1/* 2 * Copyright (c) 2002, Jeffrey Roberson <jroberson@chesapeake.net> 3 * All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions 7 * are met: 8 * 1. Redistributions of source code must retain the above copyright 9 * notice unmodified, this list of conditions, and the following 10 * disclaimer. 11 * 2. Redistributions in binary form must reproduce the above copyright 12 * notice, this list of conditions and the following disclaimer in the 13 * documentation and/or other materials provided with the distribution. 14 * 15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR 16 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 17 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 18 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, 19 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT 20 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 21 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 22 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 23 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF 24 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 25 *
|
26 * $FreeBSD: head/sys/vm/uma_int.h 99072 2002-06-29 17:26:22Z julian $
|
26 * $FreeBSD: head/sys/vm/uma_int.h 103531 2002-09-18 08:26:30Z jeff $ |
27 * 28 */ 29 30/* 31 * 32 * Jeff Roberson <jroberson@chesapeake.net> 33 * 34 * This file includes definitions, structures, prototypes, and inlines that 35 * should not be used outside of the actual implementation of UMA. 36 * 37 */ 38 39/* 40 * Here's a quick description of the relationship between the objects: 41 * 42 * Zones contain lists of slabs which are stored in either the full bin, empty 43 * bin, or partially allocated bin, to reduce fragmentation. They also contain 44 * the user supplied value for size, which is adjusted for alignment purposes 45 * and rsize is the result of that. The zone also stores information for 46 * managing a hash of page addresses that maps pages to uma_slab_t structures 47 * for pages that don't have embedded uma_slab_t's. 48 * 49 * The uma_slab_t may be embedded in a UMA_SLAB_SIZE chunk of memory or it may 50 * be allocated off the page from a special slab zone. The free list within a 51 * slab is managed with a linked list of indexes, which are 8 bit values. If 52 * UMA_SLAB_SIZE is defined to be too large I will have to switch to 16bit 53 * values. Currently on alpha you can get 250 or so 32 byte items and on x86 54 * you can get 250 or so 16byte items. For item sizes that would yield more 55 * than 10% memory waste we potentially allocate a separate uma_slab_t if this 56 * will improve the number of items per slab that will fit. 57 * 58 * Other potential space optimizations are storing the 8bit of linkage in space 59 * wasted between items due to alignment problems. This may yield a much better 60 * memory footprint for certain sizes of objects. Another alternative is to 61 * increase the UMA_SLAB_SIZE, or allow for dynamic slab sizes. I prefer 62 * dynamic slab sizes because we could stick with 8 bit indexes and only use 63 * large slab sizes for zones with a lot of waste per slab. This may create 64 * ineffeciencies in the vm subsystem due to fragmentation in the address space. 65 * 66 * The only really gross cases, with regards to memory waste, are for those 67 * items that are just over half the page size. You can get nearly 50% waste, 68 * so you fall back to the memory footprint of the power of two allocator. I 69 * have looked at memory allocation sizes on many of the machines available to 70 * me, and there does not seem to be an abundance of allocations at this range 71 * so at this time it may not make sense to optimize for it. This can, of 72 * course, be solved with dynamic slab sizes. 