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 https://opensource.org/licenses/CDDL-1.0.
  10  * See the License for the specific language governing permissions
  11  * and limitations under the License.
  12  *
  13  * When distributing Covered Code, include this CDDL HEADER in each
  14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
  15  * If applicable, add the following below this CDDL HEADER, with the
  16  * fields enclosed by brackets "[]" replaced with your own identifying
  17  * information: Portions Copyright [yyyy] [name of copyright owner]
  18  *
  19  * CDDL HEADER END
  20  */
  21 /*
  22  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
  23  * Copyright (c) 2011, 2021 by Delphix. All rights reserved.
  24  * Copyright 2011 Nexenta Systems, Inc.  All rights reserved.
  25  * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
  26  * Copyright 2013 Saso Kiselkov. All rights reserved.
  27  * Copyright (c) 2014 Integros [integros.com]
  28  * Copyright 2017 Joyent, Inc.
  29  * Copyright (c) 2017, Intel Corporation.
  30  * Copyright (c) 2019, Allan Jude
  31  * Copyright (c) 2019, Klara Inc.
  32  * Copyright (c) 2019, Datto Inc.
  33  */
  34
  35 #ifndef _SYS_SPA_H
  36 #define _SYS_SPA_H
  37
  38 #include <sys/avl.h>
  39 #include <sys/zfs_context.h>
  40 #include <sys/kstat.h>
  41 #include <sys/nvpair.h>
  42 #include <sys/sysmacros.h>
  43 #include <sys/types.h>
  44 #include <sys/fs/zfs.h>
  45 #include <sys/spa_checksum.h>
  46 #include <sys/dmu.h>
  47 #include <sys/space_map.h>
  48 #include <sys/bitops.h>
  49
  50 #ifdef  __cplusplus
  51 extern "C" {
  52 #endif
  53
  54 /*
  55  * Forward references that lots of things need.
  56  */
  57 typedef struct spa spa_t;
  58 typedef struct vdev vdev_t;
  59 typedef struct metaslab metaslab_t;
  60 typedef struct metaslab_group metaslab_group_t;
  61 typedef struct metaslab_class metaslab_class_t;
  62 typedef struct zio zio_t;
  63 typedef struct zilog zilog_t;
  64 typedef struct spa_aux_vdev spa_aux_vdev_t;
  65 typedef struct ddt ddt_t;
  66 typedef struct ddt_entry ddt_entry_t;
  67 typedef struct zbookmark_phys zbookmark_phys_t;
  68 typedef struct zbookmark_err_phys zbookmark_err_phys_t;
  69
  70 struct bpobj;
  71 struct bplist;
  72 struct dsl_pool;
  73 struct dsl_dataset;
  74 struct dsl_crypto_params;
  75
  76 /*
  77  * Alignment Shift (ashift) is an immutable, internal top-level vdev property
  78  * which can only be set at vdev creation time. Physical writes are always done
  79  * according to it, which makes 2^ashift the smallest possible IO on a vdev.
  80  *
  81  * We currently allow values ranging from 512 bytes (2^9 = 512) to 64 KiB
  82  * (2^16 = 65,536).
  83  */
  84 #define ASHIFT_MIN              9
  85 #define ASHIFT_MAX              16
  86
  87 /*
  88  * Size of block to hold the configuration data (a packed nvlist)
  89  */
  90 #define SPA_CONFIG_BLOCKSIZE    (1ULL << 14)
  91
  92 /*
  93  * The DVA size encodings for LSIZE and PSIZE support blocks up to 32MB.
  94  * The ASIZE encoding should be at least 64 times larger (6 more bits)
  95  * to support up to 4-way RAID-Z mirror mode with worst-case gang block
  96  * overhead, three DVAs per bp, plus one more bit in case we do anything
  97  * else that expands the ASIZE.
  98  */
  99 #define SPA_LSIZEBITS           16      /* LSIZE up to 32M (2^16 * 512) */
 100 #define SPA_PSIZEBITS           16      /* PSIZE up to 32M (2^16 * 512) */
 101 #define SPA_ASIZEBITS           24      /* ASIZE up to 64 times larger  */
 102
 103 #define SPA_COMPRESSBITS        7
 104 #define SPA_VDEVBITS            24
 105 #define SPA_COMPRESSMASK        ((1U << SPA_COMPRESSBITS) - 1)
 106
 107 /*
 108  * All SPA data is represented by 128-bit data virtual addresses (DVAs).
 109  * The members of the dva_t should be considered opaque outside the SPA.
 110  */
 111 typedef struct dva {
 112         uint64_t        dva_word[2];
 113 } dva_t;
 114
 115
 116 /*
 117  * Some checksums/hashes need a 256-bit initialization salt. This salt is kept
 118  * secret and is suitable for use in MAC algorithms as the key.
 119  */
 120 typedef struct zio_cksum_salt {
 121         uint8_t         zcs_bytes[32];
 122 } zio_cksum_salt_t;
 123
 124 /*
 125  * Each block is described by its DVAs, time of birth, checksum, etc.
 126  * The word-by-word, bit-by-bit layout of the blkptr is as follows:
 127  *
 128  *      64      56      48      40      32      24      16      8       0
 129  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 130  * 0    |  pad  |         vdev1         | GRID  |         ASIZE         |
 131  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 132  * 1    |G|                      offset1                                |
 133  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 134  * 2    |  pad  |         vdev2         | GRID  |         ASIZE         |
 135  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 136  * 3    |G|                      offset2                                |
 137  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 138  * 4    |  pad  |         vdev3         | GRID  |         ASIZE         |
 139  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 140  * 5    |G|                      offset3                                |
 141  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 142  * 6    |BDX|lvl| type  | cksum |E| comp|    PSIZE      |     LSIZE     |
 143  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 144  * 7    |                       padding                                 |
 145  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 146  * 8    |                       padding                                 |
 147  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 148  * 9    |                       physical birth txg                      |
 149  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 150  * a    |                       logical birth txg                       |
 151  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 152  * b    |                       fill count                              |
 153  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 154  * c    |                       checksum[0]                             |
 155  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 156  * d    |                       checksum[1]                             |
 157  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 158  * e    |                       checksum[2]                             |
 159  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 160  * f    |                       checksum[3]                             |
 161  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 162  *
 163  * Legend:
 164  *
 165  * vdev         virtual device ID
 166  * offset       offset into virtual device
 167  * LSIZE        logical size
 168  * PSIZE        physical size (after compression)
 169  * ASIZE        allocated size (including RAID-Z parity and gang block headers)
 170  * GRID         RAID-Z layout information (reserved for future use)
 171  * cksum        checksum function
 172  * comp         compression function
 173  * G            gang block indicator
 174  * B            byteorder (endianness)
 175  * D            dedup
 176  * X            encryption
 177  * E            blkptr_t contains embedded data (see below)
 178  * lvl          level of indirection
 179  * type         DMU object type
 180  * phys birth   txg when dva[0] was written; zero if same as logical birth txg
 181  *              note that typically all the dva's would be written in this
 182  *              txg, but they could be different if they were moved by
 183  *              device removal.
