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, 2018 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  * Copyright (c) 2017, Intel Corporation.
  31  */
  32
  33 #ifndef _SYS_SPA_H
  34 #define _SYS_SPA_H
  35
  36 #include <sys/avl.h>
  37 #include <sys/zfs_context.h>
  38 #include <sys/kstat.h>
  39 #include <sys/nvpair.h>
  40 #include <sys/sysmacros.h>
  41 #include <sys/types.h>
  42 #include <sys/fs/zfs.h>
  43 #include <sys/spa_checksum.h>
  44 #include <sys/dmu.h>
  45
  46 #ifdef  __cplusplus
  47 extern "C" {
  48 #endif
  49
  50 /*
  51  * Forward references that lots of things need.
  52  */
  53 typedef struct spa spa_t;
  54 typedef struct vdev vdev_t;
  55 typedef struct metaslab metaslab_t;
  56 typedef struct metaslab_group metaslab_group_t;
  57 typedef struct metaslab_class metaslab_class_t;
  58 typedef struct zio zio_t;
  59 typedef struct zilog zilog_t;
  60 typedef struct spa_aux_vdev spa_aux_vdev_t;
  61 typedef struct ddt ddt_t;
  62 typedef struct ddt_entry ddt_entry_t;
  63 typedef struct zbookmark_phys zbookmark_phys_t;
  64
  65 struct dsl_pool;
  66 struct dsl_dataset;
  67 struct dsl_crypto_params;
  68
  69 /*
  70  * General-purpose 32-bit and 64-bit bitfield encodings.
  71  */
  72 #define BF32_DECODE(x, low, len)        P2PHASE((x) >> (low), 1U << (len))
  73 #define BF64_DECODE(x, low, len)        P2PHASE((x) >> (low), 1ULL << (len))
  74 #define BF32_ENCODE(x, low, len)        (P2PHASE((x), 1U << (len)) << (low))
  75 #define BF64_ENCODE(x, low, len)        (P2PHASE((x), 1ULL << (len)) << (low))
  76
  77 #define BF32_GET(x, low, len)           BF32_DECODE(x, low, len)
  78 #define BF64_GET(x, low, len)           BF64_DECODE(x, low, len)
  79
  80 #define BF32_SET(x, low, len, val) do { \
  81         ASSERT3U(val, <, 1U << (len)); \
  82         ASSERT3U(low + len, <=, 32); \
  83         (x) ^= BF32_ENCODE((x >> low) ^ (val), low, len); \
  84 _NOTE(CONSTCOND) } while (0)
  85
  86 #define BF64_SET(x, low, len, val) do { \
  87         ASSERT3U(val, <, 1ULL << (len)); \
  88         ASSERT3U(low + len, <=, 64); \
  89         ((x) ^= BF64_ENCODE((x >> low) ^ (val), low, len)); \
  90 _NOTE(CONSTCOND) } while (0)
  91
  92 #define BF32_GET_SB(x, low, len, shift, bias)   \
  93         ((BF32_GET(x, low, len) + (bias)) << (shift))
  94 #define BF64_GET_SB(x, low, len, shift, bias)   \
  95         ((BF64_GET(x, low, len) + (bias)) << (shift))
  96
  97 #define BF32_SET_SB(x, low, len, shift, bias, val) do { \
  98         ASSERT(IS_P2ALIGNED(val, 1U << shift)); \
  99         ASSERT3S((val) >> (shift), >=, bias); \
 100         BF32_SET(x, low, len, ((val) >> (shift)) - (bias)); \
 101 _NOTE(CONSTCOND) } while (0)
 102 #define BF64_SET_SB(x, low, len, shift, bias, val) do { \
 103         ASSERT(IS_P2ALIGNED(val, 1ULL << shift)); \
 104         ASSERT3S((val) >> (shift), >=, bias); \
 105         BF64_SET(x, low, len, ((val) >> (shift)) - (bias)); \
 106 _NOTE(CONSTCOND) } while (0)
 107
 108 /*
 109  * We currently support block sizes from 512 bytes to 16MB.
 110  * The benefits of larger blocks, and thus larger IO, need to be weighed
 111  * against the cost of COWing a giant block to modify one byte, and the
 112  * large latency of reading or writing a large block.
 113  *
 114  * Note that although blocks up to 16MB are supported, the recordsize
 115  * property can not be set larger than zfs_max_recordsize (default 1MB).
 116  * See the comment near zfs_max_recordsize in dsl_dataset.c for details.
 117  *
 118  * Note that although the LSIZE field of the blkptr_t can store sizes up
 119  * to 32MB, the dnode's dn_datablkszsec can only store sizes up to
 120  * 32MB - 512 bytes.  Therefore, we limit SPA_MAXBLOCKSIZE to 16MB.
 121  */
 122 #define SPA_MINBLOCKSHIFT       9
 123 #define SPA_OLD_MAXBLOCKSHIFT   17
 124 #define SPA_MAXBLOCKSHIFT       24
 125 #define SPA_MINBLOCKSIZE        (1ULL << SPA_MINBLOCKSHIFT)
 126 #define SPA_OLD_MAXBLOCKSIZE    (1ULL << SPA_OLD_MAXBLOCKSHIFT)
 127 #define SPA_MAXBLOCKSIZE        (1ULL << SPA_MAXBLOCKSHIFT)
 128
 129 /*
 130  * Alignment Shift (ashift) is an immutable, internal top-level vdev property
 131  * which can only be set at vdev creation time. Physical writes are always done
 132  * according to it, which makes 2^ashift the smallest possible IO on a vdev.
 133  *
 134  * We currently allow values ranging from 512 bytes (2^9 = 512) to 64 KiB
 135  * (2^16 = 65,536).
 136  */
 137 #define ASHIFT_MIN              9
 138 #define ASHIFT_MAX              16
 139
 140 /*
 141  * Size of block to hold the configuration data (a packed nvlist)
 142  */
 143 #define SPA_CONFIG_BLOCKSIZE    (1ULL << 14)
 144
 145 /*
 146  * The DVA size encodings for LSIZE and PSIZE support blocks up to 32MB.
 147  * The ASIZE encoding should be at least 64 times larger (6 more bits)
 148  * to support up to 4-way RAID-Z mirror mode with worst-case gang block
 149  * overhead, three DVAs per bp, plus one more bit in case we do anything
 150  * else that expands the ASIZE.
 151  */
 152 #define SPA_LSIZEBITS           16      /* LSIZE up to 32M (2^16 * 512) */
 153 #define SPA_PSIZEBITS           16      /* PSIZE up to 32M (2^16 * 512) */
 154 #define SPA_ASIZEBITS           24      /* ASIZE up to 64 times larger  */
 155
 156 #define SPA_COMPRESSBITS        7
 157 #define SPA_VDEVBITS            24
 158
 159 /*
 160  * All SPA data is represented by 128-bit data virtual addresses (DVAs).
 161  * The members of the dva_t should be considered opaque outside the SPA.
 162  */
 163 typedef struct dva {
 164         uint64_t        dva_word[2];
 165 } dva_t;
 166
 167
 168 /*
 169  * Some checksums/hashes need a 256-bit initialization salt. This salt is kept
 170  * secret and is suitable for use in MAC algorithms as the key.
 171  */
 172 typedef struct zio_cksum_salt {
 173         uint8_t         zcs_bytes[32];
 174 } zio_cksum_salt_t;
 175
 176 /*
 177  * Each block is described by its DVAs, time of birth, checksum, etc.
