xref: /freebsd/sys/vm/uma_int.h (revision b1d046441de9053152c7cf03d6b60d9882687e1b)
1 /*-
2  * Copyright (c) 2002-2005, 2009 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 linked list of indices, which are 8 bit values.  If
49  * UMA_SLAB_SIZE is defined to be too large I will have to switch to 16bit
50  * values.  Currently on alpha you can get 250 or so 32 byte items and on x86
51  * you can get 250 or so 16byte items.  For item sizes that would yield more
52  * than 10% memory waste we potentially allocate a separate uma_slab_t if this
53  * will improve the number of items per slab that will fit.
54  *
55  * Other potential space optimizations are storing the 8bit of linkage in space
56  * wasted between items due to alignment problems.  This may yield a much better
57  * memory footprint for certain sizes of objects.  Another alternative is to
58  * increase the UMA_SLAB_SIZE, or allow for dynamic slab sizes.  I prefer
59  * dynamic slab sizes because we could stick with 8 bit indices and only use
60  * large slab sizes for zones with a lot of waste per slab.  This may create
61  * inefficiencies in the vm subsystem due to fragmentation in the address space.
62  *
63  * The only really gross cases, with regards to memory waste, are for those
64  * items that are just over half the page size.   You can get nearly 50% waste,
65  * so you fall back to the memory footprint of the power of two allocator. I
66  * have looked at memory allocation sizes on many of the machines available to
67  * me, and there does not seem to be an abundance of allocations at this range
68  * so at this time it may not make sense to optimize for it.  This can, of
69  * course, be solved with dynamic slab sizes.
70  *
71  * Kegs may serve multiple Zones but by far most of the time they only serve
72  * one.  When a Zone is created, a Keg is allocated and setup for it.  While
73  * the backing Keg stores slabs, the Zone caches Buckets of items allocated
74  * from the slabs.  Each Zone is equipped with an init/fini and ctor/dtor
75  * pair, as well as with its own set of small per-CPU caches, layered above
76  * the Zone's general Bucket cache.
77  *
78  * The PCPU caches are protected by critical sections, and may be accessed
79  * safely only from their associated CPU, while the Zones backed by the same
80  * Keg all share a common Keg lock (to coalesce contention on the backing
81  * slabs).  The backing Keg typically only serves one Zone but in the case of
82  * multiple Zones, one of the Zones is considered the Master Zone and all
83  * Zone-related stats from the Keg are done in the Master Zone.  For an
84  * example of a Multi-Zone setup, refer to the Mbuf allocation code.
85  */
86 
87 /*
88  *	This is the representation for normal (Non OFFPAGE slab)
89  *
90  *	i == item
91  *	s == slab pointer
92  *
93  *	<----------------  Page (UMA_SLAB_SIZE) ------------------>
94  *	___________________________________________________________
95  *     | _  _  _  _  _  _  _  _  _  _  _  _  _  _  _   ___________ |
96  *     ||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i| |slab header||
97  *     ||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_| |___________||
98  *     |___________________________________________________________|
99  *
100  *
101  *	This is an OFFPAGE slab. These can be larger than UMA_SLAB_SIZE.
102  *
103  *	___________________________________________________________
104  *     | _  _  _  _  _  _  _  _  _  _  _  _  _  _  _  _  _  _  _   |
105  *     ||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i|  |
106  *     ||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_|  |
107  *     |___________________________________________________________|
108  *       ___________    ^
109  *	|slab header|   |
110  *	|___________|---*
111  *
112  */
113 
114 #ifndef VM_UMA_INT_H
115 #define VM_UMA_INT_H
116 
117 #define UMA_SLAB_SIZE	PAGE_SIZE	/* How big are our slabs? */
118 #define UMA_SLAB_MASK	(PAGE_SIZE - 1)	/* Mask to get back to the page */
119 #define UMA_SLAB_SHIFT	PAGE_SHIFT	/* Number of bits PAGE_MASK */
120 
121 #define UMA_BOOT_PAGES		64	/* Pages allocated for startup */
122 
123 /* Max waste before going to off page slab management */
124 #define UMA_MAX_WASTE	(UMA_SLAB_SIZE / 10)
125 
126 /*
127  * I doubt there will be many cases where this is exceeded. This is the initial
128  * size of the hash table for uma_slabs that are managed off page. This hash
129  * does expand by powers of two.  Currently it doesn't get smaller.
130  */
131 #define UMA_HASH_SIZE_INIT	32
132 
133 /*
134  * I should investigate other hashing algorithms.  This should yield a low
135  * number of collisions if the pages are relatively contiguous.
136  *
137  * This is the same algorithm that most processor caches use.
