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