xref: /freebsd/sys/vm/uma_core.c (revision 473957941922d17be72089e385e2e2a995fd0e1c)
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  * These are the two zones from which all offpage uma_slab_ts are allocated.
121  *
122  * One zone is for slab headers that can represent a larger number of items,
123  * making the slabs themselves more efficient, and the other zone is for
124  * headers that are smaller and represent fewer items, making the headers more
125  * efficient.
126  */
127 #define	SLABZONE_SIZE(setsize)					\
128     (sizeof(struct uma_hash_slab) + BITSET_SIZE(setsize) * SLAB_BITSETS)
129 #define	SLABZONE0_SETSIZE	(PAGE_SIZE / 16)
130 #define	SLABZONE1_SETSIZE	SLAB_MAX_SETSIZE
131 #define	SLABZONE0_SIZE	SLABZONE_SIZE(SLABZONE0_SETSIZE)
132 #define	SLABZONE1_SIZE	SLABZONE_SIZE(SLABZONE1_SETSIZE)
133 static uma_zone_t slabzones[2];
134 
135 /*
136  * The initial hash tables come out of this zone so they can be allocated
137  * prior to malloc coming up.
138  */
139 static uma_zone_t hashzone;
140 
141 /* The boot-time adjusted value for cache line alignment. */
142 int uma_align_cache = 64 - 1;
143 
144 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets");
145 static MALLOC_DEFINE(M_UMA, "UMA", "UMA Misc");
146 
147 /*
148  * Are we allowed to allocate buckets?
149  */
150 static int bucketdisable = 1;
151 
152 /* Linked list of all kegs in the system */
153 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs);
154 
155 /* Linked list of all cache-only zones in the system */
156 static LIST_HEAD(,uma_zone) uma_cachezones =
157     LIST_HEAD_INITIALIZER(uma_cachezones);
158 
159 /* This RW lock protects the keg list */
160 static struct rwlock_padalign __exclusive_cache_line uma_rwlock;
161 
162 /*
163  * First available virual address for boot time allocations.
164  */
165 static vm_offset_t bootstart;
166 static vm_offset_t bootmem;
167 
168 static struct sx uma_reclaim_lock;
169 
170 /*
171  * kmem soft limit, initialized by uma_set_limit().  Ensure that early
172  * allocations don't trigger a wakeup of the reclaim thread.
173  */
174 unsigned long uma_kmem_limit = LONG_MAX;
175 SYSCTL_ULONG(_vm, OID_AUTO, uma_kmem_limit, CTLFLAG_RD, &uma_kmem_limit, 0,
176     "UMA kernel memory soft limit");
177 unsigned long uma_kmem_total;
178 SYSCTL_ULONG(_vm, OID_AUTO, uma_kmem_total, CTLFLAG_RD, &uma_kmem_total, 0,
179     "UMA kernel memory usage");
180 
181 /* Is the VM done starting up? */
182 static enum {
183 	BOOT_COLD,
184 	BOOT_KVA,
185 	BOOT_RUNNING,
186 	BOOT_SHUTDOWN,
187 } booted = BOOT_COLD;
188 
189 /*
190  * This is the handle used to schedule events that need to happen
191  * outside of the allocation fast path.
192  */
193 static struct callout uma_callout;
194 #define	UMA_TIMEOUT	20		/* Seconds for callout interval. */
195 
196 /*
197  * This structure is passed as the zone ctor arg so that I don't have to create
198  * a special allocation function just for zones.
199  */
200 struct uma_zctor_args {
201 	const char *name;
202 	size_t size;
203 	uma_ctor ctor;
204 	uma_dtor dtor;
205 	uma_init uminit;
206 	uma_fini fini;
207 	uma_import import;
208 	uma_release release;
209 	void *arg;
210 	uma_keg_t keg;
211 	int align;
212 	uint32_t flags;
213 };
214 
215 struct uma_kctor_args {
216 	uma_zone_t zone;
217 	size_t size;
218 	uma_init uminit;
219 	uma_fini fini;
220 	int align;
221 	uint32_t flags;
222 };
223 
224 struct uma_bucket_zone {
225 	uma_zone_t	ubz_zone;
226 	const char	*ubz_name;
227 	int		ubz_entries;	/* Number of items it can hold. */
228 	int		ubz_maxsize;	/* Maximum allocation size per-item. */
229 };
230 
231 /*
232  * Compute the actual number of bucket entries to pack them in power
233  * of two sizes for more efficient space utilization.
234  */
235 #define	BUCKET_SIZE(n)						\
236     (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *))
237 
238 #define	BUCKET_MAX	BUCKET_SIZE(256)
239 #define	BUCKET_MIN	2
240 
241 struct uma_bucket_zone bucket_zones[] = {
242 	/* Literal bucket sizes. */
243 	{ NULL, "2 Bucket", 2, 4096 },
244 	{ NULL, "4 Bucket", 4, 3072 },
245 	{ NULL, "8 Bucket", 8, 2048 },
246 	{ NULL, "16 Bucket", 16, 1024 },
247 	/* Rounded down power of 2 sizes for efficiency. */
248 	{ NULL, "32 Bucket", BUCKET_SIZE(32), 512 },
249 	{ NULL, "64 Bucket", BUCKET_SIZE(64), 256 },
250 	{ NULL, "128 Bucket", BUCKET_SIZE(128), 128 },
251 	{ NULL, "256 Bucket", BUCKET_SIZE(256), 64 },
252 	{ NULL, NULL, 0}
253 };
254 
255 /*
256  * Flags and enumerations to be passed to internal functions.
257  */
258 enum zfreeskip {
259 	SKIP_NONE =	0,
260 	SKIP_CNT =	0x00000001,
261 	SKIP_DTOR =	0x00010000,
262 	SKIP_FINI =	0x00020000,
263 };
264 
265 /* Prototypes.. */
266 
267 void	uma_startup1(vm_offset_t);
268 void	uma_startup2(void);
269 
270 static void *noobj_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
271 static void *page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
272 static void *pcpu_page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
273 static void *startup_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
274 static void *contig_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
275 static void page_free(void *, vm_size_t, uint8_t);
276 static void pcpu_page_free(void *, vm_size_t, uint8_t);
277 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int, int, int);
278 static void cache_drain(uma_zone_t);
279 static void bucket_drain(uma_zone_t, uma_bucket_t);
280 static void bucket_cache_reclaim(uma_zone_t zone, bool);
281 static int keg_ctor(void *, int, void *, int);
282 static void keg_dtor(void *, int, void *);
283 static int zone_ctor(void *, int, void *, int);
284 static void zone_dtor(void *, int, void *);
285 static inline void item_dtor(uma_zone_t zone, void *item, int size,
286     void *udata, enum zfreeskip skip);
287 static int zero_init(void *, int, int);
288 static void zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata,
289     int itemdomain, bool ws);
290 static void zone_foreach(void (*zfunc)(uma_zone_t, void *), void *);
291 static void zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *), void *);
292 static void zone_timeout(uma_zone_t zone, void *);
293 static int hash_alloc(struct uma_hash *, u_int);
294 static int hash_expand(struct uma_hash *, struct uma_hash *);
295 static void hash_free(struct uma_hash *hash);
296 static void uma_timeout(void *);
297 static void uma_startup3(void);
298 static void uma_shutdown(void);
299 static void *zone_alloc_item(uma_zone_t, void *, int, int);
300 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip);
301 static int zone_alloc_limit(uma_zone_t zone, int count, int flags);
302 static void zone_free_limit(uma_zone_t zone, int count);
303 static void bucket_enable(void);
304 static void bucket_init(void);
305 static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int);
306 static void bucket_free(uma_zone_t zone, uma_bucket_t, void *);
307 static void bucket_zone_drain(void);
308 static uma_bucket_t zone_alloc_bucket(uma_zone_t, void *, int, int);
309 static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab);
310 static void slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item);
311 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit,
312     uma_fini fini, int align, uint32_t flags);
313 static int zone_import(void *, void **, int, int, int);
314 static void zone_release(void *, void **, int);
315 static bool cache_alloc(uma_zone_t, uma_cache_t, void *, int);
316 static bool cache_free(uma_zone_t, uma_cache_t, void *, void *, int);
317 
318 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS);
319 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS);
320 static int sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS);
321 static int sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS);
322 static int sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS);
323 static int sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS);
324 static int sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS);
325 
326 static uint64_t uma_zone_get_allocs(uma_zone_t zone);
327 
328 static SYSCTL_NODE(_vm, OID_AUTO, debug, CTLFLAG_RD, 0,
329     "Memory allocation debugging");
330 
331 #ifdef INVARIANTS
332 static uint64_t uma_keg_get_allocs(uma_keg_t zone);
333 static inline struct noslabbits *slab_dbg_bits(uma_slab_t slab, uma_keg_t keg);
334 
335 static bool uma_dbg_kskip(uma_keg_t keg, void *mem);
336 static bool uma_dbg_zskip(uma_zone_t zone, void *mem);
337 static void uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item);
338 static void uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item);
339 
340 static u_int dbg_divisor = 1;
341 SYSCTL_UINT(_vm_debug, OID_AUTO, divisor,
342     CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &dbg_divisor, 0,
343     "Debug & thrash every this item in memory allocator");
344 
345 static counter_u64_t uma_dbg_cnt = EARLY_COUNTER;
346 static counter_u64_t uma_skip_cnt = EARLY_COUNTER;
347 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, trashed, CTLFLAG_RD,
348     &uma_dbg_cnt, "memory items debugged");
349 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, skipped, CTLFLAG_RD,
350     &uma_skip_cnt, "memory items skipped, not debugged");
351 #endif
352 
353 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
354 
355 SYSCTL_NODE(_vm, OID_AUTO, uma, CTLFLAG_RW, 0, "Universal Memory Allocator");
356 
357 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLFLAG_MPSAFE|CTLTYPE_INT,
358     0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones");
359 
360 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLFLAG_MPSAFE|CTLTYPE_STRUCT,
361     0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats");
362 
363 static int zone_warnings = 1;
364 SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RWTUN, &zone_warnings, 0,
365     "Warn when UMA zones becomes full");
366 
367 static int multipage_slabs = 1;
368 TUNABLE_INT("vm.debug.uma_multipage_slabs", &multipage_slabs);
369 SYSCTL_INT(_vm_debug, OID_AUTO, uma_multipage_slabs,
370     CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &multipage_slabs, 0,
371     "UMA may choose larger slab sizes for better efficiency");
372 
373 /*
374  * Select the slab zone for an offpage slab with the given maximum item count.
375  */
376 static inline uma_zone_t
377 slabzone(int ipers)
378 {
379 
380 	return (slabzones[ipers > SLABZONE0_SETSIZE]);
381 }
382 
383 /*
384  * This routine checks to see whether or not it's safe to enable buckets.
385  */
386 static void
387 bucket_enable(void)
388 {
389 
390 	KASSERT(booted >= BOOT_KVA, ("Bucket enable before init"));
391 	bucketdisable = vm_page_count_min();
392 }
393 
394 /*
395  * Initialize bucket_zones, the array of zones of buckets of various sizes.
396  *
397  * For each zone, calculate the memory required for each bucket, consisting
398  * of the header and an array of pointers.
399  */
400 static void
401 bucket_init(void)
402 {
403 	struct uma_bucket_zone *ubz;
404 	int size;
405 
406 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) {
407 		size = roundup(sizeof(struct uma_bucket), sizeof(void *));
408 		size += sizeof(void *) * ubz->ubz_entries;
409 		ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
410 		    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
411 		    UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET |
412 		    UMA_ZONE_FIRSTTOUCH);
413 	}
414 }
415 
416 /*
417  * Given a desired number of entries for a bucket, return the zone from which
418  * to allocate the bucket.
419  */
420 static struct uma_bucket_zone *
421 bucket_zone_lookup(int entries)
422 {
423 	struct uma_bucket_zone *ubz;
424 
425 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
426 		if (ubz->ubz_entries >= entries)
427 			return (ubz);
428 	ubz--;
429 	return (ubz);
430 }
431 
432 static struct uma_bucket_zone *
433 bucket_zone_max(uma_zone_t zone, int nitems)
434 {
435 	struct uma_bucket_zone *ubz;
436 	int bpcpu;
437 
438 	bpcpu = 2;
439 	if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
440 		/* Count the cross-domain bucket. */
441 		bpcpu++;
442 
443 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
444 		if (ubz->ubz_entries * bpcpu * mp_ncpus > nitems)
445 			break;
446 	if (ubz == &bucket_zones[0])
447 		ubz = NULL;
448 	else
449 		ubz--;
450 	return (ubz);
451 }
452 
453 static int
454 bucket_select(int size)
455 {
456 	struct uma_bucket_zone *ubz;
457 
458 	ubz = &bucket_zones[0];
459 	if (size > ubz->ubz_maxsize)
460 		return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1);
461 
462 	for (; ubz->ubz_entries != 0; ubz++)
463 		if (ubz->ubz_maxsize < size)
464 			break;
465 	ubz--;
466 	return (ubz->ubz_entries);
467 }
468 
469 static uma_bucket_t
470 bucket_alloc(uma_zone_t zone, void *udata, int flags)
471 {
472 	struct uma_bucket_zone *ubz;
473 	uma_bucket_t bucket;
474 
475 	/*
476 	 * Don't allocate buckets early in boot.
477 	 */
478 	if (__predict_false(booted < BOOT_KVA))
479 		return (NULL);
480 
481 	/*
482 	 * To limit bucket recursion we store the original zone flags
483 	 * in a cookie passed via zalloc_arg/zfree_arg.  This allows the
484 	 * NOVM flag to persist even through deep recursions.  We also
485 	 * store ZFLAG_BUCKET once we have recursed attempting to allocate
486 	 * a bucket for a bucket zone so we do not allow infinite bucket
487 	 * recursion.  This cookie will even persist to frees of unused
488 	 * buckets via the allocation path or bucket allocations in the
489 	 * free path.
490 	 */
491 	if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
492 		udata = (void *)(uintptr_t)zone->uz_flags;
493 	else {
494 		if ((uintptr_t)udata & UMA_ZFLAG_BUCKET)
495 			return (NULL);
496 		udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET);
497 	}
498 	if (((uintptr_t)udata & UMA_ZONE_VM) != 0)
499 		flags |= M_NOVM;
500 	ubz = bucket_zone_lookup(zone->uz_bucket_size);
501 	if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0)
502 		ubz++;
503 	bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags);
504 	if (bucket) {
505 #ifdef INVARIANTS
506 		bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
507 #endif
508 		bucket->ub_cnt = 0;
509 		bucket->ub_entries = ubz->ubz_entries;
510 		bucket->ub_seq = SMR_SEQ_INVALID;
511 		CTR3(KTR_UMA, "bucket_alloc: zone %s(%p) allocated bucket %p",
512 		    zone->uz_name, zone, bucket);
513 	}
514 
515 	return (bucket);
516 }
517 
518 static void
519 bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata)
520 {
521 	struct uma_bucket_zone *ubz;
522 
523 	if (bucket->ub_cnt != 0)
524 		bucket_drain(zone, bucket);
525 
526 	KASSERT(bucket->ub_cnt == 0,
527 	    ("bucket_free: Freeing a non free bucket."));
528 	KASSERT(bucket->ub_seq == SMR_SEQ_INVALID,
529 	    ("bucket_free: Freeing an SMR bucket."));
530 	if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
531 		udata = (void *)(uintptr_t)zone->uz_flags;
532 	ubz = bucket_zone_lookup(bucket->ub_entries);
533 	uma_zfree_arg(ubz->ubz_zone, bucket, udata);
534 }
535 
536 static void
537 bucket_zone_drain(void)
538 {
539 	struct uma_bucket_zone *ubz;
540 
541 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
542 		uma_zone_reclaim(ubz->ubz_zone, UMA_RECLAIM_DRAIN);
543 }
544 
545 /*
546  * Acquire the domain lock and record contention.
547  */
548 static uma_zone_domain_t
549 zone_domain_lock(uma_zone_t zone, int domain)
550 {
551 	uma_zone_domain_t zdom;
552 	bool lockfail;
553 
554 	zdom = ZDOM_GET(zone, domain);
555 	lockfail = false;
556 	if (ZDOM_OWNED(zdom))
557 		lockfail = true;
558 	ZDOM_LOCK(zdom);
559 	/* This is unsynchronized.  The counter does not need to be precise. */
560 	if (lockfail && zone->uz_bucket_size < zone->uz_bucket_size_max)
561 		zone->uz_bucket_size++;
562 	return (zdom);
563 }
564 
565 /*
566  * Search for the domain with the least cached items and return it, breaking
567  * ties with a preferred domain by returning it.
568  */
569 static __noinline int
570 zone_domain_lowest(uma_zone_t zone, int pref)
571 {
572 	long least, nitems;
573 	int domain;
574 	int i;
575 
576 	least = LONG_MAX;
577 	domain = 0;
578 	for (i = 0; i < vm_ndomains; i++) {
579 		nitems = ZDOM_GET(zone, i)->uzd_nitems;
580 		if (nitems < least) {
581 			domain = i;
582 			least = nitems;
583 		} else if (nitems == least && (i == pref || domain == pref))
584 			domain = pref;
585 	}
586 
587 	return (domain);
588 }
589 
590 /*
591  * Search for the domain with the most cached items and return it or the
592  * preferred domain if it has enough to proceed.
593  */
594 static __noinline int
595 zone_domain_highest(uma_zone_t zone, int pref)
596 {
597 	long most, nitems;
598 	int domain;
599 	int i;
600 
601 	if (ZDOM_GET(zone, pref)->uzd_nitems > BUCKET_MAX)
602 		return (pref);
603 
604 	most = 0;
605 	domain = 0;
606 	for (i = 0; i < vm_ndomains; i++) {
607 		nitems = ZDOM_GET(zone, i)->uzd_nitems;
608 		if (nitems > most) {
609 			domain = i;
610 			most = nitems;
611 		}
612 	}
613 
614 	return (domain);
615 }
616 
617 /*
618  * Safely subtract cnt from imax.
619  */
620 static void
621 zone_domain_imax_sub(uma_zone_domain_t zdom, int cnt)
622 {
623 	long new;
624 	long old;
625 
626 	old = zdom->uzd_imax;
627 	do {
628 		if (old <= cnt)
629 			new = 0;
630 		else
631 			new = old - cnt;
632 	} while (atomic_fcmpset_long(&zdom->uzd_imax, &old, new) == 0);
633 }
634 
635 /*
636  * Set the maximum imax value.
637  */
638 static void
639 zone_domain_imax_set(uma_zone_domain_t zdom, int nitems)
640 {
641 	long old;
642 
643 	old = zdom->uzd_imax;
644 	do {
645 		if (old >= nitems)
646 			break;
647 	} while (atomic_fcmpset_long(&zdom->uzd_imax, &old, nitems) == 0);
648 }
649 
650 /*
651  * Attempt to satisfy an allocation by retrieving a full bucket from one of the
652  * zone's caches.  If a bucket is found the zone is not locked on return.
653  */
654 static uma_bucket_t
655 zone_fetch_bucket(uma_zone_t zone, uma_zone_domain_t zdom, bool reclaim)
656 {
657 	uma_bucket_t bucket;
658 	int i;
659 	bool dtor = false;
660 
661 	ZDOM_LOCK_ASSERT(zdom);
662 
663 	if ((bucket = STAILQ_FIRST(&zdom->uzd_buckets)) == NULL)
664 		return (NULL);
665 
666 	/* SMR Buckets can not be re-used until readers expire. */
667 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
668 	    bucket->ub_seq != SMR_SEQ_INVALID) {
669 		if (!smr_poll(zone->uz_smr, bucket->ub_seq, false))
670 			return (NULL);
671 		bucket->ub_seq = SMR_SEQ_INVALID;
672 		dtor = (zone->uz_dtor != NULL) || UMA_ALWAYS_CTORDTOR;
673 		if (STAILQ_NEXT(bucket, ub_link) != NULL)
674 			zdom->uzd_seq = STAILQ_NEXT(bucket, ub_link)->ub_seq;
675 	}
676 	MPASS(zdom->uzd_nitems >= bucket->ub_cnt);
677 	STAILQ_REMOVE_HEAD(&zdom->uzd_buckets, ub_link);
678 	zdom->uzd_nitems -= bucket->ub_cnt;
679 
680 	/*
681 	 * Shift the bounds of the current WSS interval to avoid
682 	 * perturbing the estimate.
683 	 */
684 	if (reclaim) {
685 		zdom->uzd_imin -= lmin(zdom->uzd_imin, bucket->ub_cnt);
686 		zone_domain_imax_sub(zdom, bucket->ub_cnt);
687 	} else if (zdom->uzd_imin > zdom->uzd_nitems)
688 		zdom->uzd_imin = zdom->uzd_nitems;
689 
690 	ZDOM_UNLOCK(zdom);
691 	if (dtor)
692 		for (i = 0; i < bucket->ub_cnt; i++)
693 			item_dtor(zone, bucket->ub_bucket[i], zone->uz_size,
694 			    NULL, SKIP_NONE);
695 
696 	return (bucket);
697 }
698 
699 /*
700  * Insert a full bucket into the specified cache.  The "ws" parameter indicates
701  * whether the bucket's contents should be counted as part of the zone's working
702  * set.  The bucket may be freed if it exceeds the bucket limit.
703  */
704 static void
705 zone_put_bucket(uma_zone_t zone, int domain, uma_bucket_t bucket, void *udata,
706     const bool ws)
707 {
708 	uma_zone_domain_t zdom;
709 
710 	/* We don't cache empty buckets.  This can happen after a reclaim. */
711 	if (bucket->ub_cnt == 0)
712 		goto out;
713 	zdom = zone_domain_lock(zone, domain);
714 
715 	KASSERT(!ws || zdom->uzd_nitems < zone->uz_bucket_max,
716 	    ("%s: zone %p overflow", __func__, zone));
717 
718 	/*
719 	 * Conditionally set the maximum number of items.
720 	 */
721 	zdom->uzd_nitems += bucket->ub_cnt;
722 	if (__predict_true(zdom->uzd_nitems < zone->uz_bucket_max)) {
723 		if (ws)
724 			zone_domain_imax_set(zdom, zdom->uzd_nitems);
725 		if (STAILQ_EMPTY(&zdom->uzd_buckets))
726 			zdom->uzd_seq = bucket->ub_seq;
727 		STAILQ_INSERT_TAIL(&zdom->uzd_buckets, bucket, ub_link);
728 		ZDOM_UNLOCK(zdom);
729 		return;
730 	}
731 	zdom->uzd_nitems -= bucket->ub_cnt;
732 	ZDOM_UNLOCK(zdom);
733 out:
734 	bucket_free(zone, bucket, udata);
735 }
736 
737 /* Pops an item out of a per-cpu cache bucket. */
738 static inline void *
739 cache_bucket_pop(uma_cache_t cache, uma_cache_bucket_t bucket)
740 {
741 	void *item;
742 
743 	CRITICAL_ASSERT(curthread);
744 
745 	bucket->ucb_cnt--;
746 	item = bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt];
747 #ifdef INVARIANTS
748 	bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = NULL;
749 	KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled."));
750 #endif
751 	cache->uc_allocs++;
752 
753 	return (item);
754 }
755 
756 /* Pushes an item into a per-cpu cache bucket. */
757 static inline void
758 cache_bucket_push(uma_cache_t cache, uma_cache_bucket_t bucket, void *item)
759 {
760 
761 	CRITICAL_ASSERT(curthread);
762 	KASSERT(bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] == NULL,
763 	    ("uma_zfree: Freeing to non free bucket index."));
764 
765 	bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = item;
766 	bucket->ucb_cnt++;
767 	cache->uc_frees++;
768 }
769 
770 /*
771  * Unload a UMA bucket from a per-cpu cache.
772  */
773 static inline uma_bucket_t
774 cache_bucket_unload(uma_cache_bucket_t bucket)
775 {
776 	uma_bucket_t b;
777 
778 	b = bucket->ucb_bucket;
779 	if (b != NULL) {
780 		MPASS(b->ub_entries == bucket->ucb_entries);
781 		b->ub_cnt = bucket->ucb_cnt;
782 		bucket->ucb_bucket = NULL;
783 		bucket->ucb_entries = bucket->ucb_cnt = 0;
784 	}
785 
786 	return (b);
787 }
788 
789 static inline uma_bucket_t
790 cache_bucket_unload_alloc(uma_cache_t cache)
791 {
792 
793 	return (cache_bucket_unload(&cache->uc_allocbucket));
794 }
795 
796 static inline uma_bucket_t
797 cache_bucket_unload_free(uma_cache_t cache)
798 {
799 
800 	return (cache_bucket_unload(&cache->uc_freebucket));
801 }
802 
803 static inline uma_bucket_t
804 cache_bucket_unload_cross(uma_cache_t cache)
805 {
806 
807 	return (cache_bucket_unload(&cache->uc_crossbucket));
808 }
809 
810 /*
811  * Load a bucket into a per-cpu cache bucket.
