xref: /freebsd/sys/vm/uma_int.h (revision e0c4386e7e71d93b0edc0c8fa156263fc4a8b0b6)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2002-2019 Jeffrey Roberson <jeff@FreeBSD.org>
5  * Copyright (c) 2004, 2005 Bosko Milekic <bmilekic@FreeBSD.org>
6  * All rights reserved.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice unmodified, this list of conditions, and the following
13  *    disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
19  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
20  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
21  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
22  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
23  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
24  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
25  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
26  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
27  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
28  *
29  */
30 
31 #include <sys/counter.h>
32 #include <sys/_bitset.h>
33 #include <sys/_domainset.h>
34 #include <sys/_task.h>
35 
36 /*
37  * This file includes definitions, structures, prototypes, and inlines that
38  * should not be used outside of the actual implementation of UMA.
39  */
40 
41 /*
42  * The brief summary;  Zones describe unique allocation types.  Zones are
43  * organized into per-CPU caches which are filled by buckets.  Buckets are
44  * organized according to memory domains.  Buckets are filled from kegs which
45  * are also organized according to memory domains.  Kegs describe a unique
46  * allocation type, backend memory provider, and layout.  Kegs are associated
47  * with one or more zones and zones reference one or more kegs.  Kegs provide
48  * slabs which are virtually contiguous collections of pages.  Each slab is
49  * broken down int one or more items that will satisfy an individual allocation.
50  *
51  * Allocation is satisfied in the following order:
52  * 1) Per-CPU cache
53  * 2) Per-domain cache of buckets
54  * 3) Slab from any of N kegs
55  * 4) Backend page provider
56  *
57  * More detail on individual objects is contained below:
58  *
59  * Kegs contain lists of slabs which are stored in either the full bin, empty
60  * bin, or partially allocated bin, to reduce fragmentation.  They also contain
61  * the user supplied value for size, which is adjusted for alignment purposes
62  * and rsize is the result of that.  The Keg also stores information for
63  * managing a hash of page addresses that maps pages to uma_slab_t structures
64  * for pages that don't have embedded uma_slab_t's.
65  *
66  * Keg slab lists are organized by memory domain to support NUMA allocation
67  * policies.  By default allocations are spread across domains to reduce the
68  * potential for hotspots.  Special keg creation flags may be specified to
69  * prefer location allocation.  However there is no strict enforcement as frees
70  * may happen on any CPU and these are returned to the CPU-local cache
71  * regardless of the originating domain.
72  *
73  * The uma_slab_t may be embedded in a UMA_SLAB_SIZE chunk of memory or it may
74  * be allocated off the page from a special slab zone.  The free list within a
75  * slab is managed with a bitmask.  For item sizes that would yield more than
76  * 10% memory waste we potentially allocate a separate uma_slab_t if this will
77  * improve the number of items per slab that will fit.
78  *
79  * The only really gross cases, with regards to memory waste, are for those
80  * items that are just over half the page size.   You can get nearly 50% waste,
81  * so you fall back to the memory footprint of the power of two allocator. I
82  * have looked at memory allocation sizes on many of the machines available to
83  * me, and there does not seem to be an abundance of allocations at this range
84  * so at this time it may not make sense to optimize for it.  This can, of
85  * course, be solved with dynamic slab sizes.
86  *
87  * Kegs may serve multiple Zones but by far most of the time they only serve
88  * one.  When a Zone is created, a Keg is allocated and setup for it.  While
89  * the backing Keg stores slabs, the Zone caches Buckets of items allocated
90  * from the slabs.  Each Zone is equipped with an init/fini and ctor/dtor
91  * pair, as well as with its own set of small per-CPU caches, layered above
92  * the Zone's general Bucket cache.
