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 (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or https://opensource.org/licenses/CDDL-1.0. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21 /* 22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 23 * Copyright (c) 2011, 2024 by Delphix. All rights reserved. 24 * Copyright 2011 Nexenta Systems, Inc. All rights reserved. 25 * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved. 26 * Copyright 2013 Saso Kiselkov. All rights reserved. 27 * Copyright (c) 2014 Integros [integros.com] 28 * Copyright 2017 Joyent, Inc. 29 * Copyright (c) 2017, Intel Corporation. 30 * Copyright (c) 2019, Allan Jude 31 * Copyright (c) 2019, Klara Inc. 32 * Copyright (c) 2019, Datto Inc. 33 */ 34 35 #ifndef _SYS_SPA_H 36 #define _SYS_SPA_H 37 38 #include <sys/zfs_context.h> 39 #include <sys/avl.h> 40 #include <sys/kstat.h> 41 #include <sys/nvpair.h> 42 #include <sys/types.h> 43 #include <sys/fs/zfs.h> 44 #include <sys/spa_checksum.h> 45 #include <sys/dmu.h> 46 #include <sys/space_map.h> 47 #include <sys/bitops.h> 48 49 #ifdef __cplusplus 50 extern "C" { 51 #endif 52 53 /* 54 * Forward references that lots of things need. 55 */ 56 typedef struct spa spa_t; 57 typedef struct vdev vdev_t; 58 typedef struct metaslab metaslab_t; 59 typedef struct metaslab_group metaslab_group_t; 60 typedef struct metaslab_class metaslab_class_t; 61 typedef struct zio zio_t; 62 typedef struct zilog zilog_t; 63 typedef struct spa_aux_vdev spa_aux_vdev_t; 64 typedef struct zbookmark_phys zbookmark_phys_t; 65 typedef struct zbookmark_err_phys zbookmark_err_phys_t; 66 67 struct bpobj; 68 struct bplist; 69 struct dsl_pool; 70 struct dsl_dataset; 71 struct dsl_crypto_params; 72 73 /* 74 * Alignment Shift (ashift) is an immutable, internal top-level vdev property 75 * which can only be set at vdev creation time. Physical writes are always done 76 * according to it, which makes 2^ashift the smallest possible IO on a vdev. 77 * 78 * We currently allow values ranging from 512 bytes (2^9 = 512) to 64 KiB 79 * (2^16 = 65,536). 80 */ 81 #define ASHIFT_MIN 9 82 #define ASHIFT_MAX 16 83 84 /* 85 * Size of block to hold the configuration data (a packed nvlist) 86 */ 87 #define SPA_CONFIG_BLOCKSIZE (1ULL << 14) 88 89 /* 90 * The DVA size encodings for LSIZE and PSIZE support blocks up to 32MB. 91 * The ASIZE encoding should be at least 64 times larger (6 more bits) 92 * to support up to 4-way RAID-Z mirror mode with worst-case gang block 93 * overhead, three DVAs per bp, plus one more bit in case we do anything 94 * else that expands the ASIZE. 95 */ 96 #define SPA_LSIZEBITS 16 /* LSIZE up to 32M (2^16 * 512) */ 97 #define SPA_PSIZEBITS 16 /* PSIZE up to 32M (2^16 * 512) */ 98 #define SPA_ASIZEBITS 24 /* ASIZE up to 64 times larger */ 99 100 #define SPA_COMPRESSBITS 7 101 #define SPA_VDEVBITS 24 102 #define SPA_COMPRESSMASK ((1U << SPA_COMPRESSBITS) - 1) 103 104 /* 105 * All SPA data is represented by 128-bit data virtual addresses (DVAs). 106 * The members of the dva_t should be considered opaque outside the SPA. 107 */ 108 typedef struct dva { 109 uint64_t dva_word[2]; 110 } dva_t; 111 112 113 /* 114 * Some checksums/hashes need a 256-bit initialization salt. This salt is kept 115 * secret and is suitable for use in MAC algorithms as the key. 116 */ 117 typedef struct zio_cksum_salt { 118 uint8_t zcs_bytes[32]; 119 } zio_cksum_salt_t; 120 121 /* 122 * Each block is described by its DVAs, time of birth, checksum, etc. 123 * The word-by-word, bit-by-bit layout of the blkptr is as follows: 124 * 125 * 64 56 48 40 32 24 16 8 0 126 * +-------+-------+-------+-------+-------+-------+-------+-------+ 127 * 0 | pad | vdev1 | pad | ASIZE | 128 * +-------+-------+-------+-------+-------+-------+-------+-------+ 129 * 1 |G| offset1 | 130 * +-------+-------+-------+-------+-------+-------+-------+-------+ 131 * 2 | pad | vdev2 | pad | ASIZE | 132 * +-------+-------+-------+-------+-------+-------+-------+-------+ 133 * 3 |G| offset2 | 134 * +-------+-------+-------+-------+-------+-------+-------+-------+ 135 * 4 | pad | vdev3 | pad | ASIZE | 136 * +-------+-------+-------+-------+-------+-------+-------+-------+ 137 * 5 |G| offset3 | 138 * +-------+-------+-------+-------+-------+-------+-------+-------+ 139 * 6 |BDX|lvl| type | cksum |E| comp| PSIZE | LSIZE | 140 * +-------+-------+-------+-------+-------+-------+-------+-------+ 141 * 7 | padding | 142 * +-------+-------+-------+-------+-------+-------+-------+-------+ 143 * 8 | padding | 144 * +-------+-------+-------+-------+-------+-------+-------+-------+ 145 * 9 | physical birth txg | 146 * +-------+-------+-------+-------+-------+-------+-------+-------+ 147 * a | logical birth txg | 148 * +-------+-------+-------+-------+-------+-------+-------+-------+ 149 * b | fill count | 150 * +-------+-------+-------+-------+-------+-------+-------+-------+ 151 * c | checksum[0] | 152 * +-------+-------+-------+-------+-------+-------+-------+-------+ 153 * d | checksum[1] | 154 * +-------+-------+-------+-------+-------+-------+-------+-------+ 155 * e | checksum[2] | 156 * +-------+-------+-------+-------+-------+-------+-------+-------+ 157 * f | checksum[3] | 158 * +-------+-------+-------+-------+-------+-------+-------+-------+ 159 * 160 * Legend: 161 * 162 * vdev virtual device ID 163 * offset offset into virtual device 164 * LSIZE logical size 165 * PSIZE physical size (after compression) 166 * ASIZE allocated size (including RAID-Z parity and gang block headers) 167 * cksum checksum function 168 * comp compression function 169 * G gang block indicator 170 * B byteorder (endianness) 171 * D dedup 172 * X encryption 173 * E blkptr_t contains embedded data (see below) 174 * lvl level of indirection 175 * type DMU object type 176 * phys birth txg when dva[0] was written; zero if same as logical birth txg 177 * note that typically all the dva's would be written in this 178 * txg, but they could be different if they were moved by 179 * device removal. 