812  */
813 static inline void
814 cache_bucket_load(uma_cache_bucket_t bucket, uma_bucket_t b)
815 {
816 
817 	CRITICAL_ASSERT(curthread);
818 	MPASS(bucket->ucb_bucket == NULL);
819 	MPASS(b->ub_seq == SMR_SEQ_INVALID);
820 
821 	bucket->ucb_bucket = b;
822 	bucket->ucb_cnt = b->ub_cnt;
823 	bucket->ucb_entries = b->ub_entries;
824 }
825 
826 static inline void
827 cache_bucket_load_alloc(uma_cache_t cache, uma_bucket_t b)
828 {
829 
830 	cache_bucket_load(&cache->uc_allocbucket, b);
831 }
832 
833 static inline void
834 cache_bucket_load_free(uma_cache_t cache, uma_bucket_t b)
835 {
836 
837 	cache_bucket_load(&cache->uc_freebucket, b);
838 }
839 
840 #ifdef NUMA
841 static inline void
842 cache_bucket_load_cross(uma_cache_t cache, uma_bucket_t b)
843 {
844 
845 	cache_bucket_load(&cache->uc_crossbucket, b);
846 }
847 #endif
848 
849 /*
850  * Copy and preserve ucb_spare.
851  */
852 static inline void
853 cache_bucket_copy(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
854 {
855 
856 	b1->ucb_bucket = b2->ucb_bucket;
857 	b1->ucb_entries = b2->ucb_entries;
858 	b1->ucb_cnt = b2->ucb_cnt;
859 }
860 
861 /*
862  * Swap two cache buckets.
863  */
864 static inline void
865 cache_bucket_swap(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
866 {
867 	struct uma_cache_bucket b3;
868 
869 	CRITICAL_ASSERT(curthread);
870 
871 	cache_bucket_copy(&b3, b1);
872 	cache_bucket_copy(b1, b2);
873 	cache_bucket_copy(b2, &b3);
874 }
875 
876 /*
877  * Attempt to fetch a bucket from a zone on behalf of the current cpu cache.
878  */
879 static uma_bucket_t
880 cache_fetch_bucket(uma_zone_t zone, uma_cache_t cache, int domain)
881 {
882 	uma_zone_domain_t zdom;
883 	uma_bucket_t bucket;
884 
885 	/*
886 	 * Avoid the lock if possible.
887 	 */
888 	zdom = ZDOM_GET(zone, domain);
889 	if (zdom->uzd_nitems == 0)
890 		return (NULL);
891 
892 	if ((cache_uz_flags(cache) & UMA_ZONE_SMR) != 0 &&
893 	    !smr_poll(zone->uz_smr, zdom->uzd_seq, false))
894 		return (NULL);
895 
896 	/*
897 	 * Check the zone's cache of buckets.
898 	 */
899 	zdom = zone_domain_lock(zone, domain);
900 	if ((bucket = zone_fetch_bucket(zone, zdom, false)) != NULL) {
901 		KASSERT(bucket->ub_cnt != 0,
902 		    ("cache_fetch_bucket: Returning an empty bucket."));
903 		return (bucket);
904 	}
905 	ZDOM_UNLOCK(zdom);
906 
907 	return (NULL);
908 }
909 
910 static void
911 zone_log_warning(uma_zone_t zone)
912 {
913 	static const struct timeval warninterval = { 300, 0 };
914 
915 	if (!zone_warnings || zone->uz_warning == NULL)
916 		return;
917 
918 	if (ratecheck(&zone->uz_ratecheck, &warninterval))
919 		printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning);
920 }
921 
922 static inline void
923 zone_maxaction(uma_zone_t zone)
924 {
925 
926 	if (zone->uz_maxaction.ta_func != NULL)
927 		taskqueue_enqueue(taskqueue_thread, &zone->uz_maxaction);
928 }
929 
930 /*
931  * Routine called by timeout which is used to fire off some time interval
932  * based calculations.  (stats, hash size, etc.)
933  *
934  * Arguments:
935  *	arg   Unused
936  *
937  * Returns:
938  *	Nothing
939  */
940 static void
941 uma_timeout(void *unused)
942 {
943 	bucket_enable();
944 	zone_foreach(zone_timeout, NULL);
945 
946 	/* Reschedule this event */
947 	callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
948 }
949 
950 /*
951  * Update the working set size estimate for the zone's bucket cache.
952  * The constants chosen here are somewhat arbitrary.  With an update period of
953  * 20s (UMA_TIMEOUT), this estimate is dominated by zone activity over the
954  * last 100s.
955  */
956 static void
957 zone_domain_update_wss(uma_zone_domain_t zdom)
958 {
959 	long wss;
960 
961 	ZDOM_LOCK(zdom);
962 	MPASS(zdom->uzd_imax >= zdom->uzd_imin);
963 	wss = zdom->uzd_imax - zdom->uzd_imin;
964 	zdom->uzd_imax = zdom->uzd_imin = zdom->uzd_nitems;
965 	zdom->uzd_wss = (4 * wss + zdom->uzd_wss) / 5;
966 	ZDOM_UNLOCK(zdom);
967 }
968 
969 /*
970  * Routine to perform timeout driven calculations.  This expands the
971  * hashes and does per cpu statistics aggregation.
972  *
973  *  Returns nothing.
974  */
975 static void
976 zone_timeout(uma_zone_t zone, void *unused)
977 {
978 	uma_keg_t keg;
979 	u_int slabs, pages;
980 
981 	if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
982 		goto update_wss;
983 
984 	keg = zone->uz_keg;
985 
986 	/*
987 	 * Hash zones are non-numa by definition so the first domain
988 	 * is the only one present.
989 	 */
990 	KEG_LOCK(keg, 0);
991 	pages = keg->uk_domain[0].ud_pages;
992 
993 	/*
994 	 * Expand the keg hash table.
995 	 *
996 	 * This is done if the number of slabs is larger than the hash size.
997 	 * What I'm trying to do here is completely reduce collisions.  This
998 	 * may be a little aggressive.  Should I allow for two collisions max?
999 	 */
1000 	if ((slabs = pages / keg->uk_ppera) > keg->uk_hash.uh_hashsize) {
1001 		struct uma_hash newhash;
1002 		struct uma_hash oldhash;
1003 		int ret;
1004 
1005 		/*
1006 		 * This is so involved because allocating and freeing
1007 		 * while the keg lock is held will lead to deadlock.
1008 		 * I have to do everything in stages and check for
1009 		 * races.
1010 		 */
1011 		KEG_UNLOCK(keg, 0);
1012 		ret = hash_alloc(&newhash, 1 << fls(slabs));
1013 		KEG_LOCK(keg, 0);
1014 		if (ret) {
1015 			if (hash_expand(&keg->uk_hash, &newhash)) {
1016 				oldhash = keg->uk_hash;
1017 				keg->uk_hash = newhash;
1018 			} else
1019 				oldhash = newhash;
1020 
1021 			KEG_UNLOCK(keg, 0);
1022 			hash_free(&oldhash);
1023 			goto update_wss;
1024 		}
1025 	}
1026 	KEG_UNLOCK(keg, 0);
1027 
1028 update_wss:
1029 	for (int i = 0; i < vm_ndomains; i++)
1030 		zone_domain_update_wss(ZDOM_GET(zone, i));
1031 }
1032 
1033 /*
1034  * Allocate and zero fill the next sized hash table from the appropriate
1035  * backing store.
1036  *
1037  * Arguments:
1038  *	hash  A new hash structure with the old hash size in uh_hashsize
1039  *
1040  * Returns:
1041  *	1 on success and 0 on failure.
1042  */
1043 static int
1044 hash_alloc(struct uma_hash *hash, u_int size)
1045 {
1046 	size_t alloc;
1047 
1048 	KASSERT(powerof2(size), ("hash size must be power of 2"));
1049 	if (size > UMA_HASH_SIZE_INIT)  {
1050 		hash->uh_hashsize = size;
1051 		alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
1052 		hash->uh_slab_hash = malloc(alloc, M_UMAHASH, M_NOWAIT);
1053 	} else {
1054 		alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
1055 		hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
1056 		    UMA_ANYDOMAIN, M_WAITOK);
1057 		hash->uh_hashsize = UMA_HASH_SIZE_INIT;
1058 	}
1059 	if (hash->uh_slab_hash) {
1060 		bzero(hash->uh_slab_hash, alloc);
1061 		hash->uh_hashmask = hash->uh_hashsize - 1;
1062 		return (1);
1063 	}
1064 
1065 	return (0);
1066 }
1067 
1068 /*
1069  * Expands the hash table for HASH zones.  This is done from zone_timeout
1070  * to reduce collisions.  This must not be done in the regular allocation
1071  * path, otherwise, we can recurse on the vm while allocating pages.
1072  *
1073  * Arguments:
1074  *	oldhash  The hash you want to expand
1075  *	newhash  The hash structure for the new table
1076  *
1077  * Returns:
1078  *	Nothing
1079  *
1080  * Discussion:
1081  */
1082 static int
1083 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
1084 {
1085 	uma_hash_slab_t slab;
1086 	u_int hval;
1087 	u_int idx;
1088 
1089 	if (!newhash->uh_slab_hash)
1090 		return (0);
1091 
1092 	if (oldhash->uh_hashsize >= newhash->uh_hashsize)
1093 		return (0);
1094 
1095 	/*
1096 	 * I need to investigate hash algorithms for resizing without a
1097 	 * full rehash.
1098 	 */
1099 
1100 	for (idx = 0; idx < oldhash->uh_hashsize; idx++)
1101 		while (!LIST_EMPTY(&oldhash->uh_slab_hash[idx])) {
1102 			slab = LIST_FIRST(&oldhash->uh_slab_hash[idx]);
1103 			LIST_REMOVE(slab, uhs_hlink);
1104 			hval = UMA_HASH(newhash, slab->uhs_data);
1105 			LIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
1106 			    slab, uhs_hlink);
1107 		}
1108 
1109 	return (1);
1110 }
1111 
1112 /*
1113  * Free the hash bucket to the appropriate backing store.
1114  *
1115  * Arguments:
1116  *	slab_hash  The hash bucket we're freeing
1117  *	hashsize   The number of entries in that hash bucket
1118  *
1119  * Returns:
1120  *	Nothing
1121  */
1122 static void
1123 hash_free(struct uma_hash *hash)
1124 {
1125 	if (hash->uh_slab_hash == NULL)
1126 		return;
1127 	if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
1128 		zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE);
1129 	else
1130 		free(hash->uh_slab_hash, M_UMAHASH);
1131 }
1132 
1133 /*
1134  * Frees all outstanding items in a bucket
1135  *
1136  * Arguments:
1137  *	zone   The zone to free to, must be unlocked.
1138  *	bucket The free/alloc bucket with items.
1139  *
1140  * Returns:
1141  *	Nothing
1142  */
1143 static void
1144 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
1145 {
1146 	int i;
1147 
1148 	if (bucket->ub_cnt == 0)
1149 		return;
1150 
1151 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
1152 	    bucket->ub_seq != SMR_SEQ_INVALID) {
1153 		smr_wait(zone->uz_smr, bucket->ub_seq);
1154 		bucket->ub_seq = SMR_SEQ_INVALID;
1155 		for (i = 0; i < bucket->ub_cnt; i++)
1156 			item_dtor(zone, bucket->ub_bucket[i],
1157 			    zone->uz_size, NULL, SKIP_NONE);
1158 	}
1159 	if (zone->uz_fini)
1160 		for (i = 0; i < bucket->ub_cnt; i++)
1161 			zone->uz_fini(bucket->ub_bucket[i], zone->uz_size);
1162 	zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt);
1163 	if (zone->uz_max_items > 0)
1164 		zone_free_limit(zone, bucket->ub_cnt);
1165 #ifdef INVARIANTS
1166 	bzero(bucket->ub_bucket, sizeof(void *) * bucket->ub_cnt);
1167 #endif
1168 	bucket->ub_cnt = 0;
1169 }
1170 
1171 /*
1172  * Drains the per cpu caches for a zone.
1173  *
1174  * NOTE: This may only be called while the zone is being torn down, and not
1175  * during normal operation.  This is necessary in order that we do not have
1176  * to migrate CPUs to drain the per-CPU caches.
1177  *
1178  * Arguments:
1179  *	zone     The zone to drain, must be unlocked.
1180  *
1181  * Returns:
1182  *	Nothing
1183  */
1184 static void
1185 cache_drain(uma_zone_t zone)
1186 {
1187 	uma_cache_t cache;
1188 	uma_bucket_t bucket;
1189 	smr_seq_t seq;
1190 	int cpu;
1191 
1192 	/*
1193 	 * XXX: It is safe to not lock the per-CPU caches, because we're
1194 	 * tearing down the zone anyway.  I.e., there will be no further use
1195 	 * of the caches at this point.
1196 	 *
1197 	 * XXX: It would good to be able to assert that the zone is being
1198 	 * torn down to prevent improper use of cache_drain().
1199 	 */
1200 	seq = SMR_SEQ_INVALID;
1201 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
1202 		seq = smr_advance(zone->uz_smr);
1203 	CPU_FOREACH(cpu) {
1204 		cache = &zone->uz_cpu[cpu];
1205 		bucket = cache_bucket_unload_alloc(cache);
1206 		if (bucket != NULL)
1207 			bucket_free(zone, bucket, NULL);
1208 		bucket = cache_bucket_unload_free(cache);
1209 		if (bucket != NULL) {
1210 			bucket->ub_seq = seq;
1211 			bucket_free(zone, bucket, NULL);
1212 		}
1213 		bucket = cache_bucket_unload_cross(cache);
1214 		if (bucket != NULL) {
1215 			bucket->ub_seq = seq;
1216 			bucket_free(zone, bucket, NULL);
1217 		}
1218 	}
1219 	bucket_cache_reclaim(zone, true);
1220 }
1221 
1222 static void
1223 cache_shrink(uma_zone_t zone, void *unused)
1224 {
1225 
1226 	if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1227 		return;
1228 
1229 	zone->uz_bucket_size =
1230 	    (zone->uz_bucket_size_min + zone->uz_bucket_size) / 2;
1231 }
1232 
1233 static void
1234 cache_drain_safe_cpu(uma_zone_t zone, void *unused)
1235 {
1236 	uma_cache_t cache;
1237 	uma_bucket_t b1, b2, b3;
1238 	int domain;
1239 
1240 	if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1241 		return;
1242 
1243 	b1 = b2 = b3 = NULL;
1244 	critical_enter();
1245 	cache = &zone->uz_cpu[curcpu];
1246 	domain = PCPU_GET(domain);
1247 	b1 = cache_bucket_unload_alloc(cache);
1248 
1249 	/*
1250 	 * Don't flush SMR zone buckets.  This leaves the zone without a
1251 	 * bucket and forces every free to synchronize().
1252 	 */
1253 	if ((zone->uz_flags & UMA_ZONE_SMR) == 0) {
1254 		b2 = cache_bucket_unload_free(cache);
1255 		b3 = cache_bucket_unload_cross(cache);
1256 	}
1257 	critical_exit();
1258 
1259 	if (b1 != NULL)
1260 		zone_free_bucket(zone, b1, NULL, domain, false);
1261 	if (b2 != NULL)
1262 		zone_free_bucket(zone, b2, NULL, domain, false);
1263 	if (b3 != NULL) {
1264 		/* Adjust the domain so it goes to zone_free_cross. */
1265 		domain = (domain + 1) % vm_ndomains;
1266 		zone_free_bucket(zone, b3, NULL, domain, false);
1267 	}
1268 }
1269 
1270 /*
1271  * Safely drain per-CPU caches of a zone(s) to alloc bucket.
1272  * This is an expensive call because it needs to bind to all CPUs
1273  * one by one and enter a critical section on each of them in order
1274  * to safely access their cache buckets.
1275  * Zone lock must not be held on call this function.
1276  */
1277 static void
1278 pcpu_cache_drain_safe(uma_zone_t zone)
1279 {
1280 	int cpu;
1281 
1282 	/*
1283 	 * Polite bucket sizes shrinking was not enough, shrink aggressively.
1284 	 */
1285 	if (zone)
1286 		cache_shrink(zone, NULL);
1287 	else
1288 		zone_foreach(cache_shrink, NULL);
1289 
1290 	CPU_FOREACH(cpu) {
1291 		thread_lock(curthread);
1292 		sched_bind(curthread, cpu);
1293 		thread_unlock(curthread);
1294 
1295 		if (zone)
1296 			cache_drain_safe_cpu(zone, NULL);
1297 		else
1298 			zone_foreach(cache_drain_safe_cpu, NULL);
1299 	}
1300 	thread_lock(curthread);
1301 	sched_unbind(curthread);
1302 	thread_unlock(curthread);
1303 }
1304 
1305 /*
1306  * Reclaim cached buckets from a zone.  All buckets are reclaimed if the caller
1307  * requested a drain, otherwise the per-domain caches are trimmed to either
1308  * estimated working set size.
1309  */
1310 static void
1311 bucket_cache_reclaim(uma_zone_t zone, bool drain)
1312 {
1313 	uma_zone_domain_t zdom;
1314 	uma_bucket_t bucket;
1315 	long target;
1316 	int i;
1317 
1318 	/*
1319 	 * Shrink the zone bucket size to ensure that the per-CPU caches
1320 	 * don't grow too large.
1321 	 */
1322 	if (zone->uz_bucket_size > zone->uz_bucket_size_min)
1323 		zone->uz_bucket_size--;
1324 
1325 	for (i = 0; i < vm_ndomains; i++) {
1326 		/*
1327 		 * The cross bucket is partially filled and not part of
1328 		 * the item count.  Reclaim it individually here.
1329 		 */
1330 		zdom = ZDOM_GET(zone, i);
1331 		if ((zone->uz_flags & UMA_ZONE_SMR) == 0 || drain) {
1332 			ZONE_CROSS_LOCK(zone);
1333 			bucket = zdom->uzd_cross;
1334 			zdom->uzd_cross = NULL;
1335 			ZONE_CROSS_UNLOCK(zone);
1336 			if (bucket != NULL)
1337 				bucket_free(zone, bucket, NULL);
1338 		}
1339 
1340 		/*
1341 		 * If we were asked to drain the zone, we are done only once
1342 		 * this bucket cache is empty.  Otherwise, we reclaim items in
1343 		 * excess of the zone's estimated working set size.  If the
1344 		 * difference nitems - imin is larger than the WSS estimate,
1345 		 * then the estimate will grow at the end of this interval and
1346 		 * we ignore the historical average.
1347 		 */
1348 		ZDOM_LOCK(zdom);
1349 		target = drain ? 0 : lmax(zdom->uzd_wss, zdom->uzd_nitems -
1350 		    zdom->uzd_imin);
1351 		while (zdom->uzd_nitems > target) {
1352 			bucket = zone_fetch_bucket(zone, zdom, true);
1353 			if (bucket == NULL)
1354 				break;
1355 			bucket_free(zone, bucket, NULL);
1356 			ZDOM_LOCK(zdom);
1357 		}
1358 		ZDOM_UNLOCK(zdom);
1359 	}
1360 }
1361 
1362 static void
1363 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start)
1364 {
1365 	uint8_t *mem;
1366 	int i;
1367 	uint8_t flags;
1368 
1369 	CTR4(KTR_UMA, "keg_free_slab keg %s(%p) slab %p, returning %d bytes",
1370 	    keg->uk_name, keg, slab, PAGE_SIZE * keg->uk_ppera);
1371 
1372 	mem = slab_data(slab, keg);
1373 	flags = slab->us_flags;
1374 	i = start;
1375 	if (keg->uk_fini != NULL) {
1376 		for (i--; i > -1; i--)
1377 #ifdef INVARIANTS
1378 		/*
1379 		 * trash_fini implies that dtor was trash_dtor. trash_fini
1380 		 * would check that memory hasn't been modified since free,
1381 		 * which executed trash_dtor.
1382 		 * That's why we need to run uma_dbg_kskip() check here,
1383 		 * albeit we don't make skip check for other init/fini
1384 		 * invocations.
1385 		 */
1386 		if (!uma_dbg_kskip(keg, slab_item(slab, keg, i)) ||
1387 		    keg->uk_fini != trash_fini)
1388 #endif
1389 			keg->uk_fini(slab_item(slab, keg, i), keg->uk_size);
1390 	}
1391 	if (keg->uk_flags & UMA_ZFLAG_OFFPAGE)
1392 		zone_free_item(slabzone(keg->uk_ipers), slab_tohashslab(slab),
1393 		    NULL, SKIP_NONE);
1394 	keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags);
1395 	uma_total_dec(PAGE_SIZE * keg->uk_ppera);
1396 }
1397 
1398 /*
1399  * Frees pages from a keg back to the system.  This is done on demand from
1400  * the pageout daemon.
1401  *
1402  * Returns nothing.
1403  */
1404 static void
1405 keg_drain(uma_keg_t keg)
1406 {
1407 	struct slabhead freeslabs;
1408 	uma_domain_t dom;
1409 	uma_slab_t slab, tmp;
1410 	int i, n;
1411 
1412 	if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL)
1413 		return;
1414 
1415 	for (i = 0; i < vm_ndomains; i++) {
1416 		CTR4(KTR_UMA, "keg_drain %s(%p) domain %d free items: %u",
1417 		    keg->uk_name, keg, i, dom->ud_free_items);
1418 		dom = &keg->uk_domain[i];
1419 		LIST_INIT(&freeslabs);
1420 
1421 		KEG_LOCK(keg, i);
1422 		if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0) {
1423 			LIST_FOREACH(slab, &dom->ud_free_slab, us_link)
1424 				UMA_HASH_REMOVE(&keg->uk_hash, slab);
1425 		}
1426 		n = dom->ud_free_slabs;
1427 		LIST_SWAP(&freeslabs, &dom->ud_free_slab, uma_slab, us_link);
1428 		dom->ud_free_slabs = 0;
1429 		dom->ud_free_items -= n * keg->uk_ipers;
1430 		dom->ud_pages -= n * keg->uk_ppera;
1431 		KEG_UNLOCK(keg, i);
1432 
1433 		LIST_FOREACH_SAFE(slab, &freeslabs, us_link, tmp)
1434 			keg_free_slab(keg, slab, keg->uk_ipers);
1435 	}
1436 }
1437 
1438 static void
1439 zone_reclaim(uma_zone_t zone, int waitok, bool drain)
1440 {
1441 
1442 	/*
1443 	 * Set draining to interlock with zone_dtor() so we can release our
1444 	 * locks as we go.  Only dtor() should do a WAITOK call since it
1445 	 * is the only call that knows the structure will still be available
1446 	 * when it wakes up.
1447 	 */
1448 	ZONE_LOCK(zone);
1449 	while (zone->uz_flags & UMA_ZFLAG_RECLAIMING) {
1450 		if (waitok == M_NOWAIT)
1451 			goto out;
1452 		msleep(zone, &ZDOM_GET(zone, 0)->uzd_lock, PVM, "zonedrain",
1453 		    1);
1454 	}
1455 	zone->uz_flags |= UMA_ZFLAG_RECLAIMING;
1456 	ZONE_UNLOCK(zone);
1457 	bucket_cache_reclaim(zone, drain);
1458 
1459 	/*
1460 	 * The DRAINING flag protects us from being freed while
1461 	 * we're running.  Normally the uma_rwlock would protect us but we
1462 	 * must be able to release and acquire the right lock for each keg.
1463 	 */
1464 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0)
1465 		keg_drain(zone->uz_keg);
1466 	ZONE_LOCK(zone);
1467 	zone->uz_flags &= ~UMA_ZFLAG_RECLAIMING;
1468 	wakeup(zone);
1469 out:
1470 	ZONE_UNLOCK(zone);
1471 }
1472 
1473 static void
1474 zone_drain(uma_zone_t zone, void *unused)
1475 {
1476 
1477 	zone_reclaim(zone, M_NOWAIT, true);
1478 }
1479 
1480 static void
1481 zone_trim(uma_zone_t zone, void *unused)
1482 {
1483 
1484 	zone_reclaim(zone, M_NOWAIT, false);
1485 }
1486 
1487 /*
1488  * Allocate a new slab for a keg and inserts it into the partial slab list.
1489  * The keg should be unlocked on entry.  If the allocation succeeds it will
1490  * be locked on return.
1491  *
1492  * Arguments:
1493  *	flags   Wait flags for the item initialization routine
1494  *	aflags  Wait flags for the slab allocation
1495  *
1496  * Returns:
1497  *	The slab that was allocated or NULL if there is no memory and the
1498  *	caller specified M_NOWAIT.
