xref: /freebsd/sys/vm/uma_int.h (revision b249ce48ea5560afdcff57e72a9880b7d3132434)
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  * 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  * $FreeBSD$
30  *
31  */
32 
33 #include <sys/counter.h>
34 #include <sys/_bitset.h>
35 #include <sys/_domainset.h>
36 #include <sys/_task.h>
37 
38 /*
39  * This file includes definitions, structures, prototypes, and inlines that
40  * should not be used outside of the actual implementation of UMA.
41  */
42 
43 /*
44  * The brief summary;  Zones describe unique allocation types.  Zones are
45  * organized into per-CPU caches which are filled by buckets.  Buckets are
46  * organized according to memory domains.  Buckets are filled from kegs which
47  * are also organized according to memory domains.  Kegs describe a unique
48  * allocation type, backend memory provider, and layout.  Kegs are associated
49  * with one or more zones and zones reference one or more kegs.  Kegs provide
50  * slabs which are virtually contiguous collections of pages.  Each slab is
51  * broken down int one or more items that will satisfy an individual allocation.
52  *
53  * Allocation is satisfied in the following order:
54  * 1) Per-CPU cache
55  * 2) Per-domain cache of buckets
56  * 3) Slab from any of N kegs
57  * 4) Backend page provider
58  *
59  * More detail on individual objects is contained below:
60  *
61  * Kegs contain lists of slabs which are stored in either the full bin, empty
62  * bin, or partially allocated bin, to reduce fragmentation.  They also contain
63  * the user supplied value for size, which is adjusted for alignment purposes
64  * and rsize is the result of that.  The Keg also stores information for
65  * managing a hash of page addresses that maps pages to uma_slab_t structures
66  * for pages that don't have embedded uma_slab_t's.
67  *
68  * Keg slab lists are organized by memory domain to support NUMA allocation
69  * policies.  By default allocations are spread across domains to reduce the
70  * potential for hotspots.  Special keg creation flags may be specified to
71  * prefer location allocation.  However there is no strict enforcement as frees
72  * may happen on any CPU and these are returned to the CPU-local cache
73  * regardless of the originating domain.
74  *
75  * The uma_slab_t may be embedded in a UMA_SLAB_SIZE chunk of memory or it may
76  * be allocated off the page from a special slab zone.  The free list within a
77  * slab is managed with a bitmask.  For item sizes that would yield more than
78  * 10% memory waste we potentially allocate a separate uma_slab_t if this will
79  * improve the number of items per slab that will fit.
80  *
81  * The only really gross cases, with regards to memory waste, are for those
82  * items that are just over half the page size.   You can get nearly 50% waste,
83  * so you fall back to the memory footprint of the power of two allocator. I
84  * have looked at memory allocation sizes on many of the machines available to
85  * me, and there does not seem to be an abundance of allocations at this range
86  * so at this time it may not make sense to optimize for it.  This can, of
87  * course, be solved with dynamic slab sizes.
88  *
89  * Kegs may serve multiple Zones but by far most of the time they only serve
90  * one.  When a Zone is created, a Keg is allocated and setup for it.  While
91  * the backing Keg stores slabs, the Zone caches Buckets of items allocated
92  * from the slabs.  Each Zone is equipped with an init/fini and ctor/dtor
93  * pair, as well as with its own set of small per-CPU caches, layered above
94  * the Zone's general Bucket cache.
95  *
96  * The PCPU caches are protected by critical sections, and may be accessed
97  * safely only from their associated CPU, while the Zones backed by the same
98  * Keg all share a common Keg lock (to coalesce contention on the backing
99  * slabs).  The backing Keg typically only serves one Zone but in the case of
100  * multiple Zones, one of the Zones is considered the Master Zone and all
101  * Zone-related stats from the Keg are done in the Master Zone.  For an
102  * example of a Multi-Zone setup, refer to the Mbuf allocation code.
