 	return (nitems);
4459 }
4460 
4461 /* See uma.h */
4462 void
4463 uma_zone_set_maxcache(uma_zone_t zone, int nitems)
4464 {
4465 	struct uma_bucket_zone *ubz;
4466 	int bpcpu;
4467 
4468 	ZONE_LOCK(zone);
4469 	ubz = bucket_zone_max(zone, nitems);
4470 	if (ubz != NULL) {
4471 		bpcpu = 2;
4472 		if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4473 			/* Count the cross-domain bucket. */
4474 			bpcpu++;
4475 		nitems -= ubz->ubz_entries * bpcpu * mp_ncpus;
4476 		zone->uz_bucket_size_max = ubz->ubz_entries;
4477 	} else {
4478 		zone->uz_bucket_size_max = zone->uz_bucket_size = 0;
4479 	}
4480 	if (zone->uz_bucket_size_min > zone->uz_bucket_size_max)
4481 		zone->uz_bucket_size_min = zone->uz_bucket_size_max;
4482 	zone->uz_bucket_max = nitems / vm_ndomains;
4483 	ZONE_UNLOCK(zone);
4484 }
4485 
4486 /* See uma.h */
4487 int
4488 uma_zone_get_max(uma_zone_t zone)
4489 {
4490 	int nitems;
4491 
4492 	nitems = atomic_load_64(&zone->uz_max_items);
4493 
4494 	return (nitems);
4495 }
4496 
4497 /* See uma.h */
4498 void
4499 uma_zone_set_warning(uma_zone_t zone, const char *warning)
4500 {
4501 
4502 	ZONE_ASSERT_COLD(zone);
4503 	zone->uz_warning = warning;
4504 }
4505 
4506 /* See uma.h */
4507 void
4508 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction)
4509 {
4510 
4511 	ZONE_ASSERT_COLD(zone);
4512 	TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone);
4513 }
4514 
4515 /* See uma.h */
4516 int
4517 uma_zone_get_cur(uma_zone_t zone)
4518 {
4519 	int64_t nitems;
4520 	u_int i;
4521 
4522 	nitems = 0;
4523 	if (zone->uz_allocs != EARLY_COUNTER && zone->uz_frees != EARLY_COUNTER)
4524 		nitems = counter_u64_fetch(zone->uz_allocs) -
4525 		    counter_u64_fetch(zone->uz_frees);
4526 	CPU_FOREACH(i)
4527 		nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs) -
4528 		    atomic_load_64(&zone->uz_cpu[i].uc_frees);
4529 
4530 	return (nitems < 0 ? 0 : nitems);
4531 }
4532 
4533 static uint64_t
4534 uma_zone_get_allocs(uma_zone_t zone)
4535 {
4536 	uint64_t nitems;
4537 	u_int i;
4538 
4539 	nitems = 0;
4540 	if (zone->uz_allocs != EARLY_COUNTER)
4541 		nitems = counter_u64_fetch(zone->uz_allocs);
4542 	CPU_FOREACH(i)
4543 		nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs);
4544 
4545 	return (nitems);
4546 }
4547 
4548 static uint64_t
4549 uma_zone_get_frees(uma_zone_t zone)
4550 {
4551 	uint64_t nitems;
4552 	u_int i;
4553 
4554 	nitems = 0;
4555 	if (zone->uz_frees != EARLY_COUNTER)
4556 		nitems = counter_u64_fetch(zone->uz_frees);
4557 	CPU_FOREACH(i)
4558 		nitems += atomic_load_64(&zone->uz_cpu[i].uc_frees);
4559 
4560 	return (nitems);
4561 }
4562 
4563 #ifdef INVARIANTS
4564 /* Used only for KEG_ASSERT_COLD(). */
4565 static uint64_t
4566 uma_keg_get_allocs(uma_keg_t keg)
4567 {
4568 	uma_zone_t z;
4569 	uint64_t nitems;
4570 
4571 	nitems = 0;
4572 	LIST_FOREACH(z, &keg->uk_zones, uz_link)
4573 		nitems += uma_zone_get_allocs(z);
4574 
4575 	return (nitems);
4576 }
4577 #endif
4578 
4579 /* See uma.h */
4580 void
4581 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
4582 {
4583 	uma_keg_t keg;
4584 
4585 	KEG_GET(zone, keg);
4586 	KEG_ASSERT_COLD(keg);
4587 	keg->uk_init = uminit;
4588 }
4589 
4590 /* See uma.h */
4591 void
4592 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
4593 {
4594 	uma_keg_t keg;
4595 
4596 	KEG_GET(zone, keg);
