include/sys/spa.h

   1 /*
   2  * CDDL HEADER START
   3  *
   4  * The contents of this file are subject to the terms of the
   5  * Common Development and Distribution License (the "License").
   6  * You may not use this file except in compliance with the License.
   7  *
   8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
   9  * or https://opensource.org/licenses/CDDL-1.0.
  10  * See the License for the specific language governing permissions
  11  * and limitations under the License.
  12  *
  13  * When distributing Covered Code, include this CDDL HEADER in each
  14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
  15  * If applicable, add the following below this CDDL HEADER, with the
  16  * fields enclosed by brackets "[]" replaced with your own identifying
  17  * information: Portions Copyright [yyyy] [name of copyright owner]
  18  *
  19  * CDDL HEADER END
  20  */
  21 /*
  22  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
  23  * Copyright (c) 2011, 2021 by Delphix. All rights reserved.
  24  * Copyright 2011 Nexenta Systems, Inc.  All rights reserved.
  25  * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
  26  * Copyright 2013 Saso Kiselkov. All rights reserved.
  27  * Copyright (c) 2014 Integros [integros.com]
  28  * Copyright 2017 Joyent, Inc.
  29  * Copyright (c) 2017, Intel Corporation.
  30  * Copyright (c) 2019, Allan Jude
  31  * Copyright (c) 2019, Klara Inc.
  32  * Copyright (c) 2019, Datto Inc.
  33  */
  34
  35 #ifndef _SYS_SPA_H
  36 #define _SYS_SPA_H
  37
  38 #include <sys/avl.h>
  39 #include <sys/zfs_context.h>
  40 #include <sys/kstat.h>
  41 #include <sys/nvpair.h>
  42 #include <sys/sysmacros.h>
  43 #include <sys/types.h>
  44 #include <sys/fs/zfs.h>
  45 #include <sys/spa_checksum.h>
  46 #include <sys/dmu.h>
  47 #include <sys/space_map.h>
  48 #include <sys/bitops.h>
  49
  50 #ifdef  __cplusplus
  51 extern "C" {
  52 #endif
  53
  54 /*
  55  * Forward references that lots of things need.
  56  */
  57 typedef struct spa spa_t;
  58 typedef struct vdev vdev_t;
  59 typedef struct metaslab metaslab_t;
  60 typedef struct metaslab_group metaslab_group_t;
  61 typedef struct metaslab_class metaslab_class_t;
  62 typedef struct zio zio_t;
  63 typedef struct zilog zilog_t;
  64 typedef struct spa_aux_vdev spa_aux_vdev_t;
  65 typedef struct ddt ddt_t;
  66 typedef struct ddt_entry ddt_entry_t;
  67 typedef struct zbookmark_phys zbookmark_phys_t;
  68
  69 struct bpobj;
  70 struct bplist;
  71 struct dsl_pool;
  72 struct dsl_dataset;
  73 struct dsl_crypto_params;
  74
  75 /*
  76  * Alignment Shift (ashift) is an immutable, internal top-level vdev property
  77  * which can only be set at vdev creation time. Physical writes are always done
  78  * according to it, which makes 2^ashift the smallest possible IO on a vdev.
  79  *
  80  * We currently allow values ranging from 512 bytes (2^9 = 512) to 64 KiB
  81  * (2^16 = 65,536).
  82  */
  83 #define ASHIFT_MIN              9
  84 #define ASHIFT_MAX              16
  85
  86 /*
  87  * Size of block to hold the configuration data (a packed nvlist)
  88  */
  89 #define SPA_CONFIG_BLOCKSIZE    (1ULL << 14)
  90
  91 /*
  92  * The DVA size encodings for LSIZE and PSIZE support blocks up to 32MB.
  93  * The ASIZE encoding should be at least 64 times larger (6 more bits)
  94  * to support up to 4-way RAID-Z mirror mode with worst-case gang block
  95  * overhead, three DVAs per bp, plus one more bit in case we do anything
  96  * else that expands the ASIZE.
