xref: /freebsd/sys/vm/uma_int.h (revision 137a344c6341d1469432e9deb3a25593f96672ad)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2002-2005, 2009, 2013 Jeffrey Roberson <jeff@FreeBSD.org>
5  * Copyright (c) 2004, 2005 Bosko Milekic <bmilekic@FreeBSD.org>
6  * All rights reserved.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice unmodified, this list of conditions, and the following
13  *    disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
19  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
20  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
21  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
22  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
23  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
24  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
25  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
26  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
27  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
28  *
29  * $FreeBSD$
30  *
31  */
32 
33 #include <sys/_bitset.h>
34 #include <sys/_task.h>
35 
36 /*
37  * This file includes definitions, structures, prototypes, and inlines that
38  * should not be used outside of the actual implementation of UMA.
39  */
40 
41 /*
42  * The brief summary;  Zones describe unique allocation types.  Zones are
43  * organized into per-CPU caches which are filled by buckets.  Buckets are
44  * organized according to memory domains.  Buckets are filled from kegs which
45  * are also organized according to memory domains.  Kegs describe a unique
46  * allocation type, backend memory provider, and layout.  Kegs are associated
47  * with one or more zones and zones reference one or more kegs.  Kegs provide
48  * slabs which are virtually contiguous collections of pages.  Each slab is
49  * broken down int one or more items that will satisfy an individual allocation.
50  *
51  * Allocation is satisfied in the following order:
52  * 1) Per-CPU cache
53  * 2) Per-domain cache of buckets
54  * 3) Slab from any of N kegs
55  * 4) Backend page provider
56  *
57  * More detail on individual objects is contained below:
58  *
59  * Kegs contain lists of slabs which are stored in either the full bin, empty
60  * bin, or partially allocated bin, to reduce fragmentation.  They also contain
61  * the user supplied value for size, which is adjusted for alignment purposes
62  * and rsize is the result of that.  The Keg also stores information for
63  * managing a hash of page addresses that maps pages to uma_slab_t structures
64  * for pages that don't have embedded uma_slab_t's.
65  *
66  * Keg slab lists are organized by memory domain to support NUMA allocation
67  * policies.  By default allocations are spread across domains to reduce the
68  * potential for hotspots.  Special keg creation flags may be specified to
69  * prefer location allocation.  However there is no strict enforcement as frees
70  * may happen on any CPU and these are returned to the CPU-local cache
71  * regardless of the originating domain.
72  *
73  * The uma_slab_t may be embedded in a UMA_SLAB_SIZE chunk of memory or it may
74  * be allocated off the page from a special slab zone.  The free list within a
75  * slab is managed with a bitmask.  For item sizes that would yield more than
76  * 10% memory waste we potentially allocate a separate uma_slab_t if this will
77  * improve the number of items per slab that will fit.
78  *
79  * The only really gross cases, with regards to memory waste, are for those
80  * items that are just over half the page size.   You can get nearly 50% waste,
81  * so you fall back to the memory footprint of the power of two allocator. I
82  * have looked at memory allocation sizes on many of the machines available to
83  * me, and there does not seem to be an abundance of allocations at this range
84  * so at this time it may not make sense to optimize for it.  This can, of
85  * course, be solved with dynamic slab sizes.
86  *
87  * Kegs may serve multiple Zones but by far most of the time they only serve
88  * one.  When a Zone is created, a Keg is allocated and setup for it.  While
89  * the backing Keg stores slabs, the Zone caches Buckets of items allocated
90  * from the slabs.  Each Zone is equipped with an init/fini and ctor/dtor
91  * pair, as well as with its own set of small per-CPU caches, layered above
92  * the Zone's general Bucket cache.
93  *
94  * The PCPU caches are protected by critical sections, and may be accessed
95  * safely only from their associated CPU, while the Zones backed by the same
96  * Keg all share a common Keg lock (to coalesce contention on the backing
97  * slabs).  The backing Keg typically only serves one Zone but in the case of
98  * multiple Zones, one of the Zones is considered the Master Zone and all
99  * Zone-related stats from the Keg are done in the Master Zone.  For an
100  * example of a Multi-Zone setup, refer to the Mbuf allocation code.
