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 /* Max waste percentage before going to off page slab management */ 138 #define UMA_MAX_WASTE 10 139 140 /* 141 * Size of memory in a not offpage slab available for actual items. 142 */ 143 #define UMA_SLAB_SPACE (UMA_SLAB_SIZE - sizeof(struct uma_slab)) 144 145 /* 146 * I doubt there will be many cases where this is exceeded. This is the initial 147 * size of the hash table for uma_slabs that are managed off page. This hash 148 * does expand by powers of two. Currently it doesn't get smaller. 149 */ 150 #define UMA_HASH_SIZE_INIT 32 151 152 /* 153 * I should investigate other hashing algorithms. This should yield a low 154 * number of collisions if the pages are relatively contiguous. 155 */ 156 157 #define UMA_HASH(h, s) ((((uintptr_t)s) >> UMA_SLAB_SHIFT) & (h)->uh_hashmask) 158 159 #define UMA_HASH_INSERT(h, s, mem) \ 160 SLIST_INSERT_HEAD(&(h)->uh_slab_hash[UMA_HASH((h), \ 161 (mem))], (s), us_hlink) 162 #define UMA_HASH_REMOVE(h, s, mem) \ 163 SLIST_REMOVE(&(h)->uh_slab_hash[UMA_HASH((h), \ 164 (mem))], (s), uma_slab, us_hlink) 165 166 /* Hash table for freed address -> slab translation */ 167 168 SLIST_HEAD(slabhead, uma_slab); 169 170 struct uma_hash { 171 struct slabhead *uh_slab_hash; /* Hash table for slabs */ 172 int uh_hashsize; /* Current size of the hash table */ 173 int uh_hashmask; /* Mask used during hashing */ 174 }; 175 176 /* 177 * align field or structure to cache line 178 */ 179 #if defined(__amd64__) 180 #define UMA_ALIGN __aligned(CACHE_LINE_SIZE) 181 #else 182 #define UMA_ALIGN 183 #endif 184 185 /* 186 * Structures for per cpu queues. 187 */ 188 189 struct uma_bucket { 190 LIST_ENTRY(uma_bucket) ub_link; /* Link into the zone */ 191 int16_t ub_cnt; /* Count of free items. */ 192 int16_t ub_entries; /* Max items. */ 193 void *ub_bucket[]; /* actual allocation storage */ 194 }; 195 196 typedef struct uma_bucket * uma_bucket_t; 197 198 struct uma_cache { 199 uma_bucket_t uc_freebucket; /* Bucket we're freeing to */ 200 uma_bucket_t uc_allocbucket; /* Bucket to allocate from */ 201 uint64_t uc_allocs; /* Count of allocations */ 202 uint64_t uc_frees; /* Count of frees */ 203 } UMA_ALIGN; 204 205 typedef struct uma_cache * uma_cache_t; 206 207 /* 208 * Per-domain memory list. Embedded in the kegs. 209 */ 210 struct uma_domain { 211 LIST_HEAD(,uma_slab) ud_part_slab; /* partially allocated slabs */ 212 LIST_HEAD(,uma_slab) ud_free_slab; /* empty slab list */ 213 LIST_HEAD(,uma_slab) ud_full_slab; /* full slabs */ 214 }; 215 216 typedef struct uma_domain * uma_domain_t; 217 218 /* 219 * Keg management structure 220 * 221 * TODO: Optimize for cache line size 222 * 223 */ 224 struct uma_keg { 225 struct mtx_padalign uk_lock; /* Lock for the keg */ 226 struct uma_hash uk_hash; 227 228 LIST_HEAD(,uma_zone) uk_zones; /* Keg's zones */ 229 230 uint32_t uk_cursor; /* Domain alloc cursor. */ 231 uint32_t uk_align; /* Alignment mask */ 232 uint32_t uk_pages; /* Total page count */ 233 uint32_t uk_free; /* Count of items free in slabs */ 234 uint32_t uk_reserve; /* Number of reserved items. */ 235 uint32_t uk_size; /* Requested size of each item */ 236 uint32_t uk_rsize; /* Real size of each item */ 237 uint32_t uk_maxpages; /* Maximum number of pages to alloc */ 238 239 uma_init uk_init; /* Keg's init routine */ 240 uma_fini uk_fini; /* Keg's fini routine */ 241 uma_alloc uk_allocf; /* Allocation function */ 242 uma_free uk_freef; /* Free routine */ 243 244 u_long uk_offset; /* Next free offset from base KVA */ 245 vm_offset_t uk_kva; /* Zone base KVA */ 246 uma_zone_t uk_slabzone; /* Slab zone backing us, if OFFPAGE */ 247 248 uint32_t uk_pgoff; /* Offset to uma_slab struct */ 249 uint16_t uk_ppera; /* pages per