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, 2024 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/zfs_context.h>
  39 #include <sys/avl.h>
  40 #include <sys/kstat.h>
  41 #include <sys/nvpair.h>
  42 #include <sys/types.h>
  43 #include <sys/fs/zfs.h>
  44 #include <sys/spa_checksum.h>
  45 #include <sys/dmu.h>
  46 #include <sys/space_map.h>
  47 #include <sys/bitops.h>
  48
  49 #ifdef  __cplusplus
  50 extern "C" {
  51 #endif
  52
  53 /*
  54  * Forward references that lots of things need.
  55  */
  56 typedef struct spa spa_t;
  57 typedef struct vdev vdev_t;
  58 typedef struct metaslab metaslab_t;
  59 typedef struct metaslab_group metaslab_group_t;
  60 typedef struct metaslab_class metaslab_class_t;
  61 typedef struct zio zio_t;
  62 typedef struct zilog zilog_t;
  63 typedef struct spa_aux_vdev spa_aux_vdev_t;
  64 typedef struct zbookmark_phys zbookmark_phys_t;
  65 typedef struct zbookmark_err_phys zbookmark_err_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  * Alignment Shift (ashift) is an immutable, internal top-level vdev property
  75  * which can only be set at vdev creation time. Physical writes are always done
  76  * according to it, which makes 2^ashift the smallest possible IO on a vdev.
  77  *
  78  * We currently allow values ranging from 512 bytes (2^9 = 512) to 64 KiB
  79  * (2^16 = 65,536).
  80  */
  81 #define ASHIFT_MIN              9
  82 #define ASHIFT_MAX              16
  83
  84 /*
  85  * Size of block to hold the configuration data (a packed nvlist)
  86  */
  87 #define SPA_CONFIG_BLOCKSIZE    (1ULL << 14)
  88
  89 /*
  90  * The DVA size encodings for LSIZE and PSIZE support blocks up to 32MB.
  91  * The ASIZE encoding should be at least 64 times larger (6 more bits)
  92  * to support up to 4-way RAID-Z mirror mode with worst-case gang block
  93  * overhead, three DVAs per bp, plus one more bit in case we do anything
  94  * else that expands the ASIZE.
  95  */
  96 #define SPA_LSIZEBITS           16      /* LSIZE up to 32M (2^16 * 512) */
  97 #define SPA_PSIZEBITS           16      /* PSIZE up to 32M (2^16 * 512) */
  98 #define SPA_ASIZEBITS           24      /* ASIZE up to 64 times larger  */
  99
 100 #define SPA_COMPRESSBITS        7
 101 #define SPA_VDEVBITS            24
 102 #define SPA_COMPRESSMASK        ((1U << SPA_COMPRESSBITS) - 1)
 103
 104 /*
 105  * All SPA data is represented by 128-bit data virtual addresses (DVAs).
 106  * The members of the dva_t should be considered opaque outside the SPA.
 107  */
 108 typedef struct dva {
 109         uint64_t        dva_word[2];
 110 } dva_t;
 111
 112
 113 /*
 114  * Some checksums/hashes need a 256-bit initialization salt. This salt is kept
 115  * secret and is suitable for use in MAC algorithms as the key.
 116  */
 117 typedef struct zio_cksum_salt {
 118         uint8_t         zcs_bytes[32];
 119 } zio_cksum_salt_t;
 120
 121 /*
 122  * Each block is described by its DVAs, time of birth, checksum, etc.
 123  * The word-by-word, bit-by-bit layout of the blkptr is as follows:
 124  *
 125  *      64      56      48      40      32      24      16      8       0
 126  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 127  * 0    |  pad  |         vdev1         | pad   |         ASIZE         |
 128  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 129  * 1    |G|                      offset1                                |
 130  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 131  * 2    |  pad  |         vdev2         | pad   |         ASIZE         |
 132  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 133  * 3    |G|                      offset2                                |
 134  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 135  * 4    |  pad  |         vdev3         | pad   |         ASIZE         |
 136  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 137  * 5    |G|                      offset3                                |
 138  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 139  * 6    |BDX|lvl| type  | cksum |E| comp|    PSIZE      |     LSIZE     |
 140  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 141  * 7    |                       padding                                 |
 142  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 143  * 8    |                       padding                                 |
 144  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 145  * 9    |                       physical birth txg                      |
 146  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 147  * a    |                       logical birth txg                       |
 148  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 149  * b    |                       fill count                              |
 150  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 151  * c    |                       checksum[0]                             |
 152  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 153  * d    |                       checksum[1]                             |
 154  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 155  * e    |                       checksum[2]                             |
 156  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 157  * f    |                       checksum[3]                             |
 158  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 159  *
 160  * Legend:
 161  *
 162  * vdev         virtual device ID
 163  * offset       offset into virtual device
 164  * LSIZE        logical size
 165  * PSIZE        physical size (after compression)
 166  * ASIZE        allocated size (including RAID-Z parity and gang block headers)
 167  * cksum        checksum function
 168  * comp         compression function
 169  * G            gang block indicator
 170  * B            byteorder (endianness)
 171  * D            dedup
 172  * X            encryption
 173  * E            blkptr_t contains embedded data (see below)
 174  * lvl          level of indirection
 175  * type         DMU object type
 176  * phys birth   txg when dva[0] was written; zero if same as logical birth txg
 177  *              note that typically all the dva's would be written in this
 178  *              txg, but they could be different if they were moved by
 179  *              device removal.
 180  * log. birth   transaction group in which the block was logically born
 181  * fill count   number of non-zero blocks under this bp
 182  * checksum[4]  256-bit checksum of the data this bp describes
 183  */
 184
 185 /*
 186  * The blkptr_t's of encrypted blocks also need to store the encryption
 187  * parameters so that the block can be decrypted. This layout is as follows:
 188  *
 189  *      64      56      48      40      32      24      16      8       0
 190  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 191  * 0    |               vdev1           | pad   |         ASIZE         |
 192  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 193  * 1    |G|                      offset1                                |
 194  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 195  * 2    |               vdev2           | pad   |         ASIZE         |
 196  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 197  * 3    |G|                      offset2                                |
 198  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 199  * 4    |                       salt                                    |
 200  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 201  * 5    |                       IV1                                     |
 202  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 203  * 6    |BDX|lvl| type  | cksum |E| comp|    PSIZE      |     LSIZE     |
 204  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 205  * 7    |                       padding                                 |
 206  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 207  * 8    |                       padding                                 |
 208  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 209  * 9    |                       physical birth txg                      |
 210  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 211  * a    |                       logical birth txg                       |
 212  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 213  * b    |               IV2             |           fill count          |
 214  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 215  * c    |                       checksum[0]                             |
 216  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 217  * d    |                       checksum[1]                             |
 218  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 219  * e    |                       MAC[0]                                  |
 220  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 221  * f    |                       MAC[1]                                  |
 222  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 223  *
 224  * Legend:
 225  *
 226  * salt         Salt for generating encryption keys
 227  * IV1          First 64 bits of encryption IV
 228  * X            Block requires encryption handling (set to 1)
 229  * E            blkptr_t contains embedded data (set to 0, see below)
 230  * fill count   number of non-zero blocks under this bp (truncated to 32 bits)
 231  * IV2          Last 32 bits of encryption IV
 232  * checksum[2]  128-bit checksum of the data this bp describes
 233  * MAC[2]       128-bit message authentication code for this data
 234  *
 235  * The X bit being set indicates that this block is one of 3 types. If this is
 236  * a level 0 block with an encrypted object type, the block is encrypted
 237  * (see BP_IS_ENCRYPTED()). If this is a level 0 block with an unencrypted
 238  * object type, this block is authenticated with an HMAC (see
 239  * BP_IS_AUTHENTICATED()). Otherwise (if level > 0), this bp will use the MAC
 240  * words to store a checksum-of-MACs from the level below (see
 241  * BP_HAS_INDIRECT_MAC_CKSUM()). For convenience in the code, BP_IS_PROTECTED()
 242  * refers to both encrypted and authenticated blocks and BP_USES_CRYPT()
 243  * refers to any of these 3 kinds of blocks.
