xref: /freebsd/sys/vm/uma_core.c (revision 4e99f45480598189d49d45a825533a6c9e12f02c)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2002-2019 Jeffrey Roberson <jeff@FreeBSD.org>
5  * Copyright (c) 2004, 2005 Bosko Milekic <bmilekic@FreeBSD.org>
6  * Copyright (c) 2004-2006 Robert N. M. Watson
7  * All rights reserved.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice unmodified, this list of conditions, and the following
14  *    disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
20  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
21  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
22  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
23  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
24  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
25  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
26  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
27  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
28  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
29  */
30 
31 /*
32  * uma_core.c  Implementation of the Universal Memory allocator
33  *
34  * This allocator is intended to replace the multitude of similar object caches
35  * in the standard FreeBSD kernel.  The intent is to be flexible as well as
36  * efficient.  A primary design goal is to return unused memory to the rest of
37  * the system.  This will make the system as a whole more flexible due to the
38  * ability to move memory to subsystems which most need it instead of leaving
39  * pools of reserved memory unused.
40  *
41  * The basic ideas stem from similar slab/zone based allocators whose algorithms
42  * are well known.
43  *
44  */
45 
46 /*
47  * TODO:
48  *	- Improve memory usage for large allocations
49  *	- Investigate cache size adjustments
50  */
51 
52 #include <sys/cdefs.h>
53 __FBSDID("$FreeBSD$");
54 
55 #include "opt_ddb.h"
56 #include "opt_param.h"
57 #include "opt_vm.h"
58 
59 #include <sys/param.h>
60 #include <sys/systm.h>
61 #include <sys/bitset.h>
62 #include <sys/domainset.h>
63 #include <sys/eventhandler.h>
64 #include <sys/kernel.h>
65 #include <sys/types.h>
66 #include <sys/limits.h>
67 #include <sys/queue.h>
68 #include <sys/malloc.h>
69 #include <sys/ktr.h>
70 #include <sys/lock.h>
71 #include <sys/sysctl.h>
72 #include <sys/mutex.h>
73 #include <sys/proc.h>
74 #include <sys/random.h>
75 #include <sys/rwlock.h>
76 #include <sys/sbuf.h>
77 #include <sys/sched.h>
78 #include <sys/sleepqueue.h>
79 #include <sys/smp.h>
80 #include <sys/smr.h>
81 #include <sys/taskqueue.h>
82 #include <sys/vmmeter.h>
83 
84 #include <vm/vm.h>
85 #include <vm/vm_domainset.h>
86 #include <vm/vm_object.h>
87 #include <vm/vm_page.h>
88 #include <vm/vm_pageout.h>
89 #include <vm/vm_param.h>
90 #include <vm/vm_phys.h>
91 #include <vm/vm_pagequeue.h>
92 #include <vm/vm_map.h>
93 #include <vm/vm_kern.h>
94 #include <vm/vm_extern.h>
95 #include <vm/uma.h>
96 #include <vm/uma_int.h>
97 #include <vm/uma_dbg.h>
98 
99 #include <ddb/ddb.h>
100 
101 #ifdef DEBUG_MEMGUARD
102 #include <vm/memguard.h>
103 #endif
104 
105 #include <machine/md_var.h>
106 
107 #ifdef INVARIANTS
108 #define	UMA_ALWAYS_CTORDTOR	1
109 #else
110 #define	UMA_ALWAYS_CTORDTOR	0
111 #endif
112 
113 /*
114  * This is the zone and keg from which all zones are spawned.
115  */
116 static uma_zone_t kegs;
117 static uma_zone_t zones;
118 
119 /*
120  * On INVARIANTS builds, the slab contains a second bitset of the same size,
121  * "dbg_bits", which is laid out immediately after us_free.
122  */
123 #ifdef INVARIANTS
124 #define	SLAB_BITSETS	2
125 #else
126 #define	SLAB_BITSETS	1
127 #endif
128 
129 /*
130  * These are the two zones from which all offpage uma_slab_ts are allocated.
131  *
132  * One zone is for slab headers that can represent a larger number of items,
133  * making the slabs themselves more efficient, and the other zone is for
134  * headers that are smaller and represent fewer items, making the headers more
135  * efficient.
136  */
137 #define	SLABZONE_SIZE(setsize)					\
138     (sizeof(struct uma_hash_slab) + BITSET_SIZE(setsize) * SLAB_BITSETS)
139 #define	SLABZONE0_SETSIZE	(PAGE_SIZE / 16)
140 #define	SLABZONE1_SETSIZE	SLAB_MAX_SETSIZE
141 #define	SLABZONE0_SIZE	SLABZONE_SIZE(SLABZONE0_SETSIZE)
142 #define	SLABZONE1_SIZE	SLABZONE_SIZE(SLABZONE1_SETSIZE)
143 static uma_zone_t slabzones[2];
144 
145 /*
146  * The initial hash tables come out of this zone so they can be allocated
147  * prior to malloc coming up.
148  */
149 static uma_zone_t hashzone;
150 
151 /* The boot-time adjusted value for cache line alignment. */
152 int uma_align_cache = 64 - 1;
153 
154 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets");
155 static MALLOC_DEFINE(M_UMA, "UMA", "UMA Misc");
156 
157 /*
158  * Are we allowed to allocate buckets?
159  */
160 static int bucketdisable = 1;
161 
162 /* Linked list of all kegs in the system */
163 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs);
164 
165 /* Linked list of all cache-only zones in the system */
166 static LIST_HEAD(,uma_zone) uma_cachezones =
167     LIST_HEAD_INITIALIZER(uma_cachezones);
168 
169 /* This RW lock protects the keg list */
170 static struct rwlock_padalign __exclusive_cache_line uma_rwlock;
171 
172 /*
173  * First available virual address for boot time allocations.
174  */
175 static vm_offset_t bootstart;
176 static vm_offset_t bootmem;
177 
178 static struct sx uma_reclaim_lock;
179 
180 /*
181  * kmem soft limit, initialized by uma_set_limit().  Ensure that early
182  * allocations don't trigger a wakeup of the reclaim thread.
183  */
184 unsigned long uma_kmem_limit = LONG_MAX;
185 SYSCTL_ULONG(_vm, OID_AUTO, uma_kmem_limit, CTLFLAG_RD, &uma_kmem_limit, 0,
186     "UMA kernel memory soft limit");
187 unsigned long uma_kmem_total;
188 SYSCTL_ULONG(_vm, OID_AUTO, uma_kmem_total, CTLFLAG_RD, &uma_kmem_total, 0,
189     "UMA kernel memory usage");
190 
191 /* Is the VM done starting up? */
192 static enum {
193 	BOOT_COLD,
194 	BOOT_KVA,
195 	BOOT_PCPU,
196 	BOOT_RUNNING,
197 	BOOT_SHUTDOWN,
198 } booted = BOOT_COLD;
199 
200 /*
201  * This is the handle used to schedule events that need to happen
202  * outside of the allocation fast path.
203  */
204 static struct callout uma_callout;
205 #define	UMA_TIMEOUT	20		/* Seconds for callout interval. */
206 
207 /*
208  * This structure is passed as the zone ctor arg so that I don't have to create
209  * a special allocation function just for zones.
210  */
211 struct uma_zctor_args {
212 	const char *name;
213 	size_t size;
214 	uma_ctor ctor;
215 	uma_dtor dtor;
216 	uma_init uminit;
217 	uma_fini fini;
218 	uma_import import;
219 	uma_release release;
220 	void *arg;
221 	uma_keg_t keg;
222 	int align;
223 	uint32_t flags;
224 };
225 
226 struct uma_kctor_args {
227 	uma_zone_t zone;
228 	size_t size;
229 	uma_init uminit;
230 	uma_fini fini;
231 	int align;
232 	uint32_t flags;
233 };
234 
235 struct uma_bucket_zone {
236 	uma_zone_t	ubz_zone;
237 	const char	*ubz_name;
238 	int		ubz_entries;	/* Number of items it can hold. */
239 	int		ubz_maxsize;	/* Maximum allocation size per-item. */
240 };
241 
242 /*
243  * Compute the actual number of bucket entries to pack them in power
244  * of two sizes for more efficient space utilization.
245  */
246 #define	BUCKET_SIZE(n)						\
247     (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *))
248 
249 #define	BUCKET_MAX	BUCKET_SIZE(256)
250 #define	BUCKET_MIN	2
251 
252 struct uma_bucket_zone bucket_zones[] = {
253 	/* Literal bucket sizes. */
254 	{ NULL, "2 Bucket", 2, 4096 },
255 	{ NULL, "4 Bucket", 4, 3072 },
256 	{ NULL, "8 Bucket", 8, 2048 },
257 	{ NULL, "16 Bucket", 16, 1024 },
258 	/* Rounded down power of 2 sizes for efficiency. */
259 	{ NULL, "32 Bucket", BUCKET_SIZE(32), 512 },
260 	{ NULL, "64 Bucket", BUCKET_SIZE(64), 256 },
261 	{ NULL, "128 Bucket", BUCKET_SIZE(128), 128 },
262 	{ NULL, "256 Bucket", BUCKET_SIZE(256), 64 },
263 	{ NULL, NULL, 0}
264 };
265 
266 /*
267  * Flags and enumerations to be passed to internal functions.
268  */
269 enum zfreeskip {
270 	SKIP_NONE =	0,
271 	SKIP_CNT =	0x00000001,
272 	SKIP_DTOR =	0x00010000,
273 	SKIP_FINI =	0x00020000,
274 };
275 
276 /* Prototypes.. */
277 
278 void	uma_startup1(vm_offset_t);
279 void	uma_startup2(void);
280 
281 static void *noobj_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
282 static void *page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
283 static void *pcpu_page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
284 static void *startup_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
285 static void *contig_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
286 static void page_free(void *, vm_size_t, uint8_t);
287 static void pcpu_page_free(void *, vm_size_t, uint8_t);
288 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int, int, int);
289 static void cache_drain(uma_zone_t);
290 static void bucket_drain(uma_zone_t, uma_bucket_t);
291 static void bucket_cache_reclaim(uma_zone_t zone, bool);
292 static int keg_ctor(void *, int, void *, int);
293 static void keg_dtor(void *, int, void *);
294 static int zone_ctor(void *, int, void *, int);
295 static void zone_dtor(void *, int, void *);
296 static inline void item_dtor(uma_zone_t zone, void *item, int size,
297     void *udata, enum zfreeskip skip);
298 static int zero_init(void *, int, int);
299 static void zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata,
300     int itemdomain, bool ws);
301 static void zone_foreach(void (*zfunc)(uma_zone_t, void *), void *);
302 static void zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *), void *);
303 static void zone_timeout(uma_zone_t zone, void *);
304 static int hash_alloc(struct uma_hash *, u_int);
305 static int hash_expand(struct uma_hash *, struct uma_hash *);
306 static void hash_free(struct uma_hash *hash);
307 static void uma_timeout(void *);
308 static void uma_shutdown(void);
309 static void *zone_alloc_item(uma_zone_t, void *, int, int);
310 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip);
311 static int zone_alloc_limit(uma_zone_t zone, int count, int flags);
312 static void zone_free_limit(uma_zone_t zone, int count);
313 static void bucket_enable(void);
314 static void bucket_init(void);
315 static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int);
316 static void bucket_free(uma_zone_t zone, uma_bucket_t, void *);
317 static void bucket_zone_drain(void);
318 static uma_bucket_t zone_alloc_bucket(uma_zone_t, void *, int, int);
319 static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab);
320 static void slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item);
321 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit,
322     uma_fini fini, int align, uint32_t flags);
323 static int zone_import(void *, void **, int, int, int);
324 static void zone_release(void *, void **, int);
325 static bool cache_alloc(uma_zone_t, uma_cache_t, void *, int);
326 static bool cache_free(uma_zone_t, uma_cache_t, void *, void *, int);
327 
328 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS);
329 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS);
330 static int sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS);
331 static int sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS);
332 static int sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS);
333 static int sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS);
334 static int sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS);
335 
336 static uint64_t uma_zone_get_allocs(uma_zone_t zone);
337 
338 static SYSCTL_NODE(_vm, OID_AUTO, debug, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
339     "Memory allocation debugging");
340 
341 #ifdef INVARIANTS
342 static uint64_t uma_keg_get_allocs(uma_keg_t zone);
343 static inline struct noslabbits *slab_dbg_bits(uma_slab_t slab, uma_keg_t keg);
344 
345 static bool uma_dbg_kskip(uma_keg_t keg, void *mem);
346 static bool uma_dbg_zskip(uma_zone_t zone, void *mem);
347 static void uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item);
348 static void uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item);
349 
350 static u_int dbg_divisor = 1;
351 SYSCTL_UINT(_vm_debug, OID_AUTO, divisor,
352     CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &dbg_divisor, 0,
353     "Debug & thrash every this item in memory allocator");
354 
355 static counter_u64_t uma_dbg_cnt = EARLY_COUNTER;
356 static counter_u64_t uma_skip_cnt = EARLY_COUNTER;
357 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, trashed, CTLFLAG_RD,
358     &uma_dbg_cnt, "memory items debugged");
359 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, skipped, CTLFLAG_RD,
360     &uma_skip_cnt, "memory items skipped, not debugged");
361 #endif
362 
363 SYSCTL_NODE(_vm, OID_AUTO, uma, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
364     "Universal Memory Allocator");
365 
366 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLFLAG_MPSAFE|CTLTYPE_INT,
367     0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones");
368 
369 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLFLAG_MPSAFE|CTLTYPE_STRUCT,
370     0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats");
371 
372 static int zone_warnings = 1;
373 SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RWTUN, &zone_warnings, 0,
374     "Warn when UMA zones becomes full");
375 
376 static int multipage_slabs = 1;
377 TUNABLE_INT("vm.debug.uma_multipage_slabs", &multipage_slabs);
378 SYSCTL_INT(_vm_debug, OID_AUTO, uma_multipage_slabs,
379     CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &multipage_slabs, 0,
380     "UMA may choose larger slab sizes for better efficiency");
381 
382 /*
383  * Select the slab zone for an offpage slab with the given maximum item count.
384  */
385 static inline uma_zone_t
386 slabzone(int ipers)
387 {
388 
389 	return (slabzones[ipers > SLABZONE0_SETSIZE]);
390 }
391 
392 /*
393  * This routine checks to see whether or not it's safe to enable buckets.
394  */
395 static void
396 bucket_enable(void)
397 {
398 
399 	KASSERT(booted >= BOOT_KVA, ("Bucket enable before init"));
400 	bucketdisable = vm_page_count_min();
401 }
402 
403 /*
404  * Initialize bucket_zones, the array of zones of buckets of various sizes.
405  *
406  * For each zone, calculate the memory required for each bucket, consisting
407  * of the header and an array of pointers.
408  */
409 static void
410 bucket_init(void)
411 {
412 	struct uma_bucket_zone *ubz;
413 	int size;
414 
415 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) {
416 		size = roundup(sizeof(struct uma_bucket), sizeof(void *));
417 		size += sizeof(void *) * ubz->ubz_entries;
418 		ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
419 		    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
420 		    UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET |
421 		    UMA_ZONE_FIRSTTOUCH);
422 	}
423 }
424 
425 /*
426  * Given a desired number of entries for a bucket, return the zone from which
427  * to allocate the bucket.
428  */
429 static struct uma_bucket_zone *
430 bucket_zone_lookup(int entries)
431 {
432 	struct uma_bucket_zone *ubz;
433 
434 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
435 		if (ubz->ubz_entries >= entries)
436 			return (ubz);
437 	ubz--;
438 	return (ubz);
439 }
440 
441 static struct uma_bucket_zone *
442 bucket_zone_max(uma_zone_t zone, int nitems)
443 {
444 	struct uma_bucket_zone *ubz;
445 	int bpcpu;
446 
447 	bpcpu = 2;
448 	if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
449 		/* Count the cross-domain bucket. */
450 		bpcpu++;
451 
452 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
453 		if (ubz->ubz_entries * bpcpu * mp_ncpus > nitems)
454 			break;
455 	if (ubz == &bucket_zones[0])
456 		ubz = NULL;
457 	else
458 		ubz--;
459 	return (ubz);
460 }
461 
462 static int
463 bucket_select(int size)
464 {
465 	struct uma_bucket_zone *ubz;
466 
467 	ubz = &bucket_zones[0];
468 	if (size > ubz->ubz_maxsize)
469 		return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1);
470 
471 	for (; ubz->ubz_entries != 0; ubz++)
472 		if (ubz->ubz_maxsize < size)
473 			break;
474 	ubz--;
475 	return (ubz->ubz_entries);
476 }
477 
478 static uma_bucket_t
479 bucket_alloc(uma_zone_t zone, void *udata, int flags)
480 {
481 	struct uma_bucket_zone *ubz;
482 	uma_bucket_t bucket;
483 
484 	/*
485 	 * Don't allocate buckets early in boot.
486 	 */
487 	if (__predict_false(booted < BOOT_KVA))
488 		return (NULL);
489 
490 	/*
491 	 * To limit bucket recursion we store the original zone flags
492 	 * in a cookie passed via zalloc_arg/zfree_arg.  This allows the
493 	 * NOVM flag to persist even through deep recursions.  We also
494 	 * store ZFLAG_BUCKET once we have recursed attempting to allocate
495 	 * a bucket for a bucket zone so we do not allow infinite bucket
496 	 * recursion.  This cookie will even persist to frees of unused
497 	 * buckets via the allocation path or bucket allocations in the
498 	 * free path.
499 	 */
500 	if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
501 		udata = (void *)(uintptr_t)zone->uz_flags;
502 	else {
503 		if ((uintptr_t)udata & UMA_ZFLAG_BUCKET)
504 			return (NULL);
505 		udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET);
506 	}
507 	if (((uintptr_t)udata & UMA_ZONE_VM) != 0)
508 		flags |= M_NOVM;
509 	ubz = bucket_zone_lookup(zone->uz_bucket_size);
510 	if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0)
511 		ubz++;
512 	bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags);
513 	if (bucket) {
514 #ifdef INVARIANTS
515 		bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
516 #endif
517 		bucket->ub_cnt = 0;
518 		bucket->ub_entries = ubz->ubz_entries;
519 		bucket->ub_seq = SMR_SEQ_INVALID;
520 		CTR3(KTR_UMA, "bucket_alloc: zone %s(%p) allocated bucket %p",
521 		    zone->uz_name, zone, bucket);
522 	}
523 
524 	return (bucket);
525 }
526 
527 static void
528 bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata)
529 {
530 	struct uma_bucket_zone *ubz;
531 
532 	if (bucket->ub_cnt != 0)
533 		bucket_drain(zone, bucket);
534 
535 	KASSERT(bucket->ub_cnt == 0,
536 	    ("bucket_free: Freeing a non free bucket."));
537 	KASSERT(bucket->ub_seq == SMR_SEQ_INVALID,
538 	    ("bucket_free: Freeing an SMR bucket."));
539 	if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
540 		udata = (void *)(uintptr_t)zone->uz_flags;
541 	ubz = bucket_zone_lookup(bucket->ub_entries);
542 	uma_zfree_arg(ubz->ubz_zone, bucket, udata);
543 }
544 
545 static void
546 bucket_zone_drain(void)
547 {
548 	struct uma_bucket_zone *ubz;
549 
550 	for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
551 		uma_zone_reclaim(ubz->ubz_zone, UMA_RECLAIM_DRAIN);
552 }
553 
554 /*
555  * Acquire the domain lock and record contention.
556  */
557 static uma_zone_domain_t
558 zone_domain_lock(uma_zone_t zone, int domain)
559 {
560 	uma_zone_domain_t zdom;
561 	bool lockfail;
562 
563 	zdom = ZDOM_GET(zone, domain);
564 	lockfail = false;
565 	if (ZDOM_OWNED(zdom))
566 		lockfail = true;
567 	ZDOM_LOCK(zdom);
568 	/* This is unsynchronized.  The counter does not need to be precise. */
569 	if (lockfail && zone->uz_bucket_size < zone->uz_bucket_size_max)
570 		zone->uz_bucket_size++;
571 	return (zdom);
572 }
573 
574 /*
575  * Search for the domain with the least cached items and return it if it
576  * is out of balance with the preferred domain.
577  */
578 static __noinline int
579 zone_domain_lowest(uma_zone_t zone, int pref)
580 {
581 	long least, nitems, prefitems;
582 	int domain;
583 	int i;
584 
585 	prefitems = least = LONG_MAX;
586 	domain = 0;
587 	for (i = 0; i < vm_ndomains; i++) {
588 		nitems = ZDOM_GET(zone, i)->uzd_nitems;
589 		if (nitems < least) {
590 			domain = i;
591 			least = nitems;
592 		}
593 		if (domain == pref)
594 			prefitems = nitems;
595 	}
596 	if (prefitems < least * 2)
597 		return (pref);
598 
599 	return (domain);
600 }
601 
602 /*
603  * Search for the domain with the most cached items and return it or the
604  * preferred domain if it has enough to proceed.
605  */
606 static __noinline int
607 zone_domain_highest(uma_zone_t zone, int pref)
608 {
609 	long most, nitems;
610 	int domain;
611 	int i;
612 
613 	if (ZDOM_GET(zone, pref)->uzd_nitems > BUCKET_MAX)
614 		return (pref);
615 
616 	most = 0;
617 	domain = 0;
618 	for (i = 0; i < vm_ndomains; i++) {
619 		nitems = ZDOM_GET(zone, i)->uzd_nitems;
620 		if (nitems > most) {
621 			domain = i;
622 			most = nitems;
623 		}
624 	}
625 
626 	return (domain);
627 }
628 
629 /*
630  * Safely subtract cnt from imax.
631  */
632 static void
633 zone_domain_imax_sub(uma_zone_domain_t zdom, int cnt)
634 {
635 	long new;
636 	long old;
637 
638 	old = zdom->uzd_imax;
639 	do {
640 		if (old <= cnt)
641 			new = 0;
642 		else
643 			new = old - cnt;
644 	} while (atomic_fcmpset_long(&zdom->uzd_imax, &old, new) == 0);
645 }
646 
647 /*
648  * Set the maximum imax value.
649  */
650 static void
651 zone_domain_imax_set(uma_zone_domain_t zdom, int nitems)
652 {
653 	long old;
654 
655 	old = zdom->uzd_imax;
656 	do {
657 		if (old >= nitems)
658 			break;
659 	} while (atomic_fcmpset_long(&zdom->uzd_imax, &old, nitems) == 0);
660 }
661 
662 /*
663  * Attempt to satisfy an allocation by retrieving a full bucket from one of the
664  * zone's caches.  If a bucket is found the zone is not locked on return.
665  */
666 static uma_bucket_t
667 zone_fetch_bucket(uma_zone_t zone, uma_zone_domain_t zdom, bool reclaim)
668 {
669 	uma_bucket_t bucket;
670 	int i;
671 	bool dtor = false;
672 
673 	ZDOM_LOCK_ASSERT(zdom);
674 
675 	if ((bucket = STAILQ_FIRST(&zdom->uzd_buckets)) == NULL)
676 		return (NULL);
677 
678 	/* SMR Buckets can not be re-used until readers expire. */
679 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
680 	    bucket->ub_seq != SMR_SEQ_INVALID) {
681 		if (!smr_poll(zone->uz_smr, bucket->ub_seq, false))
682 			return (NULL);
683 		bucket->ub_seq = SMR_SEQ_INVALID;
684 		dtor = (zone->uz_dtor != NULL) || UMA_ALWAYS_CTORDTOR;
685 		if (STAILQ_NEXT(bucket, ub_link) != NULL)
686 			zdom->uzd_seq = STAILQ_NEXT(bucket, ub_link)->ub_seq;
687 	}
688 	MPASS(zdom->uzd_nitems >= bucket->ub_cnt);
689 	STAILQ_REMOVE_HEAD(&zdom->uzd_buckets, ub_link);
690 	zdom->uzd_nitems -= bucket->ub_cnt;
691 
692 	/*
693 	 * Shift the bounds of the current WSS interval to avoid
694 	 * perturbing the estimate.
695 	 */
696 	if (reclaim) {
697 		zdom->uzd_imin -= lmin(zdom->uzd_imin, bucket->ub_cnt);
698 		zone_domain_imax_sub(zdom, bucket->ub_cnt);
699 	} else if (zdom->uzd_imin > zdom->uzd_nitems)
700 		zdom->uzd_imin = zdom->uzd_nitems;
701 
702 	ZDOM_UNLOCK(zdom);
703 	if (dtor)
704 		for (i = 0; i < bucket->ub_cnt; i++)
705 			item_dtor(zone, bucket->ub_bucket[i], zone->uz_size,
706 			    NULL, SKIP_NONE);
707 
708 	return (bucket);
709 }
710 
711 /*
712  * Insert a full bucket into the specified cache.  The "ws" parameter indicates
713  * whether the bucket's contents should be counted as part of the zone's working
714  * set.  The bucket may be freed if it exceeds the bucket limit.
