xref: /freebsd/sys/vm/uma_int.h (revision a3e8fd0b7f663db7eafff527d5c3ca3bcfa8a537)
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		30	/* Number of pages allocated for startup */
107 #define UMA_WORKING_TIME	20	/* Seconds worth of items to keep */
108 
109 
110 /* Max waste before going to off page slab management */
111 #define UMA_MAX_WASTE	(UMA_SLAB_SIZE / 10)
112 
113 /*
114  * I doubt there will be many cases where this is exceeded. This is the initial
115  * size of the hash table for uma_slabs that are managed off page. This hash
116  * does expand by powers of two.  Currently it doesn't get smaller.
117  */
118 #define UMA_HASH_SIZE_INIT	32
119 
120 
121 /*
122  * I should investigate other hashing algorithms.  This should yield a low
123  * number of collisions if the pages are relatively contiguous.
124  *
125  * This is the same algorithm that most processor caches use.
126  *
127  * I'm shifting and masking instead of % because it should be faster.
128  */
129 
130 #define UMA_HASH(h, s) ((((unsigned long)s) >> UMA_SLAB_SHIFT) &	\
131     (h)->uh_hashmask)
132 
133 #define UMA_HASH_INSERT(h, s, mem)					\
134 		SLIST_INSERT_HEAD(&(h)->uh_slab_hash[UMA_HASH((h),	\
135 		    (mem))], (s), us_hlink);
136 #define UMA_HASH_REMOVE(h, s, mem)					\
137 		SLIST_REMOVE(&(h)->uh_slab_hash[UMA_HASH((h),		\
138 		    (mem))], (s), uma_slab, us_hlink);
139 
140 /* Page management structure */
141 
142 /* Sorry for the union, but space efficiency is important */
143 struct uma_slab {
144 	uma_zone_t	us_zone;		/* Zone we live in */
145 	union {
146 		LIST_ENTRY(uma_slab)	us_link;	/* slabs in zone */
147 		unsigned long	us_size;	/* Size of allocation */
148 	} us_type;
149 	SLIST_ENTRY(uma_slab)	us_hlink;	/* Link for hash table */
150 	u_int8_t	*us_data;		/* First item */
151 	u_int8_t	us_flags;		/* Page flags see uma.h */
152 	u_int8_t	us_freecount;	/* How many are free? */
153 	u_int8_t	us_firstfree;	/* First free item index */
154 	u_int8_t	us_freelist[1];	/* Free List (actually larger) */
155 };
156 
157 #define us_link	us_type.us_link
158 #define us_size	us_type.us_size
159 
160 typedef struct uma_slab * uma_slab_t;
161 
162 /* Hash table for freed address -> slab translation */
163 
164 SLIST_HEAD(slabhead, uma_slab);
165 
166 struct uma_hash {
167 	struct slabhead	*uh_slab_hash;	/* Hash table for slabs */
168 	int		uh_hashsize;	/* Current size of the hash table */
169 	int		uh_hashmask;	/* Mask used during hashing */
170 };
171 
172 /*
173  * Structures for per cpu queues.
174  */
175 
176 /*
177  * This size was chosen so that the struct bucket size is roughly
178  * 128 * sizeof(void *).  This is exactly true for x86, and for alpha
179  * it will would be 32bits smaller if it didn't have alignment adjustments.
