1 // SPDX-License-Identifier: CDDL-1.0 2 /* 3 * This file and its contents are supplied under the terms of the 4 * Common Development and Distribution License ("CDDL"), version 1.0. 5 * You may only use this file in accordance with the terms of version 6 * 1.0 of the CDDL. 7 * 8 * A full copy of the text of the CDDL should have accompanied this 9 * source. A copy of the CDDL is also available via the Internet at 10 * https://opensource.org/license/CDDL-1.0. 11 */ 12 /* 13 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 14 * Copyright (c) 2011, 2024 by Delphix. All rights reserved. 15 * Copyright 2011 Nexenta Systems, Inc. All rights reserved. 16 * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved. 17 * Copyright 2013 Saso Kiselkov. All rights reserved. 18 * Copyright (c) 2014 Integros [integros.com] 19 * Copyright 2017 Joyent, Inc. 20 * Copyright (c) 2017, Intel Corporation. 21 * Copyright (c) 2019, Allan Jude 22 * Copyright (c) 2019, 2024-2026, Klara, Inc. 23 * Copyright (c) 2019, Datto Inc. 24 * Copyright (c) 2026, TrueNAS. 25 */ 26 27 #ifndef _SYS_SPA_H 28 #define _SYS_SPA_H 29 30 #include <sys/zfs_context.h> 31 #include <sys/avl.h> 32 #include <sys/kstat.h> 33 #include <sys/nvpair.h> 34 #include <sys/types.h> 35 #include <sys/fs/zfs.h> 36 #include <sys/spa_checksum.h> 37 #include <sys/dmu.h> 38 #include <sys/space_map.h> 39 #include <sys/bitops.h> 40 41 #ifdef __cplusplus 42 extern "C" { 43 #endif 44 45 /* 46 * Forward references that lots of things need. 47 */ 48 typedef struct brt_vdev brt_vdev_t; 49 typedef struct brt_dedup_shard brt_dedup_shard_t; 50 typedef struct spa spa_t; 51 typedef struct vdev vdev_t; 52 typedef struct metaslab metaslab_t; 53 typedef struct metaslab_group metaslab_group_t; 54 typedef struct metaslab_class metaslab_class_t; 55 typedef struct zio zio_t; 56 typedef struct zilog zilog_t; 57 typedef struct spa_aux_vdev spa_aux_vdev_t; 58 typedef struct zbookmark_phys zbookmark_phys_t; 59 typedef struct zbookmark_err_phys zbookmark_err_phys_t; 60 61 struct bpobj; 62 struct bplist; 63 struct dsl_pool; 64 struct dsl_dataset; 65 struct dsl_crypto_params; 66 67 /* 68 * Alignment Shift (ashift) is an immutable, internal top-level vdev property 69 * which can only be set at vdev creation time. Physical writes are always done 70 * according to it, which makes 2^ashift the smallest possible IO on a vdev. 71 * 72 * We currently allow values ranging from 512 bytes (2^9 = 512) to 64 KiB 73 * (2^16 = 65,536). 74 */ 75 #define ASHIFT_MIN 9 76 #define ASHIFT_MAX 16 77 78 /* 79 * Size of block to hold the configuration data (a packed nvlist) 80 */ 81 #define SPA_CONFIG_BLOCKSIZE (1ULL << 14) 82 83 /* 84 * The DVA size encodings for LSIZE and PSIZE support blocks up to 32MB. 85 * The ASIZE encoding should be at least 64 times larger (6 more bits) 86 * to support up to 4-way RAID-Z mirror mode with worst-case gang block 87 * overhead, three DVAs per bp, plus one more bit in case we do anything 88 * else that expands the ASIZE. 89 */ 90 #define SPA_LSIZEBITS 16 /* LSIZE up to 32M (2^16 * 512) */ 91 #define SPA_PSIZEBITS 16 /* PSIZE up to 32M (2^16 * 512) */ 92 #define SPA_ASIZEBITS 24 /* ASIZE up to 64 times larger */ 93 94 #define SPA_COMPRESSBITS 7 95 #define SPA_VDEVBITS 24 96 #define SPA_COMPRESSMASK ((1U << SPA_COMPRESSBITS) - 1) 97 98 /* 99 * All SPA data is represented by 128-bit data virtual addresses (DVAs). 100 * The members of the dva_t should be considered opaque outside the SPA. 101 */ 102 typedef struct dva { 103 uint64_t dva_word[2]; 104 } dva_t; 105 106 107 /* 108 * Some checksums/hashes need a 256-bit initialization salt. This salt is kept 109 * secret and is suitable for use in MAC algorithms as the key. 110 */ 111 typedef struct zio_cksum_salt { 112 uint8_t zcs_bytes[32]; 113 } zio_cksum_salt_t; 114 115 /* 116 * Each block is described by its DVAs, time of birth, checksum, etc. 117 * The word-by-word, bit-by-bit layout of the blkptr is as follows: 118 * 119 * 64 56 48 40 32 24 16 8 0 120 * +-------+-------+-------+-------+-------+-------+-------+-------+ 121 * 0 | pad | vdev1 | pad | ASIZE | 122 * +-------+-------+-------+-------+-------+-------+-------+-------+ 123 * 1 |G| offset1 | 124 * +-------+-------+-------+-------+-------+-------+-------+-------+ 125 * 2 | pad | vdev2 | pad | ASIZE | 126 * +-------+-------+-------+-------+-------+-------+-------+-------+ 127 * 3 |G| offset2 | 128 * +-------+-------+-------+-------+-------+-------+-------+-------+ 129 * 4 | pad | vdev3 | pad | ASIZE | 130 * +-------+-------+-------+-------+-------+-------+-------+-------+ 131 * 5 |G| offset3 | 132 * +-------+-------+-------+-------+-------+-------+-------+-------+ 133 * 6 |BDX|lvl| type | cksum |E| comp| PSIZE | LSIZE | 134 * +-------+-------+-------+-------+-------+-------+-------+-------+ 135 * 7 |R| padding | 136 * +-------+-------+-------+-------+-------+-------+-------+-------+ 137 * 8 | padding | 138 * +-------+-------+-------+-------+-------+-------+-------+-------+ 139 * 9 | physical birth txg | 140 * +-------+-------+-------+-------+-------+-------+-------+-------+ 141 * a | logical birth txg | 142 * +-------+-------+-------+-------+-------+-------+-------+-------+ 143 * b | fill count | 144 * +-------+-------+-------+-------+-------+-------+-------+-------+ 145 * c | checksum[0] | 146 * +-------+-------+-------+-------+-------+-------+-------+-------+ 147 * d | checksum[1] | 148 * +-------+-------+-------+-------+-------+-------+-------+-------+ 149 * e | checksum[2] | 150 * +-------+-------+-------+-------+-------+-------+-------+-------+ 151 * f | checksum[3] | 152 * +-------+-------+-------+-------+-------+-------+-------+-------+ 153 * 154 * Legend: 155 * 156 * vdev virtual device ID 157 * offset offset into virtual device 158 * LSIZE logical size 159 * PSIZE physical size (after compression) 160 * ASIZE allocated size (including RAID-Z parity and gang block headers) 161 * cksum checksum function 162 * comp compression function 163 * G gang block indicator 164 * B byteorder (endianness) 165 * D dedup 166 * X encryption 167 * E blkptr_t contains embedded data (see below) 168 * lvl level of indirection 169 * type DMU object type 170 * R rewrite (reallocated/rewritten at phys birth TXG) 171 * phys birth txg when dva[0] was written; zero if same as logical birth txg 172 * note that typically all the dva's would be written in this 173 * txg, but they could be different if they were moved by 174 * device removal. 