1 // SPDX-License-Identifier: CDDL-1.0 2 /* 3 * CDDL HEADER START 4 * 5 * The contents of this file are subject to the terms of the 6 * Common Development and Distribution License (the "License"). 7 * You may not use this file except in compliance with the License. 8 * 9 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 10 * or https://opensource.org/licenses/CDDL-1.0. 11 * See the License for the specific language governing permissions 12 * and limitations under the License. 13 * 14 * When distributing Covered Code, include this CDDL HEADER in each 15 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 16 * If applicable, add the following below this CDDL HEADER, with the 17 * fields enclosed by brackets "[]" replaced with your own identifying 18 * information: Portions Copyright [yyyy] [name of copyright owner] 19 * 20 * CDDL HEADER END 21 */ 22 /* 23 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 24 * Copyright (c) 2011, 2024 by Delphix. All rights reserved. 25 * Copyright 2011 Nexenta Systems, Inc. All rights reserved. 26 * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved. 27 * Copyright 2013 Saso Kiselkov. All rights reserved. 28 * Copyright (c) 2014 Integros [integros.com] 29 * Copyright 2017 Joyent, Inc. 30 * Copyright (c) 2017, Intel Corporation. 31 * Copyright (c) 2019, Allan Jude 32 * Copyright (c) 2019, 2025, Klara, Inc. 33 * Copyright (c) 2019, Datto Inc. 34 */ 35 36 #ifndef _SYS_SPA_H 37 #define _SYS_SPA_H 38 39 #include <sys/zfs_context.h> 40 #include <sys/avl.h> 41 #include <sys/kstat.h> 42 #include <sys/nvpair.h> 43 #include <sys/types.h> 44 #include <sys/fs/zfs.h> 45 #include <sys/spa_checksum.h> 46 #include <sys/dmu.h> 47 #include <sys/space_map.h> 48 #include <sys/bitops.h> 49 50 #ifdef __cplusplus 51 extern "C" { 52 #endif 53 54 /* 55 * Forward references that lots of things need. 56 */ 57 typedef struct brt_vdev brt_vdev_t; 58 typedef struct spa spa_t; 59 typedef struct vdev vdev_t; 60 typedef struct metaslab metaslab_t; 61 typedef struct metaslab_group metaslab_group_t; 62 typedef struct metaslab_class metaslab_class_t; 63 typedef struct zio zio_t; 64 typedef struct zilog zilog_t; 65 typedef struct spa_aux_vdev spa_aux_vdev_t; 66 typedef struct zbookmark_phys zbookmark_phys_t; 67 typedef struct zbookmark_err_phys zbookmark_err_phys_t; 68 69 struct bpobj; 70 struct bplist; 71 struct dsl_pool; 72 struct dsl_dataset; 73 struct dsl_crypto_params; 74 75 /* 76 * Alignment Shift (ashift) is an immutable, internal top-level vdev property 77 * which can only be set at vdev creation time. Physical writes are always done 78 * according to it, which makes 2^ashift the smallest possible IO on a vdev. 79 * 80 * We currently allow values ranging from 512 bytes (2^9 = 512) to 64 KiB 81 * (2^16 = 65,536). 82 */ 83 #define ASHIFT_MIN 9 84 #define ASHIFT_MAX 16 85 86 /* 87 * Size of block to hold the configuration data (a packed nvlist) 88 */ 89 #define SPA_CONFIG_BLOCKSIZE (1ULL << 14) 90 91 /* 92 * The DVA size encodings for LSIZE and PSIZE support blocks up to 32MB. 93 * The ASIZE encoding should be at least 64 times larger (6 more bits) 94 * to support up to 4-way RAID-Z mirror mode with worst-case gang block 95 * overhead, three DVAs per bp, plus one more bit in case we do anything 96 * else that expands the ASIZE. 97 */ 98 #define SPA_LSIZEBITS 16 /* LSIZE up to 32M (2^16 * 512) */ 99 #define SPA_PSIZEBITS 16 /* PSIZE up to 32M (2^16 * 512) */ 100 #define SPA_ASIZEBITS 24 /* ASIZE up to 64 times larger */ 101 102 #define SPA_COMPRESSBITS 7 103 #define SPA_VDEVBITS 24 104 #define SPA_COMPRESSMASK ((1U << SPA_COMPRESSBITS) - 1) 105 106 /* 107 * All SPA data is represented by 128-bit data virtual addresses (DVAs). 108 * The members of the dva_t should be considered opaque outside the SPA. 109 */ 110 typedef struct dva { 111 uint64_t dva_word[2]; 112 } dva_t; 113 114 115 /* 116 * Some checksums/hashes need a 256-bit initialization salt. This salt is kept 117 * secret and is suitable for use in MAC algorithms as the key. 118 */ 119 typedef struct zio_cksum_salt { 120 uint8_t zcs_bytes[32]; 121 } zio_cksum_salt_t; 122 123 /* 124 * Each block is described by its DVAs, time of birth, checksum, etc. 125 * The word-by-word, bit-by-bit layout of the blkptr is as follows: 126 * 127 * 64 56 48 40 32 24 16 8 0 128 * +-------+-------+-------+-------+-------+-------+-------+-------+ 129 * 0 | pad | vdev1 | pad | ASIZE | 130 * +-------+-------+-------+-------+-------+-------+-------+-------+ 131 * 1 |G| offset1 | 132 * +-------+-------+-------+-------+-------+-------+-------+-------+ 133 * 2 | pad | vdev2 | pad | ASIZE | 134 * +-------+-------+-------+-------+-------+-------+-------+-------+ 135 * 3 |G| offset2 | 136 * +-------+-------+-------+-------+-------+-------+-------+-------+ 137 * 4 | pad | vdev3 | pad | ASIZE | 138 * +-------+-------+-------+-------+-------+-------+-------+-------+ 139 * 5 |G| offset3 | 140 * +-------+-------+-------+-------+-------+-------+-------+-------+ 141 * 6 |BDX|lvl| type | cksum |E| comp| PSIZE | LSIZE | 142 * +-------+-------+-------+-------+-------+-------+-------+-------+ 143 * 7 |R| padding | 144 * +-------+-------+-------+-------+-------+-------+-------+-------+ 145 * 8 | padding | 146 * +-------+-------+-------+-------+-------+-------+-------+-------+ 147 * 9 | physical birth txg | 148 * +-------+-------+-------+-------+-------+-------+-------+-------+ 149 * a | logical birth txg | 150 * +-------+-------+-------+-------+-------+-------+-------+-------+ 151 * b | fill count | 152 * +-------+-------+-------+-------+-------+-------+-------+-------+ 153 * c | checksum[0] | 154 * +-------+-------+-------+-------+-------+-------+-------+-------+ 155 * d | checksum[1] | 156 * +-------+-------+-------+-------+-------+-------+-------+-------+ 157 * e | checksum[2] | 158 * +-------+-------+-------+-------+-------+-------+-------+-------+ 159 * f | checksum[3] | 160 * +-------+-------+-------+-------+-------+-------+-------+-------+ 161 * 162 * Legend: 163 * 164 * vdev virtual device ID 165 * offset offset into virtual device 166 * LSIZE logical size 167 * PSIZE physical size (after compression) 168 * ASIZE allocated size (including RAID-Z parity and gang block headers) 169 * cksum checksum function 170 * comp compression function 171 * G gang block indicator 172 * B byteorder (endianness) 173 * D dedup 174 * X encryption 175 * E blkptr_t contains embedded data (see below) 176 * lvl level of indirection 177 * type DMU object type 178 * R rewrite (reallocated/rewritten at phys birth TXG) 179 * phys birth txg when dva[0] was written; zero if same as logical birth txg 180 * note that typically all the dva's would be written in this 181 * txg, but they could be different if they were moved by 182 * device removal. 