include/sys/spa.h

   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 http://www.opensolaris.org/os/licensing.
  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, 2014 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 Datto Inc.
  30  */
  31
  32 #ifndef _SYS_SPA_H
  33 #define _SYS_SPA_H
  34
  35 #include <sys/avl.h>
  36 #include <sys/zfs_context.h>
  37 #include <sys/kstat.h>
  38 #include <sys/nvpair.h>
  39 #include <sys/sysmacros.h>
  40 #include <sys/types.h>
  41 #include <sys/fs/zfs.h>
  42 #include <sys/spa_checksum.h>
  43 #include <sys/dmu.h>
  44
  45 #ifdef  __cplusplus
  46 extern "C" {
  47 #endif
  48
  49 /*
  50  * Forward references that lots of things need.
  51  */
  52 typedef struct spa spa_t;
  53 typedef struct vdev vdev_t;
  54 typedef struct metaslab metaslab_t;
  55 typedef struct metaslab_group metaslab_group_t;
  56 typedef struct metaslab_class metaslab_class_t;
  57 typedef struct zio zio_t;
  58 typedef struct zilog zilog_t;
  59 typedef struct spa_aux_vdev spa_aux_vdev_t;
  60 typedef struct ddt ddt_t;
  61 typedef struct ddt_entry ddt_entry_t;
  62 typedef struct zbookmark_phys zbookmark_phys_t;
  63
  64 struct dsl_pool;
  65 struct dsl_dataset;
  66 struct dsl_crypto_params;
  67
  68 /*
  69  * General-purpose 32-bit and 64-bit bitfield encodings.
  70  */
  71 #define BF32_DECODE(x, low, len)        P2PHASE((x) >> (low), 1U << (len))
  72 #define BF64_DECODE(x, low, len)        P2PHASE((x) >> (low), 1ULL << (len))
  73 #define BF32_ENCODE(x, low, len)        (P2PHASE((x), 1U << (len)) << (low))
  74 #define BF64_ENCODE(x, low, len)        (P2PHASE((x), 1ULL << (len)) << (low))
  75
  76 #define BF32_GET(x, low, len)           BF32_DECODE(x, low, len)
  77 #define BF64_GET(x, low, len)           BF64_DECODE(x, low, len)
  78
  79 #define BF32_SET(x, low, len, val) do { \
  80         ASSERT3U(val, <, 1U << (len)); \
  81         ASSERT3U(low + len, <=, 32); \
  82         (x) ^= BF32_ENCODE((x >> low) ^ (val), low, len); \
  83 _NOTE(CONSTCOND) } while (0)
  84
  85 #define BF64_SET(x, low, len, val) do { \
  86         ASSERT3U(val, <, 1ULL << (len)); \
  87         ASSERT3U(low + len, <=, 64); \
  88         ((x) ^= BF64_ENCODE((x >> low) ^ (val), low, len)); \
  89 _NOTE(CONSTCOND) } while (0)
  90
  91 #define BF32_GET_SB(x, low, len, shift, bias)   \
  92         ((BF32_GET(x, low, len) + (bias)) << (shift))
  93 #define BF64_GET_SB(x, low, len, shift, bias)   \
  94         ((BF64_GET(x, low, len) + (bias)) << (shift))
  95
  96 #define BF32_SET_SB(x, low, len, shift, bias, val) do { \
  97         ASSERT(IS_P2ALIGNED(val, 1U << shift)); \
  98         ASSERT3S((val) >> (shift), >=, bias); \
  99         BF32_SET(x, low, len, ((val) >> (shift)) - (bias)); \
 100 _NOTE(CONSTCOND) } while (0)
 101 #define BF64_SET_SB(x, low, len, shift, bias, val) do { \
 102         ASSERT(IS_P2ALIGNED(val, 1ULL << shift)); \
 103         ASSERT3S((val) >> (shift), >=, bias); \
 104         BF64_SET(x, low, len, ((val) >> (shift)) - (bias)); \
 105 _NOTE(CONSTCOND) } while (0)
 106
 107 /*
 108  * We currently support block sizes from 512 bytes to 16MB.
 109  * The benefits of larger blocks, and thus larger IO, need to be weighed
 110  * against the cost of COWing a giant block to modify one byte, and the
 111  * large latency of reading or writing a large block.
 112  *
 113  * Note that although blocks up to 16MB are supported, the recordsize
 114  * property can not be set larger than zfs_max_recordsize (default 1MB).
 115  * See the comment near zfs_max_recordsize in dsl_dataset.c for details.
 116  *
 117  * Note that although the LSIZE field of the blkptr_t can store sizes up
 118  * to 32MB, the dnode's dn_datablkszsec can only store sizes up to
 119  * 32MB - 512 bytes.  Therefore, we limit SPA_MAXBLOCKSIZE to 16MB.
 120  */
 121 #define SPA_MINBLOCKSHIFT       9
 122 #define SPA_OLD_MAXBLOCKSHIFT   17
 123 #define SPA_MAXBLOCKSHIFT       24
 124 #define SPA_MINBLOCKSIZE        (1ULL << SPA_MINBLOCKSHIFT)
 125 #define SPA_OLD_MAXBLOCKSIZE    (1ULL << SPA_OLD_MAXBLOCKSHIFT)
 126 #define SPA_MAXBLOCKSIZE        (1ULL << SPA_MAXBLOCKSHIFT)
 127
 128 /*
 129  * Alignment Shift (ashift) is an immutable, internal top-level vdev property
 130  * which can only be set at vdev creation time. Physical writes are always done
 131  * according to it, which makes 2^ashift the smallest possible IO on a vdev.
 132  *
 133  * We currently allow values ranging from 512 bytes (2^9 = 512) to 64 KiB
 134  * (2^16 = 65,536).
 135  */
 136 #define ASHIFT_MIN              9
 137 #define ASHIFT_MAX              16
 138
 139 /*
 140  * Size of block to hold the configuration data (a packed nvlist)
 141  */
 142 #define SPA_CONFIG_BLOCKSIZE    (1ULL << 14)
 143
 144 /*
 145  * The DVA size encodings for LSIZE and PSIZE support blocks up to 32MB.
 146  * The ASIZE encoding should be at least 64 times larger (6 more bits)
 147  * to support up to 4-way RAID-Z mirror mode with worst-case gang block
 148  * overhead, three DVAs per bp, plus one more bit in case we do anything
 149  * else that expands the ASIZE.
 150  */
 151 #define SPA_LSIZEBITS           16      /* LSIZE up to 32M (2^16 * 512) */
 152 #define SPA_PSIZEBITS           16      /* PSIZE up to 32M (2^16 * 512) */
 153 #define SPA_ASIZEBITS           24      /* ASIZE up to 64 times larger  */
 154
 155 #define SPA_COMPRESSBITS        7
 156
 157 /*
 158  * All SPA data is represented by 128-bit data virtual addresses (DVAs).
 159  * The members of the dva_t should be considered opaque outside the SPA.
 160  */
 161 typedef struct dva {
 162         uint64_t        dva_word[2];
 163 } dva_t;
 164
 165
 166 /*
 167  * Some checksums/hashes need a 256-bit initialization salt. This salt is kept
 168  * secret and is suitable for use in MAC algorithms as the key.
 169  */
 170 typedef struct zio_cksum_salt {
 171         uint8_t         zcs_bytes[32];
 172 } zio_cksum_salt_t;
 173
 174 /*
 175  * Each block is described by its DVAs, time of birth, checksum, etc.