175 * log. birth transaction group in which the block was logically born 176 * fill count number of non-zero blocks under this bp 177 * checksum[4] 256-bit checksum of the data this bp describes 178 */ 179 180 /* 181 * The blkptr_t's of encrypted blocks also need to store the encryption 182 * parameters so that the block can be decrypted. This layout is as follows: 183 * 184 * 64 56 48 40 32 24 16 8 0 185 * +-------+-------+-------+-------+-------+-------+-------+-------+ 186 * 0 | pad | vdev1 | pad | ASIZE | 187 * +-------+-------+-------+-------+-------+-------+-------+-------+ 188 * 1 |G| offset1 | 189 * +-------+-------+-------+-------+-------+-------+-------+-------+ 190 * 2 | pad | vdev2 | pad | ASIZE | 191 * +-------+-------+-------+-------+-------+-------+-------+-------+ 192 * 3 |G| offset2 | 193 * +-------+-------+-------+-------+-------+-------+-------+-------+ 194 * 4 | salt | 195 * +-------+-------+-------+-------+-------+-------+-------+-------+ 196 * 5 | IV1 | 197 * +-------+-------+-------+-------+-------+-------+-------+-------+ 198 * 6 |BDX|lvl| type | cksum |E| comp| PSIZE | LSIZE | 199 * +-------+-------+-------+-------+-------+-------+-------+-------+ 200 * 7 |R| padding | 201 * +-------+-------+-------+-------+-------+-------+-------+-------+ 202 * 8 | padding | 203 * +-------+-------+-------+-------+-------+-------+-------+-------+ 204 * 9 | physical birth txg | 205 * +-------+-------+-------+-------+-------+-------+-------+-------+ 206 * a | logical birth txg | 207 * +-------+-------+-------+-------+-------+-------+-------+-------+ 208 * b | IV2 | fill count | 209 * +-------+-------+-------+-------+-------+-------+-------+-------+ 210 * c | checksum[0] | 211 * +-------+-------+-------+-------+-------+-------+-------+-------+ 212 * d | checksum[1] | 213 * +-------+-------+-------+-------+-------+-------+-------+-------+ 214 * e | MAC[0] | 215 * +-------+-------+-------+-------+-------+-------+-------+-------+ 216 * f | MAC[1] | 217 * +-------+-------+-------+-------+-------+-------+-------+-------+ 218 * 219 * Legend: 220 * 221 * salt Salt for generating encryption keys 222 * IV1 First 64 bits of encryption IV 223 * X Block requires encryption handling (set to 1) 224 * E blkptr_t contains embedded data (set to 0, see below) 225 * fill count number of non-zero blocks under this bp (truncated to 32 bits) 226 * IV2 Last 32 bits of encryption IV 227 * checksum[2] 128-bit checksum of the data this bp describes 228 * MAC[2] 128-bit message authentication code for this data 229 * 230 * The X bit being set indicates that this block is one of 3 types. If this is 231 * a level 0 block with an encrypted object type, the block is encrypted 232 * (see BP_IS_ENCRYPTED()). If this is a level 0 block with an unencrypted 233 * object type, this block is authenticated with an HMAC (see 234 * BP_IS_AUTHENTICATED()). Otherwise (if level > 0), this bp will use the MAC 235 * words to store a checksum-of-MACs from the level below (see 236 * BP_HAS_INDIRECT_MAC_CKSUM()). For convenience in the code, BP_IS_PROTECTED() 237 * refers to both encrypted and authenticated blocks and BP_USES_CRYPT() 238 * refers to any of these 3 kinds of blocks. 239 * 240 * The additional encryption parameters are the salt, IV, and MAC which are 241 * explained in greater detail in the block comment at the top of zio_crypt.c. 242 * The MAC occupies half of the checksum space since it serves a very similar 243 * purpose: to prevent data corruption on disk. The only functional difference 244 * is that the checksum is used to detect on-disk corruption whether or not the 245 * encryption key is loaded and the MAC provides additional protection against 246 * malicious disk tampering. We use the 3rd DVA to store the salt and first 247 * 64 bits of the IV. As a result encrypted blocks can only have 2 copies 248 * maximum instead of the normal 3. The last 32 bits of the IV are stored in 249 * the upper bits of what is usually the fill count. Note that only blocks at 250 * level 0 or -2 are ever encrypted, which allows us to guarantee that these 251 * 32 bits are not trampled over by other code (see zio_crypt.c for details). 252 * The salt and IV are not used for authenticated bps or bps with an indirect 253 * MAC checksum, so these blocks can utilize all 3 DVAs and the full 64 bits 254 * for the fill count. 255 */ 256 257 /* 258 * "Embedded" blkptr_t's don't actually point to a block, instead they 259 * have a data payload embedded in the blkptr_t itself. See the comment 260 * in blkptr.c for more details. 261 * 262 * The blkptr_t is laid out as follows: 263 * 264 * 64 56 48 40 32 24 16 8 0 265 * +-------+-------+-------+-------+-------+-------+-------+-------+ 266 * 0 | payload | 267 * 1 | payload | 268 * 2 | payload | 269 * 3 | payload | 270 * 4 | payload | 271 * 5 | payload | 272 * +-------+-------+-------+-------+-------+-------+-------+-------+ 273 * 6 |BDX|lvl| type | etype |E| comp| PSIZE| LSIZE | 274 * +-------+-------+-------+-------+-------+-------+-------+-------+ 275 * 7 | payload | 276 * 8 | payload | 277 * 9 | payload | 278 * +-------+-------+-------+-------+-------+-------+-------+-------+ 279 * a | logical birth txg | 280 * +-------+-------+-------+-------+-------+-------+-------+-------+ 281 * b | payload | 282 * c | payload | 283 * d | payload | 284 * e | payload | 285 * f | payload | 286 * +-------+-------+-------+-------+-------+-------+-------+-------+ 287 * 288 * Legend: 289 * 290 * payload contains the embedded data 291 * B (byteorder) byteorder (endianness) 292 * D (dedup) padding (set to zero) 293 * X encryption (set to zero) 294 * E (embedded) set to one 295 * lvl indirection level 296 * type DMU object type 297 * etype how to interpret embedded data (BP_EMBEDDED_TYPE_*) 298 * comp compression function of payload 299 * PSIZE size of payload after compression, in bytes 300 * LSIZE logical size of payload, in bytes 301 * note that 25 bits is enough to store the largest 302 * "normal" BP's LSIZE (2^16 * 2^9) in bytes 303 * log. birth transaction group in which the block was logically born 304 * 305 * Note that LSIZE and PSIZE are stored in bytes, whereas for non-embedded 306 * bp's they are stored in units of SPA_MINBLOCKSHIFT. 307 * Generally, the generic BP_GET_*() macros can be used on embedded BP's. 308 * The B, D, X, lvl, type, and comp fields are stored the same as with normal 309 * BP's so the BP_SET_* macros can be used with them. etype, PSIZE, LSIZE must 310 * be set with the BPE_SET_* macros. BP_SET_EMBEDDED() should be called before 311 * other macros, as they assert that they are only used on BP's of the correct 312 * "embedded-ness". Encrypted blkptr_t's cannot be embedded because they use 313 * the payload space for encryption parameters (see the comment above on 314 * how encryption parameters are stored). 