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/counter.h> 34 #include <sys/_bitset.h> 35 #include <sys/_domainset.h> 36 #include <sys/_task.h> 37 38 /* 39 * This file includes definitions, structures, prototypes, and inlines that 40 * should not be used outside of the actual implementation of UMA. 41 */ 42 43 /* 44 * The brief summary; Zones describe unique allocation types. Zones are 45 * organized into per-CPU caches which are filled by buckets. Buckets are 46 * organized according to memory domains. Buckets are filled from kegs which 47 * are also organized according to memory domains. Kegs describe a unique 48 * allocation type, backend memory provider, and layout. Kegs are associated 49 * with one or more zones and zones reference one or more kegs. Kegs provide 50 * slabs which are virtually contiguous collections of pages. Each slab is 51 * broken down int one or more items that will satisfy an individual allocation. 52 * 53 * Allocation is satisfied in the following order: 54 * 1) Per-CPU cache 55 * 2) Per-domain cache of buckets 56 * 3) Slab from any of N kegs 57 * 4) Backend page provider 58 * 59 * More detail on individual objects is contained below: 60 * 61 * Kegs contain lists of slabs which are stored in either the full bin, empty 62 * bin, or partially allocated bin, to reduce fragmentation. They also contain 63 * the user supplied value for size, which is adjusted for alignment purposes 64 * and rsize is the result of that. The Keg also stores information for 65 * managing a hash of page addresses that maps pages to uma_slab_t structures 66 * for pages that don't have embedded uma_slab_t's. 67 * 68 * Keg slab lists are organized by memory domain to support NUMA allocation 69 * policies. By default allocations are spread across domains to reduce the 70 * potential for hotspots. Special keg creation flags may be specified to 71 * prefer location allocation. However there is no strict enforcement as frees 72 * may happen on any CPU and these are returned to the CPU-local cache 73 * regardless of the originating domain. 74 * 75 * The uma_slab_t may be embedded in a UMA_SLAB_SIZE chunk of memory or it may 76 * be allocated off the page from a special slab zone. The free list within a 77 * slab is managed with a bitmask. For item sizes that would yield more than 78 * 10% memory waste we potentially allocate a separate uma_slab_t if this will 79 * improve the number of items per slab that will fit. 80 * 81 * The only really gross cases, with regards to memory waste, are for those 82 * items that are just over half the page size. You can get nearly 50% waste, 83 * so you fall back to the memory footprint of the power of two allocator. I 84 * have looked at memory allocation sizes on many of the machines available to 85 * me, and there does not seem to be an abundance of allocations at this range 86 * so at this time it may not make sense to optimize for it. This can, of 87 * course, be solved with dynamic slab sizes. 88 * 89 * Kegs may serve multiple Zones but by far most of the time they only serve 90 * one. When a Zone is created, a Keg is allocated and setup for it. While 91 * the backing Keg stores slabs, the Zone caches Buckets of items allocated 92 * from the slabs. Each Zone is equipped with an init/fini and ctor/dtor 93 * pair, as well as with its own set of small per-CPU caches, layered above 94 * the Zone's general Bucket cache. 