103  */
104 
105 /*
106  *	This is the representation for normal (Non OFFPAGE slab)
107  *
108  *	i == item
109  *	s == slab pointer
110  *
111  *	<----------------  Page (UMA_SLAB_SIZE) ------------------>
112  *	___________________________________________________________
113  *     | _  _  _  _  _  _  _  _  _  _  _  _  _  _  _   ___________ |
114  *     ||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i| |slab header||
115  *     ||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_| |___________||
116  *     |___________________________________________________________|
117  *
118  *
119  *	This is an OFFPAGE slab. These can be larger than UMA_SLAB_SIZE.
120  *
121  *	___________________________________________________________
122  *     | _  _  _  _  _  _  _  _  _  _  _  _  _  _  _  _  _  _  _   |
123  *     ||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i|  |
124  *     ||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_|  |
125  *     |___________________________________________________________|
126  *       ___________    ^
127  *	|slab header|   |
128  *	|___________|---*
129  *
130  */
131 
132 #ifndef VM_UMA_INT_H
133 #define VM_UMA_INT_H
134 
135 #define UMA_SLAB_SIZE	PAGE_SIZE	/* How big are our slabs? */
136 #define UMA_SLAB_MASK	(PAGE_SIZE - 1)	/* Mask to get back to the page */
137 #define UMA_SLAB_SHIFT	PAGE_SHIFT	/* Number of bits PAGE_MASK */
138 
139 /* Max waste percentage before going to off page slab management */
140 #define UMA_MAX_WASTE	10
141 
142 
143 /*
144  * Hash table for freed address -> slab translation.
145  *
146  * Only zones with memory not touchable by the allocator use the
147  * hash table.  Otherwise slabs are found with vtoslab().
148  */
149 #define UMA_HASH_SIZE_INIT	32
150 
151 #define UMA_HASH(h, s) ((((uintptr_t)s) >> UMA_SLAB_SHIFT) & (h)->uh_hashmask)
152 
153 #define UMA_HASH_INSERT(h, s, mem)					\
154 	LIST_INSERT_HEAD(&(h)->uh_slab_hash[UMA_HASH((h),		\
155 	    (mem))], (uma_hash_slab_t)(s), uhs_hlink)
156 
157 #define UMA_HASH_REMOVE(h, s)						\
158 	LIST_REMOVE((uma_hash_slab_t)(s), uhs_hlink)
159 
160 LIST_HEAD(slabhashhead, uma_hash_slab);
161 
162 struct uma_hash {
163 	struct slabhashhead	*uh_slab_hash;	/* Hash table for slabs */
164 	u_int		uh_hashsize;	/* Current size of the hash table */
165 	u_int		uh_hashmask;	/* Mask used during hashing */
166 };
167 
168 /*
169  * align field or structure to cache line
170  */
171 #if defined(__amd64__) || defined(__powerpc64__)
172 #define UMA_ALIGN	__aligned(128)
173 #else
174 #define UMA_ALIGN	__aligned(CACHE_LINE_SIZE)
175 #endif
176 
177 /*
178  * The uma_bucket structure is used to queue and manage buckets divorced
179  * from per-cpu caches.  They are loaded into uma_cache_bucket structures
180  * for use.
181  */
182 struct uma_bucket {
183 	TAILQ_ENTRY(uma_bucket)	ub_link;	/* Link into the zone */
184 	int16_t	ub_cnt;				/* Count of items in bucket. */
185 	int16_t	ub_entries;			/* Max items. */
186 	void	*ub_bucket[];			/* actual allocation storage */
187 };
188 
189 typedef struct uma_bucket * uma_bucket_t;
190 
191 /*
192  * The uma_cache_bucket structure is statically allocated on each per-cpu
193  * cache.  Its use reduces branches and cache misses in the fast path.
194  */
195 struct uma_cache_bucket {
196 	uma_bucket_t	ucb_bucket;
197 	int16_t		ucb_cnt;
198 	int16_t		ucb_entries;
199 	uint32_t	ucb_spare;
200 };
201 
202 typedef struct uma_cache_bucket * uma_cache_bucket_t;
203 
204 /*
205  * The uma_cache structure is allocated for each cpu for every zone
206  * type.  This optimizes synchronization out of the allocator fast path.