 176  * The word-by-word, bit-by-bit layout of the blkptr is as follows:
 177  *
 178  *      64      56      48      40      32      24      16      8       0
 179  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 180  * 0    |               vdev1           | GRID  |         ASIZE         |
 181  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 182  * 1    |G|                      offset1                                |
 183  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 184  * 2    |               vdev2           | GRID  |         ASIZE         |
 185  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 186  * 3    |G|                      offset2                                |
 187  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 188  * 4    |               vdev3           | GRID  |         ASIZE         |
 189  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 190  * 5    |G|                      offset3                                |
 191  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 192  * 6    |BDX|lvl| type  | cksum |E| comp|    PSIZE      |     LSIZE     |
 193  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 194  * 7    |                       padding                                 |
 195  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 196  * 8    |                       padding                                 |
 197  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 198  * 9    |                       physical birth txg                      |
 199  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 200  * a    |                       logical birth txg                       |
 201  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 202  * b    |                       fill count                              |
 203  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 204  * c    |                       checksum[0]                             |
 205  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 206  * d    |                       checksum[1]                             |
 207  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 208  * e    |                       checksum[2]                             |
 209  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 210  * f    |                       checksum[3]                             |
 211  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 212  *
 213  * Legend:
 214  *
 215  * vdev         virtual device ID
 216  * offset       offset into virtual device
 217  * LSIZE        logical size
 218  * PSIZE        physical size (after compression)
 219  * ASIZE        allocated size (including RAID-Z parity and gang block headers)
 220  * GRID         RAID-Z layout information (reserved for future use)
 221  * cksum        checksum function
 222  * comp         compression function
 223  * G            gang block indicator
 224  * B            byteorder (endianness)
 225  * D            dedup
 226  * X            encryption
 227  * E            blkptr_t contains embedded data (see below)
 228  * lvl          level of indirection
 229  * type         DMU object type
 230  * phys birth   txg when dva[0] was written; zero if same as logical birth txg
 231  *              note that typically all the dva's would be written in this
 232  *              txg, but they could be different if they were moved by
 233  *              device removal.
 234  * log. birth   transaction group in which the block was logically born
 235  * fill count   number of non-zero blocks under this bp
 236  * checksum[4]  256-bit checksum of the data this bp describes
 237  */
 238
 239 /*
 240  * The blkptr_t's of encrypted blocks also need to store the encryption
 241  * parameters so that the block can be decrypted. This layout is as follows:
 242  *
 243  *      64      56      48      40      32      24      16      8       0
 244  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 245  * 0    |               vdev1           | GRID  |         ASIZE         |
 246  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 247  * 1    |G|                      offset1                                |
 248  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 249  * 2    |               vdev2           | GRID  |         ASIZE         |
 250  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 251  * 3    |G|                      offset2                                |
 252  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 253  * 4    |                       salt                                    |
 254  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 255  * 5    |                       IV1                                     |
 256  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 257  * 6    |BDX|lvl| type  | cksum |E| comp|    PSIZE      |     LSIZE     |
 258  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 259  * 7    |                       padding                                 |
 260  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 261  * 8    |                       padding                                 |
 262  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 263  * 9    |                       physical birth txg                      |
 264  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 265  * a    |                       logical birth txg                       |
 266  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 267  * b    |               IV2             |           fill count          |
 268  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 269  * c    |                       checksum[0]                             |
 270  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 271  * d    |                       checksum[1]                             |
 272  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 273  * e    |                       MAC[0]                                  |
 274  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 275  * f    |                       MAC[1]                                  |
 276  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 277  *
 278  * Legend:
 279  *
 280  * salt         Salt for generating encryption keys
 281  * IV1          First 64 bits of encryption IV
 282  * X            Block requires encryption handling (set to 1)
 283  * E            blkptr_t contains embedded data (set to 0, see below)
 284  * fill count   number of non-zero blocks under this bp (truncated to 32 bits)
 285  * IV2          Last 32 bits of encryption IV
 286  * checksum[2]  128-bit checksum of the data this bp describes
 287  * MAC[2]       128-bit message authentication code for this data
 288  *
 289  * The X bit being set indicates that this block is one of 3 types. If this is
 290  * a level 0 block with an encrypted object type, the block is encrypted
 291  * (see BP_IS_ENCRYPTED()). If this is a level 0 block with an unencrypted
 292  * object type, this block is authenticated with an HMAC (see
 293  * BP_IS_AUTHENTICATED()). Otherwise (if level > 0), this bp will use the MAC
 294  * words to store a checksum-of-MACs from the level below (see
 295  * BP_HAS_INDIRECT_MAC_CKSUM()). For convenience in the code, BP_IS_PROTECTED()
 296  * refers to both encrypted and authenticated blocks and BP_USES_CRYPT()
 297  * refers to any of these 3 kinds of blocks.
 298  *
 299  * The additional encryption parameters are the salt, IV, and MAC which are
 300  * explained in greater detail in the block comment at the top of zio_crypt.c.
 301  * The MAC occupies half of the checksum space since it serves a very similar
 302  * purpose: to prevent data corruption on disk. The only functional difference
 303  * is that the checksum is used to detect on-disk corruption whether or not the
 304  * encryption key is loaded and the MAC provides additional protection against
 305  * malicious disk tampering. We use the 3rd DVA to store the salt and first
 306  * 64 bits of the IV. As a result encrypted blocks can only have 2 copies
 307  * maximum instead of the normal 3. The last 32 bits of the IV are stored in
 308  * the upper bits of what is usually the fill count. Note that only blocks at
 309  * level 0 or -2 are ever encrypted, which allows us to guarantee that these
 310  * 32 bits are not trampled over by other code (see zio_crypt.c for details).
 311  * The salt and IV are not used for authenticated bps or bps with an indirect
 312  * MAC checksum, so these blocks can utilize all 3 DVAs and the full 64 bits
 313  * for the fill count.
 314  */
 315
 316 /*
 317  * "Embedded" blkptr_t's don't actually point to a block, instead they
 318  * have a data payload embedded in the blkptr_t itself.  See the comment
 319  * in blkptr.c for more details.
 320  *
 321  * The blkptr_t is laid out as follows:
 322  *
 323  *      64      56      48      40      32      24      16      8       0
 324  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 325  * 0    |      payload                                                  |
 326  * 1    |      payload                                                  |
 327  * 2    |      payload                                                  |
 328  * 3    |      payload                                                  |
 329  * 4    |      payload                                                  |
 330  * 5    |      payload                                                  |
 331  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 332  * 6    |BDX|lvl| type  | etype |E| comp| PSIZE|              LSIZE     |
 333  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 334  * 7    |      payload                                                  |
 335  * 8    |      payload                                                  |
 336  * 9    |      payload                                                  |
 337  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 338  * a    |                       logical birth txg                       |
 339  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 340  * b    |      payload                                                  |
 341  * c    |      payload                                                  |
 342  * d    |      payload                                                  |
 343  * e    |      payload                                                  |
 344  * f    |      payload                                                  |
 345  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 346  *
 347  * Legend:
 348  *
 349  * payload              contains the embedded data
 350  * B (byteorder)        byteorder (endianness)
 351  * D (dedup)            padding (set to zero)
 352  * X                    encryption (set to zero)
 353  * E (embedded)         set to one
 354  * lvl                  indirection level
 355  * type                 DMU object type
 356  * etype                how to interpret embedded data (BP_EMBEDDED_TYPE_*)
 357  * comp                 compression function of payload
 358  * PSIZE                size of payload after compression, in bytes
 359  * LSIZE                logical size of payload, in bytes
 360  *                      note that 25 bits is enough to store the largest
 361  *                      "normal" BP's LSIZE (2^16 * 2^9) in bytes
 362  * log. birth           transaction group in which the block was logically born
 363  *
 364  * Note that LSIZE and PSIZE are stored in bytes, whereas for non-embedded
 365  * bp's they are stored in units of SPA_MINBLOCKSHIFT.