93  *
94  * The PCPU caches are protected by critical sections, and may be accessed
95  * safely only from their associated CPU, while the Zones backed by the same
96  * Keg all share a common Keg lock (to coalesce contention on the backing
97  * slabs).  The backing Keg typically only serves one Zone but in the case of
98  * multiple Zones, one of the Zones is considered the Primary Zone and all
99  * Zone-related stats from the Keg are done in the Primary Zone.  For an
100  * example of a Multi-Zone setup, refer to the Mbuf allocation code.
101  */
102 
103 /*
104  *	This is the representation for normal (Non OFFPAGE slab)
105  *
106  *	i == item
107  *	s == slab pointer
108  *
109  *	<----------------  Page (UMA_SLAB_SIZE) ------------------>
110  *	___________________________________________________________
111  *     | _  _  _  _  _  _  _  _  _  _  _  _  _  _  _   ___________ |
112  *     ||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i| |slab header||
113  *     ||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_| |___________||
114  *     |___________________________________________________________|
115  *
116  *
117  *	This is an OFFPAGE slab. These can be larger than UMA_SLAB_SIZE.
118  *
119  *	___________________________________________________________
120  *     | _  _  _  _  _  _  _  _  _  _  _  _  _  _  _  _  _  _  _   |
121  *     ||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i|  |
122  *     ||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_|  |
123  *     |___________________________________________________________|
124  *       ___________    ^
125  *	|slab header|   |
126  *	|___________|---*
127  *
128  */
129 
130 #ifndef VM_UMA_INT_H
131 #define VM_UMA_INT_H
132 
133 #define UMA_SLAB_SIZE	PAGE_SIZE	/* How big are our slabs? */
134 #define UMA_SLAB_MASK	(PAGE_SIZE - 1)	/* Mask to get back to the page */
135 #define UMA_SLAB_SHIFT	PAGE_SHIFT	/* Number of bits PAGE_MASK */
136 
137 /* Max waste percentage before going to off page slab management */
138 #define UMA_MAX_WASTE	10
139 
140 /* Max size of a CACHESPREAD slab. */
141 #define	UMA_CACHESPREAD_MAX_SIZE	(128 * 1024)
142 
143 /*
144  * These flags must not overlap with the UMA_ZONE flags specified in uma.h.
145  */
146 #define	UMA_ZFLAG_OFFPAGE	0x00200000	/*
147 						 * Force the slab structure
148 						 * allocation off of the real
149 						 * memory.
150 						 */
151 #define	UMA_ZFLAG_HASH		0x00400000	/*
152 						 * Use a hash table instead of
153 						 * caching information in the
154 						 * vm_page.
155 						 */
156 #define	UMA_ZFLAG_VTOSLAB	0x00800000	/*
157 						 * Zone uses vtoslab for
158 						 * lookup.
159 						 */
160 #define	UMA_ZFLAG_CTORDTOR	0x01000000	/* Zone has ctor/dtor set. */
161 #define	UMA_ZFLAG_LIMIT		0x02000000	/* Zone has limit set. */
162 #define	UMA_ZFLAG_CACHE		0x04000000	/* uma_zcache_create()d it */
163 #define	UMA_ZFLAG_BUCKET	0x10000000	/* Bucket zone. */
164 #define	UMA_ZFLAG_INTERNAL	0x20000000	/* No offpage no PCPU. */
165 #define	UMA_ZFLAG_TRASH		0x40000000	/* Add trash ctor/dtor. */
166 
167 #define	UMA_ZFLAG_INHERIT						\
168     (UMA_ZFLAG_OFFPAGE | UMA_ZFLAG_HASH | UMA_ZFLAG_VTOSLAB |		\
169      UMA_ZFLAG_BUCKET | UMA_ZFLAG_INTERNAL)
170 
171 #define	PRINT_UMA_ZFLAGS	"\20"	\
172     "\37TRASH"				\
173     "\36INTERNAL"			\
174     "\35BUCKET"				\
175     "\33CACHE"				\
176     "\32LIMIT"				\
177     "\31CTORDTOR"			\
178     "\30VTOSLAB"			\
179     "\27HASH"				\
180     "\26OFFPAGE"			\
181     "\23SMR"				\
182     "\22ROUNDROBIN"			\
183     "\21FIRSTTOUCH"			\
184     "\20PCPU"				\
185     "\17NODUMP"				\
186     "\16CACHESPREAD"			\
187     "\14MAXBUCKET"			\
188     "\13NOBUCKET"			\
189     "\12SECONDARY"			\
190     "\11NOTPAGE"			\
191     "\10VM"				\
192     "\7MTXCLASS"			\
193     "\6NOFREE"				\
194     "\5MALLOC"				\
195     "\4NOTOUCH"				\
196     "\3CONTIG"				\
197     "\2ZINIT"
198 
199 /*
200  * Hash table for freed address -> slab translation.