715  */
716 static void
717 zone_put_bucket(uma_zone_t zone, int domain, uma_bucket_t bucket, void *udata,
718     const bool ws)
719 {
720 	uma_zone_domain_t zdom;
721 
722 	/* We don't cache empty buckets.  This can happen after a reclaim. */
723 	if (bucket->ub_cnt == 0)
724 		goto out;
725 	zdom = zone_domain_lock(zone, domain);
726 
727 	/*
728 	 * Conditionally set the maximum number of items.
729 	 */
730 	zdom->uzd_nitems += bucket->ub_cnt;
731 	if (__predict_true(zdom->uzd_nitems < zone->uz_bucket_max)) {
732 		if (ws)
733 			zone_domain_imax_set(zdom, zdom->uzd_nitems);
734 		if (STAILQ_EMPTY(&zdom->uzd_buckets))
735 			zdom->uzd_seq = bucket->ub_seq;
736 		STAILQ_INSERT_TAIL(&zdom->uzd_buckets, bucket, ub_link);
737 		ZDOM_UNLOCK(zdom);
738 		return;
739 	}
740 	zdom->uzd_nitems -= bucket->ub_cnt;
741 	ZDOM_UNLOCK(zdom);
742 out:
743 	bucket_free(zone, bucket, udata);
744 }
745 
746 /* Pops an item out of a per-cpu cache bucket. */
747 static inline void *
748 cache_bucket_pop(uma_cache_t cache, uma_cache_bucket_t bucket)
749 {
750 	void *item;
751 
752 	CRITICAL_ASSERT(curthread);
753 
754 	bucket->ucb_cnt--;
755 	item = bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt];
756 #ifdef INVARIANTS
757 	bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = NULL;
758 	KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled."));
759 #endif
760 	cache->uc_allocs++;
761 
762 	return (item);
763 }
764 
765 /* Pushes an item into a per-cpu cache bucket. */
766 static inline void
767 cache_bucket_push(uma_cache_t cache, uma_cache_bucket_t bucket, void *item)
768 {
769 
770 	CRITICAL_ASSERT(curthread);
771 	KASSERT(bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] == NULL,
772 	    ("uma_zfree: Freeing to non free bucket index."));
773 
774 	bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = item;
775 	bucket->ucb_cnt++;
776 	cache->uc_frees++;
777 }
778 
779 /*
780  * Unload a UMA bucket from a per-cpu cache.
781  */
782 static inline uma_bucket_t
783 cache_bucket_unload(uma_cache_bucket_t bucket)
784 {
785 	uma_bucket_t b;
786 
787 	b = bucket->ucb_bucket;
788 	if (b != NULL) {
789 		MPASS(b->ub_entries == bucket->ucb_entries);
790 		b->ub_cnt = bucket->ucb_cnt;
791 		bucket->ucb_bucket = NULL;
792 		bucket->ucb_entries = bucket->ucb_cnt = 0;
793 	}
794 
795 	return (b);
796 }
797 
798 static inline uma_bucket_t
799 cache_bucket_unload_alloc(uma_cache_t cache)
800 {
801 
802 	return (cache_bucket_unload(&cache->uc_allocbucket));
803 }
804 
805 static inline uma_bucket_t
806 cache_bucket_unload_free(uma_cache_t cache)
807 {
808 
809 	return (cache_bucket_unload(&cache->uc_freebucket));
810 }
811 
812 static inline uma_bucket_t
813 cache_bucket_unload_cross(uma_cache_t cache)
814 {
815 
816 	return (cache_bucket_unload(&cache->uc_crossbucket));
817 }
818 
819 /*
820  * Load a bucket into a per-cpu cache bucket.
821  */
822 static inline void
823 cache_bucket_load(uma_cache_bucket_t bucket, uma_bucket_t b)
824 {
825 
826 	CRITICAL_ASSERT(curthread);
827 	MPASS(bucket->ucb_bucket == NULL);
828 	MPASS(b->ub_seq == SMR_SEQ_INVALID);
829 
830 	bucket->ucb_bucket = b;
831 	bucket->ucb_cnt = b->ub_cnt;
832 	bucket->ucb_entries = b->ub_entries;
833 }
834 
835 static inline void
836 cache_bucket_load_alloc(uma_cache_t cache, uma_bucket_t b)
837 {
838 
839 	cache_bucket_load(&cache->uc_allocbucket, b);
840 }
841 
842 static inline void
843 cache_bucket_load_free(uma_cache_t cache, uma_bucket_t b)
844 {
845 
846 	cache_bucket_load(&cache->uc_freebucket, b);
847 }
848 
849 #ifdef NUMA
850 static inline void
851 cache_bucket_load_cross(uma_cache_t cache, uma_bucket_t b)
852 {
853 
854 	cache_bucket_load(&cache->uc_crossbucket, b);
855 }
856 #endif
857 
858 /*
859  * Copy and preserve ucb_spare.
860  */
861 static inline void
862 cache_bucket_copy(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
863 {
864 
865 	b1->ucb_bucket = b2->ucb_bucket;
866 	b1->ucb_entries = b2->ucb_entries;
867 	b1->ucb_cnt = b2->ucb_cnt;
868 }
869 
870 /*
871  * Swap two cache buckets.
872  */
873 static inline void
874 cache_bucket_swap(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
875 {
876 	struct uma_cache_bucket b3;
877 
878 	CRITICAL_ASSERT(curthread);
879 
880 	cache_bucket_copy(&b3, b1);
881 	cache_bucket_copy(b1, b2);
882 	cache_bucket_copy(b2, &b3);
883 }
884 
885 /*
886  * Attempt to fetch a bucket from a zone on behalf of the current cpu cache.
887  */
888 static uma_bucket_t
889 cache_fetch_bucket(uma_zone_t zone, uma_cache_t cache, int domain)
890 {
891 	uma_zone_domain_t zdom;
892 	uma_bucket_t bucket;
893 
894 	/*
895 	 * Avoid the lock if possible.
896 	 */
897 	zdom = ZDOM_GET(zone, domain);
898 	if (zdom->uzd_nitems == 0)
899 		return (NULL);
900 
901 	if ((cache_uz_flags(cache) & UMA_ZONE_SMR) != 0 &&
902 	    !smr_poll(zone->uz_smr, zdom->uzd_seq, false))
903 		return (NULL);
904 
905 	/*
906 	 * Check the zone's cache of buckets.
907 	 */
908 	zdom = zone_domain_lock(zone, domain);
909 	if ((bucket = zone_fetch_bucket(zone, zdom, false)) != NULL) {
910 		KASSERT(bucket->ub_cnt != 0,
911 		    ("cache_fetch_bucket: Returning an empty bucket."));
912 		return (bucket);
913 	}
914 	ZDOM_UNLOCK(zdom);
915 
916 	return (NULL);
917 }
918 
919 static void
920 zone_log_warning(uma_zone_t zone)
921 {
922 	static const struct timeval warninterval = { 300, 0 };
923 
924 	if (!zone_warnings || zone->uz_warning == NULL)
925 		return;
926 
927 	if (ratecheck(&zone->uz_ratecheck, &warninterval))
928 		printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning);
929 }
930 
931 static inline void
932 zone_maxaction(uma_zone_t zone)
933 {
934 
935 	if (zone->uz_maxaction.ta_func != NULL)
936 		taskqueue_enqueue(taskqueue_thread, &zone->uz_maxaction);
937 }
938 
939 /*
940  * Routine called by timeout which is used to fire off some time interval
941  * based calculations.  (stats, hash size, etc.)
942  *
943  * Arguments:
944  *	arg   Unused
945  *
946  * Returns:
947  *	Nothing
948  */
949 static void
950 uma_timeout(void *unused)
951 {
952 	bucket_enable();
953 	zone_foreach(zone_timeout, NULL);
954 
955 	/* Reschedule this event */
956 	callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
957 }
958 
959 /*
960  * Update the working set size estimate for the zone's bucket cache.
961  * The constants chosen here are somewhat arbitrary.  With an update period of
962  * 20s (UMA_TIMEOUT), this estimate is dominated by zone activity over the
963  * last 100s.
964  */
965 static void
966 zone_domain_update_wss(uma_zone_domain_t zdom)
967 {
968 	long wss;
969 
970 	ZDOM_LOCK(zdom);
971 	MPASS(zdom->uzd_imax >= zdom->uzd_imin);
972 	wss = zdom->uzd_imax - zdom->uzd_imin;
973 	zdom->uzd_imax = zdom->uzd_imin = zdom->uzd_nitems;
974 	zdom->uzd_wss = (4 * wss + zdom->uzd_wss) / 5;
975 	ZDOM_UNLOCK(zdom);
976 }
977 
978 /*
979  * Routine to perform timeout driven calculations.  This expands the
980  * hashes and does per cpu statistics aggregation.
981  *
982  *  Returns nothing.
983  */
984 static void
985 zone_timeout(uma_zone_t zone, void *unused)
986 {
987 	uma_keg_t keg;
988 	u_int slabs, pages;
989 
990 	if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
991 		goto update_wss;
992 
993 	keg = zone->uz_keg;
994 
995 	/*
996 	 * Hash zones are non-numa by definition so the first domain
997 	 * is the only one present.
998 	 */
999 	KEG_LOCK(keg, 0);
1000 	pages = keg->uk_domain[0].ud_pages;
1001 
1002 	/*
1003 	 * Expand the keg hash table.
1004 	 *
1005 	 * This is done if the number of slabs is larger than the hash size.
1006 	 * What I'm trying to do here is completely reduce collisions.  This
1007 	 * may be a little aggressive.  Should I allow for two collisions max?
1008 	 */
1009 	if ((slabs = pages / keg->uk_ppera) > keg->uk_hash.uh_hashsize) {
1010 		struct uma_hash newhash;
1011 		struct uma_hash oldhash;
1012 		int ret;
1013 
1014 		/*
1015 		 * This is so involved because allocating and freeing
1016 		 * while the keg lock is held will lead to deadlock.
1017 		 * I have to do everything in stages and check for
1018 		 * races.
1019 		 */
1020 		KEG_UNLOCK(keg, 0);
1021 		ret = hash_alloc(&newhash, 1 << fls(slabs));
1022 		KEG_LOCK(keg, 0);
1023 		if (ret) {
1024 			if (hash_expand(&keg->uk_hash, &newhash)) {
1025 				oldhash = keg->uk_hash;
1026 				keg->uk_hash = newhash;
1027 			} else
1028 				oldhash = newhash;
1029 
1030 			KEG_UNLOCK(keg, 0);
1031 			hash_free(&oldhash);
1032 			goto update_wss;
1033 		}
1034 	}
1035 	KEG_UNLOCK(keg, 0);
1036 
1037 update_wss:
1038 	for (int i = 0; i < vm_ndomains; i++)
1039 		zone_domain_update_wss(ZDOM_GET(zone, i));
1040 }
1041 
1042 /*
1043  * Allocate and zero fill the next sized hash table from the appropriate
1044  * backing store.
1045  *
1046  * Arguments:
1047  *	hash  A new hash structure with the old hash size in uh_hashsize
1048  *
1049  * Returns:
1050  *	1 on success and 0 on failure.
1051  */
1052 static int
1053 hash_alloc(struct uma_hash *hash, u_int size)
1054 {
1055 	size_t alloc;
1056 
1057 	KASSERT(powerof2(size), ("hash size must be power of 2"));
1058 	if (size > UMA_HASH_SIZE_INIT)  {
1059 		hash->uh_hashsize = size;
1060 		alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
1061 		hash->uh_slab_hash = malloc(alloc, M_UMAHASH, M_NOWAIT);
1062 	} else {
1063 		alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
1064 		hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
1065 		    UMA_ANYDOMAIN, M_WAITOK);
1066 		hash->uh_hashsize = UMA_HASH_SIZE_INIT;
1067 	}
1068 	if (hash->uh_slab_hash) {
1069 		bzero(hash->uh_slab_hash, alloc);
1070 		hash->uh_hashmask = hash->uh_hashsize - 1;
1071 		return (1);
1072 	}
1073 
1074 	return (0);
1075 }
1076 
1077 /*
1078  * Expands the hash table for HASH zones.  This is done from zone_timeout
1079  * to reduce collisions.  This must not be done in the regular allocation
1080  * path, otherwise, we can recurse on the vm while allocating pages.
1081  *
1082  * Arguments:
1083  *	oldhash  The hash you want to expand
1084  *	newhash  The hash structure for the new table
1085  *
1086  * Returns:
1087  *	Nothing
1088  *
1089  * Discussion:
1090  */
1091 static int
1092 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
1093 {
1094 	uma_hash_slab_t slab;
1095 	u_int hval;
1096 	u_int idx;
1097 
1098 	if (!newhash->uh_slab_hash)
1099 		return (0);
1100 
1101 	if (oldhash->uh_hashsize >= newhash->uh_hashsize)
1102 		return (0);
1103 
1104 	/*
1105 	 * I need to investigate hash algorithms for resizing without a
1106 	 * full rehash.
1107 	 */
1108 
1109 	for (idx = 0; idx < oldhash->uh_hashsize; idx++)
1110 		while (!LIST_EMPTY(&oldhash->uh_slab_hash[idx])) {
1111 			slab = LIST_FIRST(&oldhash->uh_slab_hash[idx]);
1112 			LIST_REMOVE(slab, uhs_hlink);
1113 			hval = UMA_HASH(newhash, slab->uhs_data);
1114 			LIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
1115 			    slab, uhs_hlink);
1116 		}
1117 
1118 	return (1);
1119 }
1120 
1121 /*
1122  * Free the hash bucket to the appropriate backing store.
1123  *
1124  * Arguments:
1125  *	slab_hash  The hash bucket we're freeing
1126  *	hashsize   The number of entries in that hash bucket
1127  *
1128  * Returns:
1129  *	Nothing
1130  */
1131 static void
1132 hash_free(struct uma_hash *hash)
1133 {
1134 	if (hash->uh_slab_hash == NULL)
1135 		return;
1136 	if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
1137 		zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE);
1138 	else
1139 		free(hash->uh_slab_hash, M_UMAHASH);
1140 }
1141 
1142 /*
1143  * Frees all outstanding items in a bucket
1144  *
1145  * Arguments:
1146  *	zone   The zone to free to, must be unlocked.
1147  *	bucket The free/alloc bucket with items.
1148  *
1149  * Returns:
1150  *	Nothing
1151  */
1152 static void
1153 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
1154 {
1155 	int i;
1156 
1157 	if (bucket->ub_cnt == 0)
1158 		return;
1159 
1160 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
1161 	    bucket->ub_seq != SMR_SEQ_INVALID) {
1162 		smr_wait(zone->uz_smr, bucket->ub_seq);
1163 		bucket->ub_seq = SMR_SEQ_INVALID;
1164 		for (i = 0; i < bucket->ub_cnt; i++)
1165 			item_dtor(zone, bucket->ub_bucket[i],
1166 			    zone->uz_size, NULL, SKIP_NONE);
1167 	}
1168 	if (zone->uz_fini)
1169 		for (i = 0; i < bucket->ub_cnt; i++)
1170 			zone->uz_fini(bucket->ub_bucket[i], zone->uz_size);
1171 	zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt);
1172 	if (zone->uz_max_items > 0)
1173 		zone_free_limit(zone, bucket->ub_cnt);
1174 #ifdef INVARIANTS
1175 	bzero(bucket->ub_bucket, sizeof(void *) * bucket->ub_cnt);
1176 #endif
1177 	bucket->ub_cnt = 0;
1178 }
1179 
1180 /*
1181  * Drains the per cpu caches for a zone.
1182  *
1183  * NOTE: This may only be called while the zone is being torn down, and not
1184  * during normal operation.  This is necessary in order that we do not have
1185  * to migrate CPUs to drain the per-CPU caches.
1186  *
1187  * Arguments:
1188  *	zone     The zone to drain, must be unlocked.
1189  *
1190  * Returns:
1191  *	Nothing
1192  */
1193 static void
1194 cache_drain(uma_zone_t zone)
1195 {
1196 	uma_cache_t cache;
1197 	uma_bucket_t bucket;
1198 	smr_seq_t seq;
1199 	int cpu;
1200 
1201 	/*
1202 	 * XXX: It is safe to not lock the per-CPU caches, because we're
1203 	 * tearing down the zone anyway.  I.e., there will be no further use
1204 	 * of the caches at this point.
1205 	 *
1206 	 * XXX: It would good to be able to assert that the zone is being
1207 	 * torn down to prevent improper use of cache_drain().
1208 	 */
1209 	seq = SMR_SEQ_INVALID;
1210 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
1211 		seq = smr_advance(zone->uz_smr);
1212 	CPU_FOREACH(cpu) {
1213 		cache = &zone->uz_cpu[cpu];
1214 		bucket = cache_bucket_unload_alloc(cache);
1215 		if (bucket != NULL)
1216 			bucket_free(zone, bucket, NULL);
1217 		bucket = cache_bucket_unload_free(cache);
1218 		if (bucket != NULL) {
1219 			bucket->ub_seq = seq;
1220 			bucket_free(zone, bucket, NULL);
1221 		}
1222 		bucket = cache_bucket_unload_cross(cache);
1223 		if (bucket != NULL) {
1224 			bucket->ub_seq = seq;
1225 			bucket_free(zone, bucket, NULL);
1226 		}
1227 	}
1228 	bucket_cache_reclaim(zone, true);
1229 }
1230 
1231 static void
1232 cache_shrink(uma_zone_t zone, void *unused)
1233 {
1234 
1235 	if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1236 		return;
1237 
1238 	zone->uz_bucket_size =
1239 	    (zone->uz_bucket_size_min + zone->uz_bucket_size) / 2;
1240 }
1241 
1242 static void
1243 cache_drain_safe_cpu(uma_zone_t zone, void *unused)
1244 {
1245 	uma_cache_t cache;
1246 	uma_bucket_t b1, b2, b3;
1247 	int domain;
1248 
1249 	if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1250 		return;
1251 
1252 	b1 = b2 = b3 = NULL;
1253 	critical_enter();
1254 	cache = &zone->uz_cpu[curcpu];
1255 	domain = PCPU_GET(domain);
1256 	b1 = cache_bucket_unload_alloc(cache);
1257 
1258 	/*
1259 	 * Don't flush SMR zone buckets.  This leaves the zone without a
1260 	 * bucket and forces every free to synchronize().
1261 	 */
1262 	if ((zone->uz_flags & UMA_ZONE_SMR) == 0) {
1263 		b2 = cache_bucket_unload_free(cache);
1264 		b3 = cache_bucket_unload_cross(cache);
1265 	}
1266 	critical_exit();
1267 
1268 	if (b1 != NULL)
1269 		zone_free_bucket(zone, b1, NULL, domain, false);
1270 	if (b2 != NULL)
1271 		zone_free_bucket(zone, b2, NULL, domain, false);
1272 	if (b3 != NULL) {
1273 		/* Adjust the domain so it goes to zone_free_cross. */
1274 		domain = (domain + 1) % vm_ndomains;
1275 		zone_free_bucket(zone, b3, NULL, domain, false);
1276 	}
1277 }
1278 
1279 /*
1280  * Safely drain per-CPU caches of a zone(s) to alloc bucket.
1281  * This is an expensive call because it needs to bind to all CPUs
1282  * one by one and enter a critical section on each of them in order
1283  * to safely access their cache buckets.
1284  * Zone lock must not be held on call this function.
1285  */
1286 static void
1287 pcpu_cache_drain_safe(uma_zone_t zone)
1288 {
1289 	int cpu;
1290 
1291 	/*
1292 	 * Polite bucket sizes shrinking was not enough, shrink aggressively.
1293 	 */
1294 	if (zone)
1295 		cache_shrink(zone, NULL);
1296 	else
1297 		zone_foreach(cache_shrink, NULL);
1298 
1299 	CPU_FOREACH(cpu) {
1300 		thread_lock(curthread);
1301 		sched_bind(curthread, cpu);
1302 		thread_unlock(curthread);
1303 
1304 		if (zone)
1305 			cache_drain_safe_cpu(zone, NULL);
1306 		else
1307 			zone_foreach(cache_drain_safe_cpu, NULL);
1308 	}
1309 	thread_lock(curthread);
1310 	sched_unbind(curthread);
1311 	thread_unlock(curthread);
1312 }
1313 
1314 /*
1315  * Reclaim cached buckets from a zone.  All buckets are reclaimed if the caller
1316  * requested a drain, otherwise the per-domain caches are trimmed to either
1317  * estimated working set size.
1318  */
1319 static void
1320 bucket_cache_reclaim(uma_zone_t zone, bool drain)
1321 {
1322 	uma_zone_domain_t zdom;
1323 	uma_bucket_t bucket;
1324 	long target;
1325 	int i;
1326 
1327 	/*
1328 	 * Shrink the zone bucket size to ensure that the per-CPU caches
1329 	 * don't grow too large.
1330 	 */
1331 	if (zone->uz_bucket_size > zone->uz_bucket_size_min)
1332 		zone->uz_bucket_size--;
1333 
1334 	for (i = 0; i < vm_ndomains; i++) {
1335 		/*
1336 		 * The cross bucket is partially filled and not part of
1337 		 * the item count.  Reclaim it individually here.
1338 		 */
1339 		zdom = ZDOM_GET(zone, i);
1340 		if ((zone->uz_flags & UMA_ZONE_SMR) == 0 || drain) {
1341 			ZONE_CROSS_LOCK(zone);
1342 			bucket = zdom->uzd_cross;
1343 			zdom->uzd_cross = NULL;
1344 			ZONE_CROSS_UNLOCK(zone);
1345 			if (bucket != NULL)
1346 				bucket_free(zone, bucket, NULL);
1347 		}
1348 
1349 		/*
1350 		 * If we were asked to drain the zone, we are done only once
1351 		 * this bucket cache is empty.  Otherwise, we reclaim items in
1352 		 * excess of the zone's estimated working set size.  If the
1353 		 * difference nitems - imin is larger than the WSS estimate,
1354 		 * then the estimate will grow at the end of this interval and
1355 		 * we ignore the historical average.
1356 		 */
1357 		ZDOM_LOCK(zdom);
1358 		target = drain ? 0 : lmax(zdom->uzd_wss, zdom->uzd_nitems -
1359 		    zdom->uzd_imin);
1360 		while (zdom->uzd_nitems > target) {
1361 			bucket = zone_fetch_bucket(zone, zdom, true);
1362 			if (bucket == NULL)
1363 				break;
1364 			bucket_free(zone, bucket, NULL);
1365 			ZDOM_LOCK(zdom);
1366 		}
1367 		ZDOM_UNLOCK(zdom);
1368 	}
1369 }
1370 
1371 static void
1372 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start)
1373 {
1374 	uint8_t *mem;
1375 	int i;
1376 	uint8_t flags;
1377 
1378 	CTR4(KTR_UMA, "keg_free_slab keg %s(%p) slab %p, returning %d bytes",
1379 	    keg->uk_name, keg, slab, PAGE_SIZE * keg->uk_ppera);
1380 
1381 	mem = slab_data(slab, keg);
1382 	flags = slab->us_flags;
1383 	i = start;
1384 	if (keg->uk_fini != NULL) {
1385 		for (i--; i > -1; i--)
1386 #ifdef INVARIANTS
1387 		/*
1388 		 * trash_fini implies that dtor was trash_dtor. trash_fini
1389 		 * would check that memory hasn't been modified since free,
1390 		 * which executed trash_dtor.
1391 		 * That's why we need to run uma_dbg_kskip() check here,
1392 		 * albeit we don't make skip check for other init/fini
1393 		 * invocations.
1394 		 */
1395 		if (!uma_dbg_kskip(keg, slab_item(slab, keg, i)) ||
1396 		    keg->uk_fini != trash_fini)
1397 #endif
1398 			keg->uk_fini(slab_item(slab, keg, i), keg->uk_size);
1399 	}
1400 	if (keg->uk_flags & UMA_ZFLAG_OFFPAGE)
1401 		zone_free_item(slabzone(keg->uk_ipers), slab_tohashslab(slab),
1402 		    NULL, SKIP_NONE);
1403 	keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags);
1404 	uma_total_dec(PAGE_SIZE * keg->uk_ppera);
1405 }
1406 
1407 /*
1408  * Frees pages from a keg back to the system.  This is done on demand from
1409  * the pageout daemon.
1410  *
1411  * Returns nothing.