 178  * The word-by-word, bit-by-bit layout of the blkptr is as follows:
 179  *
 180  *      64      56      48      40      32      24      16      8       0
 181  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 182  * 0    |  pad  |         vdev1         | GRID  |         ASIZE         |
 183  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 184  * 1    |G|                      offset1                                |
 185  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 186  * 2    |  pad  |         vdev2         | GRID  |         ASIZE         |
 187  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 188  * 3    |G|                      offset2                                |
 189  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 190  * 4    |  pad  |         vdev3         | GRID  |         ASIZE         |
 191  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 192  * 5    |G|                      offset3                                |
 193  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 194  * 6    |BDX|lvl| type  | cksum |E| comp|    PSIZE      |     LSIZE     |
 195  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 196  * 7    |                       padding                                 |
 197  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 198  * 8    |                       padding                                 |
 199  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 200  * 9    |                       physical birth txg                      |
 201  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 202  * a    |                       logical birth txg                       |
 203  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 204  * b    |                       fill count                              |
 205  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 206  * c    |                       checksum[0]                             |
 207  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 208  * d    |                       checksum[1]                             |
 209  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 210  * e    |                       checksum[2]                             |
 211  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 212  * f    |                       checksum[3]                             |
 213  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 214  *
 215  * Legend:
 216  *
 217  * vdev         virtual device ID
 218  * offset       offset into virtual device
 219  * LSIZE        logical size
 220  * PSIZE        physical size (after compression)
 221  * ASIZE        allocated size (including RAID-Z parity and gang block headers)
 222  * GRID         RAID-Z layout information (reserved for future use)
 223  * cksum        checksum function
 224  * comp         compression function
 225  * G            gang block indicator
 226  * B            byteorder (endianness)
 227  * D            dedup
 228  * X            encryption
 229  * E            blkptr_t contains embedded data (see below)
 230  * lvl          level of indirection
 231  * type         DMU object type
 232  * phys birth   txg when dva[0] was written; zero if same as logical birth txg
 233  *              note that typically all the dva's would be written in this
 234  *              txg, but they could be different if they were moved by
 235  *              device removal.
 236  * log. birth   transaction group in which the block was logically born
 237  * fill count   number of non-zero blocks under this bp
 238  * checksum[4]  256-bit checksum of the data this bp describes
 239  */
 240
 241 /*
 242  * The blkptr_t's of encrypted blocks also need to store the encryption
 243  * parameters so that the block can be decrypted. This layout is as follows:
 244  *
 245  *      64      56      48      40      32      24      16      8       0
 246  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 247  * 0    |               vdev1           | GRID  |         ASIZE         |
 248  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 249  * 1    |G|                      offset1                                |
 250  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 251  * 2    |               vdev2           | GRID  |         ASIZE         |
 252  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 253  * 3    |G|                      offset2                                |
 254  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 255  * 4    |                       salt                                    |
 256  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 257  * 5    |                       IV1                                     |
 258  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 259  * 6    |BDX|lvl| type  | cksum |E| comp|    PSIZE      |     LSIZE     |
 260  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 261  * 7    |                       padding                                 |
 262  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 263  * 8    |                       padding                                 |
 264  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 265  * 9    |                       physical birth txg                      |
 266  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 267  * a    |                       logical birth txg                       |
 268  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 269  * b    |               IV2             |           fill count          |
 270  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 271  * c    |                       checksum[0]                             |
 272  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 273  * d    |                       checksum[1]                             |
 274  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 275  * e    |                       MAC[0]                                  |
 276  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 277  * f    |                       MAC[1]                                  |
 278  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 279  *
 280  * Legend:
 281  *
 282  * salt         Salt for generating encryption keys
 283  * IV1          First 64 bits of encryption IV
 284  * X            Block requires encryption handling (set to 1)
 285  * E            blkptr_t contains embedded data (set to 0, see below)
 286  * fill count   number of non-zero blocks under this bp (truncated to 32 bits)
 287  * IV2          Last 32 bits of encryption IV
 288  * checksum[2]  128-bit checksum of the data this bp describes
 289  * MAC[2]       128-bit message authentication code for this data
 290  *
 291  * The X bit being set indicates that this block is one of 3 types. If this is
 292  * a level 0 block with an encrypted object type, the block is encrypted
 293  * (see BP_IS_ENCRYPTED()). If this is a level 0 block with an unencrypted
 294  * object type, this block is authenticated with an HMAC (see
 295  * BP_IS_AUTHENTICATED()). Otherwise (if level > 0), this bp will use the MAC
 296  * words to store a checksum-of-MACs from the level below (see
 297  * BP_HAS_INDIRECT_MAC_CKSUM()). For convenience in the code, BP_IS_PROTECTED()
 298  * refers to both encrypted and authenticated blocks and BP_USES_CRYPT()
 299  * refers to any of these 3 kinds of blocks.
 300  *
 301  * The additional encryption parameters are the salt, IV, and MAC which are
 302  * explained in greater detail in the block comment at the top of zio_crypt.c.
 303  * The MAC occupies half of the checksum space since it serves a very similar
 304  * purpose: to prevent data corruption on disk. The only functional difference
 305  * is that the checksum is used to detect on-disk corruption whether or not the
 306  * encryption key is loaded and the MAC provides additional protection against
 307  * malicious disk tampering. We use the 3rd DVA to store the salt and first
 308  * 64 bits of the IV. As a result encrypted blocks can only have 2 copies
 309  * maximum instead of the normal 3. The last 32 bits of the IV are stored in
 310  * the upper bits of what is usually the fill count. Note that only blocks at
 311  * level 0 or -2 are ever encrypted, which allows us to guarantee that these
 312  * 32 bits are not trampled over by other code (see zio_crypt.c for details).
 313  * The salt and IV are not used for authenticated bps or bps with an indirect
 314  * MAC checksum, so these blocks can utilize all 3 DVAs and the full 64 bits
 315  * for the fill count.
 316  */
 317
 318 /*
 319  * "Embedded" blkptr_t's don't actually point to a block, instead they
 320  * have a data payload embedded in the blkptr_t itself.  See the comment
 321  * in blkptr.c for more details.
 322  *
 323  * The blkptr_t is laid out as follows:
 324  *
 325  *      64      56      48      40      32      24      16      8       0
 326  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 327  * 0    |      payload                                                  |
 328  * 1    |      payload                                                  |
 329  * 2    |      payload                                                  |
 330  * 3    |      payload                                                  |
 331  * 4    |      payload                                                  |
 332  * 5    |      payload                                                  |
 333  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 334  * 6    |BDX|lvl| type  | etype |E| comp| PSIZE|              LSIZE     |
 335  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 336  * 7    |      payload                                                  |
 337  * 8    |      payload                                                  |
 338  * 9    |      payload                                                  |
 339  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 340  * a    |                       logical birth txg                       |
 341  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 342  * b    |      payload                                                  |
 343  * c    |      payload                                                  |
 344  * d    |      payload                                                  |
 345  * e    |      payload                                                  |
 346  * f    |      payload                                                  |
 347  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 348  *
 349  * Legend:
 350  *
 351  * payload              contains the embedded data
 352  * B (byteorder)        byteorder (endianness)
 353  * D (dedup)            padding (set to zero)
 354  * X                    encryption (set to zero)
 355  * E (embedded)         set to one
 356  * lvl                  indirection level
 357  * type                 DMU object type
 358  * etype                how to interpret embedded data (BP_EMBEDDED_TYPE_*)
 359  * comp                 compression function of payload
 360  * PSIZE                size of payload after compression, in bytes
 361  * LSIZE                logical size of payload, in bytes
 362  *                      note that 25 bits is enough to store the largest
 363  *                      "normal" BP's LSIZE (2^16 * 2^9) in bytes
 364  * log. birth           transaction group in which the block was logically born
 365  *
 366  * Note that LSIZE and PSIZE are stored in bytes, whereas for non-embedded
 367  * bp's they are stored in units of SPA_MINBLOCKSHIFT.
 368  * Generally, the generic BP_GET_*() macros can be used on embedded BP's.