183 * log. birth transaction group in which the block was logically born 184 * fill count number of non-zero blocks under this bp 185 * checksum[4] 256-bit checksum of the data this bp describes 186 */ 187 188 /* 189 * The blkptr_t's of encrypted blocks also need to store the encryption 190 * parameters so that the block can be decrypted. This layout is as follows: 191 * 192 * 64 56 48 40 32 24 16 8 0 193 * +-------+-------+-------+-------+-------+-------+-------+-------+ 194 * 0 | pad | vdev1 | pad | ASIZE | 195 * +-------+-------+-------+-------+-------+-------+-------+-------+ 196 * 1 |G| offset1 | 197 * +-------+-------+-------+-------+-------+-------+-------+-------+ 198 * 2 | pad | vdev2 | pad | ASIZE | 199 * +-------+-------+-------+-------+-------+-------+-------+-------+ 200 * 3 |G| offset2 | 201 * +-------+-------+-------+-------+-------+-------+-------+-------+ 202 * 4 | salt | 203 * +-------+-------+-------+-------+-------+-------+-------+-------+ 204 * 5 | IV1 | 205 * +-------+-------+-------+-------+-------+-------+-------+-------+ 206 * 6 |BDX|lvl| type | cksum |E| comp| PSIZE | LSIZE | 207 * +-------+-------+-------+-------+-------+-------+-------+-------+ 208 * 7 |R| padding | 209 * +-------+-------+-------+-------+-------+-------+-------+-------+ 210 * 8 | padding | 211 * +-------+-------+-------+-------+-------+-------+-------+-------+ 212 * 9 | physical birth txg | 213 * +-------+-------+-------+-------+-------+-------+-------+-------+ 214 * a | logical birth txg | 215 * +-------+-------+-------+-------+-------+-------+-------+-------+ 216 * b | IV2 | fill count | 217 * +-------+-------+-------+-------+-------+-------+-------+-------+ 218 * c | checksum[0] | 219 * +-------+-------+-------+-------+-------+-------+-------+-------+ 220 * d | checksum[1] | 221 * +-------+-------+-------+-------+-------+-------+-------+-------+ 222 * e | MAC[0] | 223 * +-------+-------+-------+-------+-------+-------+-------+-------+ 224 * f | MAC[1] | 225 * +-------+-------+-------+-------+-------+-------+-------+-------+ 226 * 227 * Legend: 228 * 229 * salt Salt for generating encryption keys 230 * IV1 First 64 bits of encryption IV 231 * X Block requires encryption handling (set to 1) 232 * E blkptr_t contains embedded data (set to 0, see below) 233 * fill count number of non-zero blocks under this bp (truncated to 32 bits) 234 * IV2 Last 32 bits of encryption IV 235 * checksum[2] 128-bit checksum of the data this bp describes 236 * MAC[2] 128-bit message authentication code for this data 237 * 238 * The X bit being set indicates that this block is one of 3 types. If this is 239 * a level 0 block with an encrypted object type, the block is encrypted 240 * (see BP_IS_ENCRYPTED()). If this is a level 0 block with an unencrypted 241 * object type, this block is authenticated with an HMAC (see 242 * BP_IS_AUTHENTICATED()). Otherwise (if level > 0), this bp will use the MAC 243 * words to store a checksum-of-MACs from the level below (see 244 * BP_HAS_INDIRECT_MAC_CKSUM()). For convenience in the code, BP_IS_PROTECTED() 245 * refers to both encrypted and authenticated blocks and BP_USES_CRYPT() 246 * refers to any of these 3 kinds of blocks. 247 * 248 * The additional encryption parameters are the salt, IV, and MAC which are 249 * explained in greater detail in the block comment at the top of zio_crypt.c. 250 * The MAC occupies half of the checksum space since it serves a very similar 251 * purpose: to prevent data corruption on disk. The only functional difference 252 * is that the checksum is used to detect on-disk corruption whether or not the 253 * encryption key is loaded and the MAC provides additional protection against 254 * malicious disk tampering. We use the 3rd DVA to store the salt and first 255 * 64 bits of the IV. As a result encrypted blocks can only have 2 copies 256 * maximum instead of the normal 3. The last 32 bits of the IV are stored in 257 * the upper bits of what is usually the fill count. Note that only blocks at 258 * level 0 or -2 are ever encrypted, which allows us to guarantee that these 259 * 32 bits are not trampled over by other code (see zio_crypt.c for details). 260 * The salt and IV are not used for authenticated bps or bps with an indirect 261 * MAC checksum, so these blocks can utilize all 3 DVAs and the full 64 bits 262 * for the fill count. 263 */ 264 265 /* 266 * "Embedded" blkptr_t's don't actually point to a block, instead they 267 * have a data payload embedded in the blkptr_t itself. See the comment 268 * in blkptr.c for more details. 269 * 270 * The blkptr_t is laid out as follows: 271 * 272 * 64 56 48 40 32 24 16 8 0 273 * +-------+-------+-------+-------+-------+-------+-------+-------+ 274 * 0 | payload | 275 * 1 | payload | 276 * 2 | payload | 277 * 3 | payload | 278 * 4 | payload | 279 * 5 | payload | 280 * +-------+-------+-------+-------+-------+-------+-------+-------+ 281 * 6 |BDX|lvl| type | etype |E| comp| PSIZE| LSIZE | 282 * +-------+-------+-------+-------+-------+-------+-------+-------+ 283 * 7 | payload | 284 * 8 | payload | 285 * 9 | payload | 286 * +-------+-------+-------+-------+-------+-------+-------+-------+ 287 * a | logical birth txg | 288 * +-------+-------+-------+-------+-------+-------+-------+-------+ 289 * b | payload | 290 * c | payload | 291 * d | payload | 292 * e | payload | 293 * f | payload | 294 * +-------+-------+-------+-------+-------+-------+-------+-------+ 295 * 296 * Legend: 297 * 298 * payload contains the embedded data 299 * B (byteorder) byteorder (endianness) 300 * D (dedup) padding (set to zero) 301 * X encryption (set to zero) 302 * E (embedded) set to one 303 * lvl indirection level 304 * type DMU object type 305 * etype how to interpret embedded data (BP_EMBEDDED_TYPE_*) 306 * comp compression function of payload 307 * PSIZE size of payload after compression, in bytes 308 * LSIZE logical size of payload, in bytes 309 * note that 25 bits is enough to store the largest 310 * "normal" BP's LSIZE (2^16 * 2^9) in bytes 311 * log. birth transaction group in which the block was logically born 312 * 313 * Note that LSIZE and PSIZE are stored in bytes, whereas for non-embedded 314 * bp's they are stored in units of SPA_MINBLOCKSHIFT. 315 * Generally, the generic BP_GET_*() macros can be used on embedded BP's. 316 * The B, D, X, lvl, type, and comp fields are stored the same as with normal 317 * BP's so the BP_SET_* macros can be used with them. etype, PSIZE, LSIZE must 318 * be set with the BPE_SET_* macros. BP_SET_EMBEDDED() should be called before 319 * other macros, as they assert that they are only used on BP's of the correct 320 * "embedded-ness". Encrypted blkptr_t's cannot be embedded because they use 321 * the payload space for encryption parameters (see the comment above on 322 * how encryption parameters are stored). 