 366  * Generally, the generic BP_GET_*() macros can be used on embedded BP's.
 367  * The B, D, X, lvl, type, and comp fields are stored the same as with normal
 368  * BP's so the BP_SET_* macros can be used with them.  etype, PSIZE, LSIZE must
 369  * be set with the BPE_SET_* macros.  BP_SET_EMBEDDED() should be called before
 370  * other macros, as they assert that they are only used on BP's of the correct
 371  * "embedded-ness". Encrypted blkptr_t's cannot be embedded because they use
 372  * the payload space for encryption parameters (see the comment above on
 373  * how encryption parameters are stored).
 374  */
 375
 376 #define BPE_GET_ETYPE(bp)       \
 377         (ASSERT(BP_IS_EMBEDDED(bp)), \
 378         BF64_GET((bp)->blk_prop, 40, 8))
 379 #define BPE_SET_ETYPE(bp, t)    do { \
 380         ASSERT(BP_IS_EMBEDDED(bp)); \
 381         BF64_SET((bp)->blk_prop, 40, 8, t); \
 382 _NOTE(CONSTCOND) } while (0)
 383
 384 #define BPE_GET_LSIZE(bp)       \
 385         (ASSERT(BP_IS_EMBEDDED(bp)), \
 386         BF64_GET_SB((bp)->blk_prop, 0, 25, 0, 1))
 387 #define BPE_SET_LSIZE(bp, x)    do { \
 388         ASSERT(BP_IS_EMBEDDED(bp)); \
 389         BF64_SET_SB((bp)->blk_prop, 0, 25, 0, 1, x); \
 390 _NOTE(CONSTCOND) } while (0)
 391
 392 #define BPE_GET_PSIZE(bp)       \
 393         (ASSERT(BP_IS_EMBEDDED(bp)), \
 394         BF64_GET_SB((bp)->blk_prop, 25, 7, 0, 1))
 395 #define BPE_SET_PSIZE(bp, x)    do { \
 396         ASSERT(BP_IS_EMBEDDED(bp)); \
 397         BF64_SET_SB((bp)->blk_prop, 25, 7, 0, 1, x); \
 398 _NOTE(CONSTCOND) } while (0)
 399
 400 typedef enum bp_embedded_type {
 401         BP_EMBEDDED_TYPE_DATA,
 402         BP_EMBEDDED_TYPE_RESERVED, /* Reserved for an unintegrated feature. */
 403         NUM_BP_EMBEDDED_TYPES = BP_EMBEDDED_TYPE_RESERVED
 404 } bp_embedded_type_t;
 405
 406 #define BPE_NUM_WORDS 14
 407 #define BPE_PAYLOAD_SIZE (BPE_NUM_WORDS * sizeof (uint64_t))
 408 #define BPE_IS_PAYLOADWORD(bp, wp) \
 409         ((wp) != &(bp)->blk_prop && (wp) != &(bp)->blk_birth)
 410
 411 #define SPA_BLKPTRSHIFT 7               /* blkptr_t is 128 bytes        */
 412 #define SPA_DVAS_PER_BP 3               /* Number of DVAs in a bp       */
 413 #define SPA_SYNC_MIN_VDEVS 3            /* min vdevs to update during sync */
 414
 415 /*
 416  * A block is a hole when it has either 1) never been written to, or
 417  * 2) is zero-filled. In both cases, ZFS can return all zeroes for all reads
 418  * without physically allocating disk space. Holes are represented in the
 419  * blkptr_t structure by zeroed blk_dva. Correct checking for holes is
 420  * done through the BP_IS_HOLE macro. For holes, the logical size, level,
 421  * DMU object type, and birth times are all also stored for holes that
 422  * were written to at some point (i.e. were punched after having been filled).
 423  */
 424 typedef struct blkptr {
 425         dva_t           blk_dva[SPA_DVAS_PER_BP]; /* Data Virtual Addresses */
 426         uint64_t        blk_prop;       /* size, compression, type, etc     */
 427         uint64_t        blk_pad[2];     /* Extra space for the future       */
 428         uint64_t        blk_phys_birth; /* txg when block was allocated     */
 429         uint64_t        blk_birth;      /* transaction group at birth       */
 430         uint64_t        blk_fill;       /* fill count                       */
 431         zio_cksum_t     blk_cksum;      /* 256-bit checksum                 */
 432 } blkptr_t;
 433
 434 /*
 435  * Macros to get and set fields in a bp or DVA.
 436  */
 437 #define DVA_GET_ASIZE(dva)      \
 438         BF64_GET_SB((dva)->dva_word[0], 0, SPA_ASIZEBITS, SPA_MINBLOCKSHIFT, 0)
 439 #define DVA_SET_ASIZE(dva, x)   \
 440         BF64_SET_SB((dva)->dva_word[0], 0, SPA_ASIZEBITS, \
 441         SPA_MINBLOCKSHIFT, 0, x)
 442
 443 #define DVA_GET_GRID(dva)       BF64_GET((dva)->dva_word[0], 24, 8)
 444 #define DVA_SET_GRID(dva, x)    BF64_SET((dva)->dva_word[0], 24, 8, x)
 445
 446 #define DVA_GET_VDEV(dva)       BF64_GET((dva)->dva_word[0], 32, 32)
 447 #define DVA_SET_VDEV(dva, x)    BF64_SET((dva)->dva_word[0], 32, 32, x)
 448
 449 #define DVA_GET_OFFSET(dva)     \
 450         BF64_GET_SB((dva)->dva_word[1], 0, 63, SPA_MINBLOCKSHIFT, 0)
 451 #define DVA_SET_OFFSET(dva, x)  \
 452         BF64_SET_SB((dva)->dva_word[1], 0, 63, SPA_MINBLOCKSHIFT, 0, x)
 453
 454 #define DVA_GET_GANG(dva)       BF64_GET((dva)->dva_word[1], 63, 1)
 455 #define DVA_SET_GANG(dva, x)    BF64_SET((dva)->dva_word[1], 63, 1, x)
 456
 457 #define BP_GET_LSIZE(bp)        \
 458         (BP_IS_EMBEDDED(bp) ?   \
 459         (BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA ? BPE_GET_LSIZE(bp) : 0): \
 460         BF64_GET_SB((bp)->blk_prop, 0, SPA_LSIZEBITS, SPA_MINBLOCKSHIFT, 1))
 461 #define BP_SET_LSIZE(bp, x)     do { \
 462         ASSERT(!BP_IS_EMBEDDED(bp)); \
 463         BF64_SET_SB((bp)->blk_prop, \
 464             0, SPA_LSIZEBITS, SPA_MINBLOCKSHIFT, 1, x); \
 465 _NOTE(CONSTCOND) } while (0)
 466
 467 #define BP_GET_PSIZE(bp)        \
 468         (BP_IS_EMBEDDED(bp) ? 0 : \
 469         BF64_GET_SB((bp)->blk_prop, 16, SPA_PSIZEBITS, SPA_MINBLOCKSHIFT, 1))
 470 #define BP_SET_PSIZE(bp, x)     do { \
 471         ASSERT(!BP_IS_EMBEDDED(bp)); \
 472         BF64_SET_SB((bp)->blk_prop, \
 473             16, SPA_PSIZEBITS, SPA_MINBLOCKSHIFT, 1, x); \
 474 _NOTE(CONSTCOND) } while (0)
 475
 476 #define BP_GET_COMPRESS(bp)             \
 477         BF64_GET((bp)->blk_prop, 32, SPA_COMPRESSBITS)
 478 #define BP_SET_COMPRESS(bp, x)          \
 479         BF64_SET((bp)->blk_prop, 32, SPA_COMPRESSBITS, x)
 480
 481 #define BP_IS_EMBEDDED(bp)              BF64_GET((bp)->blk_prop, 39, 1)
 482 #define BP_SET_EMBEDDED(bp, x)          BF64_SET((bp)->blk_prop, 39, 1, x)
 483
 484 #define BP_GET_CHECKSUM(bp)             \
 485         (BP_IS_EMBEDDED(bp) ? ZIO_CHECKSUM_OFF : \
 486         BF64_GET((bp)->blk_prop, 40, 8))
 487 #define BP_SET_CHECKSUM(bp, x)          do { \
 488         ASSERT(!BP_IS_EMBEDDED(bp)); \
 489         BF64_SET((bp)->blk_prop, 40, 8, x); \
 490 _NOTE(CONSTCOND) } while (0)
 491
 492 #define BP_GET_TYPE(bp)                 BF64_GET((bp)->blk_prop, 48, 8)