180 * log. birth transaction group in which the block was logically born 181 * fill count number of non-zero blocks under this bp 182 * checksum[4] 256-bit checksum of the data this bp describes 183 */ 184 185 /* 186 * The blkptr_t's of encrypted blocks also need to store the encryption 187 * parameters so that the block can be decrypted. This layout is as follows: 188 * 189 * 64 56 48 40 32 24 16 8 0 190 * +-------+-------+-------+-------+-------+-------+-------+-------+ 191 * 0 | vdev1 | pad | ASIZE | 192 * +-------+-------+-------+-------+-------+-------+-------+-------+ 193 * 1 |G| offset1 | 194 * +-------+-------+-------+-------+-------+-------+-------+-------+ 195 * 2 | vdev2 | pad | ASIZE | 196 * +-------+-------+-------+-------+-------+-------+-------+-------+ 197 * 3 |G| offset2 | 198 * +-------+-------+-------+-------+-------+-------+-------+-------+ 199 * 4 | salt | 200 * +-------+-------+-------+-------+-------+-------+-------+-------+ 201 * 5 | IV1 | 202 * +-------+-------+-------+-------+-------+-------+-------+-------+ 203 * 6 |BDX|lvl| type | cksum |E| comp| PSIZE | LSIZE | 204 * +-------+-------+-------+-------+-------+-------+-------+-------+ 205 * 7 | padding | 206 * +-------+-------+-------+-------+-------+-------+-------+-------+ 207 * 8 | padding | 208 * +-------+-------+-------+-------+-------+-------+-------+-------+ 209 * 9 | physical birth txg | 210 * +-------+-------+-------+-------+-------+-------+-------+-------+ 211 * a | logical birth txg | 212 * +-------+-------+-------+-------+-------+-------+-------+-------+ 213 * b | IV2 | fill count | 214 * +-------+-------+-------+-------+-------+-------+-------+-------+ 215 * c | checksum[0] | 216 * +-------+-------+-------+-------+-------+-------+-------+-------+ 217 * d | checksum[1] | 218 * +-------+-------+-------+-------+-------+-------+-------+-------+ 219 * e | MAC[0] | 220 * +-------+-------+-------+-------+-------+-------+-------+-------+ 221 * f | MAC[1] | 222 * +-------+-------+-------+-------+-------+-------+-------+-------+ 223 * 224 * Legend: 225 * 226 * salt Salt for generating encryption keys 227 * IV1 First 64 bits of encryption IV 228 * X Block requires encryption handling (set to 1) 229 * E blkptr_t contains embedded data (set to 0, see below) 230 * fill count number of non-zero blocks under this bp (truncated to 32 bits) 231 * IV2 Last 32 bits of encryption IV 232 * checksum[2] 128-bit checksum of the data this bp describes 233 * MAC[2] 128-bit message authentication code for this data 234 * 235 * The X bit being set indicates that this block is one of 3 types. If this is 236 * a level 0 block with an encrypted object type, the block is encrypted 237 * (see BP_IS_ENCRYPTED()). If this is a level 0 block with an unencrypted 238 * object type, this block is authenticated with an HMAC (see 239 * BP_IS_AUTHENTICATED()). Otherwise (if level > 0), this bp will use the MAC 240 * words to store a checksum-of-MACs from the level below (see 241 * BP_HAS_INDIRECT_MAC_CKSUM()). For convenience in the code, BP_IS_PROTECTED() 242 * refers to both encrypted and authenticated blocks and BP_USES_CRYPT() 243 * refers to any of these 3 kinds of blocks. 244 * 245 * The additional encryption parameters are the salt, IV, and MAC which are 246 * explained in greater detail in the block comment at the top of zio_crypt.c. 247 * The MAC occupies half of the checksum space since it serves a very similar 248 * purpose: to prevent data corruption on disk. The only functional difference 249 * is that the checksum is used to detect on-disk corruption whether or not the 250 * encryption key is loaded and the MAC provides additional protection against 251 * malicious disk tampering. We use the 3rd DVA to store the salt and first 252 * 64 bits of the IV. As a result encrypted blocks can only have 2 copies 253 * maximum instead of the normal 3. The last 32 bits of the IV are stored in 254 * the upper bits of what is usually the fill count. Note that only blocks at 255 * level 0 or -2 are ever encrypted, which allows us to guarantee that these 256 * 32 bits are not trampled over by other code (see zio_crypt.c for details). 257 * The salt and IV are not used for authenticated bps or bps with an indirect 258 * MAC checksum, so these blocks can utilize all 3 DVAs and the full 64 bits 259 * for the fill count. 260 */ 261 262 /* 263 * "Embedded" blkptr_t's don't actually point to a block, instead they 264 * have a data payload embedded in the blkptr_t itself. See the comment 265 * in blkptr.c for more details. 266 * 267 * The blkptr_t is laid out as follows: 268 * 269 * 64 56 48 40 32 24 16 8 0 270 * +-------+-------+-------+-------+-------+-------+-------+-------+ 271 * 0 | payload | 272 * 1 | payload | 273 * 2 | payload | 274 * 3 | payload | 275 * 4 | payload | 276 * 5 | payload | 277 * +-------+-------+-------+-------+-------+-------+-------+-------+ 278 * 6 |BDX|lvl| type | etype |E| comp| PSIZE| LSIZE | 279 * +-------+-------+-------+-------+-------+-------+-------+-------+ 280 * 7 | payload | 281 * 8 | payload | 282 * 9 | payload | 283 * +-------+-------+-------+-------+-------+-------+-------+-------+ 284 * a | logical birth txg | 285 * +-------+-------+-------+-------+-------+-------+-------+-------+ 286 * b | payload | 287 * c | payload | 288 * d | payload | 289 * e | payload | 290 * f | payload | 291 * +-------+-------+-------+-------+-------+-------+-------+-------+ 292 * 293 * Legend: 294 * 295 * payload contains the embedded data 296 * B (byteorder) byteorder (endianness) 297 * D (dedup) padding (set to zero) 298 * X encryption (set to zero) 299 * E (embedded) set to one 300 * lvl indirection level 301 * type DMU object type 302 * etype how to interpret embedded data (BP_EMBEDDED_TYPE_*) 303 * comp compression function of payload 304 * PSIZE size of payload after compression, in bytes 305 * LSIZE logical size of payload, in bytes 306 * note that 25 bits is enough to store the largest 307 * "normal" BP's LSIZE (2^16 * 2^9) in bytes 308 * log. birth transaction group in which the block was logically born 309 * 310 * Note that LSIZE and PSIZE are stored in bytes, whereas for non-embedded 311 * bp's they are stored in units of SPA_MINBLOCKSHIFT. 312 * Generally, the generic BP_GET_*() macros can be used on embedded BP's. 313 * The B, D, X, lvl, type, and comp fields are stored the same as with normal 314 * BP's so the BP_SET_* macros can be used with them. etype, PSIZE, LSIZE must 315 * be set with the BPE_SET_* macros. BP_SET_EMBEDDED() should be called before 316 * other macros, as they assert that they are only used on BP's of the correct 317 * "embedded-ness". Encrypted blkptr_t's cannot be embedded because they use 318 * the payload space for encryption parameters (see the comment above on 319 * how encryption parameters are stored). 