xref: /freebsd/sys/vm/uma_int.h (revision b0d29bc47dba79f6f38e67eabadfb4b32ffd9390)
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 /* Max size of a CACHESPREAD slab. */
143 #define	UMA_CACHESPREAD_MAX_SIZE	(128 * 1024)
144 
145 /*
146  * These flags must not overlap with the UMA_ZONE flags specified in uma.h.
147  */
148 #define	UMA_ZFLAG_OFFPAGE	0x00200000	/*
149 						 * Force the slab structure
150 						 * allocation off of the real
151 						 * memory.
152 						 */
153 #define	UMA_ZFLAG_HASH		0x00400000	/*
154 						 * Use a hash table instead of
155 						 * caching information in the
156 						 * vm_page.
157 						 */
158 #define	UMA_ZFLAG_VTOSLAB	0x00800000	/*
159 						 * Zone uses vtoslab for
160 						 * lookup.
161 						 */
162 #define	UMA_ZFLAG_CTORDTOR	0x01000000	/* Zone has ctor/dtor set. */
163 #define	UMA_ZFLAG_LIMIT		0x02000000	/* Zone has limit set. */
164 #define	UMA_ZFLAG_CACHE		0x04000000	/* uma_zcache_create()d it */
165 #define	UMA_ZFLAG_RECLAIMING	0x08000000	/* Running zone_reclaim(). */
166 #define	UMA_ZFLAG_BUCKET	0x10000000	/* Bucket zone. */
167 #define	UMA_ZFLAG_INTERNAL	0x20000000	/* No offpage no PCPU. */
168 #define	UMA_ZFLAG_TRASH		0x40000000	/* Add trash ctor/dtor. */
169 
170 #define	UMA_ZFLAG_INHERIT						\
171     (UMA_ZFLAG_OFFPAGE | UMA_ZFLAG_HASH | UMA_ZFLAG_VTOSLAB |		\
172      UMA_ZFLAG_BUCKET | UMA_ZFLAG_INTERNAL)
173 
174 #define	PRINT_UMA_ZFLAGS	"\20"	\
175     "\37TRASH"				\
176     "\36INTERNAL"			\
177     "\35BUCKET"				\
178     "\34RECLAIMING"			\
179     "\33CACHE"				\
180     "\32LIMIT"				\
181     "\31CTORDTOR"			\
182     "\30VTOSLAB"			\
183     "\27HASH"				\
184     "\26OFFPAGE"			\
185     "\23SMR"				\
186     "\22ROUNDROBIN"			\
187     "\21FIRSTTOUCH"			\
188     "\20PCPU"				\
189     "\17NODUMP"				\
190     "\16CACHESPREAD"			\
191     "\15MINBUCKET"			\
192     "\14MAXBUCKET"			\
193     "\13NOBUCKET"			\
194     "\12SECONDARY"			\
195     "\11NOTPAGE"			\
196     "\10VM"				\
197     "\7MTXCLASS"			\
198     "\6NOFREE"				\
199     "\5MALLOC"				\
200     "\4NOTOUCH"				\
201     "\3CONTIG"				\
202     "\2ZINIT"
203 
204 /*
205  * Hash table for freed address -> slab translation.
206  *
207  * Only zones with memory not touchable by the allocator use the
208  * hash table.  Otherwise slabs are found with vtoslab().
209  */
210 #define UMA_HASH_SIZE_INIT	32
211 
212 #define UMA_HASH(h, s) ((((uintptr_t)s) >> UMA_SLAB_SHIFT) & (h)->uh_hashmask)
213 
214 #define UMA_HASH_INSERT(h, s, mem)					\
215 	LIST_INSERT_HEAD(&(h)->uh_slab_hash[UMA_HASH((h),		\
216 	    (mem))], slab_tohashslab(s), uhs_hlink)
217 
218 #define UMA_HASH_REMOVE(h, s)						\
219 	LIST_REMOVE(slab_tohashslab(s), uhs_hlink)
220 
221 LIST_HEAD(slabhashhead, uma_hash_slab);
222 
223 struct uma_hash {
224 	struct slabhashhead	*uh_slab_hash;	/* Hash table for slabs */
225 	u_int		uh_hashsize;	/* Current size of the hash table */
226 	u_int		uh_hashmask;	/* Mask used during hashing */
227 };
228 
229 /*
230  * Align field or structure to cache 'sector' in intel terminology.  This
231  * is more efficient with adjacent line prefetch.