1499  */
1500 static uma_slab_t
1501 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int domain, int flags,
1502     int aflags)
1503 {
1504 	uma_domain_t dom;
1505 	uma_alloc allocf;
1506 	uma_slab_t slab;
1507 	unsigned long size;
1508 	uint8_t *mem;
1509 	uint8_t sflags;
1510 	int i;
1511 
1512 	KASSERT(domain >= 0 && domain < vm_ndomains,
1513 	    ("keg_alloc_slab: domain %d out of range", domain));
1514 
1515 	allocf = keg->uk_allocf;
1516 	slab = NULL;
1517 	mem = NULL;
1518 	if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) {
1519 		uma_hash_slab_t hslab;
1520 		hslab = zone_alloc_item(slabzone(keg->uk_ipers), NULL,
1521 		    domain, aflags);
1522 		if (hslab == NULL)
1523 			goto fail;
1524 		slab = &hslab->uhs_slab;
1525 	}
1526 
1527 	/*
1528 	 * This reproduces the old vm_zone behavior of zero filling pages the
1529 	 * first time they are added to a zone.
1530 	 *
1531 	 * Malloced items are zeroed in uma_zalloc.
1532 	 */
1533 
1534 	if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
1535 		aflags |= M_ZERO;
1536 	else
1537 		aflags &= ~M_ZERO;
1538 
1539 	if (keg->uk_flags & UMA_ZONE_NODUMP)
1540 		aflags |= M_NODUMP;
1541 
1542 	/* zone is passed for legacy reasons. */
1543 	size = keg->uk_ppera * PAGE_SIZE;
1544 	mem = allocf(zone, size, domain, &sflags, aflags);
1545 	if (mem == NULL) {
1546 		if (keg->uk_flags & UMA_ZFLAG_OFFPAGE)
1547 			zone_free_item(slabzone(keg->uk_ipers),
1548 			    slab_tohashslab(slab), NULL, SKIP_NONE);
1549 		goto fail;
1550 	}
1551 	uma_total_inc(size);
1552 
1553 	/* For HASH zones all pages go to the same uma_domain. */
1554 	if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
1555 		domain = 0;
1556 
1557 	/* Point the slab into the allocated memory */
1558 	if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE))
1559 		slab = (uma_slab_t )(mem + keg->uk_pgoff);
1560 	else
1561 		slab_tohashslab(slab)->uhs_data = mem;
1562 
1563 	if (keg->uk_flags & UMA_ZFLAG_VTOSLAB)
1564 		for (i = 0; i < keg->uk_ppera; i++)
1565 			vsetzoneslab((vm_offset_t)mem + (i * PAGE_SIZE),
1566 			    zone, slab);
1567 
1568 	slab->us_freecount = keg->uk_ipers;
1569 	slab->us_flags = sflags;
1570 	slab->us_domain = domain;
1571 
1572 	BIT_FILL(keg->uk_ipers, &slab->us_free);
1573 #ifdef INVARIANTS
1574 	BIT_ZERO(keg->uk_ipers, slab_dbg_bits(slab, keg));
1575 #endif
1576 
1577 	if (keg->uk_init != NULL) {
1578 		for (i = 0; i < keg->uk_ipers; i++)
1579 			if (keg->uk_init(slab_item(slab, keg, i),
1580 			    keg->uk_size, flags) != 0)
1581 				break;
1582 		if (i != keg->uk_ipers) {
1583 			keg_free_slab(keg, slab, i);
1584 			goto fail;
1585 		}
1586 	}
1587 	KEG_LOCK(keg, domain);
1588 
1589 	CTR3(KTR_UMA, "keg_alloc_slab: allocated slab %p for %s(%p)",
1590 	    slab, keg->uk_name, keg);
1591 
1592 	if (keg->uk_flags & UMA_ZFLAG_HASH)
1593 		UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
1594 
1595 	/*
1596 	 * If we got a slab here it's safe to mark it partially used
1597 	 * and return.  We assume that the caller is going to remove
1598 	 * at least one item.
1599 	 */
1600 	dom = &keg->uk_domain[domain];
1601 	LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
1602 	dom->ud_pages += keg->uk_ppera;
1603 	dom->ud_free_items += keg->uk_ipers;
1604 
1605 	return (slab);
1606 
1607 fail:
1608 	return (NULL);
1609 }
1610 
1611 /*
1612  * This function is intended to be used early on in place of page_alloc() so
1613  * that we may use the boot time page cache to satisfy allocations before
1614  * the VM is ready.
1615  */
1616 static void *
1617 startup_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1618     int wait)
1619 {
1620 	vm_paddr_t pa;
1621 	vm_page_t m;
1622 	void *mem;
1623 	int pages;
1624 	int i;
1625 
1626 	pages = howmany(bytes, PAGE_SIZE);
1627 	KASSERT(pages > 0, ("%s can't reserve 0 pages", __func__));
1628 
1629 	*pflag = UMA_SLAB_BOOT;
1630 	m = vm_page_alloc_contig_domain(NULL, 0, domain,
1631 	    malloc2vm_flags(wait) | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED, pages,
1632 	    (vm_paddr_t)0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT);
1633 	if (m == NULL)
1634 		return (NULL);
1635 
1636 	pa = VM_PAGE_TO_PHYS(m);
1637 	for (i = 0; i < pages; i++, pa += PAGE_SIZE) {
1638 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
1639     defined(__riscv) || defined(__powerpc64__)
1640 		if ((wait & M_NODUMP) == 0)
1641 			dump_add_page(pa);
1642 #endif
1643 	}
1644 	/* Allocate KVA and indirectly advance bootmem. */
1645 	mem = (void *)pmap_map(&bootmem, m->phys_addr,
1646 	    m->phys_addr + (pages * PAGE_SIZE), VM_PROT_READ | VM_PROT_WRITE);
1647         if ((wait & M_ZERO) != 0)
1648                 bzero(mem, pages * PAGE_SIZE);
1649 
1650         return (mem);
1651 }
1652 
1653 static void
1654 startup_free(void *mem, vm_size_t bytes)
1655 {
1656 	vm_offset_t va;
1657 	vm_page_t m;
1658 
1659 	va = (vm_offset_t)mem;
1660 	m = PHYS_TO_VM_PAGE(pmap_kextract(va));
1661 	pmap_remove(kernel_pmap, va, va + bytes);
1662 	for (; bytes != 0; bytes -= PAGE_SIZE, m++) {
1663 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
1664     defined(__riscv) || defined(__powerpc64__)
1665 		dump_drop_page(VM_PAGE_TO_PHYS(m));
1666 #endif
1667 		vm_page_unwire_noq(m);
1668 		vm_page_free(m);
1669 	}
1670 }
1671 
1672 /*
1673  * Allocates a number of pages from the system
1674  *
1675  * Arguments:
1676  *	bytes  The number of bytes requested
1677  *	wait  Shall we wait?
1678  *
1679  * Returns:
1680  *	A pointer to the alloced memory or possibly
1681  *	NULL if M_NOWAIT is set.
1682  */
1683 static void *
1684 page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1685     int wait)
1686 {
1687 	void *p;	/* Returned page */
1688 
1689 	*pflag = UMA_SLAB_KERNEL;
1690 	p = (void *)kmem_malloc_domainset(DOMAINSET_FIXED(domain), bytes, wait);
1691 
1692 	return (p);
1693 }
1694 
1695 static void *
1696 pcpu_page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1697     int wait)
1698 {
1699 	struct pglist alloctail;
1700 	vm_offset_t addr, zkva;
1701 	int cpu, flags;
1702 	vm_page_t p, p_next;
1703 #ifdef NUMA
1704 	struct pcpu *pc;
1705 #endif
1706 
1707 	MPASS(bytes == (mp_maxid + 1) * PAGE_SIZE);
1708 
1709 	TAILQ_INIT(&alloctail);
1710 	flags = VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | VM_ALLOC_NOOBJ |
1711 	    malloc2vm_flags(wait);
1712 	*pflag = UMA_SLAB_KERNEL;
1713 	for (cpu = 0; cpu <= mp_maxid; cpu++) {
1714 		if (CPU_ABSENT(cpu)) {
1715 			p = vm_page_alloc(NULL, 0, flags);
1716 		} else {
1717 #ifndef NUMA
1718 			p = vm_page_alloc(NULL, 0, flags);
1719 #else
1720 			pc = pcpu_find(cpu);
1721 			if (__predict_false(VM_DOMAIN_EMPTY(pc->pc_domain)))
1722 				p = NULL;
1723 			else
1724 				p = vm_page_alloc_domain(NULL, 0,
1725 				    pc->pc_domain, flags);
1726 			if (__predict_false(p == NULL))
1727 				p = vm_page_alloc(NULL, 0, flags);
1728 #endif
1729 		}
1730 		if (__predict_false(p == NULL))
1731 			goto fail;
1732 		TAILQ_INSERT_TAIL(&alloctail, p, listq);
1733 	}
1734 	if ((addr = kva_alloc(bytes)) == 0)
1735 		goto fail;
1736 	zkva = addr;
1737 	TAILQ_FOREACH(p, &alloctail, listq) {
1738 		pmap_qenter(zkva, &p, 1);
1739 		zkva += PAGE_SIZE;
1740 	}
1741 	return ((void*)addr);
1742 fail:
1743 	TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1744 		vm_page_unwire_noq(p);
1745 		vm_page_free(p);
1746 	}
1747 	return (NULL);
1748 }
1749 
1750 /*
1751  * Allocates a number of pages from within an object
1752  *
1753  * Arguments:
1754  *	bytes  The number of bytes requested
1755  *	wait   Shall we wait?
1756  *
1757  * Returns:
1758  *	A pointer to the alloced memory or possibly
1759  *	NULL if M_NOWAIT is set.
1760  */
1761 static void *
1762 noobj_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags,
1763     int wait)
1764 {
1765 	TAILQ_HEAD(, vm_page) alloctail;
1766 	u_long npages;
1767 	vm_offset_t retkva, zkva;
1768 	vm_page_t p, p_next;
1769 	uma_keg_t keg;
1770 
1771 	TAILQ_INIT(&alloctail);
1772 	keg = zone->uz_keg;
1773 
1774 	npages = howmany(bytes, PAGE_SIZE);
1775 	while (npages > 0) {
1776 		p = vm_page_alloc_domain(NULL, 0, domain, VM_ALLOC_INTERRUPT |
1777 		    VM_ALLOC_WIRED | VM_ALLOC_NOOBJ |
1778 		    ((wait & M_WAITOK) != 0 ? VM_ALLOC_WAITOK :
1779 		    VM_ALLOC_NOWAIT));
1780 		if (p != NULL) {
1781 			/*
1782 			 * Since the page does not belong to an object, its
1783 			 * listq is unused.
1784 			 */
1785 			TAILQ_INSERT_TAIL(&alloctail, p, listq);
1786 			npages--;
1787 			continue;
1788 		}
1789 		/*
1790 		 * Page allocation failed, free intermediate pages and
1791 		 * exit.
1792 		 */
1793 		TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1794 			vm_page_unwire_noq(p);
1795 			vm_page_free(p);
1796 		}
1797 		return (NULL);
1798 	}
1799 	*flags = UMA_SLAB_PRIV;
1800 	zkva = keg->uk_kva +
1801 	    atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
1802 	retkva = zkva;
1803 	TAILQ_FOREACH(p, &alloctail, listq) {
1804 		pmap_qenter(zkva, &p, 1);
1805 		zkva += PAGE_SIZE;
1806 	}
1807 
1808 	return ((void *)retkva);
1809 }
1810 
1811 /*
1812  * Allocate physically contiguous pages.
1813  */
1814 static void *
1815 contig_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1816     int wait)
1817 {
1818 
1819 	*pflag = UMA_SLAB_KERNEL;
1820 	return ((void *)kmem_alloc_contig_domainset(DOMAINSET_FIXED(domain),
1821 	    bytes, wait, 0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT));
1822 }
1823 
1824 /*
1825  * Frees a number of pages to the system
1826  *
1827  * Arguments:
1828  *	mem   A pointer to the memory to be freed
1829  *	size  The size of the memory being freed
1830  *	flags The original p->us_flags field
1831  *
1832  * Returns:
1833  *	Nothing
1834  */
1835 static void
1836 page_free(void *mem, vm_size_t size, uint8_t flags)
1837 {
1838 
1839 	if ((flags & UMA_SLAB_BOOT) != 0) {
1840 		startup_free(mem, size);
1841 		return;
1842 	}
1843 
1844 	KASSERT((flags & UMA_SLAB_KERNEL) != 0,
1845 	    ("UMA: page_free used with invalid flags %x", flags));
1846 
1847 	kmem_free((vm_offset_t)mem, size);
1848 }
1849 
1850 /*
1851  * Frees pcpu zone allocations
1852  *
1853  * Arguments:
1854  *	mem   A pointer to the memory to be freed
1855  *	size  The size of the memory being freed
1856  *	flags The original p->us_flags field
1857  *
1858  * Returns:
1859  *	Nothing
1860  */
1861 static void
1862 pcpu_page_free(void *mem, vm_size_t size, uint8_t flags)
1863 {
1864 	vm_offset_t sva, curva;
1865 	vm_paddr_t paddr;
1866 	vm_page_t m;
1867 
1868 	MPASS(size == (mp_maxid+1)*PAGE_SIZE);
1869 
1870 	if ((flags & UMA_SLAB_BOOT) != 0) {
1871 		startup_free(mem, size);
1872 		return;
1873 	}
1874 
1875 	sva = (vm_offset_t)mem;
1876 	for (curva = sva; curva < sva + size; curva += PAGE_SIZE) {
1877 		paddr = pmap_kextract(curva);
1878 		m = PHYS_TO_VM_PAGE(paddr);
1879 		vm_page_unwire_noq(m);
1880 		vm_page_free(m);
1881 	}
1882 	pmap_qremove(sva, size >> PAGE_SHIFT);
1883 	kva_free(sva, size);
1884 }
1885 
1886 
1887 /*
1888  * Zero fill initializer
1889  *
1890  * Arguments/Returns follow uma_init specifications
1891  */
1892 static int
1893 zero_init(void *mem, int size, int flags)
1894 {
1895 	bzero(mem, size);
1896 	return (0);
1897 }
1898 
1899 #ifdef INVARIANTS
1900 struct noslabbits *
1901 slab_dbg_bits(uma_slab_t slab, uma_keg_t keg)
1902 {
1903 
1904 	return ((void *)((char *)&slab->us_free + BITSET_SIZE(keg->uk_ipers)));
1905 }
1906 #endif
1907 
1908 /*
1909  * Actual size of embedded struct slab (!OFFPAGE).
1910  */
1911 size_t
1912 slab_sizeof(int nitems)
1913 {
1914 	size_t s;
1915 
1916 	s = sizeof(struct uma_slab) + BITSET_SIZE(nitems) * SLAB_BITSETS;
1917 	return (roundup(s, UMA_ALIGN_PTR + 1));
1918 }
1919 
1920 /*
1921  * Size of memory for embedded slabs (!OFFPAGE).
1922  */
1923 size_t
1924 slab_space(int nitems)
1925 {
1926 	return (UMA_SLAB_SIZE - slab_sizeof(nitems));
1927 }
1928 
1929 #define	UMA_FIXPT_SHIFT	31
1930 #define	UMA_FRAC_FIXPT(n, d)						\
1931 	((uint32_t)(((uint64_t)(n) << UMA_FIXPT_SHIFT) / (d)))
1932 #define	UMA_FIXPT_PCT(f)						\
1933 	((u_int)(((uint64_t)100 * (f)) >> UMA_FIXPT_SHIFT))
1934 #define	UMA_PCT_FIXPT(pct)	UMA_FRAC_FIXPT((pct), 100)
1935 #define	UMA_MIN_EFF	UMA_PCT_FIXPT(100 - UMA_MAX_WASTE)
1936 
1937 /*
1938  * Compute the number of items that will fit in a slab.  If hdr is true, the
1939  * item count may be limited to provide space in the slab for an inline slab
1940  * header.  Otherwise, all slab space will be provided for item storage.
1941  */
1942 static u_int
1943 slab_ipers_hdr(u_int size, u_int rsize, u_int slabsize, bool hdr)
1944 {
1945 	u_int ipers;
1946 	u_int padpi;
1947 
1948 	/* The padding between items is not needed after the last item. */
1949 	padpi = rsize - size;
1950 
1951 	if (hdr) {
1952 		/*
1953 		 * Start with the maximum item count and remove items until
1954 		 * the slab header first alongside the allocatable memory.
1955 		 */
1956 		for (ipers = MIN(SLAB_MAX_SETSIZE,
1957 		    (slabsize + padpi - slab_sizeof(1)) / rsize);
1958 		    ipers > 0 &&
1959 		    ipers * rsize - padpi + slab_sizeof(ipers) > slabsize;
1960 		    ipers--)
1961 			continue;
1962 	} else {
1963 		ipers = MIN((slabsize + padpi) / rsize, SLAB_MAX_SETSIZE);
1964 	}
1965 
1966 	return (ipers);
1967 }
1968 
1969 /*
1970  * Compute the number of items that will fit in a slab for a startup zone.
1971  */
1972 int
1973 slab_ipers(size_t size, int align)
1974 {
1975 	int rsize;
1976 
1977 	rsize = roundup(size, align + 1); /* Assume no CACHESPREAD */
1978 	return (slab_ipers_hdr(size, rsize, UMA_SLAB_SIZE, true));
1979 }
1980 
1981 struct keg_layout_result {
1982 	u_int format;
1983 	u_int slabsize;
1984 	u_int ipers;
1985 	u_int eff;
1986 };
1987 
1988 static void
1989 keg_layout_one(uma_keg_t keg, u_int rsize, u_int slabsize, u_int fmt,
1990     struct keg_layout_result *kl)
1991 {
1992 	u_int total;
1993 
1994 	kl->format = fmt;
1995 	kl->slabsize = slabsize;
1996 
1997 	/* Handle INTERNAL as inline with an extra page. */
1998 	if ((fmt & UMA_ZFLAG_INTERNAL) != 0) {
1999 		kl->format &= ~UMA_ZFLAG_INTERNAL;
2000 		kl->slabsize += PAGE_SIZE;
2001 	}
2002 
2003 	kl->ipers = slab_ipers_hdr(keg->uk_size, rsize, kl->slabsize,
2004 	    (fmt & UMA_ZFLAG_OFFPAGE) == 0);
2005 
2006 	/* Account for memory used by an offpage slab header. */
2007 	total = kl->slabsize;
2008 	if ((fmt & UMA_ZFLAG_OFFPAGE) != 0)
2009 		total += slabzone(kl->ipers)->uz_keg->uk_rsize;
2010 
2011 	kl->eff = UMA_FRAC_FIXPT(kl->ipers * rsize, total);
2012 }
2013 
2014 /*
2015  * Determine the format of a uma keg.  This determines where the slab header
2016  * will be placed (inline or offpage) and calculates ipers, rsize, and ppera.
2017  *
2018  * Arguments
2019  *	keg  The zone we should initialize
2020  *
2021  * Returns
2022  *	Nothing
2023  */
2024 static void
2025 keg_layout(uma_keg_t keg)
2026 {
2027 	struct keg_layout_result kl = {}, kl_tmp;
2028 	u_int fmts[2];
2029 	u_int alignsize;
2030 	u_int nfmt;
2031 	u_int pages;
2032 	u_int rsize;
2033 	u_int slabsize;
2034 	u_int i, j;
2035 
2036 	KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
2037 	    (keg->uk_size <= UMA_PCPU_ALLOC_SIZE &&
2038 	     (keg->uk_flags & UMA_ZONE_CACHESPREAD) == 0),
2039 	    ("%s: cannot configure for PCPU: keg=%s, size=%u, flags=0x%b",
2040 	     __func__, keg->uk_name, keg->uk_size, keg->uk_flags,
2041 	     PRINT_UMA_ZFLAGS));
2042 	KASSERT((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) == 0 ||
2043 	    (keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0,
2044 	    ("%s: incompatible flags 0x%b", __func__, keg->uk_flags,
2045 	     PRINT_UMA_ZFLAGS));
2046 
2047 	alignsize = keg->uk_align + 1;
2048 
2049 	/*
2050 	 * Calculate the size of each allocation (rsize) according to
2051 	 * alignment.  If the requested size is smaller than we have
2052 	 * allocation bits for we round it up.
2053 	 */
2054 	rsize = MAX(keg->uk_size, UMA_SMALLEST_UNIT);
2055 	rsize = roundup2(rsize, alignsize);
2056 
2057 	if ((keg->uk_flags & UMA_ZONE_CACHESPREAD) != 0) {
2058 		/*
2059 		 * We want one item to start on every align boundary in a page.
2060 		 * To do this we will span pages.  We will also extend the item
2061 		 * by the size of align if it is an even multiple of align.
2062 		 * Otherwise, it would fall on the same boundary every time.
2063 		 */
2064 		if ((rsize & alignsize) == 0)
2065 			rsize += alignsize;
2066 		slabsize = rsize * (PAGE_SIZE / alignsize);
2067 		slabsize = MIN(slabsize, rsize * SLAB_MAX_SETSIZE);
2068 		slabsize = MIN(slabsize, UMA_CACHESPREAD_MAX_SIZE);
2069 		slabsize = round_page(slabsize);
2070 	} else {
2071 		/*
2072 		 * Start with a slab size of as many pages as it takes to
2073 		 * represent a single item.  We will try to fit as many
2074 		 * additional items into the slab as possible.
2075 		 */
2076 		slabsize = round_page(keg->uk_size);
2077 	}
2078 
2079 	/* Build a list of all of the available formats for this keg. */
2080 	nfmt = 0;
2081 
2082 	/* Evaluate an inline slab layout. */
2083 	if ((keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0)
2084 		fmts[nfmt++] = 0;
2085 
2086 	/* TODO: vm_page-embedded slab. */
2087 
2088 	/*
2089 	 * We can't do OFFPAGE if we're internal or if we've been
2090 	 * asked to not go to the VM for buckets.  If we do this we
2091 	 * may end up going to the VM for slabs which we do not want
2092 	 * to do if we're UMA_ZONE_VM, which clearly forbids it.
2093 	 * In those cases, evaluate a pseudo-format called INTERNAL
2094 	 * which has an inline slab header and one extra page to
2095 	 * guarantee that it fits.
2096 	 *
2097 	 * Otherwise, see if using an OFFPAGE slab will improve our
2098 	 * efficiency.
2099 	 */
2100 	if ((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) != 0)
2101 		fmts[nfmt++] = UMA_ZFLAG_INTERNAL;
2102 	else
2103 		fmts[nfmt++] = UMA_ZFLAG_OFFPAGE;
2104 
2105 	/*
2106 	 * Choose a slab size and format which satisfy the minimum efficiency.
2107 	 * Prefer the smallest slab size that meets the constraints.
2108 	 *
2109 	 * Start with a minimum slab size, to accommodate CACHESPREAD.  Then,
2110 	 * for small items (up to PAGE_SIZE), the iteration increment is one
2111 	 * page; and for large items, the increment is one item.
2112 	 */
2113 	i = (slabsize + rsize - keg->uk_size) / MAX(PAGE_SIZE, rsize);
2114 	KASSERT(i >= 1, ("keg %s(%p) flags=0x%b slabsize=%u, rsize=%u, i=%u",
2115 	    keg->uk_name, keg, keg->uk_flags, PRINT_UMA_ZFLAGS, slabsize,
2116 	    rsize, i));
2117 	for ( ; ; i++) {
2118 		slabsize = (rsize <= PAGE_SIZE) ? ptoa(i) :
2119 		    round_page(rsize * (i - 1) + keg->uk_size);
2120 
2121 		for (j = 0; j < nfmt; j++) {
2122 			/* Only if we have no viable format yet. */
2123 			if ((fmts[j] & UMA_ZFLAG_INTERNAL) != 0 &&
2124 			    kl.ipers > 0)
2125 				continue;
2126 
2127 			keg_layout_one(keg, rsize, slabsize, fmts[j], &kl_tmp);
2128 			if (kl_tmp.eff <= kl.eff)
2129 				continue;
2130 
2131 			kl = kl_tmp;
2132 
2133 			CTR6(KTR_UMA, "keg %s layout: format %#x "
2134 			    "(ipers %u * rsize %u) / slabsize %#x = %u%% eff",
2135 			    keg->uk_name, kl.format, kl.ipers, rsize,
2136 			    kl.slabsize, UMA_FIXPT_PCT(kl.eff));
2137 
2138 			/* Stop when we reach the minimum efficiency. */
2139 			if (kl.eff >= UMA_MIN_EFF)
2140 				break;
2141 		}
2142 
2143 		if (kl.eff >= UMA_MIN_EFF || !multipage_slabs ||
2144 		    slabsize >= SLAB_MAX_SETSIZE * rsize ||
2145 		    (keg->uk_flags & (UMA_ZONE_PCPU | UMA_ZONE_CONTIG)) != 0)
2146 			break;
2147 	}
2148 
2149 	pages = atop(kl.slabsize);
2150 	if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
2151 		pages *= mp_maxid + 1;
2152 
2153 	keg->uk_rsize = rsize;
2154 	keg->uk_ipers = kl.ipers;
2155 	keg->uk_ppera = pages;
2156 	keg->uk_flags |= kl.format;
2157 
2158 	/*
2159 	 * How do we find the slab header if it is offpage or if not all item
2160 	 * start addresses are in the same page?  We could solve the latter
2161 	 * case with vaddr alignment, but we don't.