allocation from backend */ 250 uint16_t uk_ipers; /* Items per slab */ 251 uint32_t uk_flags; /* Internal flags */ 252 253 /* Least used fields go to the last cache line. */ 254 const char *uk_name; /* Name of creating zone. */ 255 LIST_ENTRY(uma_keg) uk_link; /* List of all kegs */ 256 257 /* Must be last, variable sized. */ 258 struct uma_domain uk_domain[]; /* Keg's slab lists. */ 259 }; 260 typedef struct uma_keg * uma_keg_t; 261 262 /* 263 * Free bits per-slab. 264 */ 265 #define SLAB_SETSIZE (PAGE_SIZE / UMA_SMALLEST_UNIT) 266 BITSET_DEFINE(slabbits, SLAB_SETSIZE); 267 268 /* 269 * The slab structure manages a single contiguous allocation from backing 270 * store and subdivides it into individually allocatable items. 271 */ 272 struct uma_slab { 273 uma_keg_t us_keg; /* Keg we live in */ 274 union { 275 LIST_ENTRY(uma_slab) _us_link; /* slabs in zone */ 276 unsigned long _us_size; /* Size of allocation */ 277 } us_type; 278 SLIST_ENTRY(uma_slab) us_hlink; /* Link for hash table */ 279 uint8_t *us_data; /* First item */ 280 struct slabbits us_free; /* Free bitmask. */ 281 #ifdef INVARIANTS 282 struct slabbits us_debugfree; /* Debug bitmask. */ 283 #endif 284 uint16_t us_freecount; /* How many are free? */ 285 uint8_t us_flags; /* Page flags see uma.h */ 286 uint8_t us_domain; /* Backing NUMA domain. */ 287 }; 288 289 #define us_link us_type._us_link 290 #define us_size us_type._us_size 291 292 #if MAXMEMDOM >= 255 293 #error "Slab domain type insufficient" 294 #endif 295 296 typedef struct uma_slab * uma_slab_t; 297 typedef uma_slab_t (*uma_slaballoc)(uma_zone_t, uma_keg_t, int, int); 298 299 struct uma_klink { 300 LIST_ENTRY(uma_klink) kl_link; 301 uma_keg_t kl_keg; 302 }; 303 typedef struct uma_klink *uma_klink_t; 304 305 struct uma_zone_domain { 306 LIST_HEAD(,uma_bucket) uzd_buckets; /* full buckets */ 307 }; 308 309 typedef struct uma_zone_domain * uma_zone_domain_t; 310 311 /* 312 * Zone management structure 313 * 314 * TODO: Optimize for cache line size 315 * 316 */ 317 struct uma_zone { 318 struct mtx_padalign uz_lock; /* Lock for the zone */ 319 struct mtx_padalign *uz_lockptr; 320 const char *uz_name; /* Text name of the zone */ 321 322 LIST_ENTRY(uma_zone) uz_link; /* List of all zones in keg */ 323 struct uma_zone_domain *uz_domain; /* per-domain buckets */ 324 325 LIST_HEAD(,uma_klink) uz_kegs; /* List of kegs. */ 326 struct uma_klink uz_klink; /* klink for first keg. */ 327 328 uma_slaballoc uz_slab; /* Allocate a slab from the backend. */ 329 uma_ctor uz_ctor; /* Constructor for each allocation */ 330 uma_dtor uz_dtor; /* Destructor */ 331 uma_init uz_init; /* Initializer for each item */ 332 uma_fini uz_fini; /* Finalizer for each item. */ 333 uma_import uz_import; /* Import new memory to cache. */ 334 uma_release uz_release; /* Release memory from cache. */ 335 void *uz_arg; /* Import/release argument. */ 336 337 uint32_t uz_flags; /* Flags inherited from kegs */ 338 uint32_t uz_size; /* Size inherited from kegs */ 339 340 volatile u_long uz_allocs UMA_ALIGN; /* Total number of allocations */ 341 volatile u_long uz_fails; /* Total number of alloc failures */ 342 volatile u_long uz_frees; /* Total number of frees */ 343 uint64_t uz_sleeps; /* Total number of alloc sleeps */ 344 uint16_t uz_count; /* Amount of items in full bucket */ 345 uint16_t uz_count_min; /* Minimal amount of items there */ 346 347 /* The next two fields are used to print a rate-limited warnings. */ 348 const char *uz_warning; /* Warning to print on failure */ 349 struct timeval uz_ratecheck; /* Warnings rate-limiting */ 350 351 struct task uz_maxaction; /* Task to run when at limit */ 352 353 /* 354 * This HAS to be the last item because we adjust the zone size 355 * based on NCPU and then allocate the space for the zones. 