xref: /freebsd/sys/vm/uma_int.h (revision f0157ce528a128e2abb181a5c766033a2ce49a5f)
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  * Other potential space optimizations are storing the 8bit of linkage in space
53  * wasted between items due to alignment problems.  This may yield a much better
54  * memory footprint for certain sizes of objects.  Another alternative is to
55  * increase the UMA_SLAB_SIZE, or allow for dynamic slab sizes.  I prefer
56  * dynamic slab sizes because we could stick with 8 bit indices and only use
57  * large slab sizes for zones with a lot of waste per slab.  This may create
58  * inefficiencies in the vm subsystem due to fragmentation in the address space.
59  *
60  * The only really gross cases, with regards to memory waste, are for those
61  * items that are just over half the page size.   You can get nearly 50% waste,
62  * so you fall back to the memory footprint of the power of two allocator. I
63  * have looked at memory allocation sizes on many of the machines available to
64  * me, and there does not seem to be an abundance of allocations at this range
65  * so at this time it may not make sense to optimize for it.  This can, of
66  * course, be solved with dynamic slab sizes.
67  *
68  * Kegs may serve multiple Zones but by far most of the time they only serve
69  * one.  When a Zone is created, a Keg is allocated and setup for it.  While
70  * the backing Keg stores slabs, the Zone caches Buckets of items allocated
71  * from the slabs.  Each Zone is equipped with an init/fini and ctor/dtor
72  * pair, as well as with its own set of small per-CPU caches, layered above
73  * the Zone's general Bucket cache.
74  *
75  * The PCPU caches are protected by critical sections, and may be accessed
76  * safely only from their associated CPU, while the Zones backed by the same
77  * Keg all share a common Keg lock (to coalesce contention on the backing
78  * slabs).  The backing Keg typically only serves one Zone but in the case of
79  * multiple Zones, one of the Zones is considered the Master Zone and all
80  * Zone-related stats from the Keg are done in the Master Zone.  For an
81  * example of a Multi-Zone setup, refer to the Mbuf allocation code.
82  */
83 
84 /*
85  *	This is the representation for normal (Non OFFPAGE slab)
86  *
87  *	i == item
88  *	s == slab pointer
89  *
90  *	<----------------  Page (UMA_SLAB_SIZE) ------------------>
91  *	___________________________________________________________
92  *     | _  _  _  _  _  _  _  _  _  _  _  _  _  _  _   ___________ |
93  *     ||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i| |slab header||
94  *     ||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_| |___________||
95  *     |___________________________________________________________|
96  *
97  *
98  *	This is an OFFPAGE slab. These can be larger than UMA_SLAB_SIZE.
99  *
100  *	___________________________________________________________
101  *     | _  _  _  _  _  _  _  _  _  _  _  _  _  _  _  _  _  _  _   |
102  *     ||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i|  |
103  *     ||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_|  |
104  *     |___________________________________________________________|
105  *       ___________    ^
106  *	|slab header|   |
107  *	|___________|---*
108  *
109  */
110 
111 #ifndef VM_UMA_INT_H
112 #define VM_UMA_INT_H
113 
114 #define UMA_SLAB_SIZE	PAGE_SIZE	/* How big are our slabs? */
115 #define UMA_SLAB_MASK	(PAGE_SIZE - 1)	/* Mask to get back to the page */
116 #define UMA_SLAB_SHIFT	PAGE_SHIFT	/* Number of bits PAGE_MASK */
117 
118 #define UMA_BOOT_PAGES		64	/* Pages allocated for startup */
119 
120 /* Max waste percentage before going to off page slab management */
121 #define UMA_MAX_WASTE	10
122 
123 /*
124  * I doubt there will be many cases where this is exceeded. This is the initial
125  * size of the hash table for uma_slabs that are managed off page. This hash
126  * does expand by powers of two.  Currently it doesn't get smaller.
127  */
128 #define UMA_HASH_SIZE_INIT	32
129 
130 /*
131  * I should investigate other hashing algorithms.  This should yield a low
132  * number of collisions if the pages are relatively contiguous.