 493 #define BP_SET_TYPE(bp, x)              BF64_SET((bp)->blk_prop, 48, 8, x)
 494
 495 #define BP_GET_LEVEL(bp)                BF64_GET((bp)->blk_prop, 56, 5)
 496 #define BP_SET_LEVEL(bp, x)             BF64_SET((bp)->blk_prop, 56, 5, x)
 497
 498 /* encrypted, authenticated, and MAC cksum bps use the same bit */
 499 #define BP_USES_CRYPT(bp)               BF64_GET((bp)->blk_prop, 61, 1)
 500 #define BP_SET_CRYPT(bp, x)             BF64_SET((bp)->blk_prop, 61, 1, x)
 501
 502 #define BP_IS_ENCRYPTED(bp)                     \
 503         (BP_USES_CRYPT(bp) &&                   \
 504         BP_GET_LEVEL(bp) <= 0 &&                \
 505         DMU_OT_IS_ENCRYPTED(BP_GET_TYPE(bp)))
 506
 507 #define BP_IS_AUTHENTICATED(bp)                 \
 508         (BP_USES_CRYPT(bp) &&                   \
 509         BP_GET_LEVEL(bp) <= 0 &&                \
 510         !DMU_OT_IS_ENCRYPTED(BP_GET_TYPE(bp)))
 511
 512 #define BP_HAS_INDIRECT_MAC_CKSUM(bp)           \
 513         (BP_USES_CRYPT(bp) && BP_GET_LEVEL(bp) > 0)
 514
 515 #define BP_IS_PROTECTED(bp)                     \
 516         (BP_IS_ENCRYPTED(bp) || BP_IS_AUTHENTICATED(bp))
 517
 518 #define BP_GET_DEDUP(bp)                BF64_GET((bp)->blk_prop, 62, 1)
 519 #define BP_SET_DEDUP(bp, x)             BF64_SET((bp)->blk_prop, 62, 1, x)
 520
 521 #define BP_GET_BYTEORDER(bp)            BF64_GET((bp)->blk_prop, 63, 1)
 522 #define BP_SET_BYTEORDER(bp, x)         BF64_SET((bp)->blk_prop, 63, 1, x)
 523
 524 #define BP_PHYSICAL_BIRTH(bp)           \
 525         (BP_IS_EMBEDDED(bp) ? 0 : \
 526         (bp)->blk_phys_birth ? (bp)->blk_phys_birth : (bp)->blk_birth)
 527
 528 #define BP_SET_BIRTH(bp, logical, physical)     \
 529 {                                               \
 530         ASSERT(!BP_IS_EMBEDDED(bp));            \
 531         (bp)->blk_birth = (logical);            \
 532         (bp)->blk_phys_birth = ((logical) == (physical) ? 0 : (physical)); \
 533 }
 534
 535 #define BP_GET_FILL(bp)                         \
 536         ((BP_IS_ENCRYPTED(bp)) ? BF64_GET((bp)->blk_fill, 0, 32) : \
 537         ((BP_IS_EMBEDDED(bp)) ? 1 : (bp)->blk_fill))
 538
 539 #define BP_SET_FILL(bp, fill)                   \
 540 {                                               \
 541         if (BP_IS_ENCRYPTED(bp))                        \
 542                 BF64_SET((bp)->blk_fill, 0, 32, fill); \
 543         else                                    \
 544                 (bp)->blk_fill = fill;          \
 545 }
 546
 547 #define BP_GET_IV2(bp)                          \
 548         (ASSERT(BP_IS_ENCRYPTED(bp)),           \
 549         BF64_GET((bp)->blk_fill, 32, 32))
 550 #define BP_SET_IV2(bp, iv2)                     \
 551 {                                               \
 552         ASSERT(BP_IS_ENCRYPTED(bp));            \
 553         BF64_SET((bp)->blk_fill, 32, 32, iv2);  \
 554 }
 555
 556 #define BP_IS_METADATA(bp)      \
 557         (BP_GET_LEVEL(bp) > 0 || DMU_OT_IS_METADATA(BP_GET_TYPE(bp)))
 558
 559 #define BP_GET_ASIZE(bp)        \
 560         (BP_IS_EMBEDDED(bp) ? 0 : \
 561         DVA_GET_ASIZE(&(bp)->blk_dva[0]) + \
 562         DVA_GET_ASIZE(&(bp)->blk_dva[1]) + \
 563         (DVA_GET_ASIZE(&(bp)->blk_dva[2]) * !BP_IS_ENCRYPTED(bp)))
 564
 565 #define BP_GET_UCSIZE(bp)       \
 566         (BP_IS_METADATA(bp) ? BP_GET_PSIZE(bp) : BP_GET_LSIZE(bp))
 567
 568 #define BP_GET_NDVAS(bp)        \
 569         (BP_IS_EMBEDDED(bp) ? 0 : \
 570         !!DVA_GET_ASIZE(&(bp)->blk_dva[0]) + \
 571         !!DVA_GET_ASIZE(&(bp)->blk_dva[1]) + \
 572         (!!DVA_GET_ASIZE(&(bp)->blk_dva[2]) * !BP_IS_ENCRYPTED(bp)))
 573
 574 #define BP_COUNT_GANG(bp)       \
 575         (BP_IS_EMBEDDED(bp) ? 0 : \
 576         (DVA_GET_GANG(&(bp)->blk_dva[0]) + \
 577         DVA_GET_GANG(&(bp)->blk_dva[1]) + \
 578         (DVA_GET_GANG(&(bp)->blk_dva[2]) * !BP_IS_ENCRYPTED(bp))))
 579
 580 #define DVA_EQUAL(dva1, dva2)   \
 581         ((dva1)->dva_word[1] == (dva2)->dva_word[1] && \
 582         (dva1)->dva_word[0] == (dva2)->dva_word[0])
 583
 584 #define BP_EQUAL(bp1, bp2)      \
 585         (BP_PHYSICAL_BIRTH(bp1) == BP_PHYSICAL_BIRTH(bp2) &&    \
 586         (bp1)->blk_birth == (bp2)->blk_birth &&                 \
 587         DVA_EQUAL(&(bp1)->blk_dva[0], &(bp2)->blk_dva[0]) &&    \
 588         DVA_EQUAL(&(bp1)->blk_dva[1], &(bp2)->blk_dva[1]) &&    \
 589         DVA_EQUAL(&(bp1)->blk_dva[2], &(bp2)->blk_dva[2]))
 590
 591
 592 #define DVA_IS_VALID(dva)       (DVA_GET_ASIZE(dva) != 0)
 593
 594 #define BP_IDENTITY(bp)         (ASSERT(!BP_IS_EMBEDDED(bp)), &(bp)->blk_dva[0])
 595 #define BP_IS_GANG(bp)          \
 596         (BP_IS_EMBEDDED(bp) ? B_FALSE : DVA_GET_GANG(BP_IDENTITY(bp)))
 597 #define DVA_IS_EMPTY(dva)       ((dva)->dva_word[0] == 0ULL &&  \
 598                                 (dva)->dva_word[1] == 0ULL)
 599 #define BP_IS_HOLE(bp) \
 600         (!BP_IS_EMBEDDED(bp) && DVA_IS_EMPTY(BP_IDENTITY(bp)))
 601
 602 /* BP_IS_RAIDZ(bp) assumes no block compression */
 603 #define BP_IS_RAIDZ(bp)         (DVA_GET_ASIZE(&(bp)->blk_dva[0]) > \
 604                                 BP_GET_PSIZE(bp))
 605
 606 #define BP_ZERO(bp)                             \
 607 {                                               \
 608         (bp)->blk_dva[0].dva_word[0] = 0;       \
 609         (bp)->blk_dva[0].dva_word[1] = 0;       \
 610         (bp)->blk_dva[1].dva_word[0] = 0;       \
 611         (bp)->blk_dva[1].dva_word[1] = 0;       \
 612         (bp)->blk_dva[2].dva_word[0] = 0;       \
 613         (bp)->blk_dva[2].dva_word[1] = 0;       \
 614         (bp)->blk_prop = 0;                     \
 615         (bp)->blk_pad[0] = 0;                   \
 616         (bp)->blk_pad[1] = 0;                   \
 617         (bp)->blk_phys_birth = 0;               \
 618         (bp)->blk_birth = 0;                    \
 619         (bp)->blk_fill = 0;                     \
 620         ZIO_SET_CHECKSUM(&(bp)->blk_cksum, 0, 0, 0, 0); \
 621 }
 622
 623 #ifdef _BIG_ENDIAN
 624 #define ZFS_HOST_BYTEORDER      (0ULL)
 625 #else
 626 #define ZFS_HOST_BYTEORDER      (1ULL)
 627 #endif
 628
 629 #define BP_SHOULD_BYTESWAP(bp)  (BP_GET_BYTEORDER(bp) != ZFS_HOST_BYTEORDER)