 184  * log. birth   transaction group in which the block was logically born
 185  * fill count   number of non-zero blocks under this bp
 186  * checksum[4]  256-bit checksum of the data this bp describes
 187  */
 188
 189 /*
 190  * The blkptr_t's of encrypted blocks also need to store the encryption
 191  * parameters so that the block can be decrypted. This layout is as follows:
 192  *
 193  *      64      56      48      40      32      24      16      8       0
 194  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 195  * 0    |               vdev1           | GRID  |         ASIZE         |
 196  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 197  * 1    |G|                      offset1                                |
 198  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 199  * 2    |               vdev2           | GRID  |         ASIZE         |
 200  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 201  * 3    |G|                      offset2                                |
 202  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 203  * 4    |                       salt                                    |
 204  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 205  * 5    |                       IV1                                     |
 206  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 207  * 6    |BDX|lvl| type  | cksum |E| comp|    PSIZE      |     LSIZE     |
 208  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 209  * 7    |                       padding                                 |
 210  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 211  * 8    |                       padding                                 |
 212  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 213  * 9    |                       physical birth txg                      |
 214  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 215  * a    |                       logical birth txg                       |
 216  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 217  * b    |               IV2             |           fill count          |
 218  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 219  * c    |                       checksum[0]                             |
 220  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 221  * d    |                       checksum[1]                             |
 222  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 223  * e    |                       MAC[0]                                  |
 224  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 225  * f    |                       MAC[1]                                  |
 226  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 227  *
 228  * Legend:
 229  *
 230  * salt         Salt for generating encryption keys
 231  * IV1          First 64 bits of encryption IV
 232  * X            Block requires encryption handling (set to 1)
 233  * E            blkptr_t contains embedded data (set to 0, see below)
 234  * fill count   number of non-zero blocks under this bp (truncated to 32 bits)
 235  * IV2          Last 32 bits of encryption IV
 236  * checksum[2]  128-bit checksum of the data this bp describes
 237  * MAC[2]       128-bit message authentication code for this data
 238  *
 239  * The X bit being set indicates that this block is one of 3 types. If this is
 240  * a level 0 block with an encrypted object type, the block is encrypted
 241  * (see BP_IS_ENCRYPTED()). If this is a level 0 block with an unencrypted
 242  * object type, this block is authenticated with an HMAC (see
 243  * BP_IS_AUTHENTICATED()). Otherwise (if level > 0), this bp will use the MAC
 244  * words to store a checksum-of-MACs from the level below (see
 245  * BP_HAS_INDIRECT_MAC_CKSUM()). For convenience in the code, BP_IS_PROTECTED()
 246  * refers to both encrypted and authenticated blocks and BP_USES_CRYPT()
 247  * refers to any of these 3 kinds of blocks.
 248  *
 249  * The additional encryption parameters are the salt, IV, and MAC which are
 250  * explained in greater detail in the block comment at the top of zio_crypt.c.
 251  * The MAC occupies half of the checksum space since it serves a very similar
 252  * purpose: to prevent data corruption on disk. The only functional difference
 253  * is that the checksum is used to detect on-disk corruption whether or not the
 254  * encryption key is loaded and the MAC provides additional protection against
 255  * malicious disk tampering. We use the 3rd DVA to store the salt and first
 256  * 64 bits of the IV. As a result encrypted blocks can only have 2 copies
 257  * maximum instead of the normal 3. The last 32 bits of the IV are stored in
 258  * the upper bits of what is usually the fill count. Note that only blocks at
 259  * level 0 or -2 are ever encrypted, which allows us to guarantee that these
 260  * 32 bits are not trampled over by other code (see zio_crypt.c for details).
 261  * The salt and IV are not used for authenticated bps or bps with an indirect
 262  * MAC checksum, so these blocks can utilize all 3 DVAs and the full 64 bits
 263  * for the fill count.
 264  */
 265
 266 /*
 267  * "Embedded" blkptr_t's don't actually point to a block, instead they
 268  * have a data payload embedded in the blkptr_t itself.  See the comment
 269  * in blkptr.c for more details.
 270  *
 271  * The blkptr_t is laid out as follows:
 272  *
 273  *      64      56      48      40      32      24      16      8       0
 274  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 275  * 0    |      payload                                                  |
 276  * 1    |      payload                                                  |
 277  * 2    |      payload                                                  |
 278  * 3    |      payload                                                  |
 279  * 4    |      payload                                                  |
 280  * 5    |      payload                                                  |
 281  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 282  * 6    |BDX|lvl| type  | etype |E| comp| PSIZE|              LSIZE     |
 283  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 284  * 7    |      payload                                                  |
 285  * 8    |      payload                                                  |
 286  * 9    |      payload                                                  |
 287  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 288  * a    |                       logical birth txg                       |
 289  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 290  * b    |      payload                                                  |
 291  * c    |      payload                                                  |
 292  * d    |      payload                                                  |
 293  * e    |      payload                                                  |
 294  * f    |      payload                                                  |
 295  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 296  *
 297  * Legend:
 298  *
 299  * payload              contains the embedded data
 300  * B (byteorder)        byteorder (endianness)
 301  * D (dedup)            padding (set to zero)
 302  * X                    encryption (set to zero)
 303  * E (embedded)         set to one
 304  * lvl                  indirection level
 305  * type                 DMU object type
 306  * etype                how to interpret embedded data (BP_EMBEDDED_TYPE_*)
 307  * comp                 compression function of payload
 308  * PSIZE                size of payload after compression, in bytes
 309  * LSIZE                logical size of payload, in bytes
 310  *                      note that 25 bits is enough to store the largest
 311  *                      "normal" BP's LSIZE (2^16 * 2^9) in bytes
 312  * log. birth           transaction group in which the block was logically born
 313  *
 314  * Note that LSIZE and PSIZE are stored in bytes, whereas for non-embedded
 315  * bp's they are stored in units of SPA_MINBLOCKSHIFT.
 316  * Generally, the generic BP_GET_*() macros can be used on embedded BP's.
 317  * The B, D, X, lvl, type, and comp fields are stored the same as with normal
 318  * BP's so the BP_SET_* macros can be used with them.  etype, PSIZE, LSIZE must
 319  * be set with the BPE_SET_* macros.  BP_SET_EMBEDDED() should be called before
 320  * other macros, as they assert that they are only used on BP's of the correct
 321  * "embedded-ness". Encrypted blkptr_t's cannot be embedded because they use
 322  * the payload space for encryption parameters (see the comment above on
 323  * how encryption parameters are stored).
 324  */
 325
 326 #define BPE_GET_ETYPE(bp)       \
 327         (ASSERT(BP_IS_EMBEDDED(bp)), \
 328         BF64_GET((bp)->blk_prop, 40, 8))
 329 #define BPE_SET_ETYPE(bp, t)    do { \
 330         ASSERT(BP_IS_EMBEDDED(bp)); \
 331         BF64_SET((bp)->blk_prop, 40, 8, t); \
 332 } while (0)
 333
 334 #define BPE_GET_LSIZE(bp)       \
 335         (ASSERT(BP_IS_EMBEDDED(bp)), \
 336         BF64_GET_SB((bp)->blk_prop, 0, 25, 0, 1))
 337 #define BPE_SET_LSIZE(bp, x)    do { \
 338         ASSERT(BP_IS_EMBEDDED(bp)); \
 339         BF64_SET_SB((bp)->blk_prop, 0, 25, 0, 1, x); \
 340 } while (0)
 341
 342 #define BPE_GET_PSIZE(bp)       \
 343         (ASSERT(BP_IS_EMBEDDED(bp)), \
 344         BF64_GET_SB((bp)->blk_prop, 25, 7, 0, 1))
 345 #define BPE_SET_PSIZE(bp, x)    do { \
 346         ASSERT(BP_IS_EMBEDDED(bp)); \
 347         BF64_SET_SB((bp)->blk_prop, 25, 7, 0, 1, x); \
 348 } while (0)
 349
 350 typedef enum bp_embedded_type {
 351         BP_EMBEDDED_TYPE_DATA,
 352         BP_EMBEDDED_TYPE_RESERVED, /* Reserved for Delphix byteswap feature. */
 353         BP_EMBEDDED_TYPE_REDACTED,
 354         NUM_BP_EMBEDDED_TYPES
 355 } bp_embedded_type_t;
 356
 357 #define BPE_NUM_WORDS 14
 358 #define BPE_PAYLOAD_SIZE (BPE_NUM_WORDS * sizeof (uint64_t))
 359 #define BPE_IS_PAYLOADWORD(bp, wp) \
 360         ((wp) != &(bp)->blk_prop && (wp) != &(bp)->blk_birth)
 361
 362 #define SPA_BLKPTRSHIFT 7               /* blkptr_t is 128 bytes        */
 363 #define SPA_DVAS_PER_BP 3               /* Number of DVAs in a bp       */
 364 #define SPA_SYNC_MIN_VDEVS 3            /* min vdevs to update during sync */
 365
 366 /*
 367  * A block is a hole when it has either 1) never been written to, or
 368  * 2) is zero-filled. In both cases, ZFS can return all zeroes for all reads
 369  * without physically allocating disk space. Holes are represented in the
 370  * blkptr_t structure by zeroed blk_dva. Correct checking for holes is
 371  * done through the BP_IS_HOLE macro. For holes, the logical size, level,
 372  * DMU object type, and birth times are all also stored for holes that
 373  * were written to at some point (i.e. were punched after having been filled).