  97  */
  98 #define SPA_LSIZEBITS           16      /* LSIZE up to 32M (2^16 * 512) */
  99 #define SPA_PSIZEBITS           16      /* PSIZE up to 32M (2^16 * 512) */
 100 #define SPA_ASIZEBITS           24      /* ASIZE up to 64 times larger  */
 101
 102 #define SPA_COMPRESSBITS        7
 103 #define SPA_VDEVBITS            24
 104 #define SPA_COMPRESSMASK        ((1U << SPA_COMPRESSBITS) - 1)
 105
 106 /*
 107  * All SPA data is represented by 128-bit data virtual addresses (DVAs).
 108  * The members of the dva_t should be considered opaque outside the SPA.
 109  */
 110 typedef struct dva {
 111         uint64_t        dva_word[2];
 112 } dva_t;
 113
 114
 115 /*
 116  * Some checksums/hashes need a 256-bit initialization salt. This salt is kept
 117  * secret and is suitable for use in MAC algorithms as the key.
 118  */
 119 typedef struct zio_cksum_salt {
 120         uint8_t         zcs_bytes[32];
 121 } zio_cksum_salt_t;
 122
 123 /*
 124  * Each block is described by its DVAs, time of birth, checksum, etc.
 125  * The word-by-word, bit-by-bit layout of the blkptr is as follows:
 126  *
 127  *      64      56      48      40      32      24      16      8       0
 128  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 129  * 0    |  pad  |         vdev1         | GRID  |         ASIZE         |
 130  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 131  * 1    |G|                      offset1                                |
 132  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 133  * 2    |  pad  |         vdev2         | GRID  |         ASIZE         |
 134  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 135  * 3    |G|                      offset2                                |
 136  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 137  * 4    |  pad  |         vdev3         | GRID  |         ASIZE         |
 138  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 139  * 5    |G|                      offset3                                |
 140  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 141  * 6    |BDX|lvl| type  | cksum |E| comp|    PSIZE      |     LSIZE     |
 142  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 143  * 7    |                       padding                                 |
 144  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 145  * 8    |                       padding                                 |
 146  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 147  * 9    |                       physical birth txg                      |
 148  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 149  * a    |                       logical birth txg                       |
 150  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 151  * b    |                       fill count                              |
 152  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 153  * c    |                       checksum[0]                             |
 154  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 155  * d    |                       checksum[1]                             |
 156  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 157  * e    |                       checksum[2]                             |
 158  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 159  * f    |                       checksum[3]                             |
 160  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 161  *
 162  * Legend:
 163  *
 164  * vdev         virtual device ID
 165  * offset       offset into virtual device
 166  * LSIZE        logical size
 167  * PSIZE        physical size (after compression)
 168  * ASIZE        allocated size (including RAID-Z parity and gang block headers)
 169  * GRID         RAID-Z layout information (reserved for future use)
 170  * cksum        checksum function
 171  * comp         compression function
 172  * G            gang block indicator
 173  * B            byteorder (endianness)
 174  * D            dedup
 175  * X            encryption
 176  * E            blkptr_t contains embedded data (see below)
 177  * lvl          level of indirection
 178  * type         DMU object type
 179  * phys birth   txg when dva[0] was written; zero if same as logical birth txg
 180  *              note that typically all the dva's would be written in this
 181  *              txg, but they could be different if they were moved by
 182  *              device removal.