4597 	KEG_ASSERT_COLD(keg);
4598 	keg->uk_fini = fini;
4599 }
4600 
4601 /* See uma.h */
4602 void
4603 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
4604 {
4605 
4606 	ZONE_ASSERT_COLD(zone);
4607 	zone->uz_init = zinit;
4608 }
4609 
4610 /* See uma.h */
4611 void
4612 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
4613 {
4614 
4615 	ZONE_ASSERT_COLD(zone);
4616 	zone->uz_fini = zfini;
4617 }
4618 
4619 /* See uma.h */
4620 void
4621 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
4622 {
4623 	uma_keg_t keg;
4624 
4625 	KEG_GET(zone, keg);
4626 	KEG_ASSERT_COLD(keg);
4627 	keg->uk_freef = freef;
4628 }
4629 
4630 /* See uma.h */
4631 void
4632 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
4633 {
4634 	uma_keg_t keg;
4635 
4636 	KEG_GET(zone, keg);
4637 	KEG_ASSERT_COLD(keg);
4638 	keg->uk_allocf = allocf;
4639 }
4640 
4641 /* See uma.h */
4642 void
4643 uma_zone_set_smr(uma_zone_t zone, smr_t smr)
4644 {
4645 
4646 	ZONE_ASSERT_COLD(zone);
4647 
4648 	KASSERT(smr != NULL, ("Got NULL smr"));
4649 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
4650 	    ("zone %p (%s) already uses SMR", zone, zone->uz_name));
4651 	zone->uz_flags |= UMA_ZONE_SMR;
4652 	zone->uz_smr = smr;
4653 	zone_update_caches(zone);
4654 }
4655 
4656 smr_t
4657 uma_zone_get_smr(uma_zone_t zone)
4658 {
4659 
4660 	return (zone->uz_smr);
4661 }
4662 
4663 /* See uma.h */
4664 void
4665 uma_zone_reserve(uma_zone_t zone, int items)
4666 {
4667 	uma_keg_t keg;
4668 
4669 	KEG_GET(zone, keg);
4670 	KEG_ASSERT_COLD(keg);
4671 	keg->uk_reserve = items;
4672 }
4673 
4674 /* See uma.h */
4675 int
4676 uma_zone_reserve_kva(uma_zone_t zone, int count)
4677 {
4678 	uma_keg_t keg;
4679 	vm_offset_t kva;
4680 	u_int pages;
4681 
4682 	KEG_GET(zone, keg);
4683 	KEG_ASSERT_COLD(keg);
4684 	ZONE_ASSERT_COLD(zone);
4685 
4686 	pages = howmany(count, keg->uk_ipers) * keg->uk_ppera;
4687 
4688 #ifdef UMA_MD_SMALL_ALLOC
4689 	if (keg->uk_ppera > 1) {
4690 #else
4691 	if (1) {
4692 #endif
4693 		kva = kva_alloc((vm_size_t)pages * PAGE_SIZE);
4694 		if (kva == 0)
4695 			return (0);
4696 	} else
4697 		kva = 0;
4698 
4699 	MPASS(keg->uk_kva == 0);
4700 	keg->uk_kva = kva;
4701 	keg->uk_offset = 0;
4702 	zone->uz_max_items = pages * keg->uk_ipers;
4703 #ifdef UMA_MD_SMALL_ALLOC
4704 	keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
4705 #else
4706 	keg->uk_allocf = noobj_alloc;
4707 #endif
4708 	keg->uk_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
4709 	zone->uz_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
4710 	zone_update_caches(zone);
4711 
4712 	return (1);
4713 }
4714 
4715 /* See uma.h */
4716 void
4717 uma_prealloc(uma_zone_t zone, int items)
4718 {
4719 	struct vm_domainset_iter di;
4720 	uma_domain_t dom;
4721 	uma_slab_t slab;
4722 	uma_keg_t keg;
4723 	int aflags, domain, slabs;
4724 
4725 	KEG_GET(zone, keg);
4726 	slabs = howmany(items, keg->uk_ipers);
4727 	while (slabs-- > 0) {
4728 		aflags = M_NOWAIT;
4729 		vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
4730 		    &aflags);
4731 		for (;;) {
4732 			slab = keg_alloc_slab(keg, zone, domain, M_WAITOK,
4733 			    aflags);
4734 			if (slab != NULL) {
4735 				dom = &keg->uk_domain[slab->us_domain];
4736 				/*
4737 				 * keg_alloc_slab() always returns a slab on the
4738 				 * partial list.
