xref: /freebsd/sys/vm/uma_int.h (revision 1f4bcc459a76b7aa664f3fd557684cd0ba6da352)
1 /*-
2  * Copyright (c) 2002-2005, 2009, 2013 Jeffrey Roberson <jeff@FreeBSD.org>
3  * Copyright (c) 2004, 2005 Bosko Milekic <bmilekic@FreeBSD.org>
4  * All rights reserved.
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice unmodified, this list of conditions, and the following
11  *    disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
17  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
18  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
19  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
20  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
21  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
22  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
23  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
24  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
25  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
26  *
27  * $FreeBSD$
28  *
29  */
30 
31 /*
32  * This file includes definitions, structures, prototypes, and inlines that
33  * should not be used outside of the actual implementation of UMA.
34  */
35 
36 /*
37  * Here's a quick description of the relationship between the objects:
38  *
39  * Kegs contain lists of slabs which are stored in either the full bin, empty
40  * bin, or partially allocated bin, to reduce fragmentation.  They also contain
41  * the user supplied value for size, which is adjusted for alignment purposes
42  * and rsize is the result of that.  The Keg also stores information for
43  * managing a hash of page addresses that maps pages to uma_slab_t structures
44  * for pages that don't have embedded uma_slab_t's.
45  *
46  * The uma_slab_t may be embedded in a UMA_SLAB_SIZE chunk of memory or it may
47  * be allocated off the page from a special slab zone.  The free list within a
48  * slab is managed with a bitmask.  For item sizes that would yield more than
49  * 10% memory waste we potentially allocate a separate uma_slab_t if this will
50  * improve the number of items per slab that will fit.
51  *
52  * The only really gross cases, with regards to memory waste, are for those
53  * items that are just over half the page size.   You can get nearly 50% waste,
54  * so you fall back to the memory footprint of the power of two allocator. I
55  * have looked at memory allocation sizes on many of the machines available to
56  * me, and there does not seem to be an abundance of allocations at this range
57  * so at this time it may not make sense to optimize for it.  This can, of
58  * course, be solved with dynamic slab sizes.
59  *
60  * Kegs may serve multiple Zones but by far most of the time they only serve
61  * one.  When a Zone is created, a Keg is allocated and setup for it.  While
62  * the backing Keg stores slabs, the Zone caches Buckets of items allocated
63  * from the slabs.  Each Zone is equipped with an init/fini and ctor/dtor
64  * pair, as well as with its own set of small per-CPU caches, layered above
65  * the Zone's general Bucket cache.
66  *
67  * The PCPU caches are protected by critical sections, and may be accessed
68  * safely only from their associated CPU, while the Zones backed by the same
69  * Keg all share a common Keg lock (to coalesce contention on the backing
70  * slabs).  The backing Keg typically only serves one Zone but in the case of
71  * multiple Zones, one of the Zones is considered the Master Zone and all
72  * Zone-related stats from the Keg are done in the Master Zone.  For an
73  * example of a Multi-Zone setup, refer to the Mbuf allocation code.
74  */
75 
76 /*
77  *	This is the representation for normal (Non OFFPAGE slab)
78  *
79  *	i == item
80  *	s == slab pointer
81  *
82  *	<----------------  Page (UMA_SLAB_SIZE) ------------------>
83  *	___________________________________________________________
84  *     | _  _  _  _  _  _  _  _  _  _  _  _  _  _  _   ___________ |
85  *     ||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i| |slab header||
86  *     ||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_| |___________||
87  *     |___________________________________________________________|
88  *
89  *
90  *	This is an OFFPAGE slab. These can be larger than UMA_SLAB_SIZE.
91  *
92  *	___________________________________________________________
93  *     | _  _  _  _  _  _  _  _  _  _  _  _  _  _  _  _  _  _  _   |
94  *     ||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i|  |
95  *     ||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_|  |
96  *     |___________________________________________________________|
97  *       ___________    ^
98  *	|slab header|   |
99  *	|___________|---*
100  *
101  */
102 
103 #ifndef VM_UMA_INT_H
104 #define VM_UMA_INT_H
105 
106 #define UMA_SLAB_SIZE	PAGE_SIZE	/* How big are our slabs? */
107 #define UMA_SLAB_MASK	(PAGE_SIZE - 1)	/* Mask to get back to the page */
108 #define UMA_SLAB_SHIFT	PAGE_SHIFT	/* Number of bits PAGE_MASK */
109 
110 #define UMA_BOOT_PAGES		64	/* Pages allocated for startup */
111 
112 /* Max waste percentage before going to off page slab management */
113 #define UMA_MAX_WASTE	10
114 
115 /*
116  * I doubt there will be many cases where this is exceeded. This is the initial
117  * size of the hash table for uma_slabs that are managed off page. This hash
118  * does expand by powers of two.  Currently it doesn't get smaller.