207  */
208 struct uma_cache {
209 	struct uma_cache_bucket	uc_freebucket;	/* Bucket we're freeing to */
210 	struct uma_cache_bucket	uc_allocbucket;	/* Bucket to allocate from */
211 	struct uma_cache_bucket	uc_crossbucket;	/* cross domain bucket */
212 	uint64_t		uc_allocs;	/* Count of allocations */
213 	uint64_t		uc_frees;	/* Count of frees */
214 } UMA_ALIGN;
215 
216 typedef struct uma_cache * uma_cache_t;
217 
218 LIST_HEAD(slabhead, uma_slab);
219 
220 /*
221  * The cache structure pads perfectly into 64 bytes so we use spare
222  * bits from the embedded cache buckets to store information from the zone
223  * and keep all fast-path allocations accessing a single per-cpu line.
224  */
225 static inline void
226 cache_set_uz_flags(uma_cache_t cache, uint32_t flags)
227 {
228 
229 	cache->uc_freebucket.ucb_spare = flags;
230 }
231 
232 static inline void
233 cache_set_uz_size(uma_cache_t cache, uint32_t size)
234 {
235 
236 	cache->uc_allocbucket.ucb_spare = size;
237 }
238 
239 static inline uint32_t
240 cache_uz_flags(uma_cache_t cache)
241 {
242 
243 	return (cache->uc_freebucket.ucb_spare);
244 }
245 
246 static inline uint32_t
247 cache_uz_size(uma_cache_t cache)
248 {
249 
250 	return (cache->uc_allocbucket.ucb_spare);
251 }
252 
253 /*
254  * Per-domain slab lists.  Embedded in the kegs.
255  */
256 struct uma_domain {
257 	struct slabhead	ud_part_slab;	/* partially allocated slabs */
258 	struct slabhead	ud_free_slab;	/* completely unallocated slabs */
259 	struct slabhead ud_full_slab;	/* fully allocated slabs */
260 };
261 
262 typedef struct uma_domain * uma_domain_t;
263 
264 /*
265  * Keg management structure
266  *
267  * TODO: Optimize for cache line size
268  *
269  */
270 struct uma_keg {
271 	struct mtx	uk_lock;	/* Lock for the keg must be first.
272 					 * See shared uz_keg/uz_lockptr
273 					 * member of struct uma_zone. */
274 	struct uma_hash	uk_hash;
275 	LIST_HEAD(,uma_zone)	uk_zones;	/* Keg's zones */
276 
277 	struct domainset_ref uk_dr;	/* Domain selection policy. */
278 	uint32_t	uk_align;	/* Alignment mask */
279 	uint32_t	uk_pages;	/* Total page count */
280 	uint32_t	uk_free;	/* Count of items free in slabs */
281 	uint32_t	uk_reserve;	/* Number of reserved items. */
282 	uint32_t	uk_size;	/* Requested size of each item */
283 	uint32_t	uk_rsize;	/* Real size of each item */
284 
285 	uma_init	uk_init;	/* Keg's init routine */
286 	uma_fini	uk_fini;	/* Keg's fini routine */
287 	uma_alloc	uk_allocf;	/* Allocation function */
288 	uma_free	uk_freef;	/* Free routine */
289 
290 	u_long		uk_offset;	/* Next free offset from base KVA */
291 	vm_offset_t	uk_kva;		/* Zone base KVA */
292 	uma_zone_t	uk_slabzone;	/* Slab zone backing us, if OFFPAGE */
293 
294 	uint32_t	uk_pgoff;	/* Offset to uma_slab struct */
295 	uint16_t	uk_ppera;	/* pages per allocation from backend */
296 	uint16_t	uk_ipers;	/* Items per slab */
297 	uint32_t	uk_flags;	/* Internal flags */
298 
299 	/* Least used fields go to the last cache line. */
300 	const char	*uk_name;		/* Name of creating zone. */
301 	LIST_ENTRY(uma_keg)	uk_link;	/* List of all kegs */
302 
303 	/* Must be last, variable sized. */
304 	struct uma_domain	uk_domain[];	/* Keg's slab lists. */
305 };
306 typedef struct uma_keg	* uma_keg_t;
307 
308 #ifdef _KERNEL
309 /*
310  * Free bits per-slab.