101  */
102 
103 /*
104  *	This is the representation for normal (Non OFFPAGE slab)
105  *
106  *	i == item
107  *	s == slab pointer
108  *
109  *	<----------------  Page (UMA_SLAB_SIZE) ------------------>
110  *	___________________________________________________________
111  *     | _  _  _  _  _  _  _  _  _  _  _  _  _  _  _   ___________ |
112  *     ||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i| |slab header||
113  *     ||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_| |___________||
114  *     |___________________________________________________________|
115  *
116  *
117  *	This is an OFFPAGE slab. These can be larger than UMA_SLAB_SIZE.
118  *
119  *	___________________________________________________________
120  *     | _  _  _  _  _  _  _  _  _  _  _  _  _  _  _  _  _  _  _   |
121  *     ||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i|  |
122  *     ||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_|  |
123  *     |___________________________________________________________|
124  *       ___________    ^
125  *	|slab header|   |
126  *	|___________|---*
127  *
128  */
129 
130 #ifndef VM_UMA_INT_H
131 #define VM_UMA_INT_H
132 
133 #define UMA_SLAB_SIZE	PAGE_SIZE	/* How big are our slabs? */
134 #define UMA_SLAB_MASK	(PAGE_SIZE - 1)	/* Mask to get back to the page */
135 #define UMA_SLAB_SHIFT	PAGE_SHIFT	/* Number of bits PAGE_MASK */
136 
137 #define UMA_BOOT_PAGES		64	/* Pages allocated for startup */
138 #define UMA_BOOT_PAGES_ZONES	32	/* Multiplier for pages to reserve */
139 					/* if uma_zone > PAGE_SIZE */
140 
141 /* Max waste percentage before going to off page slab management */
142 #define UMA_MAX_WASTE	10
143 
144 /*
145  * I doubt there will be many cases where this is exceeded. This is the initial
146  * size of the hash table for uma_slabs that are managed off page. This hash
147  * does expand by powers of two.  Currently it doesn't get smaller.
148  */
149 #define UMA_HASH_SIZE_INIT	32
150 
151 /*
152  * I should investigate other hashing algorithms.  This should yield a low
153  * number of collisions if the pages are relatively contiguous.
154  */
155 
156 #define UMA_HASH(h, s) ((((uintptr_t)s) >> UMA_SLAB_SHIFT) & (h)->uh_hashmask)
157 
158 #define UMA_HASH_INSERT(h, s, mem)					\
159 		SLIST_INSERT_HEAD(&(h)->uh_slab_hash[UMA_HASH((h),	\
160 		    (mem))], (s), us_hlink)
161 #define UMA_HASH_REMOVE(h, s, mem)					\
162 		SLIST_REMOVE(&(h)->uh_slab_hash[UMA_HASH((h),		\
163 		    (mem))], (s), uma_slab, us_hlink)
164 
165 /* Hash table for freed address -> slab translation */
166 
167 SLIST_HEAD(slabhead, uma_slab);
168 
169 struct uma_hash {
170 	struct slabhead	*uh_slab_hash;	/* Hash table for slabs */
171 	int		uh_hashsize;	/* Current size of the hash table */
172 	int		uh_hashmask;	/* Mask used during hashing */
173 };
174 
175 /*
176  * align field or structure to cache line
177  */
178 #if defined(__amd64__)
179 #define UMA_ALIGN	__aligned(CACHE_LINE_SIZE)
180 #else
181 #define UMA_ALIGN
182 #endif
183 
184 /*
185  * Structures for per cpu queues.
186  */
187 
188 struct uma_bucket {
189 	LIST_ENTRY(uma_bucket)	ub_link;	/* Link into the zone */
190 	int16_t	ub_cnt;				/* Count of free items. */
191 	int16_t	ub_entries;			/* Max items. */
192 	void	*ub_bucket[];			/* actual allocation storage */
193 };
194 
195 typedef struct uma_bucket * uma_bucket_t;
196 
197 struct uma_cache {
198 	uma_bucket_t	uc_freebucket;	/* Bucket we're freeing to */
199 	uma_bucket_t	uc_allocbucket;	/* Bucket to allocate from */
200 	uint64_t	uc_allocs;	/* Count of allocations */
201 	uint64_t	uc_frees;	/* Count of frees */
202 } UMA_ALIGN;
203 
204 typedef struct uma_cache * uma_cache_t;
205 
206 /*
207  * Per-domain memory list.  Embedded in the kegs.
