xref: /freebsd/sys/vm/uma_int.h (revision 77ebcc05eac2658a68b447e654cfdf7ff3e703b8)
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/_domainset.h>
35 #include <sys/_task.h>
36 
37 /*
38  * This file includes definitions, structures, prototypes, and inlines that
39  * should not be used outside of the actual implementation of UMA.
40  */
41 
42 /*
43  * The brief summary;  Zones describe unique allocation types.  Zones are
44  * organized into per-CPU caches which are filled by buckets.  Buckets are
45  * organized according to memory domains.  Buckets are filled from kegs which
46  * are also organized according to memory domains.  Kegs describe a unique
47  * allocation type, backend memory provider, and layout.  Kegs are associated
48  * with one or more zones and zones reference one or more kegs.  Kegs provide
49  * slabs which are virtually contiguous collections of pages.  Each slab is
50  * broken down int one or more items that will satisfy an individual allocation.
51  *
52  * Allocation is satisfied in the following order:
53  * 1) Per-CPU cache
54  * 2) Per-domain cache of buckets
55  * 3) Slab from any of N kegs
56  * 4) Backend page provider
57  *
58  * More detail on individual objects is contained below:
59  *
60  * Kegs contain lists of slabs which are stored in either the full bin, empty
61  * bin, or partially allocated bin, to reduce fragmentation.  They also contain
62  * the user supplied value for size, which is adjusted for alignment purposes
63  * and rsize is the result of that.  The Keg also stores information for
64  * managing a hash of page addresses that maps pages to uma_slab_t structures
65  * for pages that don't have embedded uma_slab_t's.
66  *
67  * Keg slab lists are organized by memory domain to support NUMA allocation
68  * policies.  By default allocations are spread across domains to reduce the
69  * potential for hotspots.  Special keg creation flags may be specified to
70  * prefer location allocation.  However there is no strict enforcement as frees
71  * may happen on any CPU and these are returned to the CPU-local cache
72  * regardless of the originating domain.
73  *
74  * The uma_slab_t may be embedded in a UMA_SLAB_SIZE chunk of memory or it may
75  * be allocated off the page from a special slab zone.  The free list within a
76  * slab is managed with a bitmask.  For item sizes that would yield more than
77  * 10% memory waste we potentially allocate a separate uma_slab_t if this will
78  * improve the number of items per slab that will fit.
79  *
80  * The only really gross cases, with regards to memory waste, are for those
81  * items that are just over half the page size.   You can get nearly 50% waste,
82  * so you fall back to the memory footprint of the power of two allocator. I
83  * have looked at memory allocation sizes on many of the machines available to
84  * me, and there does not seem to be an abundance of allocations at this range
85  * so at this time it may not make sense to optimize for it.  This can, of
86  * course, be solved with dynamic slab sizes.
87  *
88  * Kegs may serve multiple Zones but by far most of the time they only serve
89  * one.  When a Zone is created, a Keg is allocated and setup for it.  While
90  * the backing Keg stores slabs, the Zone caches Buckets of items allocated
91  * from the slabs.  Each Zone is equipped with an init/fini and ctor/dtor
92  * pair, as well as with its own set of small per-CPU caches, layered above
93  * the Zone's general Bucket cache.
94  *
95  * The PCPU caches are protected by critical sections, and may be accessed
96  * safely only from their associated CPU, while the Zones backed by the same
97  * Keg all share a common Keg lock (to coalesce contention on the backing
98  * slabs).  The backing Keg typically only serves one Zone but in the case of
99  * multiple Zones, one of the Zones is considered the Master Zone and all
100  * Zone-related stats from the Keg are done in the Master Zone.  For an
101  * example of a Multi-Zone setup, refer to the Mbuf allocation code.
102  */
103 
104 /*
105  *	This is the representation for normal (Non OFFPAGE slab)
106  *
107  *	i == item
108  *	s == slab pointer
109  *
110  *	<----------------  Page (UMA_SLAB_SIZE) ------------------>
111  *	___________________________________________________________
112  *     | _  _  _  _  _  _  _  _  _  _  _  _  _  _  _   ___________ |
113  *     ||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i| |slab header||
114  *     ||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_| |___________||
115  *     |___________________________________________________________|
116  *
117  *
118  *	This is an OFFPAGE slab. These can be larger than UMA_SLAB_SIZE.