 374  */
 375 typedef struct blkptr {
 376         dva_t           blk_dva[SPA_DVAS_PER_BP]; /* Data Virtual Addresses */
 377         uint64_t        blk_prop;       /* size, compression, type, etc     */
 378         uint64_t        blk_pad[2];     /* Extra space for the future       */
 379         uint64_t        blk_phys_birth; /* txg when block was allocated     */
 380         uint64_t        blk_birth;      /* transaction group at birth       */
 381         uint64_t        blk_fill;       /* fill count                       */
 382         zio_cksum_t     blk_cksum;      /* 256-bit checksum                 */
 383 } blkptr_t;
 384
 385 /*
 386  * Macros to get and set fields in a bp or DVA.
 387  */
 388
 389 /*
 390  * Note, for gang blocks, DVA_GET_ASIZE() is the total space allocated for
 391  * this gang DVA including its children BP's.  The space allocated at this
 392  * DVA's vdev/offset is vdev_gang_header_asize(vdev).
 393  */
 394 #define DVA_GET_ASIZE(dva)      \
 395         BF64_GET_SB((dva)->dva_word[0], 0, SPA_ASIZEBITS, SPA_MINBLOCKSHIFT, 0)
 396 #define DVA_SET_ASIZE(dva, x)   \
 397         BF64_SET_SB((dva)->dva_word[0], 0, SPA_ASIZEBITS, \
 398         SPA_MINBLOCKSHIFT, 0, x)
 399
 400 #define DVA_GET_GRID(dva)       BF64_GET((dva)->dva_word[0], 24, 8)
 401 #define DVA_SET_GRID(dva, x)    BF64_SET((dva)->dva_word[0], 24, 8, x)
 402
 403 #define DVA_GET_VDEV(dva)       BF64_GET((dva)->dva_word[0], 32, SPA_VDEVBITS)
 404 #define DVA_SET_VDEV(dva, x)    \
 405         BF64_SET((dva)->dva_word[0], 32, SPA_VDEVBITS, x)
 406
 407 #define DVA_GET_OFFSET(dva)     \
 408         BF64_GET_SB((dva)->dva_word[1], 0, 63, SPA_MINBLOCKSHIFT, 0)
 409 #define DVA_SET_OFFSET(dva, x)  \
 410         BF64_SET_SB((dva)->dva_word[1], 0, 63, SPA_MINBLOCKSHIFT, 0, x)
 411
 412 #define DVA_GET_GANG(dva)       BF64_GET((dva)->dva_word[1], 63, 1)
 413 #define DVA_SET_GANG(dva, x)    BF64_SET((dva)->dva_word[1], 63, 1, x)
 414
 415 #define BP_GET_LSIZE(bp)        \
 416         (BP_IS_EMBEDDED(bp) ?   \
 417         (BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA ? BPE_GET_LSIZE(bp) : 0): \
 418         BF64_GET_SB((bp)->blk_prop, 0, SPA_LSIZEBITS, SPA_MINBLOCKSHIFT, 1))
 419 #define BP_SET_LSIZE(bp, x)     do { \
 420         ASSERT(!BP_IS_EMBEDDED(bp)); \
 421         BF64_SET_SB((bp)->blk_prop, \
 422             0, SPA_LSIZEBITS, SPA_MINBLOCKSHIFT, 1, x); \
 423 } while (0)
 424
 425 #define BP_GET_PSIZE(bp)        \
 426         (BP_IS_EMBEDDED(bp) ? 0 : \
 427         BF64_GET_SB((bp)->blk_prop, 16, SPA_PSIZEBITS, SPA_MINBLOCKSHIFT, 1))
 428 #define BP_SET_PSIZE(bp, x)     do { \
 429         ASSERT(!BP_IS_EMBEDDED(bp)); \
 430         BF64_SET_SB((bp)->blk_prop, \
 431             16, SPA_PSIZEBITS, SPA_MINBLOCKSHIFT, 1, x); \
 432 } while (0)
 433
 434 #define BP_GET_COMPRESS(bp)             \
 435         BF64_GET((bp)->blk_prop, 32, SPA_COMPRESSBITS)
 436 #define BP_SET_COMPRESS(bp, x)          \
 437         BF64_SET((bp)->blk_prop, 32, SPA_COMPRESSBITS, x)
 438
 439 #define BP_IS_EMBEDDED(bp)              BF64_GET((bp)->blk_prop, 39, 1)
 440 #define BP_SET_EMBEDDED(bp, x)          BF64_SET((bp)->blk_prop, 39, 1, x)
 441
 442 #define BP_GET_CHECKSUM(bp)             \
 443         (BP_IS_EMBEDDED(bp) ? ZIO_CHECKSUM_OFF : \
 444         BF64_GET((bp)->blk_prop, 40, 8))
 445 #define BP_SET_CHECKSUM(bp, x)          do { \
 446         ASSERT(!BP_IS_EMBEDDED(bp)); \
 447         BF64_SET((bp)->blk_prop, 40, 8, x); \
 448 } while (0)
 449
 450 #define BP_GET_TYPE(bp)                 BF64_GET((bp)->blk_prop, 48, 8)
 451 #define BP_SET_TYPE(bp, x)              BF64_SET((bp)->blk_prop, 48, 8, x)
 452
 453 #define BP_GET_LEVEL(bp)                BF64_GET((bp)->blk_prop, 56, 5)
 454 #define BP_SET_LEVEL(bp, x)             BF64_SET((bp)->blk_prop, 56, 5, x)
 455
 456 /* encrypted, authenticated, and MAC cksum bps use the same bit */
 457 #define BP_USES_CRYPT(bp)               BF64_GET((bp)->blk_prop, 61, 1)
 458 #define BP_SET_CRYPT(bp, x)             BF64_SET((bp)->blk_prop, 61, 1, x)
 459
 460 #define BP_IS_ENCRYPTED(bp)                     \
 461         (BP_USES_CRYPT(bp) &&                   \
 462         BP_GET_LEVEL(bp) <= 0 &&                \
 463         DMU_OT_IS_ENCRYPTED(BP_GET_TYPE(bp)))
 464
 465 #define BP_IS_AUTHENTICATED(bp)                 \
 466         (BP_USES_CRYPT(bp) &&                   \
 467         BP_GET_LEVEL(bp) <= 0 &&                \
 468         !DMU_OT_IS_ENCRYPTED(BP_GET_TYPE(bp)))
 469
 470 #define BP_HAS_INDIRECT_MAC_CKSUM(bp)           \
 471         (BP_USES_CRYPT(bp) && BP_GET_LEVEL(bp) > 0)
 472
 473 #define BP_IS_PROTECTED(bp)                     \
 474         (BP_IS_ENCRYPTED(bp) || BP_IS_AUTHENTICATED(bp))
 475
 476 #define BP_GET_DEDUP(bp)                BF64_GET((bp)->blk_prop, 62, 1)
 477 #define BP_SET_DEDUP(bp, x)             BF64_SET((bp)->blk_prop, 62, 1, x)
 478
 479 #define BP_GET_BYTEORDER(bp)            BF64_GET((bp)->blk_prop, 63, 1)
 480 #define BP_SET_BYTEORDER(bp, x)         BF64_SET((bp)->blk_prop, 63, 1, x)
 481
 482 #define BP_GET_FREE(bp)                 BF64_GET((bp)->blk_fill, 0, 1)
 483 #define BP_SET_FREE(bp, x)              BF64_SET((bp)->blk_fill, 0, 1, x)
 484
 485 #define BP_PHYSICAL_BIRTH(bp)           \
 486         (BP_IS_EMBEDDED(bp) ? 0 : \