311  */
312 #define	SLAB_MAX_SETSIZE	(PAGE_SIZE / UMA_SMALLEST_UNIT)
313 #define	SLAB_MIN_SETSIZE	_BITSET_BITS
314 BITSET_DEFINE(slabbits, SLAB_MAX_SETSIZE);
315 BITSET_DEFINE(noslabbits, 0);
316 
317 /*
318  * The slab structure manages a single contiguous allocation from backing
319  * store and subdivides it into individually allocatable items.
320  */
321 struct uma_slab {
322 	LIST_ENTRY(uma_slab)	us_link;	/* slabs in zone */
323 	uint16_t	us_freecount;		/* How many are free? */
324 	uint8_t		us_flags;		/* Page flags see uma.h */
325 	uint8_t		us_domain;		/* Backing NUMA domain. */
326 	struct noslabbits us_free;		/* Free bitmask, flexible. */
327 };
328 _Static_assert(sizeof(struct uma_slab) == offsetof(struct uma_slab, us_free),
329     "us_free field must be last");
330 #if MAXMEMDOM >= 255
331 #error "Slab domain type insufficient"
332 #endif
333 
334 typedef struct uma_slab * uma_slab_t;
335 
336 /*
337  * On INVARIANTS builds, the slab contains a second bitset of the same size,
338  * "dbg_bits", which is laid out immediately after us_free.
339  */
340 #ifdef INVARIANTS
341 #define	SLAB_BITSETS	2
342 #else
343 #define	SLAB_BITSETS	1
344 #endif
345 
346 /* These three functions are for embedded (!OFFPAGE) use only. */
347 size_t slab_sizeof(int nitems);
348 size_t slab_space(int nitems);
349 int slab_ipers(size_t size, int align);
350 
351 /*
352  * Slab structure with a full sized bitset and hash link for both
353  * HASH and OFFPAGE zones.
354  */
355 struct uma_hash_slab {
356 	struct uma_slab		uhs_slab;	/* Must be first. */
357 	struct slabbits		uhs_bits1;	/* Must be second. */
358 #ifdef INVARIANTS
359 	struct slabbits		uhs_bits2;	/* Must be third. */
360 #endif
361 	LIST_ENTRY(uma_hash_slab) uhs_hlink;	/* Link for hash table */
362 	uint8_t			*uhs_data;	/* First item */
363 };
364 
365 typedef struct uma_hash_slab * uma_hash_slab_t;
366 
367 static inline void *
368 slab_data(uma_slab_t slab, uma_keg_t keg)
369 {
370 
371 	if ((keg->uk_flags & UMA_ZONE_OFFPAGE) == 0)
372 		return ((void *)((uintptr_t)slab - keg->uk_pgoff));
373 	else
374 		return (((uma_hash_slab_t)slab)->uhs_data);
375 }
376 
377 static inline void *
378 slab_item(uma_slab_t slab, uma_keg_t keg, int index)
379 {
380 	uintptr_t data;
381 
382 	data = (uintptr_t)slab_data(slab, keg);
383 	return ((void *)(data + keg->uk_rsize * index));
384 }
385 
386 static inline int
387 slab_item_index(uma_slab_t slab, uma_keg_t keg, void *item)
388 {
389 	uintptr_t data;
390 
391 	data = (uintptr_t)slab_data(slab, keg);
392 	return (((uintptr_t)item - data) / keg->uk_rsize);
393 }
394 #endif /* _KERNEL */
395 
396 TAILQ_HEAD(uma_bucketlist, uma_bucket);
397 
398 struct uma_zone_domain {
399 	struct uma_bucketlist uzd_buckets; /* full buckets */
400 	long		uzd_nitems;	/* total item count */
401 	long		uzd_imax;	/* maximum item count this period */
402 	long		uzd_imin;	/* minimum item count this period */