4739 				 */
4740 				LIST_REMOVE(slab, us_link);
4741 				LIST_INSERT_HEAD(&dom->ud_free_slab, slab,
4742 				    us_link);
4743 				dom->ud_free_slabs++;
4744 				KEG_UNLOCK(keg, slab->us_domain);
4745 				break;
4746 			}
4747 			if (vm_domainset_iter_policy(&di, &domain) != 0)
4748 				vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask);
4749 		}
4750 	}
4751 }
4752 
4753 /*
4754  * Returns a snapshot of memory consumption in bytes.
4755  */
4756 size_t
4757 uma_zone_memory(uma_zone_t zone)
4758 {
4759 	size_t sz;
4760 	int i;
4761 
4762 	sz = 0;
4763 	if (zone->uz_flags & UMA_ZFLAG_CACHE) {
4764 		for (i = 0; i < vm_ndomains; i++)
4765 			sz += ZDOM_GET(zone, i)->uzd_nitems;
4766 		return (sz * zone->uz_size);
4767 	}
4768 	for (i = 0; i < vm_ndomains; i++)
4769 		sz += zone->uz_keg->uk_domain[i].ud_pages;
4770 
4771 	return (sz * PAGE_SIZE);
4772 }
4773 
4774 /* See uma.h */
4775 void
4776 uma_reclaim(int req)
4777 {
4778 
4779 	CTR0(KTR_UMA, "UMA: vm asked us to release pages!");
4780 	sx_xlock(&uma_reclaim_lock);
4781 	bucket_enable();
4782 
4783 	switch (req) {
4784 	case UMA_RECLAIM_TRIM:
4785 		zone_foreach(zone_trim, NULL);
4786 		break;
4787 	case UMA_RECLAIM_DRAIN:
4788 	case UMA_RECLAIM_DRAIN_CPU:
4789 		zone_foreach(zone_drain, NULL);
4790 		if (req == UMA_RECLAIM_DRAIN_CPU) {
4791 			pcpu_cache_drain_safe(NULL);
4792 			zone_foreach(zone_drain, NULL);
4793 		}
4794 		break;
4795 	default:
4796 		panic("unhandled reclamation request %d", req);
4797 	}
4798 
4799 	/*
4800 	 * Some slabs may have been freed but this zone will be visited early
4801 	 * we visit again so that we can free pages that are empty once other
4802 	 * zones are drained.  We have to do the same for buckets.
4803 	 */
4804 	zone_drain(slabzones[0], NULL);
4805 	zone_drain(slabzones[1], NULL);
4806 	bucket_zone_drain();
4807 	sx_xunlock(&uma_reclaim_lock);
4808 }
4809 
4810 static volatile int uma_reclaim_needed;
4811 
4812 void
4813 uma_reclaim_wakeup(void)
4814 {
4815 
4816 	if (atomic_fetchadd_int(&uma_reclaim_needed, 1) == 0)
4817 		wakeup(uma_reclaim);
4818 }
4819 
4820 void
4821 uma_reclaim_worker(void *arg __unused)
4822 {
4823 
4824 	for (;;) {
4825 		sx_xlock(&uma_reclaim_lock);
4826 		while (atomic_load_int(&uma_reclaim_needed) == 0)
4827 			sx_sleep(uma_reclaim, &uma_reclaim_lock, PVM, "umarcl",
4828 			    hz);
4829 		sx_xunlock(&uma_reclaim_lock);
4830 		EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM);
4831 		uma_reclaim(UMA_RECLAIM_DRAIN_CPU);
4832 		atomic_store_int(&uma_reclaim_needed, 0);
4833 		/* Don't fire more than once per-second. */
4834 		pause("umarclslp", hz);
4835 	}
4836 }
4837 
4838 /* See uma.h */
4839 void
4840 uma_zone_reclaim(uma_zone_t zone, int req)
4841 {
4842 
4843 	switch (req) {
4844 	case UMA_RECLAIM_TRIM:
4845 		zone_trim(zone, NULL);
4846 		break;
4847 	case UMA_RECLAIM_DRAIN:
4848 		zone_drain(zone, NULL);
4849 		break;
4850 	case UMA_RECLAIM_DRAIN_CPU:
4851 		pcpu_cache_drain_safe(zone);
4852 		zone_drain(zone, NULL);
4853 		break;
4854 	default:
4855 		panic("unhandled reclamation request %d", req);
4856 	}
4857 }
4858 
4859 /* See uma.h */
4860 int
4861 uma_zone_exhausted(uma_zone_t zone)
4862 {
4863 
4864 	return (atomic_load_32(&zone->uz_sleepers) > 0);
4865 }
4866 
4867 unsigned long
4868 uma_limit(void)
4869 {
4870 
4871 	return (uma_kmem_limit);
4872 }
4873 
4874 void
4875 uma_set_limit(unsigned long limit)
4876 {
4877 
4878 	uma_kmem_limit = limit;
4879 }
4880 
4881 unsigned long
4882 uma_size(void)
4883 {
4884 
4885 	return (atomic_load_long(&uma_kmem_total));
4886 }
4887 
4888 long
4889 uma_avail(void)
4890 {
4891 
4892 	return (uma_kmem_limit - uma_size());
4893 }
4894 
4895 #ifdef DDB
4896 /*
4897  * Generate statistics across both the zone and its per-cpu cache's.  Return
4898  * desired statistics if the pointer is non-NULL for that statistic.