201  *
202  * Only zones with memory not touchable by the allocator use the
203  * hash table.  Otherwise slabs are found with vtoslab().
204  */
205 #define UMA_HASH_SIZE_INIT	32
206 
207 #define UMA_HASH(h, s) ((((uintptr_t)s) >> UMA_SLAB_SHIFT) & (h)->uh_hashmask)
208 
209 #define UMA_HASH_INSERT(h, s, mem)					\
210 	LIST_INSERT_HEAD(&(h)->uh_slab_hash[UMA_HASH((h),		\
211 	    (mem))], slab_tohashslab(s), uhs_hlink)
212 
213 #define UMA_HASH_REMOVE(h, s)						\
214 	LIST_REMOVE(slab_tohashslab(s), uhs_hlink)
215 
216 LIST_HEAD(slabhashhead, uma_hash_slab);
217 
218 struct uma_hash {
219 	struct slabhashhead	*uh_slab_hash;	/* Hash table for slabs */
220 	u_int		uh_hashsize;	/* Current size of the hash table */
221 	u_int		uh_hashmask;	/* Mask used during hashing */
222 };
223 
224 /*
225  * Align field or structure to cache 'sector' in intel terminology.  This
226  * is more efficient with adjacent line prefetch.
227  */
228 #if defined(__amd64__) || defined(__powerpc64__)
229 #define UMA_SUPER_ALIGN	(CACHE_LINE_SIZE * 2)
230 #else
231 #define UMA_SUPER_ALIGN	CACHE_LINE_SIZE
232 #endif
233 
234 #define	UMA_ALIGN	__aligned(UMA_SUPER_ALIGN)
235 
236 /*
237  * The uma_bucket structure is used to queue and manage buckets divorced
238  * from per-cpu caches.  They are loaded into uma_cache_bucket structures
239  * for use.
240  */
241 struct uma_bucket {
242 	STAILQ_ENTRY(uma_bucket)	ub_link; /* Link into the zone */
243 	int16_t		ub_cnt;			/* Count of items in bucket. */
244 	int16_t		ub_entries;		/* Max items. */
245 	smr_seq_t	ub_seq;			/* SMR sequence number. */
246 	void		*ub_bucket[];		/* actual allocation storage */
247 };
248 
249 typedef struct uma_bucket * uma_bucket_t;
250 
251 /*
252  * The uma_cache_bucket structure is statically allocated on each per-cpu
253  * cache.  Its use reduces branches and cache misses in the fast path.
254  */
255 struct uma_cache_bucket {
256 	uma_bucket_t	ucb_bucket;
257 	int16_t		ucb_cnt;
258 	int16_t		ucb_entries;
259 	uint32_t	ucb_spare;
260 };
261 
262 typedef struct uma_cache_bucket * uma_cache_bucket_t;
263 
264 /*
265  * The uma_cache structure is allocated for each cpu for every zone
266  * type.  This optimizes synchronization out of the allocator fast path.