95 * 96 * The PCPU caches are protected by critical sections, and may be accessed 97 * safely only from their associated CPU, while the Zones backed by the same 98 * Keg all share a common Keg lock (to coalesce contention on the backing 99 * slabs). The backing Keg typically only serves one Zone but in the case of 100 * multiple Zones, one of the Zones is considered the Master Zone and all 101 * Zone-related stats from the Keg are done in the Master Zone. For an 102 * example of a Multi-Zone setup, refer to the Mbuf allocation code. 103 */ 104 105 /* 106 * This is the representation for normal (Non OFFPAGE slab) 107 * 108 * i == item 109 * s == slab pointer 110 * 111 * <---------------- Page (UMA_SLAB_SIZE) ------------------> 112 * ___________________________________________________________ 113 * | _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ ___________ | 114 * ||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i| |slab header|| 115 * ||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_| |___________|| 116 * |___________________________________________________________| 117 * 118 * 119 * This is an OFFPAGE slab. These can be larger than UMA_SLAB_SIZE. 120 * 121 * ___________________________________________________________ 122 * | _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ | 123 * ||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i| | 124 * ||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_| | 125 * |___________________________________________________________| 126 * ___________ ^ 127 * |slab header| | 128 * |___________|---* 129 * 130 */ 131 132 #ifndef VM_UMA_INT_H 133 #define VM_UMA_INT_H 134 135 #define UMA_SLAB_SIZE PAGE_SIZE /* How big are our slabs? */ 136 #define UMA_SLAB_MASK (PAGE_SIZE - 1) /* Mask to get back to the page */ 137 #define UMA_SLAB_SHIFT PAGE_SHIFT /* Number of bits PAGE_MASK */ 138 139 /* Max waste percentage before going to off page slab management */ 140 #define UMA_MAX_WASTE 10 141 142 /* 143 * Actual size of uma_slab when it is placed at an end of a page 144 * with pointer sized alignment requirement. 145 */ 146 #define SIZEOF_UMA_SLAB ((sizeof(struct uma_slab) & UMA_ALIGN_PTR) ? \ 147 (sizeof(struct uma_slab) & ~UMA_ALIGN_PTR) + \ 148 (UMA_ALIGN_PTR + 1) : sizeof(struct uma_slab)) 149 150 /* 151 * Size of memory in a not offpage single page slab available for actual items. 152 */ 153 #define UMA_SLAB_SPACE (PAGE_SIZE - SIZEOF_UMA_SLAB) 154 155 /* 156 * I doubt there will be many cases where this is exceeded. This is the initial 157 * size of the hash table for uma_slabs that are managed off page. This hash 158 * does expand by powers of two. Currently it doesn't get smaller. 159 */ 160 #define UMA_HASH_SIZE_INIT 32 161 162 /* 163 * I should investigate other hashing algorithms. This should yield a low 164 * number of collisions if the pages are relatively contiguous. 165 */ 166 167 #define UMA_HASH(h, s) ((((uintptr_t)s) >> UMA_SLAB_SHIFT) & (h)->uh_hashmask) 168 169 #define UMA_HASH_INSERT(h, s, mem) \ 170 SLIST_INSERT_HEAD(&(h)->uh_slab_hash[UMA_HASH((h), \ 171 (mem))], (s), us_hlink) 172 #define UMA_HASH_REMOVE(h, s, mem) \ 173 SLIST_REMOVE(&(h)->uh_slab_hash[UMA_HASH((h), \ 174 (mem))], (s), uma_slab, us_hlink) 175 176 /* Hash table for freed address -> slab translation */ 177 178 SLIST_HEAD(slabhead, uma_slab); 179 180 struct uma_hash { 181 struct slabhead *uh_slab_hash; /* Hash table for slabs */ 182 u_int uh_hashsize; /* Current size of the hash table */ 183 u_int uh_hashmask; /* Mask used during hashing */ 184 }; 185 186 /* 187 * align field or structure to cache line 188 */ 189 #if defined(__amd64__) || defined(__powerpc64__) 190 #define UMA_ALIGN __aligned(128) 191 #else 192 #define UMA_ALIGN 193 #endif 194 195 /* 196 * Structures for per cpu queues. 