xref: /freebsd/sys/vm/uma_int.h (revision 2357939bc239bd5334a169b62313806178dd8f30)
1 /*
2  * Copyright (c) 2002, Jeffrey Roberson <jeff@freebsd.org>
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice unmodified, this list of conditions, and the following
10  *    disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
16  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
17  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
18  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
19  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
20  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
21  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
22  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
23  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
24  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
25  *
26  * $FreeBSD$
27  *
28  */
29 
30 /*
31  * This file includes definitions, structures, prototypes, and inlines that
32  * should not be used outside of the actual implementation of UMA.
33  */
34 
35 /*
36  * Here's a quick description of the relationship between the objects:
37  *
38  * Zones contain lists of slabs which are stored in either the full bin, empty
39  * bin, or partially allocated bin, to reduce fragmentation.  They also contain
40  * the user supplied value for size, which is adjusted for alignment purposes
41  * and rsize is the result of that.  The zone also stores information for
42  * managing a hash of page addresses that maps pages to uma_slab_t structures
43  * for pages that don't have embedded uma_slab_t's.
44  *
45  * The uma_slab_t may be embedded in a UMA_SLAB_SIZE chunk of memory or it may
46  * be allocated off the page from a special slab zone.  The free list within a
47  * slab is managed with a linked list of indexes, which are 8 bit values.  If
48  * UMA_SLAB_SIZE is defined to be too large I will have to switch to 16bit
49  * values.  Currently on alpha you can get 250 or so 32 byte items and on x86
50  * you can get 250 or so 16byte items.  For item sizes that would yield more
51  * than 10% memory waste we potentially allocate a separate uma_slab_t if this
52  * will improve the number of items per slab that will fit.
53  *
54  * Other potential space optimizations are storing the 8bit of linkage in space
55  * wasted between items due to alignment problems.  This may yield a much better
56  * memory footprint for certain sizes of objects.  Another alternative is to
57  * increase the UMA_SLAB_SIZE, or allow for dynamic slab sizes.  I prefer
58  * dynamic slab sizes because we could stick with 8 bit indexes and only use
59  * large slab sizes for zones with a lot of waste per slab.  This may create
60  * ineffeciencies in the vm subsystem due to fragmentation in the address space.
61  *
62  * The only really gross cases, with regards to memory waste, are for those
63  * items that are just over half the page size.   You can get nearly 50% waste,
64  * so you fall back to the memory footprint of the power of two allocator. I
65  * have looked at memory allocation sizes on many of the machines available to
66  * me, and there does not seem to be an abundance of allocations at this range
67  * so at this time it may not make sense to optimize for it.  This can, of
68  * course, be solved with dynamic slab sizes.
69  *
70  */
71 
72 /*
73  *	This is the representation for normal (Non OFFPAGE slab)
74  *
75  *	i == item
76  *	s == slab pointer
77  *
78  *	<----------------  Page (UMA_SLAB_SIZE) ------------------>
79  *	___________________________________________________________
80  *     | _  _  _  _  _  _  _  _  _  _  _  _  _  _  _   ___________ |
81  *     ||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i| |slab header||
82  *     ||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_| |___________||
83  *     |___________________________________________________________|
84  *
85  *
86  *	This is an OFFPAGE slab. These can be larger than UMA_SLAB_SIZE.
87  *
88  *	___________________________________________________________
89  *     | _  _  _  _  _  _  _  _  _  _  _  _  _  _  _  _  _  _  _   |
90  *     ||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i|  |
91  *     ||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_|  |
92  *     |___________________________________________________________|
93  *       ___________    ^
94  *	|slab header|   |
95  *	|___________|---*
96  *
97  */
98 
99 #ifndef VM_UMA_INT_H
100 #define VM_UMA_INT_H
101 
102 #define UMA_SLAB_SIZE	PAGE_SIZE	/* How big are our slabs? */
103 #define UMA_SLAB_MASK	(PAGE_SIZE - 1)	/* Mask to get back to the page */
104 #define UMA_SLAB_SHIFT	PAGE_SHIFT	/* Number of bits PAGE_MASK */
105 
106 #define UMA_BOOT_PAGES		40	/* Pages allocated for startup */
107 
108 /* Max waste before going to off page slab management */
109 #define UMA_MAX_WASTE	(UMA_SLAB_SIZE / 10)
110 
111 /*
112  * I doubt there will be many cases where this is exceeded. This is the initial
113  * size of the hash table for uma_slabs that are managed off page. This hash
114  * does expand by powers of two.  Currently it doesn't get smaller.
115  */
116 #define UMA_HASH_SIZE_INIT	32
117 
118 /*
119  * I should investigate other hashing algorithms.  This should yield a low
120  * number of collisions if the pages are relatively contiguous.
121  *
122  * This is the same algorithm that most processor caches use.
123  *
124  * I'm shifting and masking instead of % because it should be faster.