 183  * log. birth   transaction group in which the block was logically born
 184  * fill count   number of non-zero blocks under this bp
 185  * checksum[4]  256-bit checksum of the data this bp describes
 186  */
 187
 188 /*
 189  * The blkptr_t's of encrypted blocks also need to store the encryption
 190  * parameters so that the block can be decrypted. This layout is as follows:
 191  *
 192  *      64      56      48      40      32      24      16      8       0
 193  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 194  * 0    |               vdev1           | GRID  |         ASIZE         |
 195  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 196  * 1    |G|                      offset1                                |
 197  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 198  * 2    |               vdev2           | GRID  |         ASIZE         |
 199  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 200  * 3    |G|                      offset2                                |
 201  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 202  * 4    |                       salt                                    |
 203  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 204  * 5    |                       IV1                                     |
 205  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 206  * 6    |BDX|lvl| type  | cksum |E| comp|    PSIZE      |     LSIZE     |
 207  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 208  * 7    |                       padding                                 |
 209  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 210  * 8    |                       padding                                 |
 211  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 212  * 9    |                       physical birth txg                      |
 213  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 214  * a    |                       logical birth txg                       |
 215  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 216  * b    |               IV2             |           fill count          |
 217  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 218  * c    |                       checksum[0]                             |
 219  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 220  * d    |                       checksum[1]                             |
 221  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 222  * e    |                       MAC[0]                                  |
 223  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 224  * f    |                       MAC[1]                                  |
 225  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 226  *
 227  * Legend:
 228  *
 229  * salt         Salt for generating encryption keys
 230  * IV1          First 64 bits of encryption IV
 231  * X            Block requires encryption handling (set to 1)
 232  * E            blkptr_t contains embedded data (set to 0, see below)
 233  * fill count   number of non-zero blocks under this bp (truncated to 32 bits)
 234  * IV2          Last 32 bits of encryption IV
 235  * checksum[2]  128-bit checksum of the data this bp describes
 236  * MAC[2]       128-bit message authentication code for this data
 237  *
 238  * The X bit being set indicates that this block is one of 3 types. If this is
 239  * a level 0 block with an encrypted object type, the block is encrypted
 240  * (see BP_IS_ENCRYPTED()). If this is a level 0 block with an unencrypted
 241  * object type, this block is authenticated with an HMAC (see
 242  * BP_IS_AUTHENTICATED()). Otherwise (if level > 0), this bp will use the MAC
 243  * words to store a checksum-of-MACs from the level below (see
 244  * BP_HAS_INDIRECT_MAC_CKSUM()). For convenience in the code, BP_IS_PROTECTED()
 245  * refers to both encrypted and authenticated blocks and BP_USES_CRYPT()
 246  * refers to any of these 3 kinds of blocks.
 247  *
 248  * The additional encryption parameters are the salt, IV, and MAC which are
 249  * explained in greater detail in the block comment at the top of zio_crypt.c.
 250  * The MAC occupies half of the checksum space since it serves a very similar
 251  * purpose: to prevent data corruption on disk. The only functional difference
 252  * is that the checksum is used to detect on-disk corruption whether or not the
 253  * encryption key is loaded and the MAC provides additional protection against
 254  * malicious disk tampering. We use the 3rd DVA to store the salt and first
 255  * 64 bits of the IV. As a result encrypted blocks can only have 2 copies
 256  * maximum instead of the normal 3. The last 32 bits of the IV are stored in
 257  * the upper bits of what is usually the fill count. Note that only blocks at
 258  * level 0 or -2 are ever encrypted, which allows us to guarantee that these
 259  * 32 bits are not trampled over by other code (see zio_crypt.c for details).
 260  * The salt and IV are not used for authenticated bps or bps with an indirect
 261  * MAC checksum, so these blocks can utilize all 3 DVAs and the full 64 bits
 262  * for the fill count.
 263  */
 264
 265 /*
 266  * "Embedded" blkptr_t's don't actually point to a block, instead they
 267  * have a data payload embedded in the blkptr_t itself.  See the comment
 268  * in blkptr.c for more details.
 269  *
 270  * The blkptr_t is laid out as follows:
 271  *
 272  *      64      56      48      40      32      24      16      8       0
 273  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 274  * 0    |      payload                                                  |
 275  * 1    |      payload                                                  |
 276  * 2    |      payload                                                  |
 277  * 3    |      payload                                                  |
 278  * 4    |      payload                                                  |
 279  * 5    |      payload                                                  |
 280  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 281  * 6    |BDX|lvl| type  | etype |E| comp| PSIZE|              LSIZE     |
 282  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 283  * 7    |      payload                                                  |
 284  * 8    |      payload                                                  |
 285  * 9    |      payload                                                  |
 286  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 287  * a    |                       logical birth txg                       |
 288  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 289  * b    |      payload                                                  |
 290  * c    |      payload                                                  |
 291  * d    |      payload                                                  |
 292  * e    |      payload                                                  |
 293  * f    |      payload                                                  |
 294  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 295  *
 296  * Legend:
 297  *
 298  * payload              contains the embedded data
 299  * B (byteorder)        byteorder (endianness)
 300  * D (dedup)            padding (set to zero)
 301  * X                    encryption (set to zero)
 302  * E (embedded)         set to one
 303  * lvl                  indirection level
 304  * type                 DMU object type
 305  * etype                how to interpret embedded data (BP_EMBEDDED_TYPE_*)
 306  * comp                 compression function of payload
 307  * PSIZE                size of payload after compression, in bytes
 308  * LSIZE                logical size of payload, in bytes
 309  *                      note that 25 bits is enough to store the largest
 310  *                      "normal" BP's LSIZE (2^16 * 2^9) in bytes
 311  * log. birth           transaction group in which the block was logically born
 312  *
 313  * Note that LSIZE and PSIZE are stored in bytes, whereas for non-embedded
 314  * bp's they are stored in units of SPA_MINBLOCKSHIFT.