267  */
268 struct uma_cache {
269 	struct uma_cache_bucket	uc_freebucket;	/* Bucket we're freeing to */
270 	struct uma_cache_bucket	uc_allocbucket;	/* Bucket to allocate from */
271 	struct uma_cache_bucket	uc_crossbucket;	/* cross domain bucket */
272 	uint64_t		uc_allocs;	/* Count of allocations */
273 	uint64_t		uc_frees;	/* Count of frees */
274 } UMA_ALIGN;
275 
276 typedef struct uma_cache * uma_cache_t;
277 
278 LIST_HEAD(slabhead, uma_slab);
279 
280 /*
281  * The cache structure pads perfectly into 64 bytes so we use spare
282  * bits from the embedded cache buckets to store information from the zone
283  * and keep all fast-path allocations accessing a single per-cpu line.
284  */
285 static inline void
286 cache_set_uz_flags(uma_cache_t cache, uint32_t flags)
287 {
288 
289 	cache->uc_freebucket.ucb_spare = flags;
290 }
291 
292 static inline void
293 cache_set_uz_size(uma_cache_t cache, uint32_t size)
294 {
295 
296 	cache->uc_allocbucket.ucb_spare = size;
297 }
298 
299 static inline uint32_t
300 cache_uz_flags(uma_cache_t cache)
301 {
302 
303 	return (cache->uc_freebucket.ucb_spare);
304 }
305 
306 static inline uint32_t
307 cache_uz_size(uma_cache_t cache)
308 {
309 
310 	return (cache->uc_allocbucket.ucb_spare);
311 }
312 
313 /*
314  * Per-domain slab lists.  Embedded in the kegs.
315  */
316 struct uma_domain {
317 	struct mtx_padalign ud_lock;	/* Lock for the domain lists. */
318 	struct slabhead	ud_part_slab;	/* partially allocated slabs */
319 	struct slabhead	ud_free_slab;	/* completely unallocated slabs */
320 	struct slabhead ud_full_slab;	/* fully allocated slabs */
321 	uint32_t	ud_pages;	/* Total page count */
322 	uint32_t	ud_free_items;	/* Count of items free in all slabs */
323 	uint32_t	ud_free_slabs;	/* Count of free slabs */
324 } __aligned(CACHE_LINE_SIZE);
325 
326 typedef struct uma_domain * uma_domain_t;
327 
328 /*
329  * Keg management structure
330  *
331  * TODO: Optimize for cache line size
332  *
333  */
334 struct uma_keg {
335 	struct uma_hash	uk_hash;
336 	LIST_HEAD(,uma_zone)	uk_zones;	/* Keg's zones */
337 
338 	struct domainset_ref uk_dr;	/* Domain selection policy. */
339 	uint32_t	uk_align;	/* Alignment mask */
340 	uint32_t	uk_reserve;	/* Number of reserved items. */
341 	uint32_t	uk_size;	/* Requested size of each item */
342 	uint32_t	uk_rsize;	/* Real size of each item */
343 
344 	uma_init	uk_init;	/* Keg's init routine */
345 	uma_fini	uk_fini;	/* Keg's fini routine */
346 	uma_alloc	uk_allocf;	/* Allocation function */
347 	uma_free	uk_freef;	/* Free routine */
348 
349 	u_long		uk_offset;	/* Next free offset from base KVA */
350 	vm_offset_t	uk_kva;		/* Zone base KVA */
351 
352 	uint32_t	uk_pgoff;	/* Offset to uma_slab struct */
353 	uint16_t	uk_ppera;	/* pages per allocation from backend */
354 	uint16_t	uk_ipers;	/* Items per slab */
355 	uint32_t	uk_flags;	/* Internal flags */
356 
357 	/* Least used fields go to the last cache line. */
358 	const char	*uk_name;		/* Name of creating zone. */
359 	LIST_ENTRY(uma_keg)	uk_link;	/* List of all kegs */
360 
361 	/* Must be last, variable sized. */
362 	struct uma_domain	uk_domain[];	/* Keg's slab lists. */
363 };
364 typedef struct uma_keg	* uma_keg_t;
365 
366 /*
367  * Free bits per-slab.