 369  * The B, D, X, lvl, type, and comp fields are stored the same as with normal
 370  * BP's so the BP_SET_* macros can be used with them.  etype, PSIZE, LSIZE must
 371  * be set with the BPE_SET_* macros.  BP_SET_EMBEDDED() should be called before
 372  * other macros, as they assert that they are only used on BP's of the correct
 373  * "embedded-ness". Encrypted blkptr_t's cannot be embedded because they use
 374  * the payload space for encryption parameters (see the comment above on
 375  * how encryption parameters are stored).
 376  */
 377
 378 #define BPE_GET_ETYPE(bp)       \
 379         (ASSERT(BP_IS_EMBEDDED(bp)), \
 380         BF64_GET((bp)->blk_prop, 40, 8))
 381 #define BPE_SET_ETYPE(bp, t)    do { \
 382         ASSERT(BP_IS_EMBEDDED(bp)); \
 383         BF64_SET((bp)->blk_prop, 40, 8, t); \
 384 _NOTE(CONSTCOND) } while (0)
 385
 386 #define BPE_GET_LSIZE(bp)       \
 387         (ASSERT(BP_IS_EMBEDDED(bp)), \
 388         BF64_GET_SB((bp)->blk_prop, 0, 25, 0, 1))
 389 #define BPE_SET_LSIZE(bp, x)    do { \
 390         ASSERT(BP_IS_EMBEDDED(bp)); \
 391         BF64_SET_SB((bp)->blk_prop, 0, 25, 0, 1, x); \
 392 _NOTE(CONSTCOND) } while (0)
 393
 394 #define BPE_GET_PSIZE(bp)       \
 395         (ASSERT(BP_IS_EMBEDDED(bp)), \
 396         BF64_GET_SB((bp)->blk_prop, 25, 7, 0, 1))
 397 #define BPE_SET_PSIZE(bp, x)    do { \
 398         ASSERT(BP_IS_EMBEDDED(bp)); \
 399         BF64_SET_SB((bp)->blk_prop, 25, 7, 0, 1, x); \
 400 _NOTE(CONSTCOND) } while (0)
 401
 402 typedef enum bp_embedded_type {
 403         BP_EMBEDDED_TYPE_DATA,
 404         BP_EMBEDDED_TYPE_RESERVED, /* Reserved for an unintegrated feature. */
 405         NUM_BP_EMBEDDED_TYPES = BP_EMBEDDED_TYPE_RESERVED
 406 } bp_embedded_type_t;
 407
 408 #define BPE_NUM_WORDS 14
 409 #define BPE_PAYLOAD_SIZE (BPE_NUM_WORDS * sizeof (uint64_t))
 410 #define BPE_IS_PAYLOADWORD(bp, wp) \
 411         ((wp) != &(bp)->blk_prop && (wp) != &(bp)->blk_birth)
 412
 413 #define SPA_BLKPTRSHIFT 7               /* blkptr_t is 128 bytes        */
 414 #define SPA_DVAS_PER_BP 3               /* Number of DVAs in a bp       */
 415 #define SPA_SYNC_MIN_VDEVS 3            /* min vdevs to update during sync */
 416
 417 /*
 418  * A block is a hole when it has either 1) never been written to, or
 419  * 2) is zero-filled. In both cases, ZFS can return all zeroes for all reads
 420  * without physically allocating disk space. Holes are represented in the
 421  * blkptr_t structure by zeroed blk_dva. Correct checking for holes is
 422  * done through the BP_IS_HOLE macro. For holes, the logical size, level,
 423  * DMU object type, and birth times are all also stored for holes that
 424  * were written to at some point (i.e. were punched after having been filled).
 425  */
 426 typedef struct blkptr {
 427         dva_t           blk_dva[SPA_DVAS_PER_BP]; /* Data Virtual Addresses */
 428         uint64_t        blk_prop;       /* size, compression, type, etc     */
 429         uint64_t        blk_pad[2];     /* Extra space for the future       */
 430         uint64_t        blk_phys_birth; /* txg when block was allocated     */
 431         uint64_t        blk_birth;      /* transaction group at birth       */
 432         uint64_t        blk_fill;       /* fill count                       */
 433         zio_cksum_t     blk_cksum;      /* 256-bit checksum                 */
 434 } blkptr_t;
 435
 436 /*
 437  * Macros to get and set fields in a bp or DVA.
 438  */
 439 #define DVA_GET_ASIZE(dva)      \
 440         BF64_GET_SB((dva)->dva_word[0], 0, SPA_ASIZEBITS, SPA_MINBLOCKSHIFT, 0)
 441 #define DVA_SET_ASIZE(dva, x)   \
 442         BF64_SET_SB((dva)->dva_word[0], 0, SPA_ASIZEBITS, \
 443         SPA_MINBLOCKSHIFT, 0, x)
 444
 445 #define DVA_GET_GRID(dva)       BF64_GET((dva)->dva_word[0], 24, 8)
 446 #define DVA_SET_GRID(dva, x)    BF64_SET((dva)->dva_word[0], 24, 8, x)
 447
 448 #define DVA_GET_VDEV(dva)       BF64_GET((dva)->dva_word[0], 32, SPA_VDEVBITS)
 449 #define DVA_SET_VDEV(dva, x)    \
 450         BF64_SET((dva)->dva_word[0], 32, SPA_VDEVBITS, x)
 451
 452 #define DVA_GET_OFFSET(dva)     \
 453         BF64_GET_SB((dva)->dva_word[1], 0, 63, SPA_MINBLOCKSHIFT, 0)
 454 #define DVA_SET_OFFSET(dva, x)  \
 455         BF64_SET_SB((dva)->dva_word[1], 0, 63, SPA_MINBLOCKSHIFT, 0, x)
 456
 457 #define DVA_GET_GANG(dva)       BF64_GET((dva)->dva_word[1], 63, 1)
 458 #define DVA_SET_GANG(dva, x)    BF64_SET((dva)->dva_word[1], 63, 1, x)
 459
 460 #define BP_GET_LSIZE(bp)        \
 461         (BP_IS_EMBEDDED(bp) ?   \
 462         (BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA ? BPE_GET_LSIZE(bp) : 0): \
 463         BF64_GET_SB((bp)->blk_prop, 0, SPA_LSIZEBITS, SPA_MINBLOCKSHIFT, 1))
 464 #define BP_SET_LSIZE(bp, x)     do { \
 465         ASSERT(!BP_IS_EMBEDDED(bp)); \
 466         BF64_SET_SB((bp)->blk_prop, \
 467             0, SPA_LSIZEBITS, SPA_MINBLOCKSHIFT, 1, x); \
 468 _NOTE(CONSTCOND) } while (0)
 469
 470 #define BP_GET_PSIZE(bp)        \
 471         (BP_IS_EMBEDDED(bp) ? 0 : \
 472         BF64_GET_SB((bp)->blk_prop, 16, SPA_PSIZEBITS, SPA_MINBLOCKSHIFT, 1))
 473 #define BP_SET_PSIZE(bp, x)     do { \
 474         ASSERT(!BP_IS_EMBEDDED(bp)); \
 475         BF64_SET_SB((bp)->blk_prop, \
 476             16, SPA_PSIZEBITS, SPA_MINBLOCKSHIFT, 1, x); \
 477 _NOTE(CONSTCOND) } while (0)
 478
 479 #define BP_GET_COMPRESS(bp)             \
 480         BF64_GET((bp)->blk_prop, 32, SPA_COMPRESSBITS)
 481 #define BP_SET_COMPRESS(bp, x)          \
 482         BF64_SET((bp)->blk_prop, 32, SPA_COMPRESSBITS, x)
 483
 484 #define BP_IS_EMBEDDED(bp)              BF64_GET((bp)->blk_prop, 39, 1)
 485 #define BP_SET_EMBEDDED(bp, x)          BF64_SET((bp)->blk_prop, 39, 1, x)
 486
 487 #define BP_GET_CHECKSUM(bp)             \
 488         (BP_IS_EMBEDDED(bp) ? ZIO_CHECKSUM_OFF : \
 489         BF64_GET((bp)->blk_prop, 40, 8))
 490 #define BP_SET_CHECKSUM(bp, x)          do { \
 491         ASSERT(!BP_IS_EMBEDDED(bp)); \
 492         BF64_SET((bp)->blk_prop, 40, 8, x); \
 493 _NOTE(CONSTCOND) } while (0)
 494
 495 #define BP_GET_TYPE(bp)                 BF64_GET((bp)->blk_prop, 48, 8)
 496 #define BP_SET_TYPE(bp, x)              BF64_SET((bp)->blk_prop, 48, 8, x)
 497