208  */
209 struct uma_domain {
210 	LIST_HEAD(,uma_slab)	ud_part_slab;	/* partially allocated slabs */
211 	LIST_HEAD(,uma_slab)	ud_free_slab;	/* empty slab list */
212 	LIST_HEAD(,uma_slab)	ud_full_slab;	/* full slabs */
213 };
214 
215 typedef struct uma_domain * uma_domain_t;
216 
217 /*
218  * Keg management structure
219  *
220  * TODO: Optimize for cache line size
221  *
222  */
223 struct uma_keg {
224 	struct mtx_padalign	uk_lock;	/* Lock for the keg */
225 	struct uma_hash	uk_hash;
226 
227 	LIST_HEAD(,uma_zone)	uk_zones;	/* Keg's zones */
228 
229 	uint32_t	uk_cursor;	/* Domain alloc cursor. */
230 	uint32_t	uk_align;	/* Alignment mask */
231 	uint32_t	uk_pages;	/* Total page count */
232 	uint32_t	uk_free;	/* Count of items free in slabs */
233 	uint32_t	uk_reserve;	/* Number of reserved items. */
234 	uint32_t	uk_size;	/* Requested size of each item */
235 	uint32_t	uk_rsize;	/* Real size of each item */
236 	uint32_t	uk_maxpages;	/* Maximum number of pages to alloc */
237 
238 	uma_init	uk_init;	/* Keg's init routine */
239 	uma_fini	uk_fini;	/* Keg's fini routine */
240 	uma_alloc	uk_allocf;	/* Allocation function */
241 	uma_free	uk_freef;	/* Free routine */
242 
243 	u_long		uk_offset;	/* Next free offset from base KVA */
244 	vm_offset_t	uk_kva;		/* Zone base KVA */
245 	uma_zone_t	uk_slabzone;	/* Slab zone backing us, if OFFPAGE */
246 
247 	uint32_t	uk_pgoff;	/* Offset to uma_slab struct */
248 	uint16_t	uk_ppera;	/* pages per allocation from backend */
249 	uint16_t	uk_ipers;	/* Items per slab */
250 	uint32_t	uk_flags;	/* Internal flags */
251 
252 	/* Least used fields go to the last cache line. */
253 	const char	*uk_name;		/* Name of creating zone. */
254 	LIST_ENTRY(uma_keg)	uk_link;	/* List of all kegs */
255 
256 	/* Must be last, variable sized. */
257 	struct uma_domain	uk_domain[];	/* Keg's slab lists. */
258 };
259 typedef struct uma_keg	* uma_keg_t;
260 
261 /*
262  * Free bits per-slab.
263  */
264 #define	SLAB_SETSIZE	(PAGE_SIZE / UMA_SMALLEST_UNIT)
265 BITSET_DEFINE(slabbits, SLAB_SETSIZE);
266 
267 /*
268  * The slab structure manages a single contiguous allocation from backing
269  * store and subdivides it into individually allocatable items.