323 */ 324 325 #define BPE_GET_ETYPE(bp) \ 326 (ASSERT(BP_IS_EMBEDDED(bp)), \ 327 BF64_GET((bp)->blk_prop, 40, 8)) 328 #define BPE_SET_ETYPE(bp, t) do { \ 329 ASSERT(BP_IS_EMBEDDED(bp)); \ 330 BF64_SET((bp)->blk_prop, 40, 8, t); \ 331 } while (0) 332 333 #define BPE_GET_LSIZE(bp) \ 334 (ASSERT(BP_IS_EMBEDDED(bp)), \ 335 BF64_GET_SB((bp)->blk_prop, 0, 25, 0, 1)) 336 #define BPE_SET_LSIZE(bp, x) do { \ 337 ASSERT(BP_IS_EMBEDDED(bp)); \ 338 BF64_SET_SB((bp)->blk_prop, 0, 25, 0, 1, x); \ 339 } while (0) 340 341 #define BPE_GET_PSIZE(bp) \ 342 (ASSERT(BP_IS_EMBEDDED(bp)), \ 343 BF64_GET_SB((bp)->blk_prop, 25, 7, 0, 1)) 344 #define BPE_SET_PSIZE(bp, x) do { \ 345 ASSERT(BP_IS_EMBEDDED(bp)); \ 346 BF64_SET_SB((bp)->blk_prop, 25, 7, 0, 1, x); \ 347 } while (0) 348 349 typedef enum bp_embedded_type { 350 BP_EMBEDDED_TYPE_DATA, 351 BP_EMBEDDED_TYPE_RESERVED, /* Reserved for Delphix byteswap feature. */ 352 BP_EMBEDDED_TYPE_REDACTED, 353 NUM_BP_EMBEDDED_TYPES 354 } bp_embedded_type_t; 355 356 #define BPE_NUM_WORDS 14 357 #define BPE_PAYLOAD_SIZE (BPE_NUM_WORDS * sizeof (uint64_t)) 358 #define BPE_IS_PAYLOADWORD(bp, wp) \ 359 ((wp) != &(bp)->blk_prop && (wp) != (&(bp)->blk_birth_word[1])) 360 361 #define SPA_BLKPTRSHIFT 7 /* blkptr_t is 128 bytes */ 362 #define SPA_DVAS_PER_BP 3 /* Number of DVAs in a bp */ 363 #define SPA_SYNC_MIN_VDEVS 3 /* min vdevs to update during sync */ 364 365 /* 366 * A block is a hole when it has either 1) never been written to, or 367 * 2) is zero-filled. In both cases, ZFS can return all zeroes for all reads 368 * without physically allocating disk space. Holes are represented in the 369 * blkptr_t structure by zeroed blk_dva. Correct checking for holes is 370 * done through the BP_IS_HOLE macro. For holes, the logical size, level, 371 * DMU object type, and birth times are all also stored for holes that 372 * were written to at some point (i.e. were punched after having been filled). 373 */ 374 typedef struct blkptr { 375 dva_t blk_dva[SPA_DVAS_PER_BP]; /* Data Virtual Addresses */ 376 uint64_t blk_prop; /* size, compression, type, etc */ 377 uint64_t blk_prop2; /* additional properties */ 378 uint64_t blk_pad; /* Extra space for the future */ 379 uint64_t blk_birth_word[2]; 380 uint64_t blk_fill; /* fill count */ 381 zio_cksum_t blk_cksum; /* 256-bit checksum */ 382 } blkptr_t; 383 384 /* 385 * Macros to get and set fields in a bp or DVA. 386 */ 387 388 /* 389 * Note, for gang blocks, DVA_GET_ASIZE() is the total space allocated for 390 * this gang DVA including its children BP's. The space allocated at this 391 * DVA's vdev/offset is vdev_gang_header_asize(vdev). 392 */ 393 #define DVA_GET_ASIZE(dva) \ 394 BF64_GET_SB((dva)->dva_word[0], 0, SPA_ASIZEBITS, SPA_MINBLOCKSHIFT, 0) 395 #define DVA_SET_ASIZE(dva, x) \ 396 BF64_SET_SB((dva)->dva_word[0], 0, SPA_ASIZEBITS, \ 397 SPA_MINBLOCKSHIFT, 0, x) 398 399 #define DVA_GET_VDEV(dva) BF64_GET((dva)->dva_word[0], 32, SPA_VDEVBITS) 400 #define DVA_SET_VDEV(dva, x) \ 401 BF64_SET((dva)->dva_word[0], 32, SPA_VDEVBITS, x) 402 403 #define DVA_GET_OFFSET(dva) \ 404 BF64_GET_SB((dva)->dva_word[1], 0, 63, SPA_MINBLOCKSHIFT, 0) 405 #define DVA_SET_OFFSET(dva, x) \ 406 BF64_SET_SB((dva)->dva_word[1], 0, 63, SPA_MINBLOCKSHIFT, 0, x) 407 408 #define DVA_GET_GANG(dva) BF64_GET((dva)->dva_word[1], 63, 1) 409 #define DVA_SET_GANG(dva, x) BF64_SET((dva)->dva_word[1], 63, 1, x) 410 411 #define BP_GET_LSIZE(bp) \ 412 (BP_IS_EMBEDDED(bp) ? \ 413 (BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA ? BPE_GET_LSIZE(bp) : 0): \ 414 BF64_GET_SB((bp)->blk_prop, 0, SPA_LSIZEBITS, SPA_MINBLOCKSHIFT, 1)) 415 #define BP_SET_LSIZE(bp, x) do { \ 416 ASSERT(!BP_IS_EMBEDDED(bp)); \ 417 BF64_SET_SB((bp)->blk_prop, \ 418 0, SPA_LSIZEBITS, SPA_MINBLOCKSHIFT, 1, x); \ 419 } while (0) 420 421 #define BP_GET_PSIZE(bp) \ 422 (BP_IS_EMBEDDED(bp) ? 0 : \ 423 BF64_GET_SB((bp)->blk_prop, 16, SPA_PSIZEBITS, SPA_MINBLOCKSHIFT, 1)) 424 #define BP_SET_PSIZE(bp, x) do { \ 425 ASSERT(!BP_IS_EMBEDDED(bp)); \ 426 BF64_SET_SB((bp)->blk_prop, \ 427 16, SPA_PSIZEBITS, SPA_MINBLOCKSHIFT, 1, x); \ 428 } while (0) 429 430 #define BP_GET_COMPRESS(bp) \ 431 BF64_GET((bp)->blk_prop, 32, SPA_COMPRESSBITS) 432 #define BP_SET_COMPRESS(bp, x) \ 433 BF64_SET((bp)->blk_prop, 32, SPA_COMPRESSBITS, x) 434 435 #define BP_IS_EMBEDDED(bp) BF64_GET((bp)->blk_prop, 39, 1) 436 #define BP_SET_EMBEDDED(bp, x) BF64_SET((bp)->blk_prop, 39, 1, x) 437 438 #define BP_GET_CHECKSUM(bp) \ 439 (BP_IS_EMBEDDED(bp) ? ZIO_CHECKSUM_OFF : \ 440 BF64_GET((bp)->blk_prop, 40, 8)) 441 #define BP_SET_CHECKSUM(bp, x) do { \ 442 ASSERT(!BP_IS_EMBEDDED(bp)); \ 443 BF64_SET((bp)->blk_prop, 40, 8, x); \ 444 } while (0) 445 446 #define BP_GET_TYPE(bp) BF64_GET((bp)->blk_prop, 48, 8) 447 #define BP_SET_TYPE(bp, x) BF64_SET((bp)->blk_prop, 48, 8, x) 448 449 #define BP_GET_LEVEL(bp) BF64_GET((bp)->blk_prop, 56, 5) 450 #define BP_SET_LEVEL(bp, x) BF64_SET((bp)->blk_prop, 56, 5, x) 451 452 /* encrypted, authenticated, and MAC cksum bps use the same bit */ 453 #define BP_USES_CRYPT(bp) BF64_GET((bp)->blk_prop, 61, 1) 454 #define BP_SET_CRYPT(bp, x) BF64_SET((bp)->blk_prop, 61, 1, x) 455 456 #define BP_IS_ENCRYPTED(bp) \ 457 (BP_USES_CRYPT(bp) && \ 458 BP_GET_LEVEL(bp) <= 0 && \ 459 DMU_OT_IS_ENCRYPTED(BP_GET_TYPE(bp))) 460 461 #define BP_IS_AUTHENTICATED(bp) \ 462 (BP_USES_CRYPT(bp) && \ 463 BP_GET_LEVEL(bp) <= 0 && \ 464 !DMU_OT_IS_ENCRYPTED(BP_GET_TYPE(bp))) 465 466 #define BP_HAS_INDIRECT_MAC_CKSUM(bp) \ 467 (BP_USES_CRYPT(bp) && BP_GET_LEVEL(bp) > 0) 468 469 #define BP_IS_PROTECTED(bp) \ 470 (BP_IS_ENCRYPTED(bp) || BP_IS_AUTHENTICATED(bp)) 471 472 #define BP_GET_DEDUP(bp) BF64_GET((bp)->blk_prop, 62, 1) 473 #define BP_SET_DEDUP(bp, x) BF64_SET((bp)->blk_prop, 62, 1, x) 474 475 #define BP_GET_BYTEORDER(bp) BF64_GET((bp)->blk_prop, 63, 1) 476 #define BP_SET_BYTEORDER(bp, x) BF64_SET((bp)->blk_prop, 63, 1, x) 477 478 #define BP_GET_FREE(bp) BF64_GET((bp)->blk_fill, 0, 1) 479 #define BP_SET_FREE(bp, x) BF64_SET((bp)->blk_fill, 0, 1, x) 480 481 /* 482 * Block birth time macros for different use cases: 483 * - BP_GET_LOGICAL_BIRTH(): When the block was logically modified by user. 484 * To be used with a focus on user data, like incremental replication. 485 * - BP_GET_PHYSICAL_BIRTH(): When the block was physically written to disks. 486 * For regular writes is equal to logical birth. For dedup and block cloning 487 * can be smaller than logical birth. For remapped and rewritten blocks can 488 * be bigger. To be used with focus on physical disk content: ARC, DDT, scrub. 