 487         (bp)->blk_phys_birth ? (bp)->blk_phys_birth : (bp)->blk_birth)
 488
 489 #define BP_SET_BIRTH(bp, logical, physical)     \
 490 {                                               \
 491         ASSERT(!BP_IS_EMBEDDED(bp));            \
 492         (bp)->blk_birth = (logical);            \
 493         (bp)->blk_phys_birth = ((logical) == (physical) ? 0 : (physical)); \
 494 }
 495
 496 #define BP_GET_FILL(bp)                         \
 497         ((BP_IS_ENCRYPTED(bp)) ? BF64_GET((bp)->blk_fill, 0, 32) : \
 498         ((BP_IS_EMBEDDED(bp)) ? 1 : (bp)->blk_fill))
 499
 500 #define BP_SET_FILL(bp, fill)                   \
 501 {                                               \
 502         if (BP_IS_ENCRYPTED(bp))                        \
 503                 BF64_SET((bp)->blk_fill, 0, 32, fill); \
 504         else                                    \
 505                 (bp)->blk_fill = fill;          \
 506 }
 507
 508 #define BP_GET_IV2(bp)                          \
 509         (ASSERT(BP_IS_ENCRYPTED(bp)),           \
 510         BF64_GET((bp)->blk_fill, 32, 32))
 511 #define BP_SET_IV2(bp, iv2)                     \
 512 {                                               \
 513         ASSERT(BP_IS_ENCRYPTED(bp));            \
 514         BF64_SET((bp)->blk_fill, 32, 32, iv2);  \
 515 }
 516
 517 #define BP_IS_METADATA(bp)      \
 518         (BP_GET_LEVEL(bp) > 0 || DMU_OT_IS_METADATA(BP_GET_TYPE(bp)))
 519
 520 #define BP_GET_ASIZE(bp)        \
 521         (BP_IS_EMBEDDED(bp) ? 0 : \
 522         DVA_GET_ASIZE(&(bp)->blk_dva[0]) + \
 523         DVA_GET_ASIZE(&(bp)->blk_dva[1]) + \
 524         (DVA_GET_ASIZE(&(bp)->blk_dva[2]) * !BP_IS_ENCRYPTED(bp)))
 525
 526 #define BP_GET_UCSIZE(bp)       \
 527         (BP_IS_METADATA(bp) ? BP_GET_PSIZE(bp) : BP_GET_LSIZE(bp))
 528
 529 #define BP_GET_NDVAS(bp)        \
 530         (BP_IS_EMBEDDED(bp) ? 0 : \
 531         !!DVA_GET_ASIZE(&(bp)->blk_dva[0]) + \
 532         !!DVA_GET_ASIZE(&(bp)->blk_dva[1]) + \
 533         (!!DVA_GET_ASIZE(&(bp)->blk_dva[2]) * !BP_IS_ENCRYPTED(bp)))
 534
 535 #define BP_COUNT_GANG(bp)       \
 536         (BP_IS_EMBEDDED(bp) ? 0 : \
 537         (DVA_GET_GANG(&(bp)->blk_dva[0]) + \
 538         DVA_GET_GANG(&(bp)->blk_dva[1]) + \
 539         (DVA_GET_GANG(&(bp)->blk_dva[2]) * !BP_IS_ENCRYPTED(bp))))
 540
 541 #define DVA_EQUAL(dva1, dva2)   \
 542         ((dva1)->dva_word[1] == (dva2)->dva_word[1] && \
 543         (dva1)->dva_word[0] == (dva2)->dva_word[0])
 544
 545 #define BP_EQUAL(bp1, bp2)      \
 546         (BP_PHYSICAL_BIRTH(bp1) == BP_PHYSICAL_BIRTH(bp2) &&    \
 547         (bp1)->blk_birth == (bp2)->blk_birth &&                 \
 548         DVA_EQUAL(&(bp1)->blk_dva[0], &(bp2)->blk_dva[0]) &&    \
 549         DVA_EQUAL(&(bp1)->blk_dva[1], &(bp2)->blk_dva[1]) &&    \
 550         DVA_EQUAL(&(bp1)->blk_dva[2], &(bp2)->blk_dva[2]))
 551
 552
 553 #define DVA_IS_VALID(dva)       (DVA_GET_ASIZE(dva) != 0)
 554
 555 #define BP_IDENTITY(bp)         (ASSERT(!BP_IS_EMBEDDED(bp)), &(bp)->blk_dva[0])
 556 #define BP_IS_GANG(bp)          \
 557         (BP_IS_EMBEDDED(bp) ? B_FALSE : DVA_GET_GANG(BP_IDENTITY(bp)))
 558 #define DVA_IS_EMPTY(dva)       ((dva)->dva_word[0] == 0ULL &&  \
 559                                 (dva)->dva_word[1] == 0ULL)
 560 #define BP_IS_HOLE(bp) \
 561         (!BP_IS_EMBEDDED(bp) && DVA_IS_EMPTY(BP_IDENTITY(bp)))
 562
 563 #define BP_SET_REDACTED(bp) \
 564 {                                                       \
 565         BP_SET_EMBEDDED(bp, B_TRUE);                    \
 566         BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_REDACTED);   \
 567 }
 568 #define BP_IS_REDACTED(bp) \
 569         (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_REDACTED)
 570
 571 /* BP_IS_RAIDZ(bp) assumes no block compression */
 572 #define BP_IS_RAIDZ(bp)         (DVA_GET_ASIZE(&(bp)->blk_dva[0]) > \
 573                                 BP_GET_PSIZE(bp))
 574
 575 #define BP_ZERO(bp)                             \
 576 {                                               \
 577         (bp)->blk_dva[0].dva_word[0] = 0;       \
 578         (bp)->blk_dva[0].dva_word[1] = 0;       \
 579         (bp)->blk_dva[1].dva_word[0] = 0;       \
 580         (bp)->blk_dva[1].dva_word[1] = 0;       \
 581         (bp)->blk_dva[2].dva_word[0] = 0;       \
 582         (bp)->blk_dva[2].dva_word[1] = 0;       \
 583         (bp)->blk_prop = 0;                     \
 584         (bp)->blk_pad[0] = 0;                   \
 585         (bp)->blk_pad[1] = 0;                   \
 586         (bp)->blk_phys_birth = 0;               \
 587         (bp)->blk_birth = 0;                    \
 588         (bp)->blk_fill = 0;                     \
 589         ZIO_SET_CHECKSUM(&(bp)->blk_cksum, 0, 0, 0, 0); \
 590 }
 591
 592 #ifdef _ZFS_BIG_ENDIAN
 593 #define ZFS_HOST_BYTEORDER      (0ULL)
 594 #else
 595 #define ZFS_HOST_BYTEORDER      (1ULL)
 596 #endif
 597
 598 #define BP_SHOULD_BYTESWAP(bp)  (BP_GET_BYTEORDER(bp) != ZFS_HOST_BYTEORDER)
 599
 600 #define BP_SPRINTF_LEN  400
 601
 602 /*
 603  * This macro allows code sharing between zfs, libzpool, and mdb.
 604  * 'func' is either kmem_scnprintf() or mdb_snprintf().
 605  * 'ws' (whitespace) can be ' ' for single-line format, '\n' for multi-line.