2162 	 */
2163 	if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0 ||
2164 	    (keg->uk_ipers - 1) * rsize >= PAGE_SIZE) {
2165 		if ((keg->uk_flags & UMA_ZONE_NOTPAGE) != 0)
2166 			keg->uk_flags |= UMA_ZFLAG_HASH;
2167 		else
2168 			keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2169 	}
2170 
2171 	CTR6(KTR_UMA, "%s: keg=%s, flags=%#x, rsize=%u, ipers=%u, ppera=%u",
2172 	    __func__, keg->uk_name, keg->uk_flags, rsize, keg->uk_ipers,
2173 	    pages);
2174 	KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_MAX_SETSIZE,
2175 	    ("%s: keg=%s, flags=0x%b, rsize=%u, ipers=%u, ppera=%u", __func__,
2176 	     keg->uk_name, keg->uk_flags, PRINT_UMA_ZFLAGS, rsize,
2177 	     keg->uk_ipers, pages));
2178 }
2179 
2180 /*
2181  * Keg header ctor.  This initializes all fields, locks, etc.  And inserts
2182  * the keg onto the global keg list.
2183  *
2184  * Arguments/Returns follow uma_ctor specifications
2185  *	udata  Actually uma_kctor_args
2186  */
2187 static int
2188 keg_ctor(void *mem, int size, void *udata, int flags)
2189 {
2190 	struct uma_kctor_args *arg = udata;
2191 	uma_keg_t keg = mem;
2192 	uma_zone_t zone;
2193 	int i;
2194 
2195 	bzero(keg, size);
2196 	keg->uk_size = arg->size;
2197 	keg->uk_init = arg->uminit;
2198 	keg->uk_fini = arg->fini;
2199 	keg->uk_align = arg->align;
2200 	keg->uk_reserve = 0;
2201 	keg->uk_flags = arg->flags;
2202 
2203 	/*
2204 	 * We use a global round-robin policy by default.  Zones with
2205 	 * UMA_ZONE_FIRSTTOUCH set will use first-touch instead, in which
2206 	 * case the iterator is never run.
2207 	 */
2208 	keg->uk_dr.dr_policy = DOMAINSET_RR();
2209 	keg->uk_dr.dr_iter = 0;
2210 
2211 	/*
2212 	 * The master zone is passed to us at keg-creation time.
2213 	 */
2214 	zone = arg->zone;
2215 	keg->uk_name = zone->uz_name;
2216 
2217 	if (arg->flags & UMA_ZONE_ZINIT)
2218 		keg->uk_init = zero_init;
2219 
2220 	if (arg->flags & UMA_ZONE_MALLOC)
2221 		keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2222 
2223 #ifndef SMP
2224 	keg->uk_flags &= ~UMA_ZONE_PCPU;
2225 #endif
2226 
2227 	keg_layout(keg);
2228 
2229 	/*
2230 	 * Use a first-touch NUMA policy for kegs that pmap_extract() will
2231 	 * work on.  Use round-robin for everything else.
2232 	 *
2233 	 * Zones may override the default by specifying either.
2234 	 */
2235 #ifdef NUMA
2236 	if ((keg->uk_flags &
2237 	    (UMA_ZONE_ROUNDROBIN | UMA_ZFLAG_CACHE | UMA_ZONE_NOTPAGE)) == 0)
2238 		keg->uk_flags |= UMA_ZONE_FIRSTTOUCH;
2239 	else if ((keg->uk_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2240 		keg->uk_flags |= UMA_ZONE_ROUNDROBIN;
2241 #endif
2242 
2243 	/*
2244 	 * If we haven't booted yet we need allocations to go through the
2245 	 * startup cache until the vm is ready.
2246 	 */
2247 #ifdef UMA_MD_SMALL_ALLOC
2248 	if (keg->uk_ppera == 1)
2249 		keg->uk_allocf = uma_small_alloc;
2250 	else
2251 #endif
2252 	if (booted < BOOT_KVA)
2253 		keg->uk_allocf = startup_alloc;
2254 	else if (keg->uk_flags & UMA_ZONE_PCPU)
2255 		keg->uk_allocf = pcpu_page_alloc;
2256 	else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 && keg->uk_ppera > 1)
2257 		keg->uk_allocf = contig_alloc;
2258 	else
2259 		keg->uk_allocf = page_alloc;
2260 #ifdef UMA_MD_SMALL_ALLOC
2261 	if (keg->uk_ppera == 1)
2262 		keg->uk_freef = uma_small_free;
2263 	else
2264 #endif
2265 	if (keg->uk_flags & UMA_ZONE_PCPU)
2266 		keg->uk_freef = pcpu_page_free;
2267 	else
2268 		keg->uk_freef = page_free;
2269 
2270 	/*
2271 	 * Initialize keg's locks.
2272 	 */
2273 	for (i = 0; i < vm_ndomains; i++)
2274 		KEG_LOCK_INIT(keg, i, (arg->flags & UMA_ZONE_MTXCLASS));
2275 
2276 	/*
2277 	 * If we're putting the slab header in the actual page we need to
2278 	 * figure out where in each page it goes.  See slab_sizeof
2279 	 * definition.
2280 	 */
2281 	if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE)) {
2282 		size_t shsize;
2283 
2284 		shsize = slab_sizeof(keg->uk_ipers);
2285 		keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - shsize;
2286 		/*
2287 		 * The only way the following is possible is if with our
2288 		 * UMA_ALIGN_PTR adjustments we are now bigger than
2289 		 * UMA_SLAB_SIZE.  I haven't checked whether this is
2290 		 * mathematically possible for all cases, so we make
2291 		 * sure here anyway.
2292 		 */
2293 		KASSERT(keg->uk_pgoff + shsize <= PAGE_SIZE * keg->uk_ppera,
2294 		    ("zone %s ipers %d rsize %d size %d slab won't fit",
2295 		    zone->uz_name, keg->uk_ipers, keg->uk_rsize, keg->uk_size));
2296 	}
2297 
2298 	if (keg->uk_flags & UMA_ZFLAG_HASH)
2299 		hash_alloc(&keg->uk_hash, 0);
2300 
2301 	CTR3(KTR_UMA, "keg_ctor %p zone %s(%p)", keg, zone->uz_name, zone);
2302 
2303 	LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
2304 
2305 	rw_wlock(&uma_rwlock);
2306 	LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
2307 	rw_wunlock(&uma_rwlock);
2308 	return (0);
2309 }
2310 
2311 static void
2312 zone_kva_available(uma_zone_t zone, void *unused)
2313 {
2314 	uma_keg_t keg;
2315 
2316 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
2317 		return;
2318 	KEG_GET(zone, keg);
2319 
2320 	if (keg->uk_allocf == startup_alloc) {
2321 		/* Switch to the real allocator. */
2322 		if (keg->uk_flags & UMA_ZONE_PCPU)
2323 			keg->uk_allocf = pcpu_page_alloc;
2324 		else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 &&
2325 		    keg->uk_ppera > 1)
2326 			keg->uk_allocf = contig_alloc;
2327 		else
2328 			keg->uk_allocf = page_alloc;
2329 	}
2330 }
2331 
2332 static void
2333 zone_alloc_counters(uma_zone_t zone, void *unused)
2334 {
2335 
2336 	zone->uz_allocs = counter_u64_alloc(M_WAITOK);
2337 	zone->uz_frees = counter_u64_alloc(M_WAITOK);
2338 	zone->uz_fails = counter_u64_alloc(M_WAITOK);
2339 	zone->uz_xdomain = counter_u64_alloc(M_WAITOK);
2340 }
2341 
2342 static void
2343 zone_alloc_sysctl(uma_zone_t zone, void *unused)
2344 {
2345 	uma_zone_domain_t zdom;
2346 	uma_domain_t dom;
2347 	uma_keg_t keg;
2348 	struct sysctl_oid *oid, *domainoid;
2349 	int domains, i, cnt;
2350 	static const char *nokeg = "cache zone";
2351 	char *c;
2352 
2353 	/*
2354 	 * Make a sysctl safe copy of the zone name by removing
2355 	 * any special characters and handling dups by appending
2356 	 * an index.
2357 	 */
2358 	if (zone->uz_namecnt != 0) {
2359 		/* Count the number of decimal digits and '_' separator. */
2360 		for (i = 1, cnt = zone->uz_namecnt; cnt != 0; i++)
2361 			cnt /= 10;
2362 		zone->uz_ctlname = malloc(strlen(zone->uz_name) + i + 1,
2363 		    M_UMA, M_WAITOK);
2364 		sprintf(zone->uz_ctlname, "%s_%d", zone->uz_name,
2365 		    zone->uz_namecnt);
2366 	} else
2367 		zone->uz_ctlname = strdup(zone->uz_name, M_UMA);
2368 	for (c = zone->uz_ctlname; *c != '\0'; c++)
2369 		if (strchr("./\\ -", *c) != NULL)
2370 			*c = '_';
2371 
2372 	/*
2373 	 * Basic parameters at the root.
2374 	 */
2375 	zone->uz_oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_vm_uma),
2376 	    OID_AUTO, zone->uz_ctlname, CTLFLAG_RD, NULL, "");
2377 	oid = zone->uz_oid;
2378 	SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2379 	    "size", CTLFLAG_RD, &zone->uz_size, 0, "Allocation size");
2380 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2381 	    "flags", CTLFLAG_RD | CTLTYPE_STRING | CTLFLAG_MPSAFE,
2382 	    zone, 0, sysctl_handle_uma_zone_flags, "A",
2383 	    "Allocator configuration flags");
2384 	SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2385 	    "bucket_size", CTLFLAG_RD, &zone->uz_bucket_size, 0,
2386 	    "Desired per-cpu cache size");
2387 	SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2388 	    "bucket_size_max", CTLFLAG_RD, &zone->uz_bucket_size_max, 0,
2389 	    "Maximum allowed per-cpu cache size");
2390 
2391 	/*
2392 	 * keg if present.
2393 	 */
2394 	if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
2395 		domains = vm_ndomains;
2396 	else
2397 		domains = 1;
2398 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2399 	    "keg", CTLFLAG_RD, NULL, "");
2400 	keg = zone->uz_keg;
2401 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0) {
2402 		SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2403 		    "name", CTLFLAG_RD, keg->uk_name, "Keg name");
2404 		SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2405 		    "rsize", CTLFLAG_RD, &keg->uk_rsize, 0,
2406 		    "Real object size with alignment");
2407 		SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2408 		    "ppera", CTLFLAG_RD, &keg->uk_ppera, 0,
2409 		    "pages per-slab allocation");
2410 		SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2411 		    "ipers", CTLFLAG_RD, &keg->uk_ipers, 0,
2412 		    "items available per-slab");
2413 		SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2414 		    "align", CTLFLAG_RD, &keg->uk_align, 0,
2415 		    "item alignment mask");
2416 		SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2417 		    "efficiency", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2418 		    keg, 0, sysctl_handle_uma_slab_efficiency, "I",
2419 		    "Slab utilization (100 - internal fragmentation %)");
2420 		domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(oid),
2421 		    OID_AUTO, "domain", CTLFLAG_RD, NULL, "");
2422 		for (i = 0; i < domains; i++) {
2423 			dom = &keg->uk_domain[i];
2424 			oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2425 			    OID_AUTO, VM_DOMAIN(i)->vmd_name, CTLFLAG_RD,
2426 			    NULL, "");
2427 			SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2428 			    "pages", CTLFLAG_RD, &dom->ud_pages, 0,
2429 			    "Total pages currently allocated from VM");
2430 			SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2431 			    "free_items", CTLFLAG_RD, &dom->ud_free_items, 0,
2432 			    "items free in the slab layer");
2433 		}
2434 	} else
2435 		SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2436 		    "name", CTLFLAG_RD, nokeg, "Keg name");
2437 
2438 	/*
2439 	 * Information about zone limits.
2440 	 */
2441 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2442 	    "limit", CTLFLAG_RD, NULL, "");
2443 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2444 	    "items", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2445 	    zone, 0, sysctl_handle_uma_zone_items, "QU",
2446 	    "current number of allocated items if limit is set");
2447 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2448 	    "max_items", CTLFLAG_RD, &zone->uz_max_items, 0,
2449 	    "Maximum number of cached items");
2450 	SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2451 	    "sleepers", CTLFLAG_RD, &zone->uz_sleepers, 0,
2452 	    "Number of threads sleeping at limit");
2453 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2454 	    "sleeps", CTLFLAG_RD, &zone->uz_sleeps, 0,
2455 	    "Total zone limit sleeps");
2456 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2457 	    "bucket_max", CTLFLAG_RD, &zone->uz_bucket_max, 0,
2458 	    "Maximum number of items in each domain's bucket cache");
2459 
2460 	/*
2461 	 * Per-domain zone information.
2462 	 */
2463 	domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid),
2464 	    OID_AUTO, "domain", CTLFLAG_RD, NULL, "");
2465 	for (i = 0; i < domains; i++) {
2466 		zdom = ZDOM_GET(zone, i);
2467 		oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2468 		    OID_AUTO, VM_DOMAIN(i)->vmd_name, CTLFLAG_RD, NULL, "");
2469 		SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2470 		    "nitems", CTLFLAG_RD, &zdom->uzd_nitems,
2471 		    "number of items in this domain");
2472 		SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2473 		    "imax", CTLFLAG_RD, &zdom->uzd_imax,
2474 		    "maximum item count in this period");
2475 		SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2476 		    "imin", CTLFLAG_RD, &zdom->uzd_imin,
2477 		    "minimum item count in this period");
2478 		SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2479 		    "wss", CTLFLAG_RD, &zdom->uzd_wss,
2480 		    "Working set size");
2481 	}
2482 
2483 	/*
2484 	 * General statistics.
2485 	 */
2486 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2487 	    "stats", CTLFLAG_RD, NULL, "");
2488 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2489 	    "current", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2490 	    zone, 1, sysctl_handle_uma_zone_cur, "I",
2491 	    "Current number of allocated items");
2492 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2493 	    "allocs", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2494 	    zone, 0, sysctl_handle_uma_zone_allocs, "QU",
2495 	    "Total allocation calls");
2496 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2497 	    "frees", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2498 	    zone, 0, sysctl_handle_uma_zone_frees, "QU",
2499 	    "Total free calls");
2500 	SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2501 	    "fails", CTLFLAG_RD, &zone->uz_fails,
2502 	    "Number of allocation failures");
2503 	SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2504 	    "xdomain", CTLFLAG_RD, &zone->uz_xdomain,
2505 	    "Free calls from the wrong domain");
2506 }
2507 
2508 struct uma_zone_count {
2509 	const char	*name;
2510 	int		count;
2511 };
2512 
2513 static void
2514 zone_count(uma_zone_t zone, void *arg)
2515 {
2516 	struct uma_zone_count *cnt;
2517 
2518 	cnt = arg;
2519 	/*
2520 	 * Some zones are rapidly created with identical names and
2521 	 * destroyed out of order.  This can lead to gaps in the count.
2522 	 * Use one greater than the maximum observed for this name.
2523 	 */
2524 	if (strcmp(zone->uz_name, cnt->name) == 0)
2525 		cnt->count = MAX(cnt->count,
2526 		    zone->uz_namecnt + 1);
2527 }
2528 
2529 static void
2530 zone_update_caches(uma_zone_t zone)
2531 {
2532 	int i;
2533 
2534 	for (i = 0; i <= mp_maxid; i++) {
2535 		cache_set_uz_size(&zone->uz_cpu[i], zone->uz_size);
2536 		cache_set_uz_flags(&zone->uz_cpu[i], zone->uz_flags);
2537 	}
2538 }
2539 
2540 /*
2541  * Zone header ctor.  This initializes all fields, locks, etc.
2542  *
2543  * Arguments/Returns follow uma_ctor specifications
2544  *	udata  Actually uma_zctor_args
2545  */
2546 static int
2547 zone_ctor(void *mem, int size, void *udata, int flags)
2548 {
2549 	struct uma_zone_count cnt;
2550 	struct uma_zctor_args *arg = udata;
2551 	uma_zone_domain_t zdom;
2552 	uma_zone_t zone = mem;
2553 	uma_zone_t z;
2554 	uma_keg_t keg;
2555 	int i;
2556 
2557 	bzero(zone, size);
2558 	zone->uz_name = arg->name;
2559 	zone->uz_ctor = arg->ctor;
2560 	zone->uz_dtor = arg->dtor;
2561 	zone->uz_init = NULL;
2562 	zone->uz_fini = NULL;
2563 	zone->uz_sleeps = 0;
2564 	zone->uz_bucket_size = 0;
2565 	zone->uz_bucket_size_min = 0;
2566 	zone->uz_bucket_size_max = BUCKET_MAX;
2567 	zone->uz_flags = (arg->flags & UMA_ZONE_SMR);
2568 	zone->uz_warning = NULL;
2569 	/* The domain structures follow the cpu structures. */
2570 	zone->uz_bucket_max = ULONG_MAX;
2571 	timevalclear(&zone->uz_ratecheck);
2572 
2573 	/* Count the number of duplicate names. */
2574 	cnt.name = arg->name;
2575 	cnt.count = 0;
2576 	zone_foreach(zone_count, &cnt);
2577 	zone->uz_namecnt = cnt.count;
2578 	ZONE_CROSS_LOCK_INIT(zone);
2579 
2580 	for (i = 0; i < vm_ndomains; i++) {
2581 		zdom = ZDOM_GET(zone, i);
2582 		ZDOM_LOCK_INIT(zone, zdom, (arg->flags & UMA_ZONE_MTXCLASS));
2583 		STAILQ_INIT(&zdom->uzd_buckets);
2584 	}
2585 
2586 #ifdef INVARIANTS
2587 	if (arg->uminit == trash_init && arg->fini == trash_fini)
2588 		zone->uz_flags |= UMA_ZFLAG_TRASH | UMA_ZFLAG_CTORDTOR;
2589 #endif
2590 
2591 	/*
2592 	 * This is a pure cache zone, no kegs.
2593 	 */
2594 	if (arg->import) {
2595 		KASSERT((arg->flags & UMA_ZFLAG_CACHE) != 0,
2596 		    ("zone_ctor: Import specified for non-cache zone."));
2597 		zone->uz_flags = arg->flags;
2598 		zone->uz_size = arg->size;
2599 		zone->uz_import = arg->import;
2600 		zone->uz_release = arg->release;
2601 		zone->uz_arg = arg->arg;
2602 #ifdef NUMA
2603 		/*
2604 		 * Cache zones are round-robin unless a policy is
2605 		 * specified because they may have incompatible
2606 		 * constraints.
2607 		 */
2608 		if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2609 			zone->uz_flags |= UMA_ZONE_ROUNDROBIN;
2610 #endif
2611 		rw_wlock(&uma_rwlock);
2612 		LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link);
2613 		rw_wunlock(&uma_rwlock);
2614 		goto out;
2615 	}
2616 
2617 	/*
2618 	 * Use the regular zone/keg/slab allocator.
2619 	 */
2620 	zone->uz_import = zone_import;
2621 	zone->uz_release = zone_release;
2622 	zone->uz_arg = zone;
2623 	keg = arg->keg;
2624 
2625 	if (arg->flags & UMA_ZONE_SECONDARY) {
2626 		KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
2627 		    ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
2628 		KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
2629 		zone->uz_init = arg->uminit;
2630 		zone->uz_fini = arg->fini;
2631 		zone->uz_flags |= UMA_ZONE_SECONDARY;
2632 		rw_wlock(&uma_rwlock);
2633 		ZONE_LOCK(zone);
2634 		LIST_FOREACH(z, &keg->uk_zones, uz_link) {
2635 			if (LIST_NEXT(z, uz_link) == NULL) {
2636 				LIST_INSERT_AFTER(z, zone, uz_link);
2637 				break;
2638 			}
2639 		}
2640 		ZONE_UNLOCK(zone);
2641 		rw_wunlock(&uma_rwlock);
2642 	} else if (keg == NULL) {
2643 		if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
2644 		    arg->align, arg->flags)) == NULL)
2645 			return (ENOMEM);
2646 	} else {
2647 		struct uma_kctor_args karg;
2648 		int error;
2649 
2650 		/* We should only be here from uma_startup() */
2651 		karg.size = arg->size;
2652 		karg.uminit = arg->uminit;
2653 		karg.fini = arg->fini;
2654 		karg.align = arg->align;
2655 		karg.flags = (arg->flags & ~UMA_ZONE_SMR);
2656 		karg.zone = zone;
2657 		error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
2658 		    flags);
2659 		if (error)
2660 			return (error);
2661 	}
2662 
2663 	/* Inherit properties from the keg. */
2664 	zone->uz_keg = keg;
2665 	zone->uz_size = keg->uk_size;
2666 	zone->uz_flags |= (keg->uk_flags &
2667 	    (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
2668 
2669 out:
2670 	if (__predict_true(booted >= BOOT_RUNNING)) {
2671 		zone_alloc_counters(zone, NULL);
2672 		zone_alloc_sysctl(zone, NULL);
2673 	} else {
2674 		zone->uz_allocs = EARLY_COUNTER;
2675 		zone->uz_frees = EARLY_COUNTER;
2676 		zone->uz_fails = EARLY_COUNTER;
2677 	}
2678 
2679 	/* Caller requests a private SMR context. */
2680 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
2681 		zone->uz_smr = smr_create(zone->uz_name, 0, 0);
2682 
2683 	KASSERT((arg->flags & (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET)) !=
2684 	    (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET),
2685 	    ("Invalid zone flag combination"));
2686 	if (arg->flags & UMA_ZFLAG_INTERNAL)
2687 		zone->uz_bucket_size_max = zone->uz_bucket_size = 0;
2688 	if ((arg->flags & UMA_ZONE_MAXBUCKET) != 0)
2689 		zone->uz_bucket_size = BUCKET_MAX;
2690 	else if ((arg->flags & UMA_ZONE_MINBUCKET) != 0)
2691 		zone->uz_bucket_size_max = zone->uz_bucket_size = BUCKET_MIN;
2692 	else if ((arg->flags & UMA_ZONE_NOBUCKET) != 0)
2693 		zone->uz_bucket_size = 0;
2694 	else
2695 		zone->uz_bucket_size = bucket_select(zone->uz_size);
2696 	zone->uz_bucket_size_min = zone->uz_bucket_size;
2697 	if (zone->uz_dtor != NULL || zone->uz_ctor != NULL)
2698 		zone->uz_flags |= UMA_ZFLAG_CTORDTOR;
2699 	zone_update_caches(zone);
2700 
2701 	return (0);
2702 }
2703 
2704 /*
2705  * Keg header dtor.  This frees all data, destroys locks, frees the hash
2706  * table and removes the keg from the global list.
2707  *
2708  * Arguments/Returns follow uma_dtor specifications
2709  *	udata  unused
2710  */
2711 static void
2712 keg_dtor(void *arg, int size, void *udata)
2713 {
2714 	uma_keg_t keg;
2715 	uint32_t free, pages;
2716 	int i;
2717 
2718 	keg = (uma_keg_t)arg;
2719 	free = pages = 0;
2720 	for (i = 0; i < vm_ndomains; i++) {
2721 		free += keg->uk_domain[i].ud_free_items;
2722 		pages += keg->uk_domain[i].ud_pages;
2723 		KEG_LOCK_FINI(keg, i);
2724 	}
2725 	if (pages != 0)
2726 		printf("Freed UMA keg (%s) was not empty (%u items). "
2727 		    " Lost %u pages of memory.\n",
2728 		    keg->uk_name ? keg->uk_name : "",
2729 		    pages / keg->uk_ppera * keg->uk_ipers - free, pages);
2730 
2731 	hash_free(&keg->uk_hash);
2732 }
2733 
2734 /*
2735  * Zone header dtor.
2736  *
2737  * Arguments/Returns follow uma_dtor specifications
2738  *	udata  unused
2739  */
2740 static void
2741 zone_dtor(void *arg, int size, void *udata)
2742 {
2743 	uma_zone_t zone;
2744 	uma_keg_t keg;
2745 	int i;
2746 
2747 	zone = (uma_zone_t)arg;
2748 
2749 	sysctl_remove_oid(zone->uz_oid, 1, 1);
2750 
2751 	if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
2752 		cache_drain(zone);
2753 
2754 	rw_wlock(&uma_rwlock);
2755 	LIST_REMOVE(zone, uz_link);
2756 	rw_wunlock(&uma_rwlock);
2757 	zone_reclaim(zone, M_WAITOK, true);
2758 
2759 	/*
2760 	 * We only destroy kegs from non secondary/non cache zones.