 244  *
 245  * The additional encryption parameters are the salt, IV, and MAC which are
 246  * explained in greater detail in the block comment at the top of zio_crypt.c.
 247  * The MAC occupies half of the checksum space since it serves a very similar
 248  * purpose: to prevent data corruption on disk. The only functional difference
 249  * is that the checksum is used to detect on-disk corruption whether or not the
 250  * encryption key is loaded and the MAC provides additional protection against
 251  * malicious disk tampering. We use the 3rd DVA to store the salt and first
 252  * 64 bits of the IV. As a result encrypted blocks can only have 2 copies
 253  * maximum instead of the normal 3. The last 32 bits of the IV are stored in
 254  * the upper bits of what is usually the fill count. Note that only blocks at
 255  * level 0 or -2 are ever encrypted, which allows us to guarantee that these
 256  * 32 bits are not trampled over by other code (see zio_crypt.c for details).
 257  * The salt and IV are not used for authenticated bps or bps with an indirect
 258  * MAC checksum, so these blocks can utilize all 3 DVAs and the full 64 bits
 259  * for the fill count.
 260  */
 261
 262 /*
 263  * "Embedded" blkptr_t's don't actually point to a block, instead they
 264  * have a data payload embedded in the blkptr_t itself.  See the comment
 265  * in blkptr.c for more details.
 266  *
 267  * The blkptr_t is laid out as follows:
 268  *
 269  *      64      56      48      40      32      24      16      8       0
 270  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 271  * 0    |      payload                                                  |
 272  * 1    |      payload                                                  |
 273  * 2    |      payload                                                  |
 274  * 3    |      payload                                                  |
 275  * 4    |      payload                                                  |
 276  * 5    |      payload                                                  |
 277  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 278  * 6    |BDX|lvl| type  | etype |E| comp| PSIZE|              LSIZE     |
 279  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 280  * 7    |      payload                                                  |
 281  * 8    |      payload                                                  |
 282  * 9    |      payload                                                  |
 283  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 284  * a    |                       logical birth txg                       |
 285  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 286  * b    |      payload                                                  |
 287  * c    |      payload                                                  |
 288  * d    |      payload                                                  |
 289  * e    |      payload                                                  |
 290  * f    |      payload                                                  |
 291  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 292  *
 293  * Legend:
 294  *
 295  * payload              contains the embedded data
 296  * B (byteorder)        byteorder (endianness)
 297  * D (dedup)            padding (set to zero)
 298  * X                    encryption (set to zero)
 299  * E (embedded)         set to one
 300  * lvl                  indirection level
 301  * type                 DMU object type
 302  * etype                how to interpret embedded data (BP_EMBEDDED_TYPE_*)
 303  * comp                 compression function of payload
 304  * PSIZE                size of payload after compression, in bytes
 305  * LSIZE                logical size of payload, in bytes
 306  *                      note that 25 bits is enough to store the largest
 307  *                      "normal" BP's LSIZE (2^16 * 2^9) in bytes
 308  * log. birth           transaction group in which the block was logically born
 309  *
 310  * Note that LSIZE and PSIZE are stored in bytes, whereas for non-embedded
 311  * bp's they are stored in units of SPA_MINBLOCKSHIFT.
 312  * Generally, the generic BP_GET_*() macros can be used on embedded BP's.
 313  * The B, D, X, lvl, type, and comp fields are stored the same as with normal
 314  * BP's so the BP_SET_* macros can be used with them.  etype, PSIZE, LSIZE must
 315  * be set with the BPE_SET_* macros.  BP_SET_EMBEDDED() should be called before
 316  * other macros, as they assert that they are only used on BP's of the correct
 317  * "embedded-ness". Encrypted blkptr_t's cannot be embedded because they use
 318  * the payload space for encryption parameters (see the comment above on
 319  * how encryption parameters are stored).
 320  */
 321
 322 #define BPE_GET_ETYPE(bp)       \
 323         (ASSERT(BP_IS_EMBEDDED(bp)), \
 324         BF64_GET((bp)->blk_prop, 40, 8))
 325 #define BPE_SET_ETYPE(bp, t)    do { \
 326         ASSERT(BP_IS_EMBEDDED(bp)); \
 327         BF64_SET((bp)->blk_prop, 40, 8, t); \
 328 } while (0)
 329
 330 #define BPE_GET_LSIZE(bp)       \
 331         (ASSERT(BP_IS_EMBEDDED(bp)), \
 332         BF64_GET_SB((bp)->blk_prop, 0, 25, 0, 1))
 333 #define BPE_SET_LSIZE(bp, x)    do { \
 334         ASSERT(BP_IS_EMBEDDED(bp)); \
 335         BF64_SET_SB((bp)->blk_prop, 0, 25, 0, 1, x); \
 336 } while (0)
 337
 338 #define BPE_GET_PSIZE(bp)       \
 339         (ASSERT(BP_IS_EMBEDDED(bp)), \
 340         BF64_GET_SB((bp)->blk_prop, 25, 7, 0, 1))
 341 #define BPE_SET_PSIZE(bp, x)    do { \
 342         ASSERT(BP_IS_EMBEDDED(bp)); \
 343         BF64_SET_SB((bp)->blk_prop, 25, 7, 0, 1, x); \
 344 } while (0)
 345
 346 typedef enum bp_embedded_type {
 347         BP_EMBEDDED_TYPE_DATA,
 348         BP_EMBEDDED_TYPE_RESERVED, /* Reserved for Delphix byteswap feature. */
 349         BP_EMBEDDED_TYPE_REDACTED,
 350         NUM_BP_EMBEDDED_TYPES
 351 } bp_embedded_type_t;
 352
 353 #define BPE_NUM_WORDS 14
 354 #define BPE_PAYLOAD_SIZE (BPE_NUM_WORDS * sizeof (uint64_t))
 355 #define BPE_IS_PAYLOADWORD(bp, wp) \
 356         ((wp) != &(bp)->blk_prop && (wp) != (&(bp)->blk_birth_word[1]))
 357
 358 #define SPA_BLKPTRSHIFT 7               /* blkptr_t is 128 bytes        */
 359 #define SPA_DVAS_PER_BP 3               /* Number of DVAs in a bp       */
 360 #define SPA_SYNC_MIN_VDEVS 3            /* min vdevs to update during sync */
 361
 362 /*
 363  * A block is a hole when it has either 1) never been written to, or
 364  * 2) is zero-filled. In both cases, ZFS can return all zeroes for all reads
 365  * without physically allocating disk space. Holes are represented in the
 366  * blkptr_t structure by zeroed blk_dva. Correct checking for holes is
 367  * done through the BP_IS_HOLE macro. For holes, the logical size, level,
 368  * DMU object type, and birth times are all also stored for holes that
 369  * were written to at some point (i.e. were punched after having been filled).
 370  */
 371 typedef struct blkptr {
 372         dva_t           blk_dva[SPA_DVAS_PER_BP]; /* Data Virtual Addresses */
 373         uint64_t        blk_prop;       /* size, compression, type, etc     */
 374         uint64_t        blk_pad[2];     /* Extra space for the future       */
 375         uint64_t        blk_birth_word[2];
 376         uint64_t        blk_fill;       /* fill count                       */
 377         zio_cksum_t     blk_cksum;      /* 256-bit checksum                 */
 378 } blkptr_t;
 379
 380 /*
 381  * Macros to get and set fields in a bp or DVA.
 382  */
 383
 384 /*
 385  * Note, for gang blocks, DVA_GET_ASIZE() is the total space allocated for
 386  * this gang DVA including its children BP's.  The space allocated at this
 387  * DVA's vdev/offset is vdev_gang_header_asize(vdev).
