include/sys/spa.h

   1 /*
   2  * CDDL HEADER START
   3  *
   4  * The contents of this file are subject to the terms of the
   5  * Common Development and Distribution License (the "License").
   6  * You may not use this file except in compliance with the License.
   7  *
   8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
   9  * or http://www.opensolaris.org/os/licensing.
  10  * See the License for the specific language governing permissions
  11  * and limitations under the License.
  12  *
  13  * When distributing Covered Code, include this CDDL HEADER in each
  14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
  15  * If applicable, add the following below this CDDL HEADER, with the
  16  * fields enclosed by brackets "[]" replaced with your own identifying
  17  * information: Portions Copyright [yyyy] [name of copyright owner]
  18  *
  19  * CDDL HEADER END
  20  */
  21 /*
  22  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
  23  * Copyright (c) 2011, 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, 2019, Datto Inc. All rights reserved.
  30  * Copyright (c) 2017, Intel Corporation.
  31  * Copyright (c) 2019, Allan Jude
  32  * Copyright (c) 2019, Klara 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
  69 struct bpobj;
  70 struct bplist;
  71 struct dsl_pool;
  72 struct dsl_dataset;
  73 struct dsl_crypto_params;
  74
  75 /*
  76  * We currently support block sizes from 512 bytes to 16MB.
  77  * The benefits of larger blocks, and thus larger IO, need to be weighed
  78  * against the cost of COWing a giant block to modify one byte, and the
  79  * large latency of reading or writing a large block.
  80  *
  81  * Note that although blocks up to 16MB are supported, the recordsize
  82  * property can not be set larger than zfs_max_recordsize (default 1MB).
  83  * See the comment near zfs_max_recordsize in dsl_dataset.c for details.
  84  *
  85  * Note that although the LSIZE field of the blkptr_t can store sizes up
  86  * to 32MB, the dnode's dn_datablkszsec can only store sizes up to
  87  * 32MB - 512 bytes.  Therefore, we limit SPA_MAXBLOCKSIZE to 16MB.
  88  */
  89 #define SPA_MINBLOCKSHIFT       9
  90 #define SPA_OLD_MAXBLOCKSHIFT   17
  91 #define SPA_MAXBLOCKSHIFT       24
  92 #define SPA_MINBLOCKSIZE        (1ULL << SPA_MINBLOCKSHIFT)
  93 #define SPA_OLD_MAXBLOCKSIZE    (1ULL << SPA_OLD_MAXBLOCKSHIFT)
  94 #define SPA_MAXBLOCKSIZE        (1ULL << SPA_MAXBLOCKSHIFT)
  95
  96 /*
  97  * Alignment Shift (ashift) is an immutable, internal top-level vdev property
  98  * which can only be set at vdev creation time. Physical writes are always done
  99  * according to it, which makes 2^ashift the smallest possible IO on a vdev.
 100  *
 101  * We currently allow values ranging from 512 bytes (2^9 = 512) to 64 KiB
 102  * (2^16 = 65,536).
 103  */
 104 #define ASHIFT_MIN              9
 105 #define ASHIFT_MAX              16
 106
 107 /*
 108  * Size of block to hold the configuration data (a packed nvlist)
 109  */
 110 #define SPA_CONFIG_BLOCKSIZE    (1ULL << 14)
 111
 112 /*
 113  * The DVA size encodings for LSIZE and PSIZE support blocks up to 32MB.
 114  * The ASIZE encoding should be at least 64 times larger (6 more bits)
 115  * to support up to 4-way RAID-Z mirror mode with worst-case gang block
 116  * overhead, three DVAs per bp, plus one more bit in case we do anything
 117  * else that expands the ASIZE.
 118  */
 119 #define SPA_LSIZEBITS           16      /* LSIZE up to 32M (2^16 * 512) */
 120 #define SPA_PSIZEBITS           16      /* PSIZE up to 32M (2^16 * 512) */
 121 #define SPA_ASIZEBITS           24      /* ASIZE up to 64 times larger  */
 122
 123 #define SPA_COMPRESSBITS        7
 124 #define SPA_VDEVBITS            24
 125 #define SPA_COMPRESSMASK        ((1U << SPA_COMPRESSBITS) - 1)
 126
 127 /*
 128  * All SPA data is represented by 128-bit data virtual addresses (DVAs).
 129  * The members of the dva_t should be considered opaque outside the SPA.
 130  */
 131 typedef struct dva {
 132         uint64_t        dva_word[2];
 133 } dva_t;
 134
 135
 136 /*
 137  * Some checksums/hashes need a 256-bit initialization salt. This salt is kept
 138  * secret and is suitable for use in MAC algorithms as the key.
 139  */
 140 typedef struct zio_cksum_salt {
 141         uint8_t         zcs_bytes[32];
 142 } zio_cksum_salt_t;
 143
 144 /*
 145  * Each block is described by its DVAs, time of birth, checksum, etc.
 146  * The word-by-word, bit-by-bit layout of the blkptr is as follows:
 147  *
 148  *      64      56      48      40      32      24      16      8       0
 149  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 150  * 0    |  pad  |         vdev1         | GRID  |         ASIZE         |
 151  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 152  * 1    |G|                      offset1                                |
 153  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 154  * 2    |  pad  |         vdev2         | GRID  |         ASIZE         |
 155  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 156  * 3    |G|                      offset2                                |
 157  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 158  * 4    |  pad  |         vdev3         | GRID  |         ASIZE         |
 159  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 160  * 5    |G|                      offset3                                |
 161  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 162  * 6    |BDX|lvl| type  | cksum |E| comp|    PSIZE      |     LSIZE     |
 163  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 164  * 7    |                       padding                                 |
 165  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 166  * 8    |                       padding                                 |
 167  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 168  * 9    |                       physical birth txg                      |
 169  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 170  * a    |                       logical birth txg                       |
 171  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 172  * b    |                       fill count                              |
 173  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 174  * c    |                       checksum[0]                             |
 175  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 176  * d    |                       checksum[1]                             |
 177  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 178  * e    |                       checksum[2]                             |
 179  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 180  * f    |                       checksum[3]                             |
 181  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 182  *
 183  * Legend:
 184  *
 185  * vdev         virtual device ID
 186  * offset       offset into virtual device
 187  * LSIZE        logical size
 188  * PSIZE        physical size (after compression)
 189  * ASIZE        allocated size (including RAID-Z parity and gang block headers)
 190  * GRID         RAID-Z layout information (reserved for future use)
 191  * cksum        checksum function
 192  * comp         compression function
 193  * G            gang block indicator
 194  * B            byteorder (endianness)
 195  * D            dedup
 196  * X            encryption
 197  * E            blkptr_t contains embedded data (see below)
 198  * lvl          level of indirection
 199  * type         DMU object type
 200  * phys birth   txg when dva[0] was written; zero if same as logical birth txg
 201  *              note that typically all the dva's would be written in this
 202  *              txg, but they could be different if they were moved by
 203  *              device removal.
 204  * log. birth   transaction group in which the block was logically born
 205  * fill count   number of non-zero blocks under this bp
 206  * checksum[4]  256-bit checksum of the data this bp describes
 207  */
 208
 209 /*
 210  * The blkptr_t's of encrypted blocks also need to store the encryption
 211  * parameters so that the block can be decrypted. This layout is as follows:
 212  *
 213  *      64      56      48      40      32      24      16      8       0
 214  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 215  * 0    |               vdev1           | GRID  |         ASIZE         |
 216  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 217  * 1    |G|                      offset1                                |
 218  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 219  * 2    |               vdev2           | GRID  |         ASIZE         |
 220  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 221  * 3    |G|                      offset2                                |
 222  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 223  * 4    |                       salt                                    |
 224  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 225  * 5    |                       IV1                                     |
 226  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 227  * 6    |BDX|lvl| type  | cksum |E| comp|    PSIZE      |     LSIZE     |
 228  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 229  * 7    |                       padding                                 |
 230  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 231  * 8    |                       padding                                 |
 232  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 233  * 9    |                       physical birth txg                      |
 234  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 235  * a    |                       logical birth txg                       |
 236  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 237  * b    |               IV2             |           fill count          |
 238  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 239  * c    |                       checksum[0]                             |
 240  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 241  * d    |                       checksum[1]                             |
 242  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 243  * e    |                       MAC[0]                                  |
 244  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 245  * f    |                       MAC[1]                                  |
 246  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 247  *
 248  * Legend:
 249  *
 250  * salt         Salt for generating encryption keys
 251  * IV1          First 64 bits of encryption IV
 252  * X            Block requires encryption handling (set to 1)
 253  * E            blkptr_t contains embedded data (set to 0, see below)
 254  * fill count   number of non-zero blocks under this bp (truncated to 32 bits)
 255  * IV2          Last 32 bits of encryption IV
 256  * checksum[2]  128-bit checksum of the data this bp describes
 257  * MAC[2]       128-bit message authentication code for this data
 258  *
 259  * The X bit being set indicates that this block is one of 3 types. If this is
 260  * a level 0 block with an encrypted object type, the block is encrypted
 261  * (see BP_IS_ENCRYPTED()). If this is a level 0 block with an unencrypted
 262  * object type, this block is authenticated with an HMAC (see
 263  * BP_IS_AUTHENTICATED()). Otherwise (if level > 0), this bp will use the MAC
 264  * words to store a checksum-of-MACs from the level below (see
 265  * BP_HAS_INDIRECT_MAC_CKSUM()). For convenience in the code, BP_IS_PROTECTED()
 266  * refers to both encrypted and authenticated blocks and BP_USES_CRYPT()
 267  * refers to any of these 3 kinds of blocks.