197 */ 198 199 struct uma_bucket { 200 LIST_ENTRY(uma_bucket) ub_link; /* Link into the zone */ 201 int16_t ub_cnt; /* Count of items in bucket. */ 202 int16_t ub_entries; /* Max items. */ 203 void *ub_bucket[]; /* actual allocation storage */ 204 }; 205 206 typedef struct uma_bucket * uma_bucket_t; 207 208 struct uma_cache { 209 uma_bucket_t uc_freebucket; /* Bucket we're freeing to */ 210 uma_bucket_t uc_allocbucket; /* Bucket to allocate from */ 211 uint64_t uc_allocs; /* Count of allocations */ 212 uint64_t uc_frees; /* Count of frees */ 213 } UMA_ALIGN; 214 215 typedef struct uma_cache * uma_cache_t; 216 217 /* 218 * Per-domain memory list. Embedded in the kegs. 219 */ 220 struct uma_domain { 221 LIST_HEAD(,uma_slab) ud_part_slab; /* partially allocated slabs */ 222 LIST_HEAD(,uma_slab) ud_free_slab; /* empty slab list */ 223 LIST_HEAD(,uma_slab) ud_full_slab; /* full slabs */ 224 }; 225 226 typedef struct uma_domain * uma_domain_t; 227 228 /* 229 * Keg management structure 230 * 231 * TODO: Optimize for cache line size 232 * 233 */ 234 struct uma_keg { 235 struct mtx uk_lock; /* Lock for the keg must be first. 236 * See shared uz_keg/uz_lockptr 237 * member of struct uma_zone. */ 238 struct uma_hash uk_hash; 239 LIST_HEAD(,uma_zone) uk_zones; /* Keg's zones */ 240 241 struct domainset_ref uk_dr; /* Domain selection policy. */ 242 uint32_t uk_align; /* Alignment mask */ 243 uint32_t uk_pages; /* Total page count */ 244 uint32_t uk_free; /* Count of items free in slabs */ 245 uint32_t uk_reserve; /* Number of reserved items. */ 246 uint32_t uk_size; /* Requested size of each item */ 247 uint32_t uk_rsize; /* Real size of each item */ 248 249 uma_init uk_init; /* Keg's init routine */ 250 uma_fini uk_fini; /* Keg's fini routine */ 251 uma_alloc uk_allocf; /* Allocation function */ 252 uma_free uk_freef; /* Free routine */ 253 254 u_long uk_offset; /* Next free offset from base KVA */ 255 vm_offset_t uk_kva; /* Zone base KVA */ 256 uma_zone_t uk_slabzone; /* Slab zone backing us, if OFFPAGE */ 257 258 uint32_t uk_pgoff; /* Offset to uma_slab struct */ 259 uint16_t uk_ppera; /* pages per allocation from backend */ 260 uint16_t uk_ipers; /* Items per slab */ 261 uint32_t uk_flags; /* Internal flags */ 262 263 /* Least used fields go to the last cache line. */ 264 const char *uk_name; /* Name of creating zone. */ 265 LIST_ENTRY(uma_keg) uk_link; /* List of all kegs */ 266 267 /* Must be last, variable sized. */ 268 struct uma_domain uk_domain[]; /* Keg's slab lists. */ 269 }; 270 typedef struct uma_keg * uma_keg_t; 271 272 /* 273 * Free bits per-slab. 274 */ 275 #define SLAB_SETSIZE (PAGE_SIZE / UMA_SMALLEST_UNIT) 276 BITSET_DEFINE(slabbits, SLAB_SETSIZE); 277 278 /* 279 * The slab structure manages a single contiguous allocation from backing 280 * store and subdivides it into individually allocatable items. 