232  */
233 #if defined(__amd64__) || defined(__powerpc64__)
234 #define UMA_SUPER_ALIGN	(CACHE_LINE_SIZE * 2)
235 #else
236 #define UMA_SUPER_ALIGN	CACHE_LINE_SIZE
237 #endif
238 
239 #define	UMA_ALIGN	__aligned(UMA_SUPER_ALIGN)
240 
241 /*
242  * The uma_bucket structure is used to queue and manage buckets divorced
243  * from per-cpu caches.  They are loaded into uma_cache_bucket structures
244  * for use.
245  */
246 struct uma_bucket {
247 	STAILQ_ENTRY(uma_bucket)	ub_link; /* Link into the zone */
248 	int16_t		ub_cnt;			/* Count of items in bucket. */
249 	int16_t		ub_entries;		/* Max items. */
250 	smr_seq_t	ub_seq;			/* SMR sequence number. */
251 	void		*ub_bucket[];		/* actual allocation storage */
252 };
253 
254 typedef struct uma_bucket * uma_bucket_t;
255 
256 /*
257  * The uma_cache_bucket structure is statically allocated on each per-cpu
258  * cache.  Its use reduces branches and cache misses in the fast path.
259  */
260 struct uma_cache_bucket {
261 	uma_bucket_t	ucb_bucket;
262 	int16_t		ucb_cnt;
263 	int16_t		ucb_entries;
264 	uint32_t	ucb_spare;
265 };
266 
267 typedef struct uma_cache_bucket * uma_cache_bucket_t;
268 
269 /*
270  * The uma_cache structure is allocated for each cpu for every zone
271  * type.  This optimizes synchronization out of the allocator fast path.
272  */
273 struct uma_cache {
274 	struct uma_cache_bucket	uc_freebucket;	/* Bucket we're freeing to */
275 	struct uma_cache_bucket	uc_allocbucket;	/* Bucket to allocate from */
276 	struct uma_cache_bucket	uc_crossbucket;	/* cross domain bucket */
277 	uint64_t		uc_allocs;	/* Count of allocations */
278 	uint64_t		uc_frees;	/* Count of frees */
279 } UMA_ALIGN;
280 
281 typedef struct uma_cache * uma_cache_t;
282 
283 LIST_HEAD(slabhead, uma_slab);
284 
285 /*
286  * The cache structure pads perfectly into 64 bytes so we use spare
287  * bits from the embedded cache buckets to store information from the zone
288  * and keep all fast-path allocations accessing a single per-cpu line.
289  */
290 static inline void
291 cache_set_uz_flags(uma_cache_t cache, uint32_t flags)
292 {
293 
294 	cache->uc_freebucket.ucb_spare = flags;
295 }
296 
297 static inline void
298 cache_set_uz_size(uma_cache_t cache, uint32_t size)
299 {
300 
301 	cache->uc_allocbucket.ucb_spare = size;
302 }
303 
304 static inline uint32_t
305 cache_uz_flags(uma_cache_t cache)
306 {
307 
308 	return (cache->uc_freebucket.ucb_spare);
309 }
310 
311 static inline uint32_t
312 cache_uz_size(uma_cache_t cache)
313 {
314 
315 	return (cache->uc_allocbucket.ucb_spare);
316 }
317 
318 /*
319  * Per-domain slab lists.  Embedded in the kegs.