 630
 631 #define BP_SPRINTF_LEN  400
 632
 633 /*
 634  * This macro allows code sharing between zfs, libzpool, and mdb.
 635  * 'func' is either snprintf() or mdb_snprintf().
 636  * 'ws' (whitespace) can be ' ' for single-line format, '\n' for multi-line.
 637  */
 638 #define SNPRINTF_BLKPTR(func, ws, buf, size, bp, type, checksum, compress) \
 639 {                                                                       \
 640         static const char *copyname[] =                                 \
 641             { "zero", "single", "double", "triple" };                   \
 642         int len = 0;                                                    \
 643         int copies = 0;                                                 \
 644         const char *crypt_type;                                         \
 645         if (bp != NULL) {                                               \
 646                 if (BP_IS_ENCRYPTED(bp)) {                              \
 647                         crypt_type = "encrypted";                       \
 648                         /* LINTED E_SUSPICIOUS_COMPARISON */            \
 649                 } else if (BP_IS_AUTHENTICATED(bp)) {                   \
 650                         crypt_type = "authenticated";                   \
 651                 } else if (BP_HAS_INDIRECT_MAC_CKSUM(bp)) {             \
 652                         crypt_type = "indirect-MAC";                    \
 653                 } else {                                                \
 654                         crypt_type = "unencrypted";                     \
 655                 }                                                       \
 656         }                                                               \
 657         if (bp == NULL) {                                               \
 658                 len += func(buf + len, size - len, "<NULL>");           \
 659         } else if (BP_IS_HOLE(bp)) {                                    \
 660                 len += func(buf + len, size - len,                      \
 661                     "HOLE [L%llu %s] "                                  \
 662                     "size=%llxL birth=%lluL",                           \
 663                     (u_longlong_t)BP_GET_LEVEL(bp),                     \
 664                     type,                                               \
 665                     (u_longlong_t)BP_GET_LSIZE(bp),                     \
 666                     (u_longlong_t)bp->blk_birth);                       \
 667         } else if (BP_IS_EMBEDDED(bp)) {                                \
 668                 len = func(buf + len, size - len,                       \
 669                     "EMBEDDED [L%llu %s] et=%u %s "                     \
 670                     "size=%llxL/%llxP birth=%lluL",                     \
 671                     (u_longlong_t)BP_GET_LEVEL(bp),                     \
 672                     type,                                               \
 673                     (int)BPE_GET_ETYPE(bp),                             \
 674                     compress,                                           \
 675                     (u_longlong_t)BPE_GET_LSIZE(bp),                    \
 676                     (u_longlong_t)BPE_GET_PSIZE(bp),                    \
 677                     (u_longlong_t)bp->blk_birth);                       \
 678         } else {                                                        \
 679                 for (int d = 0; d < BP_GET_NDVAS(bp); d++) {            \
 680                         const dva_t *dva = &bp->blk_dva[d];             \
 681                         if (DVA_IS_VALID(dva))                          \
 682                                 copies++;                               \
 683                         len += func(buf + len, size - len,              \
 684                             "DVA[%d]=<%llu:%llx:%llx>%c", d,            \
 685                             (u_longlong_t)DVA_GET_VDEV(dva),            \
 686                             (u_longlong_t)DVA_GET_OFFSET(dva),          \
 687                             (u_longlong_t)DVA_GET_ASIZE(dva),           \
 688                             ws);                                        \
 689                 }                                                       \
 690                 if (BP_IS_ENCRYPTED(bp)) {                              \
 691                         len += func(buf + len, size - len,              \
 692                             "salt=%llx iv=%llx:%llx%c",                 \
 693                             (u_longlong_t)bp->blk_dva[2].dva_word[0],   \
 694                             (u_longlong_t)bp->blk_dva[2].dva_word[1],   \
 695                             (u_longlong_t)BP_GET_IV2(bp),               \
 696                             ws);                                        \
 697                 }                                                       \
 698                 if (BP_IS_GANG(bp) &&                                   \
 699                     DVA_GET_ASIZE(&bp->blk_dva[2]) <=                   \
 700                     DVA_GET_ASIZE(&bp->blk_dva[1]) / 2)                 \
 701                         copies--;                                       \
 702                 len += func(buf + len, size - len,                      \
 703                     "[L%llu %s] %s %s %s %s %s %s %s%c"                 \
 704                     "size=%llxL/%llxP birth=%lluL/%lluP fill=%llu%c"    \
 705                     "cksum=%llx:%llx:%llx:%llx",                        \
 706                     (u_longlong_t)BP_GET_LEVEL(bp),                     \
 707                     type,                                               \
 708                     checksum,                                           \
 709                     compress,                                           \
 710                     crypt_type,                                         \
 711                     BP_GET_BYTEORDER(bp) == 0 ? "BE" : "LE",            \
 712                     BP_IS_GANG(bp) ? "gang" : "contiguous",             \