 606  */
 607
 608 #define SNPRINTF_BLKPTR(func, ws, buf, size, bp, type, checksum, compress) \
 609 {                                                                       \
 610         static const char *const copyname[] =                           \
 611             { "zero", "single", "double", "triple" };                   \
 612         int len = 0;                                                    \
 613         int copies = 0;                                                 \
 614         const char *crypt_type;                                         \
 615         if (bp != NULL) {                                               \
 616                 if (BP_IS_ENCRYPTED(bp)) {                              \
 617                         crypt_type = "encrypted";                       \
 618                         /* LINTED E_SUSPICIOUS_COMPARISON */            \
 619                 } else if (BP_IS_AUTHENTICATED(bp)) {                   \
 620                         crypt_type = "authenticated";                   \
 621                 } else if (BP_HAS_INDIRECT_MAC_CKSUM(bp)) {             \
 622                         crypt_type = "indirect-MAC";                    \
 623                 } else {                                                \
 624                         crypt_type = "unencrypted";                     \
 625                 }                                                       \
 626         }                                                               \
 627         if (bp == NULL) {                                               \
 628                 len += func(buf + len, size - len, "<NULL>");           \
 629         } else if (BP_IS_HOLE(bp)) {                                    \
 630                 len += func(buf + len, size - len,                      \
 631                     "HOLE [L%llu %s] "                                  \
 632                     "size=%llxL birth=%lluL",                           \
 633                     (u_longlong_t)BP_GET_LEVEL(bp),                     \
 634                     type,                                               \
 635                     (u_longlong_t)BP_GET_LSIZE(bp),                     \
 636                     (u_longlong_t)bp->blk_birth);                       \
 637         } else if (BP_IS_EMBEDDED(bp)) {                                \
 638                 len = func(buf + len, size - len,                       \
 639                     "EMBEDDED [L%llu %s] et=%u %s "                     \
 640                     "size=%llxL/%llxP birth=%lluL",                     \
 641                     (u_longlong_t)BP_GET_LEVEL(bp),                     \
 642                     type,                                               \
 643                     (int)BPE_GET_ETYPE(bp),                             \
 644                     compress,                                           \
 645                     (u_longlong_t)BPE_GET_LSIZE(bp),                    \
 646                     (u_longlong_t)BPE_GET_PSIZE(bp),                    \
 647                     (u_longlong_t)bp->blk_birth);                       \
 648         } else if (BP_IS_REDACTED(bp)) {                                \
 649                 len += func(buf + len, size - len,                      \
 650                     "REDACTED [L%llu %s] size=%llxL birth=%lluL",       \
 651                     (u_longlong_t)BP_GET_LEVEL(bp),                     \
 652                     type,                                               \
 653                     (u_longlong_t)BP_GET_LSIZE(bp),                     \
 654                     (u_longlong_t)bp->blk_birth);                       \
 655         } else {                                                        \
 656                 for (int d = 0; d < BP_GET_NDVAS(bp); d++) {            \
 657                         const dva_t *dva = &bp->blk_dva[d];             \
 658                         if (DVA_IS_VALID(dva))                          \
 659                                 copies++;                               \
 660                         len += func(buf + len, size - len,              \
 661                             "DVA[%d]=<%llu:%llx:%llx>%c", d,            \
 662                             (u_longlong_t)DVA_GET_VDEV(dva),            \
 663                             (u_longlong_t)DVA_GET_OFFSET(dva),          \
 664                             (u_longlong_t)DVA_GET_ASIZE(dva),           \
 665                             ws);                                        \
 666                 }                                                       \
 667                 ASSERT3S(copies, >, 0);                                 \
 668                 if (BP_IS_ENCRYPTED(bp)) {                              \
 669                         len += func(buf + len, size - len,              \
 670                             "salt=%llx iv=%llx:%llx%c",                 \
 671                             (u_longlong_t)bp->blk_dva[2].dva_word[0],   \
 672                             (u_longlong_t)bp->blk_dva[2].dva_word[1],   \
 673                             (u_longlong_t)BP_GET_IV2(bp),               \
 674                             ws);                                        \
 675                 }                                                       \
 676                 if (BP_IS_GANG(bp) &&                                   \
 677                     DVA_GET_ASIZE(&bp->blk_dva[2]) <=                   \
 678                     DVA_GET_ASIZE(&bp->blk_dva[1]) / 2)                 \
 679                         copies--;                                       \
 680                 len += func(buf + len, size - len,                      \
 681                     "[L%llu %s] %s %s %s %s %s %s %s%c"                 \
 682                     "size=%llxL/%llxP birth=%lluL/%lluP fill=%llu%c"    \
 683                     "cksum=%016llx:%016llx:%016llx:%016llx",            \
 684                     (u_longlong_t)BP_GET_LEVEL(bp),                     \
 685                     type,                                               \
 686                     checksum,                                           \
 687                     compress,                                           \
 688                     crypt_type,                                         \
 689                     BP_GET_BYTEORDER(bp) == 0 ? "BE" : "LE",            \
 690                     BP_IS_GANG(bp) ? "gang" : "contiguous",             \
 691                     BP_GET_DEDUP(bp) ? "dedup" : "unique",              \
 692                     copyname[copies],                                   \
 693                     ws,                                                 \
 694                     (u_longlong_t)BP_GET_LSIZE(bp),                     \
 695                     (u_longlong_t)BP_GET_PSIZE(bp),                     \
 696                     (u_longlong_t)bp->blk_birth,                        \
 697                     (u_longlong_t)BP_PHYSICAL_BIRTH(bp),                \
 698                     (u_longlong_t)BP_GET_FILL(bp),                      \
 699                     ws,                                                 \
 700                     (u_longlong_t)bp->blk_cksum.zc_word[0],             \
 701                     (u_longlong_t)bp->blk_cksum.zc_word[1],             \
 702                     (u_longlong_t)bp->blk_cksum.zc_word[2],             \
 703                     (u_longlong_t)bp->blk_cksum.zc_word[3]);            \