 498 #define BP_GET_LEVEL(bp)                BF64_GET((bp)->blk_prop, 56, 5)
 499 #define BP_SET_LEVEL(bp, x)             BF64_SET((bp)->blk_prop, 56, 5, x)
 500
 501 /* encrypted, authenticated, and MAC cksum bps use the same bit */
 502 #define BP_USES_CRYPT(bp)               BF64_GET((bp)->blk_prop, 61, 1)
 503 #define BP_SET_CRYPT(bp, x)             BF64_SET((bp)->blk_prop, 61, 1, x)
 504
 505 #define BP_IS_ENCRYPTED(bp)                     \
 506         (BP_USES_CRYPT(bp) &&                   \
 507         BP_GET_LEVEL(bp) <= 0 &&                \
 508         DMU_OT_IS_ENCRYPTED(BP_GET_TYPE(bp)))
 509
 510 #define BP_IS_AUTHENTICATED(bp)                 \
 511         (BP_USES_CRYPT(bp) &&                   \
 512         BP_GET_LEVEL(bp) <= 0 &&                \
 513         !DMU_OT_IS_ENCRYPTED(BP_GET_TYPE(bp)))
 514
 515 #define BP_HAS_INDIRECT_MAC_CKSUM(bp)           \
 516         (BP_USES_CRYPT(bp) && BP_GET_LEVEL(bp) > 0)
 517
 518 #define BP_IS_PROTECTED(bp)                     \
 519         (BP_IS_ENCRYPTED(bp) || BP_IS_AUTHENTICATED(bp))
 520
 521 #define BP_GET_DEDUP(bp)                BF64_GET((bp)->blk_prop, 62, 1)
 522 #define BP_SET_DEDUP(bp, x)             BF64_SET((bp)->blk_prop, 62, 1, x)
 523
 524 #define BP_GET_BYTEORDER(bp)            BF64_GET((bp)->blk_prop, 63, 1)
 525 #define BP_SET_BYTEORDER(bp, x)         BF64_SET((bp)->blk_prop, 63, 1, x)
 526
 527 #define BP_PHYSICAL_BIRTH(bp)           \
 528         (BP_IS_EMBEDDED(bp) ? 0 : \
 529         (bp)->blk_phys_birth ? (bp)->blk_phys_birth : (bp)->blk_birth)
 530
 531 #define BP_SET_BIRTH(bp, logical, physical)     \
 532 {                                               \
 533         ASSERT(!BP_IS_EMBEDDED(bp));            \
 534         (bp)->blk_birth = (logical);            \
 535         (bp)->blk_phys_birth = ((logical) == (physical) ? 0 : (physical)); \
 536 }
 537
 538 #define BP_GET_FILL(bp)                         \
 539         ((BP_IS_ENCRYPTED(bp)) ? BF64_GET((bp)->blk_fill, 0, 32) : \
 540         ((BP_IS_EMBEDDED(bp)) ? 1 : (bp)->blk_fill))
 541
 542 #define BP_SET_FILL(bp, fill)                   \
 543 {                                               \
 544         if (BP_IS_ENCRYPTED(bp))                        \
 545                 BF64_SET((bp)->blk_fill, 0, 32, fill); \
 546         else                                    \
 547                 (bp)->blk_fill = fill;          \
 548 }
 549
 550 #define BP_GET_IV2(bp)                          \
 551         (ASSERT(BP_IS_ENCRYPTED(bp)),           \
 552         BF64_GET((bp)->blk_fill, 32, 32))
 553 #define BP_SET_IV2(bp, iv2)                     \
 554 {                                               \
 555         ASSERT(BP_IS_ENCRYPTED(bp));            \
 556         BF64_SET((bp)->blk_fill, 32, 32, iv2);  \
 557 }
 558
 559 #define BP_IS_METADATA(bp)      \
 560         (BP_GET_LEVEL(bp) > 0 || DMU_OT_IS_METADATA(BP_GET_TYPE(bp)))
 561
 562 #define BP_GET_ASIZE(bp)        \
 563         (BP_IS_EMBEDDED(bp) ? 0 : \
 564         DVA_GET_ASIZE(&(bp)->blk_dva[0]) + \
 565         DVA_GET_ASIZE(&(bp)->blk_dva[1]) + \
 566         (DVA_GET_ASIZE(&(bp)->blk_dva[2]) * !BP_IS_ENCRYPTED(bp)))
 567
 568 #define BP_GET_UCSIZE(bp)       \
 569         (BP_IS_METADATA(bp) ? BP_GET_PSIZE(bp) : BP_GET_LSIZE(bp))
 570
 571 #define BP_GET_NDVAS(bp)        \
 572         (BP_IS_EMBEDDED(bp) ? 0 : \
 573         !!DVA_GET_ASIZE(&(bp)->blk_dva[0]) + \
 574         !!DVA_GET_ASIZE(&(bp)->blk_dva[1]) + \
 575         (!!DVA_GET_ASIZE(&(bp)->blk_dva[2]) * !BP_IS_ENCRYPTED(bp)))
 576
 577 #define BP_COUNT_GANG(bp)       \
 578         (BP_IS_EMBEDDED(bp) ? 0 : \
 579         (DVA_GET_GANG(&(bp)->blk_dva[0]) + \
 580         DVA_GET_GANG(&(bp)->blk_dva[1]) + \
 581         (DVA_GET_GANG(&(bp)->blk_dva[2]) * !BP_IS_ENCRYPTED(bp))))
 582
 583 #define DVA_EQUAL(dva1, dva2)   \
 584         ((dva1)->dva_word[1] == (dva2)->dva_word[1] && \
 585         (dva1)->dva_word[0] == (dva2)->dva_word[0])
 586
 587 #define BP_EQUAL(bp1, bp2)      \
 588         (BP_PHYSICAL_BIRTH(bp1) == BP_PHYSICAL_BIRTH(bp2) &&    \
 589         (bp1)->blk_birth == (bp2)->blk_birth &&                 \
 590         DVA_EQUAL(&(bp1)->blk_dva[0], &(bp2)->blk_dva[0]) &&    \
 591         DVA_EQUAL(&(bp1)->blk_dva[1], &(bp2)->blk_dva[1]) &&    \
 592         DVA_EQUAL(&(bp1)->blk_dva[2], &(bp2)->blk_dva[2]))
 593
 594
 595 #define DVA_IS_VALID(dva)       (DVA_GET_ASIZE(dva) != 0)
 596
 597 #define BP_IDENTITY(bp)         (ASSERT(!BP_IS_EMBEDDED(bp)), &(bp)->blk_dva[0])
 598 #define BP_IS_GANG(bp)          \
 599         (BP_IS_EMBEDDED(bp) ? B_FALSE : DVA_GET_GANG(BP_IDENTITY(bp)))
 600 #define DVA_IS_EMPTY(dva)       ((dva)->dva_word[0] == 0ULL &&  \
 601                                 (dva)->dva_word[1] == 0ULL)
 602 #define BP_IS_HOLE(bp) \
 603         (!BP_IS_EMBEDDED(bp) && DVA_IS_EMPTY(BP_IDENTITY(bp)))
 604
 605 /* BP_IS_RAIDZ(bp) assumes no block compression */
 606 #define BP_IS_RAIDZ(bp)         (DVA_GET_ASIZE(&(bp)->blk_dva[0]) > \
 607                                 BP_GET_PSIZE(bp))
 608
 609 #define BP_ZERO(bp)                             \
 610 {                                               \
 611         (bp)->blk_dva[0].dva_word[0] = 0;       \
 612         (bp)->blk_dva[0].dva_word[1] = 0;       \
 613         (bp)->blk_dva[1].dva_word[0] = 0;       \
 614         (bp)->blk_dva[1].dva_word[1] = 0;       \
 615         (bp)->blk_dva[2].dva_word[0] = 0;       \
 616         (bp)->blk_dva[2].dva_word[1] = 0;       \
 617         (bp)->blk_prop = 0;                     \
 618         (bp)->blk_pad[0] = 0;                   \
 619         (bp)->blk_pad[1] = 0;                   \
 620         (bp)->blk_phys_birth = 0;               \
 621         (bp)->blk_birth = 0;                    \
 622         (bp)->blk_fill = 0;                     \
 623         ZIO_SET_CHECKSUM(&(bp)->blk_cksum, 0, 0, 0, 0); \
 624 }
 625
 626 #ifdef _BIG_ENDIAN
 627 #define ZFS_HOST_BYTEORDER      (0ULL)
 628 #else
 629 #define ZFS_HOST_BYTEORDER      (1ULL)
 630 #endif
 631
 632 #define BP_SHOULD_BYTESWAP(bp)  (BP_GET_BYTEORDER(bp) != ZFS_HOST_BYTEORDER)
 633
 634 #define BP_SPRINTF_LEN  400
 635
 636 /*
 637  * This macro allows code sharing between zfs, libzpool, and mdb.