320  */
321 struct uma_domain {
322 	struct mtx_padalign ud_lock;	/* Lock for the domain lists. */
323 	struct slabhead	ud_part_slab;	/* partially allocated slabs */
324 	struct slabhead	ud_free_slab;	/* completely unallocated slabs */
325 	struct slabhead ud_full_slab;	/* fully allocated slabs */
326 	uint32_t	ud_pages;	/* Total page count */
327 	uint32_t	ud_free_items;	/* Count of items free in all slabs */
328 	uint32_t	ud_free_slabs;	/* Count of free slabs */
329 } __aligned(CACHE_LINE_SIZE);
330 
331 typedef struct uma_domain * uma_domain_t;
332 
333 /*
334  * Keg management structure
335  *
336  * TODO: Optimize for cache line size
337  *
338  */
339 struct uma_keg {
340 	struct uma_hash	uk_hash;
341 	LIST_HEAD(,uma_zone)	uk_zones;	/* Keg's zones */
342 
343 	struct domainset_ref uk_dr;	/* Domain selection policy. */
344 	uint32_t	uk_align;	/* Alignment mask */
345 	uint32_t	uk_reserve;	/* Number of reserved items. */
346 	uint32_t	uk_size;	/* Requested size of each item */
347 	uint32_t	uk_rsize;	/* Real size of each item */
348 
349 	uma_init	uk_init;	/* Keg's init routine */
350 	uma_fini	uk_fini;	/* Keg's fini routine */
351 	uma_alloc	uk_allocf;	/* Allocation function */
352 	uma_free	uk_freef;	/* Free routine */
353 
354 	u_long		uk_offset;	/* Next free offset from base KVA */
355 	vm_offset_t	uk_kva;		/* Zone base KVA */
356 
357 	uint32_t	uk_pgoff;	/* Offset to uma_slab struct */
358 	uint16_t	uk_ppera;	/* pages per allocation from backend */
359 	uint16_t	uk_ipers;	/* Items per slab */
360 	uint32_t	uk_flags;	/* Internal flags */
361 
362 	/* Least used fields go to the last cache line. */
363 	const char	*uk_name;		/* Name of creating zone. */
364 	LIST_ENTRY(uma_keg)	uk_link;	/* List of all kegs */
365 
366 	/* Must be last, variable sized. */
367 	struct uma_domain	uk_domain[];	/* Keg's slab lists. */
368 };
369 typedef struct uma_keg	* uma_keg_t;
370 
371 /*
372  * Free bits per-slab.
373  */
374 #define	SLAB_MAX_SETSIZE	(PAGE_SIZE / UMA_SMALLEST_UNIT)
375 #define	SLAB_MIN_SETSIZE	_BITSET_BITS
376 BITSET_DEFINE(noslabbits, 0);
377 
378 /*
379  * The slab structure manages a single contiguous allocation from backing
380  * store and subdivides it into individually allocatable items.
381  */
382 struct uma_slab {
383 	LIST_ENTRY(uma_slab)	us_link;	/* slabs in zone */
384 	uint16_t	us_freecount;		/* How many are free? */
385 	uint8_t		us_flags;		/* Page flags see uma.h */
386 	uint8_t		us_domain;		/* Backing NUMA domain. */
387 	struct noslabbits us_free;		/* Free bitmask, flexible. */
388 };
389 _Static_assert(sizeof(struct uma_slab) == __offsetof(struct uma_slab, us_free),
390     "us_free field must be last");
391 _Static_assert(MAXMEMDOM < 255,
392     "us_domain field is not wide enough");
393 
394 typedef struct uma_slab * uma_slab_t;
395 
396 /*
397  * Slab structure with a full sized bitset and hash link for both
398  * HASH and OFFPAGE zones.