119  *
120  *	___________________________________________________________
121  *     | _  _  _  _  _  _  _  _  _  _  _  _  _  _  _  _  _  _  _   |
122  *     ||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i|  |
123  *     ||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_|  |
124  *     |___________________________________________________________|
125  *       ___________    ^
126  *	|slab header|   |
127  *	|___________|---*
128  *
129  */
130 
131 #ifndef VM_UMA_INT_H
132 #define VM_UMA_INT_H
133 
134 #define UMA_SLAB_SIZE	PAGE_SIZE	/* How big are our slabs? */
135 #define UMA_SLAB_MASK	(PAGE_SIZE - 1)	/* Mask to get back to the page */
136 #define UMA_SLAB_SHIFT	PAGE_SHIFT	/* Number of bits PAGE_MASK */
137 
138 /* Max waste percentage before going to off page slab management */
139 #define UMA_MAX_WASTE	10
140 
141 /*
142  * Actual size of uma_slab when it is placed at an end of a page
143  * with pointer sized alignment requirement.
144  */
145 #define	SIZEOF_UMA_SLAB	((sizeof(struct uma_slab) & UMA_ALIGN_PTR) ?	  \
146 			    (sizeof(struct uma_slab) & ~UMA_ALIGN_PTR) +  \
147 			    (UMA_ALIGN_PTR + 1) : sizeof(struct uma_slab))
148 
149 /*
150  * Size of memory in a not offpage single page slab available for actual items.
151  */
152 #define	UMA_SLAB_SPACE	(PAGE_SIZE - SIZEOF_UMA_SLAB)
153 
154 /*
155  * I doubt there will be many cases where this is exceeded. This is the initial
156  * size of the hash table for uma_slabs that are managed off page. This hash
157  * does expand by powers of two.  Currently it doesn't get smaller.
158  */
159 #define UMA_HASH_SIZE_INIT	32
160 
161 /*
162  * I should investigate other hashing algorithms.  This should yield a low
163  * number of collisions if the pages are relatively contiguous.
164  */
165 
166 #define UMA_HASH(h, s) ((((uintptr_t)s) >> UMA_SLAB_SHIFT) & (h)->uh_hashmask)
167 
168 #define UMA_HASH_INSERT(h, s, mem)					\
169 		SLIST_INSERT_HEAD(&(h)->uh_slab_hash[UMA_HASH((h),	\
170 		    (mem))], (s), us_hlink)
171 #define UMA_HASH_REMOVE(h, s, mem)					\
172 		SLIST_REMOVE(&(h)->uh_slab_hash[UMA_HASH((h),		\
173 		    (mem))], (s), uma_slab, us_hlink)
174 
175 /* Hash table for freed address -> slab translation */
176 
177 SLIST_HEAD(slabhead, uma_slab);
178 
179 struct uma_hash {
180 	struct slabhead	*uh_slab_hash;	/* Hash table for slabs */
181 	int		uh_hashsize;	/* Current size of the hash table */
182 	int		uh_hashmask;	/* Mask used during hashing */
183 };
184 
185 /*
186  * align field or structure to cache line
187  */
188 #if defined(__amd64__) || defined(__powerpc64__)
189 #define UMA_ALIGN	__aligned(128)
190 #else
191 #define UMA_ALIGN
192 #endif
193 
194 /*
195  * Structures for per cpu queues.
196  */
197 
198 struct uma_bucket {
199 	LIST_ENTRY(uma_bucket)	ub_link;	/* Link into the zone */
200 	int16_t	ub_cnt;				/* Count of items in bucket. */
201 	int16_t	ub_entries;			/* Max items. */
202 	void	*ub_bucket[];			/* actual allocation storage */
203 };
204 
205 typedef struct uma_bucket * uma_bucket_t;
206 
207 struct uma_cache {
208 	uma_bucket_t	uc_freebucket;	/* Bucket we're freeing to */
209 	uma_bucket_t	uc_allocbucket;	/* Bucket to allocate from */
210 	uint64_t	uc_allocs;	/* Count of allocations */
211 	uint64_t	uc_frees;	/* Count of frees */
212 } UMA_ALIGN;
213 
214 typedef struct uma_cache * uma_cache_t;
215 
216 /*
217  * Per-domain memory list.  Embedded in the kegs.