 388  */
 389 #define DVA_GET_ASIZE(dva)      \
 390         BF64_GET_SB((dva)->dva_word[0], 0, SPA_ASIZEBITS, SPA_MINBLOCKSHIFT, 0)
 391 #define DVA_SET_ASIZE(dva, x)   \
 392         BF64_SET_SB((dva)->dva_word[0], 0, SPA_ASIZEBITS, \
 393         SPA_MINBLOCKSHIFT, 0, x)
 394
 395 #define DVA_GET_VDEV(dva)       BF64_GET((dva)->dva_word[0], 32, SPA_VDEVBITS)
 396 #define DVA_SET_VDEV(dva, x)    \
 397         BF64_SET((dva)->dva_word[0], 32, SPA_VDEVBITS, x)
 398
 399 #define DVA_GET_OFFSET(dva)     \
 400         BF64_GET_SB((dva)->dva_word[1], 0, 63, SPA_MINBLOCKSHIFT, 0)
 401 #define DVA_SET_OFFSET(dva, x)  \
 402         BF64_SET_SB((dva)->dva_word[1], 0, 63, SPA_MINBLOCKSHIFT, 0, x)
 403
 404 #define DVA_GET_GANG(dva)       BF64_GET((dva)->dva_word[1], 63, 1)
 405 #define DVA_SET_GANG(dva, x)    BF64_SET((dva)->dva_word[1], 63, 1, x)
 406
 407 #define BP_GET_LSIZE(bp)        \
 408         (BP_IS_EMBEDDED(bp) ?   \
 409         (BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA ? BPE_GET_LSIZE(bp) : 0): \
 410         BF64_GET_SB((bp)->blk_prop, 0, SPA_LSIZEBITS, SPA_MINBLOCKSHIFT, 1))
 411 #define BP_SET_LSIZE(bp, x)     do { \
 412         ASSERT(!BP_IS_EMBEDDED(bp)); \
 413         BF64_SET_SB((bp)->blk_prop, \
 414             0, SPA_LSIZEBITS, SPA_MINBLOCKSHIFT, 1, x); \
 415 } while (0)
 416
 417 #define BP_GET_PSIZE(bp)        \
 418         (BP_IS_EMBEDDED(bp) ? 0 : \
 419         BF64_GET_SB((bp)->blk_prop, 16, SPA_PSIZEBITS, SPA_MINBLOCKSHIFT, 1))
 420 #define BP_SET_PSIZE(bp, x)     do { \
 421         ASSERT(!BP_IS_EMBEDDED(bp)); \
 422         BF64_SET_SB((bp)->blk_prop, \
 423             16, SPA_PSIZEBITS, SPA_MINBLOCKSHIFT, 1, x); \
 424 } while (0)
 425
 426 #define BP_GET_COMPRESS(bp)             \
 427         BF64_GET((bp)->blk_prop, 32, SPA_COMPRESSBITS)
 428 #define BP_SET_COMPRESS(bp, x)          \
 429         BF64_SET((bp)->blk_prop, 32, SPA_COMPRESSBITS, x)
 430
 431 #define BP_IS_EMBEDDED(bp)              BF64_GET((bp)->blk_prop, 39, 1)
 432 #define BP_SET_EMBEDDED(bp, x)          BF64_SET((bp)->blk_prop, 39, 1, x)
 433
 434 #define BP_GET_CHECKSUM(bp)             \
 435         (BP_IS_EMBEDDED(bp) ? ZIO_CHECKSUM_OFF : \
 436         BF64_GET((bp)->blk_prop, 40, 8))
 437 #define BP_SET_CHECKSUM(bp, x)          do { \
 438         ASSERT(!BP_IS_EMBEDDED(bp)); \
 439         BF64_SET((bp)->blk_prop, 40, 8, x); \
 440 } while (0)
 441
 442 #define BP_GET_TYPE(bp)                 BF64_GET((bp)->blk_prop, 48, 8)
 443 #define BP_SET_TYPE(bp, x)              BF64_SET((bp)->blk_prop, 48, 8, x)
 444
 445 #define BP_GET_LEVEL(bp)                BF64_GET((bp)->blk_prop, 56, 5)
 446 #define BP_SET_LEVEL(bp, x)             BF64_SET((bp)->blk_prop, 56, 5, x)
 447
 448 /* encrypted, authenticated, and MAC cksum bps use the same bit */
 449 #define BP_USES_CRYPT(bp)               BF64_GET((bp)->blk_prop, 61, 1)
 450 #define BP_SET_CRYPT(bp, x)             BF64_SET((bp)->blk_prop, 61, 1, x)
 451
 452 #define BP_IS_ENCRYPTED(bp)                     \
 453         (BP_USES_CRYPT(bp) &&                   \
 454         BP_GET_LEVEL(bp) <= 0 &&                \
 455         DMU_OT_IS_ENCRYPTED(BP_GET_TYPE(bp)))
 456
 457 #define BP_IS_AUTHENTICATED(bp)                 \
 458         (BP_USES_CRYPT(bp) &&                   \
 459         BP_GET_LEVEL(bp) <= 0 &&                \
 460         !DMU_OT_IS_ENCRYPTED(BP_GET_TYPE(bp)))
 461
 462 #define BP_HAS_INDIRECT_MAC_CKSUM(bp)           \
 463         (BP_USES_CRYPT(bp) && BP_GET_LEVEL(bp) > 0)
 464
 465 #define BP_IS_PROTECTED(bp)                     \
 466         (BP_IS_ENCRYPTED(bp) || BP_IS_AUTHENTICATED(bp))
 467
 468 #define BP_GET_DEDUP(bp)                BF64_GET((bp)->blk_prop, 62, 1)
 469 #define BP_SET_DEDUP(bp, x)             BF64_SET((bp)->blk_prop, 62, 1, x)
 470
 471 #define BP_GET_BYTEORDER(bp)            BF64_GET((bp)->blk_prop, 63, 1)
 472 #define BP_SET_BYTEORDER(bp, x)         BF64_SET((bp)->blk_prop, 63, 1, x)
 473
 474 #define BP_GET_FREE(bp)                 BF64_GET((bp)->blk_fill, 0, 1)
 475 #define BP_SET_FREE(bp, x)              BF64_SET((bp)->blk_fill, 0, 1, x)
 476
 477 #define BP_GET_LOGICAL_BIRTH(bp)        (bp)->blk_birth_word[1]
 478 #define BP_SET_LOGICAL_BIRTH(bp, x)     ((bp)->blk_birth_word[1] = (x))
 479
 480 #define BP_GET_PHYSICAL_BIRTH(bp)       (bp)->blk_birth_word[0]
 481 #define BP_SET_PHYSICAL_BIRTH(bp, x)    ((bp)->blk_birth_word[0] = (x))
 482
 483 #define BP_GET_BIRTH(bp)                                        \
 484         (BP_IS_EMBEDDED(bp) ? 0 :                               \
 485         BP_GET_PHYSICAL_BIRTH(bp) ? BP_GET_PHYSICAL_BIRTH(bp) : \
 486         BP_GET_LOGICAL_BIRTH(bp))
 487
 488 #define BP_SET_BIRTH(bp, logical, physical)     \
 489 {                                               \
 490         ASSERT(!BP_IS_EMBEDDED(bp));            \
 491         BP_SET_LOGICAL_BIRTH(bp, logical);      \
 492         BP_SET_PHYSICAL_BIRTH(bp,               \
 493             ((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_GET_BIRTH(bp1) == BP_GET_BIRTH(bp2) &&      \
 547         BP_GET_LOGICAL_BIRTH(bp1) == BP_GET_LOGICAL_BIRTH(bp2) &&       \
 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_DVAS(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 }
 584
 585 #define BP_ZERO(bp)                             \
 586 {                                               \
 587         BP_ZERO_DVAS(bp);                       \
 588         (bp)->blk_prop = 0;                     \
 589         (bp)->blk_pad[0] = 0;                   \
 590         (bp)->blk_pad[1] = 0;                   \
 591         (bp)->blk_birth_word[0] = 0;            \
 592         (bp)->blk_birth_word[1] = 0;            \
 593         (bp)->blk_fill = 0;                     \
 594         ZIO_SET_CHECKSUM(&(bp)->blk_cksum, 0, 0, 0, 0); \
 595 }
 596
 597 #ifdef _ZFS_BIG_ENDIAN
 598 #define ZFS_HOST_BYTEORDER      (0ULL)
 599 #else
 600 #define ZFS_HOST_BYTEORDER      (1ULL)
 601 #endif
 602
 603 #define BP_SHOULD_BYTESWAP(bp)  (BP_GET_BYTEORDER(bp) != ZFS_HOST_BYTEORDER)
 604
 605 #define BP_SPRINTF_LEN  400
 606
 607 /*
 608  * This macro allows code sharing between zfs, libzpool, and mdb.