1412  */
1413 static void
1414 keg_drain(uma_keg_t keg)
1415 {
1416 	struct slabhead freeslabs;
1417 	uma_domain_t dom;
1418 	uma_slab_t slab, tmp;
1419 	int i, n;
1420 
1421 	if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL)
1422 		return;
1423 
1424 	for (i = 0; i < vm_ndomains; i++) {
1425 		CTR4(KTR_UMA, "keg_drain %s(%p) domain %d free items: %u",
1426 		    keg->uk_name, keg, i, dom->ud_free_items);
1427 		dom = &keg->uk_domain[i];
1428 		LIST_INIT(&freeslabs);
1429 
1430 		KEG_LOCK(keg, i);
1431 		if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0) {
1432 			LIST_FOREACH(slab, &dom->ud_free_slab, us_link)
1433 				UMA_HASH_REMOVE(&keg->uk_hash, slab);
1434 		}
1435 		n = dom->ud_free_slabs;
1436 		LIST_SWAP(&freeslabs, &dom->ud_free_slab, uma_slab, us_link);
1437 		dom->ud_free_slabs = 0;
1438 		dom->ud_free_items -= n * keg->uk_ipers;
1439 		dom->ud_pages -= n * keg->uk_ppera;
1440 		KEG_UNLOCK(keg, i);
1441 
1442 		LIST_FOREACH_SAFE(slab, &freeslabs, us_link, tmp)
1443 			keg_free_slab(keg, slab, keg->uk_ipers);
1444 	}
1445 }
1446 
1447 static void
1448 zone_reclaim(uma_zone_t zone, int waitok, bool drain)
1449 {
1450 
1451 	/*
1452 	 * Set draining to interlock with zone_dtor() so we can release our
1453 	 * locks as we go.  Only dtor() should do a WAITOK call since it
1454 	 * is the only call that knows the structure will still be available
1455 	 * when it wakes up.
1456 	 */
1457 	ZONE_LOCK(zone);
1458 	while (zone->uz_flags & UMA_ZFLAG_RECLAIMING) {
1459 		if (waitok == M_NOWAIT)
1460 			goto out;
1461 		msleep(zone, &ZDOM_GET(zone, 0)->uzd_lock, PVM, "zonedrain",
1462 		    1);
1463 	}
1464 	zone->uz_flags |= UMA_ZFLAG_RECLAIMING;
1465 	ZONE_UNLOCK(zone);
1466 	bucket_cache_reclaim(zone, drain);
1467 
1468 	/*
1469 	 * The DRAINING flag protects us from being freed while
1470 	 * we're running.  Normally the uma_rwlock would protect us but we
1471 	 * must be able to release and acquire the right lock for each keg.
1472 	 */
1473 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0)
1474 		keg_drain(zone->uz_keg);
1475 	ZONE_LOCK(zone);
1476 	zone->uz_flags &= ~UMA_ZFLAG_RECLAIMING;
1477 	wakeup(zone);
1478 out:
1479 	ZONE_UNLOCK(zone);
1480 }
1481 
1482 static void
1483 zone_drain(uma_zone_t zone, void *unused)
1484 {
1485 
1486 	zone_reclaim(zone, M_NOWAIT, true);
1487 }
1488 
1489 static void
1490 zone_trim(uma_zone_t zone, void *unused)
1491 {
1492 
1493 	zone_reclaim(zone, M_NOWAIT, false);
1494 }
1495 
1496 /*
1497  * Allocate a new slab for a keg and inserts it into the partial slab list.
1498  * The keg should be unlocked on entry.  If the allocation succeeds it will
1499  * be locked on return.
1500  *
1501  * Arguments:
1502  *	flags   Wait flags for the item initialization routine
1503  *	aflags  Wait flags for the slab allocation
1504  *
1505  * Returns:
1506  *	The slab that was allocated or NULL if there is no memory and the
1507  *	caller specified M_NOWAIT.
1508  */
1509 static uma_slab_t
1510 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int domain, int flags,
1511     int aflags)
1512 {
1513 	uma_domain_t dom;
1514 	uma_alloc allocf;
1515 	uma_slab_t slab;
1516 	unsigned long size;
1517 	uint8_t *mem;
1518 	uint8_t sflags;
1519 	int i;
1520 
1521 	KASSERT(domain >= 0 && domain < vm_ndomains,
1522 	    ("keg_alloc_slab: domain %d out of range", domain));
1523 
1524 	allocf = keg->uk_allocf;
1525 	slab = NULL;
1526 	mem = NULL;
1527 	if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) {
1528 		uma_hash_slab_t hslab;
1529 		hslab = zone_alloc_item(slabzone(keg->uk_ipers), NULL,
1530 		    domain, aflags);
1531 		if (hslab == NULL)
1532 			goto fail;
1533 		slab = &hslab->uhs_slab;
1534 	}
1535 
1536 	/*
1537 	 * This reproduces the old vm_zone behavior of zero filling pages the
1538 	 * first time they are added to a zone.
1539 	 *
1540 	 * Malloced items are zeroed in uma_zalloc.
1541 	 */
1542 
1543 	if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
1544 		aflags |= M_ZERO;
1545 	else
1546 		aflags &= ~M_ZERO;
1547 
1548 	if (keg->uk_flags & UMA_ZONE_NODUMP)
1549 		aflags |= M_NODUMP;
1550 
1551 	/* zone is passed for legacy reasons. */
1552 	size = keg->uk_ppera * PAGE_SIZE;
1553 	mem = allocf(zone, size, domain, &sflags, aflags);
1554 	if (mem == NULL) {
1555 		if (keg->uk_flags & UMA_ZFLAG_OFFPAGE)
1556 			zone_free_item(slabzone(keg->uk_ipers),
1557 			    slab_tohashslab(slab), NULL, SKIP_NONE);
1558 		goto fail;
1559 	}
1560 	uma_total_inc(size);
1561 
1562 	/* For HASH zones all pages go to the same uma_domain. */
1563 	if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
1564 		domain = 0;
1565 
1566 	/* Point the slab into the allocated memory */
1567 	if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE))
1568 		slab = (uma_slab_t )(mem + keg->uk_pgoff);
1569 	else
1570 		slab_tohashslab(slab)->uhs_data = mem;
1571 
1572 	if (keg->uk_flags & UMA_ZFLAG_VTOSLAB)
1573 		for (i = 0; i < keg->uk_ppera; i++)
1574 			vsetzoneslab((vm_offset_t)mem + (i * PAGE_SIZE),
1575 			    zone, slab);
1576 
1577 	slab->us_freecount = keg->uk_ipers;
1578 	slab->us_flags = sflags;
1579 	slab->us_domain = domain;
1580 
1581 	BIT_FILL(keg->uk_ipers, &slab->us_free);
1582 #ifdef INVARIANTS
1583 	BIT_ZERO(keg->uk_ipers, slab_dbg_bits(slab, keg));
1584 #endif
1585 
1586 	if (keg->uk_init != NULL) {
1587 		for (i = 0; i < keg->uk_ipers; i++)
1588 			if (keg->uk_init(slab_item(slab, keg, i),
1589 			    keg->uk_size, flags) != 0)
1590 				break;
1591 		if (i != keg->uk_ipers) {
1592 			keg_free_slab(keg, slab, i);
1593 			goto fail;
1594 		}
1595 	}
1596 	KEG_LOCK(keg, domain);
1597 
1598 	CTR3(KTR_UMA, "keg_alloc_slab: allocated slab %p for %s(%p)",
1599 	    slab, keg->uk_name, keg);
1600 
1601 	if (keg->uk_flags & UMA_ZFLAG_HASH)
1602 		UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
1603 
1604 	/*
1605 	 * If we got a slab here it's safe to mark it partially used
1606 	 * and return.  We assume that the caller is going to remove
1607 	 * at least one item.
1608 	 */
1609 	dom = &keg->uk_domain[domain];
1610 	LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
1611 	dom->ud_pages += keg->uk_ppera;
1612 	dom->ud_free_items += keg->uk_ipers;
1613 
1614 	return (slab);
1615 
1616 fail:
1617 	return (NULL);
1618 }
1619 
1620 /*
1621  * This function is intended to be used early on in place of page_alloc() so
1622  * that we may use the boot time page cache to satisfy allocations before
1623  * the VM is ready.
1624  */
1625 static void *
1626 startup_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1627     int wait)
1628 {
1629 	vm_paddr_t pa;
1630 	vm_page_t m;
1631 	void *mem;
1632 	int pages;
1633 	int i;
1634 
1635 	pages = howmany(bytes, PAGE_SIZE);
1636 	KASSERT(pages > 0, ("%s can't reserve 0 pages", __func__));
1637 
1638 	*pflag = UMA_SLAB_BOOT;
1639 	m = vm_page_alloc_contig_domain(NULL, 0, domain,
1640 	    malloc2vm_flags(wait) | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED, pages,
1641 	    (vm_paddr_t)0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT);
1642 	if (m == NULL)
1643 		return (NULL);
1644 
1645 	pa = VM_PAGE_TO_PHYS(m);
1646 	for (i = 0; i < pages; i++, pa += PAGE_SIZE) {
1647 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
1648     defined(__riscv) || defined(__powerpc64__)
1649 		if ((wait & M_NODUMP) == 0)
1650 			dump_add_page(pa);
1651 #endif
1652 	}
1653 	/* Allocate KVA and indirectly advance bootmem. */
1654 	mem = (void *)pmap_map(&bootmem, m->phys_addr,
1655 	    m->phys_addr + (pages * PAGE_SIZE), VM_PROT_READ | VM_PROT_WRITE);
1656         if ((wait & M_ZERO) != 0)
1657                 bzero(mem, pages * PAGE_SIZE);
1658 
1659         return (mem);
1660 }
1661 
1662 static void
1663 startup_free(void *mem, vm_size_t bytes)
1664 {
1665 	vm_offset_t va;
1666 	vm_page_t m;
1667 
1668 	va = (vm_offset_t)mem;
1669 	m = PHYS_TO_VM_PAGE(pmap_kextract(va));
1670 	pmap_remove(kernel_pmap, va, va + bytes);
1671 	for (; bytes != 0; bytes -= PAGE_SIZE, m++) {
1672 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
1673     defined(__riscv) || defined(__powerpc64__)
1674 		dump_drop_page(VM_PAGE_TO_PHYS(m));
1675 #endif
1676 		vm_page_unwire_noq(m);
1677 		vm_page_free(m);
1678 	}
1679 }
1680 
1681 /*
1682  * Allocates a number of pages from the system
1683  *
1684  * Arguments:
1685  *	bytes  The number of bytes requested
1686  *	wait  Shall we wait?
1687  *
1688  * Returns:
1689  *	A pointer to the alloced memory or possibly
1690  *	NULL if M_NOWAIT is set.
1691  */
1692 static void *
1693 page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1694     int wait)
1695 {
1696 	void *p;	/* Returned page */
1697 
1698 	*pflag = UMA_SLAB_KERNEL;
1699 	p = (void *)kmem_malloc_domainset(DOMAINSET_FIXED(domain), bytes, wait);
1700 
1701 	return (p);
1702 }
1703 
1704 static void *
1705 pcpu_page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1706     int wait)
1707 {
1708 	struct pglist alloctail;
1709 	vm_offset_t addr, zkva;
1710 	int cpu, flags;
1711 	vm_page_t p, p_next;
1712 #ifdef NUMA
1713 	struct pcpu *pc;
1714 #endif
1715 
1716 	MPASS(bytes == (mp_maxid + 1) * PAGE_SIZE);
1717 
1718 	TAILQ_INIT(&alloctail);
1719 	flags = VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | VM_ALLOC_NOOBJ |
1720 	    malloc2vm_flags(wait);
1721 	*pflag = UMA_SLAB_KERNEL;
1722 	for (cpu = 0; cpu <= mp_maxid; cpu++) {
1723 		if (CPU_ABSENT(cpu)) {
1724 			p = vm_page_alloc(NULL, 0, flags);
1725 		} else {
1726 #ifndef NUMA
1727 			p = vm_page_alloc(NULL, 0, flags);
1728 #else
1729 			pc = pcpu_find(cpu);
1730 			if (__predict_false(VM_DOMAIN_EMPTY(pc->pc_domain)))
1731 				p = NULL;
1732 			else
1733 				p = vm_page_alloc_domain(NULL, 0,
1734 				    pc->pc_domain, flags);
1735 			if (__predict_false(p == NULL))
1736 				p = vm_page_alloc(NULL, 0, flags);
1737 #endif
1738 		}
1739 		if (__predict_false(p == NULL))
1740 			goto fail;
1741 		TAILQ_INSERT_TAIL(&alloctail, p, listq);
1742 	}
1743 	if ((addr = kva_alloc(bytes)) == 0)
1744 		goto fail;
1745 	zkva = addr;
1746 	TAILQ_FOREACH(p, &alloctail, listq) {
1747 		pmap_qenter(zkva, &p, 1);
1748 		zkva += PAGE_SIZE;
1749 	}
1750 	return ((void*)addr);
1751 fail:
1752 	TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1753 		vm_page_unwire_noq(p);
1754 		vm_page_free(p);
1755 	}
1756 	return (NULL);
1757 }
1758 
1759 /*
1760  * Allocates a number of pages from within an object
1761  *
1762  * Arguments:
1763  *	bytes  The number of bytes requested
1764  *	wait   Shall we wait?
1765  *
1766  * Returns:
1767  *	A pointer to the alloced memory or possibly
1768  *	NULL if M_NOWAIT is set.
1769  */
1770 static void *
1771 noobj_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags,
1772     int wait)
1773 {
1774 	TAILQ_HEAD(, vm_page) alloctail;
1775 	u_long npages;
1776 	vm_offset_t retkva, zkva;
1777 	vm_page_t p, p_next;
1778 	uma_keg_t keg;
1779 
1780 	TAILQ_INIT(&alloctail);
1781 	keg = zone->uz_keg;
1782 
1783 	npages = howmany(bytes, PAGE_SIZE);
1784 	while (npages > 0) {
1785 		p = vm_page_alloc_domain(NULL, 0, domain, VM_ALLOC_INTERRUPT |
1786 		    VM_ALLOC_WIRED | VM_ALLOC_NOOBJ |
1787 		    ((wait & M_WAITOK) != 0 ? VM_ALLOC_WAITOK :
1788 		    VM_ALLOC_NOWAIT));
1789 		if (p != NULL) {
1790 			/*
1791 			 * Since the page does not belong to an object, its
1792 			 * listq is unused.
1793 			 */
1794 			TAILQ_INSERT_TAIL(&alloctail, p, listq);
1795 			npages--;
1796 			continue;
1797 		}
1798 		/*
1799 		 * Page allocation failed, free intermediate pages and
1800 		 * exit.
1801 		 */
1802 		TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1803 			vm_page_unwire_noq(p);
1804 			vm_page_free(p);
1805 		}
1806 		return (NULL);
1807 	}
1808 	*flags = UMA_SLAB_PRIV;
1809 	zkva = keg->uk_kva +
1810 	    atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
1811 	retkva = zkva;
1812 	TAILQ_FOREACH(p, &alloctail, listq) {
1813 		pmap_qenter(zkva, &p, 1);
1814 		zkva += PAGE_SIZE;
1815 	}
1816 
1817 	return ((void *)retkva);
1818 }
1819 
1820 /*
1821  * Allocate physically contiguous pages.
1822  */
1823 static void *
1824 contig_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1825     int wait)
1826 {
1827 
1828 	*pflag = UMA_SLAB_KERNEL;
1829 	return ((void *)kmem_alloc_contig_domainset(DOMAINSET_FIXED(domain),
1830 	    bytes, wait, 0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT));
1831 }
1832 
1833 /*
1834  * Frees a number of pages to the system
1835  *
1836  * Arguments:
1837  *	mem   A pointer to the memory to be freed
1838  *	size  The size of the memory being freed
1839  *	flags The original p->us_flags field
1840  *
1841  * Returns:
1842  *	Nothing
1843  */
1844 static void
1845 page_free(void *mem, vm_size_t size, uint8_t flags)
1846 {
1847 
1848 	if ((flags & UMA_SLAB_BOOT) != 0) {
1849 		startup_free(mem, size);
1850 		return;
1851 	}
1852 
1853 	KASSERT((flags & UMA_SLAB_KERNEL) != 0,
1854 	    ("UMA: page_free used with invalid flags %x", flags));
1855 
1856 	kmem_free((vm_offset_t)mem, size);
1857 }
1858 
1859 /*
1860  * Frees pcpu zone allocations
1861  *
1862  * Arguments:
1863  *	mem   A pointer to the memory to be freed
1864  *	size  The size of the memory being freed
1865  *	flags The original p->us_flags field
1866  *
1867  * Returns:
1868  *	Nothing
1869  */
1870 static void
1871 pcpu_page_free(void *mem, vm_size_t size, uint8_t flags)
1872 {
1873 	vm_offset_t sva, curva;
1874 	vm_paddr_t paddr;
1875 	vm_page_t m;
1876 
1877 	MPASS(size == (mp_maxid+1)*PAGE_SIZE);
1878 
1879 	if ((flags & UMA_SLAB_BOOT) != 0) {
1880 		startup_free(mem, size);
1881 		return;
1882 	}
1883 
1884 	sva = (vm_offset_t)mem;
1885 	for (curva = sva; curva < sva + size; curva += PAGE_SIZE) {
1886 		paddr = pmap_kextract(curva);
1887 		m = PHYS_TO_VM_PAGE(paddr);
1888 		vm_page_unwire_noq(m);
1889 		vm_page_free(m);
1890 	}
1891 	pmap_qremove(sva, size >> PAGE_SHIFT);
1892 	kva_free(sva, size);
1893 }
1894 
1895 
1896 /*
1897  * Zero fill initializer
1898  *
1899  * Arguments/Returns follow uma_init specifications
1900  */
1901 static int
1902 zero_init(void *mem, int size, int flags)
1903 {
1904 	bzero(mem, size);
1905 	return (0);
1906 }
1907 
1908 #ifdef INVARIANTS
1909 static struct noslabbits *
1910 slab_dbg_bits(uma_slab_t slab, uma_keg_t keg)
1911 {
1912 
1913 	return ((void *)((char *)&slab->us_free + BITSET_SIZE(keg->uk_ipers)));
1914 }
1915 #endif
1916 
1917 /*
1918  * Actual size of embedded struct slab (!OFFPAGE).
1919  */
1920 static size_t
1921 slab_sizeof(int nitems)
1922 {
1923 	size_t s;
1924 
1925 	s = sizeof(struct uma_slab) + BITSET_SIZE(nitems) * SLAB_BITSETS;
1926 	return (roundup(s, UMA_ALIGN_PTR + 1));
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 struct keg_layout_result {
1970 	u_int format;
1971 	u_int slabsize;
1972 	u_int ipers;
1973 	u_int eff;
1974 };
1975 
1976 static void
1977 keg_layout_one(uma_keg_t keg, u_int rsize, u_int slabsize, u_int fmt,
1978     struct keg_layout_result *kl)
1979 {
1980 	u_int total;
1981 
1982 	kl->format = fmt;
1983 	kl->slabsize = slabsize;
1984 
1985 	/* Handle INTERNAL as inline with an extra page. */
1986 	if ((fmt & UMA_ZFLAG_INTERNAL) != 0) {
1987 		kl->format &= ~UMA_ZFLAG_INTERNAL;
1988 		kl->slabsize += PAGE_SIZE;
1989 	}
1990 
1991 	kl->ipers = slab_ipers_hdr(keg->uk_size, rsize, kl->slabsize,
1992 	    (fmt & UMA_ZFLAG_OFFPAGE) == 0);
1993 
1994 	/* Account for memory used by an offpage slab header. */
1995 	total = kl->slabsize;
1996 	if ((fmt & UMA_ZFLAG_OFFPAGE) != 0)
1997 		total += slabzone(kl->ipers)->uz_keg->uk_rsize;
1998 
1999 	kl->eff = UMA_FRAC_FIXPT(kl->ipers * rsize, total);
2000 }
2001 
2002 /*
2003  * Determine the format of a uma keg.  This determines where the slab header
2004  * will be placed (inline or offpage) and calculates ipers, rsize, and ppera.
2005  *
2006  * Arguments
2007  *	keg  The zone we should initialize
2008  *
2009  * Returns
2010  *	Nothing
2011  */
2012 static void
2013 keg_layout(uma_keg_t keg)
2014 {
2015 	struct keg_layout_result kl = {}, kl_tmp;
2016 	u_int fmts[2];
2017 	u_int alignsize;
2018 	u_int nfmt;
2019 	u_int pages;
2020 	u_int rsize;
2021 	u_int slabsize;
2022 	u_int i, j;
2023 
2024 	KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
2025 	    (keg->uk_size <= UMA_PCPU_ALLOC_SIZE &&
2026 	     (keg->uk_flags & UMA_ZONE_CACHESPREAD) == 0),
2027 	    ("%s: cannot configure for PCPU: keg=%s, size=%u, flags=0x%b",
2028 	     __func__, keg->uk_name, keg->uk_size, keg->uk_flags,
2029 	     PRINT_UMA_ZFLAGS));
2030 	KASSERT((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) == 0 ||
2031 	    (keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0,
2032 	    ("%s: incompatible flags 0x%b", __func__, keg->uk_flags,
2033 	     PRINT_UMA_ZFLAGS));
2034 
2035 	alignsize = keg->uk_align + 1;
2036 
2037 	/*
2038 	 * Calculate the size of each allocation (rsize) according to
2039 	 * alignment.  If the requested size is smaller than we have
2040 	 * allocation bits for we round it up.
2041 	 */
2042 	rsize = MAX(keg->uk_size, UMA_SMALLEST_UNIT);
2043 	rsize = roundup2(rsize, alignsize);
2044 
2045 	if ((keg->uk_flags & UMA_ZONE_CACHESPREAD) != 0) {
2046 		/*
2047 		 * We want one item to start on every align boundary in a page.
2048 		 * To do this we will span pages.  We will also extend the item
2049 		 * by the size of align if it is an even multiple of align.
2050 		 * Otherwise, it would fall on the same boundary every time.
2051 		 */
2052 		if ((rsize & alignsize) == 0)
2053 			rsize += alignsize;
2054 		slabsize = rsize * (PAGE_SIZE / alignsize);
2055 		slabsize = MIN(slabsize, rsize * SLAB_MAX_SETSIZE);
2056 		slabsize = MIN(slabsize, UMA_CACHESPREAD_MAX_SIZE);
2057 		slabsize = round_page(slabsize);
2058 	} else {
2059 		/*
2060 		 * Start with a slab size of as many pages as it takes to
2061 		 * represent a single item.  We will try to fit as many
2062 		 * additional items into the slab as possible.
2063 		 */
2064 		slabsize = round_page(keg->uk_size);
2065 	}
2066 
2067 	/* Build a list of all of the available formats for this keg. */
2068 	nfmt = 0;
2069 
2070 	/* Evaluate an inline slab layout. */
2071 	if ((keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0)
2072 		fmts[nfmt++] = 0;
2073 
2074 	/* TODO: vm_page-embedded slab. */
2075 
2076 	/*
2077 	 * We can't do OFFPAGE if we're internal or if we've been
2078 	 * asked to not go to the VM for buckets.  If we do this we
2079 	 * may end up going to the VM for slabs which we do not want
2080 	 * to do if we're UMA_ZONE_VM, which clearly forbids it.
2081 	 * In those cases, evaluate a pseudo-format called INTERNAL
2082 	 * which has an inline slab header and one extra page to
2083 	 * guarantee that it fits.
2084 	 *
2085 	 * Otherwise, see if using an OFFPAGE slab will improve our
2086 	 * efficiency.
2087 	 */
2088 	if ((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) != 0)
2089 		fmts[nfmt++] = UMA_ZFLAG_INTERNAL;
2090 	else
2091 		fmts[nfmt++] = UMA_ZFLAG_OFFPAGE;
2092 
2093 	/*
2094 	 * Choose a slab size and format which satisfy the minimum efficiency.
2095 	 * Prefer the smallest slab size that meets the constraints.
2096 	 *
2097 	 * Start with a minimum slab size, to accommodate CACHESPREAD.  Then,
2098 	 * for small items (up to PAGE_SIZE), the iteration increment is one
2099 	 * page; and for large items, the increment is one item.
2100 	 */
2101 	i = (slabsize + rsize - keg->uk_size) / MAX(PAGE_SIZE, rsize);
2102 	KASSERT(i >= 1, ("keg %s(%p) flags=0x%b slabsize=%u, rsize=%u, i=%u",
2103 	    keg->uk_name, keg, keg->uk_flags, PRINT_UMA_ZFLAGS, slabsize,
2104 	    rsize, i));
2105 	for ( ; ; i++) {
2106 		slabsize = (rsize <= PAGE_SIZE) ? ptoa(i) :
2107 		    round_page(rsize * (i - 1) + keg->uk_size);
2108 
2109 		for (j = 0; j < nfmt; j++) {
2110 			/* Only if we have no viable format yet. */
2111 			if ((fmts[j] & UMA_ZFLAG_INTERNAL) != 0 &&
2112 			    kl.ipers > 0)
2113 				continue;
2114 
2115 			keg_layout_one(keg, rsize, slabsize, fmts[j], &kl_tmp);
2116 			if (kl_tmp.eff <= kl.eff)
2117 				continue;
2118 
2119 			kl = kl_tmp;
2120 
2121 			CTR6(KTR_UMA, "keg %s layout: format %#x "
2122 			    "(ipers %u * rsize %u) / slabsize %#x = %u%% eff",
2123 			    keg->uk_name, kl.format, kl.ipers, rsize,
2124 			    kl.slabsize, UMA_FIXPT_PCT(kl.eff));
2125 
2126 			/* Stop when we reach the minimum efficiency. */
2127 			if (kl.eff >= UMA_MIN_EFF)
2128 				break;
2129 		}
2130 
2131 		if (kl.eff >= UMA_MIN_EFF || !multipage_slabs ||
2132 		    slabsize >= SLAB_MAX_SETSIZE * rsize ||
2133 		    (keg->uk_flags & (UMA_ZONE_PCPU | UMA_ZONE_CONTIG)) != 0)
2134 			break;
2135 	}
2136 
2137 	pages = atop(kl.slabsize);
2138 	if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
2139 		pages *= mp_maxid + 1;
2140 
2141 	keg->uk_rsize = rsize;
2142 	keg->uk_ipers = kl.ipers;
2143 	keg->uk_ppera = pages;
2144 	keg->uk_flags |= kl.format;
2145 
2146 	/*
2147 	 * How do we find the slab header if it is offpage or if not all item
2148 	 * start addresses are in the same page?  We could solve the latter
2149 	 * case with vaddr alignment, but we don't.