125  */
126 
127 #define UMA_HASH(h, s) ((((unsigned long)s) >> UMA_SLAB_SHIFT) &	\
128     (h)->uh_hashmask)
129 
130 #define UMA_HASH_INSERT(h, s, mem)					\
131 		SLIST_INSERT_HEAD(&(h)->uh_slab_hash[UMA_HASH((h),	\
132 		    (mem))], (s), us_hlink);
133 #define UMA_HASH_REMOVE(h, s, mem)					\
134 		SLIST_REMOVE(&(h)->uh_slab_hash[UMA_HASH((h),		\
135 		    (mem))], (s), uma_slab, us_hlink);
136 
137 /* Page management structure */
138 
139 /* Sorry for the union, but space efficiency is important */
140 struct uma_slab {
141 	uma_zone_t	us_zone;		/* Zone we live in */
142 	union {
143 		LIST_ENTRY(uma_slab)	_us_link;	/* slabs in zone */
144 		unsigned long	_us_size;	/* Size of allocation */
145 	} us_type;
146 	SLIST_ENTRY(uma_slab)	us_hlink;	/* Link for hash table */
147 	u_int8_t	*us_data;		/* First item */
148 	u_int8_t	us_flags;		/* Page flags see uma.h */
149 	u_int8_t	us_freecount;	/* How many are free? */
150 	u_int8_t	us_firstfree;	/* First free item index */
151 	u_int8_t	us_freelist[1];	/* Free List (actually larger) */
152 };
153 
154 #define us_link	us_type._us_link
155 #define us_size	us_type._us_size
156 
157 typedef struct uma_slab * uma_slab_t;
158 
159 /* Hash table for freed address -> slab translation */
160 
161 SLIST_HEAD(slabhead, uma_slab);
162 
163 struct uma_hash {
164 	struct slabhead	*uh_slab_hash;	/* Hash table for slabs */
165 	int		uh_hashsize;	/* Current size of the hash table */
166 	int		uh_hashmask;	/* Mask used during hashing */
167 };
168 
169 /*
170  * Structures for per cpu queues.
171  */
172 
173 struct uma_bucket {
174 	LIST_ENTRY(uma_bucket)	ub_link;	/* Link into the zone */
175 	int16_t	ub_cnt;				/* Count of free items. */
176 	int16_t	ub_entries;			/* Max items. */
177 	void	*ub_bucket[];			/* actual allocation storage */
178 };
179 
180 typedef struct uma_bucket * uma_bucket_t;
181 
182 struct uma_cache {
183 	uma_bucket_t	uc_freebucket;	/* Bucket we're freeing to */
184 	uma_bucket_t	uc_allocbucket;	/* Bucket to allocate from */
185 	u_int64_t	uc_allocs;	/* Count of allocations */
186 };
187 
188 typedef struct uma_cache * uma_cache_t;
189 
190 /*
191  * Zone management structure
192  *
193  * TODO: Optimize for cache line size
194  *
195  */
196 struct uma_zone {
197 	char		*uz_name;	/* Text name of the zone */
198 	LIST_ENTRY(uma_zone)	uz_link;	/* List of all zones */
199 	u_int32_t	uz_align;	/* Alignment mask */
200 	u_int32_t	uz_pages;	/* Total page count */
201 
202 /* Used during alloc / free */
203 	struct mtx	uz_lock;	/* Lock for the zone */
204 	u_int32_t	uz_free;	/* Count of items free in slabs */
205 	u_int16_t	uz_ipers;	/* Items per slab */
206 	u_int16_t	uz_flags;	/* Internal flags */
207 
208 	LIST_HEAD(,uma_slab)	uz_part_slab;	/* partially allocated slabs */
209 	LIST_HEAD(,uma_slab)	uz_free_slab;	/* empty slab list */
210 	LIST_HEAD(,uma_slab)	uz_full_slab;	/* full slabs */
211 	LIST_HEAD(,uma_bucket)	uz_full_bucket;	/* full buckets */
212 	LIST_HEAD(,uma_bucket)	uz_free_bucket;	/* Buckets for frees */
213 	u_int32_t	uz_size;	/* Requested size of each item */
214 	u_int32_t	uz_rsize;	/* Real size of each item */
215 
216 	struct uma_hash	uz_hash;
217 	u_int16_t	uz_pgoff;	/* Offset to uma_slab struct */
218 	u_int16_t	uz_ppera;	/* pages per allocation from backend */
219 
220 	uma_ctor	uz_ctor;	/* Constructor for each allocation */
221 	uma_dtor	uz_dtor;	/* Destructor */
222 	u_int64_t	uz_allocs;	/* Total number of allocations */
223 
224 	uma_init	uz_init;	/* Initializer for each item */
225 	uma_fini	uz_fini;	/* Discards memory */
226 	uma_alloc	uz_allocf;	/* Allocation function */
227 	uma_free	uz_freef;	/* Free routine */
228 	struct vm_object	*uz_obj;	/* Zone specific object */
229 	vm_offset_t	uz_kva;		/* Base kva for zones with objs */
230 	u_int32_t	uz_maxpages;	/* Maximum number of pages to alloc */
231 	int		uz_recurse;	/* Allocation recursion count */
232 	uint16_t	uz_fills;	/* Outstanding bucket fills */
233 	uint16_t	uz_count;	/* Highest value ub_ptr can have */
234 	/*
235 	 * This HAS to be the last item because we adjust the zone size
236 	 * based on NCPU and then allocate the space for the zones.