 609  * 'func' is either kmem_scnprintf() or mdb_snprintf().
 610  * 'ws' (whitespace) can be ' ' for single-line format, '\n' for multi-line.
 611  */
 612
 613 #define SNPRINTF_BLKPTR(func, ws, buf, size, bp, type, checksum, compress) \
 614 {                                                                       \
 615         static const char *const copyname[] =                           \
 616             { "zero", "single", "double", "triple" };                   \
 617         int len = 0;                                                    \
 618         int copies = 0;                                                 \
 619         const char *crypt_type;                                         \
 620         if (bp != NULL) {                                               \
 621                 if (BP_IS_ENCRYPTED(bp)) {                              \
 622                         crypt_type = "encrypted";                       \
 623                         /* LINTED E_SUSPICIOUS_COMPARISON */            \
 624                 } else if (BP_IS_AUTHENTICATED(bp)) {                   \
 625                         crypt_type = "authenticated";                   \
 626                 } else if (BP_HAS_INDIRECT_MAC_CKSUM(bp)) {             \
 627                         crypt_type = "indirect-MAC";                    \
 628                 } else {                                                \
 629                         crypt_type = "unencrypted";                     \
 630                 }                                                       \
 631         }                                                               \
 632         if (bp == NULL) {                                               \
 633                 len += func(buf + len, size - len, "<NULL>");           \
 634         } else if (BP_IS_HOLE(bp)) {                                    \
 635                 len += func(buf + len, size - len,                      \
 636                     "HOLE [L%llu %s] "                                  \
 637                     "size=%llxL birth=%lluL",                           \
 638                     (u_longlong_t)BP_GET_LEVEL(bp),                     \
 639                     type,                                               \
 640                     (u_longlong_t)BP_GET_LSIZE(bp),                     \
 641                     (u_longlong_t)BP_GET_LOGICAL_BIRTH(bp));            \
 642         } else if (BP_IS_EMBEDDED(bp)) {                                \
 643                 len = func(buf + len, size - len,                       \
 644                     "EMBEDDED [L%llu %s] et=%u %s "                     \
 645                     "size=%llxL/%llxP birth=%lluL",                     \
 646                     (u_longlong_t)BP_GET_LEVEL(bp),                     \
 647                     type,                                               \
 648                     (int)BPE_GET_ETYPE(bp),                             \
 649                     compress,                                           \
 650                     (u_longlong_t)BPE_GET_LSIZE(bp),                    \
 651                     (u_longlong_t)BPE_GET_PSIZE(bp),                    \
 652                     (u_longlong_t)BP_GET_LOGICAL_BIRTH(bp));            \
 653         } else if (BP_IS_REDACTED(bp)) {                                \
 654                 len += func(buf + len, size - len,                      \
 655                     "REDACTED [L%llu %s] size=%llxL birth=%lluL",       \
 656                     (u_longlong_t)BP_GET_LEVEL(bp),                     \
 657                     type,                                               \
 658                     (u_longlong_t)BP_GET_LSIZE(bp),                     \
 659                     (u_longlong_t)BP_GET_LOGICAL_BIRTH(bp));            \
 660         } else {                                                        \
 661                 for (int d = 0; d < BP_GET_NDVAS(bp); d++) {            \
 662                         const dva_t *dva = &bp->blk_dva[d];             \
 663                         if (DVA_IS_VALID(dva))                          \
 664                                 copies++;                               \
 665                         len += func(buf + len, size - len,              \
 666                             "DVA[%d]=<%llu:%llx:%llx>%c", d,            \
 667                             (u_longlong_t)DVA_GET_VDEV(dva),            \
 668                             (u_longlong_t)DVA_GET_OFFSET(dva),          \
 669                             (u_longlong_t)DVA_GET_ASIZE(dva),           \
 670                             ws);                                        \
 671                 }                                                       \
 672                 ASSERT3S(copies, >, 0);                                 \
 673                 if (BP_IS_ENCRYPTED(bp)) {                              \
 674                         len += func(buf + len, size - len,              \
 675                             "salt=%llx iv=%llx:%llx%c",                 \
 676                             (u_longlong_t)bp->blk_dva[2].dva_word[0],   \
 677                             (u_longlong_t)bp->blk_dva[2].dva_word[1],   \
 678                             (u_longlong_t)BP_GET_IV2(bp),               \
 679                             ws);                                        \
 680                 }                                                       \
 681                 if (BP_IS_GANG(bp) &&                                   \
 682                     DVA_GET_ASIZE(&bp->blk_dva[2]) <=                   \
 683                     DVA_GET_ASIZE(&bp->blk_dva[1]) / 2)                 \
 684                         copies--;                                       \
 685                 len += func(buf + len, size - len,                      \
 686                     "[L%llu %s] %s %s %s %s %s %s %s%c"                 \
 687                     "size=%llxL/%llxP birth=%lluL/%lluP fill=%llu%c"    \
 688                     "cksum=%016llx:%016llx:%016llx:%016llx",            \
 689                     (u_longlong_t)BP_GET_LEVEL(bp),                     \
 690                     type,                                               \
 691                     checksum,                                           \
 692                     compress,                                           \
 693                     crypt_type,                                         \
 694                     BP_GET_BYTEORDER(bp) == 0 ? "BE" : "LE",            \
 695                     BP_IS_GANG(bp) ? "gang" : "contiguous",             \
 696                     BP_GET_DEDUP(bp) ? "dedup" : "unique",              \
 697                     copyname[copies],                                   \
 698                     ws,                                                 \
 699                     (u_longlong_t)BP_GET_LSIZE(bp),                     \
 700                     (u_longlong_t)BP_GET_PSIZE(bp),                     \
 701                     (u_longlong_t)BP_GET_LOGICAL_BIRTH(bp),             \
 702                     (u_longlong_t)BP_GET_BIRTH(bp),                     \
 703                     (u_longlong_t)BP_GET_FILL(bp),                      \
 704                     ws,                                                 \
 705                     (u_longlong_t)bp->blk_cksum.zc_word[0],             \
 706                     (u_longlong_t)bp->blk_cksum.zc_word[1],             \
 707                     (u_longlong_t)bp->blk_cksum.zc_word[2],             \
 708                     (u_longlong_t)bp->blk_cksum.zc_word[3]);            \
 709         }                                                               \
 710         ASSERT(len < size);                                             \
 711 }
 712
 713 #define BP_GET_BUFC_TYPE(bp)                                            \
 714         (BP_IS_METADATA(bp) ? ARC_BUFC_METADATA : ARC_BUFC_DATA)
 715
 716 typedef enum spa_import_type {
 717         SPA_IMPORT_EXISTING,
 718         SPA_IMPORT_ASSEMBLE
 719 } spa_import_type_t;
 720
 721 typedef enum spa_mode {
 722         SPA_MODE_UNINIT = 0,
 723         SPA_MODE_READ = 1,
 724         SPA_MODE_WRITE = 2,
 725 } spa_mode_t;
 726
 727 /*
 728  * Send TRIM commands in-line during normal pool operation while deleting.
 729  *      OFF: no
 730  *      ON: yes
 731  */
 732 typedef enum {
 733         SPA_AUTOTRIM_OFF = 0,   /* default */
 734         SPA_AUTOTRIM_ON,
 735 } spa_autotrim_t;
 736
 737 /*
 738  * Reason TRIM command was issued, used internally for accounting purposes.