281 */ 282 struct uma_slab { 283 uma_keg_t us_keg; /* Keg we live in */ 284 union { 285 LIST_ENTRY(uma_slab) _us_link; /* slabs in zone */ 286 unsigned long _us_size; /* Size of allocation */ 287 } us_type; 288 SLIST_ENTRY(uma_slab) us_hlink; /* Link for hash table */ 289 uint8_t *us_data; /* First item */ 290 struct slabbits us_free; /* Free bitmask. */ 291 #ifdef INVARIANTS 292 struct slabbits us_debugfree; /* Debug bitmask. */ 293 #endif 294 uint16_t us_freecount; /* How many are free? */ 295 uint8_t us_flags; /* Page flags see uma.h */ 296 uint8_t us_domain; /* Backing NUMA domain. */ 297 }; 298 299 #define us_link us_type._us_link 300 #define us_size us_type._us_size 301 302 #if MAXMEMDOM >= 255 303 #error "Slab domain type insufficient" 304 #endif 305 306 typedef struct uma_slab * uma_slab_t; 307 308 struct uma_zone_domain { 309 LIST_HEAD(,uma_bucket) uzd_buckets; /* full buckets */ 310 long uzd_nitems; /* total item count */ 311 long uzd_imax; /* maximum item count this period */ 312 long uzd_imin; /* minimum item count this period */ 313 long uzd_wss; /* working set size estimate */ 314 }; 315 316 typedef struct uma_zone_domain * uma_zone_domain_t; 317 318 /* 319 * Zone management structure 320 * 321 * TODO: Optimize for cache line size 322 * 323 */ 324 struct uma_zone { 325 /* Offset 0, used in alloc/free fast/medium fast path and const. */ 326 union { 327 uma_keg_t uz_keg; /* This zone's keg */ 328 struct mtx *uz_lockptr; /* To keg or to self */ 329 }; 330 struct uma_zone_domain *uz_domain; /* per-domain buckets */ 331 uint32_t uz_flags; /* Flags inherited from kegs */ 332 uint32_t uz_size; /* Size inherited from kegs */ 333 uma_ctor uz_ctor; /* Constructor for each allocation */ 334 uma_dtor uz_dtor; /* Destructor */ 335 uint64_t uz_items; /* Total items count */ 336 uint64_t uz_max_items; /* Maximum number of items to alloc */ 337 uint32_t uz_sleepers; /* Number of sleepers on memory */ 338 uint16_t uz_count; /* Amount of items in full bucket */ 339 uint16_t uz_count_max; /* Maximum amount of items there */ 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_init uz_init; /* Initializer for each item */ 346 uma_fini uz_fini; /* Finalizer for each item. */ 347 void *uz_spare; 348 uint64_t uz_bkt_count; /* Items in bucket cache */ 349 uint64_t uz_bkt_max; /* Maximum bucket cache size */ 350 351 /* Offset 128 Rare. */ 352 /* 353 * The lock is placed here to avoid adjacent line prefetcher 354 * in fast paths and to take up space near infrequently accessed 355 * members to reduce alignment overhead. 356 */ 357 struct mtx uz_lock; /* Lock for the zone */ 358 LIST_ENTRY(uma_zone) uz_link; /* List of all zones in keg */ 359 const char *uz_name; /* Text name of the zone */ 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 uint16_t uz_count_min; /* Minimal amount of items in bucket */ 365 366 /* Offset 256, stats. */ 367 counter_u64_t uz_allocs; /* Total number of allocations */ 368 counter_u64_t uz_frees; /* Total number of frees */ 369 counter_u64_t uz_fails; /* Total number of alloc failures */ 370 uint64_t uz_sleeps; /* Total number of alloc sleeps */ 371 372 /* 373 * This HAS to be the last item because we adjust the zone size 374 * based on NCPU and then allocate the space for the zones. 