320 */ 321 322 #define BPE_GET_ETYPE(bp) \ 323 (ASSERT(BP_IS_EMBEDDED(bp)), \ 324 BF64_GET((bp)->blk_prop, 40, 8)) 325 #define BPE_SET_ETYPE(bp, t) do { \ 326 ASSERT(BP_IS_EMBEDDED(bp)); \ 327 BF64_SET((bp)->blk_prop, 40, 8, t); \ 328 } while (0) 329 330 #define BPE_GET_LSIZE(bp) \ 331 (ASSERT(BP_IS_EMBEDDED(bp)), \ 332 BF64_GET_SB((bp)->blk_prop, 0, 25, 0, 1)) 333 #define BPE_SET_LSIZE(bp, x) do { \ 334 ASSERT(BP_IS_EMBEDDED(bp)); \ 335 BF64_SET_SB((bp)->blk_prop, 0, 25, 0, 1, x); \ 336 } while (0) 337 338 #define BPE_GET_PSIZE(bp) \ 339 (ASSERT(BP_IS_EMBEDDED(bp)), \ 340 BF64_GET_SB((bp)->blk_prop, 25, 7, 0, 1)) 341 #define BPE_SET_PSIZE(bp, x) do { \ 342 ASSERT(BP_IS_EMBEDDED(bp)); \ 343 BF64_SET_SB((bp)->blk_prop, 25, 7, 0, 1, x); \ 344 } while (0) 345 346 typedef enum bp_embedded_type { 347 BP_EMBEDDED_TYPE_DATA, 348 BP_EMBEDDED_TYPE_RESERVED, /* Reserved for Delphix byteswap feature. */ 349 BP_EMBEDDED_TYPE_REDACTED, 350 NUM_BP_EMBEDDED_TYPES 351 } bp_embedded_type_t; 352 353 #define BPE_NUM_WORDS 14 354 #define BPE_PAYLOAD_SIZE (BPE_NUM_WORDS * sizeof (uint64_t)) 355 #define BPE_IS_PAYLOADWORD(bp, wp) \ 356 ((wp) != &(bp)->blk_prop && (wp) != (&(bp)->blk_birth_word[1])) 357 358 #define SPA_BLKPTRSHIFT 7 /* blkptr_t is 128 bytes */ 359 #define SPA_DVAS_PER_BP 3 /* Number of DVAs in a bp */ 360 #define SPA_SYNC_MIN_VDEVS 3 /* min vdevs to update during sync */ 361 362 /* 363 * A block is a hole when it has either 1) never been written to, or 364 * 2) is zero-filled. In both cases, ZFS can return all zeroes for all reads 365 * without physically allocating disk space. Holes are represented in the 366 * blkptr_t structure by zeroed blk_dva. Correct checking for holes is 367 * done through the BP_IS_HOLE macro. For holes, the logical size, level, 368 * DMU object type, and birth times are all also stored for holes that 369 * were written to at some point (i.e. were punched after having been filled). 370 */ 371 typedef struct blkptr { 372 dva_t blk_dva[SPA_DVAS_PER_BP]; /* Data Virtual Addresses */ 373 uint64_t blk_prop; /* size, compression, type, etc */ 374 uint64_t blk_pad[2]; /* Extra space for the future */ 375 uint64_t blk_birth_word[2]; 376 uint64_t blk_fill; /* fill count */ 377 zio_cksum_t blk_cksum; /* 256-bit checksum */ 378 } blkptr_t; 379 380 /* 381 * Macros to get and set fields in a bp or DVA. 382 */ 383 384 /* 385 * Note, for gang blocks, DVA_GET_ASIZE() is the total space allocated for 386 * this gang DVA including its children BP's. The space allocated at this 387 * DVA's vdev/offset is vdev_gang_header_asize(vdev). 388 */ 389 #define DVA_GET_ASIZE(dva) \ 390 BF64_GET_SB((dva)->dva_word[0], 0, SPA_ASIZEBITS, SPA_MINBLOCKSHIFT, 0) 391 #define DVA_SET_ASIZE(dva, x) \ 392 BF64_SET_SB((dva)->dva_word[0], 0, SPA_ASIZEBITS, \ 393 SPA_MINBLOCKSHIFT, 0, x) 394 395 #define DVA_GET_VDEV(dva) BF64_GET((dva)->dva_word[0], 32, SPA_VDEVBITS) 396 #define DVA_SET_VDEV(dva, x) \ 397 BF64_SET((dva)->dva_word[0], 32, SPA_VDEVBITS, x) 398 399 #define DVA_GET_OFFSET(dva) \ 400 BF64_GET_SB((dva)->dva_word[1], 0, 63, SPA_MINBLOCKSHIFT, 0) 401 #define DVA_SET_OFFSET(dva, x) \ 402 BF64_SET_SB((dva)->dva_word[1], 0, 63, SPA_MINBLOCKSHIFT, 0, x) 403 404 #define DVA_GET_GANG(dva) BF64_GET((dva)->dva_word[1], 63, 1) 405 #define DVA_SET_GANG(dva, x) BF64_SET((dva)->dva_word[1], 63, 1, x) 406 407 #define BP_GET_LSIZE(bp) \ 408 (BP_IS_EMBEDDED(bp) ? \ 409 (BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA ? BPE_GET_LSIZE(bp) : 0): \ 410 BF64_GET_SB((bp)->blk_prop, 0, SPA_LSIZEBITS, SPA_MINBLOCKSHIFT, 1)) 411 #define BP_SET_LSIZE(bp, x) do { \ 412 ASSERT(!BP_IS_EMBEDDED(bp)); \ 413 BF64_SET_SB((bp)->blk_prop, \ 414 0, SPA_LSIZEBITS, SPA_MINBLOCKSHIFT, 1, x); \ 415 } while (0) 416 417 #define BP_GET_PSIZE(bp) \ 418 (BP_IS_EMBEDDED(bp) ? 0 : \ 419 BF64_GET_SB((bp)->blk_prop, 16, SPA_PSIZEBITS, SPA_MINBLOCKSHIFT, 1)) 420 #define BP_SET_PSIZE(bp, x) do { \ 421 ASSERT(!BP_IS_EMBEDDED(bp)); \ 422 BF64_SET_SB((bp)->blk_prop, \ 423 16, SPA_PSIZEBITS, SPA_MINBLOCKSHIFT, 1, x); \ 424 } while (0) 425 426 #define BP_GET_COMPRESS(bp) \ 427 BF64_GET((bp)->blk_prop, 32, SPA_COMPRESSBITS) 428 #define BP_SET_COMPRESS(bp, x) \ 429 BF64_SET((bp)->blk_prop, 32, SPA_COMPRESSBITS, x) 430 431 #define BP_IS_EMBEDDED(bp) BF64_GET((bp)->blk_prop, 39, 1) 432 #define BP_SET_EMBEDDED(bp, x) BF64_SET((bp)->blk_prop, 39, 1, x) 433 434 #define BP_GET_CHECKSUM(bp) \ 435 (BP_IS_EMBEDDED(bp) ? ZIO_CHECKSUM_OFF : \ 436 BF64_GET((bp)->blk_prop, 40, 8)) 437 #define BP_SET_CHECKSUM(bp, x) do { \ 438 ASSERT(!BP_IS_EMBEDDED(bp)); \ 439 BF64_SET((bp)->blk_prop, 40, 8, x); \ 440 } while (0) 441 442 #define BP_GET_TYPE(bp) BF64_GET((bp)->blk_prop, 48, 8) 443 #define BP_SET_TYPE(bp, x) BF64_SET((bp)->blk_prop, 48, 8, x) 444 445 #define BP_GET_LEVEL(bp) BF64_GET((bp)->blk_prop, 56, 5) 446 #define BP_SET_LEVEL(bp, x) BF64_SET((bp)->blk_prop, 56, 5, x) 447 448 /* encrypted, authenticated, and MAC cksum bps use the same bit */ 449 #define BP_USES_CRYPT(bp) BF64_GET((bp)->blk_prop, 61, 1) 450 #define BP_SET_CRYPT(bp, x) BF64_SET((bp)->blk_prop, 61, 1, x) 451 452 #define BP_IS_ENCRYPTED(bp) \ 453 (BP_USES_CRYPT(bp) && \ 454 BP_GET_LEVEL(bp) <= 0 && \ 455 DMU_OT_IS_ENCRYPTED(BP_GET_TYPE(bp))) 456 457 #define BP_IS_AUTHENTICATED(bp) \ 458 (BP_USES_CRYPT(bp) && \ 459 BP_GET_LEVEL(bp) <= 0 && \ 460 !DMU_OT_IS_ENCRYPTED(BP_GET_TYPE(bp))) 461 462 #define BP_HAS_INDIRECT_MAC_CKSUM(bp) \ 463 (BP_USES_CRYPT(bp) && BP_GET_LEVEL(bp) > 0) 464 465 #define BP_IS_PROTECTED(bp) \ 466 (BP_IS_ENCRYPTED(bp) || BP_IS_AUTHENTICATED(bp)) 467 468 #define BP_GET_DEDUP(bp) BF64_GET((bp)->blk_prop, 62, 1) 469 #define BP_SET_DEDUP(bp, x) BF64_SET((bp)->blk_prop, 62, 1, x) 470 471 #define BP_GET_BYTEORDER(bp) BF64_GET((bp)->blk_prop, 63, 1) 472 #define BP_SET_BYTEORDER(bp, x) BF64_SET((bp)->blk_prop, 63, 1, x) 473 474 #define BP_GET_FREE(bp) BF64_GET((bp)->blk_fill, 0, 1) 475 #define BP_SET_FREE(bp, x) BF64_SET((bp)->blk_fill, 0, 1, x) 476 477 #define BP_GET_LOGICAL_BIRTH(bp) (bp)->blk_birth_word[1] 478 #define BP_SET_LOGICAL_BIRTH(bp, x) ((bp)->blk_birth_word[1] = (x)) 479 480 #define BP_GET_PHYSICAL_BIRTH(bp) (bp)->blk_birth_word[0] 481 #define BP_SET_PHYSICAL_BIRTH(bp, x) ((bp)->blk_birth_word[0] = (x)) 482 483 #define BP_GET_BIRTH(bp) \ 484 (BP_IS_EMBEDDED(bp) ? 