 739  */
 740 typedef enum trim_type {
 741         TRIM_TYPE_MANUAL = 0,
 742         TRIM_TYPE_AUTO = 1,
 743         TRIM_TYPE_SIMPLE = 2
 744 } trim_type_t;
 745
 746 /* state manipulation functions */
 747 extern int spa_open(const char *pool, spa_t **, const void *tag);
 748 extern int spa_open_rewind(const char *pool, spa_t **, const void *tag,
 749     nvlist_t *policy, nvlist_t **config);
 750 extern int spa_get_stats(const char *pool, nvlist_t **config, char *altroot,
 751     size_t buflen);
 752 extern int spa_create(const char *pool, nvlist_t *nvroot, nvlist_t *props,
 753     nvlist_t *zplprops, struct dsl_crypto_params *dcp);
 754 extern int spa_import(char *pool, nvlist_t *config, nvlist_t *props,
 755     uint64_t flags);
 756 extern nvlist_t *spa_tryimport(nvlist_t *tryconfig);
 757 extern int spa_destroy(const char *pool);
 758 extern int spa_checkpoint(const char *pool);
 759 extern int spa_checkpoint_discard(const char *pool);
 760 extern int spa_export(const char *pool, nvlist_t **oldconfig, boolean_t force,
 761     boolean_t hardforce);
 762 extern int spa_reset(const char *pool);
 763 extern void spa_async_request(spa_t *spa, int flag);
 764 extern void spa_async_unrequest(spa_t *spa, int flag);
 765 extern void spa_async_suspend(spa_t *spa);
 766 extern void spa_async_resume(spa_t *spa);
 767 extern int spa_async_tasks(spa_t *spa);
 768 extern spa_t *spa_inject_addref(char *pool);
 769 extern void spa_inject_delref(spa_t *spa);
 770 extern void spa_scan_stat_init(spa_t *spa);
 771 extern int spa_scan_get_stats(spa_t *spa, pool_scan_stat_t *ps);
 772 extern int bpobj_enqueue_alloc_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx);
 773 extern int bpobj_enqueue_free_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx);
 774
 775 #define SPA_ASYNC_CONFIG_UPDATE                 0x01
 776 #define SPA_ASYNC_REMOVE                        0x02
 777 #define SPA_ASYNC_FAULT_VDEV                    0x04
 778 #define SPA_ASYNC_RESILVER_DONE                 0x08
 779 #define SPA_ASYNC_RESILVER                      0x10
 780 #define SPA_ASYNC_AUTOEXPAND                    0x20
 781 #define SPA_ASYNC_REMOVE_DONE                   0x40
 782 #define SPA_ASYNC_REMOVE_STOP                   0x80
 783 #define SPA_ASYNC_INITIALIZE_RESTART            0x100
 784 #define SPA_ASYNC_TRIM_RESTART                  0x200
 785 #define SPA_ASYNC_AUTOTRIM_RESTART              0x400
 786 #define SPA_ASYNC_L2CACHE_REBUILD               0x800
 787 #define SPA_ASYNC_L2CACHE_TRIM                  0x1000
 788 #define SPA_ASYNC_REBUILD_DONE                  0x2000
 789 #define SPA_ASYNC_DETACH_SPARE                  0x4000
 790
 791 /* device manipulation */
 792 extern int spa_vdev_add(spa_t *spa, nvlist_t *nvroot, boolean_t ashift_check);
 793 extern int spa_vdev_attach(spa_t *spa, uint64_t guid, nvlist_t *nvroot,
 794     int replacing, int rebuild);
 795 extern int spa_vdev_detach(spa_t *spa, uint64_t guid, uint64_t pguid,
 796     int replace_done);
 797 extern int spa_vdev_alloc(spa_t *spa, uint64_t guid);
 798 extern int spa_vdev_noalloc(spa_t *spa, uint64_t guid);
 799 extern boolean_t spa_vdev_remove_active(spa_t *spa);
 800 extern int spa_vdev_initialize(spa_t *spa, nvlist_t *nv, uint64_t cmd_type,
 801     nvlist_t *vdev_errlist);
 802 extern int spa_vdev_trim(spa_t *spa, nvlist_t *nv, uint64_t cmd_type,
 803     uint64_t rate, boolean_t partial, boolean_t secure, nvlist_t *vdev_errlist);
 804 extern int spa_vdev_setpath(spa_t *spa, uint64_t guid, const char *newpath);
 805 extern int spa_vdev_setfru(spa_t *spa, uint64_t guid, const char *newfru);
 806 extern int spa_vdev_split_mirror(spa_t *spa, const char *newname,
 807     nvlist_t *config, nvlist_t *props, boolean_t exp);
 808
 809 /* spare state (which is global across all pools) */
 810 extern void spa_spare_add(vdev_t *vd);
 811 extern void spa_spare_remove(vdev_t *vd);
 812 extern boolean_t spa_spare_exists(uint64_t guid, uint64_t *pool, int *refcnt);
 813 extern void spa_spare_activate(vdev_t *vd);
 814
 815 /* L2ARC state (which is global across all pools) */
 816 extern void spa_l2cache_add(vdev_t *vd);
 817 extern void spa_l2cache_remove(vdev_t *vd);
 818 extern boolean_t spa_l2cache_exists(uint64_t guid, uint64_t *pool);
 819 extern void spa_l2cache_activate(vdev_t *vd);
 820 extern void spa_l2cache_drop(spa_t *spa);
 821
 822 /* scanning */
 823 extern int spa_scan(spa_t *spa, pool_scan_func_t func);
 824 extern int spa_scan_stop(spa_t *spa);
 825 extern int spa_scrub_pause_resume(spa_t *spa, pool_scrub_cmd_t flag);
 826
 827 /* spa syncing */
 828 extern void spa_sync(spa_t *spa, uint64_t txg); /* only for DMU use */
 829 extern void spa_sync_allpools(void);
 830
 831 extern uint_t zfs_sync_pass_deferred_free;
 832
 833 /* spa sync taskqueues */
 834 taskq_t *spa_sync_tq_create(spa_t *spa, const char *name);
 835 void spa_sync_tq_destroy(spa_t *spa);
 836 uint_t spa_acq_allocator(spa_t *spa);
 837 void spa_rel_allocator(spa_t *spa, uint_t allocator);
 838 void spa_select_allocator(zio_t *zio);
 839
 840 /* spa namespace global mutex */
 841 extern kmutex_t spa_namespace_lock;
 842 extern avl_tree_t spa_namespace_avl;
 843 extern kcondvar_t spa_namespace_cv;
 844
 845 /*
 846  * SPA configuration functions in spa_config.c
 847  */
 848
 849 #define SPA_CONFIG_UPDATE_POOL  0
 850 #define SPA_CONFIG_UPDATE_VDEVS 1
 851
 852 extern void spa_write_cachefile(spa_t *, boolean_t, boolean_t, boolean_t);
 853 extern void spa_config_load(void);
 854 extern int spa_all_configs(uint64_t *generation, nvlist_t **pools);
 855 extern void spa_config_set(spa_t *spa, nvlist_t *config);
 856 extern nvlist_t *spa_config_generate(spa_t *spa, vdev_t *vd, uint64_t txg,
 857     int getstats);
 858 extern void spa_config_update(spa_t *spa, int what);
 859 extern int spa_config_parse(spa_t *spa, vdev_t **vdp, nvlist_t *nv,
 860     vdev_t *parent, uint_t id, int atype);
 861
 862
 863 /*
 864  * Miscellaneous SPA routines in spa_misc.c
 865  */
 866
 867 /* Namespace manipulation */
 868 extern spa_t *spa_lookup(const char *name);
 869 extern spa_t *spa_add(const char *name, nvlist_t *config, const char *altroot);
 870 extern void spa_remove(spa_t *spa);
 871 extern spa_t *spa_next(spa_t *prev);
 872
 873 /* Refcount functions */
 874 extern void spa_open_ref(spa_t *spa, const void *tag);
 875 extern void spa_close(spa_t *spa, const void *tag);