375 */ 376 struct uma_cache uz_cpu[]; /* Per cpu caches */ 377 378 /* uz_domain follows here. */ 379 }; 380 381 /* 382 * These flags must not overlap with the UMA_ZONE flags specified in uma.h. 383 */ 384 #define UMA_ZFLAG_CACHE 0x04000000 /* uma_zcache_create()d it */ 385 #define UMA_ZFLAG_DRAINING 0x08000000 /* Running zone_drain. */ 386 #define UMA_ZFLAG_BUCKET 0x10000000 /* Bucket zone. */ 387 #define UMA_ZFLAG_INTERNAL 0x20000000 /* No offpage no PCPU. */ 388 #define UMA_ZFLAG_CACHEONLY 0x80000000 /* Don't ask VM for buckets. */ 389 390 #define UMA_ZFLAG_INHERIT \ 391 (UMA_ZFLAG_INTERNAL | UMA_ZFLAG_CACHEONLY | UMA_ZFLAG_BUCKET) 392 393 #undef UMA_ALIGN 394 395 #ifdef _KERNEL 396 /* Internal prototypes */ 397 static __inline uma_slab_t hash_sfind(struct uma_hash *hash, uint8_t *data); 398 void *uma_large_malloc(vm_size_t size, int wait); 399 void *uma_large_malloc_domain(vm_size_t size, int domain, int wait); 400 void uma_large_free(uma_slab_t slab); 401 402 /* Lock Macros */ 403 404 #define KEG_LOCK_INIT(k, lc) \ 405 do { \ 406 if ((lc)) \ 407 mtx_init(&(k)->uk_lock, (k)->uk_name, \ 408 (k)->uk_name, MTX_DEF | MTX_DUPOK); \ 409 else \ 410 mtx_init(&(k)->uk_lock, (k)->uk_name, \ 411 "UMA zone", MTX_DEF | MTX_DUPOK); \ 412 } while (0) 413 414 #define KEG_LOCK_FINI(k) mtx_destroy(&(k)->uk_lock) 415 #define KEG_LOCK(k) mtx_lock(&(k)->uk_lock) 416 #define KEG_UNLOCK(k) mtx_unlock(&(k)->uk_lock) 417 #define KEG_LOCK_ASSERT(k) mtx_assert(&(k)->uk_lock, MA_OWNED) 418 419 #define KEG_GET(zone, keg) do { \ 420 (keg) = (zone)->uz_keg; \ 421 KASSERT((void *)(keg) != (void *)&(zone)->uz_lock, \ 422 ("%s: Invalid zone %p type", __func__, (zone))); \ 423 } while (0) 424 425 #define ZONE_LOCK_INIT(z, lc) \ 426 do { \ 427 if ((lc)) \ 428 mtx_init(&(z)->uz_lock, (z)->uz_name, \ 429 (z)->uz_name, MTX_DEF | MTX_DUPOK); \ 430 else \ 431 mtx_init(&(z)->uz_lock, (z)->uz_name, \ 432 "UMA zone", MTX_DEF | MTX_DUPOK); \ 433 } while (0) 434 435 #define ZONE_LOCK(z) mtx_lock((z)->uz_lockptr) 436 #define ZONE_TRYLOCK(z) mtx_trylock((z)->uz_lockptr) 437 #define ZONE_UNLOCK(z) mtx_unlock((z)->uz_lockptr) 438 #define ZONE_LOCK_FINI(z) mtx_destroy(&(z)->uz_lock) 439 #define ZONE_LOCK_ASSERT(z) mtx_assert((z)->uz_lockptr, MA_OWNED) 440 441 /* 442 * Find a slab within a hash table. This is used for OFFPAGE zones to lookup 443 * the slab structure. 444 * 445 * Arguments: 446 * hash The hash table to search. 447 * data The base page of the item. 448 * 449 * Returns: 450 * A pointer to a slab if successful, else NULL. 451 */ 452 static __inline uma_slab_t 453 hash_sfind(struct uma_hash *hash, uint8_t *data) 454 { 455 uma_slab_t slab; 456 u_int hval; 457 458 hval = UMA_HASH(hash, data); 459 460 SLIST_FOREACH(slab, &hash->uh_slab_hash[hval], us_hlink) { 461 if ((uint8_t *)slab->us_data == data) 462 return (slab); 463 } 464 return (NULL); 465 } 466 467 static __inline uma_slab_t 468 vtoslab(vm_offset_t va) 469 { 470 vm_page_t p; 471 472 p = PHYS_TO_VM_PAGE(pmap_kextract(va)); 473 return ((uma_slab_t)p->plinks.s.pv); 474 } 475 476 static __inline void 477 vsetslab(vm_offset_t va, uma_slab_t slab) 478 { 479 vm_page_t p; 480 481 p = PHYS_TO_VM_PAGE(pmap_kextract(va)); 482 p->plinks.s.pv = slab; 483 } 484 485 /* 486 * The following two functions may be defined by architecture specific code 487 * if they can provide more efficient allocation functions. This is useful 488 * for using direct mapped addresses. 489 */ 490 void *uma_small_alloc(uma_zone_t zone, vm_size_t bytes, int domain, 491 uint8_t *pflag, int wait); 492 void uma_small_free(void *mem, vm_size_t size, uint8_t flags); 493 494 /* Set a global soft limit on UMA managed memory. */ 495 void uma_set_limit(unsigned long limit); 496 #endif /* _KERNEL */ 497 498 #endif /* VM_UMA_INT_H */ 499