399  */
400 struct uma_hash_slab {
401 	LIST_ENTRY(uma_hash_slab) uhs_hlink;	/* Link for hash table */
402 	uint8_t			*uhs_data;	/* First item */
403 	struct uma_slab		uhs_slab;	/* Must be last. */
404 };
405 
406 typedef struct uma_hash_slab * uma_hash_slab_t;
407 
408 static inline uma_hash_slab_t
409 slab_tohashslab(uma_slab_t slab)
410 {
411 
412 	return (__containerof(slab, struct uma_hash_slab, uhs_slab));
413 }
414 
415 static inline void *
416 slab_data(uma_slab_t slab, uma_keg_t keg)
417 {
418 
419 	if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) == 0)
420 		return ((void *)((uintptr_t)slab - keg->uk_pgoff));
421 	else
422 		return (slab_tohashslab(slab)->uhs_data);
423 }
424 
425 static inline void *
426 slab_item(uma_slab_t slab, uma_keg_t keg, int index)
427 {
428 	uintptr_t data;
429 
430 	data = (uintptr_t)slab_data(slab, keg);
431 	return ((void *)(data + keg->uk_rsize * index));
432 }
433 
434 static inline int
435 slab_item_index(uma_slab_t slab, uma_keg_t keg, void *item)
436 {
437 	uintptr_t data;
438 
439 	data = (uintptr_t)slab_data(slab, keg);
440 	return (((uintptr_t)item - data) / keg->uk_rsize);
441 }
442 
443 STAILQ_HEAD(uma_bucketlist, uma_bucket);
444 
445 struct uma_zone_domain {
446 	struct uma_bucketlist uzd_buckets; /* full buckets */
447 	uma_bucket_t	uzd_cross;	/* Fills from cross buckets. */
448 	long		uzd_nitems;	/* total item count */
449 	long		uzd_imax;	/* maximum item count this period */
450 	long		uzd_imin;	/* minimum item count this period */
451 	long		uzd_wss;	/* working set size estimate */
452 	smr_seq_t	uzd_seq;	/* Lowest queued seq. */
453 	struct mtx	uzd_lock;	/* Lock for the domain */
454 } __aligned(CACHE_LINE_SIZE);
455 
456 typedef struct uma_zone_domain * uma_zone_domain_t;
457 
458 /*
459  * Zone structure - per memory type.
460  */
461 struct uma_zone {
462 	/* Offset 0, used in alloc/free fast/medium fast path and const. */
463 	uint32_t	uz_flags;	/* Flags inherited from kegs */
464 	uint32_t	uz_size;	/* Size inherited from kegs */
465 	uma_ctor	uz_ctor;	/* Constructor for each allocation */
466 	uma_dtor	uz_dtor;	/* Destructor */
467 	smr_t		uz_smr;		/* Safe memory reclaim context. */
468 	uint64_t	uz_max_items;	/* Maximum number of items to alloc */
469 	uint64_t	uz_bucket_max;	/* Maximum bucket cache size */
470 	uint16_t	uz_bucket_size;	/* Number of items in full bucket */
471 	uint16_t	uz_bucket_size_max; /* Maximum number of bucket items */
472 	uint32_t	uz_sleepers;	/* Threads sleeping on limit */
473 	counter_u64_t	uz_xdomain;	/* Total number of cross-domain frees */
474 
475 	/* Offset 64, used in bucket replenish. */
476 	uma_keg_t	uz_keg;		/* This zone's keg if !CACHE */
477 	uma_import	uz_import;	/* Import new memory to cache. */
478 	uma_release	uz_release;	/* Release memory from cache. */
479 	void		*uz_arg;	/* Import/release argument. */
480 	uma_init	uz_init;	/* Initializer for each item */
481 	uma_fini	uz_fini;	/* Finalizer for each item. */
482 	volatile uint64_t uz_items;	/* Total items count & sleepers */
483 	uint64_t	uz_sleeps;	/* Total number of alloc sleeps */
484 
485 	/* Offset 128 Rare stats, misc read-only. */
486 	LIST_ENTRY(uma_zone) uz_link;	/* List of all zones in keg */
487 	counter_u64_t	uz_allocs;	/* Total number of allocations */
488 	counter_u64_t	uz_frees;	/* Total number of frees */
489 	counter_u64_t	uz_fails;	/* Total number of alloc failures */
490 	const char	*uz_name;	/* Text name of the zone */
491 	char		*uz_ctlname;	/* sysctl safe name string. */
492 	int		uz_namecnt;	/* duplicate name count. */
493 	uint16_t	uz_bucket_size_min; /* Min number of items in bucket */
494 	uint16_t	uz_pad0;
495 
496 	/* Offset 192, rare read-only. */
497 	struct sysctl_oid *uz_oid;	/* sysctl oid pointer. */
498 	const char	*uz_warning;	/* Warning to print on failure */
499 	struct timeval	uz_ratecheck;	/* Warnings rate-limiting */
500 	struct task	uz_maxaction;	/* Task to run when at limit */
501 
502 	/* Offset 256. */
503 	struct mtx	uz_cross_lock;	/* Cross domain free lock */
504 
505 	/*
506 	 * This HAS to be the last item because we adjust the zone size
507 	 * based on NCPU and then allocate the space for the zones.