2761 	 */
2762 	if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) {
2763 		keg = zone->uz_keg;
2764 		rw_wlock(&uma_rwlock);
2765 		LIST_REMOVE(keg, uk_link);
2766 		rw_wunlock(&uma_rwlock);
2767 		zone_free_item(kegs, keg, NULL, SKIP_NONE);
2768 	}
2769 	counter_u64_free(zone->uz_allocs);
2770 	counter_u64_free(zone->uz_frees);
2771 	counter_u64_free(zone->uz_fails);
2772 	counter_u64_free(zone->uz_xdomain);
2773 	free(zone->uz_ctlname, M_UMA);
2774 	for (i = 0; i < vm_ndomains; i++)
2775 		ZDOM_LOCK_FINI(ZDOM_GET(zone, i));
2776 	ZONE_CROSS_LOCK_FINI(zone);
2777 }
2778 
2779 static void
2780 zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *arg), void *arg)
2781 {
2782 	uma_keg_t keg;
2783 	uma_zone_t zone;
2784 
2785 	LIST_FOREACH(keg, &uma_kegs, uk_link) {
2786 		LIST_FOREACH(zone, &keg->uk_zones, uz_link)
2787 			zfunc(zone, arg);
2788 	}
2789 	LIST_FOREACH(zone, &uma_cachezones, uz_link)
2790 		zfunc(zone, arg);
2791 }
2792 
2793 /*
2794  * Traverses every zone in the system and calls a callback
2795  *
2796  * Arguments:
2797  *	zfunc  A pointer to a function which accepts a zone
2798  *		as an argument.
2799  *
2800  * Returns:
2801  *	Nothing
2802  */
2803 static void
2804 zone_foreach(void (*zfunc)(uma_zone_t, void *arg), void *arg)
2805 {
2806 
2807 	rw_rlock(&uma_rwlock);
2808 	zone_foreach_unlocked(zfunc, arg);
2809 	rw_runlock(&uma_rwlock);
2810 }
2811 
2812 /*
2813  * Initialize the kernel memory allocator.  This is done after pages can be
2814  * allocated but before general KVA is available.
2815  */
2816 void
2817 uma_startup1(vm_offset_t virtual_avail)
2818 {
2819 	struct uma_zctor_args args;
2820 	size_t ksize, zsize, size;
2821 	uma_keg_t masterkeg;
2822 	uintptr_t m;
2823 	uint8_t pflag;
2824 
2825 	bootstart = bootmem = virtual_avail;
2826 
2827 	rw_init(&uma_rwlock, "UMA lock");
2828 	sx_init(&uma_reclaim_lock, "umareclaim");
2829 
2830 	ksize = sizeof(struct uma_keg) +
2831 	    (sizeof(struct uma_domain) * vm_ndomains);
2832 	ksize = roundup(ksize, UMA_SUPER_ALIGN);
2833 	zsize = sizeof(struct uma_zone) +
2834 	    (sizeof(struct uma_cache) * (mp_maxid + 1)) +
2835 	    (sizeof(struct uma_zone_domain) * vm_ndomains);
2836 	zsize = roundup(zsize, UMA_SUPER_ALIGN);
2837 
2838 	/* Allocate the zone of zones, zone of kegs, and zone of zones keg. */
2839 	size = (zsize * 2) + ksize;
2840 	m = (uintptr_t)startup_alloc(NULL, size, 0, &pflag, M_NOWAIT | M_ZERO);
2841 	zones = (uma_zone_t)m;
2842 	m += zsize;
2843 	kegs = (uma_zone_t)m;
2844 	m += zsize;
2845 	masterkeg = (uma_keg_t)m;
2846 
2847 	/* "manually" create the initial zone */
2848 	memset(&args, 0, sizeof(args));
2849 	args.name = "UMA Kegs";
2850 	args.size = ksize;
2851 	args.ctor = keg_ctor;
2852 	args.dtor = keg_dtor;
2853 	args.uminit = zero_init;
2854 	args.fini = NULL;
2855 	args.keg = masterkeg;
2856 	args.align = UMA_SUPER_ALIGN - 1;
2857 	args.flags = UMA_ZFLAG_INTERNAL;
2858 	zone_ctor(kegs, zsize, &args, M_WAITOK);
2859 
2860 	args.name = "UMA Zones";
2861 	args.size = zsize;
2862 	args.ctor = zone_ctor;
2863 	args.dtor = zone_dtor;
2864 	args.uminit = zero_init;
2865 	args.fini = NULL;
2866 	args.keg = NULL;
2867 	args.align = UMA_SUPER_ALIGN - 1;
2868 	args.flags = UMA_ZFLAG_INTERNAL;
2869 	zone_ctor(zones, zsize, &args, M_WAITOK);
2870 
2871 	/* Now make zones for slab headers */
2872 	slabzones[0] = uma_zcreate("UMA Slabs 0", SLABZONE0_SIZE,
2873 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2874 	slabzones[1] = uma_zcreate("UMA Slabs 1", SLABZONE1_SIZE,
2875 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2876 
2877 	hashzone = uma_zcreate("UMA Hash",
2878 	    sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
2879 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2880 
2881 	bucket_init();
2882 	smr_init();
2883 }
2884 
2885 #ifndef UMA_MD_SMALL_ALLOC
2886 extern void vm_radix_reserve_kva(void);
2887 #endif
2888 
2889 /*
2890  * Advertise the availability of normal kva allocations and switch to
2891  * the default back-end allocator.  Marks the KVA we consumed on startup
2892  * as used in the map.
2893  */
2894 void
2895 uma_startup2(void)
2896 {
2897 
2898 	if (bootstart != bootmem) {
2899 		vm_map_lock(kernel_map);
2900 		(void)vm_map_insert(kernel_map, NULL, 0, bootstart, bootmem,
2901 		    VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT);
2902 		vm_map_unlock(kernel_map);
2903 	}
2904 
2905 #ifndef UMA_MD_SMALL_ALLOC
2906 	/* Set up radix zone to use noobj_alloc. */
2907 	vm_radix_reserve_kva();
2908 #endif
2909 
2910 	booted = BOOT_KVA;
2911 	zone_foreach_unlocked(zone_kva_available, NULL);
2912 	bucket_enable();
2913 }
2914 
2915 /*
2916  * Finish our initialization steps.
2917  */
2918 static void
2919 uma_startup3(void)
2920 {
2921 
2922 #ifdef INVARIANTS
2923 	TUNABLE_INT_FETCH("vm.debug.divisor", &dbg_divisor);
2924 	uma_dbg_cnt = counter_u64_alloc(M_WAITOK);
2925 	uma_skip_cnt = counter_u64_alloc(M_WAITOK);
2926 #endif
2927 	zone_foreach_unlocked(zone_alloc_counters, NULL);
2928 	zone_foreach_unlocked(zone_alloc_sysctl, NULL);
2929 	callout_init(&uma_callout, 1);
2930 	callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
2931 	booted = BOOT_RUNNING;
2932 
2933 	EVENTHANDLER_REGISTER(shutdown_post_sync, uma_shutdown, NULL,
2934 	    EVENTHANDLER_PRI_FIRST);
2935 }
2936 
2937 static void
2938 uma_shutdown(void)
2939 {
2940 
2941 	booted = BOOT_SHUTDOWN;
2942 }
2943 
2944 static uma_keg_t
2945 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
2946 		int align, uint32_t flags)
2947 {
2948 	struct uma_kctor_args args;
2949 
2950 	args.size = size;
2951 	args.uminit = uminit;
2952 	args.fini = fini;
2953 	args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
2954 	args.flags = flags;
2955 	args.zone = zone;
2956 	return (zone_alloc_item(kegs, &args, UMA_ANYDOMAIN, M_WAITOK));
2957 }
2958 
2959 /* Public functions */
2960 /* See uma.h */
2961 void
2962 uma_set_align(int align)
2963 {
2964 
2965 	if (align != UMA_ALIGN_CACHE)
2966 		uma_align_cache = align;
2967 }
2968 
2969 /* See uma.h */
2970 uma_zone_t
2971 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
2972 		uma_init uminit, uma_fini fini, int align, uint32_t flags)
2973 
2974 {
2975 	struct uma_zctor_args args;
2976 	uma_zone_t res;
2977 
2978 	KASSERT(powerof2(align + 1), ("invalid zone alignment %d for \"%s\"",
2979 	    align, name));
2980 
2981 	/* This stuff is essential for the zone ctor */
2982 	memset(&args, 0, sizeof(args));
2983 	args.name = name;
2984 	args.size = size;
2985 	args.ctor = ctor;
2986 	args.dtor = dtor;
2987 	args.uminit = uminit;
2988 	args.fini = fini;
2989 #ifdef  INVARIANTS
2990 	/*
2991 	 * Inject procedures which check for memory use after free if we are
2992 	 * allowed to scramble the memory while it is not allocated.  This
2993 	 * requires that: UMA is actually able to access the memory, no init
2994 	 * or fini procedures, no dependency on the initial value of the
2995 	 * memory, and no (legitimate) use of the memory after free.  Note,
2996 	 * the ctor and dtor do not need to be empty.
2997 	 */
2998 	if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOTOUCH |
2999 	    UMA_ZONE_NOFREE))) && uminit == NULL && fini == NULL) {
3000 		args.uminit = trash_init;
3001 		args.fini = trash_fini;
3002 	}
3003 #endif
3004 	args.align = align;
3005 	args.flags = flags;
3006 	args.keg = NULL;
3007 
3008 	sx_slock(&uma_reclaim_lock);
3009 	res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3010 	sx_sunlock(&uma_reclaim_lock);
3011 
3012 	return (res);
3013 }
3014 
3015 /* See uma.h */
3016 uma_zone_t
3017 uma_zsecond_create(const char *name, uma_ctor ctor, uma_dtor dtor,
3018     uma_init zinit, uma_fini zfini, uma_zone_t master)
3019 {
3020 	struct uma_zctor_args args;
3021 	uma_keg_t keg;
3022 	uma_zone_t res;
3023 
3024 	keg = master->uz_keg;
3025 	memset(&args, 0, sizeof(args));
3026 	args.name = name;
3027 	args.size = keg->uk_size;
3028 	args.ctor = ctor;
3029 	args.dtor = dtor;
3030 	args.uminit = zinit;
3031 	args.fini = zfini;
3032 	args.align = keg->uk_align;
3033 	args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
3034 	args.keg = keg;
3035 
3036 	sx_slock(&uma_reclaim_lock);
3037 	res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3038 	sx_sunlock(&uma_reclaim_lock);
3039 
3040 	return (res);
3041 }
3042 
3043 /* See uma.h */
3044 uma_zone_t
3045 uma_zcache_create(const char *name, int size, uma_ctor ctor, uma_dtor dtor,
3046     uma_init zinit, uma_fini zfini, uma_import zimport, uma_release zrelease,
3047     void *arg, int flags)
3048 {
3049 	struct uma_zctor_args args;
3050 
3051 	memset(&args, 0, sizeof(args));
3052 	args.name = name;
3053 	args.size = size;
3054 	args.ctor = ctor;
3055 	args.dtor = dtor;
3056 	args.uminit = zinit;
3057 	args.fini = zfini;
3058 	args.import = zimport;
3059 	args.release = zrelease;
3060 	args.arg = arg;
3061 	args.align = 0;
3062 	args.flags = flags | UMA_ZFLAG_CACHE;
3063 
3064 	return (zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK));
3065 }
3066 
3067 /* See uma.h */
3068 void
3069 uma_zdestroy(uma_zone_t zone)
3070 {
3071 
3072 	/*
3073 	 * Large slabs are expensive to reclaim, so don't bother doing
3074 	 * unnecessary work if we're shutting down.
3075 	 */
3076 	if (booted == BOOT_SHUTDOWN &&
3077 	    zone->uz_fini == NULL && zone->uz_release == zone_release)
3078 		return;
3079 	sx_slock(&uma_reclaim_lock);
3080 	zone_free_item(zones, zone, NULL, SKIP_NONE);
3081 	sx_sunlock(&uma_reclaim_lock);
3082 }
3083 
3084 void
3085 uma_zwait(uma_zone_t zone)
3086 {
3087 
3088 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
3089 		uma_zfree_smr(zone, uma_zalloc_smr(zone, M_WAITOK));
3090 	else if ((zone->uz_flags & UMA_ZONE_PCPU) != 0)
3091 		uma_zfree_pcpu(zone, uma_zalloc_pcpu(zone, M_WAITOK));
3092 	else
3093 		uma_zfree(zone, uma_zalloc(zone, M_WAITOK));
3094 }
3095 
3096 void *
3097 uma_zalloc_pcpu_arg(uma_zone_t zone, void *udata, int flags)
3098 {
3099 	void *item, *pcpu_item;
3100 #ifdef SMP
3101 	int i;
3102 
3103 	MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3104 #endif
3105 	item = uma_zalloc_arg(zone, udata, flags & ~M_ZERO);
3106 	if (item == NULL)
3107 		return (NULL);
3108 	pcpu_item = zpcpu_base_to_offset(item);
3109 	if (flags & M_ZERO) {
3110 #ifdef SMP
3111 		for (i = 0; i <= mp_maxid; i++)
3112 			bzero(zpcpu_get_cpu(pcpu_item, i), zone->uz_size);
3113 #else
3114 		bzero(item, zone->uz_size);
3115 #endif
3116 	}
3117 	return (pcpu_item);
3118 }
3119 
3120 /*
3121  * A stub while both regular and pcpu cases are identical.
3122  */
3123 void
3124 uma_zfree_pcpu_arg(uma_zone_t zone, void *pcpu_item, void *udata)
3125 {
3126 	void *item;
3127 
3128 #ifdef SMP
3129 	MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3130 #endif
3131 	item = zpcpu_offset_to_base(pcpu_item);
3132 	uma_zfree_arg(zone, item, udata);
3133 }
3134 
3135 static inline void *
3136 item_ctor(uma_zone_t zone, int uz_flags, int size, void *udata, int flags,
3137     void *item)
3138 {
3139 #ifdef INVARIANTS
3140 	bool skipdbg;
3141 
3142 	skipdbg = uma_dbg_zskip(zone, item);
3143 	if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3144 	    zone->uz_ctor != trash_ctor)
3145 		trash_ctor(item, size, udata, flags);
3146 #endif
3147 	/* Check flags before loading ctor pointer. */
3148 	if (__predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0) &&
3149 	    __predict_false(zone->uz_ctor != NULL) &&
3150 	    zone->uz_ctor(item, size, udata, flags) != 0) {
3151 		counter_u64_add(zone->uz_fails, 1);
3152 		zone_free_item(zone, item, udata, SKIP_DTOR | SKIP_CNT);
3153 		return (NULL);
3154 	}
3155 #ifdef INVARIANTS
3156 	if (!skipdbg)
3157 		uma_dbg_alloc(zone, NULL, item);
3158 #endif
3159 	if (__predict_false(flags & M_ZERO))
3160 		return (memset(item, 0, size));
3161 
3162 	return (item);
3163 }
3164 
3165 static inline void
3166 item_dtor(uma_zone_t zone, void *item, int size, void *udata,
3167     enum zfreeskip skip)
3168 {
3169 #ifdef INVARIANTS
3170 	bool skipdbg;
3171 
3172 	skipdbg = uma_dbg_zskip(zone, item);
3173 	if (skip == SKIP_NONE && !skipdbg) {
3174 		if ((zone->uz_flags & UMA_ZONE_MALLOC) != 0)
3175 			uma_dbg_free(zone, udata, item);
3176 		else
3177 			uma_dbg_free(zone, NULL, item);
3178 	}
3179 #endif
3180 	if (__predict_true(skip < SKIP_DTOR)) {
3181 		if (zone->uz_dtor != NULL)
3182 			zone->uz_dtor(item, size, udata);
3183 #ifdef INVARIANTS
3184 		if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3185 		    zone->uz_dtor != trash_dtor)
3186 			trash_dtor(item, size, udata);
3187 #endif
3188 	}
3189 }
3190 
3191 #if defined(INVARIANTS) || defined(DEBUG_MEMGUARD) || defined(WITNESS)
3192 #define	UMA_ZALLOC_DEBUG
3193 static int
3194 uma_zalloc_debug(uma_zone_t zone, void **itemp, void *udata, int flags)
3195 {
3196 	int error;
3197 
3198 	error = 0;
3199 #ifdef WITNESS
3200 	if (flags & M_WAITOK) {
3201 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
3202 		    "uma_zalloc_debug: zone \"%s\"", zone->uz_name);
3203 	}
3204 #endif
3205 
3206 #ifdef INVARIANTS
3207 	KASSERT((flags & M_EXEC) == 0,
3208 	    ("uma_zalloc_debug: called with M_EXEC"));
3209 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3210 	    ("uma_zalloc_debug: called within spinlock or critical section"));
3211 	KASSERT((zone->uz_flags & UMA_ZONE_PCPU) == 0 || (flags & M_ZERO) == 0,
3212 	    ("uma_zalloc_debug: allocating from a pcpu zone with M_ZERO"));
3213 #endif
3214 
3215 #ifdef DEBUG_MEMGUARD
3216 	if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && memguard_cmp_zone(zone)) {
3217 		void *item;
3218 		item = memguard_alloc(zone->uz_size, flags);
3219 		if (item != NULL) {
3220 			error = EJUSTRETURN;
3221 			if (zone->uz_init != NULL &&
3222 			    zone->uz_init(item, zone->uz_size, flags) != 0) {
3223 				*itemp = NULL;
3224 				return (error);
3225 			}
3226 			if (zone->uz_ctor != NULL &&
3227 			    zone->uz_ctor(item, zone->uz_size, udata,
3228 			    flags) != 0) {
3229 				counter_u64_add(zone->uz_fails, 1);
3230 			    	zone->uz_fini(item, zone->uz_size);
3231 				*itemp = NULL;
3232 				return (error);
3233 			}
3234 			*itemp = item;
3235 			return (error);
3236 		}
3237 		/* This is unfortunate but should not be fatal. */
3238 	}
3239 #endif
3240 	return (error);
3241 }
3242 
3243 static int
3244 uma_zfree_debug(uma_zone_t zone, void *item, void *udata)
3245 {
3246 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3247 	    ("uma_zfree_debug: called with spinlock or critical section held"));
3248 
3249 #ifdef DEBUG_MEMGUARD
3250 	if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && is_memguard_addr(item)) {
3251 		if (zone->uz_dtor != NULL)
3252 			zone->uz_dtor(item, zone->uz_size, udata);
3253 		if (zone->uz_fini != NULL)
3254 			zone->uz_fini(item, zone->uz_size);
3255 		memguard_free(item);
3256 		return (EJUSTRETURN);
3257 	}
3258 #endif
3259 	return (0);
3260 }
3261 #endif
3262 
3263 static inline void *
3264 cache_alloc_item(uma_zone_t zone, uma_cache_t cache, uma_cache_bucket_t bucket,
3265     void *udata, int flags)
3266 {
3267 	void *item;
3268 	int size, uz_flags;
3269 
3270 	item = cache_bucket_pop(cache, bucket);
3271 	size = cache_uz_size(cache);
3272 	uz_flags = cache_uz_flags(cache);
3273 	critical_exit();
3274 	return (item_ctor(zone, uz_flags, size, udata, flags, item));
3275 }
3276 
3277 static __noinline void *
3278 cache_alloc_retry(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3279 {
3280 	uma_cache_bucket_t bucket;
3281 	int domain;
3282 
3283 	while (cache_alloc(zone, cache, udata, flags)) {
3284 		cache = &zone->uz_cpu[curcpu];
3285 		bucket = &cache->uc_allocbucket;
3286 		if (__predict_false(bucket->ucb_cnt == 0))
3287 			continue;
3288 		return (cache_alloc_item(zone, cache, bucket, udata, flags));
3289 	}
3290 	critical_exit();
3291 
3292 	/*
3293 	 * We can not get a bucket so try to return a single item.
3294 	 */
3295 	if (zone->uz_flags & UMA_ZONE_FIRSTTOUCH)
3296 		domain = PCPU_GET(domain);
3297 	else
3298 		domain = UMA_ANYDOMAIN;
3299 	return (zone_alloc_item(zone, udata, domain, flags));
3300 }
3301 
3302 /* See uma.h */
3303 void *
3304 uma_zalloc_smr(uma_zone_t zone, int flags)
3305 {
3306 	uma_cache_bucket_t bucket;
3307 	uma_cache_t cache;
3308 
3309 #ifdef UMA_ZALLOC_DEBUG
3310 	void *item;
3311 
3312 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
3313 	    ("uma_zalloc_arg: called with non-SMR zone.\n"));
3314 	if (uma_zalloc_debug(zone, &item, NULL, flags) == EJUSTRETURN)
3315 		return (item);
3316 #endif
3317 
3318 	critical_enter();
3319 	cache = &zone->uz_cpu[curcpu];
3320 	bucket = &cache->uc_allocbucket;
3321 	if (__predict_false(bucket->ucb_cnt == 0))
3322 		return (cache_alloc_retry(zone, cache, NULL, flags));
3323 	return (cache_alloc_item(zone, cache, bucket, NULL, flags));
3324 }
3325 
3326 /* See uma.h */
3327 void *
3328 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
3329 {
3330 	uma_cache_bucket_t bucket;
3331 	uma_cache_t cache;
3332 
3333 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3334 	random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3335 
3336 	/* This is the fast path allocation */
3337 	CTR3(KTR_UMA, "uma_zalloc_arg zone %s(%p) flags %d", zone->uz_name,
3338 	    zone, flags);
3339 
3340 #ifdef UMA_ZALLOC_DEBUG
3341 	void *item;
3342 
3343 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3344 	    ("uma_zalloc_arg: called with SMR zone.\n"));
3345 	if (uma_zalloc_debug(zone, &item, udata, flags) == EJUSTRETURN)
3346 		return (item);
3347 #endif
3348 
3349 	/*
3350 	 * If possible, allocate from the per-CPU cache.  There are two
3351 	 * requirements for safe access to the per-CPU cache: (1) the thread
3352 	 * accessing the cache must not be preempted or yield during access,
3353 	 * and (2) the thread must not migrate CPUs without switching which
3354 	 * cache it accesses.  We rely on a critical section to prevent
3355 	 * preemption and migration.  We release the critical section in
3356 	 * order to acquire the zone mutex if we are unable to allocate from
3357 	 * the current cache; when we re-acquire the critical section, we
3358 	 * must detect and handle migration if it has occurred.
3359 	 */
3360 	critical_enter();
3361 	cache = &zone->uz_cpu[curcpu];
3362 	bucket = &cache->uc_allocbucket;
3363 	if (__predict_false(bucket->ucb_cnt == 0))
3364 		return (cache_alloc_retry(zone, cache, udata, flags));
3365 	return (cache_alloc_item(zone, cache, bucket, udata, flags));
3366 }
3367 
3368 /*
3369  * Replenish an alloc bucket and possibly restore an old one.  Called in
3370  * a critical section.  Returns in a critical section.
3371  *
3372  * A false return value indicates an allocation failure.
3373  * A true return value indicates success and the caller should retry.
3374  */
3375 static __noinline bool
3376 cache_alloc(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3377 {
3378 	uma_bucket_t bucket;
3379 	int domain;
3380 	bool new;
3381 
3382 	CRITICAL_ASSERT(curthread);
3383 
3384 	/*
3385 	 * If we have run out of items in our alloc bucket see
3386 	 * if we can switch with the free bucket.
3387 	 *
3388 	 * SMR Zones can't re-use the free bucket until the sequence has
3389 	 * expired.
3390 	 */
3391 	if ((cache_uz_flags(cache) & UMA_ZONE_SMR) == 0 &&
3392 	    cache->uc_freebucket.ucb_cnt != 0) {
3393 		cache_bucket_swap(&cache->uc_freebucket,
3394 		    &cache->uc_allocbucket);
3395 		return (true);
3396 	}
3397 
3398 	/*
3399 	 * Discard any empty allocation bucket while we hold no locks.
3400 	 */
3401 	bucket = cache_bucket_unload_alloc(cache);
3402 	critical_exit();
3403 
3404 	if (bucket != NULL) {
3405 		KASSERT(bucket->ub_cnt == 0,
3406 		    ("cache_alloc: Entered with non-empty alloc bucket."));
3407 		bucket_free(zone, bucket, udata);
3408 	}
3409 
3410 	/* Short-circuit for zones without buckets and low memory. */
3411 	if (zone->uz_bucket_size == 0 || bucketdisable) {
3412 		critical_enter();
3413 		return (false);
3414 	}
3415 
3416 	/*
3417 	 * Attempt to retrieve the item from the per-CPU cache has failed, so
3418 	 * we must go back to the zone.  This requires the zdom lock, so we
3419 	 * must drop the critical section, then re-acquire it when we go back
3420 	 * to the cache.  Since the critical section is released, we may be
3421 	 * preempted or migrate.  As such, make sure not to maintain any
3422 	 * thread-local state specific to the cache from prior to releasing
3423 	 * the critical section.