270  */
271 struct uma_slab {
272 	uma_keg_t	us_keg;			/* Keg we live in */
273 	union {
274 		LIST_ENTRY(uma_slab)	_us_link;	/* slabs in zone */
275 		unsigned long	_us_size;	/* Size of allocation */
276 	} us_type;
277 	SLIST_ENTRY(uma_slab)	us_hlink;	/* Link for hash table */
278 	uint8_t		*us_data;		/* First item */
279 	struct slabbits	us_free;		/* Free bitmask. */
280 #ifdef INVARIANTS
281 	struct slabbits	us_debugfree;		/* Debug bitmask. */
282 #endif
283 	uint16_t	us_freecount;		/* How many are free? */
284 	uint8_t		us_flags;		/* Page flags see uma.h */
285 	uint8_t		us_domain;		/* Backing NUMA domain. */
286 };
287 
288 #define	us_link	us_type._us_link
289 #define	us_size	us_type._us_size
290 
291 #if MAXMEMDOM >= 255
292 #error "Slab domain type insufficient"
293 #endif
294 
295 typedef struct uma_slab * uma_slab_t;
296 typedef uma_slab_t (*uma_slaballoc)(uma_zone_t, uma_keg_t, int, int);
297 
298 struct uma_klink {
299 	LIST_ENTRY(uma_klink)	kl_link;
300 	uma_keg_t		kl_keg;
301 };
302 typedef struct uma_klink *uma_klink_t;
303 
304 struct uma_zone_domain {
305 	LIST_HEAD(,uma_bucket)	uzd_buckets;	/* full buckets */
306 };
307 
308 typedef struct uma_zone_domain * uma_zone_domain_t;
309 
310 /*
311  * Zone management structure
312  *
313  * TODO: Optimize for cache line size
314  *
315  */
316 struct uma_zone {
317 	struct mtx_padalign	uz_lock;	/* Lock for the zone */
318 	struct mtx_padalign	*uz_lockptr;
319 	const char		*uz_name;	/* Text name of the zone */
320 
321 	LIST_ENTRY(uma_zone)	uz_link;	/* List of all zones in keg */
322 	struct uma_zone_domain	*uz_domain;	/* per-domain buckets */
323 
324 	LIST_HEAD(,uma_klink)	uz_kegs;	/* List of kegs. */
325 	struct uma_klink	uz_klink;	/* klink for first keg. */
326 
327 	uma_slaballoc	uz_slab;	/* Allocate a slab from the backend. */
328 	uma_ctor	uz_ctor;	/* Constructor for each allocation */
329 	uma_dtor	uz_dtor;	/* Destructor */
330 	uma_init	uz_init;	/* Initializer for each item */
331 	uma_fini	uz_fini;	/* Finalizer for each item. */
332 	uma_import	uz_import;	/* Import new memory to cache. */
333 	uma_release	uz_release;	/* Release memory from cache. */
334 	void		*uz_arg;	/* Import/release argument. */
335 
336 	uint32_t	uz_flags;	/* Flags inherited from kegs */
337 	uint32_t	uz_size;	/* Size inherited from kegs */
338 
339 	volatile u_long	uz_allocs UMA_ALIGN; /* Total number of allocations */
340 	volatile u_long	uz_fails;	/* Total number of alloc failures */
341 	volatile u_long	uz_frees;	/* Total number of frees */
342 	uint64_t	uz_sleeps;	/* Total number of alloc sleeps */
343 	uint16_t	uz_count;	/* Amount of items in full bucket */
344 	uint16_t	uz_count_min;	/* Minimal amount of items there */
345 
346 	/* The next two fields are used to print a rate-limited warnings. */
347 	const char	*uz_warning;	/* Warning to print on failure */
348 	struct timeval	uz_ratecheck;	/* Warnings rate-limiting */
349 
350 	struct task	uz_maxaction;	/* Task to run when at limit */
351 
352 	/*
353 	 * This HAS to be the last item because we adjust the zone size
354 	 * based on NCPU and then allocate the space for the zones.
355 	 */
356 	struct uma_cache	uz_cpu[]; /* Per cpu caches */
357 
358 	/* uz_domain follows here. */
359 };
360 
361 /*
362  * These flags must not overlap with the UMA_ZONE flags specified in uma.h.