138  *
139  * I'm shifting and masking instead of % because it should be faster.
140  */
141 
142 #define UMA_HASH(h, s) ((((unsigned long)s) >> UMA_SLAB_SHIFT) &	\
143     (h)->uh_hashmask)
144 
145 #define UMA_HASH_INSERT(h, s, mem)					\
146 		SLIST_INSERT_HEAD(&(h)->uh_slab_hash[UMA_HASH((h),	\
147 		    (mem))], (s), us_hlink)
148 #define UMA_HASH_REMOVE(h, s, mem)					\
149 		SLIST_REMOVE(&(h)->uh_slab_hash[UMA_HASH((h),		\
150 		    (mem))], (s), uma_slab, us_hlink)
151 
152 /* Hash table for freed address -> slab translation */
153 
154 SLIST_HEAD(slabhead, uma_slab);
155 
156 struct uma_hash {
157 	struct slabhead	*uh_slab_hash;	/* Hash table for slabs */
158 	int		uh_hashsize;	/* Current size of the hash table */
159 	int		uh_hashmask;	/* Mask used during hashing */
160 };
161 
162 /*
163  * align field or structure to cache line
164  */
165 #if defined(__amd64__)
166 #define UMA_ALIGN	__aligned(CACHE_LINE_SIZE)
167 #else
168 #define UMA_ALIGN
169 #endif
170 
171 /*
172  * Structures for per cpu queues.
173  */
174 
175 struct uma_bucket {
176 	LIST_ENTRY(uma_bucket)	ub_link;	/* Link into the zone */
177 	int16_t	ub_cnt;				/* Count of free items. */
178 	int16_t	ub_entries;			/* Max items. */
179 	void	*ub_bucket[];			/* actual allocation storage */
180 };
181 
182 typedef struct uma_bucket * uma_bucket_t;
183 
184 struct uma_cache {
185 	uma_bucket_t	uc_freebucket;	/* Bucket we're freeing to */
186 	uma_bucket_t	uc_allocbucket;	/* Bucket to allocate from */
187 	u_int64_t	uc_allocs;	/* Count of allocations */
188 	u_int64_t	uc_frees;	/* Count of frees */
189 } UMA_ALIGN;
190 
191 typedef struct uma_cache * uma_cache_t;
192 
193 /*
194  * Keg management structure
195  *
196  * TODO: Optimize for cache line size
197  *
198  */
199 struct uma_keg {
200 	LIST_ENTRY(uma_keg)	uk_link;	/* List of all kegs */
201 
202 	struct mtx	uk_lock;	/* Lock for the keg */
203 	struct uma_hash	uk_hash;
204 
205 	char		*uk_name;		/* Name of creating zone. */
206 	LIST_HEAD(,uma_zone)	uk_zones;	/* Keg's zones */
207 	LIST_HEAD(,uma_slab)	uk_part_slab;	/* partially allocated slabs */
208 	LIST_HEAD(,uma_slab)	uk_free_slab;	/* empty slab list */
209 	LIST_HEAD(,uma_slab)	uk_full_slab;	/* full slabs */
210 
211 	u_int32_t	uk_recurse;	/* Allocation recursion count */
212 	u_int32_t	uk_align;	/* Alignment mask */
213 	u_int32_t	uk_pages;	/* Total page count */
214 	u_int32_t	uk_free;	/* Count of items free in slabs */
215 	u_int32_t	uk_size;	/* Requested size of each item */
216 	u_int32_t	uk_rsize;	/* Real size of each item */
217 	u_int32_t	uk_maxpages;	/* Maximum number of pages to alloc */
218 
219 	uma_init	uk_init;	/* Keg's init routine */
220 	uma_fini	uk_fini;	/* Keg's fini routine */
221 	uma_alloc	uk_allocf;	/* Allocation function */
222 	uma_free	uk_freef;	/* Free routine */
223 
224 	struct vm_object	*uk_obj;	/* Zone specific object */
225 	vm_offset_t	uk_kva;		/* Base kva for zones with objs */
226 	uma_zone_t	uk_slabzone;	/* Slab zone backing us, if OFFPAGE */
227 
228 	u_int16_t	uk_pgoff;	/* Offset to uma_slab struct */
229 	u_int16_t	uk_ppera;	/* pages per allocation from backend */
230 	u_int16_t	uk_ipers;	/* Items per slab */
231 	u_int32_t	uk_flags;	/* Internal flags */
232 };
233 typedef struct uma_keg	* uma_keg_t;
234 
235 /* Page management structure */
236 
237 /* Sorry for the union, but space efficiency is important */
238 struct uma_slab_head {
239 	uma_keg_t	us_keg;			/* Keg we live in */
240 	union {
241 		LIST_ENTRY(uma_slab)	_us_link;	/* slabs in zone */
242 		unsigned long	_us_size;	/* Size of allocation */
243 	} us_type;
244 	SLIST_ENTRY(uma_slab)	us_hlink;	/* Link for hash table */
245 	u_int8_t	*us_data;		/* First item */
246 	u_int8_t	us_flags;		/* Page flags see uma.h */
247 	u_int8_t	us_freecount;	/* How many are free? */
248 	u_int8_t	us_firstfree;	/* First free item index */
249 };
250 
251 /* The standard slab structure */
252 struct uma_slab {
253 	struct uma_slab_head	us_head;	/* slab header data */
254 	struct {
255 		u_int8_t	us_item;
256 	} us_freelist[1];			/* actual number bigger */
257 };
258 
259 /*
260  * The slab structure for UMA_ZONE_REFCNT zones for whose items we
261  * maintain reference counters in the slab for.