218  */
219 struct uma_domain {
220 	LIST_HEAD(,uma_slab)	ud_part_slab;	/* partially allocated slabs */
221 	LIST_HEAD(,uma_slab)	ud_free_slab;	/* empty slab list */
222 	LIST_HEAD(,uma_slab)	ud_full_slab;	/* full slabs */
223 };
224 
225 typedef struct uma_domain * uma_domain_t;
226 
227 /*
228  * Keg management structure
229  *
230  * TODO: Optimize for cache line size
231  *
232  */
233 struct uma_keg {
234 	struct mtx	uk_lock;	/* Lock for the keg */
235 	struct uma_hash	uk_hash;
236 	LIST_HEAD(,uma_zone)	uk_zones;	/* Keg's zones */
237 
238 	struct domainset_ref uk_dr;	/* Domain selection policy. */
239 	uint32_t	uk_align;	/* Alignment mask */
240 	uint32_t	uk_pages;	/* Total page count */
241 	uint32_t	uk_free;	/* Count of items free in slabs */
242 	uint32_t	uk_reserve;	/* Number of reserved items. */
243 	uint32_t	uk_size;	/* Requested size of each item */
244 	uint32_t	uk_rsize;	/* Real size of each item */
245 	uint32_t	uk_maxpages;	/* Maximum number of pages to alloc */
246 
247 	uma_init	uk_init;	/* Keg's init routine */
248 	uma_fini	uk_fini;	/* Keg's fini routine */
249 	uma_alloc	uk_allocf;	/* Allocation function */
250 	uma_free	uk_freef;	/* Free routine */
251 
252 	u_long		uk_offset;	/* Next free offset from base KVA */
253 	vm_offset_t	uk_kva;		/* Zone base KVA */
254 	uma_zone_t	uk_slabzone;	/* Slab zone backing us, if OFFPAGE */
255 
256 	uint32_t	uk_pgoff;	/* Offset to uma_slab struct */
257 	uint16_t	uk_ppera;	/* pages per allocation from backend */
258 	uint16_t	uk_ipers;	/* Items per slab */
259 	uint32_t	uk_flags;	/* Internal flags */
260 
261 	/* Least used fields go to the last cache line. */
262 	const char	*uk_name;		/* Name of creating zone. */
263 	LIST_ENTRY(uma_keg)	uk_link;	/* List of all kegs */
264 
265 	/* Must be last, variable sized. */
266 	struct uma_domain	uk_domain[];	/* Keg's slab lists. */
267 };
268 typedef struct uma_keg	* uma_keg_t;
269 
270 /*
271  * Free bits per-slab.
272  */
273 #define	SLAB_SETSIZE	(PAGE_SIZE / UMA_SMALLEST_UNIT)
274 BITSET_DEFINE(slabbits, SLAB_SETSIZE);
275 
276 /*
277  * The slab structure manages a single contiguous allocation from backing
278  * store and subdivides it into individually allocatable items.
279  */
280 struct uma_slab {
281 	uma_keg_t	us_keg;			/* Keg we live in */
282 	union {
283 		LIST_ENTRY(uma_slab)	_us_link;	/* slabs in zone */
284 		unsigned long	_us_size;	/* Size of allocation */
285 	} us_type;
286 	SLIST_ENTRY(uma_slab)	us_hlink;	/* Link for hash table */
287 	uint8_t		*us_data;		/* First item */
288 	struct slabbits	us_free;		/* Free bitmask. */
289 #ifdef INVARIANTS
290 	struct slabbits	us_debugfree;		/* Debug bitmask. */
291 #endif
292 	uint16_t	us_freecount;		/* How many are free? */
293 	uint8_t		us_flags;		/* Page flags see uma.h */
294 	uint8_t		us_domain;		/* Backing NUMA domain. */
295 };
296 
297 #define	us_link	us_type._us_link
298 #define	us_size	us_type._us_size
299 
300 #if MAXMEMDOM >= 255
301 #error "Slab domain type insufficient"
302 #endif
303 
304 typedef struct uma_slab * uma_slab_t;