 268  *
 269  * The additional encryption parameters are the salt, IV, and MAC which are
 270  * explained in greater detail in the block comment at the top of zio_crypt.c.
 271  * The MAC occupies half of the checksum space since it serves a very similar
 272  * purpose: to prevent data corruption on disk. The only functional difference
 273  * is that the checksum is used to detect on-disk corruption whether or not the
 274  * encryption key is loaded and the MAC provides additional protection against
 275  * malicious disk tampering. We use the 3rd DVA to store the salt and first
 276  * 64 bits of the IV. As a result encrypted blocks can only have 2 copies
 277  * maximum instead of the normal 3. The last 32 bits of the IV are stored in
 278  * the upper bits of what is usually the fill count. Note that only blocks at
 279  * level 0 or -2 are ever encrypted, which allows us to guarantee that these
 280  * 32 bits are not trampled over by other code (see zio_crypt.c for details).
 281  * The salt and IV are not used for authenticated bps or bps with an indirect
 282  * MAC checksum, so these blocks can utilize all 3 DVAs and the full 64 bits
 283  * for the fill count.
 284  */
 285
 286 /*
 287  * "Embedded" blkptr_t's don't actually point to a block, instead they
 288  * have a data payload embedded in the blkptr_t itself.  See the comment
 289  * in blkptr.c for more details.
 290  *
 291  * The blkptr_t is laid out as follows:
 292  *
 293  *      64      56      48      40      32      24      16      8       0
 294  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 295  * 0    |      payload                                                  |
 296  * 1    |      payload                                                  |
 297  * 2    |      payload                                                  |
 298  * 3    |      payload                                                  |
 299  * 4    |      payload                                                  |
 300  * 5    |      payload                                                  |
 301  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 302  * 6    |BDX|lvl| type  | etype |E| comp| PSIZE|              LSIZE     |
 303  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 304  * 7    |      payload                                                  |
 305  * 8    |      payload                                                  |
 306  * 9    |      payload                                                  |
 307  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 308  * a    |                       logical birth txg                       |
 309  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 310  * b    |      payload                                                  |
 311  * c    |      payload                                                  |
 312  * d    |      payload                                                  |
 313  * e    |      payload                                                  |
 314  * f    |      payload                                                  |
 315  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 316  *
 317  * Legend:
 318  *
 319  * payload              contains the embedded data
 320  * B (byteorder)        byteorder (endianness)
 321  * D (dedup)            padding (set to zero)
 322  * X                    encryption (set to zero)
 323  * E (embedded)         set to one
 324  * lvl                  indirection level
 325  * type                 DMU object type
 326  * etype                how to interpret embedded data (BP_EMBEDDED_TYPE_*)
 327  * comp                 compression function of payload
 328  * PSIZE                size of payload after compression, in bytes
 329  * LSIZE                logical size of payload, in bytes
 330  *                      note that 25 bits is enough to store the largest
 331  *                      "normal" BP's LSIZE (2^16 * 2^9) in bytes
 332  * log. birth           transaction group in which the block was logically born
 333  *
 334  * Note that LSIZE and PSIZE are stored in bytes, whereas for non-embedded
 335  * bp's they are stored in units of SPA_MINBLOCKSHIFT.
 336  * Generally, the generic BP_GET_*() macros can be used on embedded BP's.
 337  * The B, D, X, lvl, type, and comp fields are stored the same as with normal
 338  * BP's so the BP_SET_* macros can be used with them.  etype, PSIZE, LSIZE must
 339  * be set with the BPE_SET_* macros.  BP_SET_EMBEDDED() should be called before
 340  * other macros, as they assert that they are only used on BP's of the correct
 341  * "embedded-ness". Encrypted blkptr_t's cannot be embedded because they use
 342  * the payload space for encryption parameters (see the comment above on
 343  * how encryption parameters are stored).
 344  */
 345
 346 #define BPE_GET_ETYPE(bp)       \
 347         (ASSERT(BP_IS_EMBEDDED(bp)), \
 348         BF64_GET((bp)->blk_prop, 40, 8))
 349 #define BPE_SET_ETYPE(bp, t)    do { \
 350         ASSERT(BP_IS_EMBEDDED(bp)); \
 351         BF64_SET((bp)->blk_prop, 40, 8, t); \
 352 _NOTE(CONSTCOND) } while (0)
 353
 354 #define BPE_GET_LSIZE(bp)       \
 355         (ASSERT(BP_IS_EMBEDDED(bp)), \
 356         BF64_GET_SB((bp)->blk_prop, 0, 25, 0, 1))
 357 #define BPE_SET_LSIZE(bp, x)    do { \
 358         ASSERT(BP_IS_EMBEDDED(bp)); \
 359         BF64_SET_SB((bp)->blk_prop, 0, 25, 0, 1, x); \
 360 _NOTE(CONSTCOND) } while (0)
 361
 362 #define BPE_GET_PSIZE(bp)       \
 363         (ASSERT(BP_IS_EMBEDDED(bp)), \
 364         BF64_GET_SB((bp)->blk_prop, 25, 7, 0, 1))
 365 #define BPE_SET_PSIZE(bp, x)    do { \
 366         ASSERT(BP_IS_EMBEDDED(bp)); \
 367         BF64_SET_SB((bp)->blk_prop, 25, 7, 0, 1, x); \
 368 _NOTE(CONSTCOND) } while (0)
 369
 370 typedef enum bp_embedded_type {
 371         BP_EMBEDDED_TYPE_DATA,
 372         BP_EMBEDDED_TYPE_RESERVED, /* Reserved for Delphix byteswap feature. */
 373         BP_EMBEDDED_TYPE_REDACTED,
 374         NUM_BP_EMBEDDED_TYPES
 375 } bp_embedded_type_t;
 376
 377 #define BPE_NUM_WORDS 14
 378 #define BPE_PAYLOAD_SIZE (BPE_NUM_WORDS * sizeof (uint64_t))
 379 #define BPE_IS_PAYLOADWORD(bp, wp) \
 380         ((wp) != &(bp)->blk_prop && (wp) != &(bp)->blk_birth)
 381
 382 #define SPA_BLKPTRSHIFT 7               /* blkptr_t is 128 bytes        */
 383 #define SPA_DVAS_PER_BP 3               /* Number of DVAs in a bp       */
 384 #define SPA_SYNC_MIN_VDEVS 3            /* min vdevs to update during sync */
 385
 386 /*
 387  * A block is a hole when it has either 1) never been written to, or
 388  * 2) is zero-filled. In both cases, ZFS can return all zeroes for all reads
 389  * without physically allocating disk space. Holes are represented in the
 390  * blkptr_t structure by zeroed blk_dva. Correct checking for holes is
 391  * done through the BP_IS_HOLE macro. For holes, the logical size, level,
 392  * DMU object type, and birth times are all also stored for holes that
 393  * were written to at some point (i.e. were punched after having been filled).