 638  * 'func' is either snprintf() or mdb_snprintf().
 639  * 'ws' (whitespace) can be ' ' for single-line format, '\n' for multi-line.
 640  */
 641 #define SNPRINTF_BLKPTR(func, ws, buf, size, bp, type, checksum, compress) \
 642 {                                                                       \
 643         static const char *copyname[] =                                 \
 644             { "zero", "single", "double", "triple" };                   \
 645         int len = 0;                                                    \
 646         int copies = 0;                                                 \
 647         const char *crypt_type;                                         \
 648         if (bp != NULL) {                                               \
 649                 if (BP_IS_ENCRYPTED(bp)) {                              \
 650                         crypt_type = "encrypted";                       \
 651                         /* LINTED E_SUSPICIOUS_COMPARISON */            \
 652                 } else if (BP_IS_AUTHENTICATED(bp)) {                   \
 653                         crypt_type = "authenticated";                   \
 654                 } else if (BP_HAS_INDIRECT_MAC_CKSUM(bp)) {             \
 655                         crypt_type = "indirect-MAC";                    \
 656                 } else {                                                \
 657                         crypt_type = "unencrypted";                     \
 658                 }                                                       \
 659         }                                                               \
 660         if (bp == NULL) {                                               \
 661                 len += func(buf + len, size - len, "<NULL>");           \
 662         } else if (BP_IS_HOLE(bp)) {                                    \
 663                 len += func(buf + len, size - len,                      \
 664                     "HOLE [L%llu %s] "                                  \
 665                     "size=%llxL birth=%lluL",                           \
 666                     (u_longlong_t)BP_GET_LEVEL(bp),                     \
 667                     type,                                               \
 668                     (u_longlong_t)BP_GET_LSIZE(bp),                     \
 669                     (u_longlong_t)bp->blk_birth);                       \
 670         } else if (BP_IS_EMBEDDED(bp)) {                                \
 671                 len = func(buf + len, size - len,                       \
 672                     "EMBEDDED [L%llu %s] et=%u %s "                     \
 673                     "size=%llxL/%llxP birth=%lluL",                     \
 674                     (u_longlong_t)BP_GET_LEVEL(bp),                     \
 675                     type,                                               \
 676                     (int)BPE_GET_ETYPE(bp),                             \
 677                     compress,                                           \
 678                     (u_longlong_t)BPE_GET_LSIZE(bp),                    \
 679                     (u_longlong_t)BPE_GET_PSIZE(bp),                    \
 680                     (u_longlong_t)bp->blk_birth);                       \
 681         } else {                                                        \
 682                 for (int d = 0; d < BP_GET_NDVAS(bp); d++) {            \
 683                         const dva_t *dva = &bp->blk_dva[d];             \
 684                         if (DVA_IS_VALID(dva))                          \
 685                                 copies++;                               \
 686                         len += func(buf + len, size - len,              \
 687                             "DVA[%d]=<%llu:%llx:%llx>%c", d,            \
 688                             (u_longlong_t)DVA_GET_VDEV(dva),            \
 689                             (u_longlong_t)DVA_GET_OFFSET(dva),          \
 690                             (u_longlong_t)DVA_GET_ASIZE(dva),           \
 691                             ws);                                        \
 692                 }                                                       \
 693                 if (BP_IS_ENCRYPTED(bp)) {                              \
 694                         len += func(buf + len, size - len,              \
 695                             "salt=%llx iv=%llx:%llx%c",                 \
 696                             (u_longlong_t)bp->blk_dva[2].dva_word[0],   \
 697                             (u_longlong_t)bp->blk_dva[2].dva_word[1],   \
 698                             (u_longlong_t)BP_GET_IV2(bp),               \
 699                             ws);                                        \
 700                 }                                                       \
 701                 if (BP_IS_GANG(bp) &&                                   \
 702                     DVA_GET_ASIZE(&bp->blk_dva[2]) <=                   \
 703                     DVA_GET_ASIZE(&bp->blk_dva[1]) / 2)                 \
 704                         copies--;                                       \
 705                 len += func(buf + len, size - len,                      \
 706                     "[L%llu %s] %s %s %s %s %s %s %s%c"                 \
 707                     "size=%llxL/%llxP birth=%lluL/%lluP fill=%llu%c"    \
 708                     "cksum=%llx:%llx:%llx:%llx",                        \
 709                     (u_longlong_t)BP_GET_LEVEL(bp),                     \
 710                     type,                                               \
 711                     checksum,                                           \
 712                     compress,                                           \
 713                     crypt_type,                                         \
 714                     BP_GET_BYTEORDER(bp) == 0 ? "BE" : "LE",            \
 715                     BP_IS_GANG(bp) ? "gang" : "contiguous",             \
 716                     BP_GET_DEDUP(bp) ? "dedup" : "unique",              \
 717                     copyname[copies],                                   \
 718                     ws,                                                 \
 719                     (u_longlong_t)BP_GET_LSIZE(bp),                     \
 720                     (u_longlong_t)BP_GET_PSIZE(bp),                     \