315 */ 316 317 #define BPE_GET_ETYPE(bp) \ 318 (ASSERT(BP_IS_EMBEDDED(bp)), \ 319 BF64_GET((bp)->blk_prop, 40, 8)) 320 #define BPE_SET_ETYPE(bp, t) do { \ 321 ASSERT(BP_IS_EMBEDDED(bp)); \ 322 BF64_SET((bp)->blk_prop, 40, 8, t); \ 323 } while (0) 324 325 #define BPE_GET_LSIZE(bp) \ 326 (ASSERT(BP_IS_EMBEDDED(bp)), \ 327 BF64_GET_SB((bp)->blk_prop, 0, 25, 0, 1)) 328 #define BPE_SET_LSIZE(bp, x) do { \ 329 ASSERT(BP_IS_EMBEDDED(bp)); \ 330 BF64_SET_SB((bp)->blk_prop, 0, 25, 0, 1, x); \ 331 } while (0) 332 333 #define BPE_GET_PSIZE(bp) \ 334 (ASSERT(BP_IS_EMBEDDED(bp)), \ 335 BF64_GET_SB((bp)->blk_prop, 25, 7, 0, 1)) 336 #define BPE_SET_PSIZE(bp, x) do { \ 337 ASSERT(BP_IS_EMBEDDED(bp)); \ 338 BF64_SET_SB((bp)->blk_prop, 25, 7, 0, 1, x); \ 339 } while (0) 340 341 typedef enum bp_embedded_type { 342 BP_EMBEDDED_TYPE_DATA, 343 BP_EMBEDDED_TYPE_RESERVED, /* Reserved for Delphix byteswap feature. */ 344 BP_EMBEDDED_TYPE_REDACTED, 345 NUM_BP_EMBEDDED_TYPES 346 } bp_embedded_type_t; 347 348 #define BPE_NUM_WORDS 14 349 #define BPE_PAYLOAD_SIZE (BPE_NUM_WORDS * sizeof (uint64_t)) 350 #define BPE_IS_PAYLOADWORD(bp, wp) \ 351 ((wp) != &(bp)->blk_prop && (wp) != (&(bp)->blk_birth_word[1])) 352 353 #define SPA_BLKPTRSHIFT 7 /* blkptr_t is 128 bytes */ 354 #define SPA_DVAS_PER_BP 3 /* Number of DVAs in a bp */ 355 #define SPA_SYNC_MIN_VDEVS 3 /* min vdevs to update during sync */ 356 357 /* 358 * Get number of data block pointers an indirect block could point to (given 359 * the block level and block size shift). 360 * 361 * For example, an L1 block with a blocksize of 128kb could point to: 362 * 363 * BP_SPANB(17, 1) = 1024 L0 block pointers 364 */ 365 #define BP_SPANB(indblkshift, level) \ 366 (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT))) 367 368 /* 369 * Helper function to lookup the byte range covered by a block pointer of any 370 * level (L0, L1, L2 ... etc). 371 * 372 * For example if you have an L1 block, and your blocksize is 128kb (shift 17), 373 * then your block can cover this many bytes: 374 * 375 * BP_BYTE_RANGE(17, 1) = 134217728 bytes 376 */ 377 #define BP_BYTE_RANGE(indblkshift, level) \ 378 (BP_SPANB(indblkshift, level) * ((uint64_t)1 << indblkshift)) 379 380 /* 381 * A block is a hole when it has either 1) never been written to, or 382 * 2) is zero-filled. In both cases, ZFS can return all zeroes for all reads 383 * without physically allocating disk space. Holes are represented in the 384 * blkptr_t structure by zeroed blk_dva. Correct checking for holes is 385 * done through the BP_IS_HOLE macro. For holes, the logical size, level, 386 * DMU object type, and birth times are all also stored for holes that 387 * were written to at some point (i.e. were punched after having been filled). 388 */ 389 typedef struct blkptr { 390 dva_t blk_dva[SPA_DVAS_PER_BP]; /* Data Virtual Addresses */ 391 uint64_t blk_prop; /* size, compression, type, etc */ 392 uint64_t blk_prop2; /* additional properties */ 393 uint64_t blk_pad; /* Extra space for the future */ 394 uint64_t blk_birth_word[2]; 395 uint64_t blk_fill; /* fill count */ 396 zio_cksum_t blk_cksum; /* 256-bit checksum */ 397 } blkptr_t; 398 399 /* 400 * Macros to get and set fields in a bp or DVA. 401 */ 402 403 /* 404 * Note, for gang blocks, DVA_GET_ASIZE() is the total space allocated for 405 * this gang DVA including its children BP's. The space allocated at this 406 * DVA's vdev/offset is vdev_gang_header_asize(vdev). 407 */ 408 #define DVA_GET_ASIZE(dva) \ 409 BF64_GET_SB((dva)->dva_word[0], 0, SPA_ASIZEBITS, SPA_MINBLOCKSHIFT, 0) 410 #define DVA_SET_ASIZE(dva, x) \ 411 BF64_SET_SB((dva)->dva_word[0], 0, SPA_ASIZEBITS, \ 412 SPA_MINBLOCKSHIFT, 0, x) 413 414 #define DVA_GET_VDEV(dva) BF64_GET((dva)->dva_word[0], 32, SPA_VDEVBITS) 415 #define DVA_SET_VDEV(dva, x) \ 416 BF64_SET((dva)->dva_word[0], 32, SPA_VDEVBITS, x) 417 418 #define DVA_GET_OFFSET(dva) \ 419 BF64_GET_SB((dva)->dva_word[1], 0, 63, SPA_MINBLOCKSHIFT, 0) 420 #define DVA_SET_OFFSET(dva, x) \ 421 BF64_SET_SB((dva)->dva_word[1], 0, 63, SPA_MINBLOCKSHIFT, 0, x) 422 423 #define DVA_GET_GANG(dva) BF64_GET((dva)->dva_word[1], 63, 1) 424 #define DVA_SET_GANG(dva, x) BF64_SET((dva)->dva_word[1], 63, 1, x) 425 426 #define BP_GET_LSIZE(bp) \ 427 (BP_IS_EMBEDDED(bp) ? \ 428 (BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA ? BPE_GET_LSIZE(bp) : 0): \ 429 BF64_GET_SB((bp)->blk_prop, 0, SPA_LSIZEBITS, SPA_MINBLOCKSHIFT, 1)) 430 #define BP_SET_LSIZE(bp, x) do { \ 431 ASSERT(!BP_IS_EMBEDDED(bp)); \ 432 BF64_SET_SB((bp)->blk_prop, \ 433 0, SPA_LSIZEBITS, SPA_MINBLOCKSHIFT, 1, x); \ 434 } while (0) 435 436 #define BP_GET_PSIZE(bp) \ 437 (BP_IS_EMBEDDED(bp) ? 0 : \ 438 BF64_GET_SB((bp)->blk_prop, 16, SPA_PSIZEBITS, SPA_MINBLOCKSHIFT, 1)) 439 #define BP_SET_PSIZE(bp, x) do { \ 440 ASSERT(!BP_IS_EMBEDDED(bp)); \ 441 BF64_SET_SB((bp)->blk_prop, \ 442 16, SPA_PSIZEBITS, SPA_MINBLOCKSHIFT, 1, x); \ 443 } while (0) 444 445 #define BP_GET_COMPRESS(bp) \ 446 BF64_GET((bp)->blk_prop, 32, SPA_COMPRESSBITS) 447 #define BP_SET_COMPRESS(bp, x) \ 448 BF64_SET((bp)->blk_prop, 32, SPA_COMPRESSBITS, x) 449 450 #define BP_IS_EMBEDDED(bp) BF64_GET((bp)->blk_prop, 39, 1) 451 #define BP_SET_EMBEDDED(bp, x) BF64_SET((bp)->blk_prop, 39, 1, x) 452 453 #define BP_GET_CHECKSUM(bp) \ 454 (BP_IS_EMBEDDED(bp) ? ZIO_CHECKSUM_OFF : \ 455 BF64_GET((bp)->blk_prop, 40, 8)) 456 #define BP_SET_CHECKSUM(bp, x) do { \ 457 ASSERT(!BP_IS_EMBEDDED(bp)); \ 458 BF64_SET((bp)->blk_prop, 40, 8, x); \ 459 } while (0) 460 461 #define BP_GET_TYPE(bp) BF64_GET((bp)->blk_prop, 48, 8) 462 #define BP_SET_TYPE(bp, x) BF64_SET((bp)->blk_prop, 48, 8, x) 463 464 #define BP_GET_LEVEL(bp) BF64_GET((bp)->blk_prop, 56, 5) 465 #define BP_SET_LEVEL(bp, x) BF64_SET((bp)->blk_prop, 56, 5, x) 466 467 /* encrypted, authenticated, and MAC cksum bps use the same bit */ 468 #define BP_USES_CRYPT(bp) BF64_GET((bp)->blk_prop, 61, 1) 469 #define BP_SET_CRYPT(bp, x) BF64_SET((bp)->blk_prop, 61, 1, x) 470 471 #define BP_IS_ENCRYPTED(bp) \ 472 (BP_USES_CRYPT(bp) && \ 473 BP_GET_LEVEL(bp) <= 0 && \ 474 DMU_OT_IS_ENCRYPTED(BP_GET_TYPE(bp))) 475 476 #define BP_IS_AUTHENTICATED(bp) \ 477 (BP_USES_CRYPT(bp) && \ 478 BP_GET_LEVEL(bp) <= 0 && \ 479 !DMU_OT_IS_ENCRYPTED(BP_GET_TYPE(bp))) 480 481 #define BP_HAS_INDIRECT_MAC_CKSUM(bp) \ 482 (BP_USES_CRYPT(bp) && BP_GET_LEVEL(bp) > 0) 483 484 #define BP_IS_PROTECTED(bp) \ 485 (BP_IS_ENCRYPTED(bp) || BP_IS_AUTHENTICATED(bp)) 486 487 #define BP_GET_DEDUP(bp) BF64_GET((bp)->blk_prop, 62, 1) 488 #define BP_SET_DEDUP(bp, x) BF64_SET((bp)->blk_prop, 62, 1, x) 489 490 #define BP_GET_BYTEORDER(bp) BF64_GET((bp)->blk_prop, 63, 1) 491 #define BP_SET_BYTEORDER(bp, x) BF64_SET((bp)->blk_prop, 63, 1, x) 492 493 #define BP_GET_FREE(bp) BF64_GET((bp)->blk_fill, 0, 1) 494 #define BP_SET_FREE(bp, x) BF64_SET((bp)->blk_fill, 0, 1, x) 495 496 /* 497 * Block birth time macros for different use cases: 498 * - BP_GET_LOGICAL_BIRTH(): When the block was logically modified by user. 499 * To be used with a focus on user data, like incremental replication. 500 * - BP_GET_PHYSICAL_BIRTH(): When the block was physically written to disks. 501 * For regular writes is equal to logical birth. For dedup and block cloning 502 * can be smaller than logical birth. For remapped and rewritten blocks can 503 * be bigger. To be used with focus on physical disk content: ARC, DDT, scrub. 