403 	long		uzd_wss;	/* working set size estimate */
404 };
405 
406 typedef struct uma_zone_domain * uma_zone_domain_t;
407 
408 /*
409  * Zone management structure
410  *
411  * TODO: Optimize for cache line size
412  *
413  */
414 struct uma_zone {
415 	/* Offset 0, used in alloc/free fast/medium fast path and const. */
416 	union {
417 		uma_keg_t	uz_keg;		/* This zone's keg */
418 		struct mtx 	*uz_lockptr;	/* To keg or to self */
419 	};
420 	struct uma_zone_domain	*uz_domain;	/* per-domain buckets */
421 	uint32_t	uz_flags;	/* Flags inherited from kegs */
422 	uint32_t	uz_size;	/* Size inherited from kegs */
423 	uma_ctor	uz_ctor;	/* Constructor for each allocation */
424 	uma_dtor	uz_dtor;	/* Destructor */
425 	uint64_t	uz_items;	/* Total items count */
426 	uint64_t	uz_max_items;	/* Maximum number of items to alloc */
427 	uint32_t	uz_sleepers;	/* Number of sleepers on memory */
428 	uint16_t	uz_bucket_size;	/* Number of items in full bucket */
429 	uint16_t	uz_bucket_size_max; /* Maximum number of bucket items */
430 
431 	/* Offset 64, used in bucket replenish. */
432 	uma_import	uz_import;	/* Import new memory to cache. */
433 	uma_release	uz_release;	/* Release memory from cache. */
434 	void		*uz_arg;	/* Import/release argument. */
435 	uma_init	uz_init;	/* Initializer for each item */
436 	uma_fini	uz_fini;	/* Finalizer for each item. */
437 	void		*uz_spare;
438 	uint64_t	uz_bkt_count;    /* Items in bucket cache */
439 	uint64_t	uz_bkt_max;	/* Maximum bucket cache size */
440 
441 	/* Offset 128 Rare. */
442 	/*
443 	 * The lock is placed here to avoid adjacent line prefetcher
444 	 * in fast paths and to take up space near infrequently accessed
445 	 * members to reduce alignment overhead.
446 	 */
447 	struct mtx	uz_lock;	/* Lock for the zone */
448 	LIST_ENTRY(uma_zone) uz_link;	/* List of all zones in keg */
449 	const char	*uz_name;	/* Text name of the zone */
450 	/* The next two fields are used to print a rate-limited warnings. */
451 	const char	*uz_warning;	/* Warning to print on failure */
452 	struct timeval	uz_ratecheck;	/* Warnings rate-limiting */
453 	struct task	uz_maxaction;	/* Task to run when at limit */
454 	uint16_t	uz_bucket_size_min; /* Min number of items in bucket */
455 
456 	/* Offset 256+, stats and misc. */
457 	counter_u64_t	uz_allocs;	/* Total number of allocations */
458 	counter_u64_t	uz_frees;	/* Total number of frees */
459 	counter_u64_t	uz_fails;	/* Total number of alloc failures */
460 	uint64_t	uz_sleeps;	/* Total number of alloc sleeps */
461 	uint64_t	uz_xdomain;	/* Total number of cross-domain frees */
462 	char		*uz_ctlname;	/* sysctl safe name string. */
463 	struct sysctl_oid *uz_oid;	/* sysctl oid pointer. */
464 	int		uz_namecnt;	/* duplicate name count. */
465 
466 	/*
467 	 * This HAS to be the last item because we adjust the zone size
468 	 * based on NCPU and then allocate the space for the zones.