 394  */
 395 typedef struct blkptr {
 396         dva_t           blk_dva[SPA_DVAS_PER_BP]; /* Data Virtual Addresses */
 397         uint64_t        blk_prop;       /* size, compression, type, etc     */
 398         uint64_t        blk_pad[2];     /* Extra space for the future       */
 399         uint64_t        blk_phys_birth; /* txg when block was allocated     */
 400         uint64_t        blk_birth;      /* transaction group at birth       */
 401         uint64_t        blk_fill;       /* fill count                       */
 402         zio_cksum_t     blk_cksum;      /* 256-bit checksum                 */
 403 } blkptr_t;
 404
 405 /*
 406  * Macros to get and set fields in a bp or DVA.
 407  */
 408
 409 /*
 410  * Note, for gang blocks, DVA_GET_ASIZE() is the total space allocated for
 411  * this gang DVA including its children BP's.  The space allocated at this
 412  * DVA's vdev/offset is vdev_gang_header_asize(vdev).
 413  */
 414 #define DVA_GET_ASIZE(dva)      \
 415         BF64_GET_SB((dva)->dva_word[0], 0, SPA_ASIZEBITS, SPA_MINBLOCKSHIFT, 0)
 416 #define DVA_SET_ASIZE(dva, x)   \
 417         BF64_SET_SB((dva)->dva_word[0], 0, SPA_ASIZEBITS, \
 418         SPA_MINBLOCKSHIFT, 0, x)
 419
 420 #define DVA_GET_GRID(dva)       BF64_GET((dva)->dva_word[0], 24, 8)
 421 #define DVA_SET_GRID(dva, x)    BF64_SET((dva)->dva_word[0], 24, 8, x)
 422
 423 #define DVA_GET_VDEV(dva)       BF64_GET((dva)->dva_word[0], 32, SPA_VDEVBITS)
 424 #define DVA_SET_VDEV(dva, x)    \
 425         BF64_SET((dva)->dva_word[0], 32, SPA_VDEVBITS, x)
 426
 427 #define DVA_GET_OFFSET(dva)     \
 428         BF64_GET_SB((dva)->dva_word[1], 0, 63, SPA_MINBLOCKSHIFT, 0)
 429 #define DVA_SET_OFFSET(dva, x)  \
 430         BF64_SET_SB((dva)->dva_word[1], 0, 63, SPA_MINBLOCKSHIFT, 0, x)
 431
 432 #define DVA_GET_GANG(dva)       BF64_GET((dva)->dva_word[1], 63, 1)
 433 #define DVA_SET_GANG(dva, x)    BF64_SET((dva)->dva_word[1], 63, 1, x)
 434
 435 #define BP_GET_LSIZE(bp)        \
 436         (BP_IS_EMBEDDED(bp) ?   \
 437         (BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA ? BPE_GET_LSIZE(bp) : 0): \
 438         BF64_GET_SB((bp)->blk_prop, 0, SPA_LSIZEBITS, SPA_MINBLOCKSHIFT, 1))
 439 #define BP_SET_LSIZE(bp, x)     do { \
 440         ASSERT(!BP_IS_EMBEDDED(bp)); \
 441         BF64_SET_SB((bp)->blk_prop, \
 442             0, SPA_LSIZEBITS, SPA_MINBLOCKSHIFT, 1, x); \
 443 _NOTE(CONSTCOND) } while (0)
 444
 445 #define BP_GET_PSIZE(bp)        \
 446         (BP_IS_EMBEDDED(bp) ? 0 : \
 447         BF64_GET_SB((bp)->blk_prop, 16, SPA_PSIZEBITS, SPA_MINBLOCKSHIFT, 1))
 448 #define BP_SET_PSIZE(bp, x)     do { \
 449         ASSERT(!BP_IS_EMBEDDED(bp)); \
 450         BF64_SET_SB((bp)->blk_prop, \
 451             16, SPA_PSIZEBITS, SPA_MINBLOCKSHIFT, 1, x); \
 452 _NOTE(CONSTCOND) } while (0)
 453
 454 #define BP_GET_COMPRESS(bp)             \
 455         BF64_GET((bp)->blk_prop, 32, SPA_COMPRESSBITS)
 456 #define BP_SET_COMPRESS(bp, x)          \
 457         BF64_SET((bp)->blk_prop, 32, SPA_COMPRESSBITS, x)
 458
 459 #define BP_IS_EMBEDDED(bp)              BF64_GET((bp)->blk_prop, 39, 1)
 460 #define BP_SET_EMBEDDED(bp, x)          BF64_SET((bp)->blk_prop, 39, 1, x)
 461
 462 #define BP_GET_CHECKSUM(bp)             \
 463         (BP_IS_EMBEDDED(bp) ? ZIO_CHECKSUM_OFF : \
 464         BF64_GET((bp)->blk_prop, 40, 8))
 465 #define BP_SET_CHECKSUM(bp, x)          do { \
 466         ASSERT(!BP_IS_EMBEDDED(bp)); \
 467         BF64_SET((bp)->blk_prop, 40, 8, x); \
 468 _NOTE(CONSTCOND) } while (0)
 469
 470 #define BP_GET_TYPE(bp)                 BF64_GET((bp)->blk_prop, 48, 8)
 471 #define BP_SET_TYPE(bp, x)              BF64_SET((bp)->blk_prop, 48, 8, x)
 472
 473 #define BP_GET_LEVEL(bp)                BF64_GET((bp)->blk_prop, 56, 5)
 474 #define BP_SET_LEVEL(bp, x)             BF64_SET((bp)->blk_prop, 56, 5, x)
 475
 476 /* encrypted, authenticated, and MAC cksum bps use the same bit */
 477 #define BP_USES_CRYPT(bp)               BF64_GET((bp)->blk_prop, 61, 1)
 478 #define BP_SET_CRYPT(bp, x)             BF64_SET((bp)->blk_prop, 61, 1, x)
 479
 480 #define BP_IS_ENCRYPTED(bp)                     \
 481         (BP_USES_CRYPT(bp) &&                   \
 482         BP_GET_LEVEL(bp) <= 0 &&                \
 483         DMU_OT_IS_ENCRYPTED(BP_GET_TYPE(bp)))
 484
 485 #define BP_IS_AUTHENTICATED(bp)                 \
 486         (BP_USES_CRYPT(bp) &&                   \
 487         BP_GET_LEVEL(bp) <= 0 &&                \
 488         !DMU_OT_IS_ENCRYPTED(BP_GET_TYPE(bp)))
 489
 490 #define BP_HAS_INDIRECT_MAC_CKSUM(bp)           \
 491         (BP_USES_CRYPT(bp) && BP_GET_LEVEL(bp) > 0)
 492
 493 #define BP_IS_PROTECTED(bp)                     \
 494         (BP_IS_ENCRYPTED(bp) || BP_IS_AUTHENTICATED(bp))
 495
 496 #define BP_GET_DEDUP(bp)                BF64_GET((bp)->blk_prop, 62, 1)
 497 #define BP_SET_DEDUP(bp, x)             BF64_SET((bp)->blk_prop, 62, 1, x)
 498
 499 #define BP_GET_BYTEORDER(bp)            BF64_GET((bp)->blk_prop, 63, 1)
 500 #define BP_SET_BYTEORDER(bp, x)         BF64_SET((bp)->blk_prop, 63, 1, x)
 501
 502 #define BP_GET_FREE(bp)                 BF64_GET((bp)->blk_fill, 0, 1)
 503 #define BP_SET_FREE(bp, x)              BF64_SET((bp)->blk_fill, 0, 1, x)
 504
 505 #define BP_PHYSICAL_BIRTH(bp)           \