 713                     BP_GET_DEDUP(bp) ? "dedup" : "unique",              \
 714                     copyname[copies],                                   \
 715                     ws,                                                 \
 716                     (u_longlong_t)BP_GET_LSIZE(bp),                     \
 717                     (u_longlong_t)BP_GET_PSIZE(bp),                     \
 718                     (u_longlong_t)bp->blk_birth,                        \
 719                     (u_longlong_t)BP_PHYSICAL_BIRTH(bp),                \
 720                     (u_longlong_t)BP_GET_FILL(bp),                      \
 721                     ws,                                                 \
 722                     (u_longlong_t)bp->blk_cksum.zc_word[0],             \
 723                     (u_longlong_t)bp->blk_cksum.zc_word[1],             \
 724                     (u_longlong_t)bp->blk_cksum.zc_word[2],             \
 725                     (u_longlong_t)bp->blk_cksum.zc_word[3]);            \
 726         }                                                               \
 727         ASSERT(len < size);                                             \
 728 }
 729
 730 #define BP_GET_BUFC_TYPE(bp)                                            \
 731         (BP_IS_METADATA(bp) ? ARC_BUFC_METADATA : ARC_BUFC_DATA)
 732
 733 typedef enum spa_import_type {
 734         SPA_IMPORT_EXISTING,
 735         SPA_IMPORT_ASSEMBLE
 736 } spa_import_type_t;
 737
 738 /* state manipulation functions */
 739 extern int spa_open(const char *pool, spa_t **, void *tag);
 740 extern int spa_open_rewind(const char *pool, spa_t **, void *tag,
 741     nvlist_t *policy, nvlist_t **config);
 742 extern int spa_get_stats(const char *pool, nvlist_t **config, char *altroot,
 743     size_t buflen);
 744 extern int spa_create(const char *pool, nvlist_t *nvroot, nvlist_t *props,
 745     nvlist_t *zplprops, struct dsl_crypto_params *dcp);
 746 extern int spa_import(char *pool, nvlist_t *config, nvlist_t *props,
 747     uint64_t flags);
 748 extern nvlist_t *spa_tryimport(nvlist_t *tryconfig);
 749 extern int spa_destroy(char *pool);
 750 extern int spa_export(char *pool, nvlist_t **oldconfig, boolean_t force,
 751     boolean_t hardforce);
 752 extern int spa_reset(char *pool);
 753 extern void spa_async_request(spa_t *spa, int flag);
 754 extern void spa_async_unrequest(spa_t *spa, int flag);
 755 extern void spa_async_suspend(spa_t *spa);
 756 extern void spa_async_resume(spa_t *spa);
 757 extern spa_t *spa_inject_addref(char *pool);
 758 extern void spa_inject_delref(spa_t *spa);
 759 extern void spa_scan_stat_init(spa_t *spa);
 760 extern int spa_scan_get_stats(spa_t *spa, pool_scan_stat_t *ps);
 761
 762 #define SPA_ASYNC_CONFIG_UPDATE 0x01
 763 #define SPA_ASYNC_REMOVE        0x02
 764 #define SPA_ASYNC_PROBE         0x04
 765 #define SPA_ASYNC_RESILVER_DONE 0x08
 766 #define SPA_ASYNC_RESILVER      0x10
 767 #define SPA_ASYNC_AUTOEXPAND    0x20
 768 #define SPA_ASYNC_REMOVE_DONE   0x40
 769 #define SPA_ASYNC_REMOVE_STOP   0x80
 770
 771 /*
 772  * Controls the behavior of spa_vdev_remove().
 773  */
 774 #define SPA_REMOVE_UNSPARE      0x01
 775 #define SPA_REMOVE_DONE         0x02
 776
 777 /* device manipulation */
 778 extern int spa_vdev_add(spa_t *spa, nvlist_t *nvroot);
 779 extern int spa_vdev_attach(spa_t *spa, uint64_t guid, nvlist_t *nvroot,
 780     int replacing);
 781 extern int spa_vdev_detach(spa_t *spa, uint64_t guid, uint64_t pguid,
 782     int replace_done);
 783 extern int spa_vdev_remove(spa_t *spa, uint64_t guid, boolean_t unspare);
 784 extern boolean_t spa_vdev_remove_active(spa_t *spa);
 785 extern int spa_vdev_setpath(spa_t *spa, uint64_t guid, const char *newpath);
 786 extern int spa_vdev_setfru(spa_t *spa, uint64_t guid, const char *newfru);
 787 extern int spa_vdev_split_mirror(spa_t *spa, char *newname, nvlist_t *config,
 788     nvlist_t *props, boolean_t exp);
 789
 790 /* spare state (which is global across all pools) */
 791 extern void spa_spare_add(vdev_t *vd);
 792 extern void spa_spare_remove(vdev_t *vd);
 793 extern boolean_t spa_spare_exists(uint64_t guid, uint64_t *pool, int *refcnt);
 794 extern void spa_spare_activate(vdev_t *vd);
 795
 796 /* L2ARC state (which is global across all pools) */
 797 extern void spa_l2cache_add(vdev_t *vd);
 798 extern void spa_l2cache_remove(vdev_t *vd);
 799 extern boolean_t spa_l2cache_exists(uint64_t guid, uint64_t *pool);
 800 extern void spa_l2cache_activate(vdev_t *vd);
 801 extern void spa_l2cache_drop(spa_t *spa);
 802
 803 /* scanning */
 804 extern int spa_scan(spa_t *spa, pool_scan_func_t func);
 805 extern int spa_scan_stop(spa_t *spa);
 806 extern int spa_scrub_pause_resume(spa_t *spa, pool_scrub_cmd_t flag);
 807
 808 /* spa syncing */
 809 extern void spa_sync(spa_t *spa, uint64_t txg); /* only for DMU use */
 810 extern void spa_sync_allpools(void);
 811
 812 extern int zfs_sync_pass_deferred_free;
 813
 814 /* spa namespace global mutex */
 815 extern kmutex_t spa_namespace_lock;
 816
 817 /*
 818  * SPA configuration functions in spa_config.c
 819  */
 820
 821 #define SPA_CONFIG_UPDATE_POOL  0
 822 #define SPA_CONFIG_UPDATE_VDEVS 1
 823
 824 extern void spa_write_cachefile(spa_t *, boolean_t, boolean_t);
 825 extern void spa_config_load(void);
 826 extern nvlist_t *spa_all_configs(uint64_t *);
 827 extern void spa_config_set(spa_t *spa, nvlist_t *config);
 828 extern nvlist_t *spa_config_generate(spa_t *spa, vdev_t *vd, uint64_t txg,
 829     int getstats);
 830 extern void spa_config_update(spa_t *spa, int what);
 831
 832 /*
 833  * Miscellaneous SPA routines in spa_misc.c
 834  */
 835
 836 /* Namespace manipulation */
 837 extern spa_t *spa_lookup(const char *name);
 838 extern spa_t *spa_add(const char *name, nvlist_t *config, const char *altroot);
 839 extern void spa_remove(spa_t *spa);
 840 extern spa_t *spa_next(spa_t *prev);
 841
 842 /* Refcount functions */
 843 extern void spa_open_ref(spa_t *spa, void *tag);
 844 extern void spa_close(spa_t *spa, void *tag);
 845 extern void spa_async_close(spa_t *spa, void *tag);
 846 extern boolean_t spa_refcount_zero(spa_t *spa);
 847
 848 #define SCL_NONE        0x00
 849 #define SCL_CONFIG      0x01