 721                     (u_longlong_t)bp->blk_birth,                        \
 722                     (u_longlong_t)BP_PHYSICAL_BIRTH(bp),                \
 723                     (u_longlong_t)BP_GET_FILL(bp),                      \
 724                     ws,                                                 \
 725                     (u_longlong_t)bp->blk_cksum.zc_word[0],             \
 726                     (u_longlong_t)bp->blk_cksum.zc_word[1],             \
 727                     (u_longlong_t)bp->blk_cksum.zc_word[2],             \
 728                     (u_longlong_t)bp->blk_cksum.zc_word[3]);            \
 729         }                                                               \
 730         ASSERT(len < size);                                             \
 731 }
 732
 733 #define BP_GET_BUFC_TYPE(bp)                                            \
 734         (BP_IS_METADATA(bp) ? ARC_BUFC_METADATA : ARC_BUFC_DATA)
 735
 736 typedef enum spa_import_type {
 737         SPA_IMPORT_EXISTING,
 738         SPA_IMPORT_ASSEMBLE
 739 } spa_import_type_t;
 740
 741 /* state manipulation functions */
 742 extern int spa_open(const char *pool, spa_t **, void *tag);
 743 extern int spa_open_rewind(const char *pool, spa_t **, void *tag,
 744     nvlist_t *policy, nvlist_t **config);
 745 extern int spa_get_stats(const char *pool, nvlist_t **config, char *altroot,
 746     size_t buflen);
 747 extern int spa_create(const char *pool, nvlist_t *nvroot, nvlist_t *props,
 748     nvlist_t *zplprops, struct dsl_crypto_params *dcp);
 749 extern int spa_import(char *pool, nvlist_t *config, nvlist_t *props,
 750     uint64_t flags);
 751 extern nvlist_t *spa_tryimport(nvlist_t *tryconfig);
 752 extern int spa_destroy(char *pool);
 753 extern int spa_checkpoint(const char *pool);
 754 extern int spa_checkpoint_discard(const char *pool);
 755 extern int spa_export(char *pool, nvlist_t **oldconfig, boolean_t force,
 756     boolean_t hardforce);
 757 extern int spa_reset(char *pool);
 758 extern void spa_async_request(spa_t *spa, int flag);
 759 extern void spa_async_unrequest(spa_t *spa, int flag);
 760 extern void spa_async_suspend(spa_t *spa);
 761 extern void spa_async_resume(spa_t *spa);
 762 extern spa_t *spa_inject_addref(char *pool);
 763 extern void spa_inject_delref(spa_t *spa);
 764 extern void spa_scan_stat_init(spa_t *spa);
 765 extern int spa_scan_get_stats(spa_t *spa, pool_scan_stat_t *ps);
 766
 767 #define SPA_ASYNC_CONFIG_UPDATE 0x01
 768 #define SPA_ASYNC_REMOVE        0x02
 769 #define SPA_ASYNC_PROBE         0x04
 770 #define SPA_ASYNC_RESILVER_DONE 0x08
 771 #define SPA_ASYNC_RESILVER      0x10
 772 #define SPA_ASYNC_AUTOEXPAND    0x20
 773 #define SPA_ASYNC_REMOVE_DONE   0x40
 774 #define SPA_ASYNC_REMOVE_STOP   0x80
 775
 776 /*
 777  * Controls the behavior of spa_vdev_remove().
 778  */
 779 #define SPA_REMOVE_UNSPARE      0x01
 780 #define SPA_REMOVE_DONE         0x02
 781
 782 /* device manipulation */
 783 extern int spa_vdev_add(spa_t *spa, nvlist_t *nvroot);
 784 extern int spa_vdev_attach(spa_t *spa, uint64_t guid, nvlist_t *nvroot,
 785     int replacing);
 786 extern int spa_vdev_detach(spa_t *spa, uint64_t guid, uint64_t pguid,
 787     int replace_done);
 788 extern int spa_vdev_remove(spa_t *spa, uint64_t guid, boolean_t unspare);
 789 extern boolean_t spa_vdev_remove_active(spa_t *spa);
 790 extern int spa_vdev_setpath(spa_t *spa, uint64_t guid, const char *newpath);
 791 extern int spa_vdev_setfru(spa_t *spa, uint64_t guid, const char *newfru);
 792 extern int spa_vdev_split_mirror(spa_t *spa, char *newname, nvlist_t *config,
 793     nvlist_t *props, boolean_t exp);
 794
 795 /* spare state (which is global across all pools) */
 796 extern void spa_spare_add(vdev_t *vd);
 797 extern void spa_spare_remove(vdev_t *vd);
 798 extern boolean_t spa_spare_exists(uint64_t guid, uint64_t *pool, int *refcnt);
 799 extern void spa_spare_activate(vdev_t *vd);
 800
 801 /* L2ARC state (which is global across all pools) */
 802 extern void spa_l2cache_add(vdev_t *vd);
 803 extern void spa_l2cache_remove(vdev_t *vd);
 804 extern boolean_t spa_l2cache_exists(uint64_t guid, uint64_t *pool);
 805 extern void spa_l2cache_activate(vdev_t *vd);
 806 extern void spa_l2cache_drop(spa_t *spa);
 807
 808 /* scanning */
 809 extern int spa_scan(spa_t *spa, pool_scan_func_t func);
 810 extern int spa_scan_stop(spa_t *spa);
 811 extern int spa_scrub_pause_resume(spa_t *spa, pool_scrub_cmd_t flag);
 812
 813 /* spa syncing */
 814 extern void spa_sync(spa_t *spa, uint64_t txg); /* only for DMU use */
 815 extern void spa_sync_allpools(void);
 816
 817 extern int zfs_sync_pass_deferred_free;
 818
 819 /* spa namespace global mutex */
 820 extern kmutex_t spa_namespace_lock;
 821
 822 /*
 823  * SPA configuration functions in spa_config.c
 824  */
 825
 826 #define SPA_CONFIG_UPDATE_POOL  0
 827 #define SPA_CONFIG_UPDATE_VDEVS 1
 828
 829 extern void spa_write_cachefile(spa_t *, boolean_t, boolean_t);
 830 extern void spa_config_load(void);
 831 extern nvlist_t *spa_all_configs(uint64_t *);
 832 extern void spa_config_set(spa_t *spa, nvlist_t *config);
 833 extern nvlist_t *spa_config_generate(spa_t *spa, vdev_t *vd, uint64_t txg,
 834     int getstats);
 835 extern void spa_config_update(spa_t *spa, int what);
 836
 837 /*
 838  * Miscellaneous SPA routines in spa_misc.c
 839  */
 840
 841 /* Namespace manipulation */
 842 extern spa_t *spa_lookup(const char *name);
 843 extern spa_t *spa_add(const char *name, nvlist_t *config, const char *altroot);
 844 extern void spa_remove(spa_t *spa);
 845 extern spa_t *spa_next(spa_t *prev);
 846
 847 /* Refcount functions */
 848 extern void spa_open_ref(spa_t *spa, void *tag);
 849 extern void spa_close(spa_t *spa, void *tag);
 850 extern void spa_async_close(spa_t *spa, void *tag);
 851 extern boolean_t spa_refcount_zero(spa_t *spa);
 852
 853 #define SCL_NONE        0x00
 854 #define SCL_CONFIG      0x01
 855 #define SCL_STATE       0x02
 856 #define SCL_L2ARC       0x04            /* hack until L2ARC 2.0 */
 857 #define SCL_ALLOC       0x08
 858 #define SCL_ZIO         0x10
 859 #define SCL_FREE        0x20
 860 #define SCL_VDEV        0x40
 861 #define SCL_LOCKS       7
 862 #define SCL_ALL         ((1 << SCL_LOCKS) - 1)
 863 #define SCL_STATE_ALL   (SCL_STATE | SCL_L2ARC | SCL_ZIO)
 864
 865 /* Historical pool statistics */
 866 typedef struct spa_history_kstat {
 867         kmutex_t                lock;
 868         uint64_t                count;