 506         (BP_IS_EMBEDDED(bp) ? 0 : \
 507         (bp)->blk_phys_birth ? (bp)->blk_phys_birth : (bp)->blk_birth)
 508
 509 #define BP_SET_BIRTH(bp, logical, physical)     \
 510 {                                               \
 511         ASSERT(!BP_IS_EMBEDDED(bp));            \
 512         (bp)->blk_birth = (logical);            \
 513         (bp)->blk_phys_birth = ((logical) == (physical) ? 0 : (physical)); \
 514 }
 515
 516 #define BP_GET_FILL(bp)                         \
 517         ((BP_IS_ENCRYPTED(bp)) ? BF64_GET((bp)->blk_fill, 0, 32) : \
 518         ((BP_IS_EMBEDDED(bp)) ? 1 : (bp)->blk_fill))
 519
 520 #define BP_SET_FILL(bp, fill)                   \
 521 {                                               \
 522         if (BP_IS_ENCRYPTED(bp))                        \
 523                 BF64_SET((bp)->blk_fill, 0, 32, fill); \
 524         else                                    \
 525                 (bp)->blk_fill = fill;          \
 526 }
 527
 528 #define BP_GET_IV2(bp)                          \
 529         (ASSERT(BP_IS_ENCRYPTED(bp)),           \
 530         BF64_GET((bp)->blk_fill, 32, 32))
 531 #define BP_SET_IV2(bp, iv2)                     \
 532 {                                               \
 533         ASSERT(BP_IS_ENCRYPTED(bp));            \
 534         BF64_SET((bp)->blk_fill, 32, 32, iv2);  \
 535 }
 536
 537 #define BP_IS_METADATA(bp)      \
 538         (BP_GET_LEVEL(bp) > 0 || DMU_OT_IS_METADATA(BP_GET_TYPE(bp)))
 539
 540 #define BP_GET_ASIZE(bp)        \
 541         (BP_IS_EMBEDDED(bp) ? 0 : \
 542         DVA_GET_ASIZE(&(bp)->blk_dva[0]) + \
 543         DVA_GET_ASIZE(&(bp)->blk_dva[1]) + \
 544         (DVA_GET_ASIZE(&(bp)->blk_dva[2]) * !BP_IS_ENCRYPTED(bp)))
 545
 546 #define BP_GET_UCSIZE(bp)       \
 547         (BP_IS_METADATA(bp) ? BP_GET_PSIZE(bp) : BP_GET_LSIZE(bp))
 548
 549 #define BP_GET_NDVAS(bp)        \
 550         (BP_IS_EMBEDDED(bp) ? 0 : \
 551         !!DVA_GET_ASIZE(&(bp)->blk_dva[0]) + \
 552         !!DVA_GET_ASIZE(&(bp)->blk_dva[1]) + \
 553         (!!DVA_GET_ASIZE(&(bp)->blk_dva[2]) * !BP_IS_ENCRYPTED(bp)))
 554
 555 #define BP_COUNT_GANG(bp)       \
 556         (BP_IS_EMBEDDED(bp) ? 0 : \
 557         (DVA_GET_GANG(&(bp)->blk_dva[0]) + \
 558         DVA_GET_GANG(&(bp)->blk_dva[1]) + \
 559         (DVA_GET_GANG(&(bp)->blk_dva[2]) * !BP_IS_ENCRYPTED(bp))))
 560
 561 #define DVA_EQUAL(dva1, dva2)   \
 562         ((dva1)->dva_word[1] == (dva2)->dva_word[1] && \
 563         (dva1)->dva_word[0] == (dva2)->dva_word[0])
 564
 565 #define BP_EQUAL(bp1, bp2)      \
 566         (BP_PHYSICAL_BIRTH(bp1) == BP_PHYSICAL_BIRTH(bp2) &&    \
 567         (bp1)->blk_birth == (bp2)->blk_birth &&                 \
 568         DVA_EQUAL(&(bp1)->blk_dva[0], &(bp2)->blk_dva[0]) &&    \
 569         DVA_EQUAL(&(bp1)->blk_dva[1], &(bp2)->blk_dva[1]) &&    \
 570         DVA_EQUAL(&(bp1)->blk_dva[2], &(bp2)->blk_dva[2]))
 571
 572
 573 #define DVA_IS_VALID(dva)       (DVA_GET_ASIZE(dva) != 0)
 574
 575 #define BP_IDENTITY(bp)         (ASSERT(!BP_IS_EMBEDDED(bp)), &(bp)->blk_dva[0])
 576 #define BP_IS_GANG(bp)          \
 577         (BP_IS_EMBEDDED(bp) ? B_FALSE : DVA_GET_GANG(BP_IDENTITY(bp)))
 578 #define DVA_IS_EMPTY(dva)       ((dva)->dva_word[0] == 0ULL &&  \
 579                                 (dva)->dva_word[1] == 0ULL)
 580 #define BP_IS_HOLE(bp) \
 581         (!BP_IS_EMBEDDED(bp) && DVA_IS_EMPTY(BP_IDENTITY(bp)))
 582
 583 #define BP_SET_REDACTED(bp) \
 584 {                                                       \
 585         BP_SET_EMBEDDED(bp, B_TRUE);                    \
 586         BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_REDACTED);   \
 587 }
 588 #define BP_IS_REDACTED(bp) \
 589         (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_REDACTED)
 590
 591 /* BP_IS_RAIDZ(bp) assumes no block compression */
 592 #define BP_IS_RAIDZ(bp)         (DVA_GET_ASIZE(&(bp)->blk_dva[0]) > \
 593                                 BP_GET_PSIZE(bp))
 594
 595 #define BP_ZERO(bp)                             \
 596 {                                               \
 597         (bp)->blk_dva[0].dva_word[0] = 0;       \
 598         (bp)->blk_dva[0].dva_word[1] = 0;       \
 599         (bp)->blk_dva[1].dva_word[0] = 0;       \
 600         (bp)->blk_dva[1].dva_word[1] = 0;       \
 601         (bp)->blk_dva[2].dva_word[0] = 0;       \
 602         (bp)->blk_dva[2].dva_word[1] = 0;       \
 603         (bp)->blk_prop = 0;                     \
 604         (bp)->blk_pad[0] = 0;                   \
 605         (bp)->blk_pad[1] = 0;                   \
 606         (bp)->blk_phys_birth = 0;               \
 607         (bp)->blk_birth = 0;                    \
 608         (bp)->blk_fill = 0;                     \
 609         ZIO_SET_CHECKSUM(&(bp)->blk_cksum, 0, 0, 0, 0); \
 610 }
 611
 612 #ifdef _ZFS_BIG_ENDIAN
 613 #define ZFS_HOST_BYTEORDER      (0ULL)
 614 #else
 615 #define ZFS_HOST_BYTEORDER      (1ULL)
 616 #endif
 617
 618 #define BP_SHOULD_BYTESWAP(bp)  (BP_GET_BYTEORDER(bp) != ZFS_HOST_BYTEORDER)
 619
 620 #define BP_SPRINTF_LEN  400
 621
 622 /*
 623  * This macro allows code sharing between zfs, libzpool, and mdb.
 624  * 'func' is either snprintf() or mdb_snprintf().
 625  * 'ws' (whitespace) can be ' ' for single-line format, '\n' for multi-line.