180  */
181 
182 #define UMA_BUCKET_SIZE	125
183 
184 struct uma_bucket {
185 	LIST_ENTRY(uma_bucket)	ub_link;	/* Link into the zone */
186 	int16_t	ub_ptr;				/* Pointer to current item */
187 	void	*ub_bucket[UMA_BUCKET_SIZE];	/* actual allocation storage */
188 };
189 
190 typedef struct uma_bucket * uma_bucket_t;
191 
192 struct uma_cache {
193 	struct mtx	uc_lock;	/* Spin lock on this cpu's bucket */
194 	uma_bucket_t	uc_freebucket;	/* Bucket we're freeing to */
195 	uma_bucket_t	uc_allocbucket;	/* Bucket to allocate from */
196 	u_int64_t	uc_allocs;	/* Count of allocations */
197 };
198 
199 typedef struct uma_cache * uma_cache_t;
200 
201 #define LOCKNAME_LEN	16		/* Length of the name for cpu locks */
202 
203 /*
204  * Zone management structure
205  *
206  * TODO: Optimize for cache line size
207  *
208  */
209 struct uma_zone {
210 	char		uz_lname[LOCKNAME_LEN];	/* Text name for the cpu lock */
211 	char		*uz_name;	/* Text name of the zone */
212 	LIST_ENTRY(uma_zone)	uz_link;	/* List of all zones */
213 	u_int32_t	uz_align;	/* Alignment mask */
214 	u_int32_t	uz_pages;	/* Total page count */
215 
216 /* Used during alloc / free */
217 	struct mtx	uz_lock;	/* Lock for the zone */
218 	u_int32_t	uz_free;	/* Count of items free in slabs */
219 	u_int16_t	uz_ipers;	/* Items per slab */
220 	u_int16_t	uz_flags;	/* Internal flags */
221 
222 	LIST_HEAD(,uma_slab)	uz_part_slab;	/* partially allocated slabs */
223 	LIST_HEAD(,uma_slab)	uz_free_slab;	/* empty slab list */
224 	LIST_HEAD(,uma_slab)	uz_full_slab;	/* full slabs */
225 	LIST_HEAD(,uma_bucket)	uz_full_bucket;	/* full buckets */
226 	LIST_HEAD(,uma_bucket)	uz_free_bucket;	/* Buckets for frees */
227 	u_int32_t	uz_size;	/* Requested size of each item */
228 	u_int32_t	uz_rsize;	/* Real size of each item */
229 
230 	struct uma_hash	uz_hash;
231 	u_int16_t	uz_pgoff;	/* Offset to uma_slab struct */
232 	u_int16_t	uz_ppera;	/* pages per allocation from backend */
233 	u_int16_t	uz_cacheoff;	/* Next cache offset */
234 	u_int16_t	uz_cachemax;	/* Max cache offset */
235 
236 	uma_ctor	uz_ctor;	/* Constructor for each allocation */
237 	uma_dtor	uz_dtor;	/* Destructor */
238 	u_int64_t	uz_allocs;	/* Total number of allocations */
239 
240 	uma_init	uz_init;	/* Initializer for each item */
241 	uma_fini	uz_fini;	/* Discards memory */
242 	uma_alloc	uz_allocf;	/* Allocation function */
243 	uma_free	uz_freef;	/* Free routine */
244 	struct vm_object	*uz_obj;	/* Zone specific object */
245 	vm_offset_t	uz_kva;		/* Base kva for zones with objs */
246 	u_int32_t	uz_maxpages;	/* Maximum number of pages to alloc */
247 	u_int32_t	uz_cachefree;	/* Last count of items free in caches */
248 	u_int64_t	uz_oallocs;	/* old allocs count */
249 	u_int64_t	uz_wssize;	/* Working set size */
250 	int		uz_recurse;	/* Allocation recursion count */
251 	uint16_t	uz_fills;	/* Outstanding bucket fills */
252 	uint16_t	uz_count;	/* Highest value ub_ptr can have */
253 	/*
254 	 * This HAS to be the last item because we adjust the zone size
255 	 * based on NCPU and then allocate the space for the zones.