2150 	 */
2151 	if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0 ||
2152 	    (keg->uk_ipers - 1) * rsize >= PAGE_SIZE) {
2153 		if ((keg->uk_flags & UMA_ZONE_NOTPAGE) != 0)
2154 			keg->uk_flags |= UMA_ZFLAG_HASH;
2155 		else
2156 			keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2157 	}
2158 
2159 	CTR6(KTR_UMA, "%s: keg=%s, flags=%#x, rsize=%u, ipers=%u, ppera=%u",
2160 	    __func__, keg->uk_name, keg->uk_flags, rsize, keg->uk_ipers,
2161 	    pages);
2162 	KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_MAX_SETSIZE,
2163 	    ("%s: keg=%s, flags=0x%b, rsize=%u, ipers=%u, ppera=%u", __func__,
2164 	     keg->uk_name, keg->uk_flags, PRINT_UMA_ZFLAGS, rsize,
2165 	     keg->uk_ipers, pages));
2166 }
2167 
2168 /*
2169  * Keg header ctor.  This initializes all fields, locks, etc.  And inserts
2170  * the keg onto the global keg list.
2171  *
2172  * Arguments/Returns follow uma_ctor specifications
2173  *	udata  Actually uma_kctor_args
2174  */
2175 static int
2176 keg_ctor(void *mem, int size, void *udata, int flags)
2177 {
2178 	struct uma_kctor_args *arg = udata;
2179 	uma_keg_t keg = mem;
2180 	uma_zone_t zone;
2181 	int i;
2182 
2183 	bzero(keg, size);
2184 	keg->uk_size = arg->size;
2185 	keg->uk_init = arg->uminit;
2186 	keg->uk_fini = arg->fini;
2187 	keg->uk_align = arg->align;
2188 	keg->uk_reserve = 0;
2189 	keg->uk_flags = arg->flags;
2190 
2191 	/*
2192 	 * We use a global round-robin policy by default.  Zones with
2193 	 * UMA_ZONE_FIRSTTOUCH set will use first-touch instead, in which
2194 	 * case the iterator is never run.
2195 	 */
2196 	keg->uk_dr.dr_policy = DOMAINSET_RR();
2197 	keg->uk_dr.dr_iter = 0;
2198 
2199 	/*
2200 	 * The master zone is passed to us at keg-creation time.
2201 	 */
2202 	zone = arg->zone;
2203 	keg->uk_name = zone->uz_name;
2204 
2205 	if (arg->flags & UMA_ZONE_ZINIT)
2206 		keg->uk_init = zero_init;
2207 
2208 	if (arg->flags & UMA_ZONE_MALLOC)
2209 		keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2210 
2211 #ifndef SMP
2212 	keg->uk_flags &= ~UMA_ZONE_PCPU;
2213 #endif
2214 
2215 	keg_layout(keg);
2216 
2217 	/*
2218 	 * Use a first-touch NUMA policy for kegs that pmap_extract() will
2219 	 * work on.  Use round-robin for everything else.
2220 	 *
2221 	 * Zones may override the default by specifying either.
2222 	 */
2223 #ifdef NUMA
2224 	if ((keg->uk_flags &
2225 	    (UMA_ZONE_ROUNDROBIN | UMA_ZFLAG_CACHE | UMA_ZONE_NOTPAGE)) == 0)
2226 		keg->uk_flags |= UMA_ZONE_FIRSTTOUCH;
2227 	else if ((keg->uk_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2228 		keg->uk_flags |= UMA_ZONE_ROUNDROBIN;
2229 #endif
2230 
2231 	/*
2232 	 * If we haven't booted yet we need allocations to go through the
2233 	 * startup cache until the vm is ready.
2234 	 */
2235 #ifdef UMA_MD_SMALL_ALLOC
2236 	if (keg->uk_ppera == 1)
2237 		keg->uk_allocf = uma_small_alloc;
2238 	else
2239 #endif
2240 	if (booted < BOOT_KVA)
2241 		keg->uk_allocf = startup_alloc;
2242 	else if (keg->uk_flags & UMA_ZONE_PCPU)
2243 		keg->uk_allocf = pcpu_page_alloc;
2244 	else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 && keg->uk_ppera > 1)
2245 		keg->uk_allocf = contig_alloc;
2246 	else
2247 		keg->uk_allocf = page_alloc;
2248 #ifdef UMA_MD_SMALL_ALLOC
2249 	if (keg->uk_ppera == 1)
2250 		keg->uk_freef = uma_small_free;
2251 	else
2252 #endif
2253 	if (keg->uk_flags & UMA_ZONE_PCPU)
2254 		keg->uk_freef = pcpu_page_free;
2255 	else
2256 		keg->uk_freef = page_free;
2257 
2258 	/*
2259 	 * Initialize keg's locks.
2260 	 */
2261 	for (i = 0; i < vm_ndomains; i++)
2262 		KEG_LOCK_INIT(keg, i, (arg->flags & UMA_ZONE_MTXCLASS));
2263 
2264 	/*
2265 	 * If we're putting the slab header in the actual page we need to
2266 	 * figure out where in each page it goes.  See slab_sizeof
2267 	 * definition.
2268 	 */
2269 	if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE)) {
2270 		size_t shsize;
2271 
2272 		shsize = slab_sizeof(keg->uk_ipers);
2273 		keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - shsize;
2274 		/*
2275 		 * The only way the following is possible is if with our
2276 		 * UMA_ALIGN_PTR adjustments we are now bigger than
2277 		 * UMA_SLAB_SIZE.  I haven't checked whether this is
2278 		 * mathematically possible for all cases, so we make
2279 		 * sure here anyway.
2280 		 */
2281 		KASSERT(keg->uk_pgoff + shsize <= PAGE_SIZE * keg->uk_ppera,
2282 		    ("zone %s ipers %d rsize %d size %d slab won't fit",
2283 		    zone->uz_name, keg->uk_ipers, keg->uk_rsize, keg->uk_size));
2284 	}
2285 
2286 	if (keg->uk_flags & UMA_ZFLAG_HASH)
2287 		hash_alloc(&keg->uk_hash, 0);
2288 
2289 	CTR3(KTR_UMA, "keg_ctor %p zone %s(%p)", keg, zone->uz_name, zone);
2290 
2291 	LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
2292 
2293 	rw_wlock(&uma_rwlock);
2294 	LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
2295 	rw_wunlock(&uma_rwlock);
2296 	return (0);
2297 }
2298 
2299 static void
2300 zone_kva_available(uma_zone_t zone, void *unused)
2301 {
2302 	uma_keg_t keg;
2303 
2304 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
2305 		return;
2306 	KEG_GET(zone, keg);
2307 
2308 	if (keg->uk_allocf == startup_alloc) {
2309 		/* Switch to the real allocator. */
2310 		if (keg->uk_flags & UMA_ZONE_PCPU)
2311 			keg->uk_allocf = pcpu_page_alloc;
2312 		else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 &&
2313 		    keg->uk_ppera > 1)
2314 			keg->uk_allocf = contig_alloc;
2315 		else
2316 			keg->uk_allocf = page_alloc;
2317 	}
2318 }
2319 
2320 static void
2321 zone_alloc_counters(uma_zone_t zone, void *unused)
2322 {
2323 
2324 	zone->uz_allocs = counter_u64_alloc(M_WAITOK);
2325 	zone->uz_frees = counter_u64_alloc(M_WAITOK);
2326 	zone->uz_fails = counter_u64_alloc(M_WAITOK);
2327 	zone->uz_xdomain = counter_u64_alloc(M_WAITOK);
2328 }
2329 
2330 static void
2331 zone_alloc_sysctl(uma_zone_t zone, void *unused)
2332 {
2333 	uma_zone_domain_t zdom;
2334 	uma_domain_t dom;
2335 	uma_keg_t keg;
2336 	struct sysctl_oid *oid, *domainoid;
2337 	int domains, i, cnt;
2338 	static const char *nokeg = "cache zone";
2339 	char *c;
2340 
2341 	/*
2342 	 * Make a sysctl safe copy of the zone name by removing
2343 	 * any special characters and handling dups by appending
2344 	 * an index.
2345 	 */
2346 	if (zone->uz_namecnt != 0) {
2347 		/* Count the number of decimal digits and '_' separator. */
2348 		for (i = 1, cnt = zone->uz_namecnt; cnt != 0; i++)
2349 			cnt /= 10;
2350 		zone->uz_ctlname = malloc(strlen(zone->uz_name) + i + 1,
2351 		    M_UMA, M_WAITOK);
2352 		sprintf(zone->uz_ctlname, "%s_%d", zone->uz_name,
2353 		    zone->uz_namecnt);
2354 	} else
2355 		zone->uz_ctlname = strdup(zone->uz_name, M_UMA);
2356 	for (c = zone->uz_ctlname; *c != '\0'; c++)
2357 		if (strchr("./\\ -", *c) != NULL)
2358 			*c = '_';
2359 
2360 	/*
2361 	 * Basic parameters at the root.
2362 	 */
2363 	zone->uz_oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_vm_uma),
2364 	    OID_AUTO, zone->uz_ctlname, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2365 	oid = zone->uz_oid;
2366 	SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2367 	    "size", CTLFLAG_RD, &zone->uz_size, 0, "Allocation size");
2368 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2369 	    "flags", CTLFLAG_RD | CTLTYPE_STRING | CTLFLAG_MPSAFE,
2370 	    zone, 0, sysctl_handle_uma_zone_flags, "A",
2371 	    "Allocator configuration flags");
2372 	SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2373 	    "bucket_size", CTLFLAG_RD, &zone->uz_bucket_size, 0,
2374 	    "Desired per-cpu cache size");
2375 	SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2376 	    "bucket_size_max", CTLFLAG_RD, &zone->uz_bucket_size_max, 0,
2377 	    "Maximum allowed per-cpu cache size");
2378 
2379 	/*
2380 	 * keg if present.
2381 	 */
2382 	if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
2383 		domains = vm_ndomains;
2384 	else
2385 		domains = 1;
2386 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2387 	    "keg", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2388 	keg = zone->uz_keg;
2389 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0) {
2390 		SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2391 		    "name", CTLFLAG_RD, keg->uk_name, "Keg name");
2392 		SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2393 		    "rsize", CTLFLAG_RD, &keg->uk_rsize, 0,
2394 		    "Real object size with alignment");
2395 		SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2396 		    "ppera", CTLFLAG_RD, &keg->uk_ppera, 0,
2397 		    "pages per-slab allocation");
2398 		SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2399 		    "ipers", CTLFLAG_RD, &keg->uk_ipers, 0,
2400 		    "items available per-slab");
2401 		SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2402 		    "align", CTLFLAG_RD, &keg->uk_align, 0,
2403 		    "item alignment mask");
2404 		SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2405 		    "efficiency", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2406 		    keg, 0, sysctl_handle_uma_slab_efficiency, "I",
2407 		    "Slab utilization (100 - internal fragmentation %)");
2408 		domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(oid),
2409 		    OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2410 		for (i = 0; i < domains; i++) {
2411 			dom = &keg->uk_domain[i];
2412 			oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2413 			    OID_AUTO, VM_DOMAIN(i)->vmd_name,
2414 			    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2415 			SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2416 			    "pages", CTLFLAG_RD, &dom->ud_pages, 0,
2417 			    "Total pages currently allocated from VM");
2418 			SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2419 			    "free_items", CTLFLAG_RD, &dom->ud_free_items, 0,
2420 			    "items free in the slab layer");
2421 		}
2422 	} else
2423 		SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2424 		    "name", CTLFLAG_RD, nokeg, "Keg name");
2425 
2426 	/*
2427 	 * Information about zone limits.
2428 	 */
2429 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2430 	    "limit", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2431 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2432 	    "items", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2433 	    zone, 0, sysctl_handle_uma_zone_items, "QU",
2434 	    "current number of allocated items if limit is set");
2435 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2436 	    "max_items", CTLFLAG_RD, &zone->uz_max_items, 0,
2437 	    "Maximum number of cached items");
2438 	SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2439 	    "sleepers", CTLFLAG_RD, &zone->uz_sleepers, 0,
2440 	    "Number of threads sleeping at limit");
2441 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2442 	    "sleeps", CTLFLAG_RD, &zone->uz_sleeps, 0,
2443 	    "Total zone limit sleeps");
2444 	SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2445 	    "bucket_max", CTLFLAG_RD, &zone->uz_bucket_max, 0,
2446 	    "Maximum number of items in each domain's bucket cache");
2447 
2448 	/*
2449 	 * Per-domain zone information.
2450 	 */
2451 	domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid),
2452 	    OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2453 	for (i = 0; i < domains; i++) {
2454 		zdom = ZDOM_GET(zone, i);
2455 		oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2456 		    OID_AUTO, VM_DOMAIN(i)->vmd_name,
2457 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2458 		SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2459 		    "nitems", CTLFLAG_RD, &zdom->uzd_nitems,
2460 		    "number of items in this domain");
2461 		SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2462 		    "imax", CTLFLAG_RD, &zdom->uzd_imax,
2463 		    "maximum item count in this period");
2464 		SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2465 		    "imin", CTLFLAG_RD, &zdom->uzd_imin,
2466 		    "minimum item count in this period");
2467 		SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2468 		    "wss", CTLFLAG_RD, &zdom->uzd_wss,
2469 		    "Working set size");
2470 	}
2471 
2472 	/*
2473 	 * General statistics.
2474 	 */
2475 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2476 	    "stats", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2477 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2478 	    "current", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2479 	    zone, 1, sysctl_handle_uma_zone_cur, "I",
2480 	    "Current number of allocated items");
2481 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2482 	    "allocs", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2483 	    zone, 0, sysctl_handle_uma_zone_allocs, "QU",
2484 	    "Total allocation calls");
2485 	SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2486 	    "frees", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2487 	    zone, 0, sysctl_handle_uma_zone_frees, "QU",
2488 	    "Total free calls");
2489 	SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2490 	    "fails", CTLFLAG_RD, &zone->uz_fails,
2491 	    "Number of allocation failures");
2492 	SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2493 	    "xdomain", CTLFLAG_RD, &zone->uz_xdomain,
2494 	    "Free calls from the wrong domain");
2495 }
2496 
2497 struct uma_zone_count {
2498 	const char	*name;
2499 	int		count;
2500 };
2501 
2502 static void
2503 zone_count(uma_zone_t zone, void *arg)
2504 {
2505 	struct uma_zone_count *cnt;
2506 
2507 	cnt = arg;
2508 	/*
2509 	 * Some zones are rapidly created with identical names and
2510 	 * destroyed out of order.  This can lead to gaps in the count.
2511 	 * Use one greater than the maximum observed for this name.
2512 	 */
2513 	if (strcmp(zone->uz_name, cnt->name) == 0)
2514 		cnt->count = MAX(cnt->count,
2515 		    zone->uz_namecnt + 1);
2516 }
2517 
2518 static void
2519 zone_update_caches(uma_zone_t zone)
2520 {
2521 	int i;
2522 
2523 	for (i = 0; i <= mp_maxid; i++) {
2524 		cache_set_uz_size(&zone->uz_cpu[i], zone->uz_size);
2525 		cache_set_uz_flags(&zone->uz_cpu[i], zone->uz_flags);
2526 	}
2527 }
2528 
2529 /*
2530  * Zone header ctor.  This initializes all fields, locks, etc.
2531  *
2532  * Arguments/Returns follow uma_ctor specifications
2533  *	udata  Actually uma_zctor_args
2534  */
2535 static int
2536 zone_ctor(void *mem, int size, void *udata, int flags)
2537 {
2538 	struct uma_zone_count cnt;
2539 	struct uma_zctor_args *arg = udata;
2540 	uma_zone_domain_t zdom;
2541 	uma_zone_t zone = mem;
2542 	uma_zone_t z;
2543 	uma_keg_t keg;
2544 	int i;
2545 
2546 	bzero(zone, size);
2547 	zone->uz_name = arg->name;
2548 	zone->uz_ctor = arg->ctor;
2549 	zone->uz_dtor = arg->dtor;
2550 	zone->uz_init = NULL;
2551 	zone->uz_fini = NULL;
2552 	zone->uz_sleeps = 0;
2553 	zone->uz_bucket_size = 0;
2554 	zone->uz_bucket_size_min = 0;
2555 	zone->uz_bucket_size_max = BUCKET_MAX;
2556 	zone->uz_flags = (arg->flags & UMA_ZONE_SMR);
2557 	zone->uz_warning = NULL;
2558 	/* The domain structures follow the cpu structures. */
2559 	zone->uz_bucket_max = ULONG_MAX;
2560 	timevalclear(&zone->uz_ratecheck);
2561 
2562 	/* Count the number of duplicate names. */
2563 	cnt.name = arg->name;
2564 	cnt.count = 0;
2565 	zone_foreach(zone_count, &cnt);
2566 	zone->uz_namecnt = cnt.count;
2567 	ZONE_CROSS_LOCK_INIT(zone);
2568 
2569 	for (i = 0; i < vm_ndomains; i++) {
2570 		zdom = ZDOM_GET(zone, i);
2571 		ZDOM_LOCK_INIT(zone, zdom, (arg->flags & UMA_ZONE_MTXCLASS));
2572 		STAILQ_INIT(&zdom->uzd_buckets);
2573 	}
2574 
2575 #ifdef INVARIANTS
2576 	if (arg->uminit == trash_init && arg->fini == trash_fini)
2577 		zone->uz_flags |= UMA_ZFLAG_TRASH | UMA_ZFLAG_CTORDTOR;
2578 #endif
2579 
2580 	/*
2581 	 * This is a pure cache zone, no kegs.
2582 	 */
2583 	if (arg->import) {
2584 		KASSERT((arg->flags & UMA_ZFLAG_CACHE) != 0,
2585 		    ("zone_ctor: Import specified for non-cache zone."));
2586 		zone->uz_flags = arg->flags;
2587 		zone->uz_size = arg->size;
2588 		zone->uz_import = arg->import;
2589 		zone->uz_release = arg->release;
2590 		zone->uz_arg = arg->arg;
2591 #ifdef NUMA
2592 		/*
2593 		 * Cache zones are round-robin unless a policy is
2594 		 * specified because they may have incompatible
2595 		 * constraints.
2596 		 */
2597 		if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2598 			zone->uz_flags |= UMA_ZONE_ROUNDROBIN;
2599 #endif
2600 		rw_wlock(&uma_rwlock);
2601 		LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link);
2602 		rw_wunlock(&uma_rwlock);
2603 		goto out;
2604 	}
2605 
2606 	/*
2607 	 * Use the regular zone/keg/slab allocator.
2608 	 */
2609 	zone->uz_import = zone_import;
2610 	zone->uz_release = zone_release;
2611 	zone->uz_arg = zone;
2612 	keg = arg->keg;
2613 
2614 	if (arg->flags & UMA_ZONE_SECONDARY) {
2615 		KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
2616 		    ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
2617 		KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
2618 		zone->uz_init = arg->uminit;
2619 		zone->uz_fini = arg->fini;
2620 		zone->uz_flags |= UMA_ZONE_SECONDARY;
2621 		rw_wlock(&uma_rwlock);
2622 		ZONE_LOCK(zone);
2623 		LIST_FOREACH(z, &keg->uk_zones, uz_link) {
2624 			if (LIST_NEXT(z, uz_link) == NULL) {
2625 				LIST_INSERT_AFTER(z, zone, uz_link);
2626 				break;
2627 			}
2628 		}
2629 		ZONE_UNLOCK(zone);
2630 		rw_wunlock(&uma_rwlock);
2631 	} else if (keg == NULL) {
2632 		if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
2633 		    arg->align, arg->flags)) == NULL)
2634 			return (ENOMEM);
2635 	} else {
2636 		struct uma_kctor_args karg;
2637 		int error;
2638 
2639 		/* We should only be here from uma_startup() */
2640 		karg.size = arg->size;
2641 		karg.uminit = arg->uminit;
2642 		karg.fini = arg->fini;
2643 		karg.align = arg->align;
2644 		karg.flags = (arg->flags & ~UMA_ZONE_SMR);
2645 		karg.zone = zone;
2646 		error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
2647 		    flags);
2648 		if (error)
2649 			return (error);
2650 	}
2651 
2652 	/* Inherit properties from the keg. */
2653 	zone->uz_keg = keg;
2654 	zone->uz_size = keg->uk_size;
2655 	zone->uz_flags |= (keg->uk_flags &
2656 	    (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
2657 
2658 out:
2659 	if (booted >= BOOT_PCPU) {
2660 		zone_alloc_counters(zone, NULL);
2661 		if (booted >= BOOT_RUNNING)
2662 			zone_alloc_sysctl(zone, NULL);
2663 	} else {
2664 		zone->uz_allocs = EARLY_COUNTER;
2665 		zone->uz_frees = EARLY_COUNTER;
2666 		zone->uz_fails = EARLY_COUNTER;
2667 	}
2668 
2669 	/* Caller requests a private SMR context. */
2670 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
2671 		zone->uz_smr = smr_create(zone->uz_name, 0, 0);
2672 
2673 	KASSERT((arg->flags & (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET)) !=
2674 	    (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET),
2675 	    ("Invalid zone flag combination"));
2676 	if (arg->flags & UMA_ZFLAG_INTERNAL)
2677 		zone->uz_bucket_size_max = zone->uz_bucket_size = 0;
2678 	if ((arg->flags & UMA_ZONE_MAXBUCKET) != 0)
2679 		zone->uz_bucket_size = BUCKET_MAX;
2680 	else if ((arg->flags & UMA_ZONE_MINBUCKET) != 0)
2681 		zone->uz_bucket_size_max = zone->uz_bucket_size = BUCKET_MIN;
2682 	else if ((arg->flags & UMA_ZONE_NOBUCKET) != 0)
2683 		zone->uz_bucket_size = 0;
2684 	else
2685 		zone->uz_bucket_size = bucket_select(zone->uz_size);
2686 	zone->uz_bucket_size_min = zone->uz_bucket_size;
2687 	if (zone->uz_dtor != NULL || zone->uz_ctor != NULL)
2688 		zone->uz_flags |= UMA_ZFLAG_CTORDTOR;
2689 	zone_update_caches(zone);
2690 
2691 	return (0);
2692 }
2693 
2694 /*
2695  * Keg header dtor.  This frees all data, destroys locks, frees the hash
2696  * table and removes the keg from the global list.
2697  *
2698  * Arguments/Returns follow uma_dtor specifications
2699  *	udata  unused
2700  */
2701 static void
2702 keg_dtor(void *arg, int size, void *udata)
2703 {
2704 	uma_keg_t keg;
2705 	uint32_t free, pages;
2706 	int i;
2707 
2708 	keg = (uma_keg_t)arg;
2709 	free = pages = 0;
2710 	for (i = 0; i < vm_ndomains; i++) {
2711 		free += keg->uk_domain[i].ud_free_items;
2712 		pages += keg->uk_domain[i].ud_pages;
2713 		KEG_LOCK_FINI(keg, i);
2714 	}
2715 	if (pages != 0)
2716 		printf("Freed UMA keg (%s) was not empty (%u items). "
2717 		    " Lost %u pages of memory.\n",
2718 		    keg->uk_name ? keg->uk_name : "",
2719 		    pages / keg->uk_ppera * keg->uk_ipers - free, pages);
2720 
2721 	hash_free(&keg->uk_hash);
2722 }
2723 
2724 /*
2725  * Zone header dtor.
2726  *
2727  * Arguments/Returns follow uma_dtor specifications
2728  *	udata  unused
2729  */
2730 static void
2731 zone_dtor(void *arg, int size, void *udata)
2732 {
2733 	uma_zone_t zone;
2734 	uma_keg_t keg;
2735 	int i;
2736 
2737 	zone = (uma_zone_t)arg;
2738 
2739 	sysctl_remove_oid(zone->uz_oid, 1, 1);
2740 
2741 	if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
2742 		cache_drain(zone);
2743 
2744 	rw_wlock(&uma_rwlock);
2745 	LIST_REMOVE(zone, uz_link);
2746 	rw_wunlock(&uma_rwlock);
2747 	zone_reclaim(zone, M_WAITOK, true);
2748 
2749 	/*
2750 	 * We only destroy kegs from non secondary/non cache zones.