 626  */
 627
 628 #define SNPRINTF_BLKPTR(func, ws, buf, size, bp, type, checksum, compress) \
 629 {                                                                       \
 630         static const char *copyname[] =                                 \
 631             { "zero", "single", "double", "triple" };                   \
 632         int len = 0;                                                    \
 633         int copies = 0;                                                 \
 634         const char *crypt_type;                                         \
 635         if (bp != NULL) {                                               \
 636                 if (BP_IS_ENCRYPTED(bp)) {                              \
 637                         crypt_type = "encrypted";                       \
 638                         /* LINTED E_SUSPICIOUS_COMPARISON */            \
 639                 } else if (BP_IS_AUTHENTICATED(bp)) {                   \
 640                         crypt_type = "authenticated";                   \
 641                 } else if (BP_HAS_INDIRECT_MAC_CKSUM(bp)) {             \
 642                         crypt_type = "indirect-MAC";                    \
 643                 } else {                                                \
 644                         crypt_type = "unencrypted";                     \
 645                 }                                                       \
 646         }                                                               \
 647         if (bp == NULL) {                                               \
 648                 len += func(buf + len, size - len, "<NULL>");           \
 649         } else if (BP_IS_HOLE(bp)) {                                    \
 650                 len += func(buf + len, size - len,                      \
 651                     "HOLE [L%llu %s] "                                  \
 652                     "size=%llxL birth=%lluL",                           \
 653                     (u_longlong_t)BP_GET_LEVEL(bp),                     \
 654                     type,                                               \
 655                     (u_longlong_t)BP_GET_LSIZE(bp),                     \
 656                     (u_longlong_t)bp->blk_birth);                       \
 657         } else if (BP_IS_EMBEDDED(bp)) {                                \
 658                 len = func(buf + len, size - len,                       \
 659                     "EMBEDDED [L%llu %s] et=%u %s "                     \
 660                     "size=%llxL/%llxP birth=%lluL",                     \
 661                     (u_longlong_t)BP_GET_LEVEL(bp),                     \
 662                     type,                                               \
 663                     (int)BPE_GET_ETYPE(bp),                             \
 664                     compress,                                           \
 665                     (u_longlong_t)BPE_GET_LSIZE(bp),                    \
 666                     (u_longlong_t)BPE_GET_PSIZE(bp),                    \
 667                     (u_longlong_t)bp->blk_birth);                       \
 668         } else if (BP_IS_REDACTED(bp)) {                                \
 669                 len += func(buf + len, size - len,                      \
 670                     "REDACTED [L%llu %s] size=%llxL birth=%lluL",       \
 671                     (u_longlong_t)BP_GET_LEVEL(bp),                     \
 672                     type,                                               \
 673                     (u_longlong_t)BP_GET_LSIZE(bp),                     \
 674                     (u_longlong_t)bp->blk_birth);                       \
 675         } else {                                                        \
 676                 for (int d = 0; d < BP_GET_NDVAS(bp); d++) {            \
 677                         const dva_t *dva = &bp->blk_dva[d];             \
 678                         if (DVA_IS_VALID(dva))                          \
 679                                 copies++;                               \
 680                         len += func(buf + len, size - len,              \
 681                             "DVA[%d]=<%llu:%llx:%llx>%c", d,            \
 682                             (u_longlong_t)DVA_GET_VDEV(dva),            \
 683                             (u_longlong_t)DVA_GET_OFFSET(dva),          \
 684                             (u_longlong_t)DVA_GET_ASIZE(dva),           \
 685                             ws);                                        \
 686                 }                                                       \
 687                 if (BP_IS_ENCRYPTED(bp)) {                              \
 688                         len += func(buf + len, size - len,              \
 689                             "salt=%llx iv=%llx:%llx%c",                 \
 690                             (u_longlong_t)bp->blk_dva[2].dva_word[0],   \
 691                             (u_longlong_t)bp->blk_dva[2].dva_word[1],   \
 692                             (u_longlong_t)BP_GET_IV2(bp),               \
 693                             ws);                                        \
 694                 }                                                       \
 695                 if (BP_IS_GANG(bp) &&                                   \
 696                     DVA_GET_ASIZE(&bp->blk_dva[2]) <=                   \
 697                     DVA_GET_ASIZE(&bp->blk_dva[1]) / 2)                 \
 698                         copies--;                                       \
 699                 len += func(buf + len, size - len,                      \
 700                     "[L%llu %s] %s %s %s %s %s %s %s%c"                 \
 701                     "size=%llxL/%llxP birth=%lluL/%lluP fill=%llu%c"    \
 702                     "cksum=%llx:%llx:%llx:%llx",                        \
 703                     (u_longlong_t)BP_GET_LEVEL(bp),                     \
 704                     type,                                               \
 705                     checksum,                                           \
 706                     compress,                                           \
 707                     crypt_type,                                         \
 708                     BP_GET_BYTEORDER(bp) == 0 ? "BE" : "LE",            \
 709                     BP_IS_GANG(bp) ? "gang" : "contiguous",             \
 710                     BP_GET_DEDUP(bp) ? "dedup" : "unique",              \
 711                     copyname[copies],                                   \
 712                     ws,                                                 \
 713                     (u_longlong_t)BP_GET_LSIZE(bp),                     \
 714                     (u_longlong_t)BP_GET_PSIZE(bp),                     \
 715                     (u_longlong_t)bp->blk_birth,                        \
 716                     (u_longlong_t)BP_PHYSICAL_BIRTH(bp),                \
 717                     (u_longlong_t)BP_GET_FILL(bp),                      \
 718                     ws,                                                 \
 719                     (u_longlong_t)bp->blk_cksum.zc_word[0],             \
 720                     (u_longlong_t)bp->blk_cksum.zc_word[1],             \
 721                     (u_longlong_t)bp->blk_cksum.zc_word[2],             \
 722                     (u_longlong_t)bp->blk_cksum.zc_word[3]);            \
 723         }                                                               \
 724         ASSERT(len < size);                                             \
 725 }
 726
 727 #define BP_GET_BUFC_TYPE(bp)                                            \
