2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
4 * Copyright (c) 2002-2005, 2009, 2013 Jeffrey Roberson <jeff@FreeBSD.org>
5 * Copyright (c) 2004, 2005 Bosko Milekic <bmilekic@FreeBSD.org>
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice unmodified, this list of conditions, and the following
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.
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.
33 #include <sys/_bitset.h>
34 #include <sys/_task.h>
37 * This file includes definitions, structures, prototypes, and inlines that
38 * should not be used outside of the actual implementation of UMA.
42 * Here's a quick description of the relationship between the objects:
44 * Kegs contain lists of slabs which are stored in either the full bin, empty
45 * bin, or partially allocated bin, to reduce fragmentation. They also contain
46 * the user supplied value for size, which is adjusted for alignment purposes
47 * and rsize is the result of that. The Keg also stores information for
48 * managing a hash of page addresses that maps pages to uma_slab_t structures
49 * for pages that don't have embedded uma_slab_t's.
51 * The uma_slab_t may be embedded in a UMA_SLAB_SIZE chunk of memory or it may
52 * be allocated off the page from a special slab zone. The free list within a
53 * slab is managed with a bitmask. For item sizes that would yield more than
54 * 10% memory waste we potentially allocate a separate uma_slab_t if this will
55 * improve the number of items per slab that will fit.
57 * The only really gross cases, with regards to memory waste, are for those
58 * items that are just over half the page size. You can get nearly 50% waste,
59 * so you fall back to the memory footprint of the power of two allocator. I
60 * have looked at memory allocation sizes on many of the machines available to
61 * me, and there does not seem to be an abundance of allocations at this range
62 * so at this time it may not make sense to optimize for it. This can, of
63 * course, be solved with dynamic slab sizes.
65 * Kegs may serve multiple Zones but by far most of the time they only serve
66 * one. When a Zone is created, a Keg is allocated and setup for it. While
67 * the backing Keg stores slabs, the Zone caches Buckets of items allocated
68 * from the slabs. Each Zone is equipped with an init/fini and ctor/dtor
69 * pair, as well as with its own set of small per-CPU caches, layered above
70 * the Zone's general Bucket cache.
72 * The PCPU caches are protected by critical sections, and may be accessed
73 * safely only from their associated CPU, while the Zones backed by the same
74 * Keg all share a common Keg lock (to coalesce contention on the backing
75 * slabs). The backing Keg typically only serves one Zone but in the case of
76 * multiple Zones, one of the Zones is considered the Master Zone and all
77 * Zone-related stats from the Keg are done in the Master Zone. For an
78 * example of a Multi-Zone setup, refer to the Mbuf allocation code.
82 * This is the representation for normal (Non OFFPAGE slab)
87 * <---------------- Page (UMA_SLAB_SIZE) ------------------>
88 * ___________________________________________________________
89 * | _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ ___________ |
90 * ||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i| |slab header||
91 * ||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_| |___________||
92 * |___________________________________________________________|
95 * This is an OFFPAGE slab. These can be larger than UMA_SLAB_SIZE.
97 * ___________________________________________________________
98 * | _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ |
99 * ||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i| |
100 * ||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_| |
101 * |___________________________________________________________|
111 #define UMA_SLAB_SIZE PAGE_SIZE /* How big are our slabs? */
112 #define UMA_SLAB_MASK (PAGE_SIZE - 1) /* Mask to get back to the page */
113 #define UMA_SLAB_SHIFT PAGE_SHIFT /* Number of bits PAGE_MASK */
115 #define UMA_BOOT_PAGES 64 /* Pages allocated for startup */
116 #define UMA_BOOT_PAGES_ZONES 32 /* Multiplier for pages to reserve */
117 /* if uma_zone > PAGE_SIZE */
119 /* Max waste percentage before going to off page slab management */
120 #define UMA_MAX_WASTE 10
123 * I doubt there will be many cases where this is exceeded. This is the initial
124 * size of the hash table for uma_slabs that are managed off page. This hash
125 * does expand by powers of two. Currently it doesn't get smaller.
127 #define UMA_HASH_SIZE_INIT 32
130 * I should investigate other hashing algorithms. This should yield a low
131 * number of collisions if the pages are relatively contiguous.
134 #define UMA_HASH(h, s) ((((uintptr_t)s) >> UMA_SLAB_SHIFT) & (h)->uh_hashmask)
136 #define UMA_HASH_INSERT(h, s, mem) \
137 SLIST_INSERT_HEAD(&(h)->uh_slab_hash[UMA_HASH((h), \
138 (mem))], (s), us_hlink)
139 #define UMA_HASH_REMOVE(h, s, mem) \
140 SLIST_REMOVE(&(h)->uh_slab_hash[UMA_HASH((h), \
141 (mem))], (s), uma_slab, us_hlink)
143 /* Hash table for freed address -> slab translation */
145 SLIST_HEAD(slabhead, uma_slab);
148 struct slabhead *uh_slab_hash; /* Hash table for slabs */
149 int uh_hashsize; /* Current size of the hash table */
150 int uh_hashmask; /* Mask used during hashing */
154 * align field or structure to cache line
156 #if defined(__amd64__)
157 #define UMA_ALIGN __aligned(CACHE_LINE_SIZE)
163 * Structures for per cpu queues.
