1 /* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License, Version 1.0 only 6 * (the "License"). You may not use this file except in compliance 7 * with the License. 8 * 9 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 10 * or http://www.opensolaris.org/os/licensing. 11 * See the License for the specific language governing permissions 12 * and limitations under the License. 13 * 14 * When distributing Covered Code, include this CDDL HEADER in each 15 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 16 * If applicable, add the following below this CDDL HEADER, with the 17 * fields enclosed by brackets "[]" replaced with your own identifying 18 * information: Portions Copyright [yyyy] [name of copyright owner] 19 * 20 * CDDL HEADER END 21 */ 22 /* 23 * Copyright 2005 Sun Microsystems, Inc. All rights reserved. 24 * Use is subject to license terms. 25 */ 26 27 #ifndef _SYS_SPA_H 28 #define _SYS_SPA_H 29 30 #pragma ident "%Z%%M% %I% %E% SMI" 31 32 #include <sys/avl.h> 33 #include <sys/zfs_context.h> 34 #include <sys/nvpair.h> 35 #include <sys/sysmacros.h> 36 #include <sys/types.h> 37 #include <sys/fs/zfs.h> 38 39 #ifdef __cplusplus 40 extern "C" { 41 #endif 42 43 /* 44 * Forward references that lots of things need. 45 */ 46 typedef struct spa spa_t; 47 typedef struct vdev vdev_t; 48 typedef struct metaslab metaslab_t; 49 typedef struct zilog zilog_t; 50 typedef struct traverse_handle traverse_handle_t; 51 struct dsl_pool; 52 53 /* 54 * General-purpose 32-bit and 64-bit bitfield encodings. 55 */ 56 #define BF32_DECODE(x, low, len) P2PHASE((x) >> (low), 1U << (len)) 57 #define BF64_DECODE(x, low, len) P2PHASE((x) >> (low), 1ULL << (len)) 58 #define BF32_ENCODE(x, low, len) (P2PHASE((x), 1U << (len)) << (low)) 59 #define BF64_ENCODE(x, low, len) (P2PHASE((x), 1ULL << (len)) << (low)) 60 61 #define BF32_GET(x, low, len) BF32_DECODE(x, low, len) 62 #define BF64_GET(x, low, len) BF64_DECODE(x, low, len) 63 64 #define BF32_SET(x, low, len, val) \ 65 ((x) ^= BF32_ENCODE((x >> low) ^ val, low, len)) 66 #define BF64_SET(x, low, len, val) \ 67 ((x) ^= BF64_ENCODE((x >> low) ^ val, low, len)) 68 69 #define BF32_GET_SB(x, low, len, shift, bias) \ 70 ((BF32_GET(x, low, len) + (bias)) << (shift)) 71 #define BF64_GET_SB(x, low, len, shift, bias) \ 72 ((BF64_GET(x, low, len) + (bias)) << (shift)) 73 74 #define BF32_SET_SB(x, low, len, shift, bias, val) \ 75 BF32_SET(x, low, len, ((val) >> (shift)) - (bias)) 76 #define BF64_SET_SB(x, low, len, shift, bias, val) \ 77 BF64_SET(x, low, len, ((val) >> (shift)) - (bias)) 78 79 /* 80 * We currently support nine block sizes, from 512 bytes to 128K. 81 * We could go higher, but the benefits are near-zero and the cost 82 * of COWing a giant block to modify one byte would become excessive. 