305 typedef uma_slab_t (*uma_slaballoc)(uma_zone_t, uma_keg_t, int, int);
306 
307 struct uma_klink {
308 	LIST_ENTRY(uma_klink)	kl_link;
309 	uma_keg_t		kl_keg;
310 };
311 typedef struct uma_klink *uma_klink_t;
312 
313 struct uma_zone_domain {
314 	LIST_HEAD(,uma_bucket)	uzd_buckets;	/* full buckets */
315 	long		uzd_nitems;	/* total item count */
316 	long		uzd_imax;	/* maximum item count this period */
317 	long		uzd_imin;	/* minimum item count this period */
318 	long		uzd_wss;	/* working set size estimate */
319 };
320 
321 typedef struct uma_zone_domain * uma_zone_domain_t;
322 
323 /*
324  * Zone management structure
325  *
326  * TODO: Optimize for cache line size
327  *
328  */
329 struct uma_zone {
330 	/* Offset 0, used in alloc/free fast/medium fast path and const. */
331 	struct mtx	*uz_lockptr;
332 	const char	*uz_name;	/* Text name of the zone */
333 	struct uma_zone_domain	*uz_domain;	/* per-domain buckets */
334 	uint32_t	uz_flags;	/* Flags inherited from kegs */
335 	uint32_t	uz_size;	/* Size inherited from kegs */
336 	uma_ctor	uz_ctor;	/* Constructor for each allocation */
337 	uma_dtor	uz_dtor;	/* Destructor */
338 	uma_init	uz_init;	/* Initializer for each item */
339 	uma_fini	uz_fini;	/* Finalizer for each item. */
340 
341 	/* Offset 64, used in bucket replenish. */
342 	uma_import	uz_import;	/* Import new memory to cache. */
343 	uma_release	uz_release;	/* Release memory from cache. */
344 	void		*uz_arg;	/* Import/release argument. */
345 	uma_slaballoc	uz_slab;	/* Allocate a slab from the backend. */
346 	uint16_t	uz_count;	/* Amount of items in full bucket */
347 	uint16_t	uz_count_min;	/* Minimal amount of items there */
348 	/* 32bit pad on 64bit. */
349 	LIST_ENTRY(uma_zone)	uz_link;	/* List of all zones in keg */
350 	LIST_HEAD(,uma_klink)	uz_kegs;	/* List of kegs. */
351 
352 	/* Offset 128 Rare. */
353 	/*
354 	 * The lock is placed here to avoid adjacent line prefetcher
355 	 * in fast paths and to take up space near infrequently accessed
356 	 * members to reduce alignment overhead.
357 	 */
358 	struct mtx	uz_lock;	/* Lock for the zone */
359 	struct uma_klink	uz_klink;	/* klink for first keg. */
360 	/* The next two fields are used to print a rate-limited warnings. */
361 	const char	*uz_warning;	/* Warning to print on failure */
362 	struct timeval	uz_ratecheck;	/* Warnings rate-limiting */
363 	struct task	uz_maxaction;	/* Task to run when at limit */
364 
365 	/* 16 bytes of pad. */
366 
367 	/* Offset 256, atomic stats. */
368 	volatile u_long	uz_allocs UMA_ALIGN; /* Total number of allocations */
369 	volatile u_long	uz_fails;	/* Total number of alloc failures */
370 	volatile u_long	uz_frees;	/* Total number of frees */
371 	uint64_t	uz_sleeps;	/* Total number of alloc sleeps */
372 
373 	/*
374 	 * This HAS to be the last item because we adjust the zone size
375 	 * based on NCPU and then allocate the space for the zones.
376 	 */
377 	struct uma_cache	uz_cpu[]; /* Per cpu caches */
378 
379 	/* uz_domain follows here. */
380 };
381 
382 /*
383  * These flags must not overlap with the UMA_ZONE flags specified in uma.h.