 876 extern void spa_async_close(spa_t *spa, const void *tag);
 877 extern boolean_t spa_refcount_zero(spa_t *spa);
 878
 879 #define SCL_NONE        0x00
 880 #define SCL_CONFIG      0x01
 881 #define SCL_STATE       0x02
 882 #define SCL_L2ARC       0x04            /* hack until L2ARC 2.0 */
 883 #define SCL_ALLOC       0x08
 884 #define SCL_ZIO         0x10
 885 #define SCL_FREE        0x20
 886 #define SCL_VDEV        0x40
 887 #define SCL_LOCKS       7
 888 #define SCL_ALL         ((1 << SCL_LOCKS) - 1)
 889 #define SCL_STATE_ALL   (SCL_STATE | SCL_L2ARC | SCL_ZIO)
 890
 891 /* Historical pool statistics */
 892 typedef struct spa_history_kstat {
 893         kmutex_t                lock;
 894         uint64_t                count;
 895         uint64_t                size;
 896         kstat_t                 *kstat;
 897         void                    *priv;
 898         list_t                  list;
 899 } spa_history_kstat_t;
 900
 901 typedef struct spa_history_list {
 902         uint64_t                size;
 903         procfs_list_t           procfs_list;
 904 } spa_history_list_t;
 905
 906 typedef struct spa_stats {
 907         spa_history_list_t      read_history;
 908         spa_history_list_t      txg_history;
 909         spa_history_kstat_t     tx_assign_histogram;
 910         spa_history_list_t      mmp_history;
 911         spa_history_kstat_t     state;          /* pool state */
 912         spa_history_kstat_t     guid;           /* pool guid */
 913         spa_history_kstat_t     iostats;
 914 } spa_stats_t;
 915
 916 typedef enum txg_state {
 917         TXG_STATE_BIRTH         = 0,
 918         TXG_STATE_OPEN          = 1,
 919         TXG_STATE_QUIESCED      = 2,
 920         TXG_STATE_WAIT_FOR_SYNC = 3,
 921         TXG_STATE_SYNCED        = 4,
 922         TXG_STATE_COMMITTED     = 5,
 923 } txg_state_t;
 924
 925 typedef struct txg_stat {
 926         vdev_stat_t             vs1;
 927         vdev_stat_t             vs2;
 928         uint64_t                txg;
 929         uint64_t                ndirty;
 930 } txg_stat_t;
 931
 932 /* Assorted pool IO kstats */
 933 typedef struct spa_iostats {
 934         kstat_named_t   trim_extents_written;
 935         kstat_named_t   trim_bytes_written;
 936         kstat_named_t   trim_extents_skipped;
 937         kstat_named_t   trim_bytes_skipped;
 938         kstat_named_t   trim_extents_failed;
 939         kstat_named_t   trim_bytes_failed;
 940         kstat_named_t   autotrim_extents_written;
 941         kstat_named_t   autotrim_bytes_written;
 942         kstat_named_t   autotrim_extents_skipped;
 943         kstat_named_t   autotrim_bytes_skipped;
 944         kstat_named_t   autotrim_extents_failed;
 945         kstat_named_t   autotrim_bytes_failed;
 946         kstat_named_t   simple_trim_extents_written;
 947         kstat_named_t   simple_trim_bytes_written;
 948         kstat_named_t   simple_trim_extents_skipped;
 949         kstat_named_t   simple_trim_bytes_skipped;
 950         kstat_named_t   simple_trim_extents_failed;
 951         kstat_named_t   simple_trim_bytes_failed;
 952         kstat_named_t   arc_read_count;
 953         kstat_named_t   arc_read_bytes;
 954         kstat_named_t   arc_write_count;
 955         kstat_named_t   arc_write_bytes;
 956         kstat_named_t   direct_read_count;
 957         kstat_named_t   direct_read_bytes;
 958         kstat_named_t   direct_write_count;
 959         kstat_named_t   direct_write_bytes;
 960 } spa_iostats_t;
 961
 962 extern void spa_stats_init(spa_t *spa);
 963 extern void spa_stats_destroy(spa_t *spa);
 964 extern void spa_read_history_add(spa_t *spa, const zbookmark_phys_t *zb,
 965     uint32_t aflags);
 966 extern void spa_txg_history_add(spa_t *spa, uint64_t txg, hrtime_t birth_time);
 967 extern int spa_txg_history_set(spa_t *spa,  uint64_t txg,
 968     txg_state_t completed_state, hrtime_t completed_time);
 969 extern txg_stat_t *spa_txg_history_init_io(spa_t *, uint64_t,
 970     struct dsl_pool *);
 971 extern void spa_txg_history_fini_io(spa_t *, txg_stat_t *);
 972 extern void spa_tx_assign_add_nsecs(spa_t *spa, uint64_t nsecs);
 973 extern int spa_mmp_history_set_skip(spa_t *spa, uint64_t mmp_kstat_id);
 974 extern int spa_mmp_history_set(spa_t *spa, uint64_t mmp_kstat_id, int io_error,
 975     hrtime_t duration);
 976 extern void spa_mmp_history_add(spa_t *spa, uint64_t txg, uint64_t timestamp,
 977     uint64_t mmp_delay, vdev_t *vd, int label, uint64_t mmp_kstat_id,
 978     int error);
 979 extern void spa_iostats_trim_add(spa_t *spa, trim_type_t type,
 980     uint64_t extents_written, uint64_t bytes_written,
 981     uint64_t extents_skipped, uint64_t bytes_skipped,
 982     uint64_t extents_failed, uint64_t bytes_failed);
 983 extern void spa_iostats_read_add(spa_t *spa, uint64_t size, uint64_t iops,
 984     uint32_t flags);
 985 extern void spa_iostats_write_add(spa_t *spa, uint64_t size, uint64_t iops,
 986     uint32_t flags);
 987 extern void spa_import_progress_add(spa_t *spa);
 988 extern void spa_import_progress_remove(uint64_t spa_guid);
 989 extern int spa_import_progress_set_mmp_check(uint64_t pool_guid,
 990     uint64_t mmp_sec_remaining);
 991 extern int spa_import_progress_set_max_txg(uint64_t pool_guid,
 992     uint64_t max_txg);
 993 extern int spa_import_progress_set_state(uint64_t pool_guid,
 994     spa_load_state_t spa_load_state);
 995 extern void spa_import_progress_set_notes(spa_t *spa,
 996     const char *fmt, ...) __printflike(2, 3);
 997 extern void spa_import_progress_set_notes_nolog(spa_t *spa,
 998     const char *fmt, ...) __printflike(2, 3);
 999
1000 /* Pool configuration locks */
1001 extern int spa_config_tryenter(spa_t *spa, int locks, const void *tag,
1002     krw_t rw);
1003 extern void spa_config_enter(spa_t *spa, int locks, const void *tag, krw_t rw);
1004 extern void spa_config_enter_mmp(spa_t *spa, int locks, const void *tag,
1005     krw_t rw);
1006 extern void spa_config_exit(spa_t *spa, int locks, const void *tag);
1007 extern int spa_config_held(spa_t *spa, int locks, krw_t rw);
1008
1009 /* Pool vdev add/remove lock */
1010 extern uint64_t spa_vdev_enter(spa_t *spa);
1011 extern uint64_t spa_vdev_detach_enter(spa_t *spa, uint64_t guid);
1012 extern uint64_t spa_vdev_config_enter(spa_t *spa);
1013 extern void spa_vdev_config_exit(spa_t *spa, vdev_t *vd, uint64_t txg,