262  */
263 struct uma_slab_refcnt {
264 	struct uma_slab_head	us_head;	/* slab header data */
265 	struct {
266 		u_int8_t	us_item;
267 		u_int32_t	us_refcnt;
268 	} us_freelist[1];			/* actual number bigger */
269 };
270 
271 #define	us_keg		us_head.us_keg
272 #define	us_link		us_head.us_type._us_link
273 #define	us_size		us_head.us_type._us_size
274 #define	us_hlink	us_head.us_hlink
275 #define	us_data		us_head.us_data
276 #define	us_flags	us_head.us_flags
277 #define	us_freecount	us_head.us_freecount
278 #define	us_firstfree	us_head.us_firstfree
279 
280 typedef struct uma_slab * uma_slab_t;
281 typedef struct uma_slab_refcnt * uma_slabrefcnt_t;
282 typedef uma_slab_t (*uma_slaballoc)(uma_zone_t, uma_keg_t, int);
283 
284 
285 /*
286  * These give us the size of one free item reference within our corresponding
287  * uma_slab structures, so that our calculations during zone setup are correct
288  * regardless of what the compiler decides to do with padding the structure
289  * arrays within uma_slab.
290  */
291 #define	UMA_FRITM_SZ	(sizeof(struct uma_slab) - sizeof(struct uma_slab_head))
292 #define	UMA_FRITMREF_SZ	(sizeof(struct uma_slab_refcnt) -	\
293     sizeof(struct uma_slab_head))
294 
295 struct uma_klink {
296 	LIST_ENTRY(uma_klink)	kl_link;
297 	uma_keg_t		kl_keg;
298 };
299 typedef struct uma_klink *uma_klink_t;
300 
301 /*
302  * Zone management structure
303  *
304  * TODO: Optimize for cache line size
305  *
306  */
307 struct uma_zone {
308 	char		*uz_name;	/* Text name of the zone */
309 	struct mtx	*uz_lock;	/* Lock for the zone (keg's lock) */
310 
311 	LIST_ENTRY(uma_zone)	uz_link;	/* List of all zones in keg */
312 	LIST_HEAD(,uma_bucket)	uz_full_bucket;	/* full buckets */
313 	LIST_HEAD(,uma_bucket)	uz_free_bucket;	/* Buckets for frees */
314 
315 	LIST_HEAD(,uma_klink)	uz_kegs;	/* List of kegs. */
316 	struct uma_klink	uz_klink;	/* klink for first keg. */
317 
318 	uma_slaballoc	uz_slab;	/* Allocate a slab from the backend. */
319 	uma_ctor	uz_ctor;	/* Constructor for each allocation */
320 	uma_dtor	uz_dtor;	/* Destructor */
321 	uma_init	uz_init;	/* Initializer for each item */
322 	uma_fini	uz_fini;	/* Discards memory */
323 
324 	u_int32_t	uz_flags;	/* Flags inherited from kegs */
325 	u_int32_t	uz_size;	/* Size inherited from kegs */
326 
327 	u_int64_t	uz_allocs UMA_ALIGN; /* Total number of allocations */
328 	u_int64_t	uz_frees;	/* Total number of frees */
329 	u_int64_t	uz_fails;	/* Total number of alloc failures */
330 	u_int64_t	uz_sleeps;	/* Total number of alloc sleeps */
331 	uint16_t	uz_fills;	/* Outstanding bucket fills */
332 	uint16_t	uz_count;	/* Highest value ub_ptr can have */
333 
334 	/*
335 	 * This HAS to be the last item because we adjust the zone size
336 	 * based on NCPU and then allocate the space for the zones.
337 	 */
338 	struct uma_cache	uz_cpu[1]; /* Per cpu caches */
339 };
340 
341 /*
342  * These flags must not overlap with the UMA_ZONE flags specified in uma.h.