 315  * Generally, the generic BP_GET_*() macros can be used on embedded BP's.
 316  * The B, D, X, lvl, type, and comp fields are stored the same as with normal
 317  * BP's so the BP_SET_* macros can be used with them.  etype, PSIZE, LSIZE must
 318  * be set with the BPE_SET_* macros.  BP_SET_EMBEDDED() should be called before
 319  * other macros, as they assert that they are only used on BP's of the correct
 320  * "embedded-ness". Encrypted blkptr_t's cannot be embedded because they use
 321  * the payload space for encryption parameters (see the comment above on
 322  * how encryption parameters are stored).
 323  */
 324
 325 #define BPE_GET_ETYPE(bp)       \
 326         (ASSERT(BP_IS_EMBEDDED(bp)), \
 327         BF64_GET((bp)->blk_prop, 40, 8))
 328 #define BPE_SET_ETYPE(bp, t)    do { \
 329         ASSERT(BP_IS_EMBEDDED(bp)); \
 330         BF64_SET((bp)->blk_prop, 40, 8, t); \
 331 } while (0)
 332
 333 #define BPE_GET_LSIZE(bp)       \
 334         (ASSERT(BP_IS_EMBEDDED(bp)), \
 335         BF64_GET_SB((bp)->blk_prop, 0, 25, 0, 1))
 336 #define BPE_SET_LSIZE(bp, x)    do { \
 337         ASSERT(BP_IS_EMBEDDED(bp)); \
 338         BF64_SET_SB((bp)->blk_prop, 0, 25, 0, 1, x); \
 339 } while (0)
 340
 341 #define BPE_GET_PSIZE(bp)       \
 342         (ASSERT(BP_IS_EMBEDDED(bp)), \
 343         BF64_GET_SB((bp)->blk_prop, 25, 7, 0, 1))
 344 #define BPE_SET_PSIZE(bp, x)    do { \
 345         ASSERT(BP_IS_EMBEDDED(bp)); \
 346         BF64_SET_SB((bp)->blk_prop, 25, 7, 0, 1, x); \
 347 } while (0)
 348
 349 typedef enum bp_embedded_type {
 350         BP_EMBEDDED_TYPE_DATA,
 351         BP_EMBEDDED_TYPE_RESERVED, /* Reserved for Delphix byteswap feature. */
 352         BP_EMBEDDED_TYPE_REDACTED,
 353         NUM_BP_EMBEDDED_TYPES
 354 } bp_embedded_type_t;
 355
 356 #define BPE_NUM_WORDS 14
 357 #define BPE_PAYLOAD_SIZE (BPE_NUM_WORDS * sizeof (uint64_t))
 358 #define BPE_IS_PAYLOADWORD(bp, wp) \
 359         ((wp) != &(bp)->blk_prop && (wp) != &(bp)->blk_birth)
 360
 361 #define SPA_BLKPTRSHIFT 7               /* blkptr_t is 128 bytes        */
 362 #define SPA_DVAS_PER_BP 3               /* Number of DVAs in a bp       */
 363 #define SPA_SYNC_MIN_VDEVS 3            /* min vdevs to update during sync */
 364
 365 /*
 366  * A block is a hole when it has either 1) never been written to, or
 367  * 2) is zero-filled. In both cases, ZFS can return all zeroes for all reads
 368  * without physically allocating disk space. Holes are represented in the
 369  * blkptr_t structure by zeroed blk_dva. Correct checking for holes is
 370  * done through the BP_IS_HOLE macro. For holes, the logical size, level,
 371  * DMU object type, and birth times are all also stored for holes that
 372  * were written to at some point (i.e. were punched after having been filled).