1014     int error, const char *tag);
1015 extern int spa_vdev_exit(spa_t *spa, vdev_t *vd, uint64_t txg, int error);
1016
1017 /* Pool vdev state change lock */
1018 extern void spa_vdev_state_enter(spa_t *spa, int oplock);
1019 extern int spa_vdev_state_exit(spa_t *spa, vdev_t *vd, int error);
1020
1021 /* Log state */
1022 typedef enum spa_log_state {
1023         SPA_LOG_UNKNOWN = 0,    /* unknown log state */
1024         SPA_LOG_MISSING,        /* missing log(s) */
1025         SPA_LOG_CLEAR,          /* clear the log(s) */
1026         SPA_LOG_GOOD,           /* log(s) are good */
1027 } spa_log_state_t;
1028
1029 extern spa_log_state_t spa_get_log_state(spa_t *spa);
1030 extern void spa_set_log_state(spa_t *spa, spa_log_state_t state);
1031 extern int spa_reset_logs(spa_t *spa);
1032
1033 /* Log claim callback */
1034 extern void spa_claim_notify(zio_t *zio);
1035 extern void spa_deadman(void *);
1036
1037 /* Accessor functions */
1038 extern boolean_t spa_shutting_down(spa_t *spa);
1039 extern struct dsl_pool *spa_get_dsl(spa_t *spa);
1040 extern boolean_t spa_is_initializing(spa_t *spa);
1041 extern boolean_t spa_indirect_vdevs_loaded(spa_t *spa);
1042 extern blkptr_t *spa_get_rootblkptr(spa_t *spa);
1043 extern void spa_set_rootblkptr(spa_t *spa, const blkptr_t *bp);
1044 extern void spa_altroot(spa_t *, char *, size_t);
1045 extern uint32_t spa_sync_pass(spa_t *spa);
1046 extern char *spa_name(spa_t *spa);
1047 extern uint64_t spa_guid(spa_t *spa);
1048 extern uint64_t spa_load_guid(spa_t *spa);
1049 extern uint64_t spa_last_synced_txg(spa_t *spa);
1050 extern uint64_t spa_first_txg(spa_t *spa);
1051 extern uint64_t spa_syncing_txg(spa_t *spa);
1052 extern uint64_t spa_final_dirty_txg(spa_t *spa);
1053 extern uint64_t spa_version(spa_t *spa);
1054 extern pool_state_t spa_state(spa_t *spa);
1055 extern spa_load_state_t spa_load_state(spa_t *spa);
1056 extern uint64_t spa_freeze_txg(spa_t *spa);
1057 extern uint64_t spa_get_worst_case_asize(spa_t *spa, uint64_t lsize);
1058 extern uint64_t spa_get_dspace(spa_t *spa);
1059 extern uint64_t spa_get_checkpoint_space(spa_t *spa);
1060 extern uint64_t spa_get_slop_space(spa_t *spa);
1061 extern void spa_update_dspace(spa_t *spa);
1062 extern uint64_t spa_version(spa_t *spa);
1063 extern boolean_t spa_deflate(spa_t *spa);
1064 extern metaslab_class_t *spa_normal_class(spa_t *spa);
1065 extern metaslab_class_t *spa_log_class(spa_t *spa);
1066 extern metaslab_class_t *spa_embedded_log_class(spa_t *spa);
1067 extern metaslab_class_t *spa_special_class(spa_t *spa);
1068 extern metaslab_class_t *spa_dedup_class(spa_t *spa);
1069 extern metaslab_class_t *spa_preferred_class(spa_t *spa, const zio_t *zio);
1070 extern boolean_t spa_special_has_ddt(spa_t *spa);
1071
1072 extern void spa_evicting_os_register(spa_t *, objset_t *os);
1073 extern void spa_evicting_os_deregister(spa_t *, objset_t *os);
1074 extern void spa_evicting_os_wait(spa_t *spa);
1075 extern int spa_max_replication(spa_t *spa);
1076 extern int spa_prev_software_version(spa_t *spa);
1077 extern uint64_t spa_get_failmode(spa_t *spa);
1078 extern uint64_t spa_get_deadman_failmode(spa_t *spa);
1079 extern void spa_set_deadman_failmode(spa_t *spa, const char *failmode);
1080 extern boolean_t spa_suspended(spa_t *spa);
1081 extern uint64_t spa_bootfs(spa_t *spa);
1082 extern uint64_t spa_delegation(spa_t *spa);
1083 extern objset_t *spa_meta_objset(spa_t *spa);
1084 extern space_map_t *spa_syncing_log_sm(spa_t *spa);
1085 extern uint64_t spa_deadman_synctime(spa_t *spa);
1086 extern uint64_t spa_deadman_ziotime(spa_t *spa);
1087 extern uint64_t spa_dirty_data(spa_t *spa);
1088 extern spa_autotrim_t spa_get_autotrim(spa_t *spa);
1089 extern int spa_get_allocator(spa_t *spa);
1090 extern void spa_set_allocator(spa_t *spa, const char *allocator);
1091
1092 /* Miscellaneous support routines */
1093 extern void spa_load_failed(spa_t *spa, const char *fmt, ...)
1094     __attribute__((format(printf, 2, 3)));
1095 extern void spa_load_note(spa_t *spa, const char *fmt, ...)
1096     __attribute__((format(printf, 2, 3)));
1097 extern void spa_activate_mos_feature(spa_t *spa, const char *feature,
1098     dmu_tx_t *tx);
1099 extern void spa_deactivate_mos_feature(spa_t *spa, const char *feature);
1100 extern spa_t *spa_by_guid(uint64_t pool_guid, uint64_t device_guid);
1101 extern boolean_t spa_guid_exists(uint64_t pool_guid, uint64_t device_guid);
1102 extern char *spa_strdup(const char *);
1103 extern void spa_strfree(char *);
1104 extern uint64_t spa_generate_guid(spa_t *spa);
1105 extern void snprintf_blkptr(char *buf, size_t buflen, const blkptr_t *bp);
1106 extern void spa_freeze(spa_t *spa);
1107 extern int spa_change_guid(spa_t *spa, const uint64_t *guidp);
1108 extern void spa_upgrade(spa_t *spa, uint64_t version);
1109 extern void spa_evict_all(void);
1110 extern vdev_t *spa_lookup_by_guid(spa_t *spa, uint64_t guid,
1111     boolean_t l2cache);
1112 extern boolean_t spa_has_l2cache(spa_t *, uint64_t guid);
1113 extern boolean_t spa_has_spare(spa_t *, uint64_t guid);
1114 extern uint64_t dva_get_dsize_sync(spa_t *spa, const dva_t *dva);
1115 extern uint64_t bp_get_dsize_sync(spa_t *spa, const blkptr_t *bp);
1116 extern uint64_t bp_get_dsize(spa_t *spa, const blkptr_t *bp);
1117 extern boolean_t spa_has_slogs(spa_t *spa);
1118 extern boolean_t spa_is_root(spa_t *spa);
1119 extern boolean_t spa_writeable(spa_t *spa);
1120 extern boolean_t spa_has_pending_synctask(spa_t *spa);
1121 extern int spa_maxblocksize(spa_t *spa);
1122 extern int spa_maxdnodesize(spa_t *spa);
1123 extern boolean_t spa_has_checkpoint(spa_t *spa);
1124 extern boolean_t spa_importing_readonly_checkpoint(spa_t *spa);
1125 extern boolean_t spa_suspend_async_destroy(spa_t *spa);
1126 extern uint64_t spa_min_claim_txg(spa_t *spa);
1127 extern boolean_t zfs_dva_valid(spa_t *spa, const dva_t *dva,
1128     const blkptr_t *bp);
1129 typedef void (*spa_remap_cb_t)(uint64_t vdev, uint64_t offset, uint64_t size,
1130     void *arg);
1131 extern boolean_t spa_remap_blkptr(spa_t *spa, blkptr_t *bp,
1132     spa_remap_cb_t callback, void *arg);
1133 extern uint64_t spa_get_last_removal_txg(spa_t *spa);
1134 extern boolean_t spa_trust_config(spa_t *spa);
1135 extern uint64_t spa_missing_tvds_allowed(spa_t *spa);
1136 extern void spa_set_missing_tvds(spa_t *spa, uint64_t missing);
1137 extern boolean_t spa_top_vdevs_spacemap_addressable(spa_t *spa);