133  */
134 
135 #define UMA_HASH(h, s) ((((uintptr_t)s) >> UMA_SLAB_SHIFT) & (h)->uh_hashmask)
136 
137 #define UMA_HASH_INSERT(h, s, mem)					\
138 		SLIST_INSERT_HEAD(&(h)->uh_slab_hash[UMA_HASH((h),	\
139 		    (mem))], (s), us_hlink)
140 #define UMA_HASH_REMOVE(h, s, mem)					\
141 		SLIST_REMOVE(&(h)->uh_slab_hash[UMA_HASH((h),		\
142 		    (mem))], (s), uma_slab, us_hlink)
143 
144 /* Hash table for freed address -> slab translation */
145 
146 SLIST_HEAD(slabhead, uma_slab);
147 
148 struct uma_hash {
149 	struct slabhead	*uh_slab_hash;	/* Hash table for slabs */
150 	int		uh_hashsize;	/* Current size of the hash table */
151 	int		uh_hashmask;	/* Mask used during hashing */
152 };
153 
154 /*
155  * align field or structure to cache line
156  */
157 #if defined(__amd64__)
158 #define UMA_ALIGN	__aligned(CACHE_LINE_SIZE)
159 #else
160 #define UMA_ALIGN
161 #endif
162 
163 /*
164  * Structures for per cpu queues.
165  */
166 
167 struct uma_bucket {
168 	LIST_ENTRY(uma_bucket)	ub_link;	/* Link into the zone */
169 	int16_t	ub_cnt;				/* Count of free items. */
170 	int16_t	ub_entries;			/* Max items. */
171 	void	*ub_bucket[];			/* actual allocation storage */
172 };
173 
174 typedef struct uma_bucket * uma_bucket_t;
175 
176 struct uma_cache {
177 	uma_bucket_t	uc_freebucket;	/* Bucket we're freeing to */
178 	uma_bucket_t	uc_allocbucket;	/* Bucket to allocate from */
179 	uint64_t	uc_allocs;	/* Count of allocations */
180 	uint64_t	uc_frees;	/* Count of frees */
181 } UMA_ALIGN;
182 
183 typedef struct uma_cache * uma_cache_t;
184 
185 /*
186  * Keg management structure
187  *
188  * TODO: Optimize for cache line size
189  *
190  */
191 struct uma_keg {
192 	struct mtx_padalign	uk_lock;	/* Lock for the keg */
193 	struct uma_hash	uk_hash;
194 
195 	LIST_HEAD(,uma_zone)	uk_zones;	/* Keg's zones */
196 	LIST_HEAD(,uma_slab)	uk_part_slab;	/* partially allocated slabs */
197 	LIST_HEAD(,uma_slab)	uk_free_slab;	/* empty slab list */
198 	LIST_HEAD(,uma_slab)	uk_full_slab;	/* full slabs */
199 
200 	uint32_t	uk_align;	/* Alignment mask */
201 	uint32_t	uk_pages;	/* Total page count */
202 	uint32_t	uk_free;	/* Count of items free in slabs */
203 	uint32_t	uk_reserve;	/* Number of reserved items. */
204 	uint32_t	uk_size;	/* Requested size of each item */
205 	uint32_t	uk_rsize;	/* Real size of each item */
206 	uint32_t	uk_maxpages;	/* Maximum number of pages to alloc */
207 
208 	uma_init	uk_init;	/* Keg's init routine */
209 	uma_fini	uk_fini;	/* Keg's fini routine */
210 	uma_alloc	uk_allocf;	/* Allocation function */
211 	uma_free	uk_freef;	/* Free routine */
212 
213 	u_long		uk_offset;	/* Next free offset from base KVA */
214 	vm_offset_t	uk_kva;		/* Zone base KVA */
215 	uma_zone_t	uk_slabzone;	/* Slab zone backing us, if OFFPAGE */
216 
217 	uint16_t	uk_slabsize;	/* Slab size for this keg */
218 	uint16_t	uk_pgoff;	/* Offset to uma_slab struct */
219 	uint16_t	uk_ppera;	/* pages per allocation from backend */
220 	uint16_t	uk_ipers;	/* Items per slab */
221 	uint32_t	uk_flags;	/* Internal flags */
222 
223 	/* Least used fields go to the last cache line. */
224 	const char	*uk_name;		/* Name of creating zone. */
225 	LIST_ENTRY(uma_keg)	uk_link;	/* List of all kegs */
226 };
227 typedef struct uma_keg	* uma_keg_t;
228 
229 /*
230  * Free bits per-slab.
231  */
232 #define	SLAB_SETSIZE	(PAGE_SIZE / UMA_SMALLEST_UNIT)
233 BITSET_DEFINE(slabbits, SLAB_SETSIZE);
234 
235 /*
236  * The slab structure manages a single contiguous allocation from backing
237  * store and subdivides it into individually allocatable items.