73 * 74 */ 75 76/* 77 * This is the representation for normal (Non OFFPAGE slab) 78 * 79 * i == item 80 * s == slab pointer 81 * 82 * <---------------- Page (UMA_SLAB_SIZE) ------------------> 83 * ___________________________________________________________ 84 * | _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ ___________ | 85 * ||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i| |slab header|| 86 * ||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_| |___________|| 87 * |___________________________________________________________| 88 * 89 * 90 * This is an OFFPAGE slab. These can be larger than UMA_SLAB_SIZE. 91 * 92 * ___________________________________________________________ 93 * | _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ | 94 * ||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i||i| | 95 * ||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_| | 96 * |___________________________________________________________| 97 * ___________ ^ 98 * |slab header| | 99 * |___________|---* 100 * 101 */ 102 103#ifndef VM_UMA_INT_H 104#define VM_UMA_INT_H 105
|
106#include <sys/mutex.h>
107
|
106#define UMA_SLAB_SIZE PAGE_SIZE /* How big are our slabs? */ 107#define UMA_SLAB_MASK (PAGE_SIZE - 1) /* Mask to get back to the page */ 108#define UMA_SLAB_SHIFT PAGE_SHIFT /* Number of bits PAGE_MASK */ 109 110#define UMA_BOOT_PAGES 30 /* Number of pages allocated for startup */ 111#define UMA_WORKING_TIME 20 /* Seconds worth of items to keep */ 112 113 114/* Max waste before going to off page slab management */ 115#define UMA_MAX_WASTE (UMA_SLAB_SIZE / 10) 116 117/* 118 * I doubt there will be many cases where this is exceeded. This is the initial 119 * size of the hash table for uma_slabs that are managed off page. This hash 120 * does expand by powers of two. Currently it doesn't get smaller. 121 */ 122#define UMA_HASH_SIZE_INIT 32 123 124 125/* 126 * I should investigate other hashing algorithms. This should yield a low 127 * number of collisions if the pages are relatively contiguous. 128 * 129 * This is the same algorithm that most processor caches use. 130 * 131 * I'm shifting and masking instead of % because it should be faster. 132 */ 133 134#define UMA_HASH(h, s) ((((unsigned long)s) >> UMA_SLAB_SHIFT) & \ 135 (h)->uh_hashmask) 136 137#define UMA_HASH_INSERT(h, s, mem) \ 138 SLIST_INSERT_HEAD(&(h)->uh_slab_hash[UMA_HASH((h), \ 139 (mem))], (s), us_hlink); 140#define UMA_HASH_REMOVE(h, s, mem) \ 141 SLIST_REMOVE(&(h)->uh_slab_hash[UMA_HASH((h), \ 142 (mem))], (s), uma_slab, us_hlink); 143 144/* Page management structure */ 145 146/* Sorry for the union, but space efficiency is important */ 147struct uma_slab { 148 uma_zone_t us_zone; /* Zone we live in */ 149 union { 150 LIST_ENTRY(uma_slab) us_link; /* slabs in zone */ 151 unsigned long us_size; /* Size of allocation */ 152 } us_type; 153 SLIST_ENTRY(uma_slab) us_hlink; /* Link for hash table */ 154 u_int8_t *us_data; /* First item */ 155 u_int8_t us_flags; /* Page flags see uma.h */ 156 u_int8_t us_freecount; /* How many are free? */ 157 u_int8_t us_firstfree; /* First free item index */ 158 u_int8_t us_freelist[1]; /* Free List (actually larger) */ 159}; 160 161#define us_link us_type.us_link 162#define us_size us_type.us_size 163 164typedef struct uma_slab * uma_slab_t; 165 166/* Hash table for freed address -> slab translation */ 167 168SLIST_HEAD(slabhead, uma_slab); 169 170struct 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