 704         }                                                               \
 705         ASSERT(len < size);                                             \
 706 }
 707
 708 #define BP_GET_BUFC_TYPE(bp)                                            \
 709         (BP_IS_METADATA(bp) ? ARC_BUFC_METADATA : ARC_BUFC_DATA)
 710
 711 typedef enum spa_import_type {
 712         SPA_IMPORT_EXISTING,
 713         SPA_IMPORT_ASSEMBLE
 714 } spa_import_type_t;
 715
 716 typedef enum spa_mode {
 717         SPA_MODE_UNINIT = 0,
 718         SPA_MODE_READ = 1,
 719         SPA_MODE_WRITE = 2,
 720 } spa_mode_t;
 721
 722 /*
 723  * Send TRIM commands in-line during normal pool operation while deleting.
 724  *      OFF: no
 725  *      ON: yes
 726  */
 727 typedef enum {
 728         SPA_AUTOTRIM_OFF = 0,   /* default */
 729         SPA_AUTOTRIM_ON,
 730 } spa_autotrim_t;
 731
 732 /*
 733  * Reason TRIM command was issued, used internally for accounting purposes.
 734  */
 735 typedef enum trim_type {
 736         TRIM_TYPE_MANUAL = 0,
 737         TRIM_TYPE_AUTO = 1,
 738         TRIM_TYPE_SIMPLE = 2
 739 } trim_type_t;
 740
 741 /* state manipulation functions */
 742 extern int spa_open(const char *pool, spa_t **, const void *tag);
 743 extern int spa_open_rewind(const char *pool, spa_t **, const void *tag,
 744     nvlist_t *policy, nvlist_t **config);
 745 extern int spa_get_stats(const char *pool, nvlist_t **config, char *altroot,
 746     size_t buflen);
 747 extern int spa_create(const char *pool, nvlist_t *nvroot, nvlist_t *props,
 748     nvlist_t *zplprops, struct dsl_crypto_params *dcp);
 749 extern int spa_import(char *pool, nvlist_t *config, nvlist_t *props,
 750     uint64_t flags);
 751 extern nvlist_t *spa_tryimport(nvlist_t *tryconfig);
 752 extern int spa_destroy(const char *pool);
 753 extern int spa_checkpoint(const char *pool);
 754 extern int spa_checkpoint_discard(const char *pool);
 755 extern int spa_export(const char *pool, nvlist_t **oldconfig, boolean_t force,
 756     boolean_t hardforce);
 757 extern int spa_reset(const char *pool);
 758 extern void spa_async_request(spa_t *spa, int flag);
 759 extern void spa_async_unrequest(spa_t *spa, int flag);
 760 extern void spa_async_suspend(spa_t *spa);
 761 extern void spa_async_resume(spa_t *spa);
 762 extern int spa_async_tasks(spa_t *spa);
 763 extern spa_t *spa_inject_addref(char *pool);
 764 extern void spa_inject_delref(spa_t *spa);
 765 extern void spa_scan_stat_init(spa_t *spa);
 766 extern int spa_scan_get_stats(spa_t *spa, pool_scan_stat_t *ps);
 767 extern int bpobj_enqueue_alloc_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx);
 768 extern int bpobj_enqueue_free_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx);
 769
 770 #define SPA_ASYNC_CONFIG_UPDATE                 0x01
 771 #define SPA_ASYNC_REMOVE                        0x02
 772 #define SPA_ASYNC_PROBE                         0x04
 773 #define SPA_ASYNC_RESILVER_DONE                 0x08
 774 #define SPA_ASYNC_RESILVER                      0x10
 775 #define SPA_ASYNC_AUTOEXPAND                    0x20
 776 #define SPA_ASYNC_REMOVE_DONE                   0x40
 777 #define SPA_ASYNC_REMOVE_STOP                   0x80
 778 #define SPA_ASYNC_INITIALIZE_RESTART            0x100
 779 #define SPA_ASYNC_TRIM_RESTART                  0x200
 780 #define SPA_ASYNC_AUTOTRIM_RESTART              0x400
 781 #define SPA_ASYNC_L2CACHE_REBUILD               0x800
 782 #define SPA_ASYNC_L2CACHE_TRIM                  0x1000
 783 #define SPA_ASYNC_REBUILD_DONE                  0x2000
 784 #define SPA_ASYNC_DETACH_SPARE                  0x4000
 785
 786 /* device manipulation */
 787 extern int spa_vdev_add(spa_t *spa, nvlist_t *nvroot);
 788 extern int spa_vdev_attach(spa_t *spa, uint64_t guid, nvlist_t *nvroot,
 789     int replacing, int rebuild);
 790 extern int spa_vdev_detach(spa_t *spa, uint64_t guid, uint64_t pguid,
 791     int replace_done);
 792 extern int spa_vdev_alloc(spa_t *spa, uint64_t guid);
 793 extern int spa_vdev_noalloc(spa_t *spa, uint64_t guid);
 794 extern boolean_t spa_vdev_remove_active(spa_t *spa);
 795 extern int spa_vdev_initialize(spa_t *spa, nvlist_t *nv, uint64_t cmd_type,
 796     nvlist_t *vdev_errlist);
 797 extern int spa_vdev_trim(spa_t *spa, nvlist_t *nv, uint64_t cmd_type,
 798     uint64_t rate, boolean_t partial, boolean_t secure, nvlist_t *vdev_errlist);
 799 extern int spa_vdev_setpath(spa_t *spa, uint64_t guid, const char *newpath);
 800 extern int spa_vdev_setfru(spa_t *spa, uint64_t guid, const char *newfru);
 801 extern int spa_vdev_split_mirror(spa_t *spa, const char *newname,
 802     nvlist_t *config, nvlist_t *props, boolean_t exp);
 803
 804 /* spare state (which is global across all pools) */
 805 extern void spa_spare_add(vdev_t *vd);
 806 extern void spa_spare_remove(vdev_t *vd);
 807 extern boolean_t spa_spare_exists(uint64_t guid, uint64_t *pool, int *refcnt);
 808 extern void spa_spare_activate(vdev_t *vd);
 809
 810 /* L2ARC state (which is global across all pools) */
 811 extern void spa_l2cache_add(vdev_t *vd);
 812 extern void spa_l2cache_remove(vdev_t *vd);
 813 extern boolean_t spa_l2cache_exists(uint64_t guid, uint64_t *pool);
 814 extern void spa_l2cache_activate(vdev_t *vd);
 815 extern void spa_l2cache_drop(spa_t *spa);
 816
 817 /* scanning */
 818 extern int spa_scan(spa_t *spa, pool_scan_func_t func);
 819 extern int spa_scan_stop(spa_t *spa);
 820 extern int spa_scrub_pause_resume(spa_t *spa, pool_scrub_cmd_t flag);
 821
 822 /* spa syncing */
 823 extern void spa_sync(spa_t *spa, uint64_t txg); /* only for DMU use */
 824 extern void spa_sync_allpools(void);
 825
 826 extern uint_t zfs_sync_pass_deferred_free;
 827
 828 /* spa namespace global mutex */
 829 extern kmutex_t spa_namespace_lock;
 830
 831 /*
 832  * SPA configuration functions in spa_config.c
 833  */
 834
 835 #define SPA_CONFIG_UPDATE_POOL  0
 836 #define SPA_CONFIG_UPDATE_VDEVS 1
 837
 838 extern void spa_write_cachefile(spa_t *, boolean_t, boolean_t, boolean_t);
 839 extern void spa_config_load(void);
 840 extern int spa_all_configs(uint64_t *generation, nvlist_t **pools);
 841 extern void spa_config_set(spa_t *spa, nvlist_t *config);
 842 extern nvlist_t *spa_config_generate(spa_t *spa, vdev_t *vd, uint64_t txg,
 843     int getstats);
 844 extern void spa_config_update(spa_t *spa, int what);
 845 extern int spa_config_parse(spa_t *spa, vdev_t **vdp, nvlist_t *nv,
 846     vdev_t *parent, uint_t id, int atype);
 847
 848
 849 /*
 850  * Miscellaneous SPA routines in spa_misc.c
 851  */
 852
 853 /* Namespace manipulation */
 854 extern spa_t *spa_lookup(const char *name);
 855 extern spa_t *spa_add(const char *name, nvlist_t *config, const char *altroot);
 856 extern void spa_remove(spa_t *spa);
 857 extern spa_t *spa_next(spa_t *prev);
 858
 859 /* Refcount functions */
 860 extern void spa_open_ref(spa_t *spa, const void *tag);
 861 extern void spa_close(spa_t *spa, const void *tag);
 862 extern void spa_async_close(spa_t *spa, const void *tag);
 863 extern boolean_t spa_refcount_zero(spa_t *spa);
 864
 865 #define SCL_NONE        0x00
 866 #define SCL_CONFIG      0x01
 867 #define SCL_STATE       0x02
 868 #define SCL_L2ARC       0x04            /* hack until L2ARC 2.0 */
 869 #define SCL_ALLOC       0x08
 870 #define SCL_ZIO         0x10
 871 #define SCL_FREE        0x20
 872 #define SCL_VDEV        0x40
 873 #define SCL_LOCKS       7
 874 #define SCL_ALL         ((1 << SCL_LOCKS) - 1)
 875 #define SCL_STATE_ALL   (SCL_STATE | SCL_L2ARC | SCL_ZIO)
 876
 877 /* Historical pool statistics */