504 * - BP_GET_RAW_PHYSICAL_BIRTH(): Raw physical birth value. Zero if equal 505 * to logical birth. Should only be used for BP copying and debugging. 506 * - BP_GET_BIRTH(): When the block was allocated, which is a physical birth 507 * for rewritten blocks (rewrite flag set) or logical birth otherwise. 508 */ 509 #define BP_GET_LOGICAL_BIRTH(bp) (bp)->blk_birth_word[1] 510 #define BP_SET_LOGICAL_BIRTH(bp, x) ((bp)->blk_birth_word[1] = (x)) 511 512 #define BP_GET_RAW_PHYSICAL_BIRTH(bp) (bp)->blk_birth_word[0] 513 #define BP_SET_PHYSICAL_BIRTH(bp, x) ((bp)->blk_birth_word[0] = (x)) 514 515 #define BP_GET_PHYSICAL_BIRTH(bp) \ 516 (BP_IS_EMBEDDED(bp) ? 0 : \ 517 BP_GET_RAW_PHYSICAL_BIRTH(bp) ? BP_GET_RAW_PHYSICAL_BIRTH(bp) : \ 518 BP_GET_LOGICAL_BIRTH(bp)) 519 520 #define BP_GET_BIRTH(bp) \ 521 ((BP_IS_EMBEDDED(bp) || !BP_GET_REWRITE(bp)) ? \ 522 BP_GET_LOGICAL_BIRTH(bp) : BP_GET_PHYSICAL_BIRTH(bp)) 523 524 #define BP_SET_BIRTH(bp, logical, physical) \ 525 { \ 526 ASSERT(!BP_IS_EMBEDDED(bp)); \ 527 BP_SET_LOGICAL_BIRTH(bp, logical); \ 528 BP_SET_PHYSICAL_BIRTH(bp, \ 529 ((logical) == (physical) ? 0 : (physical))); \ 530 } 531 532 #define BP_GET_FILL(bp) \ 533 (BP_IS_EMBEDDED(bp) ? 1 : \ 534 BP_IS_ENCRYPTED(bp) ? BF64_GET((bp)->blk_fill, 0, 32) : \ 535 (bp)->blk_fill) 536 537 #define BP_SET_FILL(bp, fill) \ 538 { \ 539 ASSERT(!BP_IS_EMBEDDED(bp)); \ 540 if (BP_IS_ENCRYPTED(bp)) \ 541 BF64_SET((bp)->blk_fill, 0, 32, fill); \ 542 else \ 543 (bp)->blk_fill = fill; \ 544 } 545 546 #define BP_GET_IV2(bp) \ 547 (ASSERT(BP_IS_ENCRYPTED(bp)), \ 548 BF64_GET((bp)->blk_fill, 32, 32)) 549 #define BP_SET_IV2(bp, iv2) \ 550 { \ 551 ASSERT(BP_IS_ENCRYPTED(bp)); \ 552 BF64_SET((bp)->blk_fill, 32, 32, iv2); \ 553 } 554 555 #define BP_GET_REWRITE(bp) \ 556 (BP_IS_EMBEDDED(bp) ? 0 : BF64_GET((bp)->blk_prop2, 63, 1)) 557 558 #define BP_SET_REWRITE(bp, x) \ 559 { \ 560 ASSERT(!BP_IS_EMBEDDED(bp)); \ 561 BF64_SET((bp)->blk_prop2, 63, 1, x); \ 562 } 563 564 #define BP_IS_METADATA(bp) \ 565 (BP_GET_LEVEL(bp) > 0 || DMU_OT_IS_METADATA(BP_GET_TYPE(bp))) 566 567 #define BP_GET_ASIZE(bp) \ 568 (BP_IS_EMBEDDED(bp) ? 0 : \ 569 DVA_GET_ASIZE(&(bp)->blk_dva[0]) + \ 570 DVA_GET_ASIZE(&(bp)->blk_dva[1]) + \ 571 (DVA_GET_ASIZE(&(bp)->blk_dva[2]) * !BP_IS_ENCRYPTED(bp))) 572 573 #define BP_GET_UCSIZE(bp) \ 574 (BP_IS_METADATA(bp) ? BP_GET_PSIZE(bp) : BP_GET_LSIZE(bp)) 575 576 #define BP_GET_NDVAS(bp) \ 577 (BP_IS_EMBEDDED(bp) ? 0 : \ 578 !!DVA_GET_ASIZE(&(bp)->blk_dva[0]) + \ 579 !!DVA_GET_ASIZE(&(bp)->blk_dva[1]) + \ 580 (!!DVA_GET_ASIZE(&(bp)->blk_dva[2]) * !BP_IS_ENCRYPTED(bp))) 581 582 #define BP_COUNT_GANG(bp) \ 583 (BP_IS_EMBEDDED(bp) ? 0 : \ 584 (DVA_GET_GANG(&(bp)->blk_dva[0]) + \ 585 DVA_GET_GANG(&(bp)->blk_dva[1]) + \ 586 (DVA_GET_GANG(&(bp)->blk_dva[2]) * !BP_IS_ENCRYPTED(bp)))) 587 588 #define DVA_EQUAL(dva1, dva2) \ 589 ((dva1)->dva_word[1] == (dva2)->dva_word[1] && \ 590 (dva1)->dva_word[0] == (dva2)->dva_word[0]) 591 592 #define BP_EQUAL(bp1, bp2) \ 593 (BP_GET_PHYSICAL_BIRTH(bp1) == BP_GET_PHYSICAL_BIRTH(bp2) && \ 594 BP_GET_LOGICAL_BIRTH(bp1) == BP_GET_LOGICAL_BIRTH(bp2) && \ 595 DVA_EQUAL(&(bp1)->blk_dva[0], &(bp2)->blk_dva[0]) && \ 596 DVA_EQUAL(&(bp1)->blk_dva[1], &(bp2)->blk_dva[1]) && \ 597 DVA_EQUAL(&(bp1)->blk_dva[2], &(bp2)->blk_dva[2])) 598 599 600 #define DVA_IS_VALID(dva) (DVA_GET_ASIZE(dva) != 0) 601 602 #define BP_IDENTITY(bp) (ASSERT(!BP_IS_EMBEDDED(bp)), &(bp)->blk_dva[0]) 603 #define BP_IS_GANG(bp) \ 604 (BP_IS_EMBEDDED(bp) ? B_FALSE : DVA_GET_GANG(BP_IDENTITY(bp))) 605 #define DVA_IS_EMPTY(dva) ((dva)->dva_word[0] == 0ULL && \ 606 (dva)->dva_word[1] == 0ULL) 607 #define BP_IS_HOLE(bp) \ 608 (!BP_IS_EMBEDDED(bp) && DVA_IS_EMPTY(BP_IDENTITY(bp))) 609 610 #define BP_SET_REDACTED(bp) \ 611 { \ 612 BP_SET_EMBEDDED(bp, B_TRUE); \ 613 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_REDACTED); \ 614 } 615 #define BP_IS_REDACTED(bp) \ 616 (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_REDACTED) 617 618 /* BP_IS_RAIDZ(bp) assumes no block compression */ 619 #define BP_IS_RAIDZ(bp) (DVA_GET_ASIZE(&(bp)->blk_dva[0]) > \ 620 BP_GET_PSIZE(bp)) 621 622 #define BP_ZERO_DVAS(bp) \ 623 { \ 624 (bp)->blk_dva[0].dva_word[0] = 0; \ 625 (bp)->blk_dva[0].dva_word[1] = 0; \ 626 (bp)->blk_dva[1].dva_word[0] = 0; \ 627 (bp)->blk_dva[1].dva_word[1] = 0; \ 628 (bp)->blk_dva[2].dva_word[0] = 0; \ 629 (bp)->blk_dva[2].dva_word[1] = 0; \ 630 } 631 632 #define BP_ZERO(bp) \ 633 { \ 634 BP_ZERO_DVAS(bp); \ 635 (bp)->blk_prop = 0; \ 636 (bp)->blk_prop2 = 0; \ 637 (bp)->blk_pad = 0; \ 638 (bp)->blk_birth_word[0] = 0; \ 639 (bp)->blk_birth_word[1] = 0; \ 640 (bp)->blk_fill = 0; \ 641 ZIO_SET_CHECKSUM(&(bp)->blk_cksum, 0, 0, 0, 0); \ 642 } 643 644 #ifdef _ZFS_BIG_ENDIAN 645 #define ZFS_HOST_BYTEORDER (0ULL) 646 #else 647 #define ZFS_HOST_BYTEORDER (1ULL) 648 #endif 649 650 #define BP_SHOULD_BYTESWAP(bp) (BP_GET_BYTEORDER(bp) != ZFS_HOST_BYTEORDER) 651 652 #define BP_SPRINTF_LEN 400 653 654 /* 655 * This macro allows code sharing between zfs, libzpool, and mdb. 656 * 'func' is either kmem_scnprintf() or mdb_snprintf(). 657 * 'ws' (whitespace) can be ' ' for single-line format, '\n' for multi-line. 658 */ 659 660 #define SNPRINTF_BLKPTR(func, ws, buf, size, bp, type, checksum, compress) \ 661 { \ 662 static const char *const copyname[] = \ 663 { "zero", "single", "double", "triple" }; \ 664 int len = 0; \ 665 int copies = 0; \ 666 const char *crypt_type; \ 667 if (bp != NULL) { \ 668 if (BP_IS_ENCRYPTED(bp)) { \ 669 crypt_type = "encrypted"; \ 670 /* LINTED E_SUSPICIOUS_COMPARISON */ \ 671 } else if (BP_IS_AUTHENTICATED(bp)) { \ 672 crypt_type = "authenticated"; \ 673 } else if (BP_HAS_INDIRECT_MAC_CKSUM(bp)) { \ 674 crypt_type = "indirect-MAC"; \ 675 } else { \ 676 crypt_type = "unencrypted"; \ 677 } \ 678 } \ 679 if (bp == NULL) { \ 680 len += func(buf + len, size - len, "<NULL>"); \ 681 } else if (BP_IS_HOLE(bp)) { \ 682 len += func(buf + len, size - len, \ 683 "HOLE [L%llu %s] " \ 684 "size=%llxL birth=%lluL", \ 685 (u_longlong_t)BP_GET_LEVEL(bp), \ 686 type, \ 687 (u_longlong_t)BP_GET_LSIZE(bp), \ 688 (u_longlong_t)BP_GET_LOGICAL_BIRTH(bp)); \ 689 } else if (BP_IS_EMBEDDED(bp)) { \ 690 len = func(buf + len, size - len, \ 691 "EMBEDDED [L%llu %s] et=%u %s " \ 692 "size=%llxL/%llxP birth=%lluL", \ 693 (u_longlong_t)BP_GET_LEVEL(bp), \ 694 type, \ 695 (int)BPE_GET_ETYPE(bp), \ 696 compress, \ 697 (u_longlong_t)BPE_GET_LSIZE(bp), \ 698 (u_longlong_t)BPE_GET_PSIZE(bp), \ 699 (u_longlong_t)BP_GET_LOGICAL_BIRTH(bp)); \ 700 } else if (BP_IS_REDACTED(bp)) { \ 701 len += func(buf + len, size - len, \ 702 "REDACTED [L%llu %s] size=%llxL birth=%lluL", \ 703 (u_longlong_t)BP_GET_LEVEL(bp), \ 704 type, \ 705 (u_longlong_t)BP_GET_LSIZE(bp), \ 706 (u_longlong_t)BP_GET_LOGICAL_BIRTH(bp)); \ 707 } else { \ 708 for (int d = 0; d < BP_GET_NDVAS(bp); d++) { \ 709 const dva_t *dva = &bp->blk_dva[d]; \ 710 if (DVA_IS_VALID(dva)) \ 711 copies++; \ 712 len += func(buf + len, size - len, \ 713 "DVA[%d]=<%llu:%llx:%llx>%c", d, \ 714 (u_longlong_t)DVA_GET_VDEV(dva), \ 715 (u_longlong_t)DVA_GET_OFFSET(dva), \ 716 (u_longlong_t)DVA_GET_ASIZE(dva), \ 717 ws); \ 718 } \ 719 if (BP_IS_ENCRYPTED(bp)) { \ 720 len += func(buf + len, size - len, \ 721 "salt=%llx iv=%llx:%llx%c", \ 722 (u_longlong_t)bp->blk_dva[2].dva_word[0], \ 723 (u_longlong_t)bp->blk_dva[2].dva_word[1], \ 724 (u_longlong_t)BP_GET_IV2(bp), \ 725 ws); \ 726 } \ 727 len += func(buf + len, size - len, \ 728 "[L%llu %s] %s %s %s %s %s %s %s%c" \ 729 "size=%llxL/%llxP birth=%lluL/%lluP fill=%llu%c" \ 730 "cksum=%016llx:%016llx:%016llx:%016llx", \ 731 (u_longlong_t)BP_GET_LEVEL(bp), \ 732 type, \ 733 checksum, \ 734 compress, \ 735 crypt_type, \ 736 BP_GET_BYTEORDER(bp) == 0 ? "BE" : "LE", \ 737 BP_IS_GANG(bp) ? "gang" : "contiguous", \ 738 BP_GET_DEDUP(bp) ? "dedup" : "unique", \ 739 copyname[copies], \ 740 ws, \ 741 (u_longlong_t)BP_GET_LSIZE(bp), \ 742 (u_longlong_t)BP_GET_PSIZE(bp), \ 743 (u_longlong_t)BP_GET_LOGICAL_BIRTH(bp), \ 744 (u_longlong_t)BP_GET_PHYSICAL_BIRTH(bp), \ 745 (u_longlong_t)BP_GET_FILL(bp), \ 746 ws, \ 747 (u_longlong_t)bp->blk_cksum.zc_word[0], \ 748 (u_longlong_t)bp->blk_cksum.zc_word[1], \ 749 (u_longlong_t)bp->blk_cksum.zc_word[2], \ 750 (u_longlong_t)bp->blk_cksum.zc_word[3]); \ 751 } \ 752 ASSERT(len < size); \ 753 } 754 755 #define BP_GET_BUFC_TYPE(bp) \ 756 (BP_IS_METADATA(bp) ? ARC_BUFC_METADATA : ARC_BUFC_DATA) 757 758 typedef enum spa_import_type { 759 SPA_IMPORT_EXISTING, 760 SPA_IMPORT_ASSEMBLE 761 } spa_import_type_t; 762 763 typedef enum spa_mode { 764 SPA_MODE_UNINIT = 0, 765 SPA_MODE_READ = 1, 766 SPA_MODE_WRITE = 2, 767 } spa_mode_t; 768 769 /* 770 * Send TRIM commands in-line during normal pool operation while deleting. 771 * OFF: no 772 * ON: yes 773 */ 774 typedef enum { 775 SPA_AUTOTRIM_OFF = 0, /* default */ 776 SPA_AUTOTRIM_ON, 777 } spa_autotrim_t; 778 779 /* 780 * Reason TRIM command was issued, used internally for accounting purposes. 781 */ 782 typedef enum trim_type { 783 TRIM_TYPE_MANUAL = 0, 784 TRIM_TYPE_AUTO = 1, 785 TRIM_TYPE_SIMPLE = 2 786 } trim_type_t; 787 788 /* state manipulation functions */ 789 extern int spa_open(const char *pool, spa_t **, const void *tag); 790 extern int spa_open_rewind(const char *pool, spa_t **, const void *tag, 791 nvlist_t *policy, nvlist_t **config); 792 extern int spa_get_stats(const char *pool, nvlist_t **config, char *altroot, 793 size_t buflen); 794 extern int spa_create(const char *pool, nvlist_t *nvroot, nvlist_t *props, 795 nvlist_t *zplprops, struct dsl_crypto_params *dcp, nvlist_t **errinfo); 796 extern int spa_import(char *pool, nvlist_t *config, nvlist_t *props, 797 uint64_t flags); 798 extern nvlist_t *spa_tryimport(nvlist_t *tryconfig); 799 extern int spa_destroy(const char *pool); 800 extern int spa_checkpoint(const char *pool); 801 extern int spa_checkpoint_discard(const char *pool); 802 extern int spa_export(const char *pool, nvlist_t **oldconfig, boolean_t force, 803 boolean_t hardforce); 804 extern int spa_reset(const char *pool); 805 extern void spa_async_request(spa_t *spa, int flag); 806 extern void spa_async_unrequest(spa_t *spa, int flag); 807 extern void spa_async_suspend(spa_t *spa); 808 extern void spa_async_resume(spa_t *spa); 809 extern int spa_async_tasks(spa_t *spa); 810 extern spa_t *spa_inject_addref(char *pool); 811 extern void spa_inject_delref(spa_t *spa); 812 extern void spa_scan_stat_init(spa_t *spa); 813 extern int spa_scan_get_stats(spa_t *spa, pool_scan_stat_t *ps); 814 extern int bpobj_enqueue_alloc_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx); 815 extern int bpobj_enqueue_free_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx); 816 817 #define SPA_ASYNC_CONFIG_UPDATE 0x01 818 #define SPA_ASYNC_REMOVE 0x02 819 #define SPA_ASYNC_FAULT_VDEV 0x04 820 #define SPA_ASYNC_RESILVER_DONE 0x08 821 #define SPA_ASYNC_RESILVER 0x10 822 #define SPA_ASYNC_AUTOEXPAND 0x20 823 #define SPA_ASYNC_REMOVE_DONE 0x40 824 #define SPA_ASYNC_REMOVE_STOP 0x80 825 #define SPA_ASYNC_INITIALIZE_RESTART 0x100 826 #define SPA_ASYNC_TRIM_RESTART 0x200 827 #define SPA_ASYNC_AUTOTRIM_RESTART 0x400 828 #define SPA_ASYNC_L2CACHE_REBUILD 0x800 829 #define SPA_ASYNC_L2CACHE_TRIM 0x1000 830 #define SPA_ASYNC_REBUILD_DONE 0x2000 831 #define SPA_ASYNC_DETACH_SPARE 0x4000 832 #define SPA_ASYNC_REMOVE_BY_USER 0x8000 833 #define SPA_ASYNC_DEFER_DESTROY 0x10000 834 835 /* device manipulation */ 836 extern int spa_vdev_add(spa_t *spa, nvlist_t *nvroot, boolean_t ashift_check); 837 extern int spa_vdev_attach(spa_t *spa, uint64_t guid, nvlist_t *nvroot, 838 int replacing, int rebuild); 839 extern int spa_vdev_detach(spa_t *spa, uint64_t guid, uint64_t pguid, 840 int replace_done); 841 extern int spa_vdev_alloc(spa_t *spa, uint64_t guid); 842 extern int spa_vdev_noalloc(spa_t *spa, uint64_t guid); 843 extern boolean_t spa_vdev_remove_active(spa_t *spa); 844 extern int spa_vdev_initialize(spa_t *spa, nvlist_t *nv, uint64_t cmd_type, 845 uint64_t value, boolean_t value_provided, nvlist_t *vdev_errlist); 846 extern int spa_vdev_trim(spa_t *spa, nvlist_t *nv, uint64_t cmd_type, 847 uint64_t rate, boolean_t partial, boolean_t secure, nvlist_t *vdev_errlist); 848 extern int spa_vdev_setpath(spa_t *spa, uint64_t guid, const char *newpath); 849 extern int spa_vdev_setfru(spa_t *spa, uint64_t guid, const char *newfru); 850 extern int spa_vdev_split_mirror(spa_t *spa, const char *newname, 851 nvlist_t *config, nvlist_t *props, boolean_t exp); 852 853 /* spare state (which is global across all pools) */ 854 extern void spa_spare_add(vdev_t *vd); 855 extern void spa_spare_remove(vdev_t *vd); 856 extern boolean_t spa_spare_exists(uint64_t guid, uint64_t *pool, int *refcnt); 857 extern void spa_spare_activate(vdev_t *vd); 858 859 /* L2ARC state (which is global across all pools) */ 860 extern void spa_l2cache_add(vdev_t *vd); 861 extern void spa_l2cache_remove(vdev_t *vd); 862 extern boolean_t spa_l2cache_exists(uint64_t guid, uint64_t *pool); 863 extern void spa_l2cache_activate(vdev_t *vd); 864 extern void spa_l2cache_drop(spa_t *spa); 865 866 /* scanning */ 867 extern int spa_scan(spa_t *spa, pool_scan_func_t func, 868 pool_scrub_flags_t flags); 869 extern int spa_scan_range(spa_t *spa, pool_scan_func_t func, uint64_t txgstart, 870 uint64_t txgend, pool_scrub_flags_t flags); 871 extern int spa_scan_stop(spa_t *spa); 872 extern int spa_scrub_pause_resume(spa_t *spa, pool_scrub_cmd_t flag); 873 874 /* spa syncing */ 875 extern void spa_sync(spa_t *spa, uint64_t txg); /* only for DMU use */ 876 extern void spa_sync_allpools(void); 877 878 extern uint_t