119  */
120 #define UMA_HASH_SIZE_INIT	32
121 
122 /*
123  * I should investigate other hashing algorithms.  This should yield a low
124  * number of collisions if the pages are relatively contiguous.
125  */
126 
127 #define UMA_HASH(h, s) ((((uintptr_t)s) >> UMA_SLAB_SHIFT) & (h)->uh_hashmask)
128 
129 #define UMA_HASH_INSERT(h, s, mem)					\
130 		SLIST_INSERT_HEAD(&(h)->uh_slab_hash[UMA_HASH((h),	\
131 		    (mem))], (s), us_hlink)
132 #define UMA_HASH_REMOVE(h, s, mem)					\
133 		SLIST_REMOVE(&(h)->uh_slab_hash[UMA_HASH((h),		\
134 		    (mem))], (s), uma_slab, us_hlink)
135 
136 /* Hash table for freed address -> slab translation */
137 
138 SLIST_HEAD(slabhead, uma_slab);
139 
140 struct uma_hash {
141 	struct slabhead	*uh_slab_hash;	/* Hash table for slabs */
142 	int		uh_hashsize;	/* Current size of the hash table */
143 	int		uh_hashmask;	/* Mask used during hashing */
144 };
145 
146 /*
147  * align field or structure to cache line
148  */
149 #if defined(__amd64__)
150 #define UMA_ALIGN	__aligned(CACHE_LINE_SIZE)
151 #else
152 #define UMA_ALIGN
153 #endif
154 
155 /*
156  * Structures for per cpu queues.
157  */
158 
159 struct uma_bucket {
160 	LIST_ENTRY(uma_bucket)	ub_link;	/* Link into the zone */
161 	int16_t	ub_cnt;				/* Count of free items. */
162 	int16_t	ub_entries;			/* Max items. */
163 	void	*ub_bucket[];			/* actual allocation storage */
164 };
165 
166 typedef struct uma_bucket * uma_bucket_t;
167 
168 struct uma_cache {
169 	uma_bucket_t	uc_freebucket;	/* Bucket we're freeing to */
170 	uma_bucket_t	uc_allocbucket;	/* Bucket to allocate from */
171 	uint64_t	uc_allocs;	/* Count of allocations */
172 	uint64_t	uc_frees;	/* Count of frees */
173 } UMA_ALIGN;
174 
175 typedef struct uma_cache * uma_cache_t;
176 
177 /*
178  * Keg management structure
179  *
180  * TODO: Optimize for cache line size
181  *
182  */
183 struct uma_keg {
184 	struct mtx_padalign	uk_lock;	/* Lock for the keg */
185 	struct uma_hash	uk_hash;
186 
187 	LIST_HEAD(,uma_zone)	uk_zones;	/* Keg's zones */
188 	LIST_HEAD(,uma_slab)	uk_part_slab;	/* partially allocated slabs */
189 	LIST_HEAD(,uma_slab)	uk_free_slab;	/* empty slab list */
190 	LIST_HEAD(,uma_slab)	uk_full_slab;	/* full slabs */
191 
192 	uint32_t	uk_align;	/* Alignment mask */
193 	uint32_t	uk_pages;	/* Total page count */
194 	uint32_t	uk_free;	/* Count of items free in slabs */
195 	uint32_t	uk_reserve;	/* Number of reserved items. */
196 	uint32_t	uk_size;	/* Requested size of each item */
197 	uint32_t	uk_rsize;	/* Real size of each item */
198 	uint32_t	uk_maxpages;	/* Maximum number of pages to alloc */
199 
200 	uma_init	uk_init;	/* Keg's init routine */
201 	uma_fini	uk_fini;	/* Keg's fini routine */
202 	uma_alloc	uk_allocf;	/* Allocation function */
203 	uma_free	uk_freef;	/* Free routine */
204 
205 	u_long		uk_offset;	/* Next free offset from base KVA */
206 	vm_offset_t	uk_kva;		/* Zone base KVA */
207 	uma_zone_t	uk_slabzone;	/* Slab zone backing us, if OFFPAGE */
208 
209 	uint16_t	uk_slabsize;	/* Slab size for this keg */
210 	uint16_t	uk_pgoff;	/* Offset to uma_slab struct */
211 	uint16_t	uk_ppera;	/* pages per allocation from backend */
212 	uint16_t	uk_ipers;	/* Items per slab */
213 	uint32_t	uk_flags;	/* Internal flags */
214 
215 	/* Least used fields go to the last cache line. */
216 	const char	*uk_name;		/* Name of creating zone. */
217 	LIST_ENTRY(uma_keg)	uk_link;	/* List of all kegs */
218 };
219 typedef struct uma_keg	* uma_keg_t;
220 
221 /*
222  * Free bits per-slab.