368  */
369 #define	SLAB_MAX_SETSIZE	(PAGE_SIZE / UMA_SMALLEST_UNIT)
370 #define	SLAB_MIN_SETSIZE	_BITSET_BITS
371 BITSET_DEFINE(noslabbits, 0);
372 
373 /*
374  * The slab structure manages a single contiguous allocation from backing
375  * store and subdivides it into individually allocatable items.
376  */
377 struct uma_slab {
378 	LIST_ENTRY(uma_slab)	us_link;	/* slabs in zone */
379 	uint16_t	us_freecount;		/* How many are free? */
380 	uint8_t		us_flags;		/* Page flags see uma.h */
381 	uint8_t		us_domain;		/* Backing NUMA domain. */
382 	struct noslabbits us_free;		/* Free bitmask, flexible. */
383 };
384 _Static_assert(sizeof(struct uma_slab) == __offsetof(struct uma_slab, us_free),
385     "us_free field must be last");
386 _Static_assert(MAXMEMDOM < 255,
387     "us_domain field is not wide enough");
388 
389 typedef struct uma_slab * uma_slab_t;
390 
391 /*
392  * Slab structure with a full sized bitset and hash link for both
393  * HASH and OFFPAGE zones.
394  */
395 struct uma_hash_slab {
396 	LIST_ENTRY(uma_hash_slab) uhs_hlink;	/* Link for hash table */
397 	uint8_t			*uhs_data;	/* First item */
398 	struct uma_slab		uhs_slab;	/* Must be last. */
399 };
400 
401 typedef struct uma_hash_slab * uma_hash_slab_t;
402 
403 static inline uma_hash_slab_t
404 slab_tohashslab(uma_slab_t slab)
405 {
406 
407 	return (__containerof(slab, struct uma_hash_slab, uhs_slab));
408 }
409 
410 static inline void *
411 slab_data(uma_slab_t slab, uma_keg_t keg)
412 {
413 
414 	if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) == 0)
415 		return ((void *)((uintptr_t)slab - keg->uk_pgoff));
416 	else
417 		return (slab_tohashslab(slab)->uhs_data);
418 }
419 
420 static inline void *
421 slab_item(uma_slab_t slab, uma_keg_t keg, int index)
422 {
423 	uintptr_t data;
424 
425 	data = (uintptr_t)slab_data(slab, keg);
426 	return ((void *)(data + keg->uk_rsize * index));
427 }
428 
429 static inline int
430 slab_item_index(uma_slab_t slab, uma_keg_t keg, void *item)
431 {
432 	uintptr_t data;
433 
434 	data = (uintptr_t)slab_data(slab, keg);
435 	return (((uintptr_t)item - data) / keg->uk_rsize);
436 }
437 
438 STAILQ_HEAD(uma_bucketlist, uma_bucket);
439 
440 struct uma_zone_domain {
441 	struct uma_bucketlist uzd_buckets; /* full buckets */
442 	uma_bucket_t	uzd_cross;	/* Fills from cross buckets. */
443 	long		uzd_nitems;	/* total item count */
444 	long		uzd_imax;	/* maximum item count this period */
445 	long		uzd_imin;	/* minimum item count this period */
446 	long		uzd_bimin;	/* Minimum item count this batch. */
447 	long		uzd_wss;	/* working set size estimate */
448 	long		uzd_limin;	/* Longtime minimum item count. */
449 	u_int		uzd_timin;	/* Time since uzd_limin == 0. */
450 	smr_seq_t	uzd_seq;	/* Lowest queued seq. */
451 	struct mtx	uzd_lock;	/* Lock for the domain */
452 } __aligned(CACHE_LINE_SIZE);
453 
454 typedef struct uma_zone_domain * uma_zone_domain_t;
455 
456 /*
457  * Zone structure - per memory type.