zfs_sync_pass_deferred_free; 879 880 /* spa sync taskqueues */ 881 taskq_t *spa_sync_tq_create(spa_t *spa, const char *name); 882 void spa_sync_tq_destroy(spa_t *spa); 883 uint_t spa_acq_allocator(spa_t *spa); 884 void spa_rel_allocator(spa_t *spa, uint_t allocator); 885 void spa_select_allocator(zio_t *zio); 886 887 /* spa namespace global lock */ 888 extern void spa_namespace_enter(const void *tag); 889 extern boolean_t spa_namespace_tryenter(const void *tag); 890 extern int spa_namespace_enter_interruptible(const void *tag); 891 extern void spa_namespace_exit(const void *tag); 892 extern boolean_t spa_namespace_held(void); 893 extern void spa_namespace_wait(void); 894 extern void spa_namespace_broadcast(void); 895 896 /* 897 * SPA configuration functions in spa_config.c 898 */ 899 900 #define SPA_CONFIG_UPDATE_POOL 0 901 #define SPA_CONFIG_UPDATE_VDEVS 1 902 903 extern void spa_write_cachefile(spa_t *, boolean_t, boolean_t, boolean_t); 904 extern int spa_all_configs(uint64_t *generation, nvlist_t **pools); 905 extern void spa_config_set(spa_t *spa, nvlist_t *config); 906 extern nvlist_t *spa_config_generate(spa_t *spa, vdev_t *vd, uint64_t txg, 907 int getstats); 908 extern void spa_config_update(spa_t *spa, int what); 909 extern int spa_config_parse(spa_t *spa, vdev_t **vdp, nvlist_t *nv, 910 vdev_t *parent, uint_t id, int atype); 911 912 913 /* 914 * Miscellaneous SPA routines in spa_misc.c 915 */ 916 917 /* Namespace manipulation */ 918 extern spa_t *spa_lookup(const char *name); 919 extern spa_t *spa_add(const char *name, nvlist_t *config, const char *altroot); 920 extern void spa_remove(spa_t *spa); 921 extern spa_t *spa_next(spa_t *prev); 922 923 /* Refcount functions */ 924 extern void spa_open_ref(spa_t *spa, const void *tag); 925 extern void spa_close(spa_t *spa, const void *tag); 926 extern void spa_async_close(spa_t *spa, const void *tag); 927 extern boolean_t spa_refcount_zero(spa_t *spa); 928 929 #define SCL_NONE 0x00 930 #define SCL_CONFIG 0x01 931 #define SCL_STATE 0x02 932 #define SCL_L2ARC 0x04 /* hack until L2ARC 2.0 */ 933 #define SCL_ALLOC 0x08 934 #define SCL_ZIO 0x10 935 #define SCL_FREE 0x20 936 #define SCL_VDEV 0x40 937 #define SCL_LOCKS 7 938 #define SCL_ALL ((1 << SCL_LOCKS) - 1) 939 #define SCL_STATE_ALL (SCL_STATE | SCL_L2ARC | SCL_ZIO) 940 941 /* Historical pool statistics */ 942 typedef struct spa_history_kstat { 943 kmutex_t lock; 944 uint64_t count; 945 uint64_t size; 946 kstat_t *kstat; 947 void *priv; 948 list_t list; 949 } spa_history_kstat_t; 950 951 typedef struct spa_history_list { 952 uint64_t size; 953 procfs_list_t procfs_list; 954 } spa_history_list_t; 955 956 typedef struct spa_stats { 957 spa_history_list_t read_history; 958 spa_history_list_t txg_history; 959 spa_history_kstat_t tx_assign_histogram; 960 spa_history_list_t mmp_history; 961 spa_history_kstat_t state; /* pool state */ 962 spa_history_kstat_t guid; /* pool guid */ 963 spa_history_kstat_t iostats; 964 spa_history_kstat_t log_spacemaps; 965 } spa_stats_t; 966 967 typedef enum txg_state { 968 TXG_STATE_BIRTH = 0, 969 TXG_STATE_OPEN = 1, 970 TXG_STATE_QUIESCED = 2, 971 TXG_STATE_WAIT_FOR_SYNC = 3, 972 TXG_STATE_SYNCED = 4, 973 TXG_STATE_COMMITTED = 5, 974 } txg_state_t; 975 976 typedef struct txg_stat { 977 vdev_stat_t vs1; 978 vdev_stat_t vs2; 979 uint64_t txg; 980 uint64_t ndirty; 981 } txg_stat_t; 982 983 /* Assorted pool IO kstats */ 984 typedef struct spa_iostats { 985 kstat_named_t trim_extents_written; 986 kstat_named_t trim_bytes_written; 987 kstat_named_t trim_extents_skipped; 988 kstat_named_t trim_bytes_skipped; 989 kstat_named_t trim_extents_failed; 990 kstat_named_t trim_bytes_failed; 991 kstat_named_t autotrim_extents_written; 992 kstat_named_t autotrim_bytes_written; 993 kstat_named_t autotrim_extents_skipped; 994 kstat_named_t autotrim_bytes_skipped; 995 kstat_named_t autotrim_extents_failed; 996 kstat_named_t autotrim_bytes_failed; 997 kstat_named_t simple_trim_extents_written; 998 kstat_named_t simple_trim_bytes_written; 999 kstat_named_t simple_trim_extents_skipped; 1000 kstat_named_t simple_trim_bytes_skipped; 1001 kstat_named_t simple_trim_extents_failed; 1002 kstat_named_t simple_trim_bytes_failed; 1003 kstat_named_t arc_read_count; 1004 kstat_named_t arc_read_bytes; 1005 kstat_named_t arc_write_count; 1006 kstat_named_t arc_write_bytes; 1007 kstat_named_t direct_read_count; 1008 kstat_named_t direct_read_bytes; 1009 kstat_named_t direct_write_count; 1010 kstat_named_t direct_write_bytes; 1011 } spa_iostats_t; 1012 1013 extern void spa_stats_init(spa_t *spa); 1014 extern void spa_stats_destroy(spa_t *spa); 1015 extern void spa_read_history_add(spa_t *spa, const zbookmark_phys_t *zb, 1016 uint32_t aflags); 1017 extern void spa_txg_history_add(spa_t *spa, uint64_t txg, hrtime_t birth_time); 1018 extern int spa_txg_history_set(spa_t *spa, uint64_t txg, 1019 txg_state_t completed_state, hrtime_t completed_time); 1020 extern txg_stat_t *spa_txg_history_init_io(spa_t *, uint64_t, 1021 struct dsl_pool *); 1022 extern void spa_txg_history_fini_io(spa_t *, txg_stat_t *); 1023 extern void spa_tx_assign_add_nsecs(spa_t *spa, uint64_t nsecs); 1024 extern int spa_mmp_history_set_skip(spa_t *spa, uint64_t mmp_kstat_id); 1025 extern int spa_mmp_history_set(spa_t *spa, uint64_t mmp_kstat_id, int io_error, 1026 hrtime_t duration); 1027 extern void spa_mmp_history_add(spa_t *spa, uint64_t txg, uint64_t timestamp, 1028 uint64_t mmp_delay, vdev_t *vd, int label, uint64_t mmp_kstat_id, 1029 int error); 1030 extern void spa_iostats_trim_add(spa_t *spa, trim_type_t type, 1031 uint64_t extents_written, uint64_t bytes_written, 1032 uint64_t extents_skipped, uint64_t bytes_skipped, 1033 uint64_t extents_failed, uint64_t bytes_failed); 1034 extern void spa_iostats_read_add(spa_t *spa, uint64_t size, uint64_t iops, 1035 dmu_flags_t flags); 1036 extern void spa_iostats_write_add(spa_t *spa, uint64_t size, uint64_t iops, 1037 dmu_flags_t flags); 1038 extern void spa_import_progress_add(spa_t *spa); 1039 extern void spa_import_progress_remove(uint64_t spa_guid); 1040 extern int spa_import_progress_set_mmp_check(uint64_t pool_guid, 1041 uint64_t mmp_sec_remaining); 1042 extern int spa_import_progress_set_max_txg(uint64_t pool_guid, 1043 uint64_t max_txg); 1044 extern int spa_import_progress_set_state(uint64_t pool_guid, 1045 spa_load_state_t spa_load_state); 1046 extern void spa_import_progress_set_notes(spa_t *spa, 1047 const char *fmt, ...) __printflike(2, 3); 1048 extern void spa_import_progress_set_notes_nolog(spa_t *spa, 1049 const char *fmt, ...) __printflike(2, 3); 1050 1051 /* Pool configuration locks */ 1052 extern int spa_config_tryenter(spa_t *spa, int locks, const void *tag, 1053 krw_t rw); 1054 extern void spa_config_enter(spa_t *spa, int locks, const void *tag, krw_t rw); 1055 extern void spa_config_enter_priority(spa_t *spa, int locks, const void *tag, 1056 krw_t rw); 1057 extern void spa_config_exit(spa_t *spa, int locks, const void *tag); 1058 extern int spa_config_held(spa_t *spa, int locks, krw_t rw); 1059 1060 /* Pool vdev add/remove lock */ 1061 extern uint64_t spa_vdev_enter(spa_t *spa); 1062 extern uint64_t spa_vdev_detach_enter(spa_t *spa, uint64_t guid); 1063 extern uint64_t spa_vdev_config_enter(spa_t *spa); 1064 extern void spa_vdev_config_exit(spa_t *spa, vdev_t *vd, uint64_t txg, 1065 int error, const char *tag); 1066 extern int spa_vdev_exit(spa_t *spa, vdev_t *vd, uint64_t txg, int error); 1067 1068 /* Pool vdev state change lock */ 1069 extern void spa_vdev_state_enter(spa_t *spa, int oplock); 1070 extern int spa_vdev_state_exit(spa_t *spa, vdev_t *vd, int error); 1071 1072 /* Log state */ 1073 typedef enum spa_log_state { 1074 SPA_LOG_UNKNOWN = 0, /* unknown log state */ 1075 SPA_LOG_MISSING, /* missing log(s) */ 1076 SPA_LOG_CLEAR, /* clear the log(s) */ 1077 SPA_LOG_GOOD, /* log(s) are good */ 1078 } spa_log_state_t; 1079 1080 extern spa_log_state_t spa_get_log_state(spa_t *spa); 1081 extern void spa_set_log_state(spa_t *spa, spa_log_state_t state); 1082 extern int spa_reset_logs(spa_t *spa); 1083 extern void spa_log_sm_stats_update(spa_t *spa); 1084 1085 /* Log claim callback */ 1086 extern void spa_claim_notify(zio_t *zio); 1087 extern void spa_deadman(void *); 1088 1089 /* Accessor functions */ 1090 extern boolean_t spa_shutting_down(spa_t *spa); 1091 extern struct dsl_pool *spa_get_dsl(spa_t *spa); 1092 extern boolean_t spa_is_initializing(spa_t *spa); 1093 extern boolean_t spa_indirect_vdevs_loaded(spa_t *spa); 1094 extern blkptr_t *spa_get_rootblkptr(spa_t *spa); 1095 extern void spa_set_rootblkptr(spa_t *spa, const blkptr_t *bp); 1096 extern void spa_altroot(spa_t *, char *, size_t); 1097 extern uint32_t spa_sync_pass(spa_t *spa); 1098 extern char *spa_name(spa_t *spa); 1099 extern char *spa_load_name(spa_t *spa); 1100 extern uint64_t spa_guid(spa_t *spa); 1101 extern uint64_t spa_load_guid(spa_t *spa); 1102 extern uint64_t spa_last_synced_txg(spa_t *spa); 1103 extern uint64_t spa_first_txg(spa_t *spa); 1104 extern uint64_t spa_open_txg(spa_t *spa); 1105 extern uint64_t spa_syncing_txg(spa_t *spa); 1106 extern uint64_t spa_final_dirty_txg(spa_t *spa); 1107 extern uint64_t spa_version(spa_t *spa); 1108 extern pool_state_t spa_state(spa_t *spa); 1109 extern spa_load_state_t spa_load_state(spa_t *spa); 1110 extern uint64_t spa_freeze_txg(spa_t *spa); 1111 extern uint64_t spa_get_worst_case_asize(spa_t *spa, uint64_t lsize); 1112 extern void spa_get_min_alloc_range(spa_t *spa, uint64_t *min, uint64_t *max); 1113 extern uint64_t spa_get_dspace(spa_t *spa); 1114 extern uint64_t spa_get_checkpoint_space(spa_t *spa); 1115 extern uint64_t spa_get_slop_space(spa_t *spa); 1116 extern void spa_update_dspace(spa_t *spa); 1117 extern uint64_t spa_version(spa_t *spa); 1118 extern boolean_t spa_deflate(spa_t *spa); 1119 extern metaslab_class_t *spa_normal_class(spa_t *spa); 1120 extern metaslab_class_t *spa_log_class(spa_t *spa); 1121 extern metaslab_class_t *spa_embedded_log_class(spa_t *spa); 1122 extern metaslab_class_t *spa_special_class(spa_t *spa); 1123 extern metaslab_class_t *spa_special_embedded_log_class(spa_t *spa); 1124 extern metaslab_class_t *spa_dedup_class(spa_t *spa); 1125 extern metaslab_class_t *spa_preferred_class(spa_t *spa, const zio_t *zio); 1126 extern boolean_t spa_special_has_ddt(spa_t *spa); 1127 1128 extern void spa_evicting_os_register(spa_t *, objset_t *os); 1129 extern void spa_evicting_os_deregister(spa_t *, objset_t *os); 1130 extern void spa_evicting_os_wait(spa_t *spa); 1131 extern int spa_max_replication(spa_t *spa); 1132 extern int spa_prev_software_version(spa_t *spa); 1133 extern uint64_t spa_get_failmode(spa_t *spa); 1134 extern uint64_t spa_get_deadman_failmode(spa_t *spa); 1135 extern void spa_set_deadman_failmode(spa_t *spa, const char *failmode); 1136 extern boolean_t spa_suspended(spa_t *spa); 1137 extern uint64_t spa_bootfs(spa_t *spa); 1138 extern uint64_t spa_get_last_scrubbed_txg(spa_t *spa); 1139 extern uint64_t spa_delegation(spa_t *spa); 1140 extern objset_t *spa_meta_objset(spa_t *spa); 1141 extern space_map_t *spa_syncing_log_sm(spa_t *spa); 1142 extern uint64_t spa_deadman_synctime(spa_t *spa); 1143 extern uint64_t spa_deadman_ziotime(spa_t *spa); 1144 extern uint64_t spa_dirty_data(spa_t *spa); 1145 extern spa_autotrim_t spa_get_autotrim(spa_t *spa); 1146 extern int spa_get_allocator(spa_t *spa); 1147 extern void spa_set_allocator(spa_t *spa, const char *allocator); 1148 1149 /* Miscellaneous support routines */ 1150 extern void spa_load_failed(spa_t *spa, const char *fmt, ...) 1151 __attribute__((format(__printf__, 2, 3))); 1152 extern void spa_load_note(spa_t *spa, const char *fmt, ...) 1153 __attribute__((format(__printf__, 2, 3))); 1154 extern void spa_activate_mos_feature(spa_t *spa, const char *feature, 1155 dmu_tx_t *tx); 1156 extern void spa_deactivate_mos_feature(spa_t *spa, const char *feature); 1157 extern spa_t *spa_by_guid(uint64_t pool_guid, uint64_t device_guid); 1158 extern boolean_t spa_guid_exists(uint64_t pool_guid, uint64_t device_guid); 1159 extern char *spa_strdup(const char *); 1160 extern void spa_strfree(char *); 1161 extern uint64_t spa_generate_guid(spa_t *spa); 1162 extern uint64_t spa_generate_load_guid(void); 1163 extern void snprintf_blkptr(char *buf, size_t buflen, const blkptr_t *bp); 1164 extern void spa_freeze(spa_t *spa); 1165 extern int spa_change_guid(spa_t *spa, const uint64_t *guidp); 1166 extern void spa_upgrade(spa_t *spa, uint64_t version); 1167 extern void spa_evict_all(void); 1168 extern vdev_t *spa_lookup_by_guid(spa_t *spa, uint64_t guid, 1169 boolean_t l2cache); 1170 extern boolean_t spa_has_l2cache(spa_t *, uint64_t guid); 1171 extern boolean_t spa_has_spare(spa_t *, uint64_t guid); 1172 extern uint64_t dva_get_dsize_sync(spa_t *spa, const dva_t *dva); 1173 extern uint64_t bp_get_dsize_sync(spa_t *spa, const blkptr_t *bp); 1174 extern uint64_t bp_get_dsize(spa_t *spa, const blkptr_t *bp); 1175 extern boolean_t spa_has_dedup(spa_t *spa); 1176 extern boolean_t spa_has_slogs(spa_t *spa); 1177 extern boolean_t spa_has_special(spa_t *spa); 1178 extern boolean_t spa_is_root(spa_t *spa); 1179 extern boolean_t spa_writeable(spa_t *spa); 1180 extern boolean_t spa_has_pending_synctask(spa_t *spa); 1181 extern int spa_maxblocksize(spa_t *spa); 1182 extern int spa_maxdnodesize(spa_t *spa); 1183 extern boolean_t spa_has_checkpoint(spa_t *spa); 1184 extern boolean_t spa_importing_readonly_checkpoint(spa_t *spa); 1185 extern boolean_t spa_suspend_async_destroy(spa_t *spa); 1186 extern uint64_t spa_min_claim_txg(spa_t *spa); 1187 extern boolean_t zfs_dva_valid(spa_t *spa, const dva_t *dva, 1188 const blkptr_t *bp); 1189 typedef void (*spa_remap_cb_t)(uint64_t vdev, uint64_t offset, uint64_t size, 1190 void *arg); 1191 extern boolean_t spa_remap_blkptr(spa_t *spa, blkptr_t *bp, 1192 spa_remap_cb_t callback, void *arg); 1193 extern uint64_t spa_get_last_removal_txg(spa_t *spa); 1194 extern boolean_t spa_trust_config(spa_t *spa); 1195 extern uint64_t spa_missing_tvds_allowed(spa_t *spa); 1196 extern void spa_set_missing_tvds(spa_t *spa, uint64_t missing); 1197 extern boolean_t spa_top_vdevs_spacemap_addressable(spa_t *spa); 1198 extern uint64_t spa_total_metaslabs(spa_t *spa); 1199 extern boolean_t spa_multihost(spa_t *spa); 1200 extern uint32_t spa_get_hostid(spa_t *spa); 1201 extern void spa_activate_allocation_classes(spa_t *, dmu_tx_t *); 1202 extern boolean_t spa_livelist_delete_check(spa_t *spa); 1203 1204 extern boolean_t spa_mmp_remote_host_activity(spa_t *spa); 1205 1206 extern spa_mode_t spa_mode(spa_t *spa); 1207 extern uint64_t zfs_strtonum(const char *str, char **nptr); 1208 1209 extern char *spa_his_ievent_table[]; 1210 1211 extern void spa_history_create_obj(spa_t *spa, dmu_tx_t *tx); 1212 extern int spa_history_get(spa_t *spa, uint64_t *offset, uint64_t *len_read, 1213 char *his_buf); 1214 extern int spa_history_log(spa_t *spa, const char *his_buf); 1215 extern int spa_history_log_nvl(spa_t *spa, nvlist_t *nvl); 1216 extern void spa_history_log_version(spa_t *spa, const char *operation, 1217 dmu_tx_t *tx); 1218 extern void spa_history_log_internal(spa_t *spa, const char *operation, 1219 dmu_tx_t *tx, const char *fmt, ...) __printflike(4, 5); 1220 extern void spa_history_log_internal_ds(struct dsl_dataset *ds, const char *op, 1221 dmu_tx_t *tx, const char *fmt, ...) __printflike(4, 5); 1222 extern void spa_history_log_internal_dd(dsl_dir_t *dd, const char *operation, 1223 dmu_tx_t *tx, const char *fmt, ...) __printflike(4, 5); 1224 1225 extern const char *spa_state_to_name(spa_t *spa); 1226 1227 /* error handling */ 1228 struct zbookmark_phys; 1229 extern void spa_log_error(spa_t *spa, const zbookmark_phys_t *zb, 1230 const uint64_t birth); 1231 extern void spa_remove_error(spa_t *spa, zbookmark_phys_t *zb, 1232 uint64_t birth); 1233 extern int zfs_ereport_post(const char *clazz, spa_t *spa, vdev_t *vd, 1234 const zbookmark_phys_t *zb, zio_t *zio, uint64_t state); 1235 extern boolean_t zfs_ereport_is_valid(const char *clazz, spa_t *spa, vdev_t *vd, 1236 zio_t *zio); 1237 extern void zfs_ereport_taskq_fini(void); 1238 extern void zfs_ereport_clear(spa_t *spa, vdev_t *vd); 1239 extern nvlist_t *zfs_event_create(spa_t *spa, vdev_t *vd, const char *type, 1240 const char *name, nvlist_t *aux); 1241 extern void zfs_post_remove(spa_t *spa, vdev_t *vd, boolean_t by_kernel); 1242 extern void zfs_post_state_change(spa_t *spa, vdev_t *vd, uint64_t laststate); 1243 extern void zfs_post_autoreplace(spa_t *spa, vdev_t *vd); 1244 extern uint64_t spa_approx_errlog_size(spa_t *spa); 1245 extern int spa_get_errlog(spa_t *spa, void *uaddr, uint64_t *count); 1246 extern uint64_t spa_get_last_errlog_size(spa_t *spa); 1247 extern void spa_errlog_rotate(spa_t *spa); 1248 extern void spa_errlog_drain(spa_t *spa); 1249 extern void spa_errlog_sync(spa_t *spa, uint64_t txg); 1250 extern void spa_get_errlists(spa_t *spa, avl_tree_t *last, avl_tree_t *scrub); 1251 extern void spa_delete_dataset_errlog(spa_t *spa, uint64_t ds, dmu_tx_t *tx); 1252 extern void spa_swap_errlog(spa_t *spa, uint64_t new_head_ds, 1253 uint64_t old_head_ds, dmu_tx_t *tx); 1254 extern void sync_error_list(spa_t *spa, avl_tree_t *t, uint64_t *obj, 1255 dmu_tx_t *tx); 1256 extern void spa_upgrade_errlog(spa_t *spa, dmu_tx_t *tx); 1257 extern int find_top_affected_fs(spa_t *spa, uint64_t head_ds, 1258 zbookmark_err_phys_t *zep, uint64_t *top_affected_fs); 1259 extern int find_birth_txg(struct dsl_dataset *ds, zbookmark_err_phys_t *zep, 1260 uint64_t *birth_txg); 1261 extern void zep_to_zb(uint64_t dataset, zbookmark_err_phys_t *zep, 1262 zbookmark_phys_t *zb); 1263 extern void name_to_errphys(char *buf, zbookmark_err_phys_t *zep); 1264 1265 /* vdev mirror */ 1266 extern void vdev_mirror_stat_init(void); 1267 extern void vdev_mirror_stat_fini(void); 1268 1269 /* Initialization and termination */ 1270 extern void spa_init(spa_mode_t mode); 1271 extern void spa_fini(void); 1272 1273 /* properties */ 1274 extern int spa_prop_set(spa_t *spa, nvlist_t *nvp); 1275 extern int spa_prop_get(spa_t *spa, nvlist_t *nvp); 1276 extern int spa_prop_get_nvlist(spa_t *spa, char **props, 1277 unsigned int n_props, nvlist_t *outnvl); 1278 extern void spa_prop_clear_bootfs(spa_t *spa, uint64_t obj, dmu_tx_t *tx); 1279 extern void spa_configfile_set(spa_t *, nvlist_t *, boolean_t); 1280 1281 /* asynchronous event notification */ 1282 extern void spa_event_notify(spa_t *spa, vdev_t *vdev, nvlist_t *hist_nvl, 1283 const char *name); 1284 extern void zfs_ereport_zvol_post(const char *subclass, const char *name, 1285 const char *device_name, const char *raw_name); 1286 1287 /* waiting for pool activities to complete */ 1288 extern int spa_wait(const char *pool, zpool_wait_activity_t activity, 1289 boolean_t *waited); 1290 extern int spa_wait_tag(const char *name, zpool_wait_activity_t activity, 1291 uint64_t tag, boolean_t *waited); 1292 extern void spa_notify_waiters(spa_t *spa); 1293 extern void spa_wake_waiters(spa_t *spa); 1294 1295 /* a no-op condense op for test & debug */ 1296 #ifdef ZFS_DEBUG 1297 extern void spa_condense_debug_start(spa_t *spa); 1298 extern void spa_condense_debug_cancel(spa_t *spa); 1299 #endif 1300 1301 extern void spa_import_os(spa_t *spa); 1302 extern void spa_export_os(spa_t *spa); 1303 extern void spa_activate_os(spa_t *spa); 1304 extern void spa_deactivate_os(spa_t *spa); 1305 1306 /* module param call functions */ 1307 int param_set_deadman_ziotime(ZFS_MODULE_PARAM_ARGS); 1308 int param_set_deadman_synctime(ZFS_MODULE_PARAM_ARGS); 1309 int param_set_slop_shift(ZFS_MODULE_PARAM_ARGS); 1310 int param_set_deadman_failmode(ZFS_MODULE_PARAM_ARGS); 1311 int param_set_active_allocator(ZFS_MODULE_PARAM_ARGS); 1312 1313 #ifdef ZFS_DEBUG 1314 #define dprintf_bp(bp, fmt, ...) do { \ 1315 if (zfs_flags & ZFS_DEBUG_DPRINTF) { \ 1316 char *__blkbuf = kmem_alloc(BP_SPRINTF_LEN, KM_SLEEP); \ 1317 snprintf_blkptr(__blkbuf, BP_SPRINTF_LEN, (bp)); \ 1318 dprintf(fmt " %s\n", __VA_ARGS__, __blkbuf); \ 1319 kmem_free(__blkbuf, BP_SPRINTF_LEN); \ 1320 } \ 1321 } while (0) 1322 #else 1323 #define dprintf_bp(bp, fmt, ...) 1324 #endif 1325 1326 extern spa_mode_t spa_mode_global; 1327 extern int zfs_deadman_enabled; 1328 extern uint64_t zfs_deadman_synctime_ms; 1329 extern uint64_t zfs_deadman_ziotime_ms; 1330 extern uint64_t zfs_deadman_checktime_ms; 1331 1332 extern kmem_cache_t *zio_buf_cache[]; 1333 extern kmem_cache_t *zio_data_buf_cache[]; 1334 1335 #ifdef __cplusplus 1336 } 1337 #endif 1338 1339 #endif /* _SYS_SPA_H */ 1340