 728         (BP_IS_METADATA(bp) ? ARC_BUFC_METADATA : ARC_BUFC_DATA)
 729
 730 typedef enum spa_import_type {
 731         SPA_IMPORT_EXISTING,
 732         SPA_IMPORT_ASSEMBLE
 733 } spa_import_type_t;
 734
 735 typedef enum spa_mode {
 736         SPA_MODE_UNINIT = 0,
 737         SPA_MODE_READ = 1,
 738         SPA_MODE_WRITE = 2,
 739 } spa_mode_t;
 740
 741 /*
 742  * Send TRIM commands in-line during normal pool operation while deleting.
 743  *      OFF: no
 744  *      ON: yes
 745  * NB: IN_FREEBSD_BASE is defined within the FreeBSD sources.
 746  */
 747 typedef enum {
 748         SPA_AUTOTRIM_OFF = 0,   /* default */
 749         SPA_AUTOTRIM_ON,
 750 #ifdef IN_FREEBSD_BASE
 751         SPA_AUTOTRIM_DEFAULT = SPA_AUTOTRIM_ON,
 752 #else
 753         SPA_AUTOTRIM_DEFAULT = SPA_AUTOTRIM_OFF,
 754 #endif
 755 } spa_autotrim_t;
 756
 757 /*
 758  * Reason TRIM command was issued, used internally for accounting purposes.
 759  */
 760 typedef enum trim_type {
 761         TRIM_TYPE_MANUAL = 0,
 762         TRIM_TYPE_AUTO = 1,
 763         TRIM_TYPE_SIMPLE = 2
 764 } trim_type_t;
 765
 766 /* state manipulation functions */
 767 extern int spa_open(const char *pool, spa_t **, void *tag);
 768 extern int spa_open_rewind(const char *pool, spa_t **, void *tag,
 769     nvlist_t *policy, nvlist_t **config);
 770 extern int spa_get_stats(const char *pool, nvlist_t **config, char *altroot,
 771     size_t buflen);
 772 extern int spa_create(const char *pool, nvlist_t *nvroot, nvlist_t *props,
 773     nvlist_t *zplprops, struct dsl_crypto_params *dcp);
 774 extern int spa_import(char *pool, nvlist_t *config, nvlist_t *props,
 775     uint64_t flags);
 776 extern nvlist_t *spa_tryimport(nvlist_t *tryconfig);
 777 extern int spa_destroy(const char *pool);
 778 extern int spa_checkpoint(const char *pool);
 779 extern int spa_checkpoint_discard(const char *pool);
 780 extern int spa_export(const char *pool, nvlist_t **oldconfig, boolean_t force,
 781     boolean_t hardforce);
 782 extern int spa_reset(const char *pool);
 783 extern void spa_async_request(spa_t *spa, int flag);
 784 extern void spa_async_unrequest(spa_t *spa, int flag);
 785 extern void spa_async_suspend(spa_t *spa);
 786 extern void spa_async_resume(spa_t *spa);
 787 extern int spa_async_tasks(spa_t *spa);
 788 extern spa_t *spa_inject_addref(char *pool);
 789 extern void spa_inject_delref(spa_t *spa);
 790 extern void spa_scan_stat_init(spa_t *spa);
 791 extern int spa_scan_get_stats(spa_t *spa, pool_scan_stat_t *ps);
 792 extern int bpobj_enqueue_alloc_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx);
 793 extern int bpobj_enqueue_free_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx);
 794
 795 #define SPA_ASYNC_CONFIG_UPDATE                 0x01
 796 #define SPA_ASYNC_REMOVE                        0x02
 797 #define SPA_ASYNC_PROBE                         0x04
 798 #define SPA_ASYNC_RESILVER_DONE                 0x08
 799 #define SPA_ASYNC_RESILVER                      0x10
 800 #define SPA_ASYNC_AUTOEXPAND                    0x20
 801 #define SPA_ASYNC_REMOVE_DONE                   0x40
 802 #define SPA_ASYNC_REMOVE_STOP                   0x80
 803 #define SPA_ASYNC_INITIALIZE_RESTART            0x100
 804 #define SPA_ASYNC_TRIM_RESTART                  0x200
 805 #define SPA_ASYNC_AUTOTRIM_RESTART              0x400
 806 #define SPA_ASYNC_L2CACHE_REBUILD               0x800
 807 #define SPA_ASYNC_L2CACHE_TRIM                  0x1000
 808 #define SPA_ASYNC_REBUILD_DONE                  0x2000
 809
 810 /* device manipulation */
 811 extern int spa_vdev_add(spa_t *spa, nvlist_t *nvroot);
 812 extern int spa_vdev_attach(spa_t *spa, uint64_t guid, nvlist_t *nvroot,
 813     int replacing, int rebuild);
 814 extern int spa_vdev_detach(spa_t *spa, uint64_t guid, uint64_t pguid,
 815     int replace_done);
 816 extern int spa_vdev_remove(spa_t *spa, uint64_t guid, boolean_t unspare);
 817 extern boolean_t spa_vdev_remove_active(spa_t *spa);
 818 extern int spa_vdev_initialize(spa_t *spa, nvlist_t *nv, uint64_t cmd_type,
 819     nvlist_t *vdev_errlist);
 820 extern int spa_vdev_trim(spa_t *spa, nvlist_t *nv, uint64_t cmd_type,
 821     uint64_t rate, boolean_t partial, boolean_t secure, nvlist_t *vdev_errlist);
 822 extern int spa_vdev_setpath(spa_t *spa, uint64_t guid, const char *newpath);
 823 extern int spa_vdev_setfru(spa_t *spa, uint64_t guid, const char *newfru);
 824 extern int spa_vdev_split_mirror(spa_t *spa, char *newname, nvlist_t *config,
 825     nvlist_t *props, boolean_t exp);
 826
 827 /* spare state (which is global across all pools) */
 828 extern void spa_spare_add(vdev_t *vd);
 829 extern void spa_spare_remove(vdev_t *vd);
 830 extern boolean_t spa_spare_exists(uint64_t guid, uint64_t *pool, int *refcnt);
 831 extern void spa_spare_activate(vdev_t *vd);
 832
 833 /* L2ARC state (which is global across all pools) */
 834 extern void spa_l2cache_add(vdev_t *vd);
 835 extern void spa_l2cache_remove(vdev_t *vd);
 836 extern boolean_t spa_l2cache_exists(uint64_t guid, uint64_t *pool);
 837 extern void spa_l2cache_activate(vdev_t *vd);
 838 extern void spa_l2cache_drop(spa_t *spa);
 839
 840 /* scanning */
 841 extern int spa_scan(spa_t *spa, pool_scan_func_t func);
 842 extern int spa_scan_stop(spa_t *spa);
 843 extern int spa_scrub_pause_resume(spa_t *spa, pool_scrub_cmd_t flag);
 844
 845 /* spa syncing */
 846 extern void spa_sync(spa_t *spa, uint64_t txg); /* only for DMU use */
 847 extern void spa_sync_allpools(void);
 848
 849 extern int zfs_sync_pass_deferred_free;
 850
 851 /* spa namespace global mutex */
 852 extern kmutex_t spa_namespace_lock;
 853
 854 /*
 855  * SPA configuration functions in spa_config.c
 856  */
 857
 858 #define SPA_CONFIG_UPDATE_POOL  0
 859 #define SPA_CONFIG_UPDATE_VDEVS 1
 860
 861 extern void spa_write_cachefile(spa_t *, boolean_t, boolean_t);
 862 extern void spa_config_load(void);
 863 extern nvlist_t *spa_all_configs(uint64_t *);
 864 extern void spa_config_set(spa_t *spa, nvlist_t *config);
 865 extern nvlist_t *spa_config_generate(spa_t *spa, vdev_t *vd, uint64_t txg,
 866     int getstats);
 867 extern void spa_config_update(spa_t *spa, int what);
 868 extern int spa_config_parse(spa_t *spa, vdev_t **vdp, nvlist_t *nv,
 869     vdev_t *parent, uint_t id, int atype);
 870
 871
 872 /*
 873  * Miscellaneous SPA routines in spa_misc.c
 874  */
 875
 876 /* Namespace manipulation */
 877 extern spa_t *spa_lookup(const char *name);
 878 extern spa_t *spa_add(const char *name, nvlist_t *config, const char *altroot);
 879 extern void spa_remove(spa_t *spa);
 880 extern spa_t *spa_next(spa_t *prev);
 881
 882 /* Refcount functions */
 883 extern void spa_open_ref(spa_t *spa, void *tag);
 884 extern void spa_close(spa_t *spa, void *tag);
 885 extern void spa_async_close(spa_t *spa, void *tag);
 886 extern boolean_t spa_refcount_zero(spa_t *spa);
 887
 888 #define SCL_NONE        0x00
 889 #define SCL_CONFIG      0x01
 890 #define SCL_STATE       0x02
 891 #define SCL_L2ARC       0x04            /* hack until L2ARC 2.0 */
 892 #define SCL_ALLOC       0x08
 893 #define SCL_ZIO         0x10
 894 #define SCL_FREE        0x20
 895 #define SCL_VDEV        0x40
 896 #define SCL_LOCKS       7
 897 #define SCL_ALL         ((1 << SCL_LOCKS) - 1)
 898 #define SCL_STATE_ALL   (SCL_STATE | SCL_L2ARC | SCL_ZIO)
 899
 900 /* Historical pool statistics */
 901 typedef struct spa_history_kstat {
 902         kmutex_t                lock;
 903         uint64_t                count;
 904         uint64_t                size;
 905         kstat_t                 *kstat;
 906         void                    *priv;
 907         list_t                  list;
 908 } spa_history_kstat_t;
 909
 910 typedef struct spa_history_list {