 850 #define SCL_STATE       0x02
 851 #define SCL_L2ARC       0x04            /* hack until L2ARC 2.0 */
 852 #define SCL_ALLOC       0x08
 853 #define SCL_ZIO         0x10
 854 #define SCL_FREE        0x20
 855 #define SCL_VDEV        0x40
 856 #define SCL_LOCKS       7
 857 #define SCL_ALL         ((1 << SCL_LOCKS) - 1)
 858 #define SCL_STATE_ALL   (SCL_STATE | SCL_L2ARC | SCL_ZIO)
 859
 860 /* Historical pool statistics */
 861 typedef struct spa_stats_history {
 862         kmutex_t                lock;
 863         uint64_t                count;
 864         uint64_t                size;
 865         kstat_t                 *kstat;
 866         void                    *private;
 867         list_t                  list;
 868 } spa_stats_history_t;
 869
 870 typedef struct spa_stats {
 871         spa_stats_history_t     read_history;
 872         spa_stats_history_t     txg_history;
 873         spa_stats_history_t     tx_assign_histogram;
 874         spa_stats_history_t     io_history;
 875         spa_stats_history_t     mmp_history;
 876 } spa_stats_t;
 877
 878 typedef enum txg_state {
 879         TXG_STATE_BIRTH         = 0,
 880         TXG_STATE_OPEN          = 1,
 881         TXG_STATE_QUIESCED      = 2,
 882         TXG_STATE_WAIT_FOR_SYNC = 3,
 883         TXG_STATE_SYNCED        = 4,
 884         TXG_STATE_COMMITTED     = 5,
 885 } txg_state_t;
 886
 887 typedef struct txg_stat {
 888         vdev_stat_t             vs1;
 889         vdev_stat_t             vs2;
 890         uint64_t                txg;
 891         uint64_t                ndirty;
 892 } txg_stat_t;
 893
 894 extern void spa_stats_init(spa_t *spa);
 895 extern void spa_stats_destroy(spa_t *spa);
 896 extern void spa_read_history_add(spa_t *spa, const zbookmark_phys_t *zb,
 897     uint32_t aflags);
 898 extern void spa_txg_history_add(spa_t *spa, uint64_t txg, hrtime_t birth_time);
 899 extern int spa_txg_history_set(spa_t *spa,  uint64_t txg,
 900     txg_state_t completed_state, hrtime_t completed_time);
 901 extern txg_stat_t *spa_txg_history_init_io(spa_t *, uint64_t,
 902     struct dsl_pool *);
 903 extern void spa_txg_history_fini_io(spa_t *, txg_stat_t *);
 904 extern void spa_tx_assign_add_nsecs(spa_t *spa, uint64_t nsecs);
 905 extern int spa_mmp_history_set_skip(spa_t *spa, uint64_t mmp_kstat_id);
 906 extern int spa_mmp_history_set(spa_t *spa, uint64_t mmp_kstat_id, int io_error,
 907     hrtime_t duration);
 908 extern void *spa_mmp_history_add(spa_t *spa, uint64_t txg, uint64_t timestamp,
 909     uint64_t mmp_delay, vdev_t *vd, int label, uint64_t mmp_kstat_id,
 910     int error);
 911
 912 /* Pool configuration locks */
 913 extern int spa_config_tryenter(spa_t *spa, int locks, void *tag, krw_t rw);
 914 extern void spa_config_enter(spa_t *spa, int locks, void *tag, krw_t rw);
 915 extern void spa_config_exit(spa_t *spa, int locks, void *tag);
 916 extern int spa_config_held(spa_t *spa, int locks, krw_t rw);
 917
 918 /* Pool vdev add/remove lock */
 919 extern uint64_t spa_vdev_enter(spa_t *spa);
 920 extern uint64_t spa_vdev_config_enter(spa_t *spa);
 921 extern void spa_vdev_config_exit(spa_t *spa, vdev_t *vd, uint64_t txg,
 922     int error, char *tag);
 923 extern int spa_vdev_exit(spa_t *spa, vdev_t *vd, uint64_t txg, int error);
 924
 925 /* Pool vdev state change lock */
 926 extern void spa_vdev_state_enter(spa_t *spa, int oplock);
 927 extern int spa_vdev_state_exit(spa_t *spa, vdev_t *vd, int error);
 928
 929 /* Log state */
 930 typedef enum spa_log_state {
 931         SPA_LOG_UNKNOWN = 0,    /* unknown log state */
 932         SPA_LOG_MISSING,        /* missing log(s) */
 933         SPA_LOG_CLEAR,          /* clear the log(s) */
 934         SPA_LOG_GOOD,           /* log(s) are good */
 935 } spa_log_state_t;
 936
 937 extern spa_log_state_t spa_get_log_state(spa_t *spa);
 938 extern void spa_set_log_state(spa_t *spa, spa_log_state_t state);
 939 extern int spa_reset_logs(spa_t *spa);
 940
 941 /* Log claim callback */
 942 extern void spa_claim_notify(zio_t *zio);
 943 extern void spa_deadman(void *);
 944
 945 /* Accessor functions */
 946 extern boolean_t spa_shutting_down(spa_t *spa);
 947 extern struct dsl_pool *spa_get_dsl(spa_t *spa);
 948 extern boolean_t spa_is_initializing(spa_t *spa);
 949 extern boolean_t spa_indirect_vdevs_loaded(spa_t *spa);
 950 extern blkptr_t *spa_get_rootblkptr(spa_t *spa);
 951 extern void spa_set_rootblkptr(spa_t *spa, const blkptr_t *bp);
 952 extern void spa_altroot(spa_t *, char *, size_t);
 953 extern int spa_sync_pass(spa_t *spa);
 954 extern char *spa_name(spa_t *spa);
 955 extern uint64_t spa_guid(spa_t *spa);
 956 extern uint64_t spa_load_guid(spa_t *spa);
 957 extern uint64_t spa_last_synced_txg(spa_t *spa);
 958 extern uint64_t spa_first_txg(spa_t *spa);
 959 extern uint64_t spa_syncing_txg(spa_t *spa);
 960 extern uint64_t spa_final_dirty_txg(spa_t *spa);
 961 extern uint64_t spa_version(spa_t *spa);
 962 extern pool_state_t spa_state(spa_t *spa);
 963 extern spa_load_state_t spa_load_state(spa_t *spa);
 964 extern uint64_t spa_freeze_txg(spa_t *spa);
 965 extern uint64_t spa_get_worst_case_asize(spa_t *spa, uint64_t lsize);
 966 extern uint64_t spa_get_dspace(spa_t *spa);
 967 extern uint64_t spa_get_slop_space(spa_t *spa);
 968 extern void spa_update_dspace(spa_t *spa);
 969 extern uint64_t spa_version(spa_t *spa);
 970 extern boolean_t spa_deflate(spa_t *spa);
 971 extern metaslab_class_t *spa_normal_class(spa_t *spa);
 972 extern metaslab_class_t *spa_log_class(spa_t *spa);
 973 extern void spa_evicting_os_register(spa_t *, objset_t *os);
 974 extern void spa_evicting_os_deregister(spa_t *, objset_t *os);
 975 extern void spa_evicting_os_wait(spa_t *spa);
 976 extern int spa_max_replication(spa_t *spa);
 977 extern int spa_prev_software_version(spa_t *spa);
 978 extern uint64_t spa_get_failmode(spa_t *spa);
 979 extern uint64_t spa_get_deadman_failmode(spa_t *spa);
 980 extern void spa_set_deadman_failmode(spa_t *spa, const char *failmode);
 981 extern boolean_t spa_suspended(spa_t *spa);
 982 extern uint64_t spa_bootfs(spa_t *spa);
 983 extern uint64_t spa_delegation(spa_t *spa);
 984 extern objset_t *spa_meta_objset(spa_t *spa);