0 : \ 485 BP_GET_PHYSICAL_BIRTH(bp) ? BP_GET_PHYSICAL_BIRTH(bp) : \ 486 BP_GET_LOGICAL_BIRTH(bp)) 487 488 #define BP_SET_BIRTH(bp, logical, physical) \ 489 { \ 490 ASSERT(!BP_IS_EMBEDDED(bp)); \ 491 BP_SET_LOGICAL_BIRTH(bp, logical); \ 492 BP_SET_PHYSICAL_BIRTH(bp, \ 493 ((logical) == (physical) ? 0 : (physical))); \ 494 } 495 496 #define BP_GET_FILL(bp) \ 497 ((BP_IS_ENCRYPTED(bp)) ? BF64_GET((bp)->blk_fill, 0, 32) : \ 498 ((BP_IS_EMBEDDED(bp)) ? 1 : (bp)->blk_fill)) 499 500 #define BP_SET_FILL(bp, fill) \ 501 { \ 502 if (BP_IS_ENCRYPTED(bp)) \ 503 BF64_SET((bp)->blk_fill, 0, 32, fill); \ 504 else \ 505 (bp)->blk_fill = fill; \ 506 } 507 508 #define BP_GET_IV2(bp) \ 509 (ASSERT(BP_IS_ENCRYPTED(bp)), \ 510 BF64_GET((bp)->blk_fill, 32, 32)) 511 #define BP_SET_IV2(bp, iv2) \ 512 { \ 513 ASSERT(BP_IS_ENCRYPTED(bp)); \ 514 BF64_SET((bp)->blk_fill, 32, 32, iv2); \ 515 } 516 517 #define BP_IS_METADATA(bp) \ 518 (BP_GET_LEVEL(bp) > 0 || DMU_OT_IS_METADATA(BP_GET_TYPE(bp))) 519 520 #define BP_GET_ASIZE(bp) \ 521 (BP_IS_EMBEDDED(bp) ? 0 : \ 522 DVA_GET_ASIZE(&(bp)->blk_dva[0]) + \ 523 DVA_GET_ASIZE(&(bp)->blk_dva[1]) + \ 524 (DVA_GET_ASIZE(&(bp)->blk_dva[2]) * !BP_IS_ENCRYPTED(bp))) 525 526 #define BP_GET_UCSIZE(bp) \ 527 (BP_IS_METADATA(bp) ? BP_GET_PSIZE(bp) : BP_GET_LSIZE(bp)) 528 529 #define BP_GET_NDVAS(bp) \ 530 (BP_IS_EMBEDDED(bp) ? 0 : \ 531 !!DVA_GET_ASIZE(&(bp)->blk_dva[0]) + \ 532 !!DVA_GET_ASIZE(&(bp)->blk_dva[1]) + \ 533 (!!DVA_GET_ASIZE(&(bp)->blk_dva[2]) * !BP_IS_ENCRYPTED(bp))) 534 535 #define BP_COUNT_GANG(bp) \ 536 (BP_IS_EMBEDDED(bp) ? 0 : \ 537 (DVA_GET_GANG(&(bp)->blk_dva[0]) + \ 538 DVA_GET_GANG(&(bp)->blk_dva[1]) + \ 539 (DVA_GET_GANG(&(bp)->blk_dva[2]) * !BP_IS_ENCRYPTED(bp)))) 540 541 #define DVA_EQUAL(dva1, dva2) \ 542 ((dva1)->dva_word[1] == (dva2)->dva_word[1] && \ 543 (dva1)->dva_word[0] == (dva2)->dva_word[0]) 544 545 #define BP_EQUAL(bp1, bp2) \ 546 (BP_GET_BIRTH(bp1) == BP_GET_BIRTH(bp2) && \ 547 BP_GET_LOGICAL_BIRTH(bp1) == BP_GET_LOGICAL_BIRTH(bp2) && \ 548 DVA_EQUAL(&(bp1)->blk_dva[0], &(bp2)->blk_dva[0]) && \ 549 DVA_EQUAL(&(bp1)->blk_dva[1], &(bp2)->blk_dva[1]) && \ 550 DVA_EQUAL(&(bp1)->blk_dva[2], &(bp2)->blk_dva[2])) 551 552 553 #define DVA_IS_VALID(dva) (DVA_GET_ASIZE(dva) != 0) 554 555 #define BP_IDENTITY(bp) (ASSERT(!BP_IS_EMBEDDED(bp)), &(bp)->blk_dva[0]) 556 #define BP_IS_GANG(bp) \ 557 (BP_IS_EMBEDDED(bp) ? B_FALSE : DVA_GET_GANG(BP_IDENTITY(bp))) 558 #define DVA_IS_EMPTY(dva) ((dva)->dva_word[0] == 0ULL && \ 559 (dva)->dva_word[1] == 0ULL) 560 #define BP_IS_HOLE(bp) \ 561 (!BP_IS_EMBEDDED(bp) && DVA_IS_EMPTY(BP_IDENTITY(bp))) 562 563 #define BP_SET_REDACTED(bp) \ 564 { \ 565 BP_SET_EMBEDDED(bp, B_TRUE); \ 566 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_REDACTED); \ 567 } 568 #define BP_IS_REDACTED(bp) \ 569 (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_REDACTED) 570 571 /* BP_IS_RAIDZ(bp) assumes no block compression */ 572 #define BP_IS_RAIDZ(bp) (DVA_GET_ASIZE(&(bp)->blk_dva[0]) > \ 573 BP_GET_PSIZE(bp)) 574 575 #define BP_ZERO(bp) \ 576 { \ 577 (bp)->blk_dva[0].dva_word[0] = 0; \ 578 (bp)->blk_dva[0].dva_word[1] = 0; \ 579 (bp)->blk_dva[1].dva_word[0] = 0; \ 580 (bp)->blk_dva[1].dva_word[1] = 0; \ 581 (bp)->blk_dva[2].dva_word[0] = 0; \ 582 (bp)->blk_dva[2].dva_word[1] = 0; \ 583 (bp)->blk_prop = 0; \ 584 (bp)->blk_pad[0] = 0; \ 585 (bp)->blk_pad[1] = 0; \ 586 (bp)->blk_birth_word[0] = 0; \ 587 (bp)->blk_birth_word[1] = 0; \ 588 (bp)->blk_fill = 0; \ 589 ZIO_SET_CHECKSUM(&(bp)->blk_cksum, 0, 0, 0, 0); \ 590 } 591 592 #ifdef _ZFS_BIG_ENDIAN 593 #define ZFS_HOST_BYTEORDER (0ULL) 594 #else 595 #define ZFS_HOST_BYTEORDER (1ULL) 596 #endif 597 598 #define BP_SHOULD_BYTESWAP(bp) (BP_GET_BYTEORDER(bp) != ZFS_HOST_BYTEORDER) 599 600 #define BP_SPRINTF_LEN 400 601 602 /* 603 * This macro allows code sharing between zfs, libzpool, and mdb. 604 * 'func' is either kmem_scnprintf() or mdb_snprintf(). 605 * 'ws' (whitespace) can be ' ' for single-line format, '\n' for multi-line. 606 */ 607 608 #define SNPRINTF_BLKPTR(func, ws, buf, size, bp, type, checksum, compress) \ 609 { \ 610 static const char *const copyname[] = \ 611 { "zero", "single", "double", "triple" }; \ 612 int len = 0; \ 613 int copies = 0; \ 614 const char *crypt_type; \ 615 if (bp != NULL) { \ 616 if (BP_IS_ENCRYPTED(bp)) { \ 617 crypt_type = "encrypted"; \ 618 /* LINTED E_SUSPICIOUS_COMPARISON */ \ 619 } else if (BP_IS_AUTHENTICATED(bp)) { \ 620 crypt_type = "authenticated"; \ 621 } else if (BP_HAS_INDIRECT_MAC_CKSUM(bp)) { \ 622 crypt_type = "indirect-MAC"; \ 623 } else { \ 624 crypt_type = "unencrypted"; \ 625 } \ 626 } \ 627 if (bp == NULL) { \ 628 len += func(buf + len, size - len, "<NULL>"); \ 629 } else if (BP_IS_HOLE(bp)) { \ 630 len += func(buf + len, size - len, \ 631 "HOLE [L%llu %s] " \ 632 "size=%llxL birth=%lluL", \ 633 (u_longlong_t)BP_GET_LEVEL(bp), \ 634 type, \ 635 (u_longlong_t)BP_GET_LSIZE(bp), \ 636 (u_longlong_t)BP_GET_LOGICAL_BIRTH(bp)); \ 637 } else if (BP_IS_EMBEDDED(bp)) { \ 638 len = func(buf + len, size - len, \ 639 "EMBEDDED [L%llu %s] et=%u %s " \ 640 "size=%llxL/%llxP birth=%lluL", \ 641 (u_longlong_t)BP_GET_LEVEL(bp), \ 642 type, \ 643 (int)BPE_GET_ETYPE(bp), \ 644 compress, \ 645 (u_longlong_t)BPE_GET_LSIZE(bp), \ 646 (u_longlong_t)BPE_GET_PSIZE(bp), \ 647 (u_longlong_t)BP_GET_LOGICAL_BIRTH(bp)); \ 648 } else if (BP_IS_REDACTED(bp)) { \ 649 len += func(buf + len, size - len, \ 650 "REDACTED [L%llu %s] size=%llxL birth=%lluL", \ 651 (u_longlong_t)BP_GET_LEVEL(bp), \ 652 type, \ 653 (u_longlong_t)BP_GET_LSIZE(bp), \ 654 (u_longlong_t)BP_GET_LOGICAL_BIRTH(bp)); \ 655 } else { \ 656 for (int d = 0; d < BP_GET_NDVAS(bp); d++) { \ 657 const dva_t *dva = &bp->blk_dva[d]; \ 658 if (DVA_IS_VALID(dva)) \ 659 copies++; \ 660 len += func(buf + len, size - len, \ 661 "DVA[%d]=<%llu:%llx:%llx>%c", d, \ 662 (u_longlong_t)DVA_GET_VDEV(dva), \ 663 (u_longlong_t)DVA_GET_OFFSET(dva), \ 664 (u_longlong_t)DVA_GET_ASIZE(dva), \ 665 ws); \ 666 } \ 667 ASSERT3S(copies, >, 0); \ 668 if (BP_IS_ENCRYPTED(bp)) { \ 669 len += func(buf + len, size - len, \ 670 "salt=%llx iv=%llx:%llx%c", \ 671 (u_longlong_t)bp->blk_dva[2].dva_word[0], \ 672 (u_longlong_t)bp->blk_dva[2].dva_word[1], \ 673 (u_longlong_t)BP_GET_IV2(bp), \ 674 ws); \ 675 } \ 676 if (BP_IS_GANG(bp) && \ 677 DVA_GET_ASIZE(&bp->blk_dva[2]) <= \ 678 DVA_GET_ASIZE(&bp->blk_dva[1]) / 2) \ 679 copies--; \ 680 len += func(buf + len, size - len, \ 681 "[L%llu %s] %s %s %s %s %s %s %s%c" \ 682 "size=%llxL/%llxP birth=%lluL/%lluP fill=%llu%c" \ 683 "cksum=%016llx:%016llx:%016llx:%016llx", \ 684 (u_longlong_t)BP_GET_LEVEL(bp), \ 685 type, \ 686 checksum, \ 687 compress, \ 688 crypt_type, \ 689 BP_GET_BYTEORDER(bp) == 0 ? "BE" : "LE", \ 690 BP_IS_GANG(bp) ? "gang" : "contiguous", \ 691 BP_GET_DEDUP(bp) ? "dedup" : "unique", \ 692 copyname[copies], \ 693 ws, \ 694 (u_longlong_t)BP_GET_LSIZE(bp), \ 695 (u_longlong_t)BP_GET_PSIZE(bp), \ 696 (u_longlong_t)BP_GET_LOGICAL_BIRTH(bp), \ 697 (u_longlong_t)BP_GET_BIRTH(bp), \ 698 (u_longlong_t)BP_GET_FILL(bp), \ 699 ws, \ 700 (u_longlong_t)bp->blk_cksum.zc_word[0], \ 701 (u_longlong_t)bp->blk_cksum.zc_word[1], \ 702 (u_longlong_t)bp->blk_cksum.zc_word[2], \ 703 (u_longlong_t)bp->blk_cksum.zc_word[3]); \ 704 } \ 705 ASSERT(len < size); \ 706 } 707 708 #define BP_GET_BUFC_TYPE(bp) \ 709 (BP_IS_METADATA(bp) ? ARC_BUFC_METADATA : ARC_BUFC_DATA) 710 711 typedef enum spa_import_type { 712 SPA_IMPORT_EXISTING, 713 SPA_IMPORT_ASSEMBLE 714 } spa_import_type_t; 715 716 typedef enum spa_mode { 717 SPA_MODE_UNINIT = 0, 718 SPA_MODE_READ = 1, 719 SPA_MODE_WRITE = 2, 720 } spa_mode_t; 721 722 /* 723 * Send TRIM commands in-line during normal pool operation while deleting. 724 * OFF: no 725 * ON: yes 726 */ 727 typedef enum { 728 SPA_AUTOTRIM_OFF = 0, /* default */ 729 SPA_AUTOTRIM_ON, 730 } spa_autotrim_t; 731 732 /* 733 * Reason TRIM command was issued, used internally for accounting purposes. 734 */ 735 typedef enum trim_type { 736 TRIM_TYPE_MANUAL = 0, 737 TRIM_TYPE_AUTO = 1, 738 TRIM_TYPE_SIMPLE = 2 739 } trim_type_t; 740 741 /* state manipulation functions */ 742 extern int spa_open(const char *pool, spa_t **, const void *tag); 743 extern int spa_open_rewind(const char *pool, spa_t **, const void *tag, 744 nvlist_t *policy, nvlist_t **config); 745 extern int spa_get_stats(const char *pool, nvlist_t **config, char *altroot, 746 size_t buflen); 747 extern int spa_create(const char *pool, nvlist_t *nvroot, nvlist_t *props, 748 nvlist_t *zplprops, struct dsl_crypto_params *dcp); 749 extern int spa_import(char *pool, nvlist_t *config, nvlist_t *props, 750 uint64_t flags); 751 extern nvlist_t *spa_tryimport(nvlist_t *tryconfig); 752 extern int spa_destroy(const char *pool); 753 extern int spa_checkpoint(const char *pool); 754 extern int spa_checkpoint_discard(const char *pool); 755 extern int spa_export(const char *pool, nvlist_t **oldconfig, boolean_t force, 756 boolean_t hardforce); 757 extern int spa_reset(const char *pool); 758 extern void spa_async_request(spa_t *spa, int flag); 759 extern void spa_async_unrequest(spa_t *spa, int flag); 760 extern void spa_async_suspend(spa_t *spa); 761 extern void spa_async_resume(spa_t *spa); 762 extern int spa_async_tasks(spa_t *spa); 763 extern spa_t *spa_inject_addref(char *pool); 764 extern void spa_inject_delref(spa_t *spa); 765 extern void spa_scan_stat_init(spa_t *spa); 766 extern int spa_scan_get_stats(spa_t *spa, pool_scan_stat_t *ps); 767 extern int bpobj_enqueue_alloc_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx); 768 extern int bpobj_enqueue_free_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx); 769 770 #define SPA_ASYNC_CONFIG_UPDATE 0x01 771 #define SPA_ASYNC_REMOVE 0x02 772 #define SPA_ASYNC_FAULT_VDEV 0x04 773 #define SPA_ASYNC_RESILVER_DONE 0x08 774 #define SPA_ASYNC_RESILVER 0x10 775 #define SPA_ASYNC_AUTOEXPAND 0x20 776 #define SPA_ASYNC_REMOVE_DONE 0x40 777 #define SPA_ASYNC_REMOVE_STOP 0x80 778 #define SPA_ASYNC_INITIALIZE_RESTART 0x100 779 #define SPA_ASYNC_TRIM_RESTART 0x200 780 #define SPA_ASYNC_AUTOTRIM_RESTART 0x400 781 #define SPA_ASYNC_L2CACHE_REBUILD 0x800 782 #define SPA_ASYNC_L2CACHE_TRIM 0x1000 783 #define SPA_ASYNC_REBUILD_DONE 0x2000 784 #define SPA_ASYNC_DETACH_SPARE 0x4000 785 786 /* device manipulation */ 787 extern int spa_vdev_add(spa_t *spa, nvlist_t *nvroot, boolean_t ashift_check); 788 extern int spa_vdev_attach(spa_t *spa, uint64_t guid, nvlist_t *nvroot, 789 int replacing, int rebuild); 790 extern int spa_vdev_detach(spa_t *spa, uint64_t guid, uint64_t pguid, 791 int replace_done); 792 extern int spa_vdev_alloc(spa_t *spa, uint64_t guid); 793 extern int spa_vdev_noalloc(spa_t *spa, uint64_t guid); 794 extern boolean_t spa_vdev_remove_active(spa_t *spa); 795 extern int spa_vdev_initialize(spa_t *spa, nvlist_t *nv, uint64_t cmd_type, 796 nvlist_t *vdev_errlist); 797 extern int spa_vdev_trim(spa_t *spa, nvlist_t *nv, uint64_t cmd_type, 798 uint64_t rate, boolean_t partial, boolean_t secure, nvlist_t *vdev_errlist); 799 extern int spa_vdev_setpath(spa_t *spa, uint64_t guid, const char *newpath); 800 extern int spa_vdev_setfru(spa_t *spa, uint64_t guid, const char *newfru); 801 extern int spa_vdev_split_mirror(spa_t *spa, const char *newname, 802 nvlist_t *config, nvlist_t *props, boolean_t exp); 803 804 /* spare state (which is global across all pools) */ 805 extern void spa_spare_add(vdev_t *vd); 806 extern void spa_spare_remove(vdev_t *vd); 807 extern boolean_t spa_spare_exists(uint64_t guid, uint64_t *pool, int *refcnt); 808 extern void spa_spare_activate(vdev_t *vd); 809 810 /* L2ARC state (which is global across all pools) */ 811 extern void spa_l2cache_add(vdev_t *vd); 812 extern void spa_l2cache_remove(vdev_t *vd); 813 extern boolean_t spa_l2cache_exists(uint64_t guid, uint64_t *pool); 814 extern void spa_l2cache_activate(vdev_t *vd); 815 extern void spa_l2cache_drop(spa_t *spa); 816 817 /* scanning */ 818 extern int spa_scan(spa_t *spa, pool_scan_func_t func); 819 extern int spa_scan_stop(spa_t *spa); 820 extern int spa_scrub_pause_resume(spa_t *spa, pool_scrub_cmd_t flag); 821 822 /* spa syncing */ 823 extern void spa_sync(spa_t *spa, uint64_t txg); /* only for DMU use */ 824 extern void spa_sync_allpools(void); 825 826 extern uint_t zfs_sync_pass_deferred_free; 827 828 /* spa sync taskqueues */ 829 taskq_t *spa_sync_tq_create(spa_t *spa, const char *name); 830 void spa_sync_tq_destroy(spa_t *spa); 831 uint_t spa_acq_allocator(spa_t *spa); 832 void spa_rel_allocator(spa_t *spa, uint_t allocator); 833 void spa_select_allocator(zio_t *zio); 834 835 /* spa namespace global mutex */ 836 extern kmutex_t spa_namespace_lock; 837 extern avl_tree_t spa_namespace_avl; 838 extern kcondvar_t spa_namespace_cv; 839 840 /* 841 * SPA configuration functions in spa_config.c 842 */ 843 844 #define SPA_CONFIG_UPDATE_POOL 0 845 #define SPA_CONFIG_UPDATE_VDEVS 1 846 847 extern void spa_write_cachefile(spa_t *, boolean_t, boolean_t, boolean_t); 848 extern void spa_config_load(void); 849 extern int spa_all_configs(uint64_t *generation, nvlist_t **pools); 850 extern void spa_config_set(spa_t *spa, nvlist_t *config); 851 extern nvlist_t *spa_config_generate(spa_t *spa, vdev_t *vd, uint64_t txg, 852 int getstats); 853 extern void spa_config_update(spa_t *spa, int what); 854 extern int spa_config_parse(spa_t *spa, vdev_t **vdp, nvlist_t *nv, 855 vdev_t *parent, uint_t id, int atype); 856 857 858 /* 859 * Miscellaneous SPA routines in spa_misc.c 860 */ 861 862 /* Namespace manipulation */ 863 extern spa_t *spa_lookup(const char *name); 864 extern spa_t *spa_add(const char *name, nvlist_t *config, const char *altroot); 865 extern void spa_remove(spa_t *spa); 866 extern spa_t *spa_next(spa_t *prev); 867 868 /* Refcount functions */ 869 extern void spa_open_ref(spa_t *spa, const void *tag); 870 extern void spa_close(spa_t *spa, const void *tag); 871 extern void spa_async_close(spa_t *spa, const void *tag); 872 extern boolean_t spa_refcount_zero(spa_t *spa); 873 874 #define SCL_NONE 0x00 875 #define SCL_CONFIG 0x01 876 #define SCL_STATE 0x02 877 #define SCL_L2ARC 0x04 /* hack until L2ARC 2.0 */ 878 #define SCL_ALLOC 0x08 879 #define SCL_ZIO 0x10 880 #define SCL_FREE 0x20 881 #define SCL_VDEV 0x40 882 #define SCL_LOCKS 7 883 #define SCL_ALL ((1 << SCL_LOCKS) - 1) 884 #define SCL_STATE_ALL (SCL_STATE | SCL_L2ARC | SCL_ZIO) 885 886 /* Historical pool statistics */ 887 typedef struct spa_history_kstat { 888 kmutex_t lock; 889 uint64_t count; 890 uint64_t size; 891 kstat_t *kstat; 892 void *priv; 893 list_t list; 894 } spa_history_kstat_t; 895 896 typedef struct spa_history_list { 897 uint64_t size; 898 procfs_list_t procfs_list; 899 } spa_history_list_t; 900 901 typedef struct spa_stats { 902 spa_history_list_t read_history; 903 spa_history_list_t txg_history; 904 spa_history_kstat_t tx_assign_histogram; 905 spa_history_list_t mmp_history; 906 spa_history_kstat_t state; /* pool state */ 907 spa_history_kstat_t guid; /* pool guid */ 908 spa_history_kstat_t iostats; 909 } spa_stats_t; 910 911 typedef enum txg_state { 912 TXG_STATE_BIRTH = 0, 913 TXG_STATE_OPEN = 1, 914 TXG_STATE_QUIESCED = 2, 915 TXG_STATE_WAIT_FOR_SYNC = 3, 916 TXG_STATE_SYNCED = 4, 917 TXG_STATE_COMMITTED = 5, 918 } txg_state_t; 919 920 typedef struct txg_stat { 921 vdev_stat_t vs1; 922 vdev_stat_t vs2; 923 uint64_t txg; 924 uint64_t ndirty; 925 } txg_stat_t; 926 927 /* Assorted pool IO kstats */ 928 typedef struct spa_iostats { 929 kstat_named_t trim_extents_written; 930 kstat_named_t trim_bytes_written; 931 kstat_named_t trim_extents_skipped; 932 kstat_named_t trim_bytes_skipped; 933 kstat_named_t trim_extents_failed; 934 kstat_named_t trim_bytes_failed; 935 kstat_named_t autotrim_extents_written; 936 kstat_named_t autotrim_bytes_written; 937 kstat_named_t autotrim_extents_skipped; 938 kstat_named_t autotrim_bytes_skipped; 939 kstat_named_t autotrim_extents_failed; 940 kstat_named_t autotrim_bytes_failed; 941 kstat_named_t simple_trim_extents_written; 942 kstat_named_t simple_trim_bytes_written; 943 kstat_named_t simple_trim_extents_skipped; 944 kstat_named_t simple_trim_bytes_skipped; 945 kstat_named_t simple_trim_extents_failed; 946 kstat_named_t simple_trim_bytes_failed; 947 } spa_iostats_t; 948 949 extern void spa_stats_init(spa_t *spa); 950 extern void spa_stats_destroy(spa_t *spa); 951 extern void spa_read_history_add(spa_t *spa, const zbookmark_phys_t *zb, 952 uint32_t aflags); 953 extern void spa_txg_history_add(spa_t *spa, uint64_t txg, hrtime_t birth_time); 954 extern int spa_txg_history_set(spa_t *spa, uint64_t txg, 955 txg_state_t completed_state, hrtime_t completed_time); 956 extern txg_stat_t *spa_txg_history_init_io(spa_t *, uint64_t, 957 struct dsl_pool *); 958 extern void spa_txg_history_fini_io(spa_t *, txg_stat_t *); 959 extern void spa_tx_assign_add_nsecs(spa_t *spa, uint64_t nsecs); 960 extern int spa_mmp_history_set_skip(spa_t *spa, uint64_t mmp_kstat_id); 961 extern int spa_mmp_history_set(spa_t *spa, uint64_t mmp_kstat_id, int io_error, 962 hrtime_t duration); 963 extern void spa_mmp_history_add(spa_t *spa, uint64_t txg, uint64_t timestamp, 964 uint64_t mmp_delay, vdev_t *vd, int label, uint64_t mmp_kstat_id, 965 int error); 966 extern void spa_iostats_trim_add(spa_t *spa, trim_type_t type, 967 uint64_t extents_written, uint64_t bytes_written, 968 uint64_t extents_skipped, uint64_t bytes_skipped, 969 uint64_t extents_failed, uint64_t bytes_failed); 970 extern void spa_import_progress_add(spa_t *spa); 971 extern void spa_import_progress_remove(uint64_t spa_guid); 972 extern int spa_import_progress_set_mmp_check(uint64_t pool_guid, 973 uint64_t mmp_sec_remaining); 974 extern int spa_import_progress_set_max_txg(uint64_t pool_guid, 975 uint64_t max_txg); 976 extern int spa_import_progress_set_state(uint64_t pool_guid, 977 spa_load_state_t spa_load_state); 978 extern void spa_import_progress_set_notes(spa_t *spa, 979 const char *fmt, ...) __printflike(2, 3); 980 extern void spa_import_progress_set_notes_nolog(spa_t *spa, 981 const char *fmt, ...) __printflike(2, 3); 982 983 /* Pool configuration locks */ 984 extern int spa_config_tryenter(spa_t *spa, int locks, const void *tag, 985 krw_t rw); 986 extern void spa_config_enter(spa_t *spa, int locks, const void *tag, krw_t rw); 987 extern void spa_config_enter_mmp(spa_t *spa, int locks, const void *tag, 988 krw_t rw); 989 extern void spa_config_exit(spa_t *spa, int locks, const void *tag); 990 extern int spa_config_held(spa_t *spa, int locks, krw_t rw); 991 992 /* Pool vdev add/remove lock */ 993 extern uint64_t spa_vdev_enter(spa_t *spa); 994 extern uint64_t