2751 	 */
2752 	if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) {
2753 		keg = zone->uz_keg;
2754 		rw_wlock(&uma_rwlock);
2755 		LIST_REMOVE(keg, uk_link);
2756 		rw_wunlock(&uma_rwlock);
2757 		zone_free_item(kegs, keg, NULL, SKIP_NONE);
2758 	}
2759 	counter_u64_free(zone->uz_allocs);
2760 	counter_u64_free(zone->uz_frees);
2761 	counter_u64_free(zone->uz_fails);
2762 	counter_u64_free(zone->uz_xdomain);
2763 	free(zone->uz_ctlname, M_UMA);
2764 	for (i = 0; i < vm_ndomains; i++)
2765 		ZDOM_LOCK_FINI(ZDOM_GET(zone, i));
2766 	ZONE_CROSS_LOCK_FINI(zone);
2767 }
2768 
2769 static void
2770 zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *arg), void *arg)
2771 {
2772 	uma_keg_t keg;
2773 	uma_zone_t zone;
2774 
2775 	LIST_FOREACH(keg, &uma_kegs, uk_link) {
2776 		LIST_FOREACH(zone, &keg->uk_zones, uz_link)
2777 			zfunc(zone, arg);
2778 	}
2779 	LIST_FOREACH(zone, &uma_cachezones, uz_link)
2780 		zfunc(zone, arg);
2781 }
2782 
2783 /*
2784  * Traverses every zone in the system and calls a callback
2785  *
2786  * Arguments:
2787  *	zfunc  A pointer to a function which accepts a zone
2788  *		as an argument.
2789  *
2790  * Returns:
2791  *	Nothing
2792  */
2793 static void
2794 zone_foreach(void (*zfunc)(uma_zone_t, void *arg), void *arg)
2795 {
2796 
2797 	rw_rlock(&uma_rwlock);
2798 	zone_foreach_unlocked(zfunc, arg);
2799 	rw_runlock(&uma_rwlock);
2800 }
2801 
2802 /*
2803  * Initialize the kernel memory allocator.  This is done after pages can be
2804  * allocated but before general KVA is available.
2805  */
2806 void
2807 uma_startup1(vm_offset_t virtual_avail)
2808 {
2809 	struct uma_zctor_args args;
2810 	size_t ksize, zsize, size;
2811 	uma_keg_t masterkeg;
2812 	uintptr_t m;
2813 	int domain;
2814 	uint8_t pflag;
2815 
2816 	bootstart = bootmem = virtual_avail;
2817 
2818 	rw_init(&uma_rwlock, "UMA lock");
2819 	sx_init(&uma_reclaim_lock, "umareclaim");
2820 
2821 	ksize = sizeof(struct uma_keg) +
2822 	    (sizeof(struct uma_domain) * vm_ndomains);
2823 	ksize = roundup(ksize, UMA_SUPER_ALIGN);
2824 	zsize = sizeof(struct uma_zone) +
2825 	    (sizeof(struct uma_cache) * (mp_maxid + 1)) +
2826 	    (sizeof(struct uma_zone_domain) * vm_ndomains);
2827 	zsize = roundup(zsize, UMA_SUPER_ALIGN);
2828 
2829 	/* Allocate the zone of zones, zone of kegs, and zone of zones keg. */
2830 	size = (zsize * 2) + ksize;
2831 	for (domain = 0; domain < vm_ndomains; domain++) {
2832 		m = (uintptr_t)startup_alloc(NULL, size, domain, &pflag,
2833 		    M_NOWAIT | M_ZERO);
2834 		if (m != 0)
2835 			break;
2836 	}
2837 	zones = (uma_zone_t)m;
2838 	m += zsize;
2839 	kegs = (uma_zone_t)m;
2840 	m += zsize;
2841 	masterkeg = (uma_keg_t)m;
2842 
2843 	/* "manually" create the initial zone */
2844 	memset(&args, 0, sizeof(args));
2845 	args.name = "UMA Kegs";
2846 	args.size = ksize;
2847 	args.ctor = keg_ctor;
2848 	args.dtor = keg_dtor;
2849 	args.uminit = zero_init;
2850 	args.fini = NULL;
2851 	args.keg = masterkeg;
2852 	args.align = UMA_SUPER_ALIGN - 1;
2853 	args.flags = UMA_ZFLAG_INTERNAL;
2854 	zone_ctor(kegs, zsize, &args, M_WAITOK);
2855 
2856 	args.name = "UMA Zones";
2857 	args.size = zsize;
2858 	args.ctor = zone_ctor;
2859 	args.dtor = zone_dtor;
2860 	args.uminit = zero_init;
2861 	args.fini = NULL;
2862 	args.keg = NULL;
2863 	args.align = UMA_SUPER_ALIGN - 1;
2864 	args.flags = UMA_ZFLAG_INTERNAL;
2865 	zone_ctor(zones, zsize, &args, M_WAITOK);
2866 
2867 	/* Now make zones for slab headers */
2868 	slabzones[0] = uma_zcreate("UMA Slabs 0", SLABZONE0_SIZE,
2869 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2870 	slabzones[1] = uma_zcreate("UMA Slabs 1", SLABZONE1_SIZE,
2871 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2872 
2873 	hashzone = uma_zcreate("UMA Hash",
2874 	    sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
2875 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
2876 
2877 	bucket_init();
2878 	smr_init();
2879 }
2880 
2881 #ifndef UMA_MD_SMALL_ALLOC
2882 extern void vm_radix_reserve_kva(void);
2883 #endif
2884 
2885 /*
2886  * Advertise the availability of normal kva allocations and switch to
2887  * the default back-end allocator.  Marks the KVA we consumed on startup
2888  * as used in the map.
2889  */
2890 void
2891 uma_startup2(void)
2892 {
2893 
2894 	if (bootstart != bootmem) {
2895 		vm_map_lock(kernel_map);
2896 		(void)vm_map_insert(kernel_map, NULL, 0, bootstart, bootmem,
2897 		    VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT);
2898 		vm_map_unlock(kernel_map);
2899 	}
2900 
2901 #ifndef UMA_MD_SMALL_ALLOC
2902 	/* Set up radix zone to use noobj_alloc. */
2903 	vm_radix_reserve_kva();
2904 #endif
2905 
2906 	booted = BOOT_KVA;
2907 	zone_foreach_unlocked(zone_kva_available, NULL);
2908 	bucket_enable();
2909 }
2910 
2911 /*
2912  * Allocate counters as early as possible so that boot-time allocations are
2913  * accounted more precisely.
2914  */
2915 static void
2916 uma_startup_pcpu(void *arg __unused)
2917 {
2918 
2919 	zone_foreach_unlocked(zone_alloc_counters, NULL);
2920 	booted = BOOT_PCPU;
2921 }
2922 SYSINIT(uma_startup_pcpu, SI_SUB_COUNTER, SI_ORDER_ANY, uma_startup_pcpu, NULL);
2923 
2924 /*
2925  * Finish our initialization steps.
2926  */
2927 static void
2928 uma_startup3(void *arg __unused)
2929 {
2930 
2931 #ifdef INVARIANTS
2932 	TUNABLE_INT_FETCH("vm.debug.divisor", &dbg_divisor);
2933 	uma_dbg_cnt = counter_u64_alloc(M_WAITOK);
2934 	uma_skip_cnt = counter_u64_alloc(M_WAITOK);
2935 #endif
2936 	zone_foreach_unlocked(zone_alloc_sysctl, NULL);
2937 	callout_init(&uma_callout, 1);
2938 	callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
2939 	booted = BOOT_RUNNING;
2940 
2941 	EVENTHANDLER_REGISTER(shutdown_post_sync, uma_shutdown, NULL,
2942 	    EVENTHANDLER_PRI_FIRST);
2943 }
2944 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
2945 
2946 static void
2947 uma_shutdown(void)
2948 {
2949 
2950 	booted = BOOT_SHUTDOWN;
2951 }
2952 
2953 static uma_keg_t
2954 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
2955 		int align, uint32_t flags)
2956 {
2957 	struct uma_kctor_args args;
2958 
2959 	args.size = size;
2960 	args.uminit = uminit;
2961 	args.fini = fini;
2962 	args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
2963 	args.flags = flags;
2964 	args.zone = zone;
2965 	return (zone_alloc_item(kegs, &args, UMA_ANYDOMAIN, M_WAITOK));
2966 }
2967 
2968 /* Public functions */
2969 /* See uma.h */
2970 void
2971 uma_set_align(int align)
2972 {
2973 
2974 	if (align != UMA_ALIGN_CACHE)
2975 		uma_align_cache = align;
2976 }
2977 
2978 /* See uma.h */
2979 uma_zone_t
2980 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
2981 		uma_init uminit, uma_fini fini, int align, uint32_t flags)
2982 
2983 {
2984 	struct uma_zctor_args args;
2985 	uma_zone_t res;
2986 
2987 	KASSERT(powerof2(align + 1), ("invalid zone alignment %d for \"%s\"",
2988 	    align, name));
2989 
2990 	/* This stuff is essential for the zone ctor */
2991 	memset(&args, 0, sizeof(args));
2992 	args.name = name;
2993 	args.size = size;
2994 	args.ctor = ctor;
2995 	args.dtor = dtor;
2996 	args.uminit = uminit;
2997 	args.fini = fini;
2998 #ifdef  INVARIANTS
2999 	/*
3000 	 * Inject procedures which check for memory use after free if we are
3001 	 * allowed to scramble the memory while it is not allocated.  This
3002 	 * requires that: UMA is actually able to access the memory, no init
3003 	 * or fini procedures, no dependency on the initial value of the
3004 	 * memory, and no (legitimate) use of the memory after free.  Note,
3005 	 * the ctor and dtor do not need to be empty.
3006 	 */
3007 	if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOTOUCH |
3008 	    UMA_ZONE_NOFREE))) && uminit == NULL && fini == NULL) {
3009 		args.uminit = trash_init;
3010 		args.fini = trash_fini;
3011 	}
3012 #endif
3013 	args.align = align;
3014 	args.flags = flags;
3015 	args.keg = NULL;
3016 
3017 	sx_slock(&uma_reclaim_lock);
3018 	res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3019 	sx_sunlock(&uma_reclaim_lock);
3020 
3021 	return (res);
3022 }
3023 
3024 /* See uma.h */
3025 uma_zone_t
3026 uma_zsecond_create(const char *name, uma_ctor ctor, uma_dtor dtor,
3027     uma_init zinit, uma_fini zfini, uma_zone_t master)
3028 {
3029 	struct uma_zctor_args args;
3030 	uma_keg_t keg;
3031 	uma_zone_t res;
3032 
3033 	keg = master->uz_keg;
3034 	memset(&args, 0, sizeof(args));
3035 	args.name = name;
3036 	args.size = keg->uk_size;
3037 	args.ctor = ctor;
3038 	args.dtor = dtor;
3039 	args.uminit = zinit;
3040 	args.fini = zfini;
3041 	args.align = keg->uk_align;
3042 	args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
3043 	args.keg = keg;
3044 
3045 	sx_slock(&uma_reclaim_lock);
3046 	res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3047 	sx_sunlock(&uma_reclaim_lock);
3048 
3049 	return (res);
3050 }
3051 
3052 /* See uma.h */
3053 uma_zone_t
3054 uma_zcache_create(const char *name, int size, uma_ctor ctor, uma_dtor dtor,
3055     uma_init zinit, uma_fini zfini, uma_import zimport, uma_release zrelease,
3056     void *arg, int flags)
3057 {
3058 	struct uma_zctor_args args;
3059 
3060 	memset(&args, 0, sizeof(args));
3061 	args.name = name;
3062 	args.size = size;
3063 	args.ctor = ctor;
3064 	args.dtor = dtor;
3065 	args.uminit = zinit;
3066 	args.fini = zfini;
3067 	args.import = zimport;
3068 	args.release = zrelease;
3069 	args.arg = arg;
3070 	args.align = 0;
3071 	args.flags = flags | UMA_ZFLAG_CACHE;
3072 
3073 	return (zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK));
3074 }
3075 
3076 /* See uma.h */
3077 void
3078 uma_zdestroy(uma_zone_t zone)
3079 {
3080 
3081 	/*
3082 	 * Large slabs are expensive to reclaim, so don't bother doing
3083 	 * unnecessary work if we're shutting down.
3084 	 */
3085 	if (booted == BOOT_SHUTDOWN &&
3086 	    zone->uz_fini == NULL && zone->uz_release == zone_release)
3087 		return;
3088 	sx_slock(&uma_reclaim_lock);
3089 	zone_free_item(zones, zone, NULL, SKIP_NONE);
3090 	sx_sunlock(&uma_reclaim_lock);
3091 }
3092 
3093 void
3094 uma_zwait(uma_zone_t zone)
3095 {
3096 
3097 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
3098 		uma_zfree_smr(zone, uma_zalloc_smr(zone, M_WAITOK));
3099 	else if ((zone->uz_flags & UMA_ZONE_PCPU) != 0)
3100 		uma_zfree_pcpu(zone, uma_zalloc_pcpu(zone, M_WAITOK));
3101 	else
3102 		uma_zfree(zone, uma_zalloc(zone, M_WAITOK));
3103 }
3104 
3105 void *
3106 uma_zalloc_pcpu_arg(uma_zone_t zone, void *udata, int flags)
3107 {
3108 	void *item, *pcpu_item;
3109 #ifdef SMP
3110 	int i;
3111 
3112 	MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3113 #endif
3114 	item = uma_zalloc_arg(zone, udata, flags & ~M_ZERO);
3115 	if (item == NULL)
3116 		return (NULL);
3117 	pcpu_item = zpcpu_base_to_offset(item);
3118 	if (flags & M_ZERO) {
3119 #ifdef SMP
3120 		for (i = 0; i <= mp_maxid; i++)
3121 			bzero(zpcpu_get_cpu(pcpu_item, i), zone->uz_size);
3122 #else
3123 		bzero(item, zone->uz_size);
3124 #endif
3125 	}
3126 	return (pcpu_item);
3127 }
3128 
3129 /*
3130  * A stub while both regular and pcpu cases are identical.
3131  */
3132 void
3133 uma_zfree_pcpu_arg(uma_zone_t zone, void *pcpu_item, void *udata)
3134 {
3135 	void *item;
3136 
3137 #ifdef SMP
3138 	MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3139 #endif
3140 	item = zpcpu_offset_to_base(pcpu_item);
3141 	uma_zfree_arg(zone, item, udata);
3142 }
3143 
3144 static inline void *
3145 item_ctor(uma_zone_t zone, int uz_flags, int size, void *udata, int flags,
3146     void *item)
3147 {
3148 #ifdef INVARIANTS
3149 	bool skipdbg;
3150 
3151 	skipdbg = uma_dbg_zskip(zone, item);
3152 	if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3153 	    zone->uz_ctor != trash_ctor)
3154 		trash_ctor(item, size, udata, flags);
3155 #endif
3156 	/* Check flags before loading ctor pointer. */
3157 	if (__predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0) &&
3158 	    __predict_false(zone->uz_ctor != NULL) &&
3159 	    zone->uz_ctor(item, size, udata, flags) != 0) {
3160 		counter_u64_add(zone->uz_fails, 1);
3161 		zone_free_item(zone, item, udata, SKIP_DTOR | SKIP_CNT);
3162 		return (NULL);
3163 	}
3164 #ifdef INVARIANTS
3165 	if (!skipdbg)
3166 		uma_dbg_alloc(zone, NULL, item);
3167 #endif
3168 	if (__predict_false(flags & M_ZERO))
3169 		return (memset(item, 0, size));
3170 
3171 	return (item);
3172 }
3173 
3174 static inline void
3175 item_dtor(uma_zone_t zone, void *item, int size, void *udata,
3176     enum zfreeskip skip)
3177 {
3178 #ifdef INVARIANTS
3179 	bool skipdbg;
3180 
3181 	skipdbg = uma_dbg_zskip(zone, item);
3182 	if (skip == SKIP_NONE && !skipdbg) {
3183 		if ((zone->uz_flags & UMA_ZONE_MALLOC) != 0)
3184 			uma_dbg_free(zone, udata, item);
3185 		else
3186 			uma_dbg_free(zone, NULL, item);
3187 	}
3188 #endif
3189 	if (__predict_true(skip < SKIP_DTOR)) {
3190 		if (zone->uz_dtor != NULL)
3191 			zone->uz_dtor(item, size, udata);
3192 #ifdef INVARIANTS
3193 		if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3194 		    zone->uz_dtor != trash_dtor)
3195 			trash_dtor(item, size, udata);
3196 #endif
3197 	}
3198 }
3199 
3200 #ifdef NUMA
3201 static int
3202 item_domain(void *item)
3203 {
3204 	int domain;
3205 
3206 	domain = _vm_phys_domain(vtophys(item));
3207 	KASSERT(domain >= 0 && domain < vm_ndomains,
3208 	    ("%s: unknown domain for item %p", __func__, item));
3209 	return (domain);
3210 }
3211 #endif
3212 
3213 #if defined(INVARIANTS) || defined(DEBUG_MEMGUARD) || defined(WITNESS)
3214 #define	UMA_ZALLOC_DEBUG
3215 static int
3216 uma_zalloc_debug(uma_zone_t zone, void **itemp, void *udata, int flags)
3217 {
3218 	int error;
3219 
3220 	error = 0;
3221 #ifdef WITNESS
3222 	if (flags & M_WAITOK) {
3223 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
3224 		    "uma_zalloc_debug: zone \"%s\"", zone->uz_name);
3225 	}
3226 #endif
3227 
3228 #ifdef INVARIANTS
3229 	KASSERT((flags & M_EXEC) == 0,
3230 	    ("uma_zalloc_debug: called with M_EXEC"));
3231 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3232 	    ("uma_zalloc_debug: called within spinlock or critical section"));
3233 	KASSERT((zone->uz_flags & UMA_ZONE_PCPU) == 0 || (flags & M_ZERO) == 0,
3234 	    ("uma_zalloc_debug: allocating from a pcpu zone with M_ZERO"));
3235 #endif
3236 
3237 #ifdef DEBUG_MEMGUARD
3238 	if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && memguard_cmp_zone(zone)) {
3239 		void *item;
3240 		item = memguard_alloc(zone->uz_size, flags);
3241 		if (item != NULL) {
3242 			error = EJUSTRETURN;
3243 			if (zone->uz_init != NULL &&
3244 			    zone->uz_init(item, zone->uz_size, flags) != 0) {
3245 				*itemp = NULL;
3246 				return (error);
3247 			}
3248 			if (zone->uz_ctor != NULL &&
3249 			    zone->uz_ctor(item, zone->uz_size, udata,
3250 			    flags) != 0) {
3251 				counter_u64_add(zone->uz_fails, 1);
3252 			    	zone->uz_fini(item, zone->uz_size);
3253 				*itemp = NULL;
3254 				return (error);
3255 			}
3256 			*itemp = item;
3257 			return (error);
3258 		}
3259 		/* This is unfortunate but should not be fatal. */
3260 	}
3261 #endif
3262 	return (error);
3263 }
3264 
3265 static int
3266 uma_zfree_debug(uma_zone_t zone, void *item, void *udata)
3267 {
3268 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3269 	    ("uma_zfree_debug: called with spinlock or critical section held"));
3270 
3271 #ifdef DEBUG_MEMGUARD
3272 	if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && is_memguard_addr(item)) {
3273 		if (zone->uz_dtor != NULL)
3274 			zone->uz_dtor(item, zone->uz_size, udata);
3275 		if (zone->uz_fini != NULL)
3276 			zone->uz_fini(item, zone->uz_size);
3277 		memguard_free(item);
3278 		return (EJUSTRETURN);
3279 	}
3280 #endif
3281 	return (0);
3282 }
3283 #endif
3284 
3285 static inline void *
3286 cache_alloc_item(uma_zone_t zone, uma_cache_t cache, uma_cache_bucket_t bucket,
3287     void *udata, int flags)
3288 {
3289 	void *item;
3290 	int size, uz_flags;
3291 
3292 	item = cache_bucket_pop(cache, bucket);
3293 	size = cache_uz_size(cache);
3294 	uz_flags = cache_uz_flags(cache);
3295 	critical_exit();
3296 	return (item_ctor(zone, uz_flags, size, udata, flags, item));
3297 }
3298 
3299 static __noinline void *
3300 cache_alloc_retry(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3301 {
3302 	uma_cache_bucket_t bucket;
3303 	int domain;
3304 
3305 	while (cache_alloc(zone, cache, udata, flags)) {
3306 		cache = &zone->uz_cpu[curcpu];
3307 		bucket = &cache->uc_allocbucket;
3308 		if (__predict_false(bucket->ucb_cnt == 0))
3309 			continue;
3310 		return (cache_alloc_item(zone, cache, bucket, udata, flags));
3311 	}
3312 	critical_exit();
3313 
3314 	/*
3315 	 * We can not get a bucket so try to return a single item.
3316 	 */
3317 	if (zone->uz_flags & UMA_ZONE_FIRSTTOUCH)
3318 		domain = PCPU_GET(domain);
3319 	else
3320 		domain = UMA_ANYDOMAIN;
3321 	return (zone_alloc_item(zone, udata, domain, flags));
3322 }
3323 
3324 /* See uma.h */
3325 void *
3326 uma_zalloc_smr(uma_zone_t zone, int flags)
3327 {
3328 	uma_cache_bucket_t bucket;
3329 	uma_cache_t cache;
3330 
3331 #ifdef UMA_ZALLOC_DEBUG
3332 	void *item;
3333 
3334 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
3335 	    ("uma_zalloc_arg: called with non-SMR zone.\n"));
3336 	if (uma_zalloc_debug(zone, &item, NULL, flags) == EJUSTRETURN)
3337 		return (item);
3338 #endif
3339 
3340 	critical_enter();
3341 	cache = &zone->uz_cpu[curcpu];
3342 	bucket = &cache->uc_allocbucket;
3343 	if (__predict_false(bucket->ucb_cnt == 0))
3344 		return (cache_alloc_retry(zone, cache, NULL, flags));
3345 	return (cache_alloc_item(zone, cache, bucket, NULL, flags));
3346 }
3347 
3348 /* See uma.h */
3349 void *
3350 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
3351 {
3352 	uma_cache_bucket_t bucket;
3353 	uma_cache_t cache;
3354 
3355 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3356 	random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3357 
3358 	/* This is the fast path allocation */
3359 	CTR3(KTR_UMA, "uma_zalloc_arg zone %s(%p) flags %d", zone->uz_name,
3360 	    zone, flags);
3361 
3362 #ifdef UMA_ZALLOC_DEBUG
3363 	void *item;
3364 
3365 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3366 	    ("uma_zalloc_arg: called with SMR zone.\n"));
3367 	if (uma_zalloc_debug(zone, &item, udata, flags) == EJUSTRETURN)
3368 		return (item);
3369 #endif
3370 
3371 	/*
3372 	 * If possible, allocate from the per-CPU cache.  There are two
3373 	 * requirements for safe access to the per-CPU cache: (1) the thread
3374 	 * accessing the cache must not be preempted or yield during access,
3375 	 * and (2) the thread must not migrate CPUs without switching which
3376 	 * cache it accesses.  We rely on a critical section to prevent
3377 	 * preemption and migration.  We release the critical section in
3378 	 * order to acquire the zone mutex if we are unable to allocate from
3379 	 * the current cache; when we re-acquire the critical section, we
3380 	 * must detect and handle migration if it has occurred.
3381 	 */
3382 	critical_enter();
3383 	cache = &zone->uz_cpu[curcpu];
3384 	bucket = &cache->uc_allocbucket;
3385 	if (__predict_false(bucket->ucb_cnt == 0))
3386 		return (cache_alloc_retry(zone, cache, udata, flags));
3387 	return (cache_alloc_item(zone, cache, bucket, udata, flags));
3388 }
3389 
3390 /*
3391  * Replenish an alloc bucket and possibly restore an old one.  Called in
3392  * a critical section.  Returns in a critical section.
3393  *
3394  * A false return value indicates an allocation failure.
3395  * A true return value indicates success and the caller should retry.
3396  */
3397 static __noinline bool
3398 cache_alloc(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3399 {
3400 	uma_bucket_t bucket;
3401 	int domain;
3402 	bool new;
3403 
3404 	CRITICAL_ASSERT(curthread);
3405 
3406 	/*
3407 	 * If we have run out of items in our alloc bucket see
3408 	 * if we can switch with the free bucket.
3409 	 *
3410 	 * SMR Zones can't re-use the free bucket until the sequence has
3411 	 * expired.
3412 	 */
3413 	if ((cache_uz_flags(cache) & UMA_ZONE_SMR) == 0 &&
3414 	    cache->uc_freebucket.ucb_cnt != 0) {
3415 		cache_bucket_swap(&cache->uc_freebucket,
3416 		    &cache->uc_allocbucket);
3417 		return (true);
3418 	}
3419 
3420 	/*
3421 	 * Discard any empty allocation bucket while we hold no locks.