508 	 */
509 	struct uma_cache	uz_cpu[]; /* Per cpu caches */
510 
511 	/* domains follow here. */
512 };
513 
514 /*
515  * Macros for interpreting the uz_items field.  20 bits of sleeper count
516  * and 44 bit of item count.
517  */
518 #define	UZ_ITEMS_SLEEPER_SHIFT	44LL
519 #define	UZ_ITEMS_SLEEPERS_MAX	((1 << (64 - UZ_ITEMS_SLEEPER_SHIFT)) - 1)
520 #define	UZ_ITEMS_COUNT_MASK	((1LL << UZ_ITEMS_SLEEPER_SHIFT) - 1)
521 #define	UZ_ITEMS_COUNT(x)	((x) & UZ_ITEMS_COUNT_MASK)
522 #define	UZ_ITEMS_SLEEPERS(x)	((x) >> UZ_ITEMS_SLEEPER_SHIFT)
523 #define	UZ_ITEMS_SLEEPER	(1LL << UZ_ITEMS_SLEEPER_SHIFT)
524 
525 #define	ZONE_ASSERT_COLD(z)						\
526 	KASSERT(uma_zone_get_allocs((z)) == 0,				\
527 	    ("zone %s initialization after use.", (z)->uz_name))
528 
529 #undef	UMA_ALIGN
530 
531 #ifdef _KERNEL
532 /* Internal prototypes */
533 static __inline uma_slab_t hash_sfind(struct uma_hash *hash, uint8_t *data);
534 
535 /* Lock Macros */
536 
537 #define	KEG_LOCKPTR(k, d)	(struct mtx *)&(k)->uk_domain[(d)].ud_lock
538 #define	KEG_LOCK_INIT(k, d, lc)						\
539 	do {								\
540 		if ((lc))						\
541 			mtx_init(KEG_LOCKPTR(k, d), (k)->uk_name,	\
542 			    (k)->uk_name, MTX_DEF | MTX_DUPOK);		\
543 		else							\
544 			mtx_init(KEG_LOCKPTR(k, d), (k)->uk_name,	\
545 			    "UMA zone", MTX_DEF | MTX_DUPOK);		\
546 	} while (0)
547 
548 #define	KEG_LOCK_FINI(k, d)	mtx_destroy(KEG_LOCKPTR(k, d))
549 #define	KEG_LOCK(k, d)							\
550 	({ mtx_lock(KEG_LOCKPTR(k, d)); KEG_LOCKPTR(k, d); })
551 #define	KEG_UNLOCK(k, d)	mtx_unlock(KEG_LOCKPTR(k, d))
552 #define	KEG_LOCK_ASSERT(k, d)	mtx_assert(KEG_LOCKPTR(k, d), MA_OWNED)
553 
554 #define	KEG_GET(zone, keg) do {					\
555 	(keg) = (zone)->uz_keg;					\
556 	KASSERT((void *)(keg) != NULL,				\
557 	    ("%s: Invalid zone %p type", __func__, (zone)));	\
558 	} while (0)
559 
560 #define	KEG_ASSERT_COLD(k)						\
561 	KASSERT(uma_keg_get_allocs((k)) == 0,				\
562 	    ("keg %s initialization after use.", (k)->uk_name))
563 
564 /* Domains are contiguous after the last CPU */
565 #define	ZDOM_GET(z, n)							\
566     (&((uma_zone_domain_t)&(z)->uz_cpu[mp_maxid + 1])[n])
567 
568 #define	ZDOM_LOCK_INIT(z, zdom, lc)					\
569 	do {								\
570 		if ((lc))						\
571 			mtx_init(&(zdom)->uzd_lock, (z)->uz_name,	\
572 			    (z)->uz_name, MTX_DEF | MTX_DUPOK);		\
573 		else							\
574 			mtx_init(&(zdom)->uzd_lock, (z)->uz_name,	\
575 			    "UMA zone", MTX_DEF | MTX_DUPOK);		\
576 	} while (0)
577 #define	ZDOM_LOCK_FINI(z)	mtx_destroy(&(z)->uzd_lock)
578 #define	ZDOM_LOCK_ASSERT(z)	mtx_assert(&(z)->uzd_lock, MA_OWNED)
579 
580 #define	ZDOM_LOCK(z)	mtx_lock(&(z)->uzd_lock)
581 #define	ZDOM_OWNED(z)	(mtx_owner(&(z)->uzd_lock) != NULL)
582 #define	ZDOM_UNLOCK(z)	mtx_unlock(&(z)->uzd_lock)
583 
584 #define	ZONE_LOCK(z)	ZDOM_LOCK(ZDOM_GET((z), 0))