363  */
364 #define	UMA_ZFLAG_MULTI		0x04000000	/* Multiple kegs in the zone. */
365 #define	UMA_ZFLAG_DRAINING	0x08000000	/* Running zone_drain. */
366 #define	UMA_ZFLAG_BUCKET	0x10000000	/* Bucket zone. */
367 #define UMA_ZFLAG_INTERNAL	0x20000000	/* No offpage no PCPU. */
368 #define UMA_ZFLAG_FULL		0x40000000	/* Reached uz_maxpages */
369 #define UMA_ZFLAG_CACHEONLY	0x80000000	/* Don't ask VM for buckets. */
370 
371 #define	UMA_ZFLAG_INHERIT						\
372     (UMA_ZFLAG_INTERNAL | UMA_ZFLAG_CACHEONLY | UMA_ZFLAG_BUCKET)
373 
374 static inline uma_keg_t
375 zone_first_keg(uma_zone_t zone)
376 {
377 	uma_klink_t klink;
378 
379 	klink = LIST_FIRST(&zone->uz_kegs);
380 	return (klink != NULL) ? klink->kl_keg : NULL;
381 }
382 
383 #undef UMA_ALIGN
384 
385 #ifdef _KERNEL
386 /* Internal prototypes */
387 static __inline uma_slab_t hash_sfind(struct uma_hash *hash, uint8_t *data);
388 void *uma_large_malloc(vm_size_t size, int wait);
389 void *uma_large_malloc_domain(vm_size_t size, int domain, int wait);
390 void uma_large_free(uma_slab_t slab);
391 
392 /* Lock Macros */
393 
394 #define	KEG_LOCK_INIT(k, lc)					\
395 	do {							\
396 		if ((lc))					\
397 			mtx_init(&(k)->uk_lock, (k)->uk_name,	\
398 			    (k)->uk_name, MTX_DEF | MTX_DUPOK);	\
399 		else						\
400 			mtx_init(&(k)->uk_lock, (k)->uk_name,	\
401 			    "UMA zone", MTX_DEF | MTX_DUPOK);	\
402 	} while (0)
403 
404 #define	KEG_LOCK_FINI(k)	mtx_destroy(&(k)->uk_lock)
405 #define	KEG_LOCK(k)	mtx_lock(&(k)->uk_lock)
406 #define	KEG_UNLOCK(k)	mtx_unlock(&(k)->uk_lock)
407 
408 #define	ZONE_LOCK_INIT(z, lc)					\
409 	do {							\
410 		if ((lc))					\
411 			mtx_init(&(z)->uz_lock, (z)->uz_name,	\
412 			    (z)->uz_name, MTX_DEF | MTX_DUPOK);	\
413 		else						\
414 			mtx_init(&(z)->uz_lock, (z)->uz_name,	\
415 			    "UMA zone", MTX_DEF | MTX_DUPOK);	\
416 	} while (0)
417 
418 #define	ZONE_LOCK(z)	mtx_lock((z)->uz_lockptr)
419 #define	ZONE_TRYLOCK(z)	mtx_trylock((z)->uz_lockptr)
420 #define	ZONE_UNLOCK(z)	mtx_unlock((z)->uz_lockptr)
421 #define	ZONE_LOCK_FINI(z)	mtx_destroy(&(z)->uz_lock)
422 
423 /*
424  * Find a slab within a hash table.  This is used for OFFPAGE zones to lookup
425  * the slab structure.
426  *
427  * Arguments:
428  *	hash  The hash table to search.
429  *	data  The base page of the item.
430  *
431  * Returns:
432  *	A pointer to a slab if successful, else NULL.
433  */
434 static __inline uma_slab_t
435 hash_sfind(struct uma_hash *hash, uint8_t *data)
436 {
437         uma_slab_t slab;
438         int hval;
439 
440         hval = UMA_HASH(hash, data);
441 
442         SLIST_FOREACH(slab, &hash->uh_slab_hash[hval], us_hlink) {
443                 if ((uint8_t *)slab->us_data == data)
444                         return (slab);
445         }
446         return (NULL);
447 }
448 
449 static __inline uma_slab_t
450 vtoslab(vm_offset_t va)
451 {
452 	vm_page_t p;
453 
454 	p = PHYS_TO_VM_PAGE(pmap_kextract(va));
455 	return ((uma_slab_t)p->plinks.s.pv);
456 }
457 
458 static __inline void
459 vsetslab(vm_offset_t va, uma_slab_t slab)
460 {
461 	vm_page_t p;
462 
463 	p = PHYS_TO_VM_PAGE(pmap_kextract(va));
464 	p->plinks.s.pv = slab;
465 }
466 
467 /*
468  * The following two functions may be defined by architecture specific code
469  * if they can provide more efficient allocation functions.  This is useful
470  * for using direct mapped addresses.
471  */
472 void *uma_small_alloc(uma_zone_t zone, vm_size_t bytes, int domain,
473     uint8_t *pflag, int wait);
474 void uma_small_free(void *mem, vm_size_t size, uint8_t flags);
475 
476 /* Set a global soft limit on UMA managed memory. */
477 void uma_set_limit(unsigned long limit);
478 #endif /* _KERNEL */
479 
480 #endif /* VM_UMA_INT_H */
481