237 	 */
238 	struct uma_cache	uz_cpu[1];	/* Per cpu caches */
239 };
240 
241 /*
242  * These flags must not overlap with the UMA_ZONE flags specified in uma.h.
243  */
244 #define UMA_ZFLAG_PRIVALLOC	0x1000		/* Use uz_allocf. */
245 #define UMA_ZFLAG_INTERNAL	0x2000		/* No offpage no PCPU. */
246 #define UMA_ZFLAG_FULL		0x4000		/* Reached uz_maxpages */
247 #define UMA_ZFLAG_CACHEONLY	0x8000		/* Don't ask VM for buckets. */
248 
249 /* Internal prototypes */
250 static __inline uma_slab_t hash_sfind(struct uma_hash *hash, u_int8_t *data);
251 void *uma_large_malloc(int size, int wait);
252 void uma_large_free(uma_slab_t slab);
253 
254 /* Lock Macros */
255 
256 #define	ZONE_LOCK_INIT(z, lc)					\
257 	do {							\
258 		if ((lc))					\
259 			mtx_init(&(z)->uz_lock, (z)->uz_name,	\
260 			    (z)->uz_name, MTX_DEF | MTX_DUPOK);	\
261 		else						\
262 			mtx_init(&(z)->uz_lock, (z)->uz_name,	\
263 			    "UMA zone", MTX_DEF | MTX_DUPOK);	\
264 	} while (0)
265 
266 #define	ZONE_LOCK_FINI(z)	mtx_destroy(&(z)->uz_lock)
267 #define	ZONE_LOCK(z)	mtx_lock(&(z)->uz_lock)
268 #define ZONE_UNLOCK(z)	mtx_unlock(&(z)->uz_lock)
269 
270 #define	CPU_LOCK_INIT(cpu)					\
271 	mtx_init(&uma_pcpu_mtx[(cpu)], "UMA pcpu", "UMA pcpu",	\
272 	    MTX_DEF | MTX_DUPOK)
273 
274 #define CPU_LOCK(cpu)						\
275 	mtx_lock(&uma_pcpu_mtx[(cpu)])
276 
277 #define CPU_UNLOCK(cpu)						\
278 	mtx_unlock(&uma_pcpu_mtx[(cpu)])
279 
280 /*
281  * Find a slab within a hash table.  This is used for OFFPAGE zones to lookup
282  * the slab structure.
283  *
284  * Arguments:
285  *	hash  The hash table to search.
286  *	data  The base page of the item.
287  *
288  * Returns:
289  *	A pointer to a slab if successful, else NULL.
290  */
291 static __inline uma_slab_t
292 hash_sfind(struct uma_hash *hash, u_int8_t *data)
293 {
294         uma_slab_t slab;
295         int hval;
296 
297         hval = UMA_HASH(hash, data);
298 
299         SLIST_FOREACH(slab, &hash->uh_slab_hash[hval], us_hlink) {
300                 if ((u_int8_t *)slab->us_data == data)
301                         return (slab);
302         }
303         return (NULL);
304 }
305 
306 static __inline uma_slab_t
307 vtoslab(vm_offset_t va)
308 {
309 	vm_page_t p;
310 	uma_slab_t slab;
311 
312 	p = PHYS_TO_VM_PAGE(pmap_kextract(va));
313 	slab = (uma_slab_t )p->object;
314 
315 	if (p->flags & PG_SLAB)
316 		return (slab);
317 	else
318 		return (NULL);
319 }
320 
321 static __inline void
322 vsetslab(vm_offset_t va, uma_slab_t slab)
323 {
324 	vm_page_t p;
325 
326 	p = PHYS_TO_VM_PAGE(pmap_kextract((vm_offset_t)va));
327 	p->object = (vm_object_t)slab;
328 	p->flags |= PG_SLAB;
329 }
330 
331 static __inline void
332 vsetobj(vm_offset_t va, vm_object_t obj)
333 {
334 	vm_page_t p;
335 
336 	p = PHYS_TO_VM_PAGE(pmap_kextract((vm_offset_t)va));
337 	p->object = obj;
338 	p->flags &= ~PG_SLAB;
339 }
340 
341 /*
342  * The following two functions may be defined by architecture specific code
343  * if they can provide more effecient allocation functions.  This is useful
344  * for using direct mapped addresses.
345  */
346 void *uma_small_alloc(uma_zone_t zone, int bytes, u_int8_t *pflag, int wait);
347 void uma_small_free(void *mem, int size, u_int8_t flags);
348 
349 #endif /* VM_UMA_INT_H */
350