|
178extern struct uma_hash *mallochash;
179
|
176/* 177 * Structures for per cpu queues. 178 */ 179 180/* 181 * This size was chosen so that the struct bucket size is roughly 182 * 128 * sizeof(void *). This is exactly true for x86, and for alpha 183 * it will would be 32bits smaller if it didn't have alignment adjustments. 184 */ 185 186#define UMA_BUCKET_SIZE 125 187 188struct uma_bucket { 189 LIST_ENTRY(uma_bucket) ub_link; /* Link into the zone */ 190 int16_t ub_ptr; /* Pointer to current item */ 191 void *ub_bucket[UMA_BUCKET_SIZE]; /* actual allocation storage */ 192}; 193 194typedef struct uma_bucket * uma_bucket_t; 195 196struct uma_cache { 197 struct mtx uc_lock; /* Spin lock on this cpu's bucket */ 198 uma_bucket_t uc_freebucket; /* Bucket we're freeing to */ 199 uma_bucket_t uc_allocbucket; /* Bucket to allocate from */ 200 u_int64_t uc_allocs; /* Count of allocations */ 201}; 202 203typedef struct uma_cache * uma_cache_t; 204 205#define LOCKNAME_LEN 16 /* Length of the name for cpu locks */ 206 207/* 208 * Zone management structure 209 * 210 * TODO: Optimize for cache line size 211 * 212 */ 213struct uma_zone { 214 char uz_lname[LOCKNAME_LEN]; /* Text name for the cpu lock */ 215 char *uz_name; /* Text name of the zone */ 216 LIST_ENTRY(uma_zone) uz_link; /* List of all zones */ 217 u_int32_t uz_align; /* Alignment mask */ 218 u_int32_t uz_pages; /* Total page count */ 219 220/* Used during alloc / free */ 221 struct mtx uz_lock; /* Lock for the zone */ 222 u_int32_t uz_free; /* Count of items free in slabs */ 223 u_int16_t uz_ipers; /* Items per slab */ 224 u_int16_t uz_flags; /* Internal flags */ 225 226 LIST_HEAD(,uma_slab) uz_part_slab; /* partially allocated slabs */ 227 LIST_HEAD(,uma_slab) uz_free_slab; /* empty slab list */ 228 LIST_HEAD(,uma_slab) uz_full_slab; /* full slabs */ 229 LIST_HEAD(,uma_bucket) uz_full_bucket; /* full buckets */ 230 LIST_HEAD(,uma_bucket) uz_free_bucket; /* Buckets for frees */ 231 u_int32_t uz_size; /* Requested size of each item */ 232 u_int32_t uz_rsize; /* Real size of each item */ 233 234 struct uma_hash uz_hash; 235 u_int16_t uz_pgoff; /* Offset to uma_slab struct */ 236 u_int16_t uz_ppera; /* pages per allocation from backend */ 237 u_int16_t uz_cacheoff; /* Next cache offset */ 238 u_int16_t uz_cachemax; /* Max cache offset */ 239 240 uma_ctor uz_ctor; /* Constructor for each allocation */ 241 uma_dtor uz_dtor; /* Destructor */ 242 u_int64_t uz_allocs; /* Total number of allocations */ 243 244 uma_init uz_init; /* Initializer for each item */ 245 uma_fini uz_fini; /* Discards memory */ 246 uma_alloc uz_allocf; /* Allocation function */ 247 uma_free uz_freef; /* Free routine */ 248 struct vm_object *uz_obj; /* Zone specific object */ 249 vm_offset_t uz_kva; /* Base kva for zones with objs */ 250 u_int32_t uz_maxpages; /* Maximum number of pages to alloc */ 251 u_int32_t uz_cachefree; /* Last count of items free in caches */ 252 u_int64_t uz_oallocs; /* old allocs count */ 253 u_int64_t uz_wssize; /* Working set size */ 254 int uz_recurse; /* Allocation recursion count */ 255 uint16_t uz_fills; /* Outstanding bucket fills */ 256 uint16_t uz_count; /* Highest value ub_ptr can have */ 257 /* 258 * This HAS to be the last item because we adjust the zone size 259 * based on NCPU and then allocate the space for the zones. 