4899  *
4900  * Note: does not update the zone statistics, as it can't safely clear the
4901  * per-CPU cache statistic.
4902  *
4903  */
4904 static void
4905 uma_zone_sumstat(uma_zone_t z, long *cachefreep, uint64_t *allocsp,
4906     uint64_t *freesp, uint64_t *sleepsp, uint64_t *xdomainp)
4907 {
4908 	uma_cache_t cache;
4909 	uint64_t allocs, frees, sleeps, xdomain;
4910 	int cachefree, cpu;
4911 
4912 	allocs = frees = sleeps = xdomain = 0;
4913 	cachefree = 0;
4914 	CPU_FOREACH(cpu) {
4915 		cache = &z->uz_cpu[cpu];
4916 		cachefree += cache->uc_allocbucket.ucb_cnt;
4917 		cachefree += cache->uc_freebucket.ucb_cnt;
4918 		xdomain += cache->uc_crossbucket.ucb_cnt;
4919 		cachefree += cache->uc_crossbucket.ucb_cnt;
4920 		allocs += cache->uc_allocs;
4921 		frees += cache->uc_frees;
4922 	}
4923 	allocs += counter_u64_fetch(z->uz_allocs);
4924 	frees += counter_u64_fetch(z->uz_frees);
4925 	xdomain += counter_u64_fetch(z->uz_xdomain);
4926 	sleeps += z->uz_sleeps;
4927 	if (cachefreep != NULL)
4928 		*cachefreep = cachefree;
4929 	if (allocsp != NULL)
4930 		*allocsp = allocs;
4931 	if (freesp != NULL)
4932 		*freesp = frees;
4933 	if (sleepsp != NULL)
4934 		*sleepsp = sleeps;
4935 	if (xdomainp != NULL)
4936 		*xdomainp = xdomain;
4937 }
4938 #endif /* DDB */
4939 
4940 static int
4941 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
4942 {
4943 	uma_keg_t kz;
4944 	uma_zone_t z;
4945 	int count;
4946 
4947 	count = 0;
4948 	rw_rlock(&uma_rwlock);
4949 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
4950 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
4951 			count++;
4952 	}
4953 	LIST_FOREACH(z, &uma_cachezones, uz_link)
4954 		count++;
4955 
4956 	rw_runlock(&uma_rwlock);
4957 	return (sysctl_handle_int(oidp, &count, 0, req));
4958 }
4959 
4960 static void
4961 uma_vm_zone_stats(struct uma_type_header *uth, uma_zone_t z, struct sbuf *sbuf,
4962     struct uma_percpu_stat *ups, bool internal)
4963 {
4964 	uma_zone_domain_t zdom;
4965 	uma_cache_t cache;
4966 	int i;
4967 
4968 
4969 	for (i = 0; i < vm_ndomains; i++) {
4970 		zdom = ZDOM_GET(z, i);
4971 		uth->uth_zone_free += zdom->uzd_nitems;
4972 	}
4973 	uth->uth_allocs = counter_u64_fetch(z->uz_allocs);
4974 	uth->uth_frees = counter_u64_fetch(z->uz_frees);
4975 	uth->uth_fails = counter_u64_fetch(z->uz_fails);
4976 	uth->uth_xdomain = counter_u64_fetch(z->uz_xdomain);
4977 	uth->uth_sleeps = z->uz_sleeps;
4978 
4979 	for (i = 0; i < mp_maxid + 1; i++) {
4980 		bzero(&ups[i], sizeof(*ups));
4981 		if (internal || CPU_ABSENT(i))
4982 			continue;
4983 		cache = &z->uz_cpu[i];
4984 		ups[i].ups_cache_free += cache->uc_allocbucket.ucb_cnt;
4985 		ups[i].ups_cache_free += cache->uc_freebucket.ucb_cnt;
4986 		ups[i].ups_cache_free += cache->uc_crossbucket.ucb_cnt;
4987 		ups[i].ups_allocs = cache->uc_allocs;
4988 		ups[i].ups_frees = cache->uc_frees;
4989 	}
4990 }
4991 
4992 static int
4993 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
4994 {
4995 	struct uma_stream_header ush;
4996 	struct uma_type_header uth;
4997 	struct uma_percpu_stat *ups;
4998 	struct sbuf sbuf;
4999 	uma_keg_t kz;
5000 	uma_zone_t z;
5001 	uint64_t items;
5002 	uint32_t kfree, pages;
5003 	int count, error, i;
5004 
5005 	error = sysctl_wire_old_buffer(req, 0);
5006 	if (error != 0)
5007 		return (error);