 869         uint64_t                size;
 870         kstat_t                 *kstat;
 871         void                    *private;
 872         list_t                  list;
 873 } spa_history_kstat_t;
 874
 875 typedef struct spa_history_list {
 876         uint64_t                size;
 877         procfs_list_t           procfs_list;
 878 } spa_history_list_t;
 879
 880 typedef struct spa_stats {
 881         spa_history_list_t      read_history;
 882         spa_history_list_t      txg_history;
 883         spa_history_kstat_t     tx_assign_histogram;
 884         spa_history_kstat_t     io_history;
 885         spa_history_list_t      mmp_history;
 886         spa_history_kstat_t     state;          /* pool state */
 887 } spa_stats_t;
 888
 889 typedef enum txg_state {
 890         TXG_STATE_BIRTH         = 0,
 891         TXG_STATE_OPEN          = 1,
 892         TXG_STATE_QUIESCED      = 2,
 893         TXG_STATE_WAIT_FOR_SYNC = 3,
 894         TXG_STATE_SYNCED        = 4,
 895         TXG_STATE_COMMITTED     = 5,
 896 } txg_state_t;
 897
 898 typedef struct txg_stat {
 899         vdev_stat_t             vs1;
 900         vdev_stat_t             vs2;
 901         uint64_t                txg;
 902         uint64_t                ndirty;
 903 } txg_stat_t;
 904
 905 extern void spa_stats_init(spa_t *spa);
 906 extern void spa_stats_destroy(spa_t *spa);
 907 extern void spa_read_history_add(spa_t *spa, const zbookmark_phys_t *zb,
 908     uint32_t aflags);
 909 extern void spa_txg_history_add(spa_t *spa, uint64_t txg, hrtime_t birth_time);
 910 extern int spa_txg_history_set(spa_t *spa,  uint64_t txg,
 911     txg_state_t completed_state, hrtime_t completed_time);
 912 extern txg_stat_t *spa_txg_history_init_io(spa_t *, uint64_t,
 913     struct dsl_pool *);
 914 extern void spa_txg_history_fini_io(spa_t *, txg_stat_t *);
 915 extern void spa_tx_assign_add_nsecs(spa_t *spa, uint64_t nsecs);
 916 extern int spa_mmp_history_set_skip(spa_t *spa, uint64_t mmp_kstat_id);
 917 extern int spa_mmp_history_set(spa_t *spa, uint64_t mmp_kstat_id, int io_error,
 918     hrtime_t duration);
 919 extern void spa_mmp_history_add(spa_t *spa, uint64_t txg, uint64_t timestamp,
 920     uint64_t mmp_delay, vdev_t *vd, int label, uint64_t mmp_kstat_id,
 921     int error);
 922
 923 /* Pool configuration locks */
 924 extern int spa_config_tryenter(spa_t *spa, int locks, void *tag, krw_t rw);
 925 extern void spa_config_enter(spa_t *spa, int locks, void *tag, krw_t rw);
 926 extern void spa_config_exit(spa_t *spa, int locks, void *tag);
 927 extern int spa_config_held(spa_t *spa, int locks, krw_t rw);
 928
 929 /* Pool vdev add/remove lock */
 930 extern uint64_t spa_vdev_enter(spa_t *spa);
 931 extern uint64_t spa_vdev_config_enter(spa_t *spa);
 932 extern void spa_vdev_config_exit(spa_t *spa, vdev_t *vd, uint64_t txg,
 933     int error, char *tag);
 934 extern int spa_vdev_exit(spa_t *spa, vdev_t *vd, uint64_t txg, int error);
 935
 936 /* Pool vdev state change lock */
 937 extern void spa_vdev_state_enter(spa_t *spa, int oplock);
 938 extern int spa_vdev_state_exit(spa_t *spa, vdev_t *vd, int error);
 939
 940 /* Log state */
 941 typedef enum spa_log_state {
 942         SPA_LOG_UNKNOWN = 0,    /* unknown log state */
 943         SPA_LOG_MISSING,        /* missing log(s) */
 944         SPA_LOG_CLEAR,          /* clear the log(s) */
 945         SPA_LOG_GOOD,           /* log(s) are good */
 946 } spa_log_state_t;
 947
 948 extern spa_log_state_t spa_get_log_state(spa_t *spa);
 949 extern void spa_set_log_state(spa_t *spa, spa_log_state_t state);
 950 extern int spa_reset_logs(spa_t *spa);
 951
 952 /* Log claim callback */
 953 extern void spa_claim_notify(zio_t *zio);
 954 extern void spa_deadman(void *);
 955
 956 /* Accessor functions */
 957 extern boolean_t spa_shutting_down(spa_t *spa);
 958 extern struct dsl_pool *spa_get_dsl(spa_t *spa);
 959 extern boolean_t spa_is_initializing(spa_t *spa);
 960 extern boolean_t spa_indirect_vdevs_loaded(spa_t *spa);
 961 extern blkptr_t *spa_get_rootblkptr(spa_t *spa);
 962 extern void spa_set_rootblkptr(spa_t *spa, const blkptr_t *bp);
 963 extern void spa_altroot(spa_t *, char *, size_t);
 964 extern int spa_sync_pass(spa_t *spa);
 965 extern char *spa_name(spa_t *spa);
 966 extern uint64_t spa_guid(spa_t *spa);
 967 extern uint64_t spa_load_guid(spa_t *spa);
 968 extern uint64_t spa_last_synced_txg(spa_t *spa);
 969 extern uint64_t spa_first_txg(spa_t *spa);
 970 extern uint64_t spa_syncing_txg(spa_t *spa);
 971 extern uint64_t spa_final_dirty_txg(spa_t *spa);
 972 extern uint64_t spa_version(spa_t *spa);
 973 extern pool_state_t spa_state(spa_t *spa);
 974 extern spa_load_state_t spa_load_state(spa_t *spa);
 975 extern uint64_t spa_freeze_txg(spa_t *spa);
 976 extern uint64_t spa_get_worst_case_asize(spa_t *spa, uint64_t lsize);
 977 extern uint64_t spa_get_dspace(spa_t *spa);
 978 extern uint64_t spa_get_checkpoint_space(spa_t *spa);
 979 extern uint64_t spa_get_slop_space(spa_t *spa);
 980 extern void spa_update_dspace(spa_t *spa);
 981 extern uint64_t spa_version(spa_t *spa);
 982 extern boolean_t spa_deflate(spa_t *spa);
 983 extern metaslab_class_t *spa_normal_class(spa_t *spa);
 984 extern metaslab_class_t *spa_log_class(spa_t *spa);
 985 extern metaslab_class_t *spa_special_class(spa_t *spa);
 986 extern metaslab_class_t *spa_dedup_class(spa_t *spa);
 987 extern metaslab_class_t *spa_preferred_class(spa_t *spa, uint64_t size,
 988     dmu_object_type_t objtype, uint_t level, uint_t special_smallblk);
 989
 990 extern void spa_evicting_os_register(spa_t *, objset_t *os);
 991 extern void spa_evicting_os_deregister(spa_t *, objset_t *os);
 992 extern void spa_evicting_os_wait(spa_t *spa);
 993 extern int spa_max_replication(spa_t *spa);
 994 extern int spa_prev_software_version(spa_t *spa);
 995 extern uint64_t spa_get_failmode(spa_t *spa);
 996 extern uint64_t spa_get_deadman_failmode(spa_t *spa);
 997 extern void spa_set_deadman_failmode(spa_t *spa, const char *failmode);
 998 extern boolean_t spa_suspended(spa_t *spa);
 999 extern uint64_t spa_bootfs(spa_t *spa);
1000 extern uint64_t spa_delegation(spa_t *spa);
1001 extern objset_t *spa_meta_objset(spa_t *spa);
1002 extern uint64_t spa_deadman_synctime(spa_t *spa);
1003 extern uint64_t spa_deadman_ziotime(spa_t *spa);
1004 extern uint64_t spa_dirty_data(spa_t *spa);
1005
1006 /* Miscellaneous support routines */
1007 extern void spa_load_failed(spa_t *spa, const char *fmt, ...);