458  */
459 struct uma_zone {
460 	/* Offset 0, used in alloc/free fast/medium fast path and const. */
461 	uint32_t	uz_flags;	/* Flags inherited from kegs */
462 	uint32_t	uz_size;	/* Size inherited from kegs */
463 	uma_ctor	uz_ctor;	/* Constructor for each allocation */
464 	uma_dtor	uz_dtor;	/* Destructor */
465 	smr_t		uz_smr;		/* Safe memory reclaim context. */
466 	uint64_t	uz_max_items;	/* Maximum number of items to alloc */
467 	uint64_t	uz_bucket_max;	/* Maximum bucket cache size */
468 	uint16_t	uz_bucket_size;	/* Number of items in full bucket */
469 	uint16_t	uz_bucket_size_max; /* Maximum number of bucket items */
470 	uint32_t	uz_sleepers;	/* Threads sleeping on limit */
471 	counter_u64_t	uz_xdomain;	/* Total number of cross-domain frees */
472 
473 	/* Offset 64, used in bucket replenish. */
474 	uma_keg_t	uz_keg;		/* This zone's keg if !CACHE */
475 	uma_import	uz_import;	/* Import new memory to cache. */
476 	uma_release	uz_release;	/* Release memory from cache. */
477 	void		*uz_arg;	/* Import/release argument. */
478 	uma_init	uz_init;	/* Initializer for each item */
479 	uma_fini	uz_fini;	/* Finalizer for each item. */
480 	volatile uint64_t uz_items;	/* Total items count & sleepers */
481 	uint64_t	uz_sleeps;	/* Total number of alloc sleeps */
482 
483 	/* Offset 128 Rare stats, misc read-only. */
484 	LIST_ENTRY(uma_zone) uz_link;	/* List of all zones in keg */
485 	counter_u64_t	uz_allocs;	/* Total number of allocations */
486 	counter_u64_t	uz_frees;	/* Total number of frees */
487 	counter_u64_t	uz_fails;	/* Total number of alloc failures */
488 	const char	*uz_name;	/* Text name of the zone */
489 	char		*uz_ctlname;	/* sysctl safe name string. */
490 	int		uz_namecnt;	/* duplicate name count. */
491 	uint16_t	uz_bucket_size_min; /* Min number of items in bucket */
492 	uint16_t	uz_reclaimers;	/* pending reclaim operations. */
493 
494 	/* Offset 192, rare read-only. */
495 	struct sysctl_oid *uz_oid;	/* sysctl oid pointer. */
496 	const char	*uz_warning;	/* Warning to print on failure */
497 	struct timeval	uz_ratecheck;	/* Warnings rate-limiting */
498 	struct task	uz_maxaction;	/* Task to run when at limit */
499 
500 	/* Offset 256. */
501 	struct mtx	uz_cross_lock;	/* Cross domain free lock */
502 
503 	/*
504 	 * This HAS to be the last item because we adjust the zone size
505 	 * based on NCPU and then allocate the space for the zones.
506 	 */
507 	struct uma_cache	uz_cpu[]; /* Per cpu caches */
508 
509 	/* domains follow here. */
510 };
511 
512 /*
513  * Macros for interpreting the uz_items field.  20 bits of sleeper count
514  * and 44 bit of item count.