 373  */
 374 typedef struct blkptr {
 375         dva_t           blk_dva[SPA_DVAS_PER_BP]; /* Data Virtual Addresses */
 376         uint64_t        blk_prop;       /* size, compression, type, etc     */
 377         uint64_t        blk_pad[2];     /* Extra space for the future       */
 378         uint64_t        blk_phys_birth; /* txg when block was allocated     */
 379         uint64_t        blk_birth;      /* transaction group at birth       */
 380         uint64_t        blk_fill;       /* fill count                       */
 381         zio_cksum_t     blk_cksum;      /* 256-bit checksum                 */
 382 } blkptr_t;
 383
 384 /*
 385  * Macros to get and set fields in a bp or DVA.
 386  */
 387
 388 /*
 389  * Note, for gang blocks, DVA_GET_ASIZE() is the total space allocated for
 390  * this gang DVA including its children BP's.  The space allocated at this
 391  * DVA's vdev/offset is vdev_gang_header_asize(vdev).
 392  */
 393 #define DVA_GET_ASIZE(dva)      \
 394         BF64_GET_SB((dva)->dva_word[0], 0, SPA_ASIZEBITS, SPA_MINBLOCKSHIFT, 0)
 395 #define DVA_SET_ASIZE(dva, x)   \
 396         BF64_SET_SB((dva)->dva_word[0], 0, SPA_ASIZEBITS, \
 397         SPA_MINBLOCKSHIFT, 0, x)
 398
 399 #define DVA_GET_GRID(dva)       BF64_GET((dva)->dva_word[0], 24, 8)
 400 #define DVA_SET_GRID(dva, x)    BF64_SET((dva)->dva_word[0], 24, 8, x)
 401
 402 #define DVA_GET_VDEV(dva)       BF64_GET((dva)->dva_word[0], 32, SPA_VDEVBITS)
 403 #define DVA_SET_VDEV(dva, x)    \
 404         BF64_SET((dva)->dva_word[0], 32, SPA_VDEVBITS, x)
 405
 406 #define DVA_GET_OFFSET(dva)     \
 407         BF64_GET_SB((dva)->dva_word[1], 0, 63, SPA_MINBLOCKSHIFT, 0)
 408 #define DVA_SET_OFFSET(dva, x)  \
 409         BF64_SET_SB((dva)->dva_word[1], 0, 63, SPA_MINBLOCKSHIFT, 0, x)
 410
 411 #define DVA_GET_GANG(dva)       BF64_GET((dva)->dva_word[1], 63, 1)
 412 #define DVA_SET_GANG(dva, x)    BF64_SET((dva)->dva_word[1], 63, 1, x)
 413
 414 #define BP_GET_LSIZE(bp)        \
 415         (BP_IS_EMBEDDED(bp) ?   \
 416         (BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA ? BPE_GET_LSIZE(bp) : 0): \
 417         BF64_GET_SB((bp)->blk_prop, 0, SPA_LSIZEBITS, SPA_MINBLOCKSHIFT, 1))
 418 #define BP_SET_LSIZE(bp, x)     do { \
 419         ASSERT(!BP_IS_EMBEDDED(bp)); \
 420         BF64_SET_SB((bp)->blk_prop, \
 421             0, SPA_LSIZEBITS, SPA_MINBLOCKSHIFT, 1, x); \
 422 } while (0)
 423
 424 #define BP_GET_PSIZE(bp)        \
 425         (BP_IS_EMBEDDED(bp) ? 0 : \
 426         BF64_GET_SB((bp)->blk_prop, 16, SPA_PSIZEBITS, SPA_MINBLOCKSHIFT, 1))
 427 #define BP_SET_PSIZE(bp, x)     do { \
 428         ASSERT(!BP_IS_EMBEDDED(bp)); \
 429         BF64_SET_SB((bp)->blk_prop, \
 430             16, SPA_PSIZEBITS, SPA_MINBLOCKSHIFT, 1, x); \
 431 } while (0)
 432
 433 #define BP_GET_COMPRESS(bp)             \
 434         BF64_GET((bp)->blk_prop, 32, SPA_COMPRESSBITS)