3424 	 */
3425 	domain = PCPU_GET(domain);
3426 	if ((cache_uz_flags(cache) & UMA_ZONE_ROUNDROBIN) != 0)
3427 		domain = zone_domain_highest(zone, domain);
3428 	bucket = cache_fetch_bucket(zone, cache, domain);
3429 	if (bucket == NULL) {
3430 		bucket = zone_alloc_bucket(zone, udata, domain, flags);
3431 		new = true;
3432 	} else
3433 		new = false;
3434 
3435 	CTR3(KTR_UMA, "uma_zalloc: zone %s(%p) bucket zone returned %p",
3436 	    zone->uz_name, zone, bucket);
3437 	if (bucket == NULL) {
3438 		critical_enter();
3439 		return (false);
3440 	}
3441 
3442 	/*
3443 	 * See if we lost the race or were migrated.  Cache the
3444 	 * initialized bucket to make this less likely or claim
3445 	 * the memory directly.
3446 	 */
3447 	critical_enter();
3448 	cache = &zone->uz_cpu[curcpu];
3449 	if (cache->uc_allocbucket.ucb_bucket == NULL &&
3450 	    ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) == 0 ||
3451 	    domain == PCPU_GET(domain))) {
3452 		if (new)
3453 			atomic_add_long(&ZDOM_GET(zone, domain)->uzd_imax,
3454 			    bucket->ub_cnt);
3455 		cache_bucket_load_alloc(cache, bucket);
3456 		return (true);
3457 	}
3458 
3459 	/*
3460 	 * We lost the race, release this bucket and start over.
3461 	 */
3462 	critical_exit();
3463 	zone_put_bucket(zone, domain, bucket, udata, false);
3464 	critical_enter();
3465 
3466 	return (true);
3467 }
3468 
3469 void *
3470 uma_zalloc_domain(uma_zone_t zone, void *udata, int domain, int flags)
3471 {
3472 
3473 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3474 	random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3475 
3476 	/* This is the fast path allocation */
3477 	CTR4(KTR_UMA, "uma_zalloc_domain zone %s(%p) domain %d flags %d",
3478 	    zone->uz_name, zone, domain, flags);
3479 
3480 	if (flags & M_WAITOK) {
3481 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
3482 		    "uma_zalloc_domain: zone \"%s\"", zone->uz_name);
3483 	}
3484 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3485 	    ("uma_zalloc_domain: called with spinlock or critical section held"));
3486 
3487 	return (zone_alloc_item(zone, udata, domain, flags));
3488 }
3489 
3490 /*
3491  * Find a slab with some space.  Prefer slabs that are partially used over those
3492  * that are totally full.  This helps to reduce fragmentation.
3493  *
3494  * If 'rr' is 1, search all domains starting from 'domain'.  Otherwise check
3495  * only 'domain'.
3496  */
3497 static uma_slab_t
3498 keg_first_slab(uma_keg_t keg, int domain, bool rr)
3499 {
3500 	uma_domain_t dom;
3501 	uma_slab_t slab;
3502 	int start;
3503 
3504 	KASSERT(domain >= 0 && domain < vm_ndomains,
3505 	    ("keg_first_slab: domain %d out of range", domain));
3506 	KEG_LOCK_ASSERT(keg, domain);
3507 
3508 	slab = NULL;
3509 	start = domain;
3510 	do {
3511 		dom = &keg->uk_domain[domain];
3512 		if ((slab = LIST_FIRST(&dom->ud_part_slab)) != NULL)
3513 			return (slab);
3514 		if ((slab = LIST_FIRST(&dom->ud_free_slab)) != NULL) {
3515 			LIST_REMOVE(slab, us_link);
3516 			dom->ud_free_slabs--;
3517 			LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
3518 			return (slab);
3519 		}
3520 		if (rr)
3521 			domain = (domain + 1) % vm_ndomains;
3522 	} while (domain != start);
3523 
3524 	return (NULL);
3525 }
3526 
3527 /*
3528  * Fetch an existing slab from a free or partial list.  Returns with the
3529  * keg domain lock held if a slab was found or unlocked if not.
3530  */
3531 static uma_slab_t
3532 keg_fetch_free_slab(uma_keg_t keg, int domain, bool rr, int flags)
3533 {
3534 	uma_slab_t slab;
3535 	uint32_t reserve;
3536 
3537 	/* HASH has a single free list. */
3538 	if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
3539 		domain = 0;
3540 
3541 	KEG_LOCK(keg, domain);
3542 	reserve = (flags & M_USE_RESERVE) != 0 ? 0 : keg->uk_reserve;
3543 	if (keg->uk_domain[domain].ud_free_items <= reserve ||
3544 	    (slab = keg_first_slab(keg, domain, rr)) == NULL) {
3545 		KEG_UNLOCK(keg, domain);
3546 		return (NULL);
3547 	}
3548 	return (slab);
3549 }
3550 
3551 static uma_slab_t
3552 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int rdomain, const int flags)
3553 {
3554 	struct vm_domainset_iter di;
3555 	uma_slab_t slab;
3556 	int aflags, domain;
3557 	bool rr;
3558 
3559 restart:
3560 	/*
3561 	 * Use the keg's policy if upper layers haven't already specified a
3562 	 * domain (as happens with first-touch zones).
3563 	 *
3564 	 * To avoid races we run the iterator with the keg lock held, but that
3565 	 * means that we cannot allow the vm_domainset layer to sleep.  Thus,
3566 	 * clear M_WAITOK and handle low memory conditions locally.
3567 	 */
3568 	rr = rdomain == UMA_ANYDOMAIN;
3569 	if (rr) {
3570 		aflags = (flags & ~M_WAITOK) | M_NOWAIT;
3571 		vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
3572 		    &aflags);
3573 	} else {
3574 		aflags = flags;
3575 		domain = rdomain;
3576 	}
3577 
3578 	for (;;) {
3579 		slab = keg_fetch_free_slab(keg, domain, rr, flags);
3580 		if (slab != NULL)
3581 			return (slab);
3582 
3583 		/*
3584 		 * M_NOVM means don't ask at all!
3585 		 */
3586 		if (flags & M_NOVM)
3587 			break;
3588 
3589 		slab = keg_alloc_slab(keg, zone, domain, flags, aflags);
3590 		if (slab != NULL)
3591 			return (slab);
3592 		if (!rr && (flags & M_WAITOK) == 0)
3593 			break;
3594 		if (rr && vm_domainset_iter_policy(&di, &domain) != 0) {
3595 			if ((flags & M_WAITOK) != 0) {
3596 				vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask);
3597 				goto restart;
3598 			}
3599 			break;
3600 		}
3601 	}
3602 
3603 	/*
3604 	 * We might not have been able to get a slab but another cpu
3605 	 * could have while we were unlocked.  Check again before we
3606 	 * fail.
3607 	 */
3608 	if ((slab = keg_fetch_free_slab(keg, domain, rr, flags)) != NULL)
3609 		return (slab);
3610 
3611 	return (NULL);
3612 }
3613 
3614 static void *
3615 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
3616 {
3617 	uma_domain_t dom;
3618 	void *item;
3619 	int freei;
3620 
3621 	KEG_LOCK_ASSERT(keg, slab->us_domain);
3622 
3623 	dom = &keg->uk_domain[slab->us_domain];
3624 	freei = BIT_FFS(keg->uk_ipers, &slab->us_free) - 1;
3625 	BIT_CLR(keg->uk_ipers, freei, &slab->us_free);
3626 	item = slab_item(slab, keg, freei);
3627 	slab->us_freecount--;
3628 	dom->ud_free_items--;
3629 
3630 	/*
3631 	 * Move this slab to the full list.  It must be on the partial list, so
3632 	 * we do not need to update the free slab count.  In particular,
3633 	 * keg_fetch_slab() always returns slabs on the partial list.
3634 	 */
3635 	if (slab->us_freecount == 0) {
3636 		LIST_REMOVE(slab, us_link);
3637 		LIST_INSERT_HEAD(&dom->ud_full_slab, slab, us_link);
3638 	}
3639 
3640 	return (item);
3641 }
3642 
3643 static int
3644 zone_import(void *arg, void **bucket, int max, int domain, int flags)
3645 {
3646 	uma_domain_t dom;
3647 	uma_zone_t zone;
3648 	uma_slab_t slab;
3649 	uma_keg_t keg;
3650 #ifdef NUMA
3651 	int stripe;
3652 #endif
3653 	int i;
3654 
3655 	zone = arg;
3656 	slab = NULL;
3657 	keg = zone->uz_keg;
3658 	/* Try to keep the buckets totally full */
3659 	for (i = 0; i < max; ) {
3660 		if ((slab = keg_fetch_slab(keg, zone, domain, flags)) == NULL)
3661 			break;
3662 #ifdef NUMA
3663 		stripe = howmany(max, vm_ndomains);
3664 #endif
3665 		dom = &keg->uk_domain[slab->us_domain];
3666 		while (slab->us_freecount && i < max) {
3667 			bucket[i++] = slab_alloc_item(keg, slab);
3668 			if (dom->ud_free_items <= keg->uk_reserve)
3669 				break;
3670 #ifdef NUMA
3671 			/*
3672 			 * If the zone is striped we pick a new slab for every
3673 			 * N allocations.  Eliminating this conditional will
3674 			 * instead pick a new domain for each bucket rather
3675 			 * than stripe within each bucket.  The current option
3676 			 * produces more fragmentation and requires more cpu
3677 			 * time but yields better distribution.
3678 			 */
3679 			if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0 &&
3680 			    vm_ndomains > 1 && --stripe == 0)
3681 				break;
3682 #endif
3683 		}
3684 		KEG_UNLOCK(keg, slab->us_domain);
3685 		/* Don't block if we allocated any successfully. */
3686 		flags &= ~M_WAITOK;
3687 		flags |= M_NOWAIT;
3688 	}
3689 
3690 	return i;
3691 }
3692 
3693 static int
3694 zone_alloc_limit_hard(uma_zone_t zone, int count, int flags)
3695 {
3696 	uint64_t old, new, total, max;
3697 
3698 	/*
3699 	 * The hard case.  We're going to sleep because there were existing
3700 	 * sleepers or because we ran out of items.  This routine enforces
3701 	 * fairness by keeping fifo order.
3702 	 *
3703 	 * First release our ill gotten gains and make some noise.
3704 	 */
3705 	for (;;) {
3706 		zone_free_limit(zone, count);
3707 		zone_log_warning(zone);
3708 		zone_maxaction(zone);
3709 		if (flags & M_NOWAIT)
3710 			return (0);
3711 
3712 		/*
3713 		 * We need to allocate an item or set ourself as a sleeper
3714 		 * while the sleepq lock is held to avoid wakeup races.  This
3715 		 * is essentially a home rolled semaphore.
3716 		 */
3717 		sleepq_lock(&zone->uz_max_items);
3718 		old = zone->uz_items;
3719 		do {
3720 			MPASS(UZ_ITEMS_SLEEPERS(old) < UZ_ITEMS_SLEEPERS_MAX);
3721 			/* Cache the max since we will evaluate twice. */
3722 			max = zone->uz_max_items;
3723 			if (UZ_ITEMS_SLEEPERS(old) != 0 ||
3724 			    UZ_ITEMS_COUNT(old) >= max)
3725 				new = old + UZ_ITEMS_SLEEPER;
3726 			else
3727 				new = old + MIN(count, max - old);
3728 		} while (atomic_fcmpset_64(&zone->uz_items, &old, new) == 0);
3729 
3730 		/* We may have successfully allocated under the sleepq lock. */
3731 		if (UZ_ITEMS_SLEEPERS(new) == 0) {
3732 			sleepq_release(&zone->uz_max_items);
3733 			return (new - old);
3734 		}
3735 
3736 		/*
3737 		 * This is in a different cacheline from uz_items so that we
3738 		 * don't constantly invalidate the fastpath cacheline when we
3739 		 * adjust item counts.  This could be limited to toggling on
3740 		 * transitions.
3741 		 */
3742 		atomic_add_32(&zone->uz_sleepers, 1);
3743 		atomic_add_64(&zone->uz_sleeps, 1);
3744 
3745 		/*
3746 		 * We have added ourselves as a sleeper.  The sleepq lock
3747 		 * protects us from wakeup races.  Sleep now and then retry.
3748 		 */
3749 		sleepq_add(&zone->uz_max_items, NULL, "zonelimit", 0, 0);
3750 		sleepq_wait(&zone->uz_max_items, PVM);
3751 
3752 		/*
3753 		 * After wakeup, remove ourselves as a sleeper and try
3754 		 * again.  We no longer have the sleepq lock for protection.
3755 		 *
3756 		 * Subract ourselves as a sleeper while attempting to add
3757 		 * our count.
3758 		 */
3759 		atomic_subtract_32(&zone->uz_sleepers, 1);
3760 		old = atomic_fetchadd_64(&zone->uz_items,
3761 		    -(UZ_ITEMS_SLEEPER - count));
3762 		/* We're no longer a sleeper. */
3763 		old -= UZ_ITEMS_SLEEPER;
3764 
3765 		/*
3766 		 * If we're still at the limit, restart.  Notably do not
3767 		 * block on other sleepers.  Cache the max value to protect
3768 		 * against changes via sysctl.
3769 		 */
3770 		total = UZ_ITEMS_COUNT(old);
3771 		max = zone->uz_max_items;
3772 		if (total >= max)
3773 			continue;
3774 		/* Truncate if necessary, otherwise wake other sleepers. */
3775 		if (total + count > max) {
3776 			zone_free_limit(zone, total + count - max);
3777 			count = max - total;
3778 		} else if (total + count < max && UZ_ITEMS_SLEEPERS(old) != 0)
3779 			wakeup_one(&zone->uz_max_items);
3780 
3781 		return (count);
3782 	}
3783 }
3784 
3785 /*
3786  * Allocate 'count' items from our max_items limit.  Returns the number
3787  * available.  If M_NOWAIT is not specified it will sleep until at least
3788  * one item can be allocated.
3789  */
3790 static int
3791 zone_alloc_limit(uma_zone_t zone, int count, int flags)
3792 {
3793 	uint64_t old;
3794 	uint64_t max;
3795 
3796 	max = zone->uz_max_items;
3797 	MPASS(max > 0);
3798 
3799 	/*
3800 	 * We expect normal allocations to succeed with a simple
3801 	 * fetchadd.
3802 	 */
3803 	old = atomic_fetchadd_64(&zone->uz_items, count);
3804 	if (__predict_true(old + count <= max))
3805 		return (count);
3806 
3807 	/*
3808 	 * If we had some items and no sleepers just return the
3809 	 * truncated value.  We have to release the excess space
3810 	 * though because that may wake sleepers who weren't woken
3811 	 * because we were temporarily over the limit.
3812 	 */
3813 	if (old < max) {
3814 		zone_free_limit(zone, (old + count) - max);
3815 		return (max - old);
3816 	}
3817 	return (zone_alloc_limit_hard(zone, count, flags));
3818 }
3819 
3820 /*
3821  * Free a number of items back to the limit.
3822  */
3823 static void
3824 zone_free_limit(uma_zone_t zone, int count)
3825 {
3826 	uint64_t old;
3827 
3828 	MPASS(count > 0);
3829 
3830 	/*
3831 	 * In the common case we either have no sleepers or
3832 	 * are still over the limit and can just return.
3833 	 */
3834 	old = atomic_fetchadd_64(&zone->uz_items, -count);
3835 	if (__predict_true(UZ_ITEMS_SLEEPERS(old) == 0 ||
3836 	   UZ_ITEMS_COUNT(old) - count >= zone->uz_max_items))
3837 		return;
3838 
3839 	/*
3840 	 * Moderate the rate of wakeups.  Sleepers will continue
3841 	 * to generate wakeups if necessary.
3842 	 */
3843 	wakeup_one(&zone->uz_max_items);
3844 }
3845 
3846 static uma_bucket_t
3847 zone_alloc_bucket(uma_zone_t zone, void *udata, int domain, int flags)
3848 {
3849 	uma_bucket_t bucket;
3850 	int maxbucket, cnt;
3851 
3852 	CTR3(KTR_UMA, "zone_alloc_bucket zone %s(%p) domain %d", zone->uz_name,
3853 	    zone, domain);
3854 
3855 	/* Avoid allocs targeting empty domains. */
3856 	if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
3857 		domain = UMA_ANYDOMAIN;
3858 	if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
3859 		domain = UMA_ANYDOMAIN;
3860 
3861 	if (zone->uz_max_items > 0)
3862 		maxbucket = zone_alloc_limit(zone, zone->uz_bucket_size,
3863 		    M_NOWAIT);
3864 	else
3865 		maxbucket = zone->uz_bucket_size;
3866 	if (maxbucket == 0)
3867 		return (false);
3868 
3869 	/* Don't wait for buckets, preserve caller's NOVM setting. */
3870 	bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
3871 	if (bucket == NULL) {
3872 		cnt = 0;
3873 		goto out;
3874 	}
3875 
3876 	bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
3877 	    MIN(maxbucket, bucket->ub_entries), domain, flags);
3878 
3879 	/*
3880 	 * Initialize the memory if necessary.
3881 	 */
3882 	if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
3883 		int i;
3884 
3885 		for (i = 0; i < bucket->ub_cnt; i++)
3886 			if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size,
3887 			    flags) != 0)
3888 				break;
3889 		/*
3890 		 * If we couldn't initialize the whole bucket, put the
3891 		 * rest back onto the freelist.
3892 		 */
3893 		if (i != bucket->ub_cnt) {
3894 			zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
3895 			    bucket->ub_cnt - i);
3896 #ifdef INVARIANTS
3897 			bzero(&bucket->ub_bucket[i],
3898 			    sizeof(void *) * (bucket->ub_cnt - i));
3899 #endif
3900 			bucket->ub_cnt = i;
3901 		}
3902 	}
3903 
3904 	cnt = bucket->ub_cnt;
3905 	if (bucket->ub_cnt == 0) {
3906 		bucket_free(zone, bucket, udata);
3907 		counter_u64_add(zone->uz_fails, 1);
3908 		bucket = NULL;
3909 	}
3910 out:
3911 	if (zone->uz_max_items > 0 && cnt < maxbucket)
3912 		zone_free_limit(zone, maxbucket - cnt);
3913 
3914 	return (bucket);
3915 }
3916 
3917 /*
3918  * Allocates a single item from a zone.
3919  *
3920  * Arguments
3921  *	zone   The zone to alloc for.
3922  *	udata  The data to be passed to the constructor.
3923  *	domain The domain to allocate from or UMA_ANYDOMAIN.
3924  *	flags  M_WAITOK, M_NOWAIT, M_ZERO.
3925  *
3926  * Returns
3927  *	NULL if there is no memory and M_NOWAIT is set
3928  *	An item if successful
3929  */
3930 
3931 static void *
3932 zone_alloc_item(uma_zone_t zone, void *udata, int domain, int flags)
3933 {
3934 	void *item;
3935 
3936 	if (zone->uz_max_items > 0 && zone_alloc_limit(zone, 1, flags) == 0)
3937 		return (NULL);
3938 
3939 	/* Avoid allocs targeting empty domains. */
3940 	if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
3941 		domain = UMA_ANYDOMAIN;
3942 
3943 	if (zone->uz_import(zone->uz_arg, &item, 1, domain, flags) != 1)
3944 		goto fail_cnt;
3945 
3946 	/*
3947 	 * We have to call both the zone's init (not the keg's init)
3948 	 * and the zone's ctor.  This is because the item is going from
3949 	 * a keg slab directly to the user, and the user is expecting it
3950 	 * to be both zone-init'd as well as zone-ctor'd.
3951 	 */
3952 	if (zone->uz_init != NULL) {
3953 		if (zone->uz_init(item, zone->uz_size, flags) != 0) {
3954 			zone_free_item(zone, item, udata, SKIP_FINI | SKIP_CNT);
3955 			goto fail_cnt;
3956 		}
3957 	}
3958 	item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata, flags,
3959 	    item);
3960 	if (item == NULL)
3961 		goto fail;
3962 
3963 	counter_u64_add(zone->uz_allocs, 1);
3964 	CTR3(KTR_UMA, "zone_alloc_item item %p from %s(%p)", item,
3965 	    zone->uz_name, zone);
3966 
3967 	return (item);
3968 
3969 fail_cnt:
3970 	counter_u64_add(zone->uz_fails, 1);
3971 fail:
3972 	if (zone->uz_max_items > 0)
3973 		zone_free_limit(zone, 1);
3974 	CTR2(KTR_UMA, "zone_alloc_item failed from %s(%p)",
3975 	    zone->uz_name, zone);
3976 
3977 	return (NULL);
3978 }
3979 
3980 /* See uma.h */
3981 void
3982 uma_zfree_smr(uma_zone_t zone, void *item)
3983 {
3984 	uma_cache_t cache;
3985 	uma_cache_bucket_t bucket;
3986 	int itemdomain, uz_flags;
3987 
3988 #ifdef UMA_ZALLOC_DEBUG
3989 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
3990 	    ("uma_zfree_smr: called with non-SMR zone.\n"));
3991 	KASSERT(item != NULL, ("uma_zfree_smr: Called with NULL pointer."));
3992 	SMR_ASSERT_NOT_ENTERED(zone->uz_smr);
3993 	if (uma_zfree_debug(zone, item, NULL) == EJUSTRETURN)
3994 		return;
3995 #endif
3996 	cache = &zone->uz_cpu[curcpu];
3997 	uz_flags = cache_uz_flags(cache);
3998 	itemdomain = 0;
3999 #ifdef NUMA
4000 	if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4001 		itemdomain = _vm_phys_domain(pmap_kextract((vm_offset_t)item));
4002 #endif
4003 	critical_enter();
4004 	do {
4005 		cache = &zone->uz_cpu[curcpu];
4006 		/* SMR Zones must free to the free bucket. */
4007 		bucket = &cache->uc_freebucket;
4008 #ifdef NUMA
4009 		if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4010 		    PCPU_GET(domain) != itemdomain) {
4011 			bucket = &cache->uc_crossbucket;
4012 		}
4013 #endif
4014 		if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4015 			cache_bucket_push(cache, bucket, item);
4016 			critical_exit();
4017 			return;
4018 		}
4019 	} while (cache_free(zone, cache, NULL, item, itemdomain));
4020 	critical_exit();
4021 
4022 	/*
4023 	 * If nothing else caught this, we'll just do an internal free.
4024 	 */
4025 	zone_free_item(zone, item, NULL, SKIP_NONE);
4026 }
4027 
4028 /* See uma.h */
4029 void
4030 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
4031 {
4032 	uma_cache_t cache;
4033 	uma_cache_bucket_t bucket;
4034 	int itemdomain, uz_flags;
4035 
4036 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
4037 	random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
4038 
4039 	CTR2(KTR_UMA, "uma_zfree_arg zone %s(%p)", zone->uz_name, zone);
4040 
4041 #ifdef UMA_ZALLOC_DEBUG
4042 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
4043 	    ("uma_zfree_arg: called with SMR zone.\n"));
4044 	if (uma_zfree_debug(zone, item, udata) == EJUSTRETURN)
4045 		return;
4046 #endif
4047         /* uma_zfree(..., NULL) does nothing, to match free(9). */
4048         if (item == NULL)
4049                 return;
4050 
4051 	/*
4052 	 * We are accessing the per-cpu cache without a critical section to
4053 	 * fetch size and flags.  This is acceptable, if we are preempted we
4054 	 * will simply read another cpu's line.
4055 	 */
4056 	cache = &zone->uz_cpu[curcpu];
4057 	uz_flags = cache_uz_flags(cache);
4058 	if (UMA_ALWAYS_CTORDTOR ||
4059 	    __predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0))
4060 		item_dtor(zone, item, cache_uz_size(cache), udata, SKIP_NONE);
4061 
4062 	/*
4063 	 * The race here is acceptable.  If we miss it we'll just have to wait
4064 	 * a little longer for the limits to be reset.
4065 	 */
4066 	if (__predict_false(uz_flags & UMA_ZFLAG_LIMIT)) {
4067 		if (zone->uz_sleepers > 0)
4068 			goto zfree_item;
4069 	}
4070 
4071 	/*
4072 	 * If possible, free to the per-CPU cache.  There are two
4073 	 * requirements for safe access to the per-CPU cache: (1) the thread
4074 	 * accessing the cache must not be preempted or yield during access,
4075 	 * and (2) the thread must not migrate CPUs without switching which
4076 	 * cache it accesses.  We rely on a critical section to prevent
4077 	 * preemption and migration.  We release the critical section in
4078 	 * order to acquire the zone mutex if we are unable to free to the
4079 	 * current cache; when we re-acquire the critical section, we must
4080 	 * detect and handle migration if it has occurred.