515  */
516 #define	UZ_ITEMS_SLEEPER_SHIFT	44LL
517 #define	UZ_ITEMS_SLEEPERS_MAX	((1 << (64 - UZ_ITEMS_SLEEPER_SHIFT)) - 1)
518 #define	UZ_ITEMS_COUNT_MASK	((1LL << UZ_ITEMS_SLEEPER_SHIFT) - 1)
519 #define	UZ_ITEMS_COUNT(x)	((x) & UZ_ITEMS_COUNT_MASK)
520 #define	UZ_ITEMS_SLEEPERS(x)	((x) >> UZ_ITEMS_SLEEPER_SHIFT)
521 #define	UZ_ITEMS_SLEEPER	(1LL << UZ_ITEMS_SLEEPER_SHIFT)
522 
523 #define	ZONE_ASSERT_COLD(z)						\
524 	KASSERT(uma_zone_get_allocs((z)) == 0,				\
525 	    ("zone %s initialization after use.", (z)->uz_name))
526 
527 /* Domains are contiguous after the last CPU */
528 #define	ZDOM_GET(z, n)							\
529 	(&((uma_zone_domain_t)&(z)->uz_cpu[mp_maxid + 1])[n])
530 
531 #undef	UMA_ALIGN
532 
533 #ifdef _KERNEL
534 /* Internal prototypes */
535 static __inline uma_slab_t hash_sfind(struct uma_hash *hash, uint8_t *data);
536 
537 /* Lock Macros */
538 
539 #define	KEG_LOCKPTR(k, d)	(struct mtx *)&(k)->uk_domain[(d)].ud_lock
540 #define	KEG_LOCK_INIT(k, d, lc)						\
541 	do {								\
542 		if ((lc))						\
543 			mtx_init(KEG_LOCKPTR(k, d), (k)->uk_name,	\
544 			    (k)->uk_name, MTX_DEF | MTX_DUPOK);		\
545 		else							\
546 			mtx_init(KEG_LOCKPTR(k, d), (k)->uk_name,	\
547 			    "UMA zone", MTX_DEF | MTX_DUPOK);		\
548 	} while (0)
549 
550 #define	KEG_LOCK_FINI(k, d)	mtx_destroy(KEG_LOCKPTR(k, d))
551 #define	KEG_LOCK(k, d)							\
552 	({ mtx_lock(KEG_LOCKPTR(k, d)); KEG_LOCKPTR(k, d); })
553 #define	KEG_UNLOCK(k, d)	mtx_unlock(KEG_LOCKPTR(k, d))
554 #define	KEG_LOCK_ASSERT(k, d)	mtx_assert(KEG_LOCKPTR(k, d), MA_OWNED)
555 
556 #define	KEG_GET(zone, keg) do {					\
557 	(keg) = (zone)->uz_keg;					\
558 	KASSERT((void *)(keg) != NULL,				\
559 	    ("%s: Invalid zone %p type", __func__, (zone)));	\
560 	} while (0)
561 
562 #define	KEG_ASSERT_COLD(k)						\
563 	KASSERT(uma_keg_get_allocs((k)) == 0,				\
564 	    ("keg %s initialization after use.", (k)->uk_name))
565 
566 #define	ZDOM_LOCK_INIT(z, zdom, lc)					\
567 	do {								\
568 		if ((lc))						\
569 			mtx_init(&(zdom)->uzd_lock, (z)->uz_name,	\
570 			    (z)->uz_name, MTX_DEF | MTX_DUPOK);		\
571 		else							\
572 			mtx_init(&(zdom)->uzd_lock, (z)->uz_name,	\
573 			    "UMA zone", MTX_DEF | MTX_DUPOK);		\
574 	} while (0)
575 #define	ZDOM_LOCK_FINI(z)	mtx_destroy(&(z)->uzd_lock)
576 #define	ZDOM_LOCK_ASSERT(z)	mtx_assert(&(z)->uzd_lock, MA_OWNED)
577 
578 #define	ZDOM_LOCK(z)	mtx_lock(&(z)->uzd_lock)
579 #define	ZDOM_OWNED(z)	(mtx_owner(&(z)->uzd_lock) != NULL)
580 #define	ZDOM_UNLOCK(z)	mtx_unlock(&(z)->uzd_lock)
581 
582 #define	ZONE_LOCK(z)	ZDOM_LOCK(ZDOM_GET((z), 0))
583 #define	ZONE_UNLOCK(z)	ZDOM_UNLOCK(ZDOM_GET((z), 0))
584 #define	ZONE_LOCKPTR(z)	(&ZDOM_GET((z), 0)->uzd_lock)
585 
586 #define	ZONE_CROSS_LOCK_INIT(z)					\
587 	mtx_init(&(z)->uz_cross_lock, "UMA Cross", NULL, MTX_DEF)
588 #define	ZONE_CROSS_LOCK(z)	mtx_lock(&(z)->uz_cross_lock)
589 #define	ZONE_CROSS_UNLOCK(z)	mtx_unlock(&(z)->uz_cross_lock)
590 #define	ZONE_CROSS_LOCK_FINI(z)	mtx_destroy(&(z)->uz_cross_lock)
591 
592 /*
593  * Find a slab within a hash table.  This is used for OFFPAGE zones to lookup
594  * the slab structure.