5008 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
5009 	sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL);
5010 	ups = malloc((mp_maxid + 1) * sizeof(*ups), M_TEMP, M_WAITOK);
5011 
5012 	count = 0;
5013 	rw_rlock(&uma_rwlock);
5014 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
5015 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
5016 			count++;
5017 	}
5018 
5019 	LIST_FOREACH(z, &uma_cachezones, uz_link)
5020 		count++;
5021 
5022 	/*
5023 	 * Insert stream header.
5024 	 */
5025 	bzero(&ush, sizeof(ush));
5026 	ush.ush_version = UMA_STREAM_VERSION;
5027 	ush.ush_maxcpus = (mp_maxid + 1);
5028 	ush.ush_count = count;
5029 	(void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
5030 
5031 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
5032 		kfree = pages = 0;
5033 		for (i = 0; i < vm_ndomains; i++) {
5034 			kfree += kz->uk_domain[i].ud_free_items;
5035 			pages += kz->uk_domain[i].ud_pages;
5036 		}
5037 		LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5038 			bzero(&uth, sizeof(uth));
5039 			strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5040 			uth.uth_align = kz->uk_align;
5041 			uth.uth_size = kz->uk_size;
5042 			uth.uth_rsize = kz->uk_rsize;
5043 			if (z->uz_max_items > 0) {
5044 				items = UZ_ITEMS_COUNT(z->uz_items);
5045 				uth.uth_pages = (items / kz->uk_ipers) *
5046 					kz->uk_ppera;
5047 			} else
5048 				uth.uth_pages = pages;
5049 			uth.uth_maxpages = (z->uz_max_items / kz->uk_ipers) *
5050 			    kz->uk_ppera;
5051 			uth.uth_limit = z->uz_max_items;
5052 			uth.uth_keg_free = kfree;
5053 
5054 			/*
5055 			 * A zone is secondary is it is not the first entry
5056 			 * on the keg's zone list.
5057 			 */
5058 			if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
5059 			    (LIST_FIRST(&kz->uk_zones) != z))
5060 				uth.uth_zone_flags = UTH_ZONE_SECONDARY;
5061 			uma_vm_zone_stats(&uth, z, &sbuf, ups,
5062 			    kz->uk_flags & UMA_ZFLAG_INTERNAL);
5063 			(void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5064 			for (i = 0; i < mp_maxid + 1; i++)
5065 				(void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5066 		}
5067 	}
5068 	LIST_FOREACH(z, &uma_cachezones, uz_link) {
5069 		bzero(&uth, sizeof(uth));
5070 		strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5071 		uth.uth_size = z->uz_size;
5072 		uma_vm_zone_stats(&uth, z, &sbuf, ups, false);
5073 		(void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5074 		for (i = 0; i < mp_maxid + 1; i++)
5075 			(void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5076 	}
5077 
5078 	rw_runlock(&uma_rwlock);
5079 	error = sbuf_finish(&sbuf);
5080 	sbuf_delete(&sbuf);
5081 	free(ups, M_TEMP);
5082 	return (error);
5083 }
5084 
5085 int
5086 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS)
5087 {
5088 	uma_zone_t zone = *(uma_zone_t *)arg1;
5089 	int error, max;
5090 
5091 	max = uma_zone_get_max(zone);
5092 	error = sysctl_handle_int(oidp, &max, 0, req);
5093 	if (error || !req->newptr)
5094 		return (error);
5095 
5096 	uma_zone_set_max(zone, max);
5097 
5098 	return (0);
5099 }
5100 
5101 int
5102 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS)
5103 {
5104 	uma_zone_t zone;
5105 	int cur;
5106 
5107 	/*
5108 	 * Some callers want to add sysctls for global zones that
5109 	 * may not yet exist so they pass a pointer to a pointer.