1138 extern uint64_t spa_total_metaslabs(spa_t *spa);
1139 extern boolean_t spa_multihost(spa_t *spa);
1140 extern uint32_t spa_get_hostid(spa_t *spa);
1141 extern void spa_activate_allocation_classes(spa_t *, dmu_tx_t *);
1142 extern boolean_t spa_livelist_delete_check(spa_t *spa);
1143
1144 extern boolean_t spa_mmp_remote_host_activity(spa_t *spa);
1145
1146 extern spa_mode_t spa_mode(spa_t *spa);
1147 extern uint64_t zfs_strtonum(const char *str, char **nptr);
1148
1149 extern char *spa_his_ievent_table[];
1150
1151 extern void spa_history_create_obj(spa_t *spa, dmu_tx_t *tx);
1152 extern int spa_history_get(spa_t *spa, uint64_t *offset, uint64_t *len_read,
1153     char *his_buf);
1154 extern int spa_history_log(spa_t *spa, const char *his_buf);
1155 extern int spa_history_log_nvl(spa_t *spa, nvlist_t *nvl);
1156 extern void spa_history_log_version(spa_t *spa, const char *operation,
1157     dmu_tx_t *tx);
1158 extern void spa_history_log_internal(spa_t *spa, const char *operation,
1159     dmu_tx_t *tx, const char *fmt, ...) __printflike(4, 5);
1160 extern void spa_history_log_internal_ds(struct dsl_dataset *ds, const char *op,
1161     dmu_tx_t *tx, const char *fmt, ...)  __printflike(4, 5);
1162 extern void spa_history_log_internal_dd(dsl_dir_t *dd, const char *operation,
1163     dmu_tx_t *tx, const char *fmt, ...) __printflike(4, 5);
1164
1165 extern const char *spa_state_to_name(spa_t *spa);
1166
1167 /* error handling */
1168 struct zbookmark_phys;
1169 extern void spa_log_error(spa_t *spa, const zbookmark_phys_t *zb,
1170     const uint64_t birth);
1171 extern void spa_remove_error(spa_t *spa, zbookmark_phys_t *zb,
1172     uint64_t birth);
1173 extern int zfs_ereport_post(const char *clazz, spa_t *spa, vdev_t *vd,
1174     const zbookmark_phys_t *zb, zio_t *zio, uint64_t state);
1175 extern boolean_t zfs_ereport_is_valid(const char *clazz, spa_t *spa, vdev_t *vd,
1176     zio_t *zio);
1177 extern void zfs_ereport_taskq_fini(void);
1178 extern void zfs_ereport_clear(spa_t *spa, vdev_t *vd);
1179 extern nvlist_t *zfs_event_create(spa_t *spa, vdev_t *vd, const char *type,
1180     const char *name, nvlist_t *aux);
1181 extern void zfs_post_remove(spa_t *spa, vdev_t *vd);
1182 extern void zfs_post_state_change(spa_t *spa, vdev_t *vd, uint64_t laststate);
1183 extern void zfs_post_autoreplace(spa_t *spa, vdev_t *vd);
1184 extern uint64_t spa_approx_errlog_size(spa_t *spa);
1185 extern int spa_get_errlog(spa_t *spa, void *uaddr, uint64_t *count);
1186 extern uint64_t spa_get_last_errlog_size(spa_t *spa);
1187 extern void spa_errlog_rotate(spa_t *spa);
1188 extern void spa_errlog_drain(spa_t *spa);
1189 extern void spa_errlog_sync(spa_t *spa, uint64_t txg);
1190 extern void spa_get_errlists(spa_t *spa, avl_tree_t *last, avl_tree_t *scrub);
1191 extern void spa_delete_dataset_errlog(spa_t *spa, uint64_t ds, dmu_tx_t *tx);
1192 extern void spa_swap_errlog(spa_t *spa, uint64_t new_head_ds,
1193     uint64_t old_head_ds, dmu_tx_t *tx);
1194 extern void sync_error_list(spa_t *spa, avl_tree_t *t, uint64_t *obj,
1195     dmu_tx_t *tx);
1196 extern void spa_upgrade_errlog(spa_t *spa, dmu_tx_t *tx);
1197 extern int find_top_affected_fs(spa_t *spa, uint64_t head_ds,
1198     zbookmark_err_phys_t *zep, uint64_t *top_affected_fs);
1199 extern int find_birth_txg(struct dsl_dataset *ds, zbookmark_err_phys_t *zep,
1200     uint64_t *birth_txg);
1201 extern void zep_to_zb(uint64_t dataset, zbookmark_err_phys_t *zep,
1202     zbookmark_phys_t *zb);
1203 extern void name_to_errphys(char *buf, zbookmark_err_phys_t *zep);
1204
1205 /* vdev mirror */
1206 extern void vdev_mirror_stat_init(void);
1207 extern void vdev_mirror_stat_fini(void);
1208
1209 /* Initialization and termination */
1210 extern void spa_init(spa_mode_t mode);
1211 extern void spa_fini(void);
1212 extern void spa_boot_init(void);
1213
1214 /* properties */
1215 extern int spa_prop_set(spa_t *spa, nvlist_t *nvp);
1216 extern int spa_prop_get(spa_t *spa, nvlist_t *nvp);
1217 extern int spa_prop_get_nvlist(spa_t *spa, char **props,
1218     unsigned int n_props, nvlist_t *outnvl);
1219 extern void spa_prop_clear_bootfs(spa_t *spa, uint64_t obj, dmu_tx_t *tx);
1220 extern void spa_configfile_set(spa_t *, nvlist_t *, boolean_t);
1221
1222 /* asynchronous event notification */
1223 extern void spa_event_notify(spa_t *spa, vdev_t *vdev, nvlist_t *hist_nvl,
1224     const char *name);
1225 extern void zfs_ereport_zvol_post(const char *subclass, const char *name,
1226     const char *device_name, const char *raw_name);
1227
1228 /* waiting for pool activities to complete */
1229 extern int spa_wait(const char *pool, zpool_wait_activity_t activity,
1230     boolean_t *waited);
1231 extern int spa_wait_tag(const char *name, zpool_wait_activity_t activity,
1232     uint64_t tag, boolean_t *waited);
1233 extern void spa_notify_waiters(spa_t *spa);
1234 extern void spa_wake_waiters(spa_t *spa);
1235
1236 extern void spa_import_os(spa_t *spa);
1237 extern void spa_export_os(spa_t *spa);
1238 extern void spa_activate_os(spa_t *spa);
1239 extern void spa_deactivate_os(spa_t *spa);
1240
1241 /* module param call functions */
1242 int param_set_deadman_ziotime(ZFS_MODULE_PARAM_ARGS);
1243 int param_set_deadman_synctime(ZFS_MODULE_PARAM_ARGS);
1244 int param_set_slop_shift(ZFS_MODULE_PARAM_ARGS);
1245 int param_set_deadman_failmode(ZFS_MODULE_PARAM_ARGS);
1246 int param_set_active_allocator(ZFS_MODULE_PARAM_ARGS);
1247
1248 #ifdef ZFS_DEBUG
1249 #define dprintf_bp(bp, fmt, ...) do {                           \
1250         if (zfs_flags & ZFS_DEBUG_DPRINTF) {                    \
1251         char *__blkbuf = kmem_alloc(BP_SPRINTF_LEN, KM_SLEEP);  \
1252         snprintf_blkptr(__blkbuf, BP_SPRINTF_LEN, (bp));        \
1253         dprintf(fmt " %s\n", __VA_ARGS__, __blkbuf);            \
1254         kmem_free(__blkbuf, BP_SPRINTF_LEN);                    \
1255         } \
1256 } while (0)
1257 #else
1258 #define dprintf_bp(bp, fmt, ...)
1259 #endif
1260
1261 extern spa_mode_t spa_mode_global;
1262 extern int zfs_deadman_enabled;
1263 extern uint64_t zfs_deadman_synctime_ms;
1264 extern uint64_t zfs_deadman_ziotime_ms;
1265 extern uint64_t zfs_deadman_checktime_ms;
1266
1267 extern kmem_cache_t *zio_buf_cache[];
1268 extern kmem_cache_t *zio_data_buf_cache[];
1269
1270 #ifdef  __cplusplus
1271 }
1272 #endif
1273
1274 #endif  /* _SYS_SPA_H */