167 LIST_ENTRY(uma_bucket) ub_link; /* Link into the zone */
168 int16_t ub_cnt; /* Count of free items. */
169 int16_t ub_entries; /* Max items. */
170 void *ub_bucket[]; /* actual allocation storage */
173 typedef struct uma_bucket * uma_bucket_t;
176 uma_bucket_t uc_freebucket; /* Bucket we're freeing to */
177 uma_bucket_t uc_allocbucket; /* Bucket to allocate from */
178 uint64_t uc_allocs; /* Count of allocations */
179 uint64_t uc_frees; /* Count of frees */
182 typedef struct uma_cache * uma_cache_t;
185 * Keg management structure
187 * TODO: Optimize for cache line size
191 struct mtx_padalign uk_lock; /* Lock for the keg */
192 struct uma_hash uk_hash;
194 LIST_HEAD(,uma_zone) uk_zones; /* Keg's zones */
195 LIST_HEAD(,uma_slab) uk_part_slab; /* partially allocated slabs */
196 LIST_HEAD(,uma_slab) uk_free_slab; /* empty slab list */
197 LIST_HEAD(,uma_slab) uk_full_slab; /* full slabs */
199 uint32_t uk_align; /* Alignment mask */
200 uint32_t uk_pages; /* Total page count */
201 uint32_t uk_free; /* Count of items free in slabs */
202 uint32_t uk_reserve; /* Number of reserved items. */
203 uint32_t uk_size; /* Requested size of each item */
204 uint32_t uk_rsize; /* Real size of each item */
205 uint32_t uk_maxpages; /* Maximum number of pages to alloc */
207 uma_init uk_init; /* Keg's init routine */
208 uma_fini uk_fini; /* Keg's fini routine */
209 uma_alloc uk_allocf; /* Allocation function */
210 uma_free uk_freef; /* Free routine */
212 u_long uk_offset; /* Next free offset from base KVA */
213 vm_offset_t uk_kva; /* Zone base KVA */
214 uma_zone_t uk_slabzone; /* Slab zone backing us, if OFFPAGE */
216 uint32_t uk_pgoff; /* Offset to uma_slab struct */
217 uint16_t uk_ppera; /* pages per allocation from backend */
218 uint16_t uk_ipers; /* Items per slab */
219 uint32_t uk_flags; /* Internal flags */
221 /* Least used fields go to the last cache line. */
222 const char *uk_name; /* Name of creating zone. */
223 LIST_ENTRY(uma_keg) uk_link; /* List of all kegs */
225 typedef struct uma_keg * uma_keg_t;
228 * Free bits per-slab.
230 #define SLAB_SETSIZE (PAGE_SIZE / UMA_SMALLEST_UNIT)
231 BITSET_DEFINE(slabbits, SLAB_SETSIZE);
234 * The slab structure manages a single contiguous allocation from backing
235 * store and subdivides it into individually allocatable items.
238 uma_keg_t us_keg; /* Keg we live in */
240 LIST_ENTRY(uma_slab) _us_link; /* slabs in zone */
241 unsigned long _us_size; /* Size of allocation */
243 SLIST_ENTRY(uma_slab) us_hlink; /* Link for hash table */
244 uint8_t *us_data; /* First item */
245 struct slabbits us_free; /* Free bitmask. */
247 struct slabbits us_debugfree; /* Debug bitmask. */
249 uint16_t us_freecount; /* How many are free? */
250 uint8_t us_flags; /* Page flags see uma.h */
251 uint8_t us_pad; /* Pad to 32bits, unused. */
254 #define us_link us_type._us_link
255 #define us_size us_type._us_size
257 typedef struct uma_slab * uma_slab_t;
258 typedef uma_slab_t (*uma_slaballoc)(uma_zone_t, uma_keg_t, int);
261 LIST_ENTRY(uma_klink) kl_link;
264 typedef struct uma_klink *uma_klink_t;
267 * Zone management structure
269 * TODO: Optimize for cache line size
273 struct mtx_padalign uz_lock; /* Lock for the zone */
274 struct mtx_padalign *uz_lockptr;
275 const char *uz_name; /* Text name of the zone */
277 LIST_ENTRY(uma_zone) uz_link; /* List of all zones in keg */
278 LIST_HEAD(,uma_bucket) uz_buckets; /* full buckets */
280 LIST_HEAD(,uma_klink) uz_kegs; /* List of kegs. */
281 struct uma_klink uz_klink; /* klink for first keg. */
283 uma_slaballoc uz_slab; /* Allocate a slab from the backend. */
284 uma_ctor uz_ctor; /* Constructor for each allocation */
285 uma_dtor uz_dtor; /* Destructor */
286 uma_init uz_init; /* Initializer for each item */
287 uma_fini uz_fini; /* Finalizer for each item. */
288 uma_import uz_import; /* Import new memory to cache. */
289 uma_release uz_release; /* Release memory from cache. */
290 void *uz_arg; /* Import/release argument. */
292 uint32_t uz_flags; /* Flags inherited from kegs */
293 uint32_t uz_size; /* Size inherited from kegs */
295 volatile u_long uz_allocs UMA_ALIGN; /* Total number of allocations */
296 volatile u_long uz_fails; /* Total number of alloc failures */
297 volatile u_long uz_frees; /* Total number of frees */
298 uint64_t uz_sleeps; /* Total number of alloc sleeps */
299 uint16_t uz_count; /* Amount of items in full bucket */
300 uint16_t uz_count_min; /* Minimal amount of items there */
302 /* The next two fields are used to print a rate-limited warnings. */
303 const char *uz_warning; /* Warning to print on failure */
304 struct timeval uz_ratecheck; /* Warnings rate-limiting */
306 struct task uz_maxaction; /* Task to run when at limit */
309 * This HAS to be the last item because we adjust the zone size
310 * based on NCPU and then allocate the space for the zones.