3422 	 */
3423 	bucket = cache_bucket_unload_alloc(cache);
3424 	critical_exit();
3425 
3426 	if (bucket != NULL) {
3427 		KASSERT(bucket->ub_cnt == 0,
3428 		    ("cache_alloc: Entered with non-empty alloc bucket."));
3429 		bucket_free(zone, bucket, udata);
3430 	}
3431 
3432 	/* Short-circuit for zones without buckets and low memory. */
3433 	if (zone->uz_bucket_size == 0 || bucketdisable) {
3434 		critical_enter();
3435 		return (false);
3436 	}
3437 
3438 	/*
3439 	 * Attempt to retrieve the item from the per-CPU cache has failed, so
3440 	 * we must go back to the zone.  This requires the zdom lock, so we
3441 	 * must drop the critical section, then re-acquire it when we go back
3442 	 * to the cache.  Since the critical section is released, we may be
3443 	 * preempted or migrate.  As such, make sure not to maintain any
3444 	 * thread-local state specific to the cache from prior to releasing
3445 	 * the critical section.
3446 	 */
3447 	domain = PCPU_GET(domain);
3448 	if ((cache_uz_flags(cache) & UMA_ZONE_ROUNDROBIN) != 0)
3449 		domain = zone_domain_highest(zone, domain);
3450 	bucket = cache_fetch_bucket(zone, cache, domain);
3451 	if (bucket == NULL) {
3452 		bucket = zone_alloc_bucket(zone, udata, domain, flags);
3453 		new = true;
3454 	} else
3455 		new = false;
3456 
3457 	CTR3(KTR_UMA, "uma_zalloc: zone %s(%p) bucket zone returned %p",
3458 	    zone->uz_name, zone, bucket);
3459 	if (bucket == NULL) {
3460 		critical_enter();
3461 		return (false);
3462 	}
3463 
3464 	/*
3465 	 * See if we lost the race or were migrated.  Cache the
3466 	 * initialized bucket to make this less likely or claim
3467 	 * the memory directly.
3468 	 */
3469 	critical_enter();
3470 	cache = &zone->uz_cpu[curcpu];
3471 	if (cache->uc_allocbucket.ucb_bucket == NULL &&
3472 	    ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) == 0 ||
3473 	    domain == PCPU_GET(domain))) {
3474 		if (new)
3475 			atomic_add_long(&ZDOM_GET(zone, domain)->uzd_imax,
3476 			    bucket->ub_cnt);
3477 		cache_bucket_load_alloc(cache, bucket);
3478 		return (true);
3479 	}
3480 
3481 	/*
3482 	 * We lost the race, release this bucket and start over.
3483 	 */
3484 	critical_exit();
3485 	zone_put_bucket(zone, domain, bucket, udata, false);
3486 	critical_enter();
3487 
3488 	return (true);
3489 }
3490 
3491 void *
3492 uma_zalloc_domain(uma_zone_t zone, void *udata, int domain, int flags)
3493 {
3494 
3495 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3496 	random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3497 
3498 	/* This is the fast path allocation */
3499 	CTR4(KTR_UMA, "uma_zalloc_domain zone %s(%p) domain %d flags %d",
3500 	    zone->uz_name, zone, domain, flags);
3501 
3502 	if (flags & M_WAITOK) {
3503 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
3504 		    "uma_zalloc_domain: zone \"%s\"", zone->uz_name);
3505 	}
3506 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3507 	    ("uma_zalloc_domain: called with spinlock or critical section held"));
3508 
3509 	return (zone_alloc_item(zone, udata, domain, flags));
3510 }
3511 
3512 /*
3513  * Find a slab with some space.  Prefer slabs that are partially used over those
3514  * that are totally full.  This helps to reduce fragmentation.
3515  *
3516  * If 'rr' is 1, search all domains starting from 'domain'.  Otherwise check
3517  * only 'domain'.
3518  */
3519 static uma_slab_t
3520 keg_first_slab(uma_keg_t keg, int domain, bool rr)
3521 {
3522 	uma_domain_t dom;
3523 	uma_slab_t slab;
3524 	int start;
3525 
3526 	KASSERT(domain >= 0 && domain < vm_ndomains,
3527 	    ("keg_first_slab: domain %d out of range", domain));
3528 	KEG_LOCK_ASSERT(keg, domain);
3529 
3530 	slab = NULL;
3531 	start = domain;
3532 	do {
3533 		dom = &keg->uk_domain[domain];
3534 		if ((slab = LIST_FIRST(&dom->ud_part_slab)) != NULL)
3535 			return (slab);
3536 		if ((slab = LIST_FIRST(&dom->ud_free_slab)) != NULL) {
3537 			LIST_REMOVE(slab, us_link);
3538 			dom->ud_free_slabs--;
3539 			LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
3540 			return (slab);
3541 		}
3542 		if (rr)
3543 			domain = (domain + 1) % vm_ndomains;
3544 	} while (domain != start);
3545 
3546 	return (NULL);
3547 }
3548 
3549 /*
3550  * Fetch an existing slab from a free or partial list.  Returns with the
3551  * keg domain lock held if a slab was found or unlocked if not.
3552  */
3553 static uma_slab_t
3554 keg_fetch_free_slab(uma_keg_t keg, int domain, bool rr, int flags)
3555 {
3556 	uma_slab_t slab;
3557 	uint32_t reserve;
3558 
3559 	/* HASH has a single free list. */
3560 	if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
3561 		domain = 0;
3562 
3563 	KEG_LOCK(keg, domain);
3564 	reserve = (flags & M_USE_RESERVE) != 0 ? 0 : keg->uk_reserve;
3565 	if (keg->uk_domain[domain].ud_free_items <= reserve ||
3566 	    (slab = keg_first_slab(keg, domain, rr)) == NULL) {
3567 		KEG_UNLOCK(keg, domain);
3568 		return (NULL);
3569 	}
3570 	return (slab);
3571 }
3572 
3573 static uma_slab_t
3574 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int rdomain, const int flags)
3575 {
3576 	struct vm_domainset_iter di;
3577 	uma_slab_t slab;
3578 	int aflags, domain;
3579 	bool rr;
3580 
3581 restart:
3582 	/*
3583 	 * Use the keg's policy if upper layers haven't already specified a
3584 	 * domain (as happens with first-touch zones).
3585 	 *
3586 	 * To avoid races we run the iterator with the keg lock held, but that
3587 	 * means that we cannot allow the vm_domainset layer to sleep.  Thus,
3588 	 * clear M_WAITOK and handle low memory conditions locally.
3589 	 */
3590 	rr = rdomain == UMA_ANYDOMAIN;
3591 	if (rr) {
3592 		aflags = (flags & ~M_WAITOK) | M_NOWAIT;
3593 		vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
3594 		    &aflags);
3595 	} else {
3596 		aflags = flags;
3597 		domain = rdomain;
3598 	}
3599 
3600 	for (;;) {
3601 		slab = keg_fetch_free_slab(keg, domain, rr, flags);
3602 		if (slab != NULL)
3603 			return (slab);
3604 
3605 		/*
3606 		 * M_NOVM means don't ask at all!
3607 		 */
3608 		if (flags & M_NOVM)
3609 			break;
3610 
3611 		slab = keg_alloc_slab(keg, zone, domain, flags, aflags);
3612 		if (slab != NULL)
3613 			return (slab);
3614 		if (!rr && (flags & M_WAITOK) == 0)
3615 			break;
3616 		if (rr && vm_domainset_iter_policy(&di, &domain) != 0) {
3617 			if ((flags & M_WAITOK) != 0) {
3618 				vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask);
3619 				goto restart;
3620 			}
3621 			break;
3622 		}
3623 	}
3624 
3625 	/*
3626 	 * We might not have been able to get a slab but another cpu
3627 	 * could have while we were unlocked.  Check again before we
3628 	 * fail.
3629 	 */
3630 	if ((slab = keg_fetch_free_slab(keg, domain, rr, flags)) != NULL)
3631 		return (slab);
3632 
3633 	return (NULL);
3634 }
3635 
3636 static void *
3637 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
3638 {
3639 	uma_domain_t dom;
3640 	void *item;
3641 	int freei;
3642 
3643 	KEG_LOCK_ASSERT(keg, slab->us_domain);
3644 
3645 	dom = &keg->uk_domain[slab->us_domain];
3646 	freei = BIT_FFS(keg->uk_ipers, &slab->us_free) - 1;
3647 	BIT_CLR(keg->uk_ipers, freei, &slab->us_free);
3648 	item = slab_item(slab, keg, freei);
3649 	slab->us_freecount--;
3650 	dom->ud_free_items--;
3651 
3652 	/*
3653 	 * Move this slab to the full list.  It must be on the partial list, so
3654 	 * we do not need to update the free slab count.  In particular,
3655 	 * keg_fetch_slab() always returns slabs on the partial list.
3656 	 */
3657 	if (slab->us_freecount == 0) {
3658 		LIST_REMOVE(slab, us_link);
3659 		LIST_INSERT_HEAD(&dom->ud_full_slab, slab, us_link);
3660 	}
3661 
3662 	return (item);
3663 }
3664 
3665 static int
3666 zone_import(void *arg, void **bucket, int max, int domain, int flags)
3667 {
3668 	uma_domain_t dom;
3669 	uma_zone_t zone;
3670 	uma_slab_t slab;
3671 	uma_keg_t keg;
3672 #ifdef NUMA
3673 	int stripe;
3674 #endif
3675 	int i;
3676 
3677 	zone = arg;
3678 	slab = NULL;
3679 	keg = zone->uz_keg;
3680 	/* Try to keep the buckets totally full */
3681 	for (i = 0; i < max; ) {
3682 		if ((slab = keg_fetch_slab(keg, zone, domain, flags)) == NULL)
3683 			break;
3684 #ifdef NUMA
3685 		stripe = howmany(max, vm_ndomains);
3686 #endif
3687 		dom = &keg->uk_domain[slab->us_domain];
3688 		while (slab->us_freecount && i < max) {
3689 			bucket[i++] = slab_alloc_item(keg, slab);
3690 			if (dom->ud_free_items <= keg->uk_reserve)
3691 				break;
3692 #ifdef NUMA
3693 			/*
3694 			 * If the zone is striped we pick a new slab for every
3695 			 * N allocations.  Eliminating this conditional will
3696 			 * instead pick a new domain for each bucket rather
3697 			 * than stripe within each bucket.  The current option
3698 			 * produces more fragmentation and requires more cpu
3699 			 * time but yields better distribution.
3700 			 */
3701 			if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0 &&
3702 			    vm_ndomains > 1 && --stripe == 0)
3703 				break;
3704 #endif
3705 		}
3706 		KEG_UNLOCK(keg, slab->us_domain);
3707 		/* Don't block if we allocated any successfully. */
3708 		flags &= ~M_WAITOK;
3709 		flags |= M_NOWAIT;
3710 	}
3711 
3712 	return i;
3713 }
3714 
3715 static int
3716 zone_alloc_limit_hard(uma_zone_t zone, int count, int flags)
3717 {
3718 	uint64_t old, new, total, max;
3719 
3720 	/*
3721 	 * The hard case.  We're going to sleep because there were existing
3722 	 * sleepers or because we ran out of items.  This routine enforces
3723 	 * fairness by keeping fifo order.
3724 	 *
3725 	 * First release our ill gotten gains and make some noise.
3726 	 */
3727 	for (;;) {
3728 		zone_free_limit(zone, count);
3729 		zone_log_warning(zone);
3730 		zone_maxaction(zone);
3731 		if (flags & M_NOWAIT)
3732 			return (0);
3733 
3734 		/*
3735 		 * We need to allocate an item or set ourself as a sleeper
3736 		 * while the sleepq lock is held to avoid wakeup races.  This
3737 		 * is essentially a home rolled semaphore.
3738 		 */
3739 		sleepq_lock(&zone->uz_max_items);
3740 		old = zone->uz_items;
3741 		do {
3742 			MPASS(UZ_ITEMS_SLEEPERS(old) < UZ_ITEMS_SLEEPERS_MAX);
3743 			/* Cache the max since we will evaluate twice. */
3744 			max = zone->uz_max_items;
3745 			if (UZ_ITEMS_SLEEPERS(old) != 0 ||
3746 			    UZ_ITEMS_COUNT(old) >= max)
3747 				new = old + UZ_ITEMS_SLEEPER;
3748 			else
3749 				new = old + MIN(count, max - old);
3750 		} while (atomic_fcmpset_64(&zone->uz_items, &old, new) == 0);
3751 
3752 		/* We may have successfully allocated under the sleepq lock. */
3753 		if (UZ_ITEMS_SLEEPERS(new) == 0) {
3754 			sleepq_release(&zone->uz_max_items);
3755 			return (new - old);
3756 		}
3757 
3758 		/*
3759 		 * This is in a different cacheline from uz_items so that we
3760 		 * don't constantly invalidate the fastpath cacheline when we
3761 		 * adjust item counts.  This could be limited to toggling on
3762 		 * transitions.
3763 		 */
3764 		atomic_add_32(&zone->uz_sleepers, 1);
3765 		atomic_add_64(&zone->uz_sleeps, 1);
3766 
3767 		/*
3768 		 * We have added ourselves as a sleeper.  The sleepq lock
3769 		 * protects us from wakeup races.  Sleep now and then retry.
3770 		 */
3771 		sleepq_add(&zone->uz_max_items, NULL, "zonelimit", 0, 0);
3772 		sleepq_wait(&zone->uz_max_items, PVM);
3773 
3774 		/*
3775 		 * After wakeup, remove ourselves as a sleeper and try
3776 		 * again.  We no longer have the sleepq lock for protection.
3777 		 *
3778 		 * Subract ourselves as a sleeper while attempting to add
3779 		 * our count.
3780 		 */
3781 		atomic_subtract_32(&zone->uz_sleepers, 1);
3782 		old = atomic_fetchadd_64(&zone->uz_items,
3783 		    -(UZ_ITEMS_SLEEPER - count));
3784 		/* We're no longer a sleeper. */
3785 		old -= UZ_ITEMS_SLEEPER;
3786 
3787 		/*
3788 		 * If we're still at the limit, restart.  Notably do not
3789 		 * block on other sleepers.  Cache the max value to protect
3790 		 * against changes via sysctl.
3791 		 */
3792 		total = UZ_ITEMS_COUNT(old);
3793 		max = zone->uz_max_items;
3794 		if (total >= max)
3795 			continue;
3796 		/* Truncate if necessary, otherwise wake other sleepers. */
3797 		if (total + count > max) {
3798 			zone_free_limit(zone, total + count - max);
3799 			count = max - total;
3800 		} else if (total + count < max && UZ_ITEMS_SLEEPERS(old) != 0)
3801 			wakeup_one(&zone->uz_max_items);
3802 
3803 		return (count);
3804 	}
3805 }
3806 
3807 /*
3808  * Allocate 'count' items from our max_items limit.  Returns the number
3809  * available.  If M_NOWAIT is not specified it will sleep until at least
3810  * one item can be allocated.
3811  */
3812 static int
3813 zone_alloc_limit(uma_zone_t zone, int count, int flags)
3814 {
3815 	uint64_t old;
3816 	uint64_t max;
3817 
3818 	max = zone->uz_max_items;
3819 	MPASS(max > 0);
3820 
3821 	/*
3822 	 * We expect normal allocations to succeed with a simple
3823 	 * fetchadd.
3824 	 */
3825 	old = atomic_fetchadd_64(&zone->uz_items, count);
3826 	if (__predict_true(old + count <= max))
3827 		return (count);
3828 
3829 	/*
3830 	 * If we had some items and no sleepers just return the
3831 	 * truncated value.  We have to release the excess space
3832 	 * though because that may wake sleepers who weren't woken
3833 	 * because we were temporarily over the limit.
3834 	 */
3835 	if (old < max) {
3836 		zone_free_limit(zone, (old + count) - max);
3837 		return (max - old);
3838 	}
3839 	return (zone_alloc_limit_hard(zone, count, flags));
3840 }
3841 
3842 /*
3843  * Free a number of items back to the limit.
3844  */
3845 static void
3846 zone_free_limit(uma_zone_t zone, int count)
3847 {
3848 	uint64_t old;
3849 
3850 	MPASS(count > 0);
3851 
3852 	/*
3853 	 * In the common case we either have no sleepers or
3854 	 * are still over the limit and can just return.
3855 	 */
3856 	old = atomic_fetchadd_64(&zone->uz_items, -count);
3857 	if (__predict_true(UZ_ITEMS_SLEEPERS(old) == 0 ||
3858 	   UZ_ITEMS_COUNT(old) - count >= zone->uz_max_items))
3859 		return;
3860 
3861 	/*
3862 	 * Moderate the rate of wakeups.  Sleepers will continue
3863 	 * to generate wakeups if necessary.
3864 	 */
3865 	wakeup_one(&zone->uz_max_items);
3866 }
3867 
3868 static uma_bucket_t
3869 zone_alloc_bucket(uma_zone_t zone, void *udata, int domain, int flags)
3870 {
3871 	uma_bucket_t bucket;
3872 	int maxbucket, cnt;
3873 
3874 	CTR3(KTR_UMA, "zone_alloc_bucket zone %s(%p) domain %d", zone->uz_name,
3875 	    zone, domain);
3876 
3877 	/* Avoid allocs targeting empty domains. */
3878 	if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
3879 		domain = UMA_ANYDOMAIN;
3880 	if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
3881 		domain = UMA_ANYDOMAIN;
3882 
3883 	if (zone->uz_max_items > 0)
3884 		maxbucket = zone_alloc_limit(zone, zone->uz_bucket_size,
3885 		    M_NOWAIT);
3886 	else
3887 		maxbucket = zone->uz_bucket_size;
3888 	if (maxbucket == 0)
3889 		return (false);
3890 
3891 	/* Don't wait for buckets, preserve caller's NOVM setting. */
3892 	bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
3893 	if (bucket == NULL) {
3894 		cnt = 0;
3895 		goto out;
3896 	}
3897 
3898 	bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
3899 	    MIN(maxbucket, bucket->ub_entries), domain, flags);
3900 
3901 	/*
3902 	 * Initialize the memory if necessary.
3903 	 */
3904 	if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
3905 		int i;
3906 
3907 		for (i = 0; i < bucket->ub_cnt; i++)
3908 			if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size,
3909 			    flags) != 0)
3910 				break;
3911 		/*
3912 		 * If we couldn't initialize the whole bucket, put the
3913 		 * rest back onto the freelist.
3914 		 */
3915 		if (i != bucket->ub_cnt) {
3916 			zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
3917 			    bucket->ub_cnt - i);
3918 #ifdef INVARIANTS
3919 			bzero(&bucket->ub_bucket[i],
3920 			    sizeof(void *) * (bucket->ub_cnt - i));
3921 #endif
3922 			bucket->ub_cnt = i;
3923 		}
3924 	}
3925 
3926 	cnt = bucket->ub_cnt;
3927 	if (bucket->ub_cnt == 0) {
3928 		bucket_free(zone, bucket, udata);
3929 		counter_u64_add(zone->uz_fails, 1);
3930 		bucket = NULL;
3931 	}
3932 out:
3933 	if (zone->uz_max_items > 0 && cnt < maxbucket)
3934 		zone_free_limit(zone, maxbucket - cnt);
3935 
3936 	return (bucket);
3937 }
3938 
3939 /*
3940  * Allocates a single item from a zone.
3941  *
3942  * Arguments
3943  *	zone   The zone to alloc for.
3944  *	udata  The data to be passed to the constructor.
3945  *	domain The domain to allocate from or UMA_ANYDOMAIN.
3946  *	flags  M_WAITOK, M_NOWAIT, M_ZERO.
3947  *
3948  * Returns
3949  *	NULL if there is no memory and M_NOWAIT is set
3950  *	An item if successful
3951  */
3952 
3953 static void *
3954 zone_alloc_item(uma_zone_t zone, void *udata, int domain, int flags)
3955 {
3956 	void *item;
3957 
3958 	if (zone->uz_max_items > 0 && zone_alloc_limit(zone, 1, flags) == 0)
3959 		return (NULL);
3960 
3961 	/* Avoid allocs targeting empty domains. */
3962 	if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
3963 		domain = UMA_ANYDOMAIN;
3964 
3965 	if (zone->uz_import(zone->uz_arg, &item, 1, domain, flags) != 1)
3966 		goto fail_cnt;
3967 
3968 	/*
3969 	 * We have to call both the zone's init (not the keg's init)
3970 	 * and the zone's ctor.  This is because the item is going from
3971 	 * a keg slab directly to the user, and the user is expecting it
3972 	 * to be both zone-init'd as well as zone-ctor'd.
3973 	 */
3974 	if (zone->uz_init != NULL) {
3975 		if (zone->uz_init(item, zone->uz_size, flags) != 0) {
3976 			zone_free_item(zone, item, udata, SKIP_FINI | SKIP_CNT);
3977 			goto fail_cnt;
3978 		}
3979 	}
3980 	item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata, flags,
3981 	    item);
3982 	if (item == NULL)
3983 		goto fail;
3984 
3985 	counter_u64_add(zone->uz_allocs, 1);
3986 	CTR3(KTR_UMA, "zone_alloc_item item %p from %s(%p)", item,
3987 	    zone->uz_name, zone);
3988 
3989 	return (item);
3990 
3991 fail_cnt:
3992 	counter_u64_add(zone->uz_fails, 1);
3993 fail:
3994 	if (zone->uz_max_items > 0)
3995 		zone_free_limit(zone, 1);
3996 	CTR2(KTR_UMA, "zone_alloc_item failed from %s(%p)",
3997 	    zone->uz_name, zone);
3998 
3999 	return (NULL);
4000 }
4001 
4002 /* See uma.h */
4003 void
4004 uma_zfree_smr(uma_zone_t zone, void *item)
4005 {
4006 	uma_cache_t cache;
4007 	uma_cache_bucket_t bucket;
4008 	int itemdomain, uz_flags;
4009 
4010 #ifdef UMA_ZALLOC_DEBUG
4011 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
4012 	    ("uma_zfree_smr: called with non-SMR zone.\n"));
4013 	KASSERT(item != NULL, ("uma_zfree_smr: Called with NULL pointer."));
4014 	SMR_ASSERT_NOT_ENTERED(zone->uz_smr);
4015 	if (uma_zfree_debug(zone, item, NULL) == EJUSTRETURN)
4016 		return;
4017 #endif
4018 	cache = &zone->uz_cpu[curcpu];
4019 	uz_flags = cache_uz_flags(cache);
4020 	itemdomain = 0;
4021 #ifdef NUMA
4022 	if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4023 		itemdomain = item_domain(item);
4024 #endif
4025 	critical_enter();
4026 	do {
4027 		cache = &zone->uz_cpu[curcpu];
4028 		/* SMR Zones must free to the free bucket. */
4029 		bucket = &cache->uc_freebucket;
4030 #ifdef NUMA
4031 		if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4032 		    PCPU_GET(domain) != itemdomain) {
4033 			bucket = &cache->uc_crossbucket;
4034 		}
4035 #endif
4036 		if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4037 			cache_bucket_push(cache, bucket, item);
4038 			critical_exit();
4039 			return;
4040 		}
4041 	} while (cache_free(zone, cache, NULL, item, itemdomain));
4042 	critical_exit();
4043 
4044 	/*
4045 	 * If nothing else caught this, we'll just do an internal free.
4046 	 */
4047 	zone_free_item(zone, item, NULL, SKIP_NONE);
4048 }
4049 
4050 /* See uma.h */
4051 void
4052 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
4053 {
4054 	uma_cache_t cache;
4055 	uma_cache_bucket_t bucket;
4056 	int itemdomain, uz_flags;
4057 
4058 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
4059 	random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
4060 
4061 	CTR2(KTR_UMA, "uma_zfree_arg zone %s(%p)", zone->uz_name, zone);
4062 
4063 #ifdef UMA_ZALLOC_DEBUG
4064 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
4065 	    ("uma_zfree_arg: called with SMR zone.\n"));
4066 	if (uma_zfree_debug(zone, item, udata) == EJUSTRETURN)
4067 		return;
4068 #endif
4069         /* uma_zfree(..., NULL) does nothing, to match free(9). */
4070         if (item == NULL)
4071                 return;
4072 
4073 	/*
4074 	 * We are accessing the per-cpu cache without a critical section to
4075 	 * fetch size and flags.  This is acceptable, if we are preempted we
4076 	 * will simply read another cpu's line.
4077 	 */
4078 	cache = &zone->uz_cpu[curcpu];
4079 	uz_flags = cache_uz_flags(cache);
4080 	if (UMA_ALWAYS_CTORDTOR ||
4081 	    __predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0))
4082 		item_dtor(zone, item, cache_uz_size(cache), udata, SKIP_NONE);
4083 
4084 	/*
4085 	 * The race here is acceptable.  If we miss it we'll just have to wait
4086 	 * a little longer for the limits to be reset.