356 */ 357 struct uma_cache uz_cpu[]; /* Per cpu caches */ 358 359 /* uz_domain follows here. */ 360 }; 361 362 /* 363 * These flags must not overlap with the UMA_ZONE flags specified in uma.h. 364 */ 365 #define UMA_ZFLAG_MULTI 0x04000000 /* Multiple kegs in the zone. */ 366 #define UMA_ZFLAG_DRAINING 0x08000000 /* Running zone_drain. */ 367 #define UMA_ZFLAG_BUCKET 0x10000000 /* Bucket zone. */ 368 #define UMA_ZFLAG_INTERNAL 0x20000000 /* No offpage no PCPU. */ 369 #define UMA_ZFLAG_FULL 0x40000000 /* Reached uz_maxpages */ 370 #define UMA_ZFLAG_CACHEONLY 0x80000000 /* Don't ask VM for buckets. */ 371 372 #define UMA_ZFLAG_INHERIT \ 373 (UMA_ZFLAG_INTERNAL | UMA_ZFLAG_CACHEONLY | UMA_ZFLAG_BUCKET) 374 375 static inline uma_keg_t 376 zone_first_keg(uma_zone_t zone) 377 { 378 uma_klink_t klink; 379 380 klink = LIST_FIRST(&zone->uz_kegs); 381 return (klink != NULL) ? klink->kl_keg : NULL; 382 } 383 384 #undef UMA_ALIGN 385 386 #ifdef _KERNEL 387 /* Internal prototypes */ 388 static __inline uma_slab_t hash_sfind(struct uma_hash *hash, uint8_t *data); 389 void *uma_large_malloc(vm_size_t size, int wait); 390 void *uma_large_malloc_domain(vm_size_t size, int domain, int wait); 391 void uma_large_free(uma_slab_t slab); 392 393 /* Lock Macros */ 394 395 #define KEG_LOCK_INIT(k, lc) \ 396 do { \ 397 if ((lc)) \ 398 mtx_init(&(k)->uk_lock, (k)->uk_name, \ 399 (k)->uk_name, MTX_DEF | MTX_DUPOK); \ 400 else \ 401 mtx_init(&(k)->uk_lock, (k)->uk_name, \ 402 "UMA zone", MTX_DEF | MTX_DUPOK); \ 403 } while (0) 404 405 #define KEG_LOCK_FINI(k) mtx_destroy(&(k)->uk_lock) 406 #define KEG_LOCK(k) mtx_lock(&(k)->uk_lock) 407 #define KEG_UNLOCK(k) mtx_unlock(&(k)->uk_lock) 408 409 #define ZONE_LOCK_INIT(z, lc) \ 410 do { \ 411 if ((lc)) \ 412 mtx_init(&(z)->uz_lock, (z)->uz_name, \ 413 (z)->uz_name, MTX_DEF | MTX_DUPOK); \ 414 else \ 415 mtx_init(&(z)->uz_lock, (z)->uz_name, \ 416 "UMA zone", MTX_DEF | MTX_DUPOK); \ 417 } while (0) 418 419 #define ZONE_LOCK(z) mtx_lock((z)->uz_lockptr) 420 #define ZONE_TRYLOCK(z) mtx_trylock((z)->uz_lockptr) 421 #define ZONE_UNLOCK(z) mtx_unlock((z)->uz_lockptr) 422 #define ZONE_LOCK_FINI(z) mtx_destroy(&(z)->uz_lock) 423 424 /* 425 * Find a slab within a hash table. This is used for OFFPAGE zones to lookup 426 * the slab structure. 427 * 428 * Arguments: 429 * hash The hash table to search. 430 * data The base page of the item. 431 * 432 * Returns: 433 * A pointer to a slab if successful, else NULL. 434 */ 435 static __inline uma_slab_t 436 hash_sfind(struct uma_hash *hash, uint8_t *data) 437 { 438 uma_slab_t slab; 439 int hval; 440 441 hval = UMA_HASH(hash, data); 442 443 SLIST_FOREACH(slab, &hash->uh_slab_hash[hval], us_hlink) { 444 if ((uint8_t *)slab->us_data == data) 445 return (slab); 446 } 447 return (NULL); 448 } 449 450 static __inline uma_slab_t 451 vtoslab(vm_offset_t va) 452 { 453 vm_page_t p; 454 455 p = PHYS_TO_VM_PAGE(pmap_kextract(va)); 456 return ((uma_slab_t)p->plinks.s.pv); 457 } 458 459 static __inline void 460 vsetslab(vm_offset_t va, uma_slab_t slab) 461 { 462 vm_page_t p; 463 464 p = PHYS_TO_VM_PAGE(pmap_kextract(va)); 465 p->plinks.s.pv = slab; 466 } 467 468 /* 469 * The following two functions may be defined by architecture specific code 470 * if they can provide more efficient allocation functions. This is useful 471 * for using direct mapped addresses. 472 */ 473 void *uma_small_alloc(uma_zone_t zone, vm_size_t bytes, int domain, 474 uint8_t *pflag, int wait); 475 void uma_small_free(void *mem, vm_size_t size, uint8_t flags); 476 477 /* Set a global soft limit on UMA managed memory. */ 478 void uma_set_limit(unsigned long limit); 479 #endif /* _KERNEL */ 480 481 #endif /* VM_UMA_INT_H */ 482