5110 	 */
5111 	if (arg2 == 0)
5112 		zone = *(uma_zone_t *)arg1;
5113 	else
5114 		zone = arg1;
5115 	cur = uma_zone_get_cur(zone);
5116 	return (sysctl_handle_int(oidp, &cur, 0, req));
5117 }
5118 
5119 static int
5120 sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS)
5121 {
5122 	uma_zone_t zone = arg1;
5123 	uint64_t cur;
5124 
5125 	cur = uma_zone_get_allocs(zone);
5126 	return (sysctl_handle_64(oidp, &cur, 0, req));
5127 }
5128 
5129 static int
5130 sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS)
5131 {
5132 	uma_zone_t zone = arg1;
5133 	uint64_t cur;
5134 
5135 	cur = uma_zone_get_frees(zone);
5136 	return (sysctl_handle_64(oidp, &cur, 0, req));
5137 }
5138 
5139 static int
5140 sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS)
5141 {
5142 	struct sbuf sbuf;
5143 	uma_zone_t zone = arg1;
5144 	int error;
5145 
5146 	sbuf_new_for_sysctl(&sbuf, NULL, 0, req);
5147 	if (zone->uz_flags != 0)
5148 		sbuf_printf(&sbuf, "0x%b", zone->uz_flags, PRINT_UMA_ZFLAGS);
5149 	else
5150 		sbuf_printf(&sbuf, "0");
5151 	error = sbuf_finish(&sbuf);
5152 	sbuf_delete(&sbuf);
5153 
5154 	return (error);
5155 }
5156 
5157 static int
5158 sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS)
5159 {
5160 	uma_keg_t keg = arg1;
5161 	int avail, effpct, total;
5162 
5163 	total = keg->uk_ppera * PAGE_SIZE;
5164 	if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0)
5165 		total += slabzone(keg->uk_ipers)->uz_keg->uk_rsize;
5166 	/*
5167 	 * We consider the client's requested size and alignment here, not the
5168 	 * real size determination uk_rsize, because we also adjust the real
5169 	 * size for internal implementation reasons (max bitset size).
5170 	 */
5171 	avail = keg->uk_ipers * roundup2(keg->uk_size, keg->uk_align + 1);
5172 	if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
5173 		avail *= mp_maxid + 1;
5174 	effpct = 100 * avail / total;
5175 	return (sysctl_handle_int(oidp, &effpct, 0, req));
5176 }
5177 
5178 static int
5179 sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS)
5180 {
5181 	uma_zone_t zone = arg1;
5182 	uint64_t cur;
5183 
5184 	cur = UZ_ITEMS_COUNT(atomic_load_64(&zone->uz_items));
5185 	return (sysctl_handle_64(oidp, &cur, 0, req));
5186 }
5187 
5188 #ifdef INVARIANTS
5189 static uma_slab_t
5190 uma_dbg_getslab(uma_zone_t zone, void *item)
5191 {
5192 	uma_slab_t slab;
5193 	uma_keg_t keg;
5194 	uint8_t *mem;
5195 
5196 	/*
5197 	 * It is safe to return the slab here even though the
5198 	 * zone is unlocked because the item's allocation state
5199 	 * essentially holds a reference.