469 	 */
470 	struct uma_cache	uz_cpu[]; /* Per cpu caches */
471 
472 	/* uz_domain follows here. */
473 };
474 
475 /*
476  * These flags must not overlap with the UMA_ZONE flags specified in uma.h.
477  */
478 #define	UMA_ZFLAG_CTORDTOR	0x01000000	/* Zone has ctor/dtor set. */
479 #define	UMA_ZFLAG_LIMIT		0x02000000	/* Zone has limit set. */
480 #define	UMA_ZFLAG_CACHE		0x04000000	/* uma_zcache_create()d it */
481 #define	UMA_ZFLAG_RECLAIMING	0x08000000	/* Running zone_reclaim(). */
482 #define	UMA_ZFLAG_BUCKET	0x10000000	/* Bucket zone. */
483 #define UMA_ZFLAG_INTERNAL	0x20000000	/* No offpage no PCPU. */
484 #define UMA_ZFLAG_TRASH		0x40000000	/* Add trash ctor/dtor. */
485 #define UMA_ZFLAG_CACHEONLY	0x80000000	/* Don't ask VM for buckets. */
486 
487 #define	UMA_ZFLAG_INHERIT						\
488     (UMA_ZFLAG_INTERNAL | UMA_ZFLAG_CACHEONLY | UMA_ZFLAG_BUCKET)
489 
490 #define	PRINT_UMA_ZFLAGS	"\20"	\
491     "\40CACHEONLY"			\
492     "\37TRASH"				\
493     "\36INTERNAL"			\
494     "\35BUCKET"				\
495     "\34RECLAIMING"			\
496     "\33CACHE"				\
497     "\32LIMIT"				\
498     "\31CTORDTOR"			\
499     "\22MINBUCKET"			\
500     "\21NUMA"				\
501     "\20PCPU"				\
502     "\17NODUMP"				\
503     "\16VTOSLAB"			\
504     "\15CACHESPREAD"			\
505     "\14MAXBUCKET"			\
506     "\13NOBUCKET"			\
507     "\12SECONDARY"			\
508     "\11HASH"				\
509     "\10VM"				\
510     "\7MTXCLASS"			\
511     "\6NOFREE"				\
512     "\5MALLOC"				\
513     "\4OFFPAGE"				\
514     "\3STATIC"				\
515     "\2ZINIT"				\
516     "\1PAGEABLE"
517 
518 #undef UMA_ALIGN
519 
520 #ifdef _KERNEL
521 /* Internal prototypes */
522 static __inline uma_slab_t hash_sfind(struct uma_hash *hash, uint8_t *data);
523 
524 /* Lock Macros */
525 
526 #define	KEG_LOCK_INIT(k, lc)					\
527 	do {							\
528 		if ((lc))					\
529 			mtx_init(&(k)->uk_lock, (k)->uk_name,	\
530 			    (k)->uk_name, MTX_DEF | MTX_DUPOK);	\
531 		else						\
532 			mtx_init(&(k)->uk_lock, (k)->uk_name,	\
533 			    "UMA zone", MTX_DEF | MTX_DUPOK);	\
534 	} while (0)
535 
536 #define	KEG_LOCK_FINI(k)	mtx_destroy(&(k)->uk_lock)
537 #define	KEG_LOCK(k)	mtx_lock(&(k)->uk_lock)
538 #define	KEG_UNLOCK(k)	mtx_unlock(&(k)->uk_lock)
539 #define	KEG_LOCK_ASSERT(k)	mtx_assert(&(k)->uk_lock, MA_OWNED)
540 
541 #define	KEG_GET(zone, keg) do {					\
542 	(keg) = (zone)->uz_keg;					\
543 	KASSERT((void *)(keg) != (void *)&(zone)->uz_lock,	\
544 	    ("%s: Invalid zone %p type", __func__, (zone)));	\
545 	} while (0)
546 
547 #define	ZONE_LOCK_INIT(z, lc)					\
548 	do {							\
549 		if ((lc))					\
550 			mtx_init(&(z)->uz_lock, (z)->uz_name,	\
551 			    (z)->uz_name, MTX_DEF | MTX_DUPOK);	\
552 		else						\
553 			mtx_init(&(z)->uz_lock, (z)->uz_name,	\
554 			    "UMA zone", MTX_DEF | MTX_DUPOK);	\
555 	} while (0)
556 
557 #define	ZONE_LOCK(z)	mtx_lock((z)->uz_lockptr)
558 #define	ZONE_TRYLOCK(z)	mtx_trylock((z)->uz_lockptr)
559 #define	ZONE_UNLOCK(z)	mtx_unlock((z)->uz_lockptr)
560 #define	ZONE_LOCK_FINI(z)	mtx_destroy(&(z)->uz_lock)
561 #define	ZONE_LOCK_ASSERT(z)	mtx_assert((z)->uz_lockptr, MA_OWNED)
562 
563 /*
564  * Find a slab within a hash table.  This is used for OFFPAGE zones to lookup
565  * the slab structure.