260 */ 261 struct uma_cache uz_cpu[1]; /* Per cpu caches */ 262}; 263 264#define UMA_CACHE_INC 16 /* How much will we move data */ 265 266#define UMA_ZFLAG_OFFPAGE 0x0001 /* Struct slab/freelist off page */ 267#define UMA_ZFLAG_PRIVALLOC 0x0002 /* Zone has supplied it's own alloc */ 268#define UMA_ZFLAG_INTERNAL 0x0004 /* Internal zone, no offpage no PCPU */ 269#define UMA_ZFLAG_MALLOC 0x0008 /* Zone created by malloc */ 270#define UMA_ZFLAG_NOFREE 0x0010 /* Don't free data from this zone */ 271#define UMA_ZFLAG_FULL 0x0020 /* This zone reached uz_maxpages */ 272#define UMA_ZFLAG_BUCKETCACHE 0x0040 /* Only allocate buckets from cache */
|
273#define UMA_ZFLAG_HASH 0x0080 /* Look up slab via hash */ |
274 275/* This lives in uflags */ 276#define UMA_ZONE_INTERNAL 0x1000 /* Internal zone for uflags */ 277 278/* Internal prototypes */ 279static __inline uma_slab_t hash_sfind(struct uma_hash *hash, u_int8_t *data); 280void *uma_large_malloc(int size, int wait); 281void uma_large_free(uma_slab_t slab); 282 283/* Lock Macros */ 284 285#define ZONE_LOCK_INIT(z, lc) \ 286 do { \ 287 if ((lc)) \ 288 mtx_init(&(z)->uz_lock, (z)->uz_name, \ 289 (z)->uz_name, MTX_DEF | MTX_DUPOK); \ 290 else \ 291 mtx_init(&(z)->uz_lock, (z)->uz_name, \ 292 "UMA zone", MTX_DEF | MTX_DUPOK); \ 293 } while (0) 294 295#define ZONE_LOCK_FINI(z) mtx_destroy(&(z)->uz_lock) 296#define ZONE_LOCK(z) mtx_lock(&(z)->uz_lock) 297#define ZONE_UNLOCK(z) mtx_unlock(&(z)->uz_lock) 298 299#define CPU_LOCK_INIT(z, cpu, lc) \ 300 do { \ 301 if ((lc)) \ 302 mtx_init(&(z)->uz_cpu[(cpu)].uc_lock, \ 303 (z)->uz_lname, (z)->uz_lname, \ 304 MTX_DEF | MTX_DUPOK); \ 305 else \ 306 mtx_init(&(z)->uz_cpu[(cpu)].uc_lock, \ 307 (z)->uz_lname, "UMA cpu", \ 308 MTX_DEF | MTX_DUPOK); \ 309 } while (0) 310 311#define CPU_LOCK_FINI(z, cpu) \ 312 mtx_destroy(&(z)->uz_cpu[(cpu)].uc_lock) 313 314#define CPU_LOCK(z, cpu) \ 315 mtx_lock(&(z)->uz_cpu[(cpu)].uc_lock) 316 317#define CPU_UNLOCK(z, cpu) \ 318 mtx_unlock(&(z)->uz_cpu[(cpu)].uc_lock) 319 320/* 321 * Find a slab within a hash table. This is used for OFFPAGE zones to lookup 322 * the slab structure. 323 * 324 * Arguments: 325 * hash The hash table to search. 326 * data The base page of the item. 327 * 328 * Returns: 329 * A pointer to a slab if successful, else NULL. 330 */ 331static __inline uma_slab_t 332hash_sfind(struct uma_hash *hash, u_int8_t *data) 333{ 334 uma_slab_t slab; 335 int hval; 336 337 hval = UMA_HASH(hash, data); 338 339 SLIST_FOREACH(slab, &hash->uh_slab_hash[hval], us_hlink) { 340 if ((u_int8_t *)slab->us_data == data) 341 return (slab); 342 } 343 return (NULL); 344} 345
|
346static __inline uma_slab_t 347vtoslab(vm_offset_t va) 348{ 349 vm_page_t p; 350 uma_slab_t slab; |
351
|
352 p = PHYS_TO_VM_PAGE(pmap_kextract(va)); 353 slab = (uma_slab_t )p->object; 354 355 if (p->flags & PG_SLAB) 356 return (slab); 357 else 358 return (NULL); 359} 360 361static __inline void 362vsetslab(vm_offset_t va, uma_slab_t slab) 363{ 364 vm_page_t p; 365 366 p = PHYS_TO_VM_PAGE(pmap_kextract((vm_offset_t)va)); 367 p->object = (vm_object_t)slab; 368 p->flags |= PG_SLAB; 369} 370 371static __inline void 372vsetobj(vm_offset_t va, vm_object_t obj) 373{ 374 vm_page_t p; 375 376 p = PHYS_TO_VM_PAGE(pmap_kextract((vm_offset_t)va)); 377 p->object = obj; 378 p->flags &= ~PG_SLAB; 379} 380 |
381#endif /* VM_UMA_INT_H */
|