 435 #define BP_SET_COMPRESS(bp, x)          \
 436         BF64_SET((bp)->blk_prop, 32, SPA_COMPRESSBITS, x)
 437
 438 #define BP_IS_EMBEDDED(bp)              BF64_GET((bp)->blk_prop, 39, 1)
 439 #define BP_SET_EMBEDDED(bp, x)          BF64_SET((bp)->blk_prop, 39, 1, x)
 440
 441 #define BP_GET_CHECKSUM(bp)             \
 442         (BP_IS_EMBEDDED(bp) ? ZIO_CHECKSUM_OFF : \
 443         BF64_GET((bp)->blk_prop, 40, 8))
 444 #define BP_SET_CHECKSUM(bp, x)          do { \
 445         ASSERT(!BP_IS_EMBEDDED(bp)); \
 446         BF64_SET((bp)->blk_prop, 40, 8, x); \
 447 } while (0)
 448
 449 #define BP_GET_TYPE(bp)                 BF64_GET((bp)->blk_prop, 48, 8)
 450 #define BP_SET_TYPE(bp, x)              BF64_SET((bp)->blk_prop, 48, 8, x)
 451
 452 #define BP_GET_LEVEL(bp)                BF64_GET((bp)->blk_prop, 56, 5)
 453 #define BP_SET_LEVEL(bp, x)             BF64_SET((bp)->blk_prop, 56, 5, x)
 454
 455 /* encrypted, authenticated, and MAC cksum bps use the same bit */
 456 #define BP_USES_CRYPT(bp)               BF64_GET((bp)->blk_prop, 61, 1)
 457 #define BP_SET_CRYPT(bp, x)             BF64_SET((bp)->blk_prop, 61, 1, x)
 458
 459 #define BP_IS_ENCRYPTED(bp)                     \
 460         (BP_USES_CRYPT(bp) &&                   \
 461         BP_GET_LEVEL(bp) <= 0 &&                \
 462         DMU_OT_IS_ENCRYPTED(BP_GET_TYPE(bp)))
 463
 464 #define BP_IS_AUTHENTICATED(bp)                 \
 465         (BP_USES_CRYPT(bp) &&                   \
 466         BP_GET_LEVEL(bp) <= 0 &&                \
 467         !DMU_OT_IS_ENCRYPTED(BP_GET_TYPE(bp)))
 468
 469 #define BP_HAS_INDIRECT_MAC_CKSUM(bp)           \
 470         (BP_USES_CRYPT(bp) && BP_GET_LEVEL(bp) > 0)
 471
 472 #define BP_IS_PROTECTED(bp)                     \
 473         (BP_IS_ENCRYPTED(bp) || BP_IS_AUTHENTICATED(bp))
 474
 475 #define BP_GET_DEDUP(bp)                BF64_GET((bp)->blk_prop, 62, 1)
 476 #define BP_SET_DEDUP(bp, x)             BF64_SET((bp)->blk_prop, 62, 1, x)
 477
 478 #define BP_GET_BYTEORDER(bp)            BF64_GET((bp)->blk_prop, 63, 1)
 479 #define BP_SET_BYTEORDER(bp, x)         BF64_SET((bp)->blk_prop, 63, 1, x)
 480
 481 #define BP_GET_FREE(bp)                 BF64_GET((bp)->blk_fill, 0, 1)
 482 #define BP_SET_FREE(bp, x)              BF64_SET((bp)->blk_fill, 0, 1, x)
 483
 484 #define BP_PHYSICAL_BIRTH(bp)           \
 485         (BP_IS_EMBEDDED(bp) ? 0 : \
 486         (bp)->blk_phys_birth ? (bp)->blk_phys_birth : (bp)->blk_birth)
 487
 488 #define BP_SET_BIRTH(bp, logical, physical)     \
 489 {                                               \
 490         ASSERT(!BP_IS_EMBEDDED(bp));            \
 491         (bp)->blk_birth = (logical);            \
 492         (bp)->blk_phys_birth = ((logical) == (physical) ? 0 : (physical)); \
 493 }
 494
 495 #define BP_GET_FILL(bp)                         \
 496         ((BP_IS_ENCRYPTED(bp)) ? BF64_GET((bp)->blk_fill, 0, 32) : \