312 struct uma_cache uz_cpu[1]; /* Per cpu caches */
316 * These flags must not overlap with the UMA_ZONE flags specified in uma.h.
318 #define UMA_ZFLAG_MULTI 0x04000000 /* Multiple kegs in the zone. */
319 #define UMA_ZFLAG_DRAINING 0x08000000 /* Running zone_drain. */
320 #define UMA_ZFLAG_BUCKET 0x10000000 /* Bucket zone. */
321 #define UMA_ZFLAG_INTERNAL 0x20000000 /* No offpage no PCPU. */
322 #define UMA_ZFLAG_FULL 0x40000000 /* Reached uz_maxpages */
323 #define UMA_ZFLAG_CACHEONLY 0x80000000 /* Don't ask VM for buckets. */
325 #define UMA_ZFLAG_INHERIT \
326 (UMA_ZFLAG_INTERNAL | UMA_ZFLAG_CACHEONLY | UMA_ZFLAG_BUCKET)
328 static inline uma_keg_t
329 zone_first_keg(uma_zone_t zone)
333 klink = LIST_FIRST(&zone->uz_kegs);
334 return (klink != NULL) ? klink->kl_keg : NULL;
340 /* Internal prototypes */
341 static __inline uma_slab_t hash_sfind(struct uma_hash *hash, uint8_t *data);
342 void *uma_large_malloc(vm_size_t size, int wait);
343 void uma_large_free(uma_slab_t slab);
347 #define KEG_LOCK_INIT(k, lc) \
350 mtx_init(&(k)->uk_lock, (k)->uk_name, \
351 (k)->uk_name, MTX_DEF | MTX_DUPOK); \
353 mtx_init(&(k)->uk_lock, (k)->uk_name, \
354 "UMA zone", MTX_DEF | MTX_DUPOK); \
357 #define KEG_LOCK_FINI(k) mtx_destroy(&(k)->uk_lock)
358 #define KEG_LOCK(k) mtx_lock(&(k)->uk_lock)
359 #define KEG_UNLOCK(k) mtx_unlock(&(k)->uk_lock)
361 #define ZONE_LOCK_INIT(z, lc) \
364 mtx_init(&(z)->uz_lock, (z)->uz_name, \
365 (z)->uz_name, MTX_DEF | MTX_DUPOK); \
367 mtx_init(&(z)->uz_lock, (z)->uz_name, \
368 "UMA zone", MTX_DEF | MTX_DUPOK); \
371 #define ZONE_LOCK(z) mtx_lock((z)->uz_lockptr)
372 #define ZONE_TRYLOCK(z) mtx_trylock((z)->uz_lockptr)
373 #define ZONE_UNLOCK(z) mtx_unlock((z)->uz_lockptr)
374 #define ZONE_LOCK_FINI(z) mtx_destroy(&(z)->uz_lock)
377 * Find a slab within a hash table. This is used for OFFPAGE zones to lookup
378 * the slab structure.
381 * hash The hash table to search.
382 * data The base page of the item.
385 * A pointer to a slab if successful, else NULL.
387 static __inline uma_slab_t
388 hash_sfind(struct uma_hash *hash, uint8_t *data)
393 hval = UMA_HASH(hash, data);
395 SLIST_FOREACH(slab, &hash->uh_slab_hash[hval], us_hlink) {
396 if ((uint8_t *)slab->us_data == data)
402 static __inline uma_slab_t
403 vtoslab(vm_offset_t va)
407 p = PHYS_TO_VM_PAGE(pmap_kextract(va));
408 return ((uma_slab_t)p->plinks.s.pv);
412 vsetslab(vm_offset_t va, uma_slab_t slab)
416 p = PHYS_TO_VM_PAGE(pmap_kextract(va));
417 p->plinks.s.pv = slab;
421 * The following two functions may be defined by architecture specific code
422 * if they can provide more efficient allocation functions. This is useful
423 * for using direct mapped addresses.
425 void *uma_small_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *pflag,
427 void uma_small_free(void *mem, vm_size_t size, uint8_t flags);
430 #endif /* VM_UMA_INT_H */