585 #define	ZONE_UNLOCK(z)	ZDOM_UNLOCK(ZDOM_GET((z), 0))
586 
587 #define	ZONE_CROSS_LOCK_INIT(z)					\
588 	mtx_init(&(z)->uz_cross_lock, "UMA Cross", NULL, MTX_DEF)
589 #define	ZONE_CROSS_LOCK(z)	mtx_lock(&(z)->uz_cross_lock)
590 #define	ZONE_CROSS_UNLOCK(z)	mtx_unlock(&(z)->uz_cross_lock)
591 #define	ZONE_CROSS_LOCK_FINI(z)	mtx_destroy(&(z)->uz_cross_lock)
592 
593 /*
594  * Find a slab within a hash table.  This is used for OFFPAGE zones to lookup
595  * the slab structure.
596  *
597  * Arguments:
598  *	hash  The hash table to search.
599  *	data  The base page of the item.
600  *
601  * Returns:
602  *	A pointer to a slab if successful, else NULL.
603  */
604 static __inline uma_slab_t
605 hash_sfind(struct uma_hash *hash, uint8_t *data)
606 {
607         uma_hash_slab_t slab;
608         u_int hval;
609 
610         hval = UMA_HASH(hash, data);
611 
612         LIST_FOREACH(slab, &hash->uh_slab_hash[hval], uhs_hlink) {
613                 if ((uint8_t *)slab->uhs_data == data)
614                         return (&slab->uhs_slab);
615         }
616         return (NULL);
617 }
618 
619 static __inline uma_slab_t
620 vtoslab(vm_offset_t va)
621 {
622 	vm_page_t p;
623 
624 	p = PHYS_TO_VM_PAGE(pmap_kextract(va));
625 	return (p->plinks.uma.slab);
626 }
627 
628 static __inline void
629 vtozoneslab(vm_offset_t va, uma_zone_t *zone, uma_slab_t *slab)
630 {
631 	vm_page_t p;
632 
633 	p = PHYS_TO_VM_PAGE(pmap_kextract(va));
634 	*slab = p->plinks.uma.slab;
635 	*zone = p->plinks.uma.zone;
636 }
637 
638 static __inline void
639 vsetzoneslab(vm_offset_t va, uma_zone_t zone, uma_slab_t slab)
640 {
641 	vm_page_t p;
642 
643 	p = PHYS_TO_VM_PAGE(pmap_kextract(va));
644 	p->plinks.uma.slab = slab;
645 	p->plinks.uma.zone = zone;
646 }
647 
648 extern unsigned long uma_kmem_limit;
649 extern unsigned long uma_kmem_total;
650 
651 /* Adjust bytes under management by UMA. */
652 static inline void
653 uma_total_dec(unsigned long size)
654 {
655 
656 	atomic_subtract_long(&uma_kmem_total, size);
657 }
658 
659 static inline void
660 uma_total_inc(unsigned long size)
661 {
662 
663 	if (atomic_fetchadd_long(&uma_kmem_total, size) > uma_kmem_limit)
664 		uma_reclaim_wakeup();
665 }
666 
667 /*
668  * The following two functions may be defined by architecture specific code
669  * if they can provide more efficient allocation functions.  This is useful
670  * for using direct mapped addresses.
671  */
672 void *uma_small_alloc(uma_zone_t zone, vm_size_t bytes, int domain,
673     uint8_t *pflag, int wait);
674 void uma_small_free(void *mem, vm_size_t size, uint8_t flags);
675 
676 /* Set a global soft limit on UMA managed memory. */
677 void uma_set_limit(unsigned long limit);
678 
679 #endif /* _KERNEL */
680 
681 #endif /* VM_UMA_INT_H */
682