1008 extern void spa_load_note(spa_t *spa, const char *fmt, ...);
1009 extern void spa_activate_mos_feature(spa_t *spa, const char *feature,
1010     dmu_tx_t *tx);
1011 extern void spa_deactivate_mos_feature(spa_t *spa, const char *feature);
1012 extern int spa_rename(const char *oldname, const char *newname);
1013 extern spa_t *spa_by_guid(uint64_t pool_guid, uint64_t device_guid);
1014 extern boolean_t spa_guid_exists(uint64_t pool_guid, uint64_t device_guid);
1015 extern char *spa_strdup(const char *);
1016 extern void spa_strfree(char *);
1017 extern uint64_t spa_get_random(uint64_t range);
1018 extern uint64_t spa_generate_guid(spa_t *spa);
1019 extern void snprintf_blkptr(char *buf, size_t buflen, const blkptr_t *bp);
1020 extern void spa_freeze(spa_t *spa);
1021 extern int spa_change_guid(spa_t *spa);
1022 extern void spa_upgrade(spa_t *spa, uint64_t version);
1023 extern void spa_evict_all(void);
1024 extern vdev_t *spa_lookup_by_guid(spa_t *spa, uint64_t guid,
1025     boolean_t l2cache);
1026 extern boolean_t spa_has_spare(spa_t *, uint64_t guid);
1027 extern uint64_t dva_get_dsize_sync(spa_t *spa, const dva_t *dva);
1028 extern uint64_t bp_get_dsize_sync(spa_t *spa, const blkptr_t *bp);
1029 extern uint64_t bp_get_dsize(spa_t *spa, const blkptr_t *bp);
1030 extern boolean_t spa_has_slogs(spa_t *spa);
1031 extern boolean_t spa_is_root(spa_t *spa);
1032 extern boolean_t spa_writeable(spa_t *spa);
1033 extern boolean_t spa_has_pending_synctask(spa_t *spa);
1034 extern int spa_maxblocksize(spa_t *spa);
1035 extern int spa_maxdnodesize(spa_t *spa);
1036 extern boolean_t spa_has_checkpoint(spa_t *spa);
1037 extern boolean_t spa_importing_readonly_checkpoint(spa_t *spa);
1038 extern boolean_t spa_suspend_async_destroy(spa_t *spa);
1039 extern uint64_t spa_min_claim_txg(spa_t *spa);
1040 extern void zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp);
1041 extern boolean_t zfs_dva_valid(spa_t *spa, const dva_t *dva,
1042     const blkptr_t *bp);
1043 typedef void (*spa_remap_cb_t)(uint64_t vdev, uint64_t offset, uint64_t size,
1044     void *arg);
1045 extern boolean_t spa_remap_blkptr(spa_t *spa, blkptr_t *bp,
1046     spa_remap_cb_t callback, void *arg);
1047 extern uint64_t spa_get_last_removal_txg(spa_t *spa);
1048 extern boolean_t spa_trust_config(spa_t *spa);
1049 extern uint64_t spa_missing_tvds_allowed(spa_t *spa);
1050 extern void spa_set_missing_tvds(spa_t *spa, uint64_t missing);
1051 extern boolean_t spa_top_vdevs_spacemap_addressable(spa_t *spa);
1052 extern boolean_t spa_multihost(spa_t *spa);
1053 extern unsigned long spa_get_hostid(void);
1054 extern void spa_activate_allocation_classes(spa_t *, dmu_tx_t *);
1055
1056 extern int spa_mode(spa_t *spa);
1057 extern uint64_t zfs_strtonum(const char *str, char **nptr);
1058
1059 extern char *spa_his_ievent_table[];
1060
1061 extern void spa_history_create_obj(spa_t *spa, dmu_tx_t *tx);
1062 extern int spa_history_get(spa_t *spa, uint64_t *offset, uint64_t *len_read,
1063     char *his_buf);
1064 extern int spa_history_log(spa_t *spa, const char *his_buf);
1065 extern int spa_history_log_nvl(spa_t *spa, nvlist_t *nvl);
1066 extern void spa_history_log_version(spa_t *spa, const char *operation,
1067     dmu_tx_t *tx);
1068 extern void spa_history_log_internal(spa_t *spa, const char *operation,
1069     dmu_tx_t *tx, const char *fmt, ...);
1070 extern void spa_history_log_internal_ds(struct dsl_dataset *ds, const char *op,
1071     dmu_tx_t *tx, const char *fmt, ...);
1072 extern void spa_history_log_internal_dd(dsl_dir_t *dd, const char *operation,
1073     dmu_tx_t *tx, const char *fmt, ...);
1074
1075 extern const char *spa_state_to_name(spa_t *spa);
1076
1077 /* error handling */
1078 struct zbookmark_phys;
1079 extern void spa_log_error(spa_t *spa, const zbookmark_phys_t *zb);
1080 extern void zfs_ereport_post(const char *class, spa_t *spa, vdev_t *vd,
1081     const zbookmark_phys_t *zb, zio_t *zio, uint64_t stateoroffset,
1082     uint64_t length);
1083 extern nvlist_t *zfs_event_create(spa_t *spa, vdev_t *vd, const char *type,
1084     const char *name, nvlist_t *aux);
1085 extern void zfs_post_remove(spa_t *spa, vdev_t *vd);
1086 extern void zfs_post_state_change(spa_t *spa, vdev_t *vd, uint64_t laststate);
1087 extern void zfs_post_autoreplace(spa_t *spa, vdev_t *vd);
1088 extern uint64_t spa_get_errlog_size(spa_t *spa);
1089 extern int spa_get_errlog(spa_t *spa, void *uaddr, size_t *count);
1090 extern void spa_errlog_rotate(spa_t *spa);
1091 extern void spa_errlog_drain(spa_t *spa);
1092 extern void spa_errlog_sync(spa_t *spa, uint64_t txg);
1093 extern void spa_get_errlists(spa_t *spa, avl_tree_t *last, avl_tree_t *scrub);
1094
1095 /* vdev cache */
1096 extern void vdev_cache_stat_init(void);
1097 extern void vdev_cache_stat_fini(void);
1098
1099 /* vdev mirror */
1100 extern void vdev_mirror_stat_init(void);
1101 extern void vdev_mirror_stat_fini(void);
1102
1103 /* Initialization and termination */
1104 extern void spa_init(int flags);
1105 extern void spa_fini(void);
1106 extern void spa_boot_init(void);
1107
1108 /* properties */
1109 extern int spa_prop_set(spa_t *spa, nvlist_t *nvp);
1110 extern int spa_prop_get(spa_t *spa, nvlist_t **nvp);
1111 extern void spa_prop_clear_bootfs(spa_t *spa, uint64_t obj, dmu_tx_t *tx);
1112 extern void spa_configfile_set(spa_t *, nvlist_t *, boolean_t);
1113
1114 /* asynchronous event notification */
1115 extern void spa_event_notify(spa_t *spa, vdev_t *vdev, nvlist_t *hist_nvl,
1116     const char *name);
1117
1118 #ifdef ZFS_DEBUG
1119 #define dprintf_bp(bp, fmt, ...) do {                           \
1120         if (zfs_flags & ZFS_DEBUG_DPRINTF) {                    \
1121         char *__blkbuf = kmem_alloc(BP_SPRINTF_LEN, KM_SLEEP);  \
1122         snprintf_blkptr(__blkbuf, BP_SPRINTF_LEN, (bp));        \
1123         dprintf(fmt " %s\n", __VA_ARGS__, __blkbuf);            \
1124         kmem_free(__blkbuf, BP_SPRINTF_LEN);                    \
1125         } \
1126 _NOTE(CONSTCOND) } while (0)
1127 #else
1128 #define dprintf_bp(bp, fmt, ...)
1129 #endif
1130
1131 extern int spa_mode_global;                     /* mode, e.g. FREAD | FWRITE */
1132 extern int zfs_deadman_enabled;
1133 extern unsigned long zfs_deadman_synctime_ms;
1134 extern unsigned long zfs_deadman_ziotime_ms;
1135 extern unsigned long zfs_deadman_checktime_ms;
1136
1137 #ifdef  __cplusplus
1138 }
1139 #endif
1140
1141 #endif  /* _SYS_SPA_H */