238  */
239 struct uma_slab {
240 	uma_keg_t	us_keg;			/* Keg we live in */
241 	union {
242 		LIST_ENTRY(uma_slab)	_us_link;	/* slabs in zone */
243 		unsigned long	_us_size;	/* Size of allocation */
244 	} us_type;
245 	SLIST_ENTRY(uma_slab)	us_hlink;	/* Link for hash table */
246 	uint8_t		*us_data;		/* First item */
247 	struct slabbits	us_free;		/* Free bitmask. */
248 #ifdef INVARIANTS
249 	struct slabbits	us_debugfree;		/* Debug bitmask. */
250 #endif
251 	uint16_t	us_freecount;		/* How many are free? */
252 	uint8_t		us_flags;		/* Page flags see uma.h */
253 	uint8_t		us_pad;			/* Pad to 32bits, unused. */
254 };
255 
256 #define	us_link	us_type._us_link
257 #define	us_size	us_type._us_size
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		us_head;	/* slab header data */
265 	uint32_t		us_refcnt[0];	/* Actually larger. */
266 };
267 
268 typedef struct uma_slab * uma_slab_t;
269 typedef struct uma_slab_refcnt * uma_slabrefcnt_t;
270 typedef uma_slab_t (*uma_slaballoc)(uma_zone_t, uma_keg_t, int);
271 
272 struct uma_klink {
273 	LIST_ENTRY(uma_klink)	kl_link;
274 	uma_keg_t		kl_keg;
275 };
276 typedef struct uma_klink *uma_klink_t;
277 
278 /*
279  * Zone management structure
280  *
281  * TODO: Optimize for cache line size
282  *
283  */
284 struct uma_zone {
285 	struct mtx_padalign	uz_lock;	/* Lock for the zone */
286 	struct mtx_padalign	*uz_lockptr;
287 	const char		*uz_name;	/* Text name of the zone */
288 
289 	LIST_ENTRY(uma_zone)	uz_link;	/* List of all zones in keg */
290 	LIST_HEAD(,uma_bucket)	uz_buckets;	/* full buckets */
291 
292 	LIST_HEAD(,uma_klink)	uz_kegs;	/* List of kegs. */
293 	struct uma_klink	uz_klink;	/* klink for first keg. */
294 
295 	uma_slaballoc	uz_slab;	/* Allocate a slab from the backend. */
296 	uma_ctor	uz_ctor;	/* Constructor for each allocation */
297 	uma_dtor	uz_dtor;	/* Destructor */
298 	uma_init	uz_init;	/* Initializer for each item */
299 	uma_fini	uz_fini;	/* Finalizer for each item. */
300 	uma_import	uz_import;	/* Import new memory to cache. */
301 	uma_release	uz_release;	/* Release memory from cache. */
302 	void		*uz_arg;	/* Import/release argument. */
303 
304 	uint32_t	uz_flags;	/* Flags inherited from kegs */
305 	uint32_t	uz_size;	/* Size inherited from kegs */
306 
307 	volatile u_long	uz_allocs UMA_ALIGN; /* Total number of allocations */
308 	volatile u_long	uz_fails;	/* Total number of alloc failures */
309 	volatile u_long	uz_frees;	/* Total number of frees */
310 	uint64_t	uz_sleeps;	/* Total number of alloc sleeps */
311 	uint16_t	uz_count;	/* Highest amount of items in bucket */
312 
313 	/* The next three fields are used to print a rate-limited warnings. */
314 	const char	*uz_warning;	/* Warning to print on failure */
315 	struct timeval	uz_ratecheck;	/* Warnings rate-limiting */
316 
317 	/*
318 	 * This HAS to be the last item because we adjust the zone size
319 	 * based on NCPU and then allocate the space for the zones.
320 	 */
321 	struct uma_cache	uz_cpu[1]; /* Per cpu caches */
322 };
323 
324 /*
325  * These flags must not overlap with the UMA_ZONE flags specified in uma.h.