384  */
385 #define	UMA_ZFLAG_MULTI		0x04000000	/* Multiple kegs in the zone. */
386 #define	UMA_ZFLAG_DRAINING	0x08000000	/* Running zone_drain. */
387 #define	UMA_ZFLAG_BUCKET	0x10000000	/* Bucket zone. */
388 #define UMA_ZFLAG_INTERNAL	0x20000000	/* No offpage no PCPU. */
389 #define UMA_ZFLAG_FULL		0x40000000	/* Reached uz_maxpages */
390 #define UMA_ZFLAG_CACHEONLY	0x80000000	/* Don't ask VM for buckets. */
391 
392 #define	UMA_ZFLAG_INHERIT						\
393     (UMA_ZFLAG_INTERNAL | UMA_ZFLAG_CACHEONLY | UMA_ZFLAG_BUCKET)
394 
395 static inline uma_keg_t
396 zone_first_keg(uma_zone_t zone)
397 {
398 	uma_klink_t klink;
399 
400 	klink = LIST_FIRST(&zone->uz_kegs);
401 	return (klink != NULL) ? klink->kl_keg : NULL;
402 }
403 
404 #undef UMA_ALIGN
405 
406 #ifdef _KERNEL
407 /* Internal prototypes */
408 static __inline uma_slab_t hash_sfind(struct uma_hash *hash, uint8_t *data);
409 void *uma_large_malloc(vm_size_t size, int wait);
410 void *uma_large_malloc_domain(vm_size_t size, int domain, int wait);
411 void uma_large_free(uma_slab_t slab);
412 
413 /* Lock Macros */
414 
415 #define	KEG_LOCK_INIT(k, lc)					\
416 	do {							\
417 		if ((lc))					\
418 			mtx_init(&(k)->uk_lock, (k)->uk_name,	\
419 			    (k)->uk_name, MTX_DEF | MTX_DUPOK);	\
420 		else						\
421 			mtx_init(&(k)->uk_lock, (k)->uk_name,	\
422 			    "UMA zone", MTX_DEF | MTX_DUPOK);	\
423 	} while (0)
424 
425 #define	KEG_LOCK_FINI(k)	mtx_destroy(&(k)->uk_lock)
426 #define	KEG_LOCK(k)	mtx_lock(&(k)->uk_lock)
427 #define	KEG_UNLOCK(k)	mtx_unlock(&(k)->uk_lock)
428 
429 #define	ZONE_LOCK_INIT(z, lc)					\
430 	do {							\
431 		if ((lc))					\
432 			mtx_init(&(z)->uz_lock, (z)->uz_name,	\
433 			    (z)->uz_name, MTX_DEF | MTX_DUPOK);	\
434 		else						\
435 			mtx_init(&(z)->uz_lock, (z)->uz_name,	\
436 			    "UMA zone", MTX_DEF | MTX_DUPOK);	\
437 	} while (0)
438 
439 #define	ZONE_LOCK(z)	mtx_lock((z)->uz_lockptr)
440 #define	ZONE_TRYLOCK(z)	mtx_trylock((z)->uz_lockptr)
441 #define	ZONE_UNLOCK(z)	mtx_unlock((z)->uz_lockptr)
442 #define	ZONE_LOCK_FINI(z)	mtx_destroy(&(z)->uz_lock)
443 #define	ZONE_LOCK_ASSERT(z)	mtx_assert((z)->uz_lockptr, MA_OWNED)
444 
445 /*
446  * Find a slab within a hash table.  This is used for OFFPAGE zones to lookup
447  * the slab structure.
448  *
449  * Arguments:
450  *	hash  The hash table to search.
451  *	data  The base page of the item.
452  *
453  * Returns:
454  *	A pointer to a slab if successful, else NULL.
455  */
456 static __inline uma_slab_t
457 hash_sfind(struct uma_hash *hash, uint8_t *data)
458 {
459         uma_slab_t slab;
460         int hval;
461 
462         hval = UMA_HASH(hash, data);
463 
464         SLIST_FOREACH(slab, &hash->uh_slab_hash[hval], us_hlink) {
465                 if ((uint8_t *)slab->us_data == data)
466                         return (slab);
467         }
468         return (NULL);
469 }
470 
471 static __inline uma_slab_t
472 vtoslab(vm_offset_t va)
473 {
474 	vm_page_t p;
475 
476 	p = PHYS_TO_VM_PAGE(pmap_kextract(va));
477 	return ((uma_slab_t)p->plinks.s.pv);
478 }
479 
480 static __inline void
481 vsetslab(vm_offset_t va, uma_slab_t slab)
482 {
483 	vm_page_t p;
484 
485 	p = PHYS_TO_VM_PAGE(pmap_kextract(va));
486 	p->plinks.s.pv = slab;
487 }
488 
489 /*
490  * The following two functions may be defined by architecture specific code
491  * if they can provide more efficient allocation functions.  This is useful
492  * for using direct mapped addresses.
493  */
494 void *uma_small_alloc(uma_zone_t zone, vm_size_t bytes, int domain,
495     uint8_t *pflag, int wait);
496 void uma_small_free(void *mem, vm_size_t size, uint8_t flags);
497 
498 /* Set a global soft limit on UMA managed memory. */
499 void uma_set_limit(unsigned long limit);
500 #endif /* _KERNEL */
501 
502 #endif /* VM_UMA_INT_H */
503