 878 typedef struct spa_history_kstat {
 879         kmutex_t                lock;
 880         uint64_t                count;
 881         uint64_t                size;
 882         kstat_t                 *kstat;
 883         void                    *priv;
 884         list_t                  list;
 885 } spa_history_kstat_t;
 886
 887 typedef struct spa_history_list {
 888         uint64_t                size;
 889         procfs_list_t           procfs_list;
 890 } spa_history_list_t;
 891
 892 typedef struct spa_stats {
 893         spa_history_list_t      read_history;
 894         spa_history_list_t      txg_history;
 895         spa_history_kstat_t     tx_assign_histogram;
 896         spa_history_list_t      mmp_history;
 897         spa_history_kstat_t     state;          /* pool state */
 898         spa_history_kstat_t     guid;           /* pool guid */
 899         spa_history_kstat_t     iostats;
 900 } spa_stats_t;
 901
 902 typedef enum txg_state {
 903         TXG_STATE_BIRTH         = 0,
 904         TXG_STATE_OPEN          = 1,
 905         TXG_STATE_QUIESCED      = 2,
 906         TXG_STATE_WAIT_FOR_SYNC = 3,
 907         TXG_STATE_SYNCED        = 4,
 908         TXG_STATE_COMMITTED     = 5,
 909 } txg_state_t;
 910
 911 typedef struct txg_stat {
 912         vdev_stat_t             vs1;
 913         vdev_stat_t             vs2;
 914         uint64_t                txg;
 915         uint64_t                ndirty;
 916 } txg_stat_t;
 917
 918 /* Assorted pool IO kstats */
 919 typedef struct spa_iostats {
 920         kstat_named_t   trim_extents_written;
 921         kstat_named_t   trim_bytes_written;
 922         kstat_named_t   trim_extents_skipped;
 923         kstat_named_t   trim_bytes_skipped;
 924         kstat_named_t   trim_extents_failed;
 925         kstat_named_t   trim_bytes_failed;
 926         kstat_named_t   autotrim_extents_written;
 927         kstat_named_t   autotrim_bytes_written;
 928         kstat_named_t   autotrim_extents_skipped;
 929         kstat_named_t   autotrim_bytes_skipped;
 930         kstat_named_t   autotrim_extents_failed;
 931         kstat_named_t   autotrim_bytes_failed;
 932         kstat_named_t   simple_trim_extents_written;
 933         kstat_named_t   simple_trim_bytes_written;
 934         kstat_named_t   simple_trim_extents_skipped;
 935         kstat_named_t   simple_trim_bytes_skipped;
 936         kstat_named_t   simple_trim_extents_failed;
 937         kstat_named_t   simple_trim_bytes_failed;
 938 } spa_iostats_t;
 939
 940 extern void spa_stats_init(spa_t *spa);
 941 extern void spa_stats_destroy(spa_t *spa);
 942 extern void spa_read_history_add(spa_t *spa, const zbookmark_phys_t *zb,
 943     uint32_t aflags);
 944 extern void spa_txg_history_add(spa_t *spa, uint64_t txg, hrtime_t birth_time);
 945 extern int spa_txg_history_set(spa_t *spa,  uint64_t txg,
 946     txg_state_t completed_state, hrtime_t completed_time);
 947 extern txg_stat_t *spa_txg_history_init_io(spa_t *, uint64_t,
 948     struct dsl_pool *);
 949 extern void spa_txg_history_fini_io(spa_t *, txg_stat_t *);
 950 extern void spa_tx_assign_add_nsecs(spa_t *spa, uint64_t nsecs);
 951 extern int spa_mmp_history_set_skip(spa_t *spa, uint64_t mmp_kstat_id);
 952 extern int spa_mmp_history_set(spa_t *spa, uint64_t mmp_kstat_id, int io_error,
 953     hrtime_t duration);
 954 extern void spa_mmp_history_add(spa_t *spa, uint64_t txg, uint64_t timestamp,
 955     uint64_t mmp_delay, vdev_t *vd, int label, uint64_t mmp_kstat_id,
 956     int error);
 957 extern void spa_iostats_trim_add(spa_t *spa, trim_type_t type,
 958     uint64_t extents_written, uint64_t bytes_written,
 959     uint64_t extents_skipped, uint64_t bytes_skipped,
 960     uint64_t extents_failed, uint64_t bytes_failed);
 961 extern void spa_import_progress_add(spa_t *spa);
 962 extern void spa_import_progress_remove(uint64_t spa_guid);
 963 extern int spa_import_progress_set_mmp_check(uint64_t pool_guid,
 964     uint64_t mmp_sec_remaining);
 965 extern int spa_import_progress_set_max_txg(uint64_t pool_guid,
 966     uint64_t max_txg);
 967 extern int spa_import_progress_set_state(uint64_t pool_guid,
 968     spa_load_state_t spa_load_state);
 969
 970 /* Pool configuration locks */
 971 extern int spa_config_tryenter(spa_t *spa, int locks, const void *tag,
 972     krw_t rw);
 973 extern void spa_config_enter(spa_t *spa, int locks, const void *tag, krw_t rw);
 974 extern void spa_config_enter_mmp(spa_t *spa, int locks, const void *tag,
 975     krw_t rw);
 976 extern void spa_config_exit(spa_t *spa, int locks, const void *tag);
 977 extern int spa_config_held(spa_t *spa, int locks, krw_t rw);
 978
 979 /* Pool vdev add/remove lock */
 980 extern uint64_t spa_vdev_enter(spa_t *spa);
 981 extern uint64_t spa_vdev_detach_enter(spa_t *spa, uint64_t guid);
 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, const 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 uint32_t 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_embedded_log_class(spa_t *spa);
1037 extern metaslab_class_t *spa_special_class(spa_t *spa);
1038 extern metaslab_class_t *spa_dedup_class(spa_t *spa);
1039 extern metaslab_class_t *spa_preferred_class(spa_t *spa, uint64_t size,
1040     dmu_object_type_t objtype, uint_t level, uint_t special_smallblk);
1041
1042 extern void spa_evicting_os_register(spa_t *, objset_t *os);
1043 extern void spa_evicting_os_deregister(spa_t *, objset_t *os);
1044 extern void spa_evicting_os_wait(spa_t *spa);
1045 extern int spa_max_replication(spa_t *spa);
1046 extern int spa_prev_software_version(spa_t *spa);
1047 extern uint64_t spa_get_failmode(spa_t *spa);
1048 extern uint64_t spa_get_deadman_failmode(spa_t *spa);
1049 extern void spa_set_deadman_failmode(spa_t *spa, const char *failmode);
1050 extern boolean_t spa_suspended(spa_t *spa);
1051 extern uint64_t spa_bootfs(spa_t *spa);
1052 extern uint64_t spa_delegation(spa_t *spa);
1053 extern objset_t *spa_meta_objset(spa_t *spa);
1054 extern space_map_t *spa_syncing_log_sm(spa_t *spa);
1055 extern uint64_t spa_deadman_synctime(spa_t *spa);
1056 extern uint64_t spa_deadman_ziotime(spa_t *spa);
1057 extern uint64_t spa_dirty_data(spa_t *spa);
1058 extern spa_autotrim_t spa_get_autotrim(spa_t *spa);
1059 extern int spa_get_allocator(spa_t *spa);
1060 extern void spa_set_allocator(spa_t *spa, const char *allocator);
1061
1062 /* Miscellaneous support routines */
1063 extern void spa_load_failed(spa_t *spa, const char *fmt, ...)
1064     __attribute__((format(printf, 2, 3)));
1065 extern void spa_load_note(spa_t *spa, const char *fmt, ...)
1066     __attribute__((format(printf, 2, 3)));
1067 extern void spa_activate_mos_feature(spa_t *spa, const char *feature,
1068     dmu_tx_t *tx);
1069 extern void spa_deactivate_mos_feature(spa_t *spa, const char *feature);
1070 extern spa_t *spa_by_guid(uint64_t pool_guid, uint64_t device_guid);
1071 extern boolean_t spa_guid_exists(uint64_t pool_guid, uint64_t device_guid);
1072 extern char *spa_strdup(const char *);
1073 extern void spa_strfree(char *);
1074 extern uint64_t spa_generate_guid(spa_t *spa);
1075 extern void snprintf_blkptr(char *buf, size_t buflen, const blkptr_t *bp);
1076 extern void spa_freeze(spa_t *spa);
1077 extern int spa_change_guid(spa_t *spa);
1078 extern void spa_upgrade(spa_t *spa, uint64_t version);
1079 extern void spa_evict_all(void);
1080 extern vdev_t *spa_lookup_by_guid(spa_t *spa, uint64_t guid,
1081     boolean_t l2cache);
1082 extern boolean_t spa_has_l2cache(spa_t *, uint64_t guid);
1083 extern boolean_t spa_has_spare(spa_t *, uint64_t guid);