spa_vdev_detach_enter(spa_t *spa, uint64_t guid); 995 extern uint64_t spa_vdev_config_enter(spa_t *spa); 996 extern void spa_vdev_config_exit(spa_t *spa, vdev_t *vd, uint64_t txg, 997 int error, const char *tag); 998 extern int spa_vdev_exit(spa_t *spa, vdev_t *vd, uint64_t txg, int error); 999 1000 /* Pool vdev state change lock */ 1001 extern void spa_vdev_state_enter(spa_t *spa, int oplock); 1002 extern int spa_vdev_state_exit(spa_t *spa, vdev_t *vd, int error); 1003 1004 /* Log state */ 1005 typedef enum spa_log_state { 1006 SPA_LOG_UNKNOWN = 0, /* unknown log state */ 1007 SPA_LOG_MISSING, /* missing log(s) */ 1008 SPA_LOG_CLEAR, /* clear the log(s) */ 1009 SPA_LOG_GOOD, /* log(s) are good */ 1010 } spa_log_state_t; 1011 1012 extern spa_log_state_t spa_get_log_state(spa_t *spa); 1013 extern void spa_set_log_state(spa_t *spa, spa_log_state_t state); 1014 extern int spa_reset_logs(spa_t *spa); 1015 1016 /* Log claim callback */ 1017 extern void spa_claim_notify(zio_t *zio); 1018 extern void spa_deadman(void *); 1019 1020 /* Accessor functions */ 1021 extern boolean_t spa_shutting_down(spa_t *spa); 1022 extern struct dsl_pool *spa_get_dsl(spa_t *spa); 1023 extern boolean_t spa_is_initializing(spa_t *spa); 1024 extern boolean_t spa_indirect_vdevs_loaded(spa_t *spa); 1025 extern blkptr_t *spa_get_rootblkptr(spa_t *spa); 1026 extern void spa_set_rootblkptr(spa_t *spa, const blkptr_t *bp); 1027 extern void spa_altroot(spa_t *, char *, size_t); 1028 extern uint32_t spa_sync_pass(spa_t *spa); 1029 extern char *spa_name(spa_t *spa); 1030 extern uint64_t spa_guid(spa_t *spa); 1031 extern uint64_t spa_load_guid(spa_t *spa); 1032 extern uint64_t spa_last_synced_txg(spa_t *spa); 1033 extern uint64_t spa_first_txg(spa_t *spa); 1034 extern uint64_t spa_syncing_txg(spa_t *spa); 1035 extern uint64_t spa_final_dirty_txg(spa_t *spa); 1036 extern uint64_t spa_version(spa_t *spa); 1037 extern pool_state_t spa_state(spa_t *spa); 1038 extern spa_load_state_t spa_load_state(spa_t *spa); 1039 extern uint64_t spa_freeze_txg(spa_t *spa); 1040 extern uint64_t spa_get_worst_case_asize(spa_t *spa, uint64_t lsize); 1041 extern uint64_t spa_get_dspace(spa_t *spa); 1042 extern uint64_t spa_get_checkpoint_space(spa_t *spa); 1043 extern uint64_t spa_get_slop_space(spa_t *spa); 1044 extern void spa_update_dspace(spa_t *spa); 1045 extern uint64_t spa_version(spa_t *spa); 1046 extern boolean_t spa_deflate(spa_t *spa); 1047 extern metaslab_class_t *spa_normal_class(spa_t *spa); 1048 extern metaslab_class_t *spa_log_class(spa_t *spa); 1049 extern metaslab_class_t *spa_embedded_log_class(spa_t *spa); 1050 extern metaslab_class_t *spa_special_class(spa_t *spa); 1051 extern metaslab_class_t *spa_dedup_class(spa_t *spa); 1052 extern metaslab_class_t *spa_preferred_class(spa_t *spa, uint64_t size, 1053 dmu_object_type_t objtype, uint_t level, uint_t special_smallblk); 1054 1055 extern void spa_evicting_os_register(spa_t *, objset_t *os); 1056 extern void spa_evicting_os_deregister(spa_t *, objset_t *os); 1057 extern void spa_evicting_os_wait(spa_t *spa); 1058 extern int spa_max_replication(spa_t *spa); 1059 extern int spa_prev_software_version(spa_t *spa); 1060 extern uint64_t spa_get_failmode(spa_t *spa); 1061 extern uint64_t spa_get_deadman_failmode(spa_t *spa); 1062 extern void spa_set_deadman_failmode(spa_t *spa, const char *failmode); 1063 extern boolean_t spa_suspended(spa_t *spa); 1064 extern uint64_t spa_bootfs(spa_t *spa); 1065 extern uint64_t spa_delegation(spa_t *spa); 1066 extern objset_t *spa_meta_objset(spa_t *spa); 1067 extern space_map_t *spa_syncing_log_sm(spa_t *spa); 1068 extern uint64_t spa_deadman_synctime(spa_t *spa); 1069 extern uint64_t spa_deadman_ziotime(spa_t *spa); 1070 extern uint64_t spa_dirty_data(spa_t *spa); 1071 extern spa_autotrim_t spa_get_autotrim(spa_t *spa); 1072 extern int spa_get_allocator(spa_t *spa); 1073 extern void spa_set_allocator(spa_t *spa, const char *allocator); 1074 1075 /* Miscellaneous support routines */ 1076 extern void spa_load_failed(spa_t *spa, const char *fmt, ...) 1077 __attribute__((format(printf, 2, 3))); 1078 extern void spa_load_note(spa_t *spa, const char *fmt, ...) 1079 __attribute__((format(printf, 2, 3))); 1080 extern void spa_activate_mos_feature(spa_t *spa, const char *feature, 1081 dmu_tx_t *tx); 1082 extern void spa_deactivate_mos_feature(spa_t *spa, const char *feature); 1083 extern spa_t *spa_by_guid(uint64_t pool_guid, uint64_t device_guid); 1084 extern boolean_t spa_guid_exists(uint64_t pool_guid, uint64_t device_guid); 1085 extern char *spa_strdup(const char *); 1086 extern void spa_strfree(char *); 1087 extern uint64_t spa_generate_guid(spa_t *spa); 1088 extern void snprintf_blkptr(char *buf, size_t buflen, const blkptr_t *bp); 1089 extern void spa_freeze(spa_t *spa); 1090 extern int spa_change_guid(spa_t *spa); 1091 extern void spa_upgrade(spa_t *spa, uint64_t version); 1092 extern void spa_evict_all(void); 1093 extern vdev_t *spa_lookup_by_guid(spa_t *spa, uint64_t guid, 1094 boolean_t l2cache); 1095 extern boolean_t spa_has_l2cache(spa_t *, uint64_t guid); 1096 extern boolean_t spa_has_spare(spa_t *, uint64_t guid); 1097 extern uint64_t dva_get_dsize_sync(spa_t *spa, const dva_t *dva); 1098 extern uint64_t bp_get_dsize_sync(spa_t *spa, const blkptr_t *bp); 1099 extern uint64_t bp_get_dsize(spa_t *spa, const blkptr_t *bp); 1100 extern boolean_t spa_has_slogs(spa_t *spa); 1101 extern boolean_t spa_is_root(spa_t *spa); 1102 extern boolean_t spa_writeable(spa_t *spa); 1103 extern boolean_t spa_has_pending_synctask(spa_t *spa); 1104 extern int spa_maxblocksize(spa_t *spa); 1105 extern int spa_maxdnodesize(spa_t *spa); 1106 extern boolean_t spa_has_checkpoint(spa_t *spa); 1107 extern boolean_t spa_importing_readonly_checkpoint(spa_t *spa); 1108 extern boolean_t spa_suspend_async_destroy(spa_t *spa); 1109 extern uint64_t spa_min_claim_txg(spa_t *spa); 1110 extern boolean_t zfs_dva_valid(spa_t *spa, const dva_t *dva, 1111 const blkptr_t *bp); 1112 typedef void (*spa_remap_cb_t)(uint64_t vdev, uint64_t offset, uint64_t size, 1113 void *arg); 1114 extern boolean_t spa_remap_blkptr(spa_t *spa, blkptr_t *bp, 1115 spa_remap_cb_t callback, void *arg); 1116 extern uint64_t spa_get_last_removal_txg(spa_t *spa); 1117 extern boolean_t spa_trust_config(spa_t *spa); 1118 extern uint64_t spa_missing_tvds_allowed(spa_t *spa); 1119 extern void spa_set_missing_tvds(spa_t *spa, uint64_t missing); 1120 extern boolean_t spa_top_vdevs_spacemap_addressable(spa_t *spa); 1121 extern uint64_t spa_total_metaslabs(spa_t *spa); 1122 extern boolean_t spa_multihost(spa_t *spa); 1123 extern uint32_t spa_get_hostid(spa_t *spa); 1124 extern void spa_activate_allocation_classes(spa_t *, dmu_tx_t *); 1125 extern boolean_t spa_livelist_delete_check(spa_t *spa); 1126 1127 extern boolean_t spa_mmp_remote_host_activity(spa_t *spa); 1128 1129 extern spa_mode_t spa_mode(spa_t *spa); 1130 extern uint64_t zfs_strtonum(const char *str, char **nptr); 1131 1132 extern char *spa_his_ievent_table[]; 1133 1134 extern void spa_history_create_obj(spa_t *spa, dmu_tx_t *tx); 1135 extern int spa_history_get(spa_t *spa, uint64_t *offset, uint64_t *len_read, 1136 char *his_buf); 1137 extern int spa_history_log(spa_t *spa, const char *his_buf); 1138 extern int spa_history_log_nvl(spa_t *spa, nvlist_t *nvl); 1139 extern void spa_history_log_version(spa_t *spa, const char *operation, 1140 dmu_tx_t *tx); 1141 extern void spa_history_log_internal(spa_t *spa, const char *operation, 1142 dmu_tx_t *tx, const char *fmt, ...) __printflike(4, 5); 1143 extern void spa_history_log_internal_ds(struct dsl_dataset *ds, const char *op, 1144 dmu_tx_t *tx, const char *fmt, ...) __printflike(4, 5); 1145 extern void spa_history_log_internal_dd(dsl_dir_t *dd, const char *operation, 1146 dmu_tx_t *tx, const char *fmt, ...) __printflike(4, 5); 1147 1148 extern const char *spa_state_to_name(spa_t *spa); 1149 1150 /* error handling */ 1151 struct zbookmark_phys; 1152 extern void spa_log_error(spa_t *spa, const zbookmark_phys_t *zb, 1153 const uint64_t birth); 1154 extern void spa_remove_error(spa_t *spa, zbookmark_phys_t *zb, 1155 uint64_t birth); 1156 extern int zfs_ereport_post(const char *clazz, spa_t *spa, vdev_t *vd, 1157 const zbookmark_phys_t *zb, zio_t *zio, uint64_t state); 1158 extern boolean_t zfs_ereport_is_valid(const char *clazz, spa_t *spa, vdev_t *vd, 1159 zio_t *zio); 1160 extern void zfs_ereport_taskq_fini(void); 1161 extern void zfs_ereport_clear(spa_t *spa, vdev_t *vd); 1162 extern nvlist_t *zfs_event_create(spa_t *spa, vdev_t *vd, const char *type, 1163 const char *name, nvlist_t *aux); 1164 extern void zfs_post_remove(spa_t *spa, vdev_t *vd); 1165 extern void zfs_post_state_change(spa_t *spa, vdev_t *vd, uint64_t laststate); 1166 extern void zfs_post_autoreplace(spa_t *spa, vdev_t *vd); 1167 extern uint64_t spa_approx_errlog_size(spa_t *spa); 1168 extern int spa_get_errlog(spa_t *spa, void *uaddr, uint64_t *count); 1169 extern uint64_t spa_get_last_errlog_size(spa_t *spa); 1170 extern void spa_errlog_rotate(spa_t *spa); 1171 extern void spa_errlog_drain(spa_t *spa); 1172 extern void spa_errlog_sync(spa_t *spa, uint64_t txg); 1173 extern void spa_get_errlists(spa_t *spa, avl_tree_t *last, avl_tree_t *scrub); 1174 extern void spa_delete_dataset_errlog(spa_t *spa, uint64_t ds, dmu_tx_t *tx); 1175 extern void spa_swap_errlog(spa_t *spa, uint64_t new_head_ds, 1176 uint64_t old_head_ds, dmu_tx_t *tx); 1177 extern void sync_error_list(spa_t *spa, avl_tree_t *t, uint64_t *obj, 1178 dmu_tx_t *tx); 1179 extern void spa_upgrade_errlog(spa_t *spa, dmu_tx_t *tx); 1180 extern int find_top_affected_fs(spa_t *spa, uint64_t head_ds, 1181 zbookmark_err_phys_t *zep, uint64_t *top_affected_fs); 1182 extern int find_birth_txg(struct dsl_dataset *ds, zbookmark_err_phys_t *zep, 1183 uint64_t *birth_txg); 1184 extern void zep_to_zb(uint64_t dataset, zbookmark_err_phys_t *zep, 1185 zbookmark_phys_t *zb); 1186 extern void name_to_errphys(char *buf, zbookmark_err_phys_t *zep); 1187 1188 /* vdev mirror */ 1189 extern void vdev_mirror_stat_init(void); 1190 extern void vdev_mirror_stat_fini(void); 1191 1192 /* Initialization and termination */ 1193 extern void spa_init(spa_mode_t mode); 1194 extern void spa_fini(void); 1195 extern void spa_boot_init(void); 1196 1197 /* properties */ 1198 extern int spa_prop_set(spa_t *spa, nvlist_t *nvp); 1199 extern int spa_prop_get(spa_t *spa, nvlist_t **nvp); 1200 extern void spa_prop_clear_bootfs(spa_t *spa, uint64_t obj, dmu_tx_t *tx); 1201 extern void spa_configfile_set(spa_t *, nvlist_t *, boolean_t); 1202 1203 /* asynchronous event notification */ 1204 extern void spa_event_notify(spa_t *spa, vdev_t *vdev, nvlist_t *hist_nvl, 1205 const char *name); 1206 extern void zfs_ereport_zvol_post(const char *subclass, const char *name, 1207 const char *device_name, const char *raw_name); 1208 1209 /* waiting for pool activities to complete */ 1210 extern int spa_wait(const char *pool, zpool_wait_activity_t activity, 1211 boolean_t *waited); 1212 extern int spa_wait_tag(const char *name, zpool_wait_activity_t activity, 1213 uint64_t tag, boolean_t *waited); 1214 extern void spa_notify_waiters(spa_t *spa); 1215 extern void spa_wake_waiters(spa_t *spa); 1216 1217 extern void spa_import_os(spa_t *spa); 1218 extern void spa_export_os(spa_t *spa); 1219 extern void spa_activate_os(spa_t *spa); 1220 extern void spa_deactivate_os(spa_t *spa); 1221 1222 /* module param call functions */ 1223 int param_set_deadman_ziotime(ZFS_MODULE_PARAM_ARGS); 1224 int param_set_deadman_synctime(ZFS_MODULE_PARAM_ARGS); 1225 int param_set_slop_shift(ZFS_MODULE_PARAM_ARGS); 1226 int param_set_deadman_failmode(ZFS_MODULE_PARAM_ARGS); 1227 int param_set_active_allocator(ZFS_MODULE_PARAM_ARGS); 1228 1229 #ifdef ZFS_DEBUG 1230 #define dprintf_bp(bp, fmt, ...) do { \ 1231 if (zfs_flags & ZFS_DEBUG_DPRINTF) { \ 1232 char *__blkbuf = kmem_alloc(BP_SPRINTF_LEN, KM_SLEEP); \ 1233 snprintf_blkptr(__blkbuf, BP_SPRINTF_LEN, (bp)); \ 1234 dprintf(fmt " %s\n", __VA_ARGS__, __blkbuf); \ 1235 kmem_free(__blkbuf, BP_SPRINTF_LEN); \ 1236 } \ 1237 } while (0) 1238 #else 1239 #define dprintf_bp(bp, fmt, ...) 1240 #endif 1241 1242 extern spa_mode_t spa_mode_global; 1243 extern int zfs_deadman_enabled; 1244 extern uint64_t zfs_deadman_synctime_ms; 1245 extern uint64_t zfs_deadman_ziotime_ms; 1246 extern uint64_t zfs_deadman_checktime_ms; 1247 1248 extern kmem_cache_t *zio_buf_cache[]; 1249 extern kmem_cache_t *zio_data_buf_cache[]; 1250 1251 #ifdef __cplusplus 1252 } 1253 #endif 1254 1255 #endif /* _SYS_SPA_H */ 1256