566  *
567  * Arguments:
568  *	hash  The hash table to search.
569  *	data  The base page of the item.
570  *
571  * Returns:
572  *	A pointer to a slab if successful, else NULL.
573  */
574 static __inline uma_slab_t
575 hash_sfind(struct uma_hash *hash, uint8_t *data)
576 {
577         uma_hash_slab_t slab;
578         u_int hval;
579 
580         hval = UMA_HASH(hash, data);
581 
582         LIST_FOREACH(slab, &hash->uh_slab_hash[hval], uhs_hlink) {
583                 if ((uint8_t *)slab->uhs_data == data)
584                         return (&slab->uhs_slab);
585         }
586         return (NULL);
587 }
588 
589 static __inline uma_slab_t
590 vtoslab(vm_offset_t va)
591 {
592 	vm_page_t p;
593 
594 	p = PHYS_TO_VM_PAGE(pmap_kextract(va));
595 	return (p->plinks.uma.slab);
596 }
597 
598 static __inline void
599 vtozoneslab(vm_offset_t va, uma_zone_t *zone, uma_slab_t *slab)
600 {
601 	vm_page_t p;
602 
603 	p = PHYS_TO_VM_PAGE(pmap_kextract(va));
604 	*slab = p->plinks.uma.slab;
605 	*zone = p->plinks.uma.zone;
606 }
607 
608 static __inline void
609 vsetzoneslab(vm_offset_t va, uma_zone_t zone, uma_slab_t slab)
610 {
611 	vm_page_t p;
612 
613 	p = PHYS_TO_VM_PAGE(pmap_kextract(va));
614 	p->plinks.uma.slab = slab;
615 	p->plinks.uma.zone = zone;
616 }
617 
618 extern unsigned long uma_kmem_limit;
619 extern unsigned long uma_kmem_total;
620 
621 /* Adjust bytes under management by UMA. */
622 static inline void
623 uma_total_dec(unsigned long size)
624 {
625 
626 	atomic_subtract_long(&uma_kmem_total, size);
627 }
628 
629 static inline void
630 uma_total_inc(unsigned long size)
631 {
632 
633 	if (atomic_fetchadd_long(&uma_kmem_total, size) > uma_kmem_limit)
634 		uma_reclaim_wakeup();
635 }
636 
637 /*
638  * The following two functions may be defined by architecture specific code
639  * if they can provide more efficient allocation functions.  This is useful
640  * for using direct mapped addresses.
641  */
642 void *uma_small_alloc(uma_zone_t zone, vm_size_t bytes, int domain,
643     uint8_t *pflag, int wait);
644 void uma_small_free(void *mem, vm_size_t size, uint8_t flags);
645 
646 /* Set a global soft limit on UMA managed memory. */
647 void uma_set_limit(unsigned long limit);
648 #endif /* _KERNEL */
649 
650 #endif /* VM_UMA_INT_H */
651