1084 extern uint64_t dva_get_dsize_sync(spa_t *spa, const dva_t *dva);
1085 extern uint64_t bp_get_dsize_sync(spa_t *spa, const blkptr_t *bp);
1086 extern uint64_t bp_get_dsize(spa_t *spa, const blkptr_t *bp);
1087 extern boolean_t spa_has_slogs(spa_t *spa);
1088 extern boolean_t spa_is_root(spa_t *spa);
1089 extern boolean_t spa_writeable(spa_t *spa);
1090 extern boolean_t spa_has_pending_synctask(spa_t *spa);
1091 extern int spa_maxblocksize(spa_t *spa);
1092 extern int spa_maxdnodesize(spa_t *spa);
1093 extern boolean_t spa_has_checkpoint(spa_t *spa);
1094 extern boolean_t spa_importing_readonly_checkpoint(spa_t *spa);
1095 extern boolean_t spa_suspend_async_destroy(spa_t *spa);
1096 extern uint64_t spa_min_claim_txg(spa_t *spa);
1097 extern boolean_t zfs_dva_valid(spa_t *spa, const dva_t *dva,
1098     const blkptr_t *bp);
1099 typedef void (*spa_remap_cb_t)(uint64_t vdev, uint64_t offset, uint64_t size,
1100     void *arg);
1101 extern boolean_t spa_remap_blkptr(spa_t *spa, blkptr_t *bp,
1102     spa_remap_cb_t callback, void *arg);
1103 extern uint64_t spa_get_last_removal_txg(spa_t *spa);
1104 extern boolean_t spa_trust_config(spa_t *spa);
1105 extern uint64_t spa_missing_tvds_allowed(spa_t *spa);
1106 extern void spa_set_missing_tvds(spa_t *spa, uint64_t missing);
1107 extern boolean_t spa_top_vdevs_spacemap_addressable(spa_t *spa);
1108 extern uint64_t spa_total_metaslabs(spa_t *spa);
1109 extern boolean_t spa_multihost(spa_t *spa);
1110 extern uint32_t spa_get_hostid(spa_t *spa);
1111 extern void spa_activate_allocation_classes(spa_t *, dmu_tx_t *);
1112 extern boolean_t spa_livelist_delete_check(spa_t *spa);
1113
1114 extern spa_mode_t spa_mode(spa_t *spa);
1115 extern uint64_t zfs_strtonum(const char *str, char **nptr);
1116
1117 extern char *spa_his_ievent_table[];
1118
1119 extern void spa_history_create_obj(spa_t *spa, dmu_tx_t *tx);
1120 extern int spa_history_get(spa_t *spa, uint64_t *offset, uint64_t *len_read,
1121     char *his_buf);
1122 extern int spa_history_log(spa_t *spa, const char *his_buf);
1123 extern int spa_history_log_nvl(spa_t *spa, nvlist_t *nvl);
1124 extern void spa_history_log_version(spa_t *spa, const char *operation,
1125     dmu_tx_t *tx);
1126 extern void spa_history_log_internal(spa_t *spa, const char *operation,
1127     dmu_tx_t *tx, const char *fmt, ...) __printflike(4, 5);
1128 extern void spa_history_log_internal_ds(struct dsl_dataset *ds, const char *op,
1129     dmu_tx_t *tx, const char *fmt, ...)  __printflike(4, 5);
1130 extern void spa_history_log_internal_dd(dsl_dir_t *dd, const char *operation,
1131     dmu_tx_t *tx, const char *fmt, ...) __printflike(4, 5);
1132
1133 extern const char *spa_state_to_name(spa_t *spa);
1134
1135 /* error handling */
1136 struct zbookmark_phys;
1137 extern void spa_log_error(spa_t *spa, const zbookmark_phys_t *zb,
1138     const uint64_t *birth);
1139 extern void spa_remove_error(spa_t *spa, zbookmark_phys_t *zb,
1140     const uint64_t *birth);
1141 extern int zfs_ereport_post(const char *clazz, spa_t *spa, vdev_t *vd,
1142     const zbookmark_phys_t *zb, zio_t *zio, uint64_t state);
1143 extern boolean_t zfs_ereport_is_valid(const char *clazz, spa_t *spa, vdev_t *vd,
1144     zio_t *zio);
1145 extern void zfs_ereport_taskq_fini(void);
1146 extern void zfs_ereport_clear(spa_t *spa, vdev_t *vd);
1147 extern nvlist_t *zfs_event_create(spa_t *spa, vdev_t *vd, const char *type,
1148     const char *name, nvlist_t *aux);
1149 extern void zfs_post_remove(spa_t *spa, vdev_t *vd);
1150 extern void zfs_post_state_change(spa_t *spa, vdev_t *vd, uint64_t laststate);
1151 extern void zfs_post_autoreplace(spa_t *spa, vdev_t *vd);
1152 extern uint64_t spa_approx_errlog_size(spa_t *spa);
1153 extern int spa_get_errlog(spa_t *spa, void *uaddr, uint64_t *count);
1154 extern uint64_t spa_get_last_errlog_size(spa_t *spa);
1155 extern void spa_errlog_rotate(spa_t *spa);
1156 extern void spa_errlog_drain(spa_t *spa);
1157 extern void spa_errlog_sync(spa_t *spa, uint64_t txg);
1158 extern void spa_get_errlists(spa_t *spa, avl_tree_t *last, avl_tree_t *scrub);
1159 extern void spa_delete_dataset_errlog(spa_t *spa, uint64_t ds, dmu_tx_t *tx);
1160 extern void spa_swap_errlog(spa_t *spa, uint64_t new_head_ds,
1161     uint64_t old_head_ds, dmu_tx_t *tx);
1162 extern void sync_error_list(spa_t *spa, avl_tree_t *t, uint64_t *obj,
1163     dmu_tx_t *tx);
1164 extern void spa_upgrade_errlog(spa_t *spa, dmu_tx_t *tx);
1165 extern int find_top_affected_fs(spa_t *spa, uint64_t head_ds,
1166     zbookmark_err_phys_t *zep, uint64_t *top_affected_fs);
1167 extern int find_birth_txg(struct dsl_dataset *ds, zbookmark_err_phys_t *zep,
1168     uint64_t *birth_txg);
1169 extern void zep_to_zb(uint64_t dataset, zbookmark_err_phys_t *zep,
1170     zbookmark_phys_t *zb);
1171 extern void name_to_errphys(char *buf, zbookmark_err_phys_t *zep);
1172
1173 /* vdev mirror */
1174 extern void vdev_mirror_stat_init(void);
1175 extern void vdev_mirror_stat_fini(void);
1176
1177 /* Initialization and termination */
1178 extern void spa_init(spa_mode_t mode);
1179 extern void spa_fini(void);
1180 extern void spa_boot_init(void);
1181
1182 /* properties */
1183 extern int spa_prop_set(spa_t *spa, nvlist_t *nvp);
1184 extern int spa_prop_get(spa_t *spa, nvlist_t **nvp);
1185 extern void spa_prop_clear_bootfs(spa_t *spa, uint64_t obj, dmu_tx_t *tx);
1186 extern void spa_configfile_set(spa_t *, nvlist_t *, boolean_t);
1187
1188 /* asynchronous event notification */
1189 extern void spa_event_notify(spa_t *spa, vdev_t *vdev, nvlist_t *hist_nvl,
1190     const char *name);
1191 extern void zfs_ereport_zvol_post(const char *subclass, const char *name,
1192     const char *device_name, const char *raw_name);
1193
1194 /* waiting for pool activities to complete */
1195 extern int spa_wait(const char *pool, zpool_wait_activity_t activity,
1196     boolean_t *waited);
1197 extern int spa_wait_tag(const char *name, zpool_wait_activity_t activity,
1198     uint64_t tag, boolean_t *waited);
1199 extern void spa_notify_waiters(spa_t *spa);
1200 extern void spa_wake_waiters(spa_t *spa);
1201
1202 extern void spa_import_os(spa_t *spa);
1203 extern void spa_export_os(spa_t *spa);
1204 extern void spa_activate_os(spa_t *spa);
1205 extern void spa_deactivate_os(spa_t *spa);
1206
1207 /* module param call functions */
1208 int param_set_deadman_ziotime(ZFS_MODULE_PARAM_ARGS);
1209 int param_set_deadman_synctime(ZFS_MODULE_PARAM_ARGS);
1210 int param_set_slop_shift(ZFS_MODULE_PARAM_ARGS);
1211 int param_set_deadman_failmode(ZFS_MODULE_PARAM_ARGS);
1212 int param_set_active_allocator(ZFS_MODULE_PARAM_ARGS);
1213
1214 #ifdef ZFS_DEBUG
1215 #define dprintf_bp(bp, fmt, ...) do {                           \
1216         if (zfs_flags & ZFS_DEBUG_DPRINTF) {                    \
1217         char *__blkbuf = kmem_alloc(BP_SPRINTF_LEN, KM_SLEEP);  \
1218         snprintf_blkptr(__blkbuf, BP_SPRINTF_LEN, (bp));        \
1219         dprintf(fmt " %s\n", __VA_ARGS__, __blkbuf);            \
1220         kmem_free(__blkbuf, BP_SPRINTF_LEN);                    \
1221         } \
1222 } while (0)
1223 #else
1224 #define dprintf_bp(bp, fmt, ...)
1225 #endif
1226
1227 extern spa_mode_t spa_mode_global;
1228 extern int zfs_deadman_enabled;
1229 extern uint64_t zfs_deadman_synctime_ms;
1230 extern uint64_t zfs_deadman_ziotime_ms;
1231 extern uint64_t zfs_deadman_checktime_ms;
1232
1233 extern kmem_cache_t *zio_buf_cache[];
1234 extern kmem_cache_t *zio_data_buf_cache[];
1235
1236 #ifdef  __cplusplus
1237 }
1238 #endif
1239
1240 #endif  /* _SYS_SPA_H */