223  */
224 #define	SLAB_SETSIZE	(PAGE_SIZE / UMA_SMALLEST_UNIT)
225 BITSET_DEFINE(slabbits, SLAB_SETSIZE);
226 
227 /*
228  * The slab structure manages a single contiguous allocation from backing
229  * store and subdivides it into individually allocatable items.
230  */
231 struct uma_slab {
232 	uma_keg_t	us_keg;			/* Keg we live in */
233 	union {
234 		LIST_ENTRY(uma_slab)	_us_link;	/* slabs in zone */
235 		unsigned long	_us_size;	/* Size of allocation */
236 	} us_type;
237 	SLIST_ENTRY(uma_slab)	us_hlink;	/* Link for hash table */
238 	uint8_t		*us_data;		/* First item */
239 	struct slabbits	us_free;		/* Free bitmask. */
240 #ifdef INVARIANTS
241 	struct slabbits	us_debugfree;		/* Debug bitmask. */
242 #endif
243 	uint16_t	us_freecount;		/* How many are free? */
244 	uint8_t		us_flags;		/* Page flags see uma.h */
245 	uint8_t		us_pad;			/* Pad to 32bits, unused. */
246 };
247 
248 #define	us_link	us_type._us_link
249 #define	us_size	us_type._us_size
250 
251 /*
252  * The slab structure for UMA_ZONE_REFCNT zones for whose items we
253  * maintain reference counters in the slab for.
254  */
255 struct uma_slab_refcnt {
256 	struct uma_slab		us_head;	/* slab header data */
257 	uint32_t		us_refcnt[0];	/* Actually larger. */
258 };
259 
260 typedef struct uma_slab * uma_slab_t;
261 typedef struct uma_slab_refcnt * uma_slabrefcnt_t;
262 typedef uma_slab_t (*uma_slaballoc)(uma_zone_t, uma_keg_t, int);
263 
264 struct uma_klink {
265 	LIST_ENTRY(uma_klink)	kl_link;
266 	uma_keg_t		kl_keg;
267 };
268 typedef struct uma_klink *uma_klink_t;
269 
270 /*
271  * Zone management structure
272  *
273  * TODO: Optimize for cache line size
274  *
275  */
276 struct uma_zone {
277 	struct mtx_padalign	uz_lock;	/* Lock for the zone */
278 	struct mtx_padalign	*uz_lockptr;
279 	const char		*uz_name;	/* Text name of the zone */
280 
281 	LIST_ENTRY(uma_zone)	uz_link;	/* List of all zones in keg */
282 	LIST_HEAD(,uma_bucket)	uz_buckets;	/* full buckets */
283 
284 	LIST_HEAD(,uma_klink)	uz_kegs;	/* List of kegs. */
285 	struct uma_klink	uz_klink;	/* klink for first keg. */
286 
287 	uma_slaballoc	uz_slab;	/* Allocate a slab from the backend. */
288 	uma_ctor	uz_ctor;	/* Constructor for each allocation */
289 	uma_dtor	uz_dtor;	/* Destructor */
290 	uma_init	uz_init;	/* Initializer for each item */
291 	uma_fini	uz_fini;	/* Finalizer for each item. */
292 	uma_import	uz_import;	/* Import new memory to cache. */
293 	uma_release	uz_release;	/* Release memory from cache. */
294 	void		*uz_arg;	/* Import/release argument. */
295 
296 	uint32_t	uz_flags;	/* Flags inherited from kegs */
297 	uint32_t	uz_size;	/* Size inherited from kegs */
298 
299 	volatile u_long	uz_allocs UMA_ALIGN; /* Total number of allocations */
300 	volatile u_long	uz_fails;	/* Total number of alloc failures */
301 	volatile u_long	uz_frees;	/* Total number of frees */
302 	uint64_t	uz_sleeps;	/* Total number of alloc sleeps */
303 	uint16_t	uz_count;	/* Amount of items in full bucket */
304 	uint16_t	uz_count_min;	/* Minimal amount of items there */
305 
306 	/* The next two fields are used to print a rate-limited warnings. */
307 	const char	*uz_warning;	/* Warning to print on failure */
308 	struct timeval	uz_ratecheck;	/* Warnings rate-limiting */
309 
310 	uma_maxaction_t	uz_maxaction;	/* Function to run when at limit */
311 
312 	/*
313 	 * This HAS to be the last item because we adjust the zone size
314 	 * based on NCPU and then allocate the space for the zones.
315 	 */
316 	struct uma_cache	uz_cpu[1]; /* Per cpu caches */
317 };
318 
319 /*
320  * These flags must not overlap with the UMA_ZONE flags specified in uma.h.