5200 	 */
5201 	mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
5202 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5203 		return (NULL);
5204 	if (zone->uz_flags & UMA_ZFLAG_VTOSLAB)
5205 		return (vtoslab((vm_offset_t)mem));
5206 	keg = zone->uz_keg;
5207 	if ((keg->uk_flags & UMA_ZFLAG_HASH) == 0)
5208 		return ((uma_slab_t)(mem + keg->uk_pgoff));
5209 	KEG_LOCK(keg, 0);
5210 	slab = hash_sfind(&keg->uk_hash, mem);
5211 	KEG_UNLOCK(keg, 0);
5212 
5213 	return (slab);
5214 }
5215 
5216 static bool
5217 uma_dbg_zskip(uma_zone_t zone, void *mem)
5218 {
5219 
5220 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5221 		return (true);
5222 
5223 	return (uma_dbg_kskip(zone->uz_keg, mem));
5224 }
5225 
5226 static bool
5227 uma_dbg_kskip(uma_keg_t keg, void *mem)
5228 {
5229 	uintptr_t idx;
5230 
5231 	if (dbg_divisor == 0)
5232 		return (true);
5233 
5234 	if (dbg_divisor == 1)
5235 		return (false);
5236 
5237 	idx = (uintptr_t)mem >> PAGE_SHIFT;
5238 	if (keg->uk_ipers > 1) {
5239 		idx *= keg->uk_ipers;
5240 		idx += ((uintptr_t)mem & PAGE_MASK) / keg->uk_rsize;
5241 	}
5242 
5243 	if ((idx / dbg_divisor) * dbg_divisor != idx) {
5244 		counter_u64_add(uma_skip_cnt, 1);
5245 		return (true);
5246 	}
5247 	counter_u64_add(uma_dbg_cnt, 1);
5248 
5249 	return (false);
5250 }
5251 
5252 /*
5253  * Set up the slab's freei data such that uma_dbg_free can function.
5254  *
5255  */
5256 static void
5257 uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item)
5258 {
5259 	uma_keg_t keg;
5260 	int freei;
5261 
5262 	if (slab == NULL) {
5263 		slab = uma_dbg_getslab(zone, item);
5264 		if (slab == NULL)
5265 			panic("uma: item %p did not belong to zone %s\n",
5266 			    item, zone->uz_name);
5267 	}
5268 	keg = zone->uz_keg;
5269 	freei = slab_item_index(slab, keg, item);
5270 
5271 	if (BIT_ISSET(keg->uk_ipers, freei, slab_dbg_bits(slab, keg)))
5272 		panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)\n",
5273 		    item, zone, zone->uz_name, slab, freei);
5274 	BIT_SET_ATOMIC(keg->uk_ipers, freei, slab_dbg_bits(slab, keg));
5275 }
5276 
5277 /*
5278  * Verifies freed addresses.  Checks for alignment, valid slab membership
5279  * and duplicate frees.
5280  *
5281  */
5282 static void
5283 uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item)
5284 {
5285 	uma_keg_t keg;
5286 	int freei;
5287 
5288 	if (slab == NULL) {
5289 		slab = uma_dbg_getslab(zone, item);
5290 		if (slab == NULL)
5291 			panic("uma: Freed item %p did not belong to zone %s\n",
5292 			    item, zone->uz_name);
5293 	}
5294 	keg = zone->uz_keg;
5295 	freei = slab_item_index(slab, keg, item);
5296 
5297 	if (freei >= keg->uk_ipers)
5298 		panic("Invalid free of %p from zone %p(%s) slab %p(%d)\n",
5299 		    item, zone, zone->uz_name, slab, freei);
5300 
5301 	if (slab_item(slab, keg, freei) != item)
5302 		panic("Unaligned free of %p from zone %p(%s) slab %p(%d)\n",
5303 		    item, zone, zone->uz_name, slab, freei);
5304 
5305 	if (!BIT_ISSET(keg->uk_ipers, freei, slab_dbg_bits(slab, keg)))
5306 		panic("Duplicate free of %p from zone %p(%s) slab %p(%d)\n",
5307 		    item, zone, zone->uz_name, slab, freei);
5308 
5309 	BIT_CLR_ATOMIC(keg->uk_ipers, freei, slab_dbg_bits(slab, keg));
5310 }
5311 #endif /* INVARIANTS */
5312 
5313 #ifdef DDB
5314 static int64_t
5315 get_uma_stats(uma_keg_t kz, uma_zone_t z, uint64_t *allocs, uint64_t *used,
5316     uint64_t *sleeps, long *cachefree, uint64_t *xdomain)
5317 {
5318 	uint64_t frees;
5319 	int i;
5320 
5321 	if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
5322 		*allocs = counter_u64_fetch(z->uz_allocs);