256 	 */
257 	struct uma_cache	uz_cpu[1];	/* Per cpu caches */
258 };
259 
260 #define UMA_CACHE_INC	16	/* How much will we move data */
261 
262 #define UMA_ZFLAG_OFFPAGE	0x0001	/* Struct slab/freelist off page */
263 #define UMA_ZFLAG_PRIVALLOC	0x0002	/* Zone has supplied it's own alloc */
264 #define UMA_ZFLAG_INTERNAL	0x0004	/* Internal zone, no offpage no PCPU */
265 #define UMA_ZFLAG_MALLOC	0x0008	/* Zone created by malloc */
266 #define UMA_ZFLAG_NOFREE	0x0010	/* Don't free data from this zone */
267 #define UMA_ZFLAG_FULL		0x0020	/* This zone reached uz_maxpages */
268 #define UMA_ZFLAG_BUCKETCACHE	0x0040	/* Only allocate buckets from cache */
269 #define	UMA_ZFLAG_HASH		0x0080	/* Look up slab via hash */
270 
271 /* This lives in uflags */
272 #define UMA_ZONE_INTERNAL	0x1000	/* Internal zone for uflags */
273 
274 /* Internal prototypes */
275 static __inline uma_slab_t hash_sfind(struct uma_hash *hash, u_int8_t *data);
276 void *uma_large_malloc(int size, int wait);
277 void uma_large_free(uma_slab_t slab);
278 
279 /* Lock Macros */
280 
281 #define	ZONE_LOCK_INIT(z, lc)					\
282 	do {							\
283 		if ((lc))					\
284 			mtx_init(&(z)->uz_lock, (z)->uz_name,	\
285 			    (z)->uz_name, MTX_DEF | MTX_DUPOK);	\
286 		else						\
287 			mtx_init(&(z)->uz_lock, (z)->uz_name,	\
288 			    "UMA zone", MTX_DEF | MTX_DUPOK);	\
289 	} while (0)
290 
291 #define	ZONE_LOCK_FINI(z)	mtx_destroy(&(z)->uz_lock)
292 #define	ZONE_LOCK(z)	mtx_lock(&(z)->uz_lock)
293 #define ZONE_UNLOCK(z)	mtx_unlock(&(z)->uz_lock)
294 
295 #define	CPU_LOCK_INIT(z, cpu, lc)				\
296 	do {							\
297 		if ((lc))					\
298 			mtx_init(&(z)->uz_cpu[(cpu)].uc_lock,	\
299 			    (z)->uz_lname, (z)->uz_lname,	\
300 			    MTX_DEF | MTX_DUPOK);		\
301 		else						\
302 			mtx_init(&(z)->uz_cpu[(cpu)].uc_lock,	\
303 			    (z)->uz_lname, "UMA cpu",		\
304 			    MTX_DEF | MTX_DUPOK);		\
305 	} while (0)
306 
307 #define	CPU_LOCK_FINI(z, cpu)	\
308 	mtx_destroy(&(z)->uz_cpu[(cpu)].uc_lock)
309 
310 #define CPU_LOCK(z, cpu)	\
311 	mtx_lock(&(z)->uz_cpu[(cpu)].uc_lock)
312 
313 #define CPU_UNLOCK(z, cpu)	\
314 	mtx_unlock(&(z)->uz_cpu[(cpu)].uc_lock)
315 
316 /*
317  * Find a slab within a hash table.  This is used for OFFPAGE zones to lookup
318  * the slab structure.
319  *
320  * Arguments:
321  *	hash  The hash table to search.
322  *	data  The base page of the item.
323  *
324  * Returns:
325  *	A pointer to a slab if successful, else NULL.
326  */
327 static __inline uma_slab_t
328 hash_sfind(struct uma_hash *hash, u_int8_t *data)
329 {
330         uma_slab_t slab;
331         int hval;
332 
333         hval = UMA_HASH(hash, data);
334 
335         SLIST_FOREACH(slab, &hash->uh_slab_hash[hval], us_hlink) {
336                 if ((u_int8_t *)slab->us_data == data)
337                         return (slab);
338         }
339         return (NULL);
340 }
341 
342 static __inline uma_slab_t
343 vtoslab(vm_offset_t va)
344 {
345 	vm_page_t p;
346 	uma_slab_t slab;
347 
348 	p = PHYS_TO_VM_PAGE(pmap_kextract(va));
349 	slab = (uma_slab_t )p->object;
350 
351 	if (p->flags & PG_SLAB)
352 		return (slab);
353 	else
354 		return (NULL);
355 }
356 
357 static __inline void
358 vsetslab(vm_offset_t va, uma_slab_t slab)
359 {
360 	vm_page_t p;
361 
362 	p = PHYS_TO_VM_PAGE(pmap_kextract((vm_offset_t)va));
363 	p->object = (vm_object_t)slab;
364 	p->flags |= PG_SLAB;
365 }
366 
367 static __inline void
368 vsetobj(vm_offset_t va, vm_object_t obj)
369 {
370 	vm_page_t p;
371 
372 	p = PHYS_TO_VM_PAGE(pmap_kextract((vm_offset_t)va));
373 	p->object = obj;
374 	p->flags &= ~PG_SLAB;
375 }
376 
377 #endif /* VM_UMA_INT_H */
378