4087 	 */
4088 	if (__predict_false(uz_flags & UMA_ZFLAG_LIMIT)) {
4089 		if (zone->uz_sleepers > 0)
4090 			goto zfree_item;
4091 	}
4092 
4093 	/*
4094 	 * If possible, free to the per-CPU cache.  There are two
4095 	 * requirements for safe access to the per-CPU cache: (1) the thread
4096 	 * accessing the cache must not be preempted or yield during access,
4097 	 * and (2) the thread must not migrate CPUs without switching which
4098 	 * cache it accesses.  We rely on a critical section to prevent
4099 	 * preemption and migration.  We release the critical section in
4100 	 * order to acquire the zone mutex if we are unable to free to the
4101 	 * current cache; when we re-acquire the critical section, we must
4102 	 * detect and handle migration if it has occurred.
4103 	 */
4104 	itemdomain = 0;
4105 #ifdef NUMA
4106 	if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4107 		itemdomain = item_domain(item);
4108 #endif
4109 	critical_enter();
4110 	do {
4111 		cache = &zone->uz_cpu[curcpu];
4112 		/*
4113 		 * Try to free into the allocbucket first to give LIFO
4114 		 * ordering for cache-hot datastructures.  Spill over
4115 		 * into the freebucket if necessary.  Alloc will swap
4116 		 * them if one runs dry.
4117 		 */
4118 		bucket = &cache->uc_allocbucket;
4119 #ifdef NUMA
4120 		if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4121 		    PCPU_GET(domain) != itemdomain) {
4122 			bucket = &cache->uc_crossbucket;
4123 		} else
4124 #endif
4125 		if (bucket->ucb_cnt == bucket->ucb_entries &&
4126 		   cache->uc_freebucket.ucb_cnt <
4127 		   cache->uc_freebucket.ucb_entries)
4128 			cache_bucket_swap(&cache->uc_freebucket,
4129 			    &cache->uc_allocbucket);
4130 		if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4131 			cache_bucket_push(cache, bucket, item);
4132 			critical_exit();
4133 			return;
4134 		}
4135 	} while (cache_free(zone, cache, udata, item, itemdomain));
4136 	critical_exit();
4137 
4138 	/*
4139 	 * If nothing else caught this, we'll just do an internal free.
4140 	 */
4141 zfree_item:
4142 	zone_free_item(zone, item, udata, SKIP_DTOR);
4143 }
4144 
4145 #ifdef NUMA
4146 /*
4147  * sort crossdomain free buckets to domain correct buckets and cache
4148  * them.
4149  */
4150 static void
4151 zone_free_cross(uma_zone_t zone, uma_bucket_t bucket, void *udata)
4152 {
4153 	struct uma_bucketlist fullbuckets;
4154 	uma_zone_domain_t zdom;
4155 	uma_bucket_t b;
4156 	smr_seq_t seq;
4157 	void *item;
4158 	int domain;
4159 
4160 	CTR3(KTR_UMA,
4161 	    "uma_zfree: zone %s(%p) draining cross bucket %p",
4162 	    zone->uz_name, zone, bucket);
4163 
4164 	/*
4165 	 * It is possible for buckets to arrive here out of order so we fetch
4166 	 * the current smr seq rather than accepting the bucket's.
4167 	 */
4168 	seq = SMR_SEQ_INVALID;
4169 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
4170 		seq = smr_advance(zone->uz_smr);
4171 
4172 	/*
4173 	 * To avoid having ndomain * ndomain buckets for sorting we have a
4174 	 * lock on the current crossfree bucket.  A full matrix with
4175 	 * per-domain locking could be used if necessary.
4176 	 */
4177 	STAILQ_INIT(&fullbuckets);
4178 	ZONE_CROSS_LOCK(zone);
4179 	while (bucket->ub_cnt > 0) {
4180 		item = bucket->ub_bucket[bucket->ub_cnt - 1];
4181 		domain = item_domain(item);
4182 		zdom = ZDOM_GET(zone, domain);
4183 		if (zdom->uzd_cross == NULL) {
4184 			zdom->uzd_cross = bucket_alloc(zone, udata, M_NOWAIT);
4185 			if (zdom->uzd_cross == NULL)
4186 				break;
4187 		}
4188 		b = zdom->uzd_cross;
4189 		b->ub_bucket[b->ub_cnt++] = item;
4190 		b->ub_seq = seq;
4191 		if (b->ub_cnt == b->ub_entries) {
4192 			STAILQ_INSERT_HEAD(&fullbuckets, b, ub_link);
4193 			zdom->uzd_cross = NULL;
4194 		}
4195 		bucket->ub_cnt--;
4196 	}
4197 	ZONE_CROSS_UNLOCK(zone);
4198 	if (bucket->ub_cnt == 0)
4199 		bucket->ub_seq = SMR_SEQ_INVALID;
4200 	bucket_free(zone, bucket, udata);
4201 
4202 	while ((b = STAILQ_FIRST(&fullbuckets)) != NULL) {
4203 		STAILQ_REMOVE_HEAD(&fullbuckets, ub_link);
4204 		domain = item_domain(b->ub_bucket[0]);
4205 		zone_put_bucket(zone, domain, b, udata, true);
4206 	}
4207 }
4208 #endif
4209 
4210 static void
4211 zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata,
4212     int itemdomain, bool ws)
4213 {
4214 
4215 #ifdef NUMA
4216 	/*
4217 	 * Buckets coming from the wrong domain will be entirely for the
4218 	 * only other domain on two domain systems.  In this case we can
4219 	 * simply cache them.  Otherwise we need to sort them back to
4220 	 * correct domains.
4221 	 */
4222 	if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4223 	    vm_ndomains > 2 && PCPU_GET(domain) != itemdomain) {
4224 		zone_free_cross(zone, bucket, udata);
4225 		return;
4226 	}
4227 #endif
4228 
4229 	/*
4230 	 * Attempt to save the bucket in the zone's domain bucket cache.
4231 	 */
4232 	CTR3(KTR_UMA,
4233 	    "uma_zfree: zone %s(%p) putting bucket %p on free list",
4234 	    zone->uz_name, zone, bucket);
4235 	/* ub_cnt is pointing to the last free item */
4236 	if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
4237 		itemdomain = zone_domain_lowest(zone, itemdomain);
4238 	zone_put_bucket(zone, itemdomain, bucket, udata, ws);
4239 }
4240 
4241 /*
4242  * Populate a free or cross bucket for the current cpu cache.  Free any
4243  * existing full bucket either to the zone cache or back to the slab layer.
4244  *
4245  * Enters and returns in a critical section.  false return indicates that
4246  * we can not satisfy this free in the cache layer.  true indicates that
4247  * the caller should retry.
4248  */
4249 static __noinline bool
4250 cache_free(uma_zone_t zone, uma_cache_t cache, void *udata, void *item,
4251     int itemdomain)
4252 {
4253 	uma_cache_bucket_t cbucket;
4254 	uma_bucket_t newbucket, bucket;
4255 
4256 	CRITICAL_ASSERT(curthread);
4257 
4258 	if (zone->uz_bucket_size == 0)
4259 		return false;
4260 
4261 	cache = &zone->uz_cpu[curcpu];
4262 	newbucket = NULL;
4263 
4264 	/*
4265 	 * FIRSTTOUCH domains need to free to the correct zdom.  When
4266 	 * enabled this is the zdom of the item.   The bucket is the
4267 	 * cross bucket if the current domain and itemdomain do not match.
4268 	 */
4269 	cbucket = &cache->uc_freebucket;
4270 #ifdef NUMA
4271 	if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4272 		if (PCPU_GET(domain) != itemdomain) {
4273 			cbucket = &cache->uc_crossbucket;
4274 			if (cbucket->ucb_cnt != 0)
4275 				counter_u64_add(zone->uz_xdomain,
4276 				    cbucket->ucb_cnt);
4277 		}
4278 	}
4279 #endif
4280 	bucket = cache_bucket_unload(cbucket);
4281 	KASSERT(bucket == NULL || bucket->ub_cnt == bucket->ub_entries,
4282 	    ("cache_free: Entered with non-full free bucket."));
4283 
4284 	/* We are no longer associated with this CPU. */
4285 	critical_exit();
4286 
4287 	/*
4288 	 * Don't let SMR zones operate without a free bucket.  Force
4289 	 * a synchronize and re-use this one.  We will only degrade
4290 	 * to a synchronize every bucket_size items rather than every
4291 	 * item if we fail to allocate a bucket.
4292 	 */
4293 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0) {
4294 		if (bucket != NULL)
4295 			bucket->ub_seq = smr_advance(zone->uz_smr);
4296 		newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4297 		if (newbucket == NULL && bucket != NULL) {
4298 			bucket_drain(zone, bucket);
4299 			newbucket = bucket;
4300 			bucket = NULL;
4301 		}
4302 	} else if (!bucketdisable)
4303 		newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4304 
4305 	if (bucket != NULL)
4306 		zone_free_bucket(zone, bucket, udata, itemdomain, true);
4307 
4308 	critical_enter();
4309 	if ((bucket = newbucket) == NULL)
4310 		return (false);
4311 	cache = &zone->uz_cpu[curcpu];
4312 #ifdef NUMA
4313 	/*
4314 	 * Check to see if we should be populating the cross bucket.  If it
4315 	 * is already populated we will fall through and attempt to populate
4316 	 * the free bucket.
4317 	 */
4318 	if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4319 		if (PCPU_GET(domain) != itemdomain &&
4320 		    cache->uc_crossbucket.ucb_bucket == NULL) {
4321 			cache_bucket_load_cross(cache, bucket);
4322 			return (true);
4323 		}
4324 	}
4325 #endif
4326 	/*
4327 	 * We may have lost the race to fill the bucket or switched CPUs.
4328 	 */
4329 	if (cache->uc_freebucket.ucb_bucket != NULL) {
4330 		critical_exit();
4331 		bucket_free(zone, bucket, udata);
4332 		critical_enter();
4333 	} else
4334 		cache_bucket_load_free(cache, bucket);
4335 
4336 	return (true);
4337 }
4338 
4339 void
4340 uma_zfree_domain(uma_zone_t zone, void *item, void *udata)
4341 {
4342 
4343 	/* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
4344 	random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
4345 
4346 	CTR2(KTR_UMA, "uma_zfree_domain zone %s(%p)", zone->uz_name, zone);
4347 
4348 	KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
4349 	    ("uma_zfree_domain: called with spinlock or critical section held"));
4350 
4351         /* uma_zfree(..., NULL) does nothing, to match free(9). */
4352         if (item == NULL)
4353                 return;
4354 	zone_free_item(zone, item, udata, SKIP_NONE);
4355 }
4356 
4357 static void
4358 slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item)
4359 {
4360 	uma_keg_t keg;
4361 	uma_domain_t dom;
4362 	int freei;
4363 
4364 	keg = zone->uz_keg;
4365 	KEG_LOCK_ASSERT(keg, slab->us_domain);
4366 
4367 	/* Do we need to remove from any lists? */
4368 	dom = &keg->uk_domain[slab->us_domain];
4369 	if (slab->us_freecount + 1 == keg->uk_ipers) {
4370 		LIST_REMOVE(slab, us_link);
4371 		LIST_INSERT_HEAD(&dom->ud_free_slab, slab, us_link);
4372 		dom->ud_free_slabs++;
4373 	} else if (slab->us_freecount == 0) {
4374 		LIST_REMOVE(slab, us_link);
4375 		LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
4376 	}
4377 
4378 	/* Slab management. */
4379 	freei = slab_item_index(slab, keg, item);
4380 	BIT_SET(keg->uk_ipers, freei, &slab->us_free);
4381 	slab->us_freecount++;
4382 
4383 	/* Keg statistics. */
4384 	dom->ud_free_items++;
4385 }
4386 
4387 static void
4388 zone_release(void *arg, void **bucket, int cnt)
4389 {
4390 	struct mtx *lock;
4391 	uma_zone_t zone;
4392 	uma_slab_t slab;
4393 	uma_keg_t keg;
4394 	uint8_t *mem;
4395 	void *item;
4396 	int i;
4397 
4398 	zone = arg;
4399 	keg = zone->uz_keg;
4400 	lock = NULL;
4401 	if (__predict_false((zone->uz_flags & UMA_ZFLAG_HASH) != 0))
4402 		lock = KEG_LOCK(keg, 0);
4403 	for (i = 0; i < cnt; i++) {
4404 		item = bucket[i];
4405 		if (__predict_true((zone->uz_flags & UMA_ZFLAG_VTOSLAB) != 0)) {
4406 			slab = vtoslab((vm_offset_t)item);
4407 		} else {
4408 			mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
4409 			if ((zone->uz_flags & UMA_ZFLAG_HASH) != 0)
4410 				slab = hash_sfind(&keg->uk_hash, mem);
4411 			else
4412 				slab = (uma_slab_t)(mem + keg->uk_pgoff);
4413 		}
4414 		if (lock != KEG_LOCKPTR(keg, slab->us_domain)) {
4415 			if (lock != NULL)
4416 				mtx_unlock(lock);
4417 			lock = KEG_LOCK(keg, slab->us_domain);
4418 		}
4419 		slab_free_item(zone, slab, item);
4420 	}
4421 	if (lock != NULL)
4422 		mtx_unlock(lock);
4423 }
4424 
4425 /*
4426  * Frees a single item to any zone.
4427  *
4428  * Arguments:
4429  *	zone   The zone to free to
4430  *	item   The item we're freeing
4431  *	udata  User supplied data for the dtor
4432  *	skip   Skip dtors and finis
4433  */
4434 static __noinline void
4435 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
4436 {
4437 
4438 	/*
4439 	 * If a free is sent directly to an SMR zone we have to
4440 	 * synchronize immediately because the item can instantly
4441 	 * be reallocated. This should only happen in degenerate
4442 	 * cases when no memory is available for per-cpu caches.
4443 	 */
4444 	if ((zone->uz_flags & UMA_ZONE_SMR) != 0 && skip == SKIP_NONE)
4445 		smr_synchronize(zone->uz_smr);
4446 
4447 	item_dtor(zone, item, zone->uz_size, udata, skip);
4448 
4449 	if (skip < SKIP_FINI && zone->uz_fini)
4450 		zone->uz_fini(item, zone->uz_size);
4451 
4452 	zone->uz_release(zone->uz_arg, &item, 1);
4453 
4454 	if (skip & SKIP_CNT)
4455 		return;
4456 
4457 	counter_u64_add(zone->uz_frees, 1);
4458 
4459 	if (zone->uz_max_items > 0)
4460 		zone_free_limit(zone, 1);
4461 }
4462 
4463 /* See uma.h */
4464 int
4465 uma_zone_set_max(uma_zone_t zone, int nitems)
4466 {
4467 	struct uma_bucket_zone *ubz;
4468 	int count;
4469 
4470 	/*
4471 	 * XXX This can misbehave if the zone has any allocations with
4472 	 * no limit and a limit is imposed.  There is currently no
4473 	 * way to clear a limit.
4474 	 */
4475 	ZONE_LOCK(zone);
4476 	ubz = bucket_zone_max(zone, nitems);
4477 	count = ubz != NULL ? ubz->ubz_entries : 0;
4478 	zone->uz_bucket_size_max = zone->uz_bucket_size = count;
4479 	if (zone->uz_bucket_size_min > zone->uz_bucket_size_max)
4480 		zone->uz_bucket_size_min = zone->uz_bucket_size_max;
4481 	zone->uz_max_items = nitems;
4482 	zone->uz_flags |= UMA_ZFLAG_LIMIT;
4483 	zone_update_caches(zone);
4484 	/* We may need to wake waiters. */
4485 	wakeup(&zone->uz_max_items);
4486 	ZONE_UNLOCK(zone);
4487 
4488 	return (nitems);
4489 }
4490 
4491 /* See uma.h */
4492 void
4493 uma_zone_set_maxcache(uma_zone_t zone, int nitems)
4494 {
4495 	struct uma_bucket_zone *ubz;
4496 	int bpcpu;
4497 
4498 	ZONE_LOCK(zone);
4499 	ubz = bucket_zone_max(zone, nitems);
4500 	if (ubz != NULL) {
4501 		bpcpu = 2;
4502 		if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4503 			/* Count the cross-domain bucket. */
4504 			bpcpu++;
4505 		nitems -= ubz->ubz_entries * bpcpu * mp_ncpus;
4506 		zone->uz_bucket_size_max = ubz->ubz_entries;
4507 	} else {
4508 		zone->uz_bucket_size_max = zone->uz_bucket_size = 0;
4509 	}
4510 	if (zone->uz_bucket_size_min > zone->uz_bucket_size_max)
4511 		zone->uz_bucket_size_min = zone->uz_bucket_size_max;
4512 	zone->uz_bucket_max = nitems / vm_ndomains;
4513 	ZONE_UNLOCK(zone);
4514 }
4515 
4516 /* See uma.h */
4517 int
4518 uma_zone_get_max(uma_zone_t zone)
4519 {
4520 	int nitems;
4521 
4522 	nitems = atomic_load_64(&zone->uz_max_items);
4523 
4524 	return (nitems);
4525 }
4526 
4527 /* See uma.h */
4528 void
4529 uma_zone_set_warning(uma_zone_t zone, const char *warning)
4530 {
4531 
4532 	ZONE_ASSERT_COLD(zone);
4533 	zone->uz_warning = warning;
4534 }
4535 
4536 /* See uma.h */
4537 void
4538 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction)
4539 {
4540 
4541 	ZONE_ASSERT_COLD(zone);
4542 	TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone);
4543 }
4544 
4545 /* See uma.h */
4546 int
4547 uma_zone_get_cur(uma_zone_t zone)
4548 {
4549 	int64_t nitems;
4550 	u_int i;
4551 
4552 	nitems = 0;
4553 	if (zone->uz_allocs != EARLY_COUNTER && zone->uz_frees != EARLY_COUNTER)
4554 		nitems = counter_u64_fetch(zone->uz_allocs) -
4555 		    counter_u64_fetch(zone->uz_frees);
4556 	CPU_FOREACH(i)
4557 		nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs) -
4558 		    atomic_load_64(&zone->uz_cpu[i].uc_frees);
4559 
4560 	return (nitems < 0 ? 0 : nitems);
4561 }
4562 
4563 static uint64_t
4564 uma_zone_get_allocs(uma_zone_t zone)
4565 {
4566 	uint64_t nitems;
4567 	u_int i;
4568 
4569 	nitems = 0;
4570 	if (zone->uz_allocs != EARLY_COUNTER)
4571 		nitems = counter_u64_fetch(zone->uz_allocs);
4572 	CPU_FOREACH(i)
4573 		nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs);
4574 
4575 	return (nitems);
4576 }
4577 
4578 static uint64_t
4579 uma_zone_get_frees(uma_zone_t zone)
4580 {
4581 	uint64_t nitems;
4582 	u_int i;
4583 
4584 	nitems = 0;
4585 	if (zone->uz_frees != EARLY_COUNTER)
4586 		nitems = counter_u64_fetch(zone->uz_frees);
4587 	CPU_FOREACH(i)
4588 		nitems += atomic_load_64(&zone->uz_cpu[i].uc_frees);
4589 
4590 	return (nitems);
4591 }
4592 
4593 #ifdef INVARIANTS
4594 /* Used only for KEG_ASSERT_COLD(). */
4595 static uint64_t
4596 uma_keg_get_allocs(uma_keg_t keg)
4597 {
4598 	uma_zone_t z;
4599 	uint64_t nitems;
4600 
4601 	nitems = 0;
4602 	LIST_FOREACH(z, &keg->uk_zones, uz_link)
4603 		nitems += uma_zone_get_allocs(z);
4604 
4605 	return (nitems);
4606 }
4607 #endif
4608 
4609 /* See uma.h */
4610 void
4611 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
4612 {
4613 	uma_keg_t keg;
4614 
4615 	KEG_GET(zone, keg);
4616 	KEG_ASSERT_COLD(keg);
4617 	keg->uk_init = uminit;
4618 }
4619 
4620 /* See uma.h */
4621 void
4622 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
4623 {
4624 	uma_keg_t keg;
4625 
4626 	KEG_GET(zone, keg);
4627 	KEG_ASSERT_COLD(keg);
4628 	keg->uk_fini = fini;
4629 }
4630 
4631 /* See uma.h */
4632 void
4633 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
4634 {
4635 
4636 	ZONE_ASSERT_COLD(zone);
4637 	zone->uz_init = zinit;
4638 }
4639 
4640 /* See uma.h */
4641 void
4642 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
4643 {
4644 
4645 	ZONE_ASSERT_COLD(zone);
4646 	zone->uz_fini = zfini;
4647 }
4648 
4649 /* See uma.h */
4650 void
4651 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
4652 {
4653 	uma_keg_t keg;
4654 
4655 	KEG_GET(zone, keg);
4656 	KEG_ASSERT_COLD(keg);
4657 	keg->uk_freef = freef;
4658 }
4659 
4660 /* See uma.h */
4661 void
4662 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
4663 {
4664 	uma_keg_t keg;
4665 
4666 	KEG_GET(zone, keg);
4667 	KEG_ASSERT_COLD(keg);
4668 	keg->uk_allocf = allocf;
4669 }
4670 
4671 /* See uma.h */
4672 void
4673 uma_zone_set_smr(uma_zone_t zone, smr_t smr)
4674 {
4675 
4676 	ZONE_ASSERT_COLD(zone);
4677 
4678 	KASSERT(smr != NULL, ("Got NULL smr"));
4679 	KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
4680 	    ("zone %p (%s) already uses SMR", zone, zone->uz_name));
4681 	zone->uz_flags |= UMA_ZONE_SMR;
4682 	zone->uz_smr = smr;
4683 	zone_update_caches(zone);
4684 }
4685 
4686 smr_t
4687 uma_zone_get_smr(uma_zone_t zone)
4688 {
4689 
4690 	return (zone->uz_smr);
4691 }
4692 
4693 /* See uma.h */
4694 void
4695 uma_zone_reserve(uma_zone_t zone, int items)
4696 {
4697 	uma_keg_t keg;
4698 
4699 	KEG_GET(zone, keg);
4700 	KEG_ASSERT_COLD(keg);
4701 	keg->uk_reserve = items;
4702 }
4703 
4704 /* See uma.h */
4705 int
4706 uma_zone_reserve_kva(uma_zone_t zone, int count)
4707 {
4708 	uma_keg_t keg;
4709 	vm_offset_t kva;
4710 	u_int pages;
4711 
4712 	KEG_GET(zone, keg);
4713 	KEG_ASSERT_COLD(keg);
4714 	ZONE_ASSERT_COLD(zone);
4715 
4716 	pages = howmany(count, keg->uk_ipers) * keg->uk_ppera;
4717 
4718 #ifdef UMA_MD_SMALL_ALLOC
4719 	if (keg->uk_ppera > 1) {
4720 #else
4721 	if (1) {
4722 #endif
4723 		kva = kva_alloc((vm_size_t)pages * PAGE_SIZE);
4724 		if (kva == 0)
4725 			return (0);
4726 	} else
4727 		kva = 0;
4728 
4729 	MPASS(keg->uk_kva == 0);
4730 	keg->uk_kva = kva;
4731 	keg->uk_offset = 0;
4732 	zone->uz_max_items = pages * keg->uk_ipers;
4733 #ifdef UMA_MD_SMALL_ALLOC
4734 	keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
4735 #else
4736 	keg->uk_allocf = noobj_alloc;
4737 #endif
4738 	keg->uk_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
4739 	zone->uz_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
4740 	zone_update_caches(zone);
4741 
4742 	return (1);
4743 }
4744 
4745 /* See uma.h */
4746 void
4747 uma_prealloc(uma_zone_t zone, int items)
4748 {
4749 	struct vm_domainset_iter di;
4750 	uma_domain_t dom;
4751 	uma_slab_t slab;
4752 	uma_keg_t keg;
4753 	int aflags, domain, slabs;
4754 
4755 	KEG_GET(zone, keg);
4756 	slabs = howmany(items, keg->uk_ipers);
4757 	while (slabs-- > 0) {
4758 		aflags = M_NOWAIT;
4759 		vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
4760 		    &aflags);
4761 		for (;;) {
4762 			slab = keg_alloc_slab(keg, zone, domain, M_WAITOK,
4763 			    aflags);
4764 			if (slab != NULL) {
4765 				dom = &keg->uk_domain[slab->us_domain];
4766 				/*
4767 				 * keg_alloc_slab() always returns a slab on the
4768 				 * partial list.
4769 				 */
4770 				LIST_REMOVE(slab, us_link);
4771 				LIST_INSERT_HEAD(&dom->ud_free_slab, slab,
4772 				    us_link);
4773 				dom->ud_free_slabs++;
4774 				KEG_UNLOCK(keg, slab->us_domain);
4775 				break;
4776 			}
4777 			if (vm_domainset_iter_policy(&di, &domain) != 0)
4778 				vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask);
4779 		}
4780 	}
4781 }
4782 
4783 /*
4784  * Returns a snapshot of memory consumption in bytes.