489 * - BP_GET_RAW_PHYSICAL_BIRTH(): Raw physical birth value. Zero if equal 490 * to logical birth. Should only be used for BP copying and debugging. 491 * - BP_GET_BIRTH(): When the block was allocated, which is a physical birth 492 * for rewritten blocks (rewrite flag set) or logical birth otherwise. 493 */ 494 #define BP_GET_LOGICAL_BIRTH(bp) (bp)->blk_birth_word[1] 495 #define BP_SET_LOGICAL_BIRTH(bp, x) ((bp)->blk_birth_word[1] = (x)) 496 497 #define BP_GET_RAW_PHYSICAL_BIRTH(bp) (bp)->blk_birth_word[0] 498 #define BP_SET_PHYSICAL_BIRTH(bp, x) ((bp)->blk_birth_word[0] = (x)) 499 500 #define BP_GET_PHYSICAL_BIRTH(bp) \ 501 (BP_IS_EMBEDDED(bp) ? 0 : \ 502 BP_GET_RAW_PHYSICAL_BIRTH(bp) ? BP_GET_RAW_PHYSICAL_BIRTH(bp) : \ 503 BP_GET_LOGICAL_BIRTH(bp)) 504 505 #define BP_GET_BIRTH(bp) \ 506 ((BP_IS_EMBEDDED(bp) || !BP_GET_REWRITE(bp)) ? \ 507 BP_GET_LOGICAL_BIRTH(bp) : BP_GET_PHYSICAL_BIRTH(bp)) 508 509 #define BP_SET_BIRTH(bp, logical, physical) \ 510 { \ 511 ASSERT(!BP_IS_EMBEDDED(bp)); \ 512 BP_SET_LOGICAL_BIRTH(bp, logical); \ 513 BP_SET_PHYSICAL_BIRTH(bp, \ 514 ((logical) == (physical) ? 0 : (physical))); \ 515 } 516 517 #define BP_GET_FILL(bp) \ 518 (BP_IS_EMBEDDED(bp) ? 1 : \ 519 BP_IS_ENCRYPTED(bp) ? BF64_GET((bp)->blk_fill, 0, 32) : \ 520 (bp)->blk_fill) 521 522 #define BP_SET_FILL(bp, fill) \ 523 { \ 524 ASSERT(!BP_IS_EMBEDDED(bp)); \ 525 if (BP_IS_ENCRYPTED(bp)) \ 526 BF64_SET((bp)->blk_fill, 0, 32, fill); \ 527 else \ 528 (bp)->blk_fill = fill; \ 529 } 530 531 #define BP_GET_IV2(bp) \ 532 (ASSERT(BP_IS_ENCRYPTED(bp)), \ 533 BF64_GET((bp)->blk_fill, 32, 32)) 534 #define BP_SET_IV2(bp, iv2) \ 535 { \ 536 ASSERT(BP_IS_ENCRYPTED(bp)); \ 537 BF64_SET((bp)->blk_fill, 32, 32, iv2); \ 538 } 539 540 #define BP_GET_REWRITE(bp) \ 541 (BP_IS_EMBEDDED(bp) ? 0 : BF64_GET((bp)->blk_prop2, 63, 1)) 542 543 #define BP_SET_REWRITE(bp, x) \ 544 { \ 545 ASSERT(!BP_IS_EMBEDDED(bp)); \ 546 BF64_SET((bp)->blk_prop2, 63, 1, x); \ 547 } 548 549 #define BP_IS_METADATA(bp) \ 550 (BP_GET_LEVEL(bp) > 0 || DMU_OT_IS_METADATA(BP_GET_TYPE(bp))) 551 552 #define BP_GET_ASIZE(bp) \ 553 (BP_IS_EMBEDDED(bp) ? 0 : \ 554 DVA_GET_ASIZE(&(bp)->blk_dva[0]) + \ 555 DVA_GET_ASIZE(&(bp)->blk_dva[1]) + \ 556 (DVA_GET_ASIZE(&(bp)->blk_dva[2]) * !BP_IS_ENCRYPTED(bp))) 557 558 #define BP_GET_UCSIZE(bp) \ 559 (BP_IS_METADATA(bp) ? BP_GET_PSIZE(bp) : BP_GET_LSIZE(bp)) 560 561 #define BP_GET_NDVAS(bp) \ 562 (BP_IS_EMBEDDED(bp) ? 0 : \ 563 !!DVA_GET_ASIZE(&(bp)->blk_dva[0]) + \ 564 !!DVA_GET_ASIZE(&(bp)->blk_dva[1]) + \ 565 (!!DVA_GET_ASIZE(&(bp)->blk_dva[2]) * !BP_IS_ENCRYPTED(bp))) 566 567 #define BP_COUNT_GANG(bp) \ 568 (BP_IS_EMBEDDED(bp) ? 0 : \ 569 (DVA_GET_GANG(&(bp)->blk_dva[0]) + \ 570 DVA_GET_GANG(&(bp)->blk_dva[1]) + \ 571 (DVA_GET_GANG(&(bp)->blk_dva[2]) * !BP_IS_ENCRYPTED(bp)))) 572 573 #define DVA_EQUAL(dva1, dva2) \ 574 ((dva1)->dva_word[1] == (dva2)->dva_word[1] && \ 575 (dva1)->dva_word[0] == (dva2)->dva_word[0]) 576 577 #define BP_EQUAL(bp1, bp2) \ 578 (BP_GET_PHYSICAL_BIRTH(bp1) == BP_GET_PHYSICAL_BIRTH(bp2) && \ 579 BP_GET_LOGICAL_BIRTH(bp1) == BP_GET_LOGICAL_BIRTH(bp2) && \ 580 DVA_EQUAL(&(bp1)->blk_dva[0], &(bp2)->blk_dva[0]) && \ 581 DVA_EQUAL(&(bp1)->blk_dva[1], &(bp2)->blk_dva[1]) && \ 582 DVA_EQUAL(&(bp1)->blk_dva[2], &(bp2)->blk_dva[2])) 583 584 585 #define DVA_IS_VALID(dva) (DVA_GET_ASIZE(dva) != 0) 586 587 #define BP_IDENTITY(bp) (ASSERT(!BP_IS_EMBEDDED(bp)), &(bp)->blk_dva[0]) 588 #define BP_IS_GANG(bp) \ 589 (BP_IS_EMBEDDED(bp) ? B_FALSE : DVA_GET_GANG(BP_IDENTITY(bp))) 590 #define DVA_IS_EMPTY(dva) ((dva)->dva_word[0] == 0ULL && \ 591 (dva)->dva_word[1] == 0ULL) 592 #define BP_IS_HOLE(bp) \ 593 (!BP_IS_EMBEDDED(bp) && DVA_IS_EMPTY(BP_IDENTITY(bp))) 594 595 #define BP_SET_REDACTED(bp) \ 596 { \ 597 BP_SET_EMBEDDED(bp, B_TRUE); \ 598 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_REDACTED); \ 599 } 600 #define BP_IS_REDACTED(bp) \ 601 (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_REDACTED) 602 603 /* BP_IS_RAIDZ(bp) assumes no block compression */ 604 #define BP_IS_RAIDZ(bp) (DVA_GET_ASIZE(&(bp)->blk_dva[0]) > \ 605 BP_GET_PSIZE(bp)) 606 607 #define BP_ZERO_DVAS(bp) \ 608 { \ 609 (bp)->blk_dva[0].dva_word[0] = 0; \ 610 (bp)->blk_dva[0].dva_word[1] = 0; \ 611 (bp)->blk_dva[1].dva_word[0] = 0; \ 612 (bp)->blk_dva[1].dva_word[1] = 0; \ 613 (bp)->blk_dva[2].dva_word[0] = 0; \ 614 (bp)->blk_dva[2].dva_word[1] = 0; \ 615 } 616 617 #define BP_ZERO(bp) \ 618 { \ 619 BP_ZERO_DVAS(bp); \ 620 (bp)->blk_prop = 0; \ 621 (bp)->blk_prop2 = 0; \ 622 (bp)->blk_pad = 0; \ 623 (bp)->blk_birth_word[0] = 0; \ 624 (bp)->blk_birth_word[1] = 0; \ 625 (bp)->blk_fill = 0; \ 626 ZIO_SET_CHECKSUM(&(bp)->blk_cksum, 0, 0, 0, 0); \ 627 } 628 629 #ifdef _ZFS_BIG_ENDIAN 630 #define ZFS_HOST_BYTEORDER (0ULL) 631 #else 632 #define ZFS_HOST_BYTEORDER (1ULL) 633 #endif 634 635 #define BP_SHOULD_BYTESWAP(bp) (BP_GET_BYTEORDER(bp) != ZFS_HOST_BYTEORDER) 636 637 #define BP_SPRINTF_LEN 400 638 639 /* 640 * This macro allows code sharing between zfs, libzpool, and mdb. 641 * 'func' is either kmem_scnprintf() or mdb_snprintf(). 642 * 'ws' (whitespace) can be ' ' for single-line format, '\n' for multi-line. 643 */ 644 645 #define SNPRINTF_BLKPTR(func, ws, buf, size, bp, type, checksum, compress) \ 646 { \ 647 static const char *const copyname[] = \ 648 { "zero", "single", "double", "triple" }; \ 649 int len = 0; \ 650 int copies = 0; \ 651 const char *crypt_type; \ 652 if (bp != NULL) { \ 653 if (BP_IS_ENCRYPTED(bp)) { \ 654 crypt_type = "encrypted"; \ 655 /* LINTED E_SUSPICIOUS_COMPARISON */ \ 656 } else if (BP_IS_AUTHENTICATED(bp)) { \ 657 crypt_type = "authenticated"; \ 658 } else if (BP_HAS_INDIRECT_MAC_CKSUM(bp)) { \ 659 crypt_type = "indirect-MAC"; \ 660 } else { \ 661 crypt_type = "unencrypted"; \ 662 } \ 663 } \ 664 if (bp == NULL) { \ 665 len += func(buf + len, size - len, "<NULL>"); \ 666 } else if (BP_IS_HOLE(bp)) { \ 667 len += func(buf + len, size - len, \ 668 "HOLE [L%llu %s] " \ 669 "size=%llxL birth=%lluL", \ 670 (u_longlong_t)BP_GET_LEVEL(bp), \ 671 type, \ 672 (u_longlong_t)BP_GET_LSIZE(bp), \ 673 (u_longlong_t)BP_GET_LOGICAL_BIRTH(bp)); \ 674 } else if (BP_IS_EMBEDDED(bp)) { \ 675 len = func(buf + len, size - len, \ 676 "EMBEDDED [L%llu %s] et=%u %s " \ 677 "size=%llxL/%llxP birth=%lluL", \ 678 (u_longlong_t)BP_GET_LEVEL(bp), \ 679 type, \ 680 (int)BPE_GET_ETYPE(bp), \ 681 compress, \ 682 (u_longlong_t)BPE_GET_LSIZE(bp), \ 683 (u_longlong_t)BPE_GET_PSIZE(bp), \ 684 (u_longlong_t)BP_GET_LOGICAL_BIRTH(bp)); \ 685 } else if (BP_IS_REDACTED(bp)) { \ 686 len += func(buf + len, size - len, \ 687 "REDACTED [L%llu %s] size=%llxL birth=%lluL", \ 688 (u_longlong_t)BP_GET_LEVEL(bp), \ 689 type, \ 690 (u_longlong_t)BP_GET_LSIZE(bp), \ 691 (u_longlong_t)BP_GET_LOGICAL_BIRTH(bp)); \ 692 } else { \ 693 for (int d = 0; d < BP_GET_NDVAS(bp); d++) { \ 694 const dva_t *dva = &bp->blk_dva[d]; \ 695 if (DVA_IS_VALID(dva)) \ 696 copies++; \ 697 len += func(buf + len, size - len, \ 698 "DVA[%d]=<%llu:%llx:%llx>%c", d, \ 699 (u_longlong_t)DVA_GET_VDEV(dva), \ 700 (u_longlong_t)DVA_GET_OFFSET(dva), \ 701 (u_longlong_t)DVA_GET_ASIZE(dva), \ 702 ws); \ 703 } \ 704 if (BP_IS_ENCRYPTED(bp)) { \ 705 len += func(buf + len, size - len, \ 706 "salt=%llx iv=%llx:%llx%c", \ 707 (u_longlong_t)bp->blk_dva[2].dva_word[0], \ 708 (u_longlong_t)bp->blk_dva[2].dva_word[1], \ 709 (u_longlong_t)BP_GET_IV2(bp), \ 710 ws); \ 711 } \ 712 len += func(buf + len, size - len, \ 713 "[L%llu %s] %s %s %s %s %s %s %s%c" \ 714 "size=%llxL/%llxP birth=%lluL/%lluP fill=%llu%c" \ 715 "cksum=%016llx:%016llx:%016llx:%016llx", \ 716 (u_longlong_t)BP_GET_LEVEL(bp), \ 717 type, \ 718 checksum, \ 719 compress, \ 720 crypt_type, \ 721 BP_GET_BYTEORDER(bp) == 0 ? "BE" : "LE", \ 722 BP_IS_GANG(bp) ? "gang" : "contiguous", \ 723 BP_GET_DEDUP(bp) ? "dedup" : "unique", \ 724 copyname[copies], \ 725 ws, \ 726 (u_longlong_t)BP_GET_LSIZE(bp), \ 727 (u_longlong_t)BP_GET_PSIZE(bp), \ 728 (u_longlong_t)BP_GET_LOGICAL_BIRTH(bp), \ 729 (u_longlong_t)BP_GET_PHYSICAL_BIRTH(bp), \ 730 (u_longlong_t)BP_GET_FILL(bp), \ 731 ws, \ 732 (u_longlong_t)bp->blk_cksum.zc_word[0], \ 733 (u_longlong_t)bp->blk_cksum.zc_word[1], \ 734 (u_longlong_t)bp->blk_cksum.zc_word[2], \ 735 (u_longlong_t)bp->blk_cksum.zc_word[3]); \ 736 } \ 737 ASSERT(len < size); \ 738 } 739 740 #define BP_GET_BUFC_TYPE(bp) \ 741 (BP_IS_METADATA(bp) ? ARC_BUFC_METADATA : ARC_BUFC_DATA) 742 743 typedef enum spa_import_type { 744 SPA_IMPORT_EXISTING, 745 SPA_IMPORT_ASSEMBLE 746 } spa_import_type_t; 747 748 typedef enum spa_mode { 749 SPA_MODE_UNINIT = 0, 750 SPA_MODE_READ = 1, 751 SPA_MODE_WRITE = 2, 752 } spa_mode_t; 753 754 /* 755 * Send TRIM commands in-line during normal pool operation while deleting. 756 * OFF: no 757 * ON: yes 758 */ 759 typedef enum { 760 SPA_AUTOTRIM_OFF = 0, /* default */ 761 SPA_AUTOTRIM_ON, 762 } spa_autotrim_t; 763 764 /* 765 * Reason TRIM command was issued, used internally for accounting purposes. 766 */ 767 typedef enum trim_type { 768 TRIM_TYPE_MANUAL = 0, 769 TRIM_TYPE_AUTO = 1, 770 TRIM_TYPE_SIMPLE = 2 771 } trim_type_t; 772 773 /* state manipulation functions */ 774 extern int spa_open(const char *pool, spa_t **, const void *tag); 775 extern int spa_open_rewind(const char *pool, spa_t **, const void *tag, 776 nvlist_t *policy, nvlist_t **config); 777 extern int spa_get_stats(const char *pool, nvlist_t **config, char *altroot, 778 size_t buflen); 779 extern int spa_create(const char *pool, nvlist_t *nvroot, nvlist_t *props, 780 nvlist_t *zplprops, struct dsl_crypto_params *dcp); 781 extern int spa_import(char *pool, nvlist_t *config, nvlist_t *props, 782 uint64_t flags); 783 extern nvlist_t *spa_tryimport(nvlist_t *tryconfig); 784 extern int spa_destroy(const char *pool); 785 extern int spa_checkpoint(const char *pool); 786 extern int spa_checkpoint_discard(const char *pool); 787 extern int spa_export(const char *pool, nvlist_t **oldconfig, boolean_t force, 788 boolean_t hardforce); 789 extern int spa_reset(const char *pool); 790 extern void spa_async_request(spa_t *spa, int flag); 791 extern void spa_async_unrequest(spa_t *spa, int flag); 792 extern void spa_async_suspend(spa_t *spa); 793 extern void spa_async_resume(spa_t *spa); 794 extern int spa_async_tasks(spa_t *spa); 795 extern spa_t *spa_inject_addref(char *pool); 796 extern void spa_inject_delref(spa_t *spa); 797 extern void spa_scan_stat_init(spa_t *spa); 798 extern int spa_scan_get_stats(spa_t *spa, pool_scan_stat_t *ps); 799 extern int bpobj_enqueue_alloc_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx); 800 extern int bpobj_enqueue_free_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx); 801 802 #define SPA_ASYNC_CONFIG_UPDATE 0x01 803 #define SPA_ASYNC_REMOVE 0x02 804 #define SPA_ASYNC_FAULT_VDEV 0x04 805 #define SPA_ASYNC_RESILVER_DONE 0x08 806 #define SPA_ASYNC_RESILVER 0x10 807 #define SPA_ASYNC_AUTOEXPAND 0x20 808 #define SPA_ASYNC_REMOVE_DONE 0x40 809 #define SPA_ASYNC_REMOVE_STOP 0x80 810 #define SPA_ASYNC_INITIALIZE_RESTART 0x100 811 #define SPA_ASYNC_TRIM_RESTART 0x200 812 #define SPA_ASYNC_AUTOTRIM_RESTART 0x400 813 #define SPA_ASYNC_L2CACHE_REBUILD 0x800 814 #define SPA_ASYNC_L2CACHE_TRIM 0x1000 815 #define SPA_ASYNC_REBUILD_DONE 0x2000 816 #define SPA_ASYNC_DETACH_SPARE 0x4000 817 #define SPA_ASYNC_REMOVE_BY_USER 0x8000 818 819 /* device manipulation */ 820 extern int spa_vdev_add(spa_t *spa, nvlist_t *nvroot, boolean_t ashift_check); 821 extern int spa_vdev_attach(spa_t *spa, uint64_t guid, nvlist_t *nvroot, 822 int replacing, int rebuild); 823 extern int spa_vdev_detach(spa_t *spa, uint64_t guid, uint64_t pguid, 824 int replace_done); 825 extern int spa_vdev_alloc(spa_t *spa, uint64_t guid); 826 extern int spa_vdev_noalloc(spa_t *spa, uint64_t guid); 827 extern boolean_t spa_vdev_remove_active(spa_t *spa); 828 extern int spa_vdev_initialize(spa_t *spa, nvlist_t *nv, uint64_t cmd_type, 829 nvlist_t *vdev_errlist); 830 extern int spa_vdev_trim(spa_t *spa, nvlist_t *nv, uint64_t cmd_type, 831 uint64_t rate, boolean_t partial, boolean_t secure, nvlist_t *vdev_errlist); 832 extern int spa_vdev_setpath(spa_t *spa, uint64_t guid, const char *newpath); 833 extern int spa_vdev_setfru(spa_t *spa, uint64_t guid, const char *newfru); 834 extern int spa_vdev_split_mirror(spa_t *spa, const char *newname, 835 nvlist_t *config, nvlist_t *props, boolean_t exp); 836 837 /* spare state (which is global across all pools) */ 838 extern void spa_spare_add(vdev_t *vd); 839 extern void spa_spare_remove(vdev_t *vd); 840 extern boolean_t spa_spare_exists(uint64_t guid, uint64_t *pool, int *refcnt); 841 extern void spa_spare_activate(vdev_t *vd); 842 843 /* L2ARC state (which is global across all pools) */ 844 extern void spa_l2cache_add(vdev_t *vd); 845 extern void spa_l2cache_remove(vdev_t *vd); 846 extern boolean_t spa_l2cache_exists(uint64_t guid, uint64_t *pool); 847 extern void spa_l2cache_activate(vdev_t *vd); 848 extern void spa_l2cache_drop(spa_t *spa); 849 850 /* scanning */ 851 extern int spa_scan(spa_t *spa, pool_scan_func_t func); 852 extern int spa_scan_range(spa_t *spa, pool_scan_func_t func, uint64_t txgstart, 853 uint64_t txgend); 854 extern int spa_scan_stop(spa_t *spa); 855 extern int spa_scrub_pause_resume(spa_t *spa, pool_scrub_cmd_t flag); 856 857 /* spa syncing */ 858 extern void spa_sync(spa_t *spa, uint64_t txg); /* only for DMU use */ 859 extern void spa_sync_allpools(void); 860 861 extern uint_t