83 */ 84 #define SPA_MINBLOCKSHIFT 9 85 #define SPA_MAXBLOCKSHIFT 17 86 #define SPA_MINBLOCKSIZE (1ULL << SPA_MINBLOCKSHIFT) 87 #define SPA_MAXBLOCKSIZE (1ULL << SPA_MAXBLOCKSHIFT) 88 89 #define SPA_BLOCKSIZES (SPA_MAXBLOCKSHIFT - SPA_MINBLOCKSHIFT + 1) 90 91 /* 92 * The DVA size encodings for LSIZE and PSIZE support blocks up to 32MB. 93 * The ASIZE encoding should be at least 64 times larger (6 more bits) 94 * to support up to 4-way RAID-Z mirror mode with worst-case gang block 95 * overhead, three DVAs per bp, plus one more bit in case we do anything 96 * else that expands the ASIZE. 97 */ 98 #define SPA_LSIZEBITS 16 /* LSIZE up to 32M (2^16 * 512) */ 99 #define SPA_PSIZEBITS 16 /* PSIZE up to 32M (2^16 * 512) */ 100 #define SPA_ASIZEBITS 24 /* ASIZE up to 64 times larger */ 101 102 /* 103 * All SPA data is represented by 128-bit data virtual addresses (DVAs). 104 * The members of the dva_t should be considered opaque outside the SPA. 105 */ 106 typedef struct dva { 107 uint64_t dva_word[2]; 108 } dva_t; 109 110 /* 111 * Each block has a 256-bit checksum -- strong enough for cryptographic hashes. 112 */ 113 typedef struct zio_cksum { 114 uint64_t zc_word[4]; 115 } zio_cksum_t; 116 117 /* 118 * Each block is described by its DVAs, time of birth, checksum, etc. 119 * The word-by-word, bit-by-bit layout of the blkptr is as follows: 120 * 121 * 64 56 48 40 32 24 16 8 0 122 * +-------+-------+-------+-------+-------+-------+-------+-------+ 123 * 0 | vdev1 | GRID | ASIZE | 124 * +-------+-------+-------+-------+-------+-------+-------+-------+ 125 * 1 |G| offset1 | 126 * +-------+-------+-------+-------+-------+-------+-------+-------+ 127 * 2 | vdev2 | GRID | ASIZE | 128 * +-------+-------+-------+-------+-------+-------+-------+-------+ 129 * 3 |G| offset2 | 130 * +-------+-------+-------+-------+-------+-------+-------+-------+ 131 * 4 | vdev3 | GRID | ASIZE | 132 * +-------+-------+-------+-------+-------+-------+-------+-------+ 133 * 5 |G| offset3 | 134 * +-------+-------+-------+-------+-------+-------+-------+-------+ 135 * 6 |E| lvl | type | cksum | comp | PSIZE | LSIZE | 136 * +-------+-------+-------+-------+-------+-------+-------+-------+ 137 * 7 | padding | 138 * +-------+-------+-------+-------+-------+-------+-------+-------+ 139 * 8 | padding | 140 * +-------+-------+-------+-------+-------+-------+-------+-------+ 141 * 9 | padding | 142 * +-------+-------+-------+-------+-------+-------+-------+-------+ 143 * a | birth txg | 144 * +-------+-------+-------+-------+-------+-------+-------+-------+ 145 * b | fill count | 146 * +-------+-------+-------+-------+-------+-------+-------+-------+ 147 * c | checksum[0] | 148 * +-------+-------+-------+-------+-------+-------+-------+-------+ 149 * d | checksum[1] | 150 * +-------+-------+-------+-------+-------+-------+-------+-------+ 151 * e | checksum[2] | 152 * +-------+-------+-------+-------+-------+-------+-------+-------+ 153 * f | checksum[3] | 154 * +-------+-------+-------+-------+-------+-------+-------+-------+ 155 * 156 * Legend: 157 * 158 * vdev virtual device ID 159 * offset offset into virtual device 160 * LSIZE logical size 161 * PSIZE physical size (after compression) 162 * ASIZE allocated size (including RAID-Z parity and gang block headers) 163 * GRID RAID-Z layout information (reserved for future use) 164 * cksum checksum function 165 * comp compression function 166 * G gang block indicator 167 * E endianness 168 * type DMU object type 169 * lvl level of indirection 170 * birth txg transaction group in which the block was born 171 * fill count number of non-zero blocks under this bp 172 * checksum[4] 256-bit checksum of the data this bp describes 173 */ 174 typedef struct blkptr { 175 dva_t blk_dva[3]; /* 128-bit Data Virtual Address */ 176 uint64_t blk_prop; /* size, compression, type, etc */ 177 uint64_t blk_pad[3]; /* Extra space for the future */ 178 uint64_t blk_birth; /* transaction group at birth */ 179 uint64_t blk_fill; /* fill count */ 180 zio_cksum_t blk_cksum; /* 256-bit checksum */ 181 } blkptr_t; 182 183 #define SPA_BLKPTRSHIFT 7 /* blkptr_t is 128 bytes */ 184 #define SPA_DVAS_PER_BP 3 /* Number of DVAs in a bp */ 185 186 /* 187 * Macros to get and set fields in a bp or DVA. 188 */ 189 #define DVA_GET_ASIZE(dva) \ 190 BF64_GET_SB((dva)->dva_word[0], 0, 24, SPA_MINBLOCKSHIFT, 0) 191 #define DVA_SET_ASIZE(dva, x) \ 192 BF64_SET_SB((dva)->dva_word[0], 0, 24, SPA_MINBLOCKSHIFT, 0, x) 193 194 #define DVA_GET_GRID(dva) BF64_GET((dva)->dva_word[0], 24, 8) 195 #define DVA_SET_GRID(dva, x) BF64_SET((dva)->dva_word[0], 24, 8, x) 196 197 #define DVA_GET_VDEV(dva) BF64_GET((dva)->dva_word[0], 32, 32) 198 #define DVA_SET_VDEV(dva, x) BF64_SET((dva)->dva_word[0], 32, 32, x) 199 200 #define DVA_GET_OFFSET(dva) \ 201 BF64_GET_SB((dva)->dva_word[1], 0, 63, SPA_MINBLOCKSHIFT, 0) 202 #define DVA_SET_OFFSET(dva, x) \ 203 BF64_SET_SB((dva)->dva_word[1], 0, 63, SPA_MINBLOCKSHIFT, 0, x) 204 205 #define DVA_GET_GANG(dva) BF64_GET((dva)->dva_word[1], 63, 1) 206 #define DVA_SET_GANG(dva, x) BF64_SET((dva)->dva_word[1], 63, 1, x) 207 208 #define BP_GET_LSIZE(bp) \ 209 (BP_IS_HOLE(bp) ? 0 : \ 210 BF64_GET_SB((bp)->blk_prop, 0, 16, SPA_MINBLOCKSHIFT, 1)) 211 #define BP_SET_LSIZE(bp, x) \ 212 BF64_SET_SB((bp)->blk_prop, 0, 16, SPA_MINBLOCKSHIFT, 1, x) 213 214 #define BP_GET_PSIZE(bp) \ 215 BF64_GET_SB((bp)->blk_prop, 16, 16, SPA_MINBLOCKSHIFT, 1) 216 #define BP_SET_PSIZE(bp, x) \ 217 BF64_SET_SB((bp)->blk_prop, 16, 16, SPA_MINBLOCKSHIFT, 1, x) 218 219 #define BP_GET_COMPRESS(bp) BF64_GET((bp)->blk_prop, 32, 8) 220 #define BP_SET_COMPRESS(bp, x) BF64_SET((bp)->blk_prop, 32, 8, x) 221 222 #define BP_GET_CHECKSUM(bp) BF64_GET((bp)->blk_prop, 40, 8) 223 #define BP_SET_CHECKSUM(bp, x) BF64_SET((bp)->blk_prop, 40, 8, x) 224 225 #define BP_GET_TYPE(bp) BF64_GET((bp)->blk_prop, 48, 8) 226 #define BP_SET_TYPE(bp, x) BF64_SET((bp)->blk_prop, 48, 8, x) 227 228 #define BP_GET_LEVEL(bp) BF64_GET((bp)->blk_prop, 56, 5) 229 #define BP_SET_LEVEL(bp, x) BF64_SET((bp)->blk_prop, 56, 5, x) 230 231 #define BP_GET_BYTEORDER(bp) (0 - BF64_GET((bp)->blk_prop, 63, 1)) 232 #define BP_SET_BYTEORDER(bp, x) BF64_SET((bp)->blk_prop, 63, 1, x) 233 234 #define BP_GET_ASIZE(bp) \ 