595  *
596  * Arguments:
597  *	hash  The hash table to search.
598  *	data  The base page of the item.
599  *
600  * Returns:
601  *	A pointer to a slab if successful, else NULL.
602  */
603 static __inline uma_slab_t
604 hash_sfind(struct uma_hash *hash, uint8_t *data)
605 {
606         uma_hash_slab_t slab;
607         u_int hval;
608 
609         hval = UMA_HASH(hash, data);
610 
611         LIST_FOREACH(slab, &hash->uh_slab_hash[hval], uhs_hlink) {
612                 if ((uint8_t *)slab->uhs_data == data)
613                         return (&slab->uhs_slab);
614         }
615         return (NULL);
616 }
617 
618 static __inline uma_slab_t
619 vtoslab(vm_offset_t va)
620 {
621 	vm_page_t p;
622 
623 	p = PHYS_TO_VM_PAGE(pmap_kextract(va));
624 	return (p->plinks.uma.slab);
625 }
626 
627 static __inline void
628 vtozoneslab(vm_offset_t va, uma_zone_t *zone, uma_slab_t *slab)
629 {
630 	vm_page_t p;
631 
632 	p = PHYS_TO_VM_PAGE(pmap_kextract(va));
633 	*slab = p->plinks.uma.slab;
634 	*zone = p->plinks.uma.zone;
635 }
636 
637 static __inline void
638 vsetzoneslab(vm_offset_t va, uma_zone_t zone, uma_slab_t slab)
639 {
640 	vm_page_t p;
641 
642 	p = PHYS_TO_VM_PAGE(pmap_kextract(va));
643 	p->plinks.uma.slab = slab;
644 	p->plinks.uma.zone = zone;
645 }
646 
647 extern unsigned long uma_kmem_limit;
648 extern unsigned long uma_kmem_total;
649 
650 /* Adjust bytes under management by UMA. */
651 static inline void
652 uma_total_dec(unsigned long size)
653 {
654 
655 	atomic_subtract_long(&uma_kmem_total, size);
656 }
657 
658 static inline void
659 uma_total_inc(unsigned long size)
660 {
661 
662 	if (atomic_fetchadd_long(&uma_kmem_total, size) > uma_kmem_limit)
663 		uma_reclaim_wakeup();
664 }
665 
666 /*
667  * The following two functions may be defined by architecture specific code
668  * if they can provide more efficient allocation functions.  This is useful
669  * for using direct mapped addresses.
670  */
671 void *uma_small_alloc(uma_zone_t zone, vm_size_t bytes, int domain,
672     uint8_t *pflag, int wait);
673 void uma_small_free(void *mem, vm_size_t size, uint8_t flags);
674 
675 /* Set a global soft limit on UMA managed memory. */
676 void uma_set_limit(unsigned long limit);
677 
678 #endif /* _KERNEL */
679 
680 #endif /* VM_UMA_INT_H */
681