zfs_sync_pass_deferred_free; 862 863 /* spa sync taskqueues */ 864 taskq_t *spa_sync_tq_create(spa_t *spa, const char *name); 865 void spa_sync_tq_destroy(spa_t *spa); 866 uint_t spa_acq_allocator(spa_t *spa); 867 void spa_rel_allocator(spa_t *spa, uint_t allocator); 868 void spa_select_allocator(zio_t *zio); 869 870 /* spa namespace global lock */ 871 extern void spa_namespace_enter(const void *tag); 872 extern boolean_t spa_namespace_tryenter(const void *tag); 873 extern int spa_namespace_enter_interruptible(const void *tag); 874 extern void spa_namespace_exit(const void *tag); 875 extern boolean_t spa_namespace_held(void); 876 extern void spa_namespace_wait(void); 877 extern void spa_namespace_broadcast(void); 878 879 /* 880 * SPA configuration functions in spa_config.c 881 */ 882 883 #define SPA_CONFIG_UPDATE_POOL 0 884 #define SPA_CONFIG_UPDATE_VDEVS 1 885 886 extern void spa_write_cachefile(spa_t *, boolean_t, boolean_t, boolean_t); 887 extern int spa_all_configs(uint64_t *generation, nvlist_t **pools); 888 extern void spa_config_set(spa_t *spa, nvlist_t *config); 889 extern nvlist_t *spa_config_generate(spa_t *spa, vdev_t *vd, uint64_t txg, 890 int getstats); 891 extern void spa_config_update(spa_t *spa, int what); 892 extern int spa_config_parse(spa_t *spa, vdev_t **vdp, nvlist_t *nv, 893 vdev_t *parent, uint_t id, int atype); 894 895 896 /* 897 * Miscellaneous SPA routines in spa_misc.c 898 */ 899 900 /* Namespace manipulation */ 901 extern spa_t *spa_lookup(const char *name); 902 extern spa_t *spa_add(const char *name, nvlist_t *config, const char *altroot); 903 extern void spa_remove(spa_t *spa); 904 extern spa_t *spa_next(spa_t *prev); 905 906 /* Refcount functions */ 907 extern void spa_open_ref(spa_t *spa, const void *tag); 908 extern void spa_close(spa_t *spa, const void *tag); 909 extern void spa_async_close(spa_t *spa, const void *tag); 910 extern boolean_t spa_refcount_zero(spa_t *spa); 911 912 #define SCL_NONE 0x00 913 #define SCL_CONFIG 0x01 914 #define SCL_STATE 0x02 915 #define SCL_L2ARC 0x04 /* hack until L2ARC 2.0 */ 916 #define SCL_ALLOC 0x08 917 #define SCL_ZIO 0x10 918 #define SCL_FREE 0x20 919 #define SCL_VDEV 0x40 920 #define SCL_LOCKS 7 921 #define SCL_ALL ((1 << SCL_LOCKS) - 1) 922 #define SCL_STATE_ALL (SCL_STATE | SCL_L2ARC | SCL_ZIO) 923 924 /* Historical pool statistics */ 925 typedef struct spa_history_kstat { 926 kmutex_t lock; 927 uint64_t count; 928 uint64_t size; 929 kstat_t *kstat; 930 void *priv; 931 list_t list; 932 } spa_history_kstat_t; 933 934 typedef struct spa_history_list { 935 uint64_t size; 936 procfs_list_t procfs_list; 937 } spa_history_list_t; 938 939 typedef struct spa_stats { 940 spa_history_list_t read_history; 941 spa_history_list_t txg_history; 942 spa_history_kstat_t tx_assign_histogram; 943 spa_history_list_t mmp_history; 944 spa_history_kstat_t state; /* pool state */ 945 spa_history_kstat_t guid; /* pool guid */ 946 spa_history_kstat_t iostats; 947 } spa_stats_t; 948 949 typedef enum txg_state { 950 TXG_STATE_BIRTH = 0, 951 TXG_STATE_OPEN = 1, 952 TXG_STATE_QUIESCED = 2, 953 TXG_STATE_WAIT_FOR_SYNC = 3, 954 TXG_STATE_SYNCED = 4, 955 TXG_STATE_COMMITTED = 5, 956 } txg_state_t; 957 958 typedef struct txg_stat { 959 vdev_stat_t vs1; 960 vdev_stat_t vs2; 961 uint64_t txg; 962 uint64_t ndirty; 963 } txg_stat_t; 964 965 /* Assorted pool IO kstats */ 966 typedef struct spa_iostats { 967 kstat_named_t trim_extents_written; 968 kstat_named_t trim_bytes_written; 969 kstat_named_t trim_extents_skipped; 970 kstat_named_t trim_bytes_skipped; 971 kstat_named_t trim_extents_failed; 972 kstat_named_t trim_bytes_failed; 973 kstat_named_t autotrim_extents_written; 974 kstat_named_t autotrim_bytes_written; 975 kstat_named_t autotrim_extents_skipped; 976 kstat_named_t autotrim_bytes_skipped; 977 kstat_named_t autotrim_extents_failed; 978 kstat_named_t autotrim_bytes_failed; 979 kstat_named_t simple_trim_extents_written; 980 kstat_named_t simple_trim_bytes_written; 981 kstat_named_t simple_trim_extents_skipped; 982 kstat_named_t simple_trim_bytes_skipped; 983 kstat_named_t simple_trim_extents_failed; 984 kstat_named_t simple_trim_bytes_failed; 985 kstat_named_t arc_read_count; 986 kstat_named_t arc_read_bytes; 987 kstat_named_t arc_write_count; 988 kstat_named_t arc_write_bytes; 989 kstat_named_t direct_read_count; 990 kstat_named_t direct_read_bytes; 991 kstat_named_t direct_write_count; 992 kstat_named_t direct_write_bytes; 993 } spa_iostats_t; 994 995 extern void spa_stats_init(spa_t *spa); 996 extern void spa_stats_destroy(spa_t *spa); 997 extern void spa_read_history_add(spa_t *spa, const zbookmark_phys_t *zb, 998 uint32_t aflags); 999 extern void spa_txg_history_add(spa_t *spa, uint64_t txg, hrtime_t birth_time); 1000 extern int spa_txg_history_set(spa_t *spa, uint64_t txg, 1001 txg_state_t completed_state, hrtime_t completed_time); 1002 extern txg_stat_t *spa_txg_history_init_io(spa_t *, uint64_t, 1003 struct dsl_pool *); 1004 extern void spa_txg_history_fini_io(spa_t *, txg_stat_t *); 1005 extern void spa_tx_assign_add_nsecs(spa_t *spa, uint64_t nsecs); 1006 extern int spa_mmp_history_set_skip(spa_t *spa, uint64_t mmp_kstat_id); 1007 extern int spa_mmp_history_set(spa_t *spa, uint64_t mmp_kstat_id, int io_error, 1008 hrtime_t duration); 1009 extern void spa_mmp_history_add(spa_t *spa, uint64_t txg, uint64_t timestamp, 1010 uint64_t mmp_delay, vdev_t *vd, int label, uint64_t mmp_kstat_id, 1011 int error); 1012 extern void spa_iostats_trim_add(spa_t *spa, trim_type_t type, 1013 uint64_t extents_written, uint64_t bytes_written, 1014 uint64_t extents_skipped, uint64_t bytes_skipped, 1015 uint64_t extents_failed, uint64_t bytes_failed); 1016 extern void spa_iostats_read_add(spa_t *spa, uint64_t size, uint64_t iops, 1017 dmu_flags_t flags); 1018 extern void spa_iostats_write_add(spa_t *spa, uint64_t size, uint64_t iops, 1019 dmu_flags_t flags); 1020 extern void spa_import_progress_add(spa_t *spa); 1021 extern void spa_import_progress_remove(uint64_t spa_guid); 1022 extern int spa_import_progress_set_mmp_check(uint64_t pool_guid, 1023 uint64_t mmp_sec_remaining); 1024 extern int spa_import_progress_set_max_txg(uint64_t pool_guid, 1025 uint64_t max_txg); 1026 extern int spa_import_progress_set_state(uint64_t pool_guid, 1027 spa_load_state_t spa_load_state); 1028 extern void spa_import_progress_set_notes(spa_t *spa, 1029 const char *fmt, ...) __printflike(2, 3); 1030 extern void spa_import_progress_set_notes_nolog(spa_t *spa, 1031 const char *fmt, ...) __printflike(2, 3); 1032 1033 /* Pool configuration locks */ 1034 extern int spa_config_tryenter(spa_t *spa, int locks, const void *tag, 1035 krw_t rw); 1036 extern void spa_config_enter(spa_t *spa, int