4081 	 */
4082 	itemdomain = 0;
4083 #ifdef NUMA
4084 	if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4085 		itemdomain = _vm_phys_domain(pmap_kextract((vm_offset_t)item));
4086 #endif
4087 	critical_enter();
4088 	do {
4089 		cache = &zone->uz_cpu[curcpu];
4090 		/*
4091 		 * Try to free into the allocbucket first to give LIFO
4092 		 * ordering for cache-hot datastructures.  Spill over
4093 		 * into the freebucket if necessary.  Alloc will swap
4094 		 * them if one runs dry.
4095 		 */
4096 		bucket = &cache->uc_allocbucket;
4097 #ifdef NUMA
4098 		if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4099 		    PCPU_GET(domain) != itemdomain) {
4100 			bucket = &cache->uc_crossbucket;
4101 		} else
4102 #endif
4103 		if (bucket->ucb_cnt >= bucket->ucb_entries)
4104 			bucket = &cache->uc_freebucket;
4105 		if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4106 			cache_bucket_push(cache, bucket, item);
4107 			critical_exit();
4108 			return;
4109 		}
4110 	} while (cache_free(zone, cache, udata, item, itemdomain));
4111 	critical_exit();
4112 
4113 	/*
4114 	 * If nothing else caught this, we'll just do an internal free.
4115 	 */
4116 zfree_item:
4117 	zone_free_item(zone, item, udata, SKIP_DTOR);
4118 }
4119 
4120 #ifdef NUMA
4121 /*
4122  * sort crossdomain free buckets to domain correct buckets and cache
4123  * them.
4124  */
4125 static void
4126 zone_free_cross(uma_zone_t zone, uma_bucket_t bucket, void *udata)
4127 {
4128 	struct uma_bucketlist fullbuckets;
4129 	uma_zone_domain_t zdom;
4130 	uma_bucket_t b;
4131 	smr_seq_t seq;
4132 	void *item;
4133 	int domain;
4134 
4135 	CTR3(KTR_UMA,
4136 	    "uma_zfree: zone %s(%p) draining cross bucket %p",
4137 	    zone->uz_name, zone, bucket);
4138 
4139 	/*
4140 	 * It is possible for buckets to arrive here out of order so we fetch
4141 	 * the current smr seq rather than accepting the bucket's.
4142 	 */
4143 	seq = SMR_SEQ_INVALID;
4144 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
4145 		seq = smr_advance(zone->uz_smr);
4146 
4147 	/*
4148 	 * To avoid having ndomain * ndomain buckets for sorting we have a
4149 	 * lock on the current crossfree bucket.  A full matrix with
4150 	 * per-domain locking could be used if necessary.
4151 	 */
4152 	STAILQ_INIT(&fullbuckets);
4153 	ZONE_CROSS_LOCK(zone);
4154 	while (bucket->ub_cnt > 0) {
4155 		item = bucket->ub_bucket[bucket->ub_cnt - 1];
4156 		domain = _vm_phys_domain(pmap_kextract((vm_offset_t)item));
4157 		zdom = ZDOM_GET(zone, domain);
4158 		if (zdom->uzd_cross == NULL) {
4159 			zdom->uzd_cross = bucket_alloc(zone, udata, M_NOWAIT);
4160 			if (zdom->uzd_cross == NULL)
4161 				break;
4162 		}
4163 		b = zdom->uzd_cross;
4164 		b->ub_bucket[b->ub_cnt++] = item;
4165 		b->ub_seq = seq;
4166 		if (b->ub_cnt == b->ub_entries) {
4167 			STAILQ_INSERT_HEAD(&fullbuckets, b, ub_link);
4168 			zdom->uzd_cross = NULL;
4169 		}
4170 		bucket->ub_cnt--;
4171 	}
4172 	ZONE_CROSS_UNLOCK(zone);
4173 	if (bucket->ub_cnt == 0)
4174 		bucket->ub_seq = SMR_SEQ_INVALID;
4175 	bucket_free(zone, bucket, udata);
4176 
4177 	while ((b = STAILQ_FIRST(&fullbuckets)) != NULL) {
4178 		STAILQ_REMOVE_HEAD(&fullbuckets, ub_link);
4179 		domain = _vm_phys_domain(pmap_kextract(
4180 		    (vm_offset_t)b->ub_bucket[0]));
4181 		zone_put_bucket(zone, domain, b, udata, true);
4182 	}
4183 }
4184 #endif
4185 
4186 static void
4187 zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata,
4188     int itemdomain, bool ws)
4189 {
4190 
4191 #ifdef NUMA
4192 	/*
4193 	 * Buckets coming from the wrong domain will be entirely for the
4194 	 * only other domain on two domain systems.  In this case we can
4195 	 * simply cache them.  Otherwise we need to sort them back to
4196 	 * correct domains.
4197 	 */
4198 	if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4199 	    vm_ndomains > 2 && PCPU_GET(domain) != itemdomain) {
4200 		zone_free_cross(zone, bucket, udata);
4201 		return;
4202 	}
4203 #endif
4204 
4205 	/*
4206 	 * Attempt to save the bucket in the zone's domain bucket cache.
4207 	 */
4208 	CTR3(KTR_UMA,
4209 	    "uma_zfree: zone %s(%p) putting bucket %p on free list",
4210 	    zone->uz_name, zone, bucket);
4211 	/* ub_cnt is pointing to the last free item */
4212 	if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
4213 		itemdomain = zone_domain_lowest(zone, itemdomain);
4214 	zone_put_bucket(zone, itemdomain, bucket, udata, ws);
4215 }
4216 
4217 /*
4218  * Populate a free or cross bucket for the current cpu cache.  Free any
4219  * existing full bucket either to the zone cache or back to the slab layer.
4220  *
4221  * Enters and returns in a critical section.  false return indicates that
4222  * we can not satisfy this free in the cache layer.  true indicates that
4223  * the caller should retry.
4224  */
4225 static __noinline bool
4226 cache_free(uma_zone_t zone, uma_cache_t cache, void *udata, void *item,
4227     int itemdomain)
4228 {
4229 	uma_cache_bucket_t cbucket;
4230 	uma_bucket_t newbucket, bucket;
4231 
4232 	CRITICAL_ASSERT(curthread);
4233 
4234 	if (zone->uz_bucket_size == 0)
4235 		return false;
4236 
4237 	cache = &zone->uz_cpu[curcpu];
4238 	newbucket = NULL;
4239 
4240 	/*
4241 	 * FIRSTTOUCH domains need to free to the correct zdom.  When
4242 	 * enabled this is the zdom of the item.   The bucket is the
4243 	 * cross bucket if the current domain and itemdomain do not match.
4244 	 */
4245 	cbucket = &cache->uc_freebucket;
4246 #ifdef NUMA
4247 	if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4248 		if (PCPU_GET(domain) != itemdomain) {
4249 			cbucket = &cache->uc_crossbucket;
4250 			if (cbucket->ucb_cnt != 0)
4251 				counter_u64_add(zone->uz_xdomain,
4252 				    cbucket->ucb_cnt);
4253 		}
4254 	}
4255 #endif
4256 	bucket = cache_bucket_unload(cbucket);
4257 	KASSERT(bucket == NULL || bucket->ub_cnt == bucket->ub_entries,
4258 	    ("cache_free: Entered with non-full free bucket."));
4259 
4260 	/* We are no longer associated with this CPU. */
4261 	critical_exit();
4262 
4263 	/*
4264 	 * Don't let SMR zones operate without a free bucket.  Force
4265 	 * a synchronize and re-use this one.  We will only degrade
4266 	 * to a synchronize every bucket_size items rather than every
4267 	 * item if we fail to allocate a bucket.
4268 	 */
4269 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0) {
4270 		if (bucket != NULL)
4271 			bucket->ub_seq = smr_advance(zone->uz_smr);
4272 		newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4273 		if (newbucket == NULL && bucket != NULL) {
4274 			bucket_drain(zone, bucket);
4275 			newbucket = bucket;
4276 			bucket = NULL;
4277 		}
4278 	} else if (!bucketdisable)
4279 		newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4280 
4281 	if (bucket != NULL)
4282 		zone_free_bucket(zone, bucket, udata, itemdomain, true);
4283 
4284 	critical_enter();
4285 	if ((bucket = newbucket) == NULL)
4286 		return (false);
4287 	cache = &zone->uz_cpu[curcpu];
4288 #ifdef NUMA
4289 	/*
4290 	 * Check to see if we should be populating the cross bucket.  If it
4291 	 * is already populated we will fall through and attempt to populate
4292 	 * the free bucket.
4293 	 */
4294 	if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4295 		if (PCPU_GET(domain) != itemdomain &&
4296 		    cache->uc_crossbucket.ucb_bucket == NULL) {
4297 			cache_bucket_load_cross(cache, bucket);
4298 			return (true);
4299 		}
4300 	}
4301 #endif
4302 	/*
4303 	 * We may have lost the race to fill the bucket or switched CPUs.
4304 	 */
4305 	if (cache->uc_freebucket.ucb_bucket != NULL) {
4306 		critical_exit();
4307 		bucket_free(zone, bucket, udata);
4308 		critical_enter();
4309 	} else
4310 		cache_bucket_load_free(cache, bucket);
4311 
4312 	return (true);
4313 }
4314 
4315 void
4316 uma_zfree_domain(uma_zone_t zone, void *item, void *udata)
4317 {
4318 
4319 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
4320 	random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
4321 
4322 	CTR2(KTR_UMA, "uma_zfree_domain zone %s(%p)", zone->uz_name, zone);
4323 
4324 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
4325 	    ("uma_zfree_domain: called with spinlock or critical section held"));
4326 
4327         /* uma_zfree(..., NULL) does nothing, to match free(9). */
4328         if (item == NULL)
4329                 return;
4330 	zone_free_item(zone, item, udata, SKIP_NONE);
4331 }
4332 
4333 static void
4334 slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item)
4335 {
4336 	uma_keg_t keg;
4337 	uma_domain_t dom;
4338 	int freei;
4339 
4340 	keg = zone->uz_keg;
4341 	KEG_LOCK_ASSERT(keg, slab->us_domain);
4342 
4343 	/* Do we need to remove from any lists? */
4344 	dom = &keg->uk_domain[slab->us_domain];
4345 	if (slab->us_freecount + 1 == keg->uk_ipers) {
4346 		LIST_REMOVE(slab, us_link);
4347 		LIST_INSERT_HEAD(&dom->ud_free_slab, slab, us_link);
4348 		dom->ud_free_slabs++;
4349 	} else if (slab->us_freecount == 0) {
4350 		LIST_REMOVE(slab, us_link);
4351 		LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
4352 	}
4353 
4354 	/* Slab management. */
4355 	freei = slab_item_index(slab, keg, item);
4356 	BIT_SET(keg->uk_ipers, freei, &slab->us_free);
4357 	slab->us_freecount++;
4358 
4359 	/* Keg statistics. */
4360 	dom->ud_free_items++;
4361 }
4362 
4363 static void
4364 zone_release(void *arg, void **bucket, int cnt)
4365 {
4366 	struct mtx *lock;
4367 	uma_zone_t zone;
4368 	uma_slab_t slab;
4369 	uma_keg_t keg;
4370 	uint8_t *mem;
4371 	void *item;
4372 	int i;
4373 
4374 	zone = arg;
4375 	keg = zone->uz_keg;
4376 	lock = NULL;
4377 	if (__predict_false((zone->uz_flags & UMA_ZFLAG_HASH) != 0))
4378 		lock = KEG_LOCK(keg, 0);
4379 	for (i = 0; i < cnt; i++) {
4380 		item = bucket[i];
4381 		if (__predict_true((zone->uz_flags & UMA_ZFLAG_VTOSLAB) != 0)) {
4382 			slab = vtoslab((vm_offset_t)item);
4383 		} else {
4384 			mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
4385 			if ((zone->uz_flags & UMA_ZFLAG_HASH) != 0)
4386 				slab = hash_sfind(&keg->uk_hash, mem);
4387 			else
4388 				slab = (uma_slab_t)(mem + keg->uk_pgoff);
4389 		}
4390 		if (lock != KEG_LOCKPTR(keg, slab->us_domain)) {
4391 			if (lock != NULL)
4392 				mtx_unlock(lock);
4393 			lock = KEG_LOCK(keg, slab->us_domain);
4394 		}
4395 		slab_free_item(zone, slab, item);
4396 	}
4397 	if (lock != NULL)
4398 		mtx_unlock(lock);
4399 }
4400 
4401 /*
4402  * Frees a single item to any zone.
4403  *
4404  * Arguments:
4405  *	zone   The zone to free to
4406  *	item   The item we're freeing
4407  *	udata  User supplied data for the dtor
4408  *	skip   Skip dtors and finis
4409  */
4410 static __noinline void
4411 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
4412 {
4413 
4414 	/*
4415 	 * If a free is sent directly to an SMR zone we have to
4416 	 * synchronize immediately because the item can instantly
4417 	 * be reallocated. This should only happen in degenerate
4418 	 * cases when no memory is available for per-cpu caches.
4419 	 */
4420 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0 && skip == SKIP_NONE)
4421 		smr_synchronize(zone->uz_smr);
4422 
4423 	item_dtor(zone, item, zone->uz_size, udata, skip);
4424 
4425 	if (skip < SKIP_FINI && zone->uz_fini)
4426 		zone->uz_fini(item, zone->uz_size);
4427 
4428 	zone->uz_release(zone->uz_arg, &item, 1);
4429 
4430 	if (skip & SKIP_CNT)
4431 		return;
4432 
4433 	counter_u64_add(zone->uz_frees, 1);
4434 
4435 	if (zone->uz_max_items > 0)
4436 		zone_free_limit(zone, 1);
4437 }
4438 
4439 /* See uma.h */
4440 int
4441 uma_zone_set_max(uma_zone_t zone, int nitems)
4442 {
4443 	struct uma_bucket_zone *ubz;
4444 	int count;
4445 
4446 	/*
4447 	 * XXX This can misbehave if the zone has any allocations with
4448 	 * no limit and a limit is imposed.  There is currently no
4449 	 * way to clear a limit.
4450 	 */
4451 	ZONE_LOCK(zone);
4452 	ubz = bucket_zone_max(zone, nitems);
4453 	count = ubz != NULL ? ubz->ubz_entries : 0;
4454 	zone->uz_bucket_size_max = zone->uz_bucket_size = count;
4455 	if (zone->uz_bucket_size_min > zone->uz_bucket_size_max)
4456 		zone->uz_bucket_size_min = zone->uz_bucket_size_max;
4457 	zone->uz_max_items = nitems;
4458 	zone->uz_flags |= UMA_ZFLAG_LIMIT;
4459 	zone_update_caches(zone);
4460 	/* We may need to wake waiters. */
4461 	wakeup(&zone->uz_max_items);
4462 	ZONE_UNLOCK(zone);
4463 
4464 	return (nitems);
4465 }
4466 
4467 /* See uma.h */
4468 void
4469 uma_zone_set_maxcache(uma_zone_t zone, int nitems)
4470 {
4471 	struct uma_bucket_zone *ubz;
4472 	int bpcpu;
4473 
4474 	ZONE_LOCK(zone);
4475 	ubz = bucket_zone_max(zone, nitems);
4476 	if (ubz != NULL) {
4477 		bpcpu = 2;
4478 		if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4479 			/* Count the cross-domain bucket. */
4480 			bpcpu++;
4481 		nitems -= ubz->ubz_entries * bpcpu * mp_ncpus;
4482 		zone->uz_bucket_size_max = ubz->ubz_entries;
4483 	} else {
4484 		zone->uz_bucket_size_max = zone->uz_bucket_size = 0;
4485 	}
4486 	if (zone->uz_bucket_size_min > zone->uz_bucket_size_max)
4487 		zone->uz_bucket_size_min = zone->uz_bucket_size_max;
4488 	zone->uz_bucket_max = nitems / vm_ndomains;
4489 	ZONE_UNLOCK(zone);
4490 }
4491 
4492 /* See uma.h */
4493 int
4494 uma_zone_get_max(uma_zone_t zone)
4495 {
4496 	int nitems;
4497 
4498 	nitems = atomic_load_64(&zone->uz_max_items);
4499 
4500 	return (nitems);
4501 }
4502 
4503 /* See uma.h */
4504 void
4505 uma_zone_set_warning(uma_zone_t zone, const char *warning)
4506 {
4507 
4508 	ZONE_ASSERT_COLD(zone);
4509 	zone->uz_warning = warning;
4510 }
4511 
4512 /* See uma.h */
4513 void
4514 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction)
4515 {
4516 
4517 	ZONE_ASSERT_COLD(zone);
4518 	TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone);
4519 }
4520 
4521 /* See uma.h */
4522 int
4523 uma_zone_get_cur(uma_zone_t zone)
4524 {
4525 	int64_t nitems;
4526 	u_int i;
4527 
4528 	nitems = 0;
4529 	if (zone->uz_allocs != EARLY_COUNTER && zone->uz_frees != EARLY_COUNTER)
4530 		nitems = counter_u64_fetch(zone->uz_allocs) -
4531 		    counter_u64_fetch(zone->uz_frees);
4532 	CPU_FOREACH(i)
4533 		nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs) -
4534 		    atomic_load_64(&zone->uz_cpu[i].uc_frees);
4535 
4536 	return (nitems < 0 ? 0 : nitems);
4537 }
4538 
4539 static uint64_t
4540 uma_zone_get_allocs(uma_zone_t zone)
4541 {
4542 	uint64_t nitems;
4543 	u_int i;
4544 
4545 	nitems = 0;
4546 	if (zone->uz_allocs != EARLY_COUNTER)
4547 		nitems = counter_u64_fetch(zone->uz_allocs);
4548 	CPU_FOREACH(i)
4549 		nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs);
4550 
4551 	return (nitems);
4552 }
4553 
4554 static uint64_t
4555 uma_zone_get_frees(uma_zone_t zone)
4556 {
4557 	uint64_t nitems;
4558 	u_int i;
4559 
4560 	nitems = 0;
4561 	if (zone->uz_frees != EARLY_COUNTER)
4562 		nitems = counter_u64_fetch(zone->uz_frees);
4563 	CPU_FOREACH(i)
4564 		nitems += atomic_load_64(&zone->uz_cpu[i].uc_frees);
4565 
4566 	return (nitems);
4567 }
4568 
4569 #ifdef INVARIANTS
4570 /* Used only for KEG_ASSERT_COLD(). */
4571 static uint64_t
4572 uma_keg_get_allocs(uma_keg_t keg)
4573 {
4574 	uma_zone_t z;
4575 	uint64_t nitems;
4576 
4577 	nitems = 0;
4578 	LIST_FOREACH(z, &keg->uk_zones, uz_link)
4579 		nitems += uma_zone_get_allocs(z);
4580 
4581 	return (nitems);
4582 }
4583 #endif
4584 
4585 /* See uma.h */
4586 void
4587 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
4588 {
4589 	uma_keg_t keg;
4590 
4591 	KEG_GET(zone, keg);
4592 	KEG_ASSERT_COLD(keg);
4593 	keg->uk_init = uminit;
4594 }
4595 
4596 /* See uma.h */
4597 void
4598 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
4599 {
4600 	uma_keg_t keg;
4601 
4602 	KEG_GET(zone, keg);
4603 	KEG_ASSERT_COLD(keg);
4604 	keg->uk_fini = fini;
4605 }
4606 
4607 /* See uma.h */
4608 void
4609 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
4610 {
4611 
4612 	ZONE_ASSERT_COLD(zone);
4613 	zone->uz_init = zinit;
4614 }
4615 
4616 /* See uma.h */
4617 void
4618 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
4619 {
4620 
4621 	ZONE_ASSERT_COLD(zone);
4622 	zone->uz_fini = zfini;
4623 }
4624 
4625 /* See uma.h */
4626 void
4627 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
4628 {
4629 	uma_keg_t keg;
4630 
4631 	KEG_GET(zone, keg);
4632 	KEG_ASSERT_COLD(keg);
4633 	keg->uk_freef = freef;
4634 }
4635 
4636 /* See uma.h */
4637 void
4638 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
4639 {
4640 	uma_keg_t keg;
4641 
4642 	KEG_GET(zone, keg);
4643 	KEG_ASSERT_COLD(keg);
4644 	keg->uk_allocf = allocf;
4645 }
4646 
4647 /* See uma.h */
4648 void
4649 uma_zone_set_smr(uma_zone_t zone, smr_t smr)
4650 {
4651 
4652 	ZONE_ASSERT_COLD(zone);
4653 
4654 	zone->uz_flags |= UMA_ZONE_SMR;
4655 	zone->uz_smr = smr;
4656 	zone_update_caches(zone);
4657 }
4658 
4659 smr_t
4660 uma_zone_get_smr(uma_zone_t zone)
4661 {
4662 
4663 	return (zone->uz_smr);
4664 }
4665 
4666 /* See uma.h */
4667 void
4668 uma_zone_reserve(uma_zone_t zone, int items)
4669 {
4670 	uma_keg_t keg;
4671 
4672 	KEG_GET(zone, keg);
4673 	KEG_ASSERT_COLD(keg);
4674 	keg->uk_reserve = items;
4675 }
4676 
4677 /* See uma.h */
4678 int
4679 uma_zone_reserve_kva(uma_zone_t zone, int count)
4680 {
4681 	uma_keg_t keg;
4682 	vm_offset_t kva;
4683 	u_int pages;
4684 
4685 	KEG_GET(zone, keg);
4686 	KEG_ASSERT_COLD(keg);
4687 	ZONE_ASSERT_COLD(zone);
4688 
4689 	pages = howmany(count, keg->uk_ipers) * keg->uk_ppera;
4690 
4691 #ifdef UMA_MD_SMALL_ALLOC
4692 	if (keg->uk_ppera > 1) {
4693 #else
4694 	if (1) {
4695 #endif
4696 		kva = kva_alloc((vm_size_t)pages * PAGE_SIZE);
4697 		if (kva == 0)
4698 			return (0);
4699 	} else
4700 		kva = 0;
4701 
4702 	MPASS(keg->uk_kva == 0);
4703 	keg->uk_kva = kva;
4704 	keg->uk_offset = 0;
4705 	zone->uz_max_items = pages * keg->uk_ipers;
4706 #ifdef UMA_MD_SMALL_ALLOC
4707 	keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
4708 #else
4709 	keg->uk_allocf = noobj_alloc;
4710 #endif
4711 	keg->uk_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
4712 	zone->uz_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
4713 	zone_update_caches(zone);
4714 
4715 	return (1);
4716 }
4717 
4718 /* See uma.h */
4719 void
4720 uma_prealloc(uma_zone_t zone, int items)
4721 {
4722 	struct vm_domainset_iter di;
4723 	uma_domain_t dom;
4724 	uma_slab_t slab;
4725 	uma_keg_t keg;
4726 	int aflags, domain, slabs;
4727 
4728 	KEG_GET(zone, keg);
4729 	slabs = howmany(items, keg->uk_ipers);
4730 	while (slabs-- > 0) {
4731 		aflags = M_NOWAIT;
4732 		vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
4733 		    &aflags);
4734 		for (;;) {
4735 			slab = keg_alloc_slab(keg, zone, domain, M_WAITOK,
4736 			    aflags);
4737 			if (slab != NULL) {
4738 				dom = &keg->uk_domain[slab->us_domain];
4739 				/*
4740 				 * keg_alloc_slab() always returns a slab on the
4741 				 * partial list.
4742 				 */
4743 				LIST_REMOVE(slab, us_link);
4744 				LIST_INSERT_HEAD(&dom->ud_free_slab, slab,
4745 				    us_link);
4746 				dom->ud_free_slabs++;
4747 				KEG_UNLOCK(keg, slab->us_domain);
4748 				break;
4749 			}
4750 			if (vm_domainset_iter_policy(&di, &domain) != 0)
4751 				vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask);
4752 		}
4753 	}
4754 }
4755 
4756 /*
4757  * Returns a snapshot of memory consumption in bytes.