4785  */
4786 size_t
4787 uma_zone_memory(uma_zone_t zone)
4788 {
4789 	size_t sz;
4790 	int i;
4791 
4792 	sz = 0;
4793 	if (zone->uz_flags & UMA_ZFLAG_CACHE) {
4794 		for (i = 0; i < vm_ndomains; i++)
4795 			sz += ZDOM_GET(zone, i)->uzd_nitems;
4796 		return (sz * zone->uz_size);
4797 	}
4798 	for (i = 0; i < vm_ndomains; i++)
4799 		sz += zone->uz_keg->uk_domain[i].ud_pages;
4800 
4801 	return (sz * PAGE_SIZE);
4802 }
4803 
4804 /* See uma.h */
4805 void
4806 uma_reclaim(int req)
4807 {
4808 
4809 	CTR0(KTR_UMA, "UMA: vm asked us to release pages!");
4810 	sx_xlock(&uma_reclaim_lock);
4811 	bucket_enable();
4812 
4813 	switch (req) {
4814 	case UMA_RECLAIM_TRIM:
4815 		zone_foreach(zone_trim, NULL);
4816 		break;
4817 	case UMA_RECLAIM_DRAIN:
4818 	case UMA_RECLAIM_DRAIN_CPU:
4819 		zone_foreach(zone_drain, NULL);
4820 		if (req == UMA_RECLAIM_DRAIN_CPU) {
4821 			pcpu_cache_drain_safe(NULL);
4822 			zone_foreach(zone_drain, NULL);
4823 		}
4824 		break;
4825 	default:
4826 		panic("unhandled reclamation request %d", req);
4827 	}
4828 
4829 	/*
4830 	 * Some slabs may have been freed but this zone will be visited early
4831 	 * we visit again so that we can free pages that are empty once other
4832 	 * zones are drained.  We have to do the same for buckets.
4833 	 */
4834 	zone_drain(slabzones[0], NULL);
4835 	zone_drain(slabzones[1], NULL);
4836 	bucket_zone_drain();
4837 	sx_xunlock(&uma_reclaim_lock);
4838 }
4839 
4840 static volatile int uma_reclaim_needed;
4841 
4842 void
4843 uma_reclaim_wakeup(void)
4844 {
4845 
4846 	if (atomic_fetchadd_int(&uma_reclaim_needed, 1) == 0)
4847 		wakeup(uma_reclaim);
4848 }
4849 
4850 void
4851 uma_reclaim_worker(void *arg __unused)
4852 {
4853 
4854 	for (;;) {
4855 		sx_xlock(&uma_reclaim_lock);
4856 		while (atomic_load_int(&uma_reclaim_needed) == 0)
4857 			sx_sleep(uma_reclaim, &uma_reclaim_lock, PVM, "umarcl",
4858 			    hz);
4859 		sx_xunlock(&uma_reclaim_lock);
4860 		EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM);
4861 		uma_reclaim(UMA_RECLAIM_DRAIN_CPU);
4862 		atomic_store_int(&uma_reclaim_needed, 0);
4863 		/* Don't fire more than once per-second. */
4864 		pause("umarclslp", hz);
4865 	}
4866 }
4867 
4868 /* See uma.h */
4869 void
4870 uma_zone_reclaim(uma_zone_t zone, int req)
4871 {
4872 
4873 	switch (req) {
4874 	case UMA_RECLAIM_TRIM:
4875 		zone_trim(zone, NULL);
4876 		break;
4877 	case UMA_RECLAIM_DRAIN:
4878 		zone_drain(zone, NULL);
4879 		break;
4880 	case UMA_RECLAIM_DRAIN_CPU:
4881 		pcpu_cache_drain_safe(zone);
4882 		zone_drain(zone, NULL);
4883 		break;
4884 	default:
4885 		panic("unhandled reclamation request %d", req);
4886 	}
4887 }
4888 
4889 /* See uma.h */
4890 int
4891 uma_zone_exhausted(uma_zone_t zone)
4892 {
4893 
4894 	return (atomic_load_32(&zone->uz_sleepers) > 0);
4895 }
4896 
4897 unsigned long
4898 uma_limit(void)
4899 {
4900 
4901 	return (uma_kmem_limit);
4902 }
4903 
4904 void
4905 uma_set_limit(unsigned long limit)
4906 {
4907 
4908 	uma_kmem_limit = limit;
4909 }
4910 
4911 unsigned long
4912 uma_size(void)
4913 {
4914 
4915 	return (atomic_load_long(&uma_kmem_total));
4916 }
4917 
4918 long
4919 uma_avail(void)
4920 {
4921 
4922 	return (uma_kmem_limit - uma_size());
4923 }
4924 
4925 #ifdef DDB
4926 /*
4927  * Generate statistics across both the zone and its per-cpu cache's.  Return
4928  * desired statistics if the pointer is non-NULL for that statistic.
4929  *
4930  * Note: does not update the zone statistics, as it can't safely clear the
4931  * per-CPU cache statistic.
4932  *
4933  */
4934 static void
4935 uma_zone_sumstat(uma_zone_t z, long *cachefreep, uint64_t *allocsp,
4936     uint64_t *freesp, uint64_t *sleepsp, uint64_t *xdomainp)
4937 {
4938 	uma_cache_t cache;
4939 	uint64_t allocs, frees, sleeps, xdomain;
4940 	int cachefree, cpu;
4941 
4942 	allocs = frees = sleeps = xdomain = 0;
4943 	cachefree = 0;
4944 	CPU_FOREACH(cpu) {
4945 		cache = &z->uz_cpu[cpu];
4946 		cachefree += cache->uc_allocbucket.ucb_cnt;
4947 		cachefree += cache->uc_freebucket.ucb_cnt;
4948 		xdomain += cache->uc_crossbucket.ucb_cnt;
4949 		cachefree += cache->uc_crossbucket.ucb_cnt;
4950 		allocs += cache->uc_allocs;
4951 		frees += cache->uc_frees;
4952 	}
4953 	allocs += counter_u64_fetch(z->uz_allocs);
4954 	frees += counter_u64_fetch(z->uz_frees);
4955 	xdomain += counter_u64_fetch(z->uz_xdomain);
4956 	sleeps += z->uz_sleeps;
4957 	if (cachefreep != NULL)
4958 		*cachefreep = cachefree;
4959 	if (allocsp != NULL)
4960 		*allocsp = allocs;
4961 	if (freesp != NULL)
4962 		*freesp = frees;
4963 	if (sleepsp != NULL)
4964 		*sleepsp = sleeps;
4965 	if (xdomainp != NULL)
4966 		*xdomainp = xdomain;
4967 }
4968 #endif /* DDB */
4969 
4970 static int
4971 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
4972 {
4973 	uma_keg_t kz;
4974 	uma_zone_t z;
4975 	int count;
4976 
4977 	count = 0;
4978 	rw_rlock(&uma_rwlock);
4979 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
4980 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
4981 			count++;
4982 	}
4983 	LIST_FOREACH(z, &uma_cachezones, uz_link)
4984 		count++;
4985 
4986 	rw_runlock(&uma_rwlock);
4987 	return (sysctl_handle_int(oidp, &count, 0, req));
4988 }
4989 
4990 static void
4991 uma_vm_zone_stats(struct uma_type_header *uth, uma_zone_t z, struct sbuf *sbuf,
4992     struct uma_percpu_stat *ups, bool internal)
4993 {
4994 	uma_zone_domain_t zdom;
4995 	uma_cache_t cache;
4996 	int i;
4997 
4998 
4999 	for (i = 0; i < vm_ndomains; i++) {
5000 		zdom = ZDOM_GET(z, i);
5001 		uth->uth_zone_free += zdom->uzd_nitems;
5002 	}
5003 	uth->uth_allocs = counter_u64_fetch(z->uz_allocs);
5004 	uth->uth_frees = counter_u64_fetch(z->uz_frees);
5005 	uth->uth_fails = counter_u64_fetch(z->uz_fails);
5006 	uth->uth_xdomain = counter_u64_fetch(z->uz_xdomain);
5007 	uth->uth_sleeps = z->uz_sleeps;
5008 
5009 	for (i = 0; i < mp_maxid + 1; i++) {
5010 		bzero(&ups[i], sizeof(*ups));
5011 		if (internal || CPU_ABSENT(i))
5012 			continue;
5013 		cache = &z->uz_cpu[i];
5014 		ups[i].ups_cache_free += cache->uc_allocbucket.ucb_cnt;
5015 		ups[i].ups_cache_free += cache->uc_freebucket.ucb_cnt;
5016 		ups[i].ups_cache_free += cache->uc_crossbucket.ucb_cnt;
5017 		ups[i].ups_allocs = cache->uc_allocs;
5018 		ups[i].ups_frees = cache->uc_frees;
5019 	}
5020 }
5021 
5022 static int
5023 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
5024 {
5025 	struct uma_stream_header ush;
5026 	struct uma_type_header uth;
5027 	struct uma_percpu_stat *ups;
5028 	struct sbuf sbuf;
5029 	uma_keg_t kz;
5030 	uma_zone_t z;
5031 	uint64_t items;
5032 	uint32_t kfree, pages;
5033 	int count, error, i;
5034 
5035 	error = sysctl_wire_old_buffer(req, 0);
5036 	if (error != 0)
5037 		return (error);
5038 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
5039 	sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL);
5040 	ups = malloc((mp_maxid + 1) * sizeof(*ups), M_TEMP, M_WAITOK);
5041 
5042 	count = 0;
5043 	rw_rlock(&uma_rwlock);
5044 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
5045 		LIST_FOREACH(z, &kz->uk_zones, uz_link)
5046 			count++;
5047 	}
5048 
5049 	LIST_FOREACH(z, &uma_cachezones, uz_link)
5050 		count++;
5051 
5052 	/*
5053 	 * Insert stream header.
5054 	 */
5055 	bzero(&ush, sizeof(ush));
5056 	ush.ush_version = UMA_STREAM_VERSION;
5057 	ush.ush_maxcpus = (mp_maxid + 1);
5058 	ush.ush_count = count;
5059 	(void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
5060 
5061 	LIST_FOREACH(kz, &uma_kegs, uk_link) {
5062 		kfree = pages = 0;
5063 		for (i = 0; i < vm_ndomains; i++) {
5064 			kfree += kz->uk_domain[i].ud_free_items;
5065 			pages += kz->uk_domain[i].ud_pages;
5066 		}
5067 		LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5068 			bzero(&uth, sizeof(uth));
5069 			strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5070 			uth.uth_align = kz->uk_align;
5071 			uth.uth_size = kz->uk_size;
5072 			uth.uth_rsize = kz->uk_rsize;
5073 			if (z->uz_max_items > 0) {
5074 				items = UZ_ITEMS_COUNT(z->uz_items);
5075 				uth.uth_pages = (items / kz->uk_ipers) *
5076 					kz->uk_ppera;
5077 			} else
5078 				uth.uth_pages = pages;
5079 			uth.uth_maxpages = (z->uz_max_items / kz->uk_ipers) *
5080 			    kz->uk_ppera;
5081 			uth.uth_limit = z->uz_max_items;
5082 			uth.uth_keg_free = kfree;
5083 
5084 			/*
5085 			 * A zone is secondary is it is not the first entry
5086 			 * on the keg's zone list.
5087 			 */
5088 			if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
5089 			    (LIST_FIRST(&kz->uk_zones) != z))
5090 				uth.uth_zone_flags = UTH_ZONE_SECONDARY;
5091 			uma_vm_zone_stats(&uth, z, &sbuf, ups,
5092 			    kz->uk_flags & UMA_ZFLAG_INTERNAL);
5093 			(void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5094 			for (i = 0; i < mp_maxid + 1; i++)
5095 				(void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5096 		}
5097 	}
5098 	LIST_FOREACH(z, &uma_cachezones, uz_link) {
5099 		bzero(&uth, sizeof(uth));
5100 		strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5101 		uth.uth_size = z->uz_size;
5102 		uma_vm_zone_stats(&uth, z, &sbuf, ups, false);
5103 		(void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5104 		for (i = 0; i < mp_maxid + 1; i++)
5105 			(void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5106 	}
5107 
5108 	rw_runlock(&uma_rwlock);
5109 	error = sbuf_finish(&sbuf);
5110 	sbuf_delete(&sbuf);
5111 	free(ups, M_TEMP);
5112 	return (error);
5113 }
5114 
5115 int
5116 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS)
5117 {
5118 	uma_zone_t zone = *(uma_zone_t *)arg1;
5119 	int error, max;
5120 
5121 	max = uma_zone_get_max(zone);
5122 	error = sysctl_handle_int(oidp, &max, 0, req);
5123 	if (error || !req->newptr)
5124 		return (error);
5125 
5126 	uma_zone_set_max(zone, max);
5127 
5128 	return (0);
5129 }
5130 
5131 int
5132 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS)
5133 {
5134 	uma_zone_t zone;
5135 	int cur;
5136 
5137 	/*
5138 	 * Some callers want to add sysctls for global zones that
5139 	 * may not yet exist so they pass a pointer to a pointer.
5140 	 */
5141 	if (arg2 == 0)
5142 		zone = *(uma_zone_t *)arg1;
5143 	else
5144 		zone = arg1;
5145 	cur = uma_zone_get_cur(zone);
5146 	return (sysctl_handle_int(oidp, &cur, 0, req));
5147 }
5148 
5149 static int
5150 sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS)
5151 {
5152 	uma_zone_t zone = arg1;
5153 	uint64_t cur;
5154 
5155 	cur = uma_zone_get_allocs(zone);
5156 	return (sysctl_handle_64(oidp, &cur, 0, req));
5157 }
5158 
5159 static int
5160 sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS)
5161 {
5162 	uma_zone_t zone = arg1;
5163 	uint64_t cur;
5164 
5165 	cur = uma_zone_get_frees(zone);
5166 	return (sysctl_handle_64(oidp, &cur, 0, req));
5167 }
5168 
5169 static int
5170 sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS)
5171 {
5172 	struct sbuf sbuf;
5173 	uma_zone_t zone = arg1;
5174 	int error;
5175 
5176 	sbuf_new_for_sysctl(&sbuf, NULL, 0, req);
5177 	if (zone->uz_flags != 0)
5178 		sbuf_printf(&sbuf, "0x%b", zone->uz_flags, PRINT_UMA_ZFLAGS);
5179 	else
5180 		sbuf_printf(&sbuf, "0");
5181 	error = sbuf_finish(&sbuf);
5182 	sbuf_delete(&sbuf);
5183 
5184 	return (error);
5185 }
5186 
5187 static int
5188 sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS)
5189 {
5190 	uma_keg_t keg = arg1;
5191 	int avail, effpct, total;
5192 
5193 	total = keg->uk_ppera * PAGE_SIZE;
5194 	if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0)
5195 		total += slabzone(keg->uk_ipers)->uz_keg->uk_rsize;
5196 	/*
5197 	 * We consider the client's requested size and alignment here, not the
5198 	 * real size determination uk_rsize, because we also adjust the real
5199 	 * size for internal implementation reasons (max bitset size).
5200 	 */
5201 	avail = keg->uk_ipers * roundup2(keg->uk_size, keg->uk_align + 1);
5202 	if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
5203 		avail *= mp_maxid + 1;
5204 	effpct = 100 * avail / total;
5205 	return (sysctl_handle_int(oidp, &effpct, 0, req));
5206 }
5207 
5208 static int
5209 sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS)
5210 {
5211 	uma_zone_t zone = arg1;
5212 	uint64_t cur;
5213 
5214 	cur = UZ_ITEMS_COUNT(atomic_load_64(&zone->uz_items));
5215 	return (sysctl_handle_64(oidp, &cur, 0, req));
5216 }
5217 
5218 #ifdef INVARIANTS
5219 static uma_slab_t
5220 uma_dbg_getslab(uma_zone_t zone, void *item)
5221 {
5222 	uma_slab_t slab;
5223 	uma_keg_t keg;
5224 	uint8_t *mem;
5225 
5226 	/*
5227 	 * It is safe to return the slab here even though the
5228 	 * zone is unlocked because the item's allocation state
5229 	 * essentially holds a reference.
5230 	 */
5231 	mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
5232 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5233 		return (NULL);
5234 	if (zone->uz_flags & UMA_ZFLAG_VTOSLAB)
5235 		return (vtoslab((vm_offset_t)mem));
5236 	keg = zone->uz_keg;
5237 	if ((keg->uk_flags & UMA_ZFLAG_HASH) == 0)
5238 		return ((uma_slab_t)(mem + keg->uk_pgoff));
5239 	KEG_LOCK(keg, 0);
5240 	slab = hash_sfind(&keg->uk_hash, mem);
5241 	KEG_UNLOCK(keg, 0);
5242 
5243 	return (slab);
5244 }
5245 
5246 static bool
5247 uma_dbg_zskip(uma_zone_t zone, void *mem)
5248 {
5249 
5250 	if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5251 		return (true);
5252 
5253 	return (uma_dbg_kskip(zone->uz_keg, mem));
5254 }
5255 
5256 static bool
5257 uma_dbg_kskip(uma_keg_t keg, void *mem)
5258 {
5259 	uintptr_t idx;
5260 
5261 	if (dbg_divisor == 0)
5262 		return (true);
5263 
5264 	if (dbg_divisor == 1)
5265 		return (false);
5266 
5267 	idx = (uintptr_t)mem >> PAGE_SHIFT;
5268 	if (keg->uk_ipers > 1) {
5269 		idx *= keg->uk_ipers;
5270 		idx += ((uintptr_t)mem & PAGE_MASK) / keg->uk_rsize;
5271 	}
5272 
5273 	if ((idx / dbg_divisor) * dbg_divisor != idx) {
5274 		counter_u64_add(uma_skip_cnt, 1);
5275 		return (true);
5276 	}
5277 	counter_u64_add(uma_dbg_cnt, 1);
5278 
5279 	return (false);
5280 }
5281 
5282 /*
5283  * Set up the slab's freei data such that uma_dbg_free can function.
5284  *
5285  */
5286 static void
5287 uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item)
5288 {
5289 	uma_keg_t keg;
5290 	int freei;
5291 
5292 	if (slab == NULL) {
5293 		slab = uma_dbg_getslab(zone, item);
5294 		if (slab == NULL)
5295 			panic("uma: item %p did not belong to zone %s\n",
5296 			    item, zone->uz_name);
5297 	}
5298 	keg = zone->uz_keg;
5299 	freei = slab_item_index(slab, keg, item);
5300 
5301 	if (BIT_ISSET(keg->uk_ipers, freei, slab_dbg_bits(slab, keg)))
5302 		panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)\n",
5303 		    item, zone, zone->uz_name, slab, freei);
5304 	BIT_SET_ATOMIC(keg->uk_ipers, freei, slab_dbg_bits(slab, keg));
5305 }
5306 
5307 /*
5308  * Verifies freed addresses.  Checks for alignment, valid slab membership
5309  * and duplicate frees.
5310  *
5311  */
5312 static void
5313 uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item)
5314 {
5315 	uma_keg_t keg;
5316 	int freei;
5317 
5318 	if (slab == NULL) {
5319 		slab = uma_dbg_getslab(zone, item);
5320 		if (slab == NULL)
5321 			panic("uma: Freed item %p did not belong to zone %s\n",
5322 			    item, zone->uz_name);
5323 	}
5324 	keg = zone->uz_keg;
5325 	freei = slab_item_index(slab, keg, item);
5326 
5327 	if (freei >= keg->uk_ipers)
5328 		panic("Invalid free of %p from zone %p(%s) slab %p(%d)\n",
5329 		    item, zone, zone->uz_name, slab, freei);
5330 
5331 	if (slab_item(slab, keg, freei) != item)
5332 		panic("Unaligned free of %p from zone %p(%s) slab %p(%d)\n",
5333 		    item, zone, zone->uz_name, slab, freei);
5334 
5335 	if (!BIT_ISSET(keg->uk_ipers, freei, slab_dbg_bits(slab, keg)))
5336 		panic("Duplicate free of %p from zone %p(%s) slab %p(%d)\n",
5337 		    item, zone, zone->uz_name, slab, freei);
5338 
5339 	BIT_CLR_ATOMIC(keg->uk_ipers, freei, slab_dbg_bits(slab, keg));
5340 }
5341 #endif /* INVARIANTS */
5342 
5343 #ifdef DDB
5344 static int64_t
5345 get_uma_stats(uma_keg_t kz, uma_zone_t z, uint64_t *allocs, uint64_t *used,
5346     uint64_t *sleeps, long *cachefree, uint64_t *xdomain)
5347 {
5348 	uint64_t frees;
5349 	int i;
5350 
5351 	if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
5352 		*allocs = counter_u64_fetch(z->uz_allocs);
5353 		frees = counter_u64_fetch(z->uz_frees);
5354 		*sleeps = z->uz_sleeps;
5355 		*cachefree = 0;
5356 		*xdomain = 0;
5357 	} else
5358 		uma_zone_sumstat(z, cachefree, allocs, &frees, sleeps,
5359 		    xdomain);
5360 	for (i = 0; i < vm_ndomains; i++) {
5361 		*cachefree += ZDOM_GET(z, i)->uzd_nitems;
5362 		if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
5363 		    (LIST_FIRST(&kz->uk_zones) != z)))
5364 			*cachefree += kz->uk_domain[i].ud_free_items;
5365 	}
5366 	*used = *allocs - frees;
5367 	return (((int64_t)*used + *cachefree) * kz->uk_size);
5368 }
5369 
5370 DB_SHOW_COMMAND(uma, db_show_uma)
5371 {
5372 	const char *fmt_hdr, *fmt_entry;
5373 	uma_keg_t kz;
5374 	uma_zone_t z;
5375 	uint64_t allocs, used, sleeps, xdomain;
5376 	long cachefree;
5377 	/* variables for sorting */
5378 	uma_keg_t cur_keg;
5379 	uma_zone_t cur_zone, last_zone;
5380 	int64_t cur_size, last_size, size;
5381 	int ties;
5382 
5383 	/* /i option produces machine-parseable CSV output */
5384 	if (modif[0] == 'i') {
5385 		fmt_hdr = "%s,%s,%s,%s,%s,%s,%s,%s,%s\n";
5386 		fmt_entry = "\"%s\",%ju,%jd,%ld,%ju,%ju,%u,%jd,%ju\n";
5387 	} else {
5388 		fmt_hdr = "%18s %6s %7s %7s %11s %7s %7s %10s %8s\n";
5389 		fmt_entry = "%18s %6ju %7jd %7ld %11ju %7ju %7u %10jd %8ju\n";
5390 	}
5391 
5392 	db_printf(fmt_hdr, "Zone", "Size", "Used", "Free", "Requests",
5393 	    "Sleeps", "Bucket", "Total Mem", "XFree");
5394 
5395 	/* Sort the zones with largest size first. */
5396 	last_zone = NULL;
5397 	last_size = INT64_MAX;
5398 	for (;;) {
5399 		cur_zone = NULL;
5400 		cur_size = -1;
5401 		ties = 0;
5402 		LIST_FOREACH(kz, &uma_kegs, uk_link) {
5403 			LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5404 				/*
5405 				 * In the case of size ties, print out zones
5406 				 * in the order they are encountered.  That is,
5407 				 * when we encounter the most recently output
5408 				 * zone, we have already printed all preceding
5409 				 * ties, and we must print all following ties.
5410 				 */
5411 				if (z == last_zone) {
5412 					ties = 1;
5413 					continue;
5414 				}
5415 				size = get_uma_stats(kz, z, &allocs, &used,
5416 				    &sleeps, &cachefree, &xdomain);
5417 				if (size > cur_size && size < last_size + ties)
5418 				{
5419 					cur_size = size;
5420 					cur_zone = z;
5421 					cur_keg = kz;
5422 				}
5423 			}
5424 		}
5425 		if (cur_zone == NULL)
5426 			break;
5427 
5428 		size = get_uma_stats(cur_keg, cur_zone, &allocs, &used,
5429 		    &sleeps, &cachefree, &xdomain);
5430 		db_printf(fmt_entry, cur_zone->uz_name,
5431 		    (uintmax_t)cur_keg->uk_size, (intmax_t)used, cachefree,
5432 		    (uintmax_t)allocs, (uintmax_t)sleeps,
5433 		    (unsigned)cur_zone->uz_bucket_size, (intmax_t)size,
5434 		    xdomain);
5435 
5436 		if (db_pager_quit)
5437 			return;
5438 		last_zone = cur_zone;
5439 		last_size = cur_size;
5440 	}
5441 }
5442 
5443 DB_SHOW_COMMAND(umacache, db_show_umacache)
5444 {
5445 	uma_zone_t z;
5446 	uint64_t allocs, frees;
5447 	long cachefree;
5448 	int i;
5449 
5450 	db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
5451 	    "Requests", "Bucket");
5452 	LIST_FOREACH(z, &uma_cachezones, uz_link) {
5453 		uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL, NULL);
5454 		for (i = 0; i < vm_ndomains; i++)
5455 			cachefree += ZDOM_GET(z, i)->uzd_nitems;
5456 		db_printf("%18s %8ju %8jd %8ld %12ju %8u\n",
5457 		    z->uz_name, (uintmax_t)z->uz_size,
5458 		    (intmax_t)(allocs - frees), cachefree,
5459 		    (uintmax_t)allocs, z->uz_bucket_size);
5460 		if (db_pager_quit)
5461 			return;
5462 	}
5463 }
5464 #endif	/* DDB */
5465