locks, const void *tag, krw_t rw); 1037 extern void spa_config_enter_priority(spa_t *spa, int locks, const void *tag, 1038 krw_t rw); 1039 extern void spa_config_exit(spa_t *spa, int locks, const void *tag); 1040 extern int spa_config_held(spa_t *spa, int locks, krw_t rw); 1041 1042 /* Pool vdev add/remove lock */ 1043 extern uint64_t spa_vdev_enter(spa_t *spa); 1044 extern uint64_t spa_vdev_detach_enter(spa_t *spa, uint64_t guid); 1045 extern uint64_t spa_vdev_config_enter(spa_t *spa); 1046 extern void spa_vdev_config_exit(spa_t *spa, vdev_t *vd, uint64_t txg, 1047 int error, const char *tag); 1048 extern int spa_vdev_exit(spa_t *spa, vdev_t *vd, uint64_t txg, int error); 1049 1050 /* Pool vdev state change lock */ 1051 extern void spa_vdev_state_enter(spa_t *spa, int oplock); 1052 extern int spa_vdev_state_exit(spa_t *spa, vdev_t *vd, int error); 1053 1054 /* Log state */ 1055 typedef enum spa_log_state { 1056 SPA_LOG_UNKNOWN = 0, /* unknown log state */ 1057 SPA_LOG_MISSING, /* missing log(s) */ 1058 SPA_LOG_CLEAR, /* clear the log(s) */ 1059 SPA_LOG_GOOD, /* log(s) are good */ 1060 } spa_log_state_t; 1061 1062 extern spa_log_state_t spa_get_log_state(spa_t *spa); 1063 extern void spa_set_log_state(spa_t *spa, spa_log_state_t state); 1064 extern int spa_reset_logs(spa_t *spa); 1065 1066 /* Log claim callback */ 1067 extern void spa_claim_notify(zio_t *zio); 1068 extern void spa_deadman(void *); 1069 1070 /* Accessor functions */ 1071 extern boolean_t spa_shutting_down(spa_t *spa); 1072 extern struct dsl_pool *spa_get_dsl(spa_t *spa); 1073 extern boolean_t spa_is_initializing(spa_t *spa); 1074 extern boolean_t spa_indirect_vdevs_loaded(spa_t *spa); 1075 extern blkptr_t *spa_get_rootblkptr(spa_t *spa); 1076 extern void spa_set_rootblkptr(spa_t *spa, const blkptr_t *bp); 1077 extern void spa_altroot(spa_t *, char *, size_t); 1078 extern uint32_t spa_sync_pass(spa_t *spa); 1079 extern char *spa_name(spa_t *spa); 1080 extern uint64_t spa_guid(spa_t *spa); 1081 extern uint64_t spa_load_guid(spa_t *spa); 1082 extern uint64_t spa_last_synced_txg(spa_t *spa); 1083 extern uint64_t spa_first_txg(spa_t *spa); 1084 extern uint64_t spa_syncing_txg(spa_t *spa); 1085 extern uint64_t spa_final_dirty_txg(spa_t *spa); 1086 extern uint64_t spa_version(spa_t *spa); 1087 extern pool_state_t spa_state(spa_t *spa); 1088 extern spa_load_state_t spa_load_state(spa_t *spa); 1089 extern uint64_t spa_freeze_txg(spa_t *spa); 1090 extern uint64_t spa_get_worst_case_asize(spa_t *spa, uint64_t lsize); 1091 extern void spa_get_min_alloc_range(spa_t *spa, uint64_t *min, uint64_t *max); 1092 extern uint64_t spa_get_dspace(spa_t *spa); 1093 extern uint64_t spa_get_checkpoint_space(spa_t *spa); 1094 extern uint64_t spa_get_slop_space(spa_t *spa); 1095 extern void spa_update_dspace(spa_t *spa); 1096 extern uint64_t spa_version(spa_t *spa); 1097 extern boolean_t spa_deflate(spa_t *spa); 1098 extern metaslab_class_t *spa_normal_class(spa_t *spa); 1099 extern metaslab_class_t *spa_log_class(spa_t *spa); 1100 extern metaslab_class_t *spa_embedded_log_class(spa_t *spa); 1101 extern metaslab_class_t *spa_special_class(spa_t *spa); 1102 extern metaslab_class_t *spa_special_embedded_log_class(spa_t *spa); 1103 extern metaslab_class_t *spa_dedup_class(spa_t *spa); 1104 extern metaslab_class_t *spa_preferred_class(spa_t *spa, const zio_t *zio); 1105 extern boolean_t spa_special_has_ddt(spa_t *spa); 1106 1107 extern void spa_evicting_os_register(spa_t *, objset_t *os); 1108 extern void spa_evicting_os_deregister(spa_t *, objset_t *os); 1109 extern void spa_evicting_os_wait(spa_t *spa); 1110 extern int spa_max_replication(spa_t *spa); 1111 extern int spa_prev_software_version(spa_t *spa); 1112 extern uint64_t spa_get_failmode(spa_t *spa); 1113 extern uint64_t spa_get_deadman_failmode(spa_t *spa); 1114 extern void spa_set_deadman_failmode(spa_t *spa, const char *failmode); 1115 extern boolean_t spa_suspended(spa_t *spa); 1116 extern uint64_t spa_bootfs(spa_t *spa); 1117 extern uint64_t spa_get_last_scrubbed_txg(spa_t *spa); 1118 extern uint64_t spa_delegation(spa_t *spa); 1119 extern objset_t *spa_meta_objset(spa_t *spa); 1120 extern space_map_t *spa_syncing_log_sm(spa_t *spa); 1121 extern uint64_t spa_deadman_synctime(spa_t *spa); 1122 extern uint64_t spa_deadman_ziotime(spa_t *spa); 1123 extern uint64_t spa_dirty_data(spa_t *spa); 1124 extern spa_autotrim_t spa_get_autotrim(spa_t *spa); 1125 extern int spa_get_allocator(spa_t *spa); 1126 extern void spa_set_allocator(spa_t *spa, const char *allocator); 1127 1128 /* Miscellaneous support routines */ 1129 extern void spa_load_failed(spa_t *spa, const char *fmt, ...) 1130 __attribute__((format(printf, 2, 3))); 1131 extern void spa_load_note(spa_t *spa, const char *fmt, ...) 1132 __attribute__((format(printf, 2, 3))); 1133 extern void spa_activate_mos_feature(spa_t *spa, const char *feature, 1134 dmu_tx_t *tx); 1135 extern void spa_deactivate_mos_feature(spa_t *spa, const char *feature); 1136 extern spa_t *spa_by_guid(uint64_t pool_guid, uint64_t device_guid); 1137 extern boolean_t spa_guid_exists(uint64_t pool_guid, uint64_t device_guid); 1138 extern char *spa_strdup(const char *); 1139 extern void spa_strfree(char *); 1140 extern uint64_t spa_generate_guid(spa_t *spa); 1141 extern uint64_t spa_generate_load_guid(void); 1142 extern void snprintf_blkptr(char *buf, size_t buflen, const blkptr_t *bp); 1143 extern void spa_freeze(spa_t *spa); 1144 extern int spa_change_guid(spa_t *spa, const uint64_t *guidp); 1145 extern void spa_upgrade(spa_t *spa, uint64_t version); 1146 extern void spa_evict_all(void); 1147 extern vdev_t *spa_lookup_by_guid(spa_t *spa, uint64_t guid, 1148 boolean_t l2cache); 1149 extern boolean_t spa_has_l2cache(spa_t *, uint64_t guid); 1150 extern boolean_t spa_has_spare(spa_t *, uint64_t guid); 1151 extern uint64_t dva_get_dsize_sync(spa_t *spa, const dva_t *dva); 1152 extern uint64_t bp_get_dsize_sync(spa_t *spa, const blkptr_t *bp); 1153 extern uint64_t bp_get_dsize(spa_t *spa, const blkptr_t *bp); 1154 extern boolean_t spa_has_dedup(spa_t *spa); 1155 extern boolean_t spa_has_slogs(spa_t *spa); 1156 extern boolean_t spa_has_special(spa_t *spa); 1157 extern boolean_t spa_is_root(spa_t *spa); 1158 extern boolean_t spa_writeable(spa_t *spa); 1159 extern boolean_t spa_has_pending_synctask(spa_t *spa); 1160 extern int spa_maxblocksize(spa_t *spa); 1161 extern int spa_maxdnodesize(spa_t *spa); 1162 extern boolean_t spa_has_checkpoint(spa_t *spa); 1163 extern boolean_t spa_importing_readonly_checkpoint(spa_t *spa); 1164 extern boolean_t spa_suspend_async_destroy(spa_t *spa); 1165 extern uint64_t spa_min_claim_txg(spa_t *spa); 1166 extern boolean_t zfs_dva_valid(spa_t *spa, const dva_t *dva, 1167 const blkptr_t *bp); 1168 typedef void (*spa_remap_cb_t)(uint64_t vdev, uint64_t offset, uint64_t size, 1169 void *arg); 1170 extern boolean_t spa_remap_blkptr(spa_t *spa, blkptr_t *bp, 1171 spa_remap_cb_t callback, void *arg); 1172 extern uint64_t spa_get_last_removal_txg(spa_t *spa); 1173 extern boolean_t spa_trust_config(spa_t *spa); 1174 extern uint64_t spa_missing_tvds_allowed(spa_t *spa); 1175 extern void spa_set_missing_tvds(spa_t *spa, uint64_t missing); 1176 extern boolean_t spa_top_vdevs_spacemap_addressable(spa_t *spa); 1177 extern uint64_t spa_total_metaslabs(spa_t *spa); 1178 extern boolean_t spa_multihost(spa_t *spa); 1179 extern uint32_t spa_get_hostid(spa_t *spa); 1180 extern void spa_activate_allocation_classes(spa_t *, dmu_tx_t *); 1181 extern boolean_t spa_livelist_delete_check(spa_t *spa); 1182 1183 extern boolean_t spa_mmp_remote_host_activity(spa_t *spa); 1184 1185 extern spa_mode_t spa_mode(spa_t *spa); 1186 extern uint64_t zfs_strtonum(const char *str, char **nptr); 1187 1188 extern char *spa_his_ievent_table[]; 1189 1190 extern void spa_history_create_obj(spa_t *spa, dmu_tx_t *tx); 1191 extern int spa_history_get(spa_t *spa, uint64_t *offset, uint64_t *len_read, 1192 char *his_buf); 1193 extern int spa_history_log(spa_t *spa, const char *his_buf); 1194 extern int spa_history_log_nvl(spa_t *spa, nvlist_t *nvl); 1195 extern void spa_history_log_version(spa_t *spa, const char *operation, 1196 dmu_tx_t *tx); 1197 extern void spa_history_log_internal(spa_t *spa, const char *operation, 1198 dmu_tx_t *tx, const char *fmt, ...) __printflike(4, 5); 1199 extern void spa_history_log_internal_ds(struct dsl_dataset *ds, const char *op, 1200 dmu_tx_t *tx, const char *fmt, ...) __printflike(4, 5); 1201 extern void spa_history_log_internal_dd(dsl_dir_t *dd, const char *operation, 1202 dmu_tx_t *tx, const char *fmt, ...) __printflike(4, 5); 1203 1204 extern const char *spa_state_to_name(spa_t *spa); 1205 1206 /* error handling */ 1207 struct zbookmark_phys; 1208 extern void spa_log_error(spa_t *spa, const zbookmark_phys_t *zb, 1209 const uint64_t birth); 1210 extern void spa_remove_error(spa_t *spa, zbookmark_phys_t *zb, 1211 uint64_t birth); 1212 extern int zfs_ereport_post(const char *clazz, spa_t *spa, vdev_t *vd, 1213 const zbookmark_phys_t *zb, zio_t *zio, uint64_t state); 1214 extern boolean_t zfs_ereport_is_valid(const char *clazz, spa_t *spa, vdev_t *vd, 1215 zio_t *zio); 1216 extern void zfs_ereport_taskq_fini(void); 1217 extern void zfs_ereport_clear(spa_t *spa, vdev_t *vd); 1218 extern nvlist_t *zfs_event_create(spa_t *spa, vdev_t *vd, const char *type, 1219 const char *name, nvlist_t *aux); 1220 extern void zfs_post_remove(spa_t *spa, vdev_t *vd, boolean_t by_kernel); 1221 extern void zfs_post_state_change(spa_t *spa, vdev_t *vd, uint64_t laststate); 1222 extern void zfs_post_autoreplace(spa_t *spa, vdev_t *vd); 1223 extern uint64_t spa_approx_errlog_size(spa_t *spa); 1224 extern int spa_get_errlog(spa_t *spa, void *uaddr, uint64_t *count); 1225 extern uint64_t spa_get_last_errlog_size(spa_t *spa); 1226 extern void spa_errlog_rotate(spa_t *spa); 1227 extern void spa_errlog_drain(spa_t *spa); 1228 extern void spa_errlog_sync(spa_t *spa, uint64_t txg); 1229 extern void spa_get_errlists(spa_t *spa, avl_tree_t *last, avl_tree_t *scrub); 1230 extern void spa_delete_dataset_errlog(spa_t *spa, uint64_t ds, dmu_tx_t *tx); 1231 extern void spa_swap_errlog(spa_t *spa, uint64_t new_head_ds, 1232 uint64_t old_head_ds, dmu_tx_t *tx); 1233 extern void sync_error_list(spa_t *spa, avl_tree_t *t, uint64_t *obj, 1234 dmu_tx_t *tx); 1235 extern void spa_upgrade_errlog(spa_t *spa, dmu_tx_t *tx); 1236 extern int find_top_affected_fs(spa_t *spa, uint64_t head_ds, 1237 zbookmark_err_phys_t *zep, uint64_t *top_affected_fs); 1238 extern int find_birth_txg(struct dsl_dataset *ds, zbookmark_err_phys_t *zep, 1239 uint64_t *birth_txg); 1240 extern void zep_to_zb(uint64_t dataset, zbookmark_err_phys_t *zep, 1241 zbookmark_phys_t *zb); 1242 extern void name_to_errphys(char *buf, zbookmark_err_phys_t *zep); 1243 1244 /* vdev mirror */ 1245 extern void vdev_mirror_stat_init(void); 1246 extern void vdev_mirror_stat_fini(void); 1247 1248 /* Initialization and termination */ 1249 extern void spa_init(spa_mode_t mode); 1250 extern void spa_fini(void); 1251 1252 /* properties */ 1253 extern int spa_prop_set(spa_t *spa, nvlist_t *nvp); 1254 extern int spa_prop_get(spa_t *spa, nvlist_t *nvp); 1255 extern int spa_prop_get_nvlist(spa_t *spa, char **props, 1256 unsigned int n_props, nvlist_t *outnvl); 1257 extern void spa_prop_clear_bootfs(spa_t *spa, uint64_t obj, dmu_tx_t *tx); 1258 extern void spa_configfile_set(spa_t *, nvlist_t *, boolean_t); 1259 1260 /* asynchronous event notification */ 1261 extern void spa_event_notify(spa_t *spa, vdev_t *vdev, nvlist_t *hist_nvl, 1262 const char *name); 1263 extern void zfs_ereport_zvol_post(const char *subclass, const char *name, 1264 const char *device_name, const char *raw_name); 1265 1266 /* waiting for pool activities to complete */ 1267 extern int spa_wait(const char *pool, zpool_wait_activity_t activity, 1268 boolean_t *waited); 1269 extern int spa_wait_tag(const char *name, zpool_wait_activity_t activity, 1270 uint64_t tag, boolean_t *waited); 1271 extern void spa_notify_waiters(spa_t *spa); 1272 extern void spa_wake_waiters(spa_t *spa); 1273 1274 extern void spa_import_os(spa_t *spa); 1275 extern void spa_export_os(spa_t *spa); 1276 extern void spa_activate_os(spa_t *spa); 1277 extern void spa_deactivate_os(spa_t *spa); 1278 1279 /* module param call functions */ 1280 int param_set_deadman_ziotime(ZFS_MODULE_PARAM_ARGS); 1281 int param_set_deadman_synctime(ZFS_MODULE_PARAM_ARGS); 1282 int param_set_slop_shift(ZFS_MODULE_PARAM_ARGS); 1283 int param_set_deadman_failmode(ZFS_MODULE_PARAM_ARGS); 1284 int param_set_active_allocator(ZFS_MODULE_PARAM_ARGS); 1285 1286 #ifdef ZFS_DEBUG 1287 #define dprintf_bp(bp, fmt, ...) do { \ 1288 if (zfs_flags & ZFS_DEBUG_DPRINTF) { \ 1289 char *__blkbuf = kmem_alloc(BP_SPRINTF_LEN, KM_SLEEP); \ 1290 snprintf_blkptr(__blkbuf, BP_SPRINTF_LEN, (bp)); \ 1291 dprintf(fmt " %s\n", __VA_ARGS__, __blkbuf); \ 1292 kmem_free(__blkbuf, BP_SPRINTF_LEN); \ 1293 } \ 1294 } while (0) 1295 #else 1296 #define dprintf_bp(bp, fmt, ...) 1297 #endif 1298 1299 extern spa_mode_t spa_mode_global; 1300 extern int zfs_deadman_enabled; 1301 extern uint64_t zfs_deadman_synctime_ms; 1302 extern uint64_t zfs_deadman_ziotime_ms; 1303 extern uint64_t zfs_deadman_checktime_ms; 1304 1305 extern kmem_cache_t *zio_buf_cache[]; 1306 extern kmem_cache_t *zio_data_buf_cache[]; 1307 1308 #ifdef __cplusplus 1309 } 1310 #endif 1311 1312 #endif /* _SYS_SPA_H */ 1313