321  */
322 #define	UMA_ZFLAG_MULTI		0x04000000	/* Multiple kegs in the zone. */
323 #define	UMA_ZFLAG_DRAINING	0x08000000	/* Running zone_drain. */
324 #define	UMA_ZFLAG_BUCKET	0x10000000	/* Bucket zone. */
325 #define UMA_ZFLAG_INTERNAL	0x20000000	/* No offpage no PCPU. */
326 #define UMA_ZFLAG_FULL		0x40000000	/* Reached uz_maxpages */
327 #define UMA_ZFLAG_CACHEONLY	0x80000000	/* Don't ask VM for buckets. */
328 
329 #define	UMA_ZFLAG_INHERIT						\
330     (UMA_ZFLAG_INTERNAL | UMA_ZFLAG_CACHEONLY | UMA_ZFLAG_BUCKET)
331 
332 static inline uma_keg_t
333 zone_first_keg(uma_zone_t zone)
334 {
335 	uma_klink_t klink;
336 
337 	klink = LIST_FIRST(&zone->uz_kegs);
338 	return (klink != NULL) ? klink->kl_keg : NULL;
339 }
340 
341 #undef UMA_ALIGN
342 
343 #ifdef _KERNEL
344 /* Internal prototypes */
345 static __inline uma_slab_t hash_sfind(struct uma_hash *hash, uint8_t *data);
346 void *uma_large_malloc(vm_size_t size, int wait);
347 void uma_large_free(uma_slab_t slab);
348 
349 /* Lock Macros */
350 
351 #define	KEG_LOCK_INIT(k, lc)					\
352 	do {							\
353 		if ((lc))					\
354 			mtx_init(&(k)->uk_lock, (k)->uk_name,	\
355 			    (k)->uk_name, MTX_DEF | MTX_DUPOK);	\
356 		else						\
357 			mtx_init(&(k)->uk_lock, (k)->uk_name,	\
358 			    "UMA zone", MTX_DEF | MTX_DUPOK);	\
359 	} while (0)
360 
361 #define	KEG_LOCK_FINI(k)	mtx_destroy(&(k)->uk_lock)
362 #define	KEG_LOCK(k)	mtx_lock(&(k)->uk_lock)
363 #define	KEG_UNLOCK(k)	mtx_unlock(&(k)->uk_lock)
364 
365 #define	ZONE_LOCK_INIT(z, lc)					\
366 	do {							\
367 		if ((lc))					\
368 			mtx_init(&(z)->uz_lock, (z)->uz_name,	\
369 			    (z)->uz_name, MTX_DEF | MTX_DUPOK);	\
370 		else						\
371 			mtx_init(&(z)->uz_lock, (z)->uz_name,	\
372 			    "UMA zone", MTX_DEF | MTX_DUPOK);	\
373 	} while (0)
374 
375 #define	ZONE_LOCK(z)	mtx_lock((z)->uz_lockptr)
376 #define	ZONE_TRYLOCK(z)	mtx_trylock((z)->uz_lockptr)
377 #define	ZONE_UNLOCK(z)	mtx_unlock((z)->uz_lockptr)
378 #define	ZONE_LOCK_FINI(z)	mtx_destroy(&(z)->uz_lock)
379 
380 /*
381  * Find a slab within a hash table.  This is used for OFFPAGE zones to lookup
382  * the slab structure.
383  *
384  * Arguments:
385  *	hash  The hash table to search.
386  *	data  The base page of the item.
387  *
388  * Returns:
389  *	A pointer to a slab if successful, else NULL.
390  */
391 static __inline uma_slab_t
392 hash_sfind(struct uma_hash *hash, uint8_t *data)
393 {
394         uma_slab_t slab;
395         int hval;
396 
397         hval = UMA_HASH(hash, data);
398 
399         SLIST_FOREACH(slab, &hash->uh_slab_hash[hval], us_hlink) {
400                 if ((uint8_t *)slab->us_data == data)
401                         return (slab);
402         }
403         return (NULL);
404 }
405 
406 static __inline uma_slab_t
407 vtoslab(vm_offset_t va)
408 {
409 	vm_page_t p;
410 
411 	p = PHYS_TO_VM_PAGE(pmap_kextract(va));
412 	return ((uma_slab_t)p->plinks.s.pv);
413 }
414 
415 static __inline void
416 vsetslab(vm_offset_t va, uma_slab_t slab)
417 {
418 	vm_page_t p;
419 
420 	p = PHYS_TO_VM_PAGE(pmap_kextract(va));
421 	p->plinks.s.pv = slab;
422 }
423 
424 /*
425  * The following two functions may be defined by architecture specific code
426  * if they can provide more effecient allocation functions.  This is useful
427  * for using direct mapped addresses.
428  */
429 void *uma_small_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *pflag,
430     int wait);
431 void uma_small_free(void *mem, vm_size_t size, uint8_t flags);
432 #endif /* _KERNEL */
433 
434 #endif /* VM_UMA_INT_H */
435