235 (DVA_GET_ASIZE(&(bp)->blk_dva[0]) + DVA_GET_ASIZE(&(bp)->blk_dva[1]) + \ 236 DVA_GET_ASIZE(&(bp)->blk_dva[2])) 237 238 #define DVA_EQUAL(dva1, dva2) \ 239 ((dva1)->dva_word[1] == (dva2)->dva_word[1] && \ 240 (dva1)->dva_word[0] == (dva2)->dva_word[0]) 241 242 #define DVA_IS_VALID(dva) (DVA_GET_ASIZE(dva) != 0) 243 244 #define ZIO_SET_CHECKSUM(zcp, w0, w1, w2, w3) \ 245 { \ 246 (zcp)->zc_word[0] = w0; \ 247 (zcp)->zc_word[1] = w1; \ 248 (zcp)->zc_word[2] = w2; \ 249 (zcp)->zc_word[3] = w3; \ 250 } 251 252 #define BP_IS_HOLE(bp) ((bp)->blk_birth == 0) 253 254 #define BP_IDENTITY(bp) (&(bp)->blk_dva[0]) 255 256 #define BP_ZERO(bp) \ 257 { \ 258 (bp)->blk_dva[0].dva_word[0] = 0; \ 259 (bp)->blk_dva[0].dva_word[1] = 0; \ 260 (bp)->blk_dva[1].dva_word[0] = 0; \ 261 (bp)->blk_dva[1].dva_word[1] = 0; \ 262 (bp)->blk_dva[2].dva_word[0] = 0; \ 263 (bp)->blk_dva[2].dva_word[1] = 0; \ 264 (bp)->blk_prop = 0; \ 265 (bp)->blk_pad[0] = 0; \ 266 (bp)->blk_pad[1] = 0; \ 267 (bp)->blk_pad[2] = 0; \ 268 (bp)->blk_birth = 0; \ 269 (bp)->blk_fill = 0; \ 270 ZIO_SET_CHECKSUM(&(bp)->blk_cksum, 0, 0, 0, 0); \ 271 } 272 273 /* 274 * Note: the byteorder is either 0 or -1, both of which are palindromes. 275 * This simplifies the endianness handling a bit. 276 */ 277 #ifdef _BIG_ENDIAN 278 #define ZFS_HOST_BYTEORDER (0ULL) 279 #else 280 #define ZFS_HOST_BYTEORDER (-1ULL) 281 #endif 282 283 #define BP_SHOULD_BYTESWAP(bp) (BP_GET_BYTEORDER(bp) != ZFS_HOST_BYTEORDER) 284 285 #include <sys/dmu.h> 286 287 /* 288 * Routines found in spa.c 289 */ 290 291 /* state manipulation functions */ 292 extern int spa_open(const char *pool, spa_t **, void *tag); 293 extern int spa_get_stats(const char *pool, nvlist_t **config); 294 extern int spa_create(const char *pool, nvlist_t *config, char *altroot); 295 extern int spa_import(const char *pool, nvlist_t *config, char *altroot); 296 extern nvlist_t *spa_tryimport(nvlist_t *tryconfig); 297 extern int spa_destroy(char *pool); 298 extern int spa_export(char *pool); 299 300 /* device manipulation */ 301 extern int spa_vdev_add(spa_t *spa, nvlist_t *nvroot); 302 extern int spa_vdev_add_unlocked(spa_t *spa, nvlist_t *nvroot); 303 extern int spa_vdev_attach(spa_t *spa, const char *path, nvlist_t *nvroot, 304 int replacing); 305 extern int spa_vdev_detach(spa_t *spa, const char *path, uint64_t guid, 306 int replace_done); 307 extern void spa_vdev_replace_done(spa_t *spa); 308 309 /* scrubbing */ 310 extern int spa_scrub(spa_t *spa, pool_scrub_type_t type, boolean_t force); 311 extern void spa_scrub_suspend(spa_t *spa); 312 extern void spa_scrub_resume(spa_t *spa); 313 extern void spa_scrub_restart(spa_t *spa, uint64_t txg); 314 315 /* spa syncing */ 316 extern void spa_sync(spa_t *spa, uint64_t txg); /* only for DMU use */ 317 extern void spa_sync_allpools(void); 318 319 /* 320 * SPA configuration functions in spa_config.c 321 */ 322 extern void spa_config_sync(void); 323 extern void spa_config_load(void); 324 extern nvlist_t *spa_all_configs(uint64_t *); 325 extern void spa_config_set(spa_t *spa, nvlist_t *config); 326 extern nvlist_t *spa_config_generate(spa_t *spa, vdev_t *vd, uint64_t txg, 327 int getstats); 328 329 /* 330 * Miscellaneous SPA routines in spa_misc.c 331 */ 332 333 /* Namespace manipulation */ 334 extern spa_t *spa_lookup(const char *name); 335 extern spa_t *spa_add(const char *name); 336 extern void spa_remove(spa_t *spa); 337 extern spa_t *spa_next(spa_t *prev); 338 339 /* Refcount functions */ 340 extern void spa_open_ref(spa_t *spa, void *tag); 341 extern void spa_close(spa_t *spa, void *tag); 342 extern boolean_t spa_refcount_zero(spa_t *spa); 343 344 /* Pool configuration lock */ 345 extern void spa_config_enter(spa_t *spa, krw_t rw); 346 extern void spa_config_exit(spa_t *spa); 347 extern boolean_t spa_config_held(spa_t *spa, krw_t rw); 348 349 /* Pool vdev add/remove lock */ 350 extern uint64_t spa_vdev_enter(spa_t *spa); 351 extern int spa_vdev_exit(spa_t *spa, vdev_t *vd, uint64_t txg, int error); 352 353 /* Accessor functions */ 354 extern krwlock_t *spa_traverse_rwlock(spa_t *spa); 355 extern int spa_traverse_wanted(spa_t *spa); 356 extern struct dsl_pool *spa_get_dsl(spa_t *spa); 357 extern blkptr_t *spa_get_rootblkptr(spa_t *spa); 358 extern void spa_set_rootblkptr(spa_t *spa, const blkptr_t *bp); 359 extern void spa_altroot(spa_t *, char *, size_t); 360 extern int spa_sync_pass(spa_t *spa); 361 extern char *spa_name(spa_t *spa); 362 extern uint64_t spa_guid(spa_t *spa); 363 extern uint64_t spa_last_synced_txg(spa_t *spa); 364 extern uint64_t spa_first_txg(spa_t *spa); 365 extern int spa_state(spa_t *spa); 366 extern uint64_t spa_freeze_txg(spa_t *spa); 367 struct metaslab_class; 368 extern struct metaslab_class *spa_metaslab_class_select(spa_t *spa); 369 extern uint64_t spa_get_alloc(spa_t *spa); 370 extern uint64_t spa_get_space(spa_t *spa); 371 extern uint64_t spa_get_asize(spa_t *spa, uint64_t lsize); 372 extern int spa_busy(void); 373 374 /* Miscellaneous support routines */ 375 extern int spa_rename(const char *oldname, const char *newname); 376 extern boolean_t spa_guid_exists(uint64_t pool_guid, uint64_t device_guid); 377 extern char *spa_strdup(const char *); 378 extern void spa_strfree(char *); 379 extern uint64_t spa_get_random(uint64_t range); 380 extern void sprintf_blkptr(char *buf, blkptr_t *bp); 381 extern void spa_freeze(spa_t *spa); 382 extern void spa_evict_all(void); 383 384 /* Initialization and termination */ 385 extern void spa_init(int flags); 386 extern void spa_fini(void); 387 388 #ifdef ZFS_DEBUG 389 #define dprintf_bp(bp, fmt, ...) do { \ 390 if (zfs_flags & ZFS_DEBUG_DPRINTF) { \ 391 char __blkbuf[200]; \ 392 sprintf_blkptr(__blkbuf, (bp)); \ 393 dprintf(fmt " %s\n", __VA_ARGS__, __blkbuf); \ 394 } \ 395 _NOTE(CONSTCOND) } while (0) 396 #else 397 #define dprintf_bp(bp, fmt, ...) 398 #endif 399 400 extern int spa_mode; /* mode, e.g. FREAD | FWRITE */ 401 402 #ifdef __cplusplus 403 } 404 #endif 405 406 #endif /* _SYS_SPA_H */ 407