343  */
344 #define	UMA_ZFLAG_BUCKET	0x02000000	/* Bucket zone. */
345 #define	UMA_ZFLAG_MULTI		0x04000000	/* Multiple kegs in the zone. */
346 #define	UMA_ZFLAG_DRAINING	0x08000000	/* Running zone_drain. */
347 #define UMA_ZFLAG_PRIVALLOC	0x10000000	/* Use uz_allocf. */
348 #define UMA_ZFLAG_INTERNAL	0x20000000	/* No offpage no PCPU. */
349 #define UMA_ZFLAG_FULL		0x40000000	/* Reached uz_maxpages */
350 #define UMA_ZFLAG_CACHEONLY	0x80000000	/* Don't ask VM for buckets. */
351 
352 #define	UMA_ZFLAG_INHERIT	(UMA_ZFLAG_INTERNAL | UMA_ZFLAG_CACHEONLY | \
353 				    UMA_ZFLAG_BUCKET)
354 
355 #undef UMA_ALIGN
356 
357 #ifdef _KERNEL
358 /* Internal prototypes */
359 static __inline uma_slab_t hash_sfind(struct uma_hash *hash, u_int8_t *data);
360 void *uma_large_malloc(int size, int wait);
361 void uma_large_free(uma_slab_t slab);
362 
363 /* Lock Macros */
364 
365 #define	KEG_LOCK_INIT(k, lc)					\
366 	do {							\
367 		if ((lc))					\
368 			mtx_init(&(k)->uk_lock, (k)->uk_name,	\
369 			    (k)->uk_name, MTX_DEF | MTX_DUPOK);	\
370 		else						\
371 			mtx_init(&(k)->uk_lock, (k)->uk_name,	\
372 			    "UMA zone", MTX_DEF | MTX_DUPOK);	\
373 	} while (0)
374 
375 #define	KEG_LOCK_FINI(k)	mtx_destroy(&(k)->uk_lock)
376 #define	KEG_LOCK(k)	mtx_lock(&(k)->uk_lock)
377 #define	KEG_UNLOCK(k)	mtx_unlock(&(k)->uk_lock)
378 #define	ZONE_LOCK(z)	mtx_lock((z)->uz_lock)
379 #define ZONE_UNLOCK(z)	mtx_unlock((z)->uz_lock)
380 
381 /*
382  * Find a slab within a hash table.  This is used for OFFPAGE zones to lookup
383  * the slab structure.
384  *
385  * Arguments:
386  *	hash  The hash table to search.
387  *	data  The base page of the item.
388  *
389  * Returns:
390  *	A pointer to a slab if successful, else NULL.
391  */
392 static __inline uma_slab_t
393 hash_sfind(struct uma_hash *hash, u_int8_t *data)
394 {
395         uma_slab_t slab;
396         int hval;
397 
398         hval = UMA_HASH(hash, data);
399 
400         SLIST_FOREACH(slab, &hash->uh_slab_hash[hval], us_hlink) {
401                 if ((u_int8_t *)slab->us_data == data)
402                         return (slab);
403         }
404         return (NULL);
405 }
406 
407 static __inline uma_slab_t
408 vtoslab(vm_offset_t va)
409 {
410 	vm_page_t p;
411 	uma_slab_t slab;
412 
413 	p = PHYS_TO_VM_PAGE(pmap_kextract(va));
414 	slab = (uma_slab_t )p->object;
415 
416 	if (p->flags & PG_SLAB)
417 		return (slab);
418 	else
419 		return (NULL);
420 }
421 
422 static __inline void
423 vsetslab(vm_offset_t va, uma_slab_t slab)
424 {
425 	vm_page_t p;
426 
427 	p = PHYS_TO_VM_PAGE(pmap_kextract(va));
428 	p->object = (vm_object_t)slab;
429 	p->flags |= PG_SLAB;
430 }
431 
432 static __inline void
433 vsetobj(vm_offset_t va, vm_object_t obj)
434 {
435 	vm_page_t p;
436 
437 	p = PHYS_TO_VM_PAGE(pmap_kextract(va));
438 	p->object = obj;
439 	p->flags &= ~PG_SLAB;
440 }
441 
442 /*
443  * The following two functions may be defined by architecture specific code
444  * if they can provide more effecient allocation functions.  This is useful
445  * for using direct mapped addresses.
446  */
447 void *uma_small_alloc(uma_zone_t zone, int bytes, u_int8_t *pflag, int wait);
448 void uma_small_free(void *mem, int size, u_int8_t flags);
449 #endif /* _KERNEL */
450 
451 #endif /* VM_UMA_INT_H */
452