 985 extern uint64_t spa_deadman_synctime(spa_t *spa);
 986 extern uint64_t spa_deadman_ziotime(spa_t *spa);
 987
 988 /* Miscellaneous support routines */
 989 extern void spa_load_failed(spa_t *spa, const char *fmt, ...);
 990 extern void spa_load_note(spa_t *spa, const char *fmt, ...);
 991 extern void spa_activate_mos_feature(spa_t *spa, const char *feature,
 992     dmu_tx_t *tx);
 993 extern void spa_deactivate_mos_feature(spa_t *spa, const char *feature);
 994 extern int spa_rename(const char *oldname, const char *newname);
 995 extern spa_t *spa_by_guid(uint64_t pool_guid, uint64_t device_guid);
 996 extern boolean_t spa_guid_exists(uint64_t pool_guid, uint64_t device_guid);
 997 extern char *spa_strdup(const char *);
 998 extern void spa_strfree(char *);
 999 extern uint64_t spa_get_random(uint64_t range);
1000 extern uint64_t spa_generate_guid(spa_t *spa);
1001 extern void snprintf_blkptr(char *buf, size_t buflen, const blkptr_t *bp);
1002 extern void spa_freeze(spa_t *spa);
1003 extern int spa_change_guid(spa_t *spa);
1004 extern void spa_upgrade(spa_t *spa, uint64_t version);
1005 extern void spa_evict_all(void);
1006 extern vdev_t *spa_lookup_by_guid(spa_t *spa, uint64_t guid,
1007     boolean_t l2cache);
1008 extern boolean_t spa_has_spare(spa_t *, uint64_t guid);
1009 extern uint64_t dva_get_dsize_sync(spa_t *spa, const dva_t *dva);
1010 extern uint64_t bp_get_dsize_sync(spa_t *spa, const blkptr_t *bp);
1011 extern uint64_t bp_get_dsize(spa_t *spa, const blkptr_t *bp);
1012 extern boolean_t spa_has_slogs(spa_t *spa);
1013 extern boolean_t spa_is_root(spa_t *spa);
1014 extern boolean_t spa_writeable(spa_t *spa);
1015 extern boolean_t spa_has_pending_synctask(spa_t *spa);
1016 extern int spa_maxblocksize(spa_t *spa);
1017 extern int spa_maxdnodesize(spa_t *spa);
1018 extern void zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp);
1019 extern boolean_t zfs_dva_valid(spa_t *spa, const dva_t *dva,
1020     const blkptr_t *bp);
1021 typedef void (*spa_remap_cb_t)(uint64_t vdev, uint64_t offset, uint64_t size,
1022     void *arg);
1023 extern boolean_t spa_remap_blkptr(spa_t *spa, blkptr_t *bp,
1024     spa_remap_cb_t callback, void *arg);
1025 extern uint64_t spa_get_last_removal_txg(spa_t *spa);
1026 extern boolean_t spa_trust_config(spa_t *spa);
1027 extern uint64_t spa_missing_tvds_allowed(spa_t *spa);
1028 extern void spa_set_missing_tvds(spa_t *spa, uint64_t missing);
1029 extern boolean_t spa_multihost(spa_t *spa);
1030 extern unsigned long spa_get_hostid(void);
1031
1032 extern int spa_mode(spa_t *spa);
1033 extern uint64_t zfs_strtonum(const char *str, char **nptr);
1034
1035 extern char *spa_his_ievent_table[];
1036
1037 extern void spa_history_create_obj(spa_t *spa, dmu_tx_t *tx);
1038 extern int spa_history_get(spa_t *spa, uint64_t *offset, uint64_t *len_read,
1039     char *his_buf);
1040 extern int spa_history_log(spa_t *spa, const char *his_buf);
1041 extern int spa_history_log_nvl(spa_t *spa, nvlist_t *nvl);
1042 extern void spa_history_log_version(spa_t *spa, const char *operation,
1043     dmu_tx_t *tx);
1044 extern void spa_history_log_internal(spa_t *spa, const char *operation,
1045     dmu_tx_t *tx, const char *fmt, ...);
1046 extern void spa_history_log_internal_ds(struct dsl_dataset *ds, const char *op,
1047     dmu_tx_t *tx, const char *fmt, ...);
1048 extern void spa_history_log_internal_dd(dsl_dir_t *dd, const char *operation,
1049     dmu_tx_t *tx, const char *fmt, ...);
1050
1051 /* error handling */
1052 struct zbookmark_phys;
1053 extern void spa_log_error(spa_t *spa, const zbookmark_phys_t *zb);
1054 extern void zfs_ereport_post(const char *class, spa_t *spa, vdev_t *vd,
1055     const zbookmark_phys_t *zb, zio_t *zio, uint64_t stateoroffset,
1056     uint64_t length);
1057 extern nvlist_t *zfs_event_create(spa_t *spa, vdev_t *vd, const char *type,
1058     const char *name, nvlist_t *aux);
1059 extern void zfs_post_remove(spa_t *spa, vdev_t *vd);
1060 extern void zfs_post_state_change(spa_t *spa, vdev_t *vd, uint64_t laststate);
1061 extern void zfs_post_autoreplace(spa_t *spa, vdev_t *vd);
1062 extern uint64_t spa_get_errlog_size(spa_t *spa);
1063 extern int spa_get_errlog(spa_t *spa, void *uaddr, size_t *count);
1064 extern void spa_errlog_rotate(spa_t *spa);
1065 extern void spa_errlog_drain(spa_t *spa);
1066 extern void spa_errlog_sync(spa_t *spa, uint64_t txg);
1067 extern void spa_get_errlists(spa_t *spa, avl_tree_t *last, avl_tree_t *scrub);
1068
1069 /* vdev cache */
1070 extern void vdev_cache_stat_init(void);
1071 extern void vdev_cache_stat_fini(void);
1072
1073 /* vdev mirror */
1074 extern void vdev_mirror_stat_init(void);
1075 extern void vdev_mirror_stat_fini(void);
1076
1077 /* Initialization and termination */
1078 extern void spa_init(int flags);
1079 extern void spa_fini(void);
1080 extern void spa_boot_init(void);
1081
1082 /* properties */
1083 extern int spa_prop_set(spa_t *spa, nvlist_t *nvp);
1084 extern int spa_prop_get(spa_t *spa, nvlist_t **nvp);
1085 extern void spa_prop_clear_bootfs(spa_t *spa, uint64_t obj, dmu_tx_t *tx);
1086 extern void spa_configfile_set(spa_t *, nvlist_t *, boolean_t);
1087
1088 /* asynchronous event notification */
1089 extern void spa_event_notify(spa_t *spa, vdev_t *vdev, nvlist_t *hist_nvl,
1090     const char *name);
1091
1092 #ifdef ZFS_DEBUG
1093 #define dprintf_bp(bp, fmt, ...) do {                           \
1094         if (zfs_flags & ZFS_DEBUG_DPRINTF) {                    \
1095         char *__blkbuf = kmem_alloc(BP_SPRINTF_LEN, KM_SLEEP);  \
1096         snprintf_blkptr(__blkbuf, BP_SPRINTF_LEN, (bp));        \
1097         dprintf(fmt " %s\n", __VA_ARGS__, __blkbuf);            \
1098         kmem_free(__blkbuf, BP_SPRINTF_LEN);                    \
1099         } \
1100 _NOTE(CONSTCOND) } while (0)
1101 #else
1102 #define dprintf_bp(bp, fmt, ...)
1103 #endif
1104
1105 extern int spa_mode_global;                     /* mode, e.g. FREAD | FWRITE */
1106 extern int zfs_deadman_enabled;
1107 extern unsigned long zfs_deadman_synctime_ms;
1108 extern unsigned long zfs_deadman_ziotime_ms;
1109 extern unsigned long zfs_deadman_checktime_ms;
1110
1111 #ifdef  __cplusplus
1112 }
1113 #endif
1114
1115 #endif  /* _SYS_SPA_H */