326  */
327 #define	UMA_ZFLAG_MULTI		0x04000000	/* Multiple kegs in the zone. */
328 #define	UMA_ZFLAG_DRAINING	0x08000000	/* Running zone_drain. */
329 #define	UMA_ZFLAG_BUCKET	0x10000000	/* Bucket zone. */
330 #define UMA_ZFLAG_INTERNAL	0x20000000	/* No offpage no PCPU. */
331 #define UMA_ZFLAG_FULL		0x40000000	/* Reached uz_maxpages */
332 #define UMA_ZFLAG_CACHEONLY	0x80000000	/* Don't ask VM for buckets. */
333 
334 #define	UMA_ZFLAG_INHERIT						\
335     (UMA_ZFLAG_INTERNAL | UMA_ZFLAG_CACHEONLY | UMA_ZFLAG_BUCKET)
336 
337 static inline uma_keg_t
338 zone_first_keg(uma_zone_t zone)
339 {
340 	uma_klink_t klink;
341 
342 	klink = LIST_FIRST(&zone->uz_kegs);
343 	return (klink != NULL) ? klink->kl_keg : NULL;
344 }
345 
346 #undef UMA_ALIGN
347 
348 #ifdef _KERNEL
349 /* Internal prototypes */
350 static __inline uma_slab_t hash_sfind(struct uma_hash *hash, uint8_t *data);
351 void *uma_large_malloc(int size, int wait);
352 void uma_large_free(uma_slab_t slab);
353 
354 /* Lock Macros */
355 
356 #define	KEG_LOCK_INIT(k, lc)					\
357 	do {							\
358 		if ((lc))					\
359 			mtx_init(&(k)->uk_lock, (k)->uk_name,	\
360 			    (k)->uk_name, MTX_DEF | MTX_DUPOK);	\
361 		else						\
362 			mtx_init(&(k)->uk_lock, (k)->uk_name,	\
363 			    "UMA zone", MTX_DEF | MTX_DUPOK);	\
364 	} while (0)
365 
366 #define	KEG_LOCK_FINI(k)	mtx_destroy(&(k)->uk_lock)
367 #define	KEG_LOCK(k)	mtx_lock(&(k)->uk_lock)
368 #define	KEG_UNLOCK(k)	mtx_unlock(&(k)->uk_lock)
369 
370 #define	ZONE_LOCK_INIT(z, lc)					\
371 	do {							\
372 		if ((lc))					\
373 			mtx_init(&(z)->uz_lock, (z)->uz_name,	\
374 			    (z)->uz_name, MTX_DEF | MTX_DUPOK);	\
375 		else						\
376 			mtx_init(&(z)->uz_lock, (z)->uz_name,	\
377 			    "UMA zone", MTX_DEF | MTX_DUPOK);	\
378 	} while (0)
379 
380 #define	ZONE_LOCK(z)	mtx_lock((z)->uz_lockptr)
381 #define	ZONE_TRYLOCK(z)	mtx_trylock((z)->uz_lockptr)
382 #define	ZONE_UNLOCK(z)	mtx_unlock((z)->uz_lockptr)
383 #define	ZONE_LOCK_FINI(z)	mtx_destroy(&(z)->uz_lock)
384 
385 /*
386  * Find a slab within a hash table.  This is used for OFFPAGE zones to lookup
387  * the slab structure.
388  *
389  * Arguments:
390  *	hash  The hash table to search.
391  *	data  The base page of the item.
392  *
393  * Returns:
394  *	A pointer to a slab if successful, else NULL.
395  */
396 static __inline uma_slab_t
397 hash_sfind(struct uma_hash *hash, uint8_t *data)
398 {
399         uma_slab_t slab;
400         int hval;
401 
402         hval = UMA_HASH(hash, data);
403 
404         SLIST_FOREACH(slab, &hash->uh_slab_hash[hval], us_hlink) {
405                 if ((uint8_t *)slab->us_data == data)
406                         return (slab);
407         }
408         return (NULL);
409 }
410 
411 static __inline uma_slab_t
412 vtoslab(vm_offset_t va)
413 {
414 	vm_page_t p;
415 	uma_slab_t slab;
416 
417 	p = PHYS_TO_VM_PAGE(pmap_kextract(va));
418 	slab = (uma_slab_t )p->object;
419 
420 	if (p->flags & PG_SLAB)
421 		return (slab);
422 	else
423 		return (NULL);
424 }
425 
426 static __inline void
427 vsetslab(vm_offset_t va, uma_slab_t slab)
428 {
429 	vm_page_t p;
430 
431 	p = PHYS_TO_VM_PAGE(pmap_kextract(va));
432 	p->object = (vm_object_t)slab;
433 	p->flags |= PG_SLAB;
434 }
435 
436 static __inline void
437 vsetobj(vm_offset_t va, vm_object_t obj)
438 {
439 	vm_page_t p;
440 
441 	p = PHYS_TO_VM_PAGE(pmap_kextract(va));
442 	p->object = obj;
443 	p->flags &= ~PG_SLAB;
444 }
445 
446 /*
447  * The following two functions may be defined by architecture specific code
448  * if they can provide more effecient allocation functions.  This is useful
449  * for using direct mapped addresses.
450  */
451 void *uma_small_alloc(uma_zone_t zone, int bytes, uint8_t *pflag, int wait);
452 void uma_small_free(void *mem, int size, uint8_t flags);
453 #endif /* _KERNEL */
454 
455 #endif /* VM_UMA_INT_H */
456