 497         ((BP_IS_EMBEDDED(bp)) ? 1 : (bp)->blk_fill))
 498
 499 #define BP_SET_FILL(bp, fill)                   \
 500 {                                               \
 501         if (BP_IS_ENCRYPTED(bp))                        \
 502                 BF64_SET((bp)->blk_fill, 0, 32, fill); \
 503         else                                    \
 504                 (bp)->blk_fill = fill;          \
 505 }
 506
 507 #define BP_GET_IV2(bp)                          \
 508         (ASSERT(BP_IS_ENCRYPTED(bp)),           \
 509         BF64_GET((bp)->blk_fill, 32, 32))
 510 #define BP_SET_IV2(bp, iv2)                     \
 511 {                                               \
 512         ASSERT(BP_IS_ENCRYPTED(bp));            \
 513         BF64_SET((bp)->blk_fill, 32, 32, iv2);  \
 514 }
 515
 516 #define BP_IS_METADATA(bp)      \
 517         (BP_GET_LEVEL(bp) > 0 || DMU_OT_IS_METADATA(BP_GET_TYPE(bp)))
 518
 519 #define BP_GET_ASIZE(bp)        \
 520         (BP_IS_EMBEDDED(bp) ? 0 : \
 521         DVA_GET_ASIZE(&(bp)->blk_dva[0]) + \
 522         DVA_GET_ASIZE(&(bp)->blk_dva[1]) + \
 523         (DVA_GET_ASIZE(&(bp)->blk_dva[2]) * !BP_IS_ENCRYPTED(bp)))
 524
 525 #define BP_GET_UCSIZE(bp)       \
 526         (BP_IS_METADATA(bp) ? BP_GET_PSIZE(bp) : BP_GET_LSIZE(bp))
 527
 528 #define BP_GET_NDVAS(bp)        \
 529         (BP_IS_EMBEDDED(bp) ? 0 : \
 530         !!DVA_GET_ASIZE(&(bp)->blk_dva[0]) + \
 531         !!DVA_GET_ASIZE(&(bp)->blk_dva[1]) + \
 532         (!!DVA_GET_ASIZE(&(bp)->blk_dva[2]) * !BP_IS_ENCRYPTED(bp)))
 533
 534 #define BP_COUNT_GANG(bp)       \
 535         (BP_IS_EMBEDDED(bp) ? 0 : \
 536         (DVA_GET_GANG(&(bp)->blk_dva[0]) + \
 537         DVA_GET_GANG(&(bp)->blk_dva[1]) + \
 538         (DVA_GET_GANG(&(bp)->blk_dva[2]) * !BP_IS_ENCRYPTED(bp))))
 539
 540 #define DVA_EQUAL(dva1, dva2)   \
 541         ((dva1)->dva_word[1] == (dva2)->dva_word[1] && \
 542         (dva1)->dva_word[0] == (dva2)->dva_word[0])
 543
 544 #define BP_EQUAL(bp1, bp2)      \
 545         (BP_PHYSICAL_BIRTH(bp1) == BP_PHYSICAL_BIRTH(bp2) &&    \
 546         (bp1)->blk_birth == (bp2)->blk_birth &&                 \
 547         DVA_EQUAL(&(bp1)->blk_dva[0], &(bp2)->blk_dva[0]) &&    \
 548         DVA_EQUAL(&(bp1)->blk_dva[1], &(bp2)->blk_dva[1]) &&    \
 549         DVA_EQUAL(&(bp1)->blk_dva[2], &(bp2)->blk_dva[2]))
 550
 551
 552 #define DVA_IS_VALID(dva)       (DVA_GET_ASIZE(dva) != 0)
 553
 554 #define BP_IDENTITY(bp)         (ASSERT(!BP_IS_EMBEDDED(bp)), &(bp)->blk_dva[0])
 555 #define BP_IS_GANG(bp)          \
 556         (BP_IS_EMBEDDED(bp) ? B_FALSE : DVA_GET_GANG(BP_IDENTITY(bp)))
 557 #define DVA_IS_EMPTY(dva)       ((dva)->dva_word[0] == 0ULL &&  \
 558                                 (dva)->dva_word[1] == 0ULL)
 559 #define BP_IS_HOLE(bp) \
 560         (!BP_IS_EMBEDDED(bp) && DVA_IS_EMPTY(BP_IDENTITY(bp)))