5323 		frees = counter_u64_fetch(z->uz_frees);
5324 		*sleeps = z->uz_sleeps;
5325 		*cachefree = 0;
5326 		*xdomain = 0;
5327 	} else
5328 		uma_zone_sumstat(z, cachefree, allocs, &frees, sleeps,
5329 		    xdomain);
5330 	for (i = 0; i < vm_ndomains; i++) {
5331 		*cachefree += ZDOM_GET(z, i)->uzd_nitems;
5332 		if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
5333 		    (LIST_FIRST(&kz->uk_zones) != z)))
5334 			*cachefree += kz->uk_domain[i].ud_free_items;
5335 	}
5336 	*used = *allocs - frees;
5337 	return (((int64_t)*used + *cachefree) * kz->uk_size);
5338 }
5339 
5340 DB_SHOW_COMMAND(uma, db_show_uma)
5341 {
5342 	const char *fmt_hdr, *fmt_entry;
5343 	uma_keg_t kz;
5344 	uma_zone_t z;
5345 	uint64_t allocs, used, sleeps, xdomain;
5346 	long cachefree;
5347 	/* variables for sorting */
5348 	uma_keg_t cur_keg;
5349 	uma_zone_t cur_zone, last_zone;
5350 	int64_t cur_size, last_size, size;
5351 	int ties;
5352 
5353 	/* /i option produces machine-parseable CSV output */
5354 	if (modif[0] == 'i') {
5355 		fmt_hdr = "%s,%s,%s,%s,%s,%s,%s,%s,%s\n";
5356 		fmt_entry = "\"%s\",%ju,%jd,%ld,%ju,%ju,%u,%jd,%ju\n";
5357 	} else {
5358 		fmt_hdr = "%18s %6s %7s %7s %11s %7s %7s %10s %8s\n";
5359 		fmt_entry = "%18s %6ju %7jd %7ld %11ju %7ju %7u %10jd %8ju\n";
5360 	}
5361 
5362 	db_printf(fmt_hdr, "Zone", "Size", "Used", "Free", "Requests",
5363 	    "Sleeps", "Bucket", "Total Mem", "XFree");
5364 
5365 	/* Sort the zones with largest size first. */
5366 	last_zone = NULL;
5367 	last_size = INT64_MAX;
5368 	for (;;) {
5369 		cur_zone = NULL;
5370 		cur_size = -1;
5371 		ties = 0;
5372 		LIST_FOREACH(kz, &uma_kegs, uk_link) {
5373 			LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5374 				/*
5375 				 * In the case of size ties, print out zones
5376 				 * in the order they are encountered.  That is,
5377 				 * when we encounter the most recently output
5378 				 * zone, we have already printed all preceding
5379 				 * ties, and we must print all following ties.
5380 				 */
5381 				if (z == last_zone) {
5382 					ties = 1;
5383 					continue;
5384 				}
5385 				size = get_uma_stats(kz, z, &allocs, &used,
5386 				    &sleeps, &cachefree, &xdomain);
5387 				if (size > cur_size && size < last_size + ties)
5388 				{
5389 					cur_size = size;
5390 					cur_zone = z;
5391 					cur_keg = kz;
5392 				}
5393 			}
5394 		}
5395 		if (cur_zone == NULL)
5396 			break;
5397 
5398 		size = get_uma_stats(cur_keg, cur_zone, &allocs, &used,
5399 		    &sleeps, &cachefree, &xdomain);
5400 		db_printf(fmt_entry, cur_zone->uz_name,
5401 		    (uintmax_t)cur_keg->uk_size, (intmax_t)used, cachefree,
5402 		    (uintmax_t)allocs, (uintmax_t)sleeps,
5403 		    (unsigned)cur_zone->uz_bucket_size, (intmax_t)size,
5404 		    xdomain);
5405 
5406 		if (db_pager_quit)
5407 			return;
5408 		last_zone = cur_zone;
5409 		last_size = cur_size;
5410 	}
5411 }
5412 
5413 DB_SHOW_COMMAND(umacache, db_show_umacache)
5414 {
5415 	uma_zone_t z;
5416 	uint64_t allocs, frees;
5417 	long cachefree;
5418 	int i;
5419 
5420 	db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
5421 	    "Requests", "Bucket");
5422 	LIST_FOREACH(z, &uma_cachezones, uz_link) {
5423 		uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL, NULL);
5424 		for (i = 0; i < vm_ndomains; i++)
5425 			cachefree += ZDOM_GET(z, i)->uzd_nitems;
5426 		db_printf("%18s %8ju %8jd %8ld %12ju %8u\n",
5427 		    z->uz_name, (uintmax_t)z->uz_size,
5428 		    (intmax_t)(allocs - frees), cachefree,
5429 		    (uintmax_t)allocs, z->uz_bucket_size);
5430 		if (db_pager_quit)
5431 			return;
5432 	}
5433 }
5434 #endif	/* DDB */
5435