 911         uint64_t                size;
 912         procfs_list_t           procfs_list;
 913 } spa_history_list_t;
 914
 915 typedef struct spa_stats {
 916         spa_history_list_t      read_history;
 917         spa_history_list_t      txg_history;
 918         spa_history_kstat_t     tx_assign_histogram;
 919         spa_history_kstat_t     io_history;
 920         spa_history_list_t      mmp_history;
 921         spa_history_kstat_t     state;          /* pool state */
 922         spa_history_kstat_t     iostats;
 923 } spa_stats_t;
 924
 925 typedef enum txg_state {
 926         TXG_STATE_BIRTH         = 0,
 927         TXG_STATE_OPEN          = 1,
 928         TXG_STATE_QUIESCED      = 2,
 929         TXG_STATE_WAIT_FOR_SYNC = 3,
 930         TXG_STATE_SYNCED        = 4,
 931         TXG_STATE_COMMITTED     = 5,
 932 } txg_state_t;
 933
 934 typedef struct txg_stat {
 935         vdev_stat_t             vs1;
 936         vdev_stat_t             vs2;
 937         uint64_t                txg;
 938         uint64_t                ndirty;
 939 } txg_stat_t;
 940
 941 /* Assorted pool IO kstats */
 942 typedef struct spa_iostats {
 943         kstat_named_t   trim_extents_written;
 944         kstat_named_t   trim_bytes_written;
 945         kstat_named_t   trim_extents_skipped;
 946         kstat_named_t   trim_bytes_skipped;
 947         kstat_named_t   trim_extents_failed;
 948         kstat_named_t   trim_bytes_failed;
 949         kstat_named_t   autotrim_extents_written;
 950         kstat_named_t   autotrim_bytes_written;
 951         kstat_named_t   autotrim_extents_skipped;
 952         kstat_named_t   autotrim_bytes_skipped;
 953         kstat_named_t   autotrim_extents_failed;
 954         kstat_named_t   autotrim_bytes_failed;
 955         kstat_named_t   simple_trim_extents_written;
 956         kstat_named_t   simple_trim_bytes_written;
 957         kstat_named_t   simple_trim_extents_skipped;
 958         kstat_named_t   simple_trim_bytes_skipped;
 959         kstat_named_t   simple_trim_extents_failed;
 960         kstat_named_t   simple_trim_bytes_failed;
 961 } spa_iostats_t;
 962
 963 extern void spa_stats_init(spa_t *spa);
 964 extern void spa_stats_destroy(spa_t *spa);
 965 extern void spa_read_history_add(spa_t *spa, const zbookmark_phys_t *zb,
 966     uint32_t aflags);
 967 extern void spa_txg_history_add(spa_t *spa, uint64_t txg, hrtime_t birth_time);
 968 extern int spa_txg_history_set(spa_t *spa,  uint64_t txg,
 969     txg_state_t completed_state, hrtime_t completed_time);
 970 extern txg_stat_t *spa_txg_history_init_io(spa_t *, uint64_t,
 971     struct dsl_pool *);
 972 extern void spa_txg_history_fini_io(spa_t *, txg_stat_t *);
 973 extern void spa_tx_assign_add_nsecs(spa_t *spa, uint64_t nsecs);
 974 extern int spa_mmp_history_set_skip(spa_t *spa, uint64_t mmp_kstat_id);
 975 extern int spa_mmp_history_set(spa_t *spa, uint64_t mmp_kstat_id, int io_error,
 976     hrtime_t duration);
 977 extern void spa_mmp_history_add(spa_t *spa, uint64_t txg, uint64_t timestamp,
 978     uint64_t mmp_delay, vdev_t *vd, int label, uint64_t mmp_kstat_id,
 979     int error);
 980 extern void spa_iostats_trim_add(spa_t *spa, trim_type_t type,
 981     uint64_t extents_written, uint64_t bytes_written,
 982     uint64_t extents_skipped, uint64_t bytes_skipped,
 983     uint64_t extents_failed, uint64_t bytes_failed);
 984 extern void spa_import_progress_add(spa_t *spa);
 985 extern void spa_import_progress_remove(uint64_t spa_guid);
 986 extern int spa_import_progress_set_mmp_check(uint64_t pool_guid,
 987     uint64_t mmp_sec_remaining);
 988 extern int spa_import_progress_set_max_txg(uint64_t pool_guid,
 989     uint64_t max_txg);
 990 extern int spa_import_progress_set_state(uint64_t pool_guid,
 991     spa_load_state_t spa_load_state);
 992
 993 /* Pool configuration locks */
 994 extern int spa_config_tryenter(spa_t *spa, int locks, void *tag, krw_t rw);
 995 extern void spa_config_enter(spa_t *spa, int locks, const void *tag, krw_t rw);
 996 extern void spa_config_exit(spa_t *spa, int locks, const void *tag);
 997 extern int spa_config_held(spa_t *spa, int locks, krw_t rw);
 998
 999 /* Pool vdev add/remove lock */
1000 extern uint64_t spa_vdev_enter(spa_t *spa);
1001 extern uint64_t spa_vdev_detach_enter(spa_t *spa, uint64_t guid);
1002 extern uint64_t spa_vdev_config_enter(spa_t *spa);
1003 extern void spa_vdev_config_exit(spa_t *spa, vdev_t *vd, uint64_t txg,
1004     int error, char *tag);
1005 extern int spa_vdev_exit(spa_t *spa, vdev_t *vd, uint64_t txg, int error);
1006
1007 /* Pool vdev state change lock */
1008 extern void spa_vdev_state_enter(spa_t *spa, int oplock);
1009 extern int spa_vdev_state_exit(spa_t *spa, vdev_t *vd, int error);
1010
1011 /* Log state */
1012 typedef enum spa_log_state {
1013         SPA_LOG_UNKNOWN = 0,    /* unknown log state */
1014         SPA_LOG_MISSING,        /* missing log(s) */
1015         SPA_LOG_CLEAR,          /* clear the log(s) */
1016         SPA_LOG_GOOD,           /* log(s) are good */
1017 } spa_log_state_t;
1018
1019 extern spa_log_state_t spa_get_log_state(spa_t *spa);
1020 extern void spa_set_log_state(spa_t *spa, spa_log_state_t state);
1021 extern int spa_reset_logs(spa_t *spa);
1022
1023 /* Log claim callback */
1024 extern void spa_claim_notify(zio_t *zio);
1025 extern void spa_deadman(void *);
1026
1027 /* Accessor functions */
1028 extern boolean_t spa_shutting_down(spa_t *spa);
1029 extern struct dsl_pool *spa_get_dsl(spa_t *spa);
1030 extern boolean_t spa_is_initializing(spa_t *spa);
1031 extern boolean_t spa_indirect_vdevs_loaded(spa_t *spa);
1032 extern blkptr_t *spa_get_rootblkptr(spa_t *spa);
1033 extern void spa_set_rootblkptr(spa_t *spa, const blkptr_t *bp);
1034 extern void spa_altroot(spa_t *, char *, size_t);
1035 extern int spa_sync_pass(spa_t *spa);
1036 extern char *spa_name(spa_t *spa);
1037 extern uint64_t spa_guid(spa_t *spa);
1038 extern uint64_t spa_load_guid(spa_t *spa);
1039 extern uint64_t spa_last_synced_txg(spa_t *spa);
1040 extern uint64_t spa_first_txg(spa_t *spa);
1041 extern uint64_t spa_syncing_txg(spa_t *spa);
1042 extern uint64_t spa_final_dirty_txg(spa_t *spa);
1043 extern uint64_t spa_version(spa_t *spa);
1044 extern pool_state_t spa_state(spa_t *spa);
1045 extern spa_load_state_t spa_load_state(spa_t *spa);
1046 extern uint64_t spa_freeze_txg(spa_t *spa);
1047 extern uint64_t spa_get_worst_case_asize(spa_t *spa, uint64_t lsize);
1048 extern uint64_t spa_get_dspace(spa_t *spa);
1049 extern uint64_t spa_get_checkpoint_space(spa_t *spa);
1050 extern uint64_t spa_get_slop_space(spa_t *spa);
1051 extern void spa_update_dspace(spa_t *spa);
1052 extern uint64_t spa_version(spa_t *spa);
1053 extern boolean_t spa_deflate(spa_t *spa);
1054 extern metaslab_class_t *spa_normal_class(spa_t *spa);
1055 extern metaslab_class_t *spa_log_class(spa_t *spa);
1056 extern metaslab_class_t *spa_embedded_log_class(spa_t *spa);
1057 extern metaslab_class_t *spa_special_class(spa_t *spa);
1058 extern metaslab_class_t *spa_dedup_class(spa_t *spa);
1059 extern metaslab_class_t *spa_preferred_class(spa_t *spa, uint64_t size,
1060     dmu_object_type_t objtype, uint_t level, uint_t special_smallblk);
1061
1062 extern void spa_evicting_os_register(spa_t *, objset_t *os);
1063 extern void spa_evicting_os_deregister(spa_t *, objset_t *os);
1064 extern void spa_evicting_os_wait(spa_t *spa);
1065 extern int spa_max_replication(spa_t *spa);
1066 extern int spa_prev_software_version(spa_t *spa);
1067 extern uint64_t spa_get_failmode(spa_t *spa);
1068 extern uint64_t spa_get_deadman_failmode(spa_t *spa);
1069 extern void spa_set_deadman_failmode(spa_t *spa, const char *failmode);
1070 extern boolean_t spa_suspended(spa_t *spa);
1071 extern uint64_t spa_bootfs(spa_t *spa);