4758  */
4759 size_t
4760 uma_zone_memory(uma_zone_t zone)
4761 {
4762 	size_t sz;
4763 	int i;
4764 
4765 	sz = 0;
4766 	if (zone->uz_flags & UMA_ZFLAG_CACHE) {
4767 		for (i = 0; i < vm_ndomains; i++)
4768 			sz += ZDOM_GET(zone, i)->uzd_nitems;
4769 		return (sz * zone->uz_size);
4770 	}
4771 	for (i = 0; i < vm_ndomains; i++)
4772 		sz += zone->uz_keg->uk_domain[i].ud_pages;
4773 
4774 	return (sz * PAGE_SIZE);
4775 }
4776 
4777 /* See uma.h */
4778 void
4779 uma_reclaim(int req)
4780 {
4781 
4782 	CTR0(KTR_UMA, "UMA: vm asked us to release pages!");
4783 	sx_xlock(&uma_reclaim_lock);
4784 	bucket_enable();
4785 
4786 	switch (req) {
4787 	case UMA_RECLAIM_TRIM:
4788 		zone_foreach(zone_trim, NULL);
4789 		break;
4790 	case UMA_RECLAIM_DRAIN:
4791 	case UMA_RECLAIM_DRAIN_CPU:
4792 		zone_foreach(zone_drain, NULL);
4793 		if (req == UMA_RECLAIM_DRAIN_CPU) {
4794 			pcpu_cache_drain_safe(NULL);
4795 			zone_foreach(zone_drain, NULL);
4796 		}
4797 		break;
4798 	default:
4799 		panic("unhandled reclamation request %d", req);
4800 	}
4801 
4802 	/*
4803 	 * Some slabs may have been freed but this zone will be visited early
4804 	 * we visit again so that we can free pages that are empty once other
4805 	 * zones are drained.  We have to do the same for buckets.
4806 	 */
4807 	zone_drain(slabzones[0], NULL);
4808 	zone_drain(slabzones[1], NULL);
4809 	bucket_zone_drain();
4810 	sx_xunlock(&uma_reclaim_lock);
4811 }
4812 
4813 static volatile int uma_reclaim_needed;
4814 
4815 void
4816 uma_reclaim_wakeup(void)
4817 {
4818 
4819 	if (atomic_fetchadd_int(&uma_reclaim_needed, 1) == 0)
4820 		wakeup(uma_reclaim);
4821 }
4822 
4823 void
4824 uma_reclaim_worker(void *arg __unused)
4825 {
4826 
4827 	for (;;) {
4828 		sx_xlock(&uma_reclaim_lock);
4829 		while (atomic_load_int(&uma_reclaim_needed) == 0)
4830 			sx_sleep(uma_reclaim, &uma_reclaim_lock, PVM, "umarcl",
4831 			    hz);
4832 		sx_xunlock(&uma_reclaim_lock);
4833 		EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM);
4834 		uma_reclaim(UMA_RECLAIM_DRAIN_CPU);
4835 		atomic_store_int(&uma_reclaim_needed, 0);
4836 		/* Don't fire more than once per-second. */
4837 		pause("umarclslp", hz);
4838 	}
4839 }
4840 
4841 /* See uma.h */
4842 void
4843 uma_zone_reclaim(uma_zone_t zone, int req)
4844 {
4845 
4846 	switch (req) {
4847 	case UMA_RECLAIM_TRIM:
4848 		zone_trim(zone, NULL);
4849 		break;
4850 	case UMA_RECLAIM_DRAIN:
4851 		zone_drain(zone, NULL);
4852 		break;
4853 	case UMA_RECLAIM_DRAIN_CPU:
4854 		pcpu_cache_drain_safe(zone);
4855 		zone_drain(zone, NULL);
4856 		break;
4857 	default:
4858 		panic("unhandled reclamation request %d", req);
4859 	}
4860 }
4861 
4862 /* See uma.h */
4863 int
4864 uma_zone_exhausted(uma_zone_t zone)
4865 {
4866 
4867 	return (atomic_load_32(&zone->uz_sleepers) > 0);
4868 }
4869 
4870 unsigned long
4871 uma_limit(void)
4872 {
4873 
4874 	return (uma_kmem_limit);
4875 }
4876 
4877 void
4878 uma_set_limit(unsigned long limit)
4879 {
4880 
4881 	uma_kmem_limit = limit;
4882 }
4883 
4884 unsigned long
4885 uma_size(void)
4886 {
4887 
4888 	return (atomic_load_long(&uma_kmem_total));
4889 }
4890 
4891 long
4892 uma_avail(void)
4893 {
4894 
4895 	return (uma_kmem_limit - uma_size());
4896 }
4897 
4898 #ifdef DDB
4899 /*
4900  * Generate statistics across both the zone and its per-cpu cache's.  Return
4901  * desired statistics if the pointer is non-NULL for that statistic.
4902  *
4903  * Note: does not update the zone statistics, as it can't safely clear the
4904  * per-CPU cache statistic.
4905  *
4906  */
4907 static void
4908 uma_zone_sumstat(uma_zone_t z, long *cachefreep, uint64_t *allocsp,
4909     uint64_t *freesp, uint64_t *sleepsp, uint64_t *xdomainp)
4910 {
4911 	uma_cache_t cache;
4912 	uint64_t allocs, frees, sleeps, xdomain;
4913 	int cachefree, cpu;
4914 
4915 	allocs = frees = sleeps = xdomain = 0;
4916 	cachefree = 0;
4917 	CPU_FOREACH(cpu) {
4918 		cache = &z->uz_cpu[cpu];
4919 		cachefree += cache->uc_allocbucket.ucb_cnt;
4920 		cachefree += cache->uc_freebucket.ucb_cnt;
4921 		xdomain += cache->uc_crossbucket.ucb_cnt;
4922 		cachefree += cache->uc_crossbucket.ucb_cnt;
4923 		allocs += cache->uc_allocs;
4924 		frees += cache->uc_frees;
4925 	}
4926 	allocs += counter_u64_fetch(z->uz_allocs);
4927 	frees += counter_u64_fetch(z->uz_frees);
4928 	xdomain += counter_u64_fetch(z->uz_xdomain);
4929 	sleeps += z->uz_sleeps;
4930 	if (cachefreep != NULL)
4931 		*cachefreep = cachefree;
4932 	if (allocsp != NULL)
4933 		*allocsp = allocs;
4934 	if (freesp != NULL)
4935 		*freesp = frees;
4936 	if (sleepsp != NULL)
4937 		*sleepsp = sleeps;
4938 	if (xdomainp != NULL)
4939 		*xdomainp = xdomain;
4940 }
4941 #endif /* DDB */
4942 
4943 static int
4944 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
4945 {
4946 	uma_keg_t kz;
4947 	uma_zone_t z;
4948 	int count;
4949 
4950 	count = 0;
4951 	rw_rlock(&uma_rwlock);
4952 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
4953 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
4954 			count++;
4955 	}
4956 	LIST_FOREACH(z, &uma_cachezones, uz_link)
4957 		count++;
4958 
4959 	rw_runlock(&uma_rwlock);
4960 	return (sysctl_handle_int(oidp, &count, 0, req));
4961 }
4962 
4963 static void
4964 uma_vm_zone_stats(struct uma_type_header *uth, uma_zone_t z, struct sbuf *sbuf,
4965     struct uma_percpu_stat *ups, bool internal)
4966 {
4967 	uma_zone_domain_t zdom;
4968 	uma_cache_t cache;
4969 	int i;
4970 
4971 
4972 	for (i = 0; i < vm_ndomains; i++) {
4973 		zdom = ZDOM_GET(z, i);
4974 		uth->uth_zone_free += zdom->uzd_nitems;
4975 	}
4976 	uth->uth_allocs = counter_u64_fetch(z->uz_allocs);
4977 	uth->uth_frees = counter_u64_fetch(z->uz_frees);
4978 	uth->uth_fails = counter_u64_fetch(z->uz_fails);
4979 	uth->uth_xdomain = counter_u64_fetch(z->uz_xdomain);
4980 	uth->uth_sleeps = z->uz_sleeps;
4981 
4982 	for (i = 0; i < mp_maxid + 1; i++) {
4983 		bzero(&ups[i], sizeof(*ups));
4984 		if (internal || CPU_ABSENT(i))
4985 			continue;
4986 		cache = &z->uz_cpu[i];
4987 		ups[i].ups_cache_free += cache->uc_allocbucket.ucb_cnt;
4988 		ups[i].ups_cache_free += cache->uc_freebucket.ucb_cnt;
4989 		ups[i].ups_cache_free += cache->uc_crossbucket.ucb_cnt;
4990 		ups[i].ups_allocs = cache->uc_allocs;
4991 		ups[i].ups_frees = cache->uc_frees;
4992 	}
4993 }
4994 
4995 static int
4996 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
4997 {
4998 	struct uma_stream_header ush;
4999 	struct uma_type_header uth;
5000 	struct uma_percpu_stat *ups;
5001 	struct sbuf sbuf;
5002 	uma_keg_t kz;
5003 	uma_zone_t z;
5004 	uint64_t items;
5005 	uint32_t kfree, pages;
5006 	int count, error, i;
5007 
5008 	error = sysctl_wire_old_buffer(req, 0);
5009 	if (error != 0)
5010 		return (error);
5011 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
5012 	sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL);
5013 	ups = malloc((mp_maxid + 1) * sizeof(*ups), M_TEMP, M_WAITOK);
5014 
5015 	count = 0;
5016 	rw_rlock(&uma_rwlock);
5017 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
5018 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
5019 			count++;
5020 	}
5021 
5022 	LIST_FOREACH(z, &uma_cachezones, uz_link)
5023 		count++;
5024 
5025 	/*
5026 	 * Insert stream header.
5027 	 */
5028 	bzero(&ush, sizeof(ush));
5029 	ush.ush_version = UMA_STREAM_VERSION;
5030 	ush.ush_maxcpus = (mp_maxid + 1);
5031 	ush.ush_count = count;
5032 	(void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
5033 
5034 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
5035 		kfree = pages = 0;
5036 		for (i = 0; i < vm_ndomains; i++) {
5037 			kfree += kz->uk_domain[i].ud_free_items;
5038 			pages += kz->uk_domain[i].ud_pages;
5039 		}
5040 		LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5041 			bzero(&uth, sizeof(uth));
5042 			strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5043 			uth.uth_align = kz->uk_align;
5044 			uth.uth_size = kz->uk_size;
5045 			uth.uth_rsize = kz->uk_rsize;
5046 			if (z->uz_max_items > 0) {
5047 				items = UZ_ITEMS_COUNT(z->uz_items);
5048 				uth.uth_pages = (items / kz->uk_ipers) *
5049 					kz->uk_ppera;
5050 			} else
5051 				uth.uth_pages = pages;
5052 			uth.uth_maxpages = (z->uz_max_items / kz->uk_ipers) *
5053 			    kz->uk_ppera;
5054 			uth.uth_limit = z->uz_max_items;
5055 			uth.uth_keg_free = kfree;
5056 
5057 			/*
5058 			 * A zone is secondary is it is not the first entry
5059 			 * on the keg's zone list.
5060 			 */
5061 			if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
5062 			    (LIST_FIRST(&kz->uk_zones) != z))
5063 				uth.uth_zone_flags = UTH_ZONE_SECONDARY;
5064 			uma_vm_zone_stats(&uth, z, &sbuf, ups,
5065 			    kz->uk_flags & UMA_ZFLAG_INTERNAL);
5066 			(void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5067 			for (i = 0; i < mp_maxid + 1; i++)
5068 				(void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5069 		}
5070 	}
5071 	LIST_FOREACH(z, &uma_cachezones, uz_link) {
5072 		bzero(&uth, sizeof(uth));
5073 		strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5074 		uth.uth_size = z->uz_size;
5075 		uma_vm_zone_stats(&uth, z, &sbuf, ups, false);
5076 		(void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5077 		for (i = 0; i < mp_maxid + 1; i++)
5078 			(void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5079 	}
5080 
5081 	rw_runlock(&uma_rwlock);
5082 	error = sbuf_finish(&sbuf);
5083 	sbuf_delete(&sbuf);
5084 	free(ups, M_TEMP);
5085 	return (error);
5086 }
5087 
5088 int
5089 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS)
5090 {
5091 	uma_zone_t zone = *(uma_zone_t *)arg1;
5092 	int error, max;
5093 
5094 	max = uma_zone_get_max(zone);
5095 	error = sysctl_handle_int(oidp, &max, 0, req);
5096 	if (error || !req->newptr)
5097 		return (error);
5098 
5099 	uma_zone_set_max(zone, max);
5100 
5101 	return (0);
5102 }
5103 
5104 int
5105 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS)
5106 {
5107 	uma_zone_t zone;
5108 	int cur;
5109 
5110 	/*
5111 	 * Some callers want to add sysctls for global zones that
5112 	 * may not yet exist so they pass a pointer to a pointer.
5113 	 */
5114 	if (arg2 == 0)
5115 		zone = *(uma_zone_t *)arg1;
5116 	else
5117 		zone = arg1;
5118 	cur = uma_zone_get_cur(zone);
5119 	return (sysctl_handle_int(oidp, &cur, 0, req));
5120 }
5121 
5122 static int
5123 sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS)
5124 {
5125 	uma_zone_t zone = arg1;
5126 	uint64_t cur;
5127 
5128 	cur = uma_zone_get_allocs(zone);
5129 	return (sysctl_handle_64(oidp, &cur, 0, req));
5130 }
5131 
5132 static int
5133 sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS)
5134 {
5135 	uma_zone_t zone = arg1;
5136 	uint64_t cur;
5137 
5138 	cur = uma_zone_get_frees(zone);
5139 	return (sysctl_handle_64(oidp, &cur, 0, req));
5140 }
5141 
5142 static int
5143 sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS)
5144 {
5145 	struct sbuf sbuf;
5146 	uma_zone_t zone = arg1;
5147 	int error;
5148 
5149 	sbuf_new_for_sysctl(&sbuf, NULL, 0, req);
5150 	if (zone->uz_flags != 0)
5151 		sbuf_printf(&sbuf, "0x%b", zone->uz_flags, PRINT_UMA_ZFLAGS);
5152 	else
5153 		sbuf_printf(&sbuf, "0");
5154 	error = sbuf_finish(&sbuf);
5155 	sbuf_delete(&sbuf);
5156 
5157 	return (error);
5158 }
5159 
5160 static int
5161 sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS)
5162 {
5163 	uma_keg_t keg = arg1;
5164 	int avail, effpct, total;
5165 
5166 	total = keg->uk_ppera * PAGE_SIZE;
5167 	if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0)
5168 		total += slabzone(keg->uk_ipers)->uz_keg->uk_rsize;
5169 	/*
5170 	 * We consider the client's requested size and alignment here, not the
5171 	 * real size determination uk_rsize, because we also adjust the real
5172 	 * size for internal implementation reasons (max bitset size).
5173 	 */
5174 	avail = keg->uk_ipers * roundup2(keg->uk_size, keg->uk_align + 1);
5175 	if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
5176 		avail *= mp_maxid + 1;
5177 	effpct = 100 * avail / total;
5178 	return (sysctl_handle_int(oidp, &effpct, 0, req));
5179 }
5180 
5181 static int
5182 sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS)
5183 {
5184 	uma_zone_t zone = arg1;
5185 	uint64_t cur;
5186 
5187 	cur = UZ_ITEMS_COUNT(atomic_load_64(&zone->uz_items));
5188 	return (sysctl_handle_64(oidp, &cur, 0, req));
5189 }
5190 
5191 #ifdef INVARIANTS
5192 static uma_slab_t
5193 uma_dbg_getslab(uma_zone_t zone, void *item)
5194 {
5195 	uma_slab_t slab;
5196 	uma_keg_t keg;
5197 	uint8_t *mem;
5198 
5199 	/*
5200 	 * It is safe to return the slab here even though the
5201 	 * zone is unlocked because the item's allocation state
5202 	 * essentially holds a reference.
5203 	 */
5204 	mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
5205 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5206 		return (NULL);
5207 	if (zone->uz_flags & UMA_ZFLAG_VTOSLAB)
5208 		return (vtoslab((vm_offset_t)mem));
5209 	keg = zone->uz_keg;
5210 	if ((keg->uk_flags & UMA_ZFLAG_HASH) == 0)
5211 		return ((uma_slab_t)(mem + keg->uk_pgoff));
5212 	KEG_LOCK(keg, 0);
5213 	slab = hash_sfind(&keg->uk_hash, mem);
5214 	KEG_UNLOCK(keg, 0);
5215 
5216 	return (slab);
5217 }
5218 
5219 static bool
5220 uma_dbg_zskip(uma_zone_t zone, void *mem)
5221 {
5222 
5223 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5224 		return (true);
5225 
5226 	return (uma_dbg_kskip(zone->uz_keg, mem));
5227 }
5228 
5229 static bool
5230 uma_dbg_kskip(uma_keg_t keg, void *mem)
5231 {
5232 	uintptr_t idx;
5233 
5234 	if (dbg_divisor == 0)
5235 		return (true);
5236 
5237 	if (dbg_divisor == 1)
5238 		return (false);
5239 
5240 	idx = (uintptr_t)mem >> PAGE_SHIFT;
5241 	if (keg->uk_ipers > 1) {
5242 		idx *= keg->uk_ipers;
5243 		idx += ((uintptr_t)mem & PAGE_MASK) / keg->uk_rsize;
5244 	}
5245 
5246 	if ((idx / dbg_divisor) * dbg_divisor != idx) {
5247 		counter_u64_add(uma_skip_cnt, 1);
5248 		return (true);
5249 	}
5250 	counter_u64_add(uma_dbg_cnt, 1);
5251 
5252 	return (false);
5253 }
5254 
5255 /*
5256  * Set up the slab's freei data such that uma_dbg_free can function.
5257  *
5258  */
5259 static void
5260 uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item)
5261 {
5262 	uma_keg_t keg;
5263 	int freei;
5264 
5265 	if (slab == NULL) {
5266 		slab = uma_dbg_getslab(zone, item);
5267 		if (slab == NULL)
5268 			panic("uma: item %p did not belong to zone %s\n",
5269 			    item, zone->uz_name);
5270 	}
5271 	keg = zone->uz_keg;
5272 	freei = slab_item_index(slab, keg, item);
5273 
5274 	if (BIT_ISSET(keg->uk_ipers, freei, slab_dbg_bits(slab, keg)))
5275 		panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)\n",
5276 		    item, zone, zone->uz_name, slab, freei);
5277 	BIT_SET_ATOMIC(keg->uk_ipers, freei, slab_dbg_bits(slab, keg));
5278 }
5279 
5280 /*
5281  * Verifies freed addresses.  Checks for alignment, valid slab membership
5282  * and duplicate frees.
5283  *
5284  */
5285 static void
5286 uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item)
5287 {
5288 	uma_keg_t keg;
5289 	int freei;
5290 
5291 	if (slab == NULL) {
5292 		slab = uma_dbg_getslab(zone, item);
5293 		if (slab == NULL)
5294 			panic("uma: Freed item %p did not belong to zone %s\n",
5295 			    item, zone->uz_name);
5296 	}
5297 	keg = zone->uz_keg;
5298 	freei = slab_item_index(slab, keg, item);
5299 
5300 	if (freei >= keg->uk_ipers)
5301 		panic("Invalid free of %p from zone %p(%s) slab %p(%d)\n",
5302 		    item, zone, zone->uz_name, slab, freei);
5303 
5304 	if (slab_item(slab, keg, freei) != item)
5305 		panic("Unaligned free of %p from zone %p(%s) slab %p(%d)\n",
5306 		    item, zone, zone->uz_name, slab, freei);
5307 
5308 	if (!BIT_ISSET(keg->uk_ipers, freei, slab_dbg_bits(slab, keg)))
5309 		panic("Duplicate free of %p from zone %p(%s) slab %p(%d)\n",
5310 		    item, zone, zone->uz_name, slab, freei);
5311 
5312 	BIT_CLR_ATOMIC(keg->uk_ipers, freei, slab_dbg_bits(slab, keg));
5313 }
5314 #endif /* INVARIANTS */
5315 
5316 #ifdef DDB
5317 static int64_t
5318 get_uma_stats(uma_keg_t kz, uma_zone_t z, uint64_t *allocs, uint64_t *used,
5319     uint64_t *sleeps, long *cachefree, uint64_t *xdomain)
5320 {
5321 	uint64_t frees;
5322 	int i;
5323 
5324 	if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
5325 		*allocs = counter_u64_fetch(z->uz_allocs);
5326 		frees = counter_u64_fetch(z->uz_frees);
5327 		*sleeps = z->uz_sleeps;
5328 		*cachefree = 0;
5329 		*xdomain = 0;
5330 	} else
5331 		uma_zone_sumstat(z, cachefree, allocs, &frees, sleeps,
5332 		    xdomain);
5333 	for (i = 0; i < vm_ndomains; i++) {
5334 		*cachefree += ZDOM_GET(z, i)->uzd_nitems;
5335 		if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
5336 		    (LIST_FIRST(&kz->uk_zones) != z)))
5337 			*cachefree += kz->uk_domain[i].ud_free_items;
5338 	}
5339 	*used = *allocs - frees;
5340 	return (((int64_t)*used + *cachefree) * kz->uk_size);
5341 }
5342 
5343 DB_SHOW_COMMAND(uma, db_show_uma)
5344 {
5345 	const char *fmt_hdr, *fmt_entry;
5346 	uma_keg_t kz;
5347 	uma_zone_t z;
5348 	uint64_t allocs, used, sleeps, xdomain;
5349 	long cachefree;
5350 	/* variables for sorting */
5351 	uma_keg_t cur_keg;
5352 	uma_zone_t cur_zone, last_zone;
5353 	int64_t cur_size, last_size, size;
5354 	int ties;
5355 
5356 	/* /i option produces machine-parseable CSV output */
5357 	if (modif[0] == 'i') {
5358 		fmt_hdr = "%s,%s,%s,%s,%s,%s,%s,%s,%s\n";
5359 		fmt_entry = "\"%s\",%ju,%jd,%ld,%ju,%ju,%u,%jd,%ju\n";
5360 	} else {
5361 		fmt_hdr = "%18s %6s %7s %7s %11s %7s %7s %10s %8s\n";
5362 		fmt_entry = "%18s %6ju %7jd %7ld %11ju %7ju %7u %10jd %8ju\n";
5363 	}
5364 
5365 	db_printf(fmt_hdr, "Zone", "Size", "Used", "Free", "Requests",
5366 	    "Sleeps", "Bucket", "Total Mem", "XFree");
5367 
5368 	/* Sort the zones with largest size first. */
5369 	last_zone = NULL;
5370 	last_size = INT64_MAX;
5371 	for (;;) {
5372 		cur_zone = NULL;
5373 		cur_size = -1;
5374 		ties = 0;
5375 		LIST_FOREACH(kz, &uma_kegs, uk_link) {
5376 			LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5377 				/*
5378 				 * In the case of size ties, print out zones
5379 				 * in the order they are encountered.  That is,
5380 				 * when we encounter the most recently output
5381 				 * zone, we have already printed all preceding
5382 				 * ties, and we must print all following ties.
5383 				 */
5384 				if (z == last_zone) {
5385 					ties = 1;
5386 					continue;
5387 				}
5388 				size = get_uma_stats(kz, z, &allocs, &used,
5389 				    &sleeps, &cachefree, &xdomain);
5390 				if (size > cur_size && size < last_size + ties)
5391 				{
5392 					cur_size = size;
5393 					cur_zone = z;
5394 					cur_keg = kz;
5395 				}
5396 			}
5397 		}
5398 		if (cur_zone == NULL)
5399 			break;
5400 
5401 		size = get_uma_stats(cur_keg, cur_zone, &allocs, &used,
5402 		    &sleeps, &cachefree, &xdomain);
5403 		db_printf(fmt_entry, cur_zone->uz_name,
5404 		    (uintmax_t)cur_keg->uk_size, (intmax_t)used, cachefree,
5405 		    (uintmax_t)allocs, (uintmax_t)sleeps,
5406 		    (unsigned)cur_zone->uz_bucket_size, (intmax_t)size,
5407 		    xdomain);
5408 
5409 		if (db_pager_quit)
5410 			return;
5411 		last_zone = cur_zone;
5412 		last_size = cur_size;
5413 	}
5414 }
5415 
5416 DB_SHOW_COMMAND(umacache, db_show_umacache)
5417 {
5418 	uma_zone_t z;
5419 	uint64_t allocs, frees;
5420 	long cachefree;
5421 	int i;
5422 
5423 	db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
5424 	    "Requests", "Bucket");
5425 	LIST_FOREACH(z, &uma_cachezones, uz_link) {
5426 		uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL, NULL);
5427 		for (i = 0; i < vm_ndomains; i++)
5428 			cachefree += ZDOM_GET(z, i)->uzd_nitems;
5429 		db_printf("%18s %8ju %8jd %8ld %12ju %8u\n",
5430 		    z->uz_name, (uintmax_t)z->uz_size,
5431 		    (intmax_t)(allocs - frees), cachefree,
5432 		    (uintmax_t)allocs, z->uz_bucket_size);
5433 		if (db_pager_quit)
5434 			return;
5435 	}
5436 }
5437 #endif	/* DDB */
5438