1072 extern uint64_t spa_delegation(spa_t *spa);
1073 extern objset_t *spa_meta_objset(spa_t *spa);
1074 extern space_map_t *spa_syncing_log_sm(spa_t *spa);
1075 extern uint64_t spa_deadman_synctime(spa_t *spa);
1076 extern uint64_t spa_deadman_ziotime(spa_t *spa);
1077 extern uint64_t spa_dirty_data(spa_t *spa);
1078 extern spa_autotrim_t spa_get_autotrim(spa_t *spa);
1079
1080 /* Miscellaneous support routines */
1081 extern void spa_load_failed(spa_t *spa, const char *fmt, ...);
1082 extern void spa_load_note(spa_t *spa, const char *fmt, ...);
1083 extern void spa_activate_mos_feature(spa_t *spa, const char *feature,
1084     dmu_tx_t *tx);
1085 extern void spa_deactivate_mos_feature(spa_t *spa, const char *feature);
1086 extern spa_t *spa_by_guid(uint64_t pool_guid, uint64_t device_guid);
1087 extern boolean_t spa_guid_exists(uint64_t pool_guid, uint64_t device_guid);
1088 extern char *spa_strdup(const char *);
1089 extern void spa_strfree(char *);
1090 extern uint64_t spa_get_random(uint64_t range);
1091 extern uint64_t spa_generate_guid(spa_t *spa);
1092 extern void snprintf_blkptr(char *buf, size_t buflen, const blkptr_t *bp);
1093 extern void spa_freeze(spa_t *spa);
1094 extern int spa_change_guid(spa_t *spa);
1095 extern void spa_upgrade(spa_t *spa, uint64_t version);
1096 extern void spa_evict_all(void);
1097 extern vdev_t *spa_lookup_by_guid(spa_t *spa, uint64_t guid,
1098     boolean_t l2cache);
1099 extern boolean_t spa_has_spare(spa_t *, uint64_t guid);
1100 extern uint64_t dva_get_dsize_sync(spa_t *spa, const dva_t *dva);
1101 extern uint64_t bp_get_dsize_sync(spa_t *spa, const blkptr_t *bp);
1102 extern uint64_t bp_get_dsize(spa_t *spa, const blkptr_t *bp);
1103 extern boolean_t spa_has_slogs(spa_t *spa);
1104 extern boolean_t spa_is_root(spa_t *spa);
1105 extern boolean_t spa_writeable(spa_t *spa);
1106 extern boolean_t spa_has_pending_synctask(spa_t *spa);
1107 extern int spa_maxblocksize(spa_t *spa);
1108 extern int spa_maxdnodesize(spa_t *spa);
1109 extern boolean_t spa_has_checkpoint(spa_t *spa);
1110 extern boolean_t spa_importing_readonly_checkpoint(spa_t *spa);
1111 extern boolean_t spa_suspend_async_destroy(spa_t *spa);
1112 extern uint64_t spa_min_claim_txg(spa_t *spa);
1113 extern boolean_t zfs_dva_valid(spa_t *spa, const dva_t *dva,
1114     const blkptr_t *bp);
1115 typedef void (*spa_remap_cb_t)(uint64_t vdev, uint64_t offset, uint64_t size,
1116     void *arg);
1117 extern boolean_t spa_remap_blkptr(spa_t *spa, blkptr_t *bp,
1118     spa_remap_cb_t callback, void *arg);
1119 extern uint64_t spa_get_last_removal_txg(spa_t *spa);
1120 extern boolean_t spa_trust_config(spa_t *spa);
1121 extern uint64_t spa_missing_tvds_allowed(spa_t *spa);
1122 extern void spa_set_missing_tvds(spa_t *spa, uint64_t missing);
1123 extern boolean_t spa_top_vdevs_spacemap_addressable(spa_t *spa);
1124 extern uint64_t spa_total_metaslabs(spa_t *spa);
1125 extern boolean_t spa_multihost(spa_t *spa);
1126 extern uint32_t spa_get_hostid(spa_t *spa);
1127 extern void spa_activate_allocation_classes(spa_t *, dmu_tx_t *);
1128 extern boolean_t spa_livelist_delete_check(spa_t *spa);
1129
1130 extern spa_mode_t spa_mode(spa_t *spa);
1131 extern uint64_t zfs_strtonum(const char *str, char **nptr);
1132
1133 extern char *spa_his_ievent_table[];
1134
1135 extern void spa_history_create_obj(spa_t *spa, dmu_tx_t *tx);
1136 extern int spa_history_get(spa_t *spa, uint64_t *offset, uint64_t *len_read,
1137     char *his_buf);
1138 extern int spa_history_log(spa_t *spa, const char *his_buf);
1139 extern int spa_history_log_nvl(spa_t *spa, nvlist_t *nvl);
1140 extern void spa_history_log_version(spa_t *spa, const char *operation,
1141     dmu_tx_t *tx);
1142 extern void spa_history_log_internal(spa_t *spa, const char *operation,
1143     dmu_tx_t *tx, const char *fmt, ...) __printflike(4, 5);
1144 extern void spa_history_log_internal_ds(struct dsl_dataset *ds, const char *op,
1145     dmu_tx_t *tx, const char *fmt, ...)  __printflike(4, 5);
1146 extern void spa_history_log_internal_dd(dsl_dir_t *dd, const char *operation,
1147     dmu_tx_t *tx, const char *fmt, ...) __printflike(4, 5);
1148
1149 extern const char *spa_state_to_name(spa_t *spa);
1150
1151 /* error handling */
1152 struct zbookmark_phys;
1153 extern void spa_log_error(spa_t *spa, const zbookmark_phys_t *zb);
1154 extern int zfs_ereport_post(const char *clazz, spa_t *spa, vdev_t *vd,
1155     const zbookmark_phys_t *zb, zio_t *zio, uint64_t state);
1156 extern boolean_t zfs_ereport_is_valid(const char *clazz, spa_t *spa, vdev_t *vd,
1157     zio_t *zio);
1158 extern void zfs_ereport_taskq_fini(void);
1159 extern void zfs_ereport_clear(spa_t *spa, vdev_t *vd);
1160 extern nvlist_t *zfs_event_create(spa_t *spa, vdev_t *vd, const char *type,
1161     const char *name, nvlist_t *aux);
1162 extern void zfs_post_remove(spa_t *spa, vdev_t *vd);
1163 extern void zfs_post_state_change(spa_t *spa, vdev_t *vd, uint64_t laststate);
1164 extern void zfs_post_autoreplace(spa_t *spa, vdev_t *vd);
1165 extern uint64_t spa_get_errlog_size(spa_t *spa);
1166 extern int spa_get_errlog(spa_t *spa, void *uaddr, size_t *count);
1167 extern void spa_errlog_rotate(spa_t *spa);
1168 extern void spa_errlog_drain(spa_t *spa);
1169 extern void spa_errlog_sync(spa_t *spa, uint64_t txg);
1170 extern void spa_get_errlists(spa_t *spa, avl_tree_t *last, avl_tree_t *scrub);
1171
1172 /* vdev cache */
1173 extern void vdev_cache_stat_init(void);
1174 extern void vdev_cache_stat_fini(void);
1175
1176 /* vdev mirror */
1177 extern void vdev_mirror_stat_init(void);
1178 extern void vdev_mirror_stat_fini(void);
1179
1180 /* Initialization and termination */
1181 extern void spa_init(spa_mode_t mode);
1182 extern void spa_fini(void);
1183 extern void spa_boot_init(void);
1184
1185 /* properties */
1186 extern int spa_prop_set(spa_t *spa, nvlist_t *nvp);
1187 extern int spa_prop_get(spa_t *spa, nvlist_t **nvp);
1188 extern void spa_prop_clear_bootfs(spa_t *spa, uint64_t obj, dmu_tx_t *tx);
1189 extern void spa_configfile_set(spa_t *, nvlist_t *, boolean_t);
1190
1191 /* asynchronous event notification */
1192 extern void spa_event_notify(spa_t *spa, vdev_t *vdev, nvlist_t *hist_nvl,
1193     const char *name);
1194
1195 /* waiting for pool activities to complete */
1196 extern int spa_wait(const char *pool, zpool_wait_activity_t activity,
1197     boolean_t *waited);
1198 extern int spa_wait_tag(const char *name, zpool_wait_activity_t activity,
1199     uint64_t tag, boolean_t *waited);
1200 extern void spa_notify_waiters(spa_t *spa);
1201 extern void spa_wake_waiters(spa_t *spa);
1202
1203 /* module param call functions */
1204 int param_set_deadman_ziotime(ZFS_MODULE_PARAM_ARGS);
1205 int param_set_deadman_synctime(ZFS_MODULE_PARAM_ARGS);
1206 int param_set_slop_shift(ZFS_MODULE_PARAM_ARGS);
1207 int param_set_deadman_failmode(ZFS_MODULE_PARAM_ARGS);
1208
1209 #ifdef ZFS_DEBUG
1210 #define dprintf_bp(bp, fmt, ...) do {                           \
1211         if (zfs_flags & ZFS_DEBUG_DPRINTF) {                    \
1212         char *__blkbuf = kmem_alloc(BP_SPRINTF_LEN, KM_SLEEP);  \
1213         snprintf_blkptr(__blkbuf, BP_SPRINTF_LEN, (bp));        \
1214         dprintf(fmt " %s\n", __VA_ARGS__, __blkbuf);            \
1215         kmem_free(__blkbuf, BP_SPRINTF_LEN);                    \
1216         } \
1217 _NOTE(CONSTCOND) } while (0)
1218 #else
1219 #define dprintf_bp(bp, fmt, ...)
1220 #endif
1221
1222 extern spa_mode_t spa_mode_global;
1223 extern int zfs_deadman_enabled;
1224 extern unsigned long zfs_deadman_synctime_ms;
1225 extern unsigned long zfs_deadman_ziotime_ms;
1226 extern unsigned long zfs_deadman_checktime_ms;
1227
1228 #ifdef  __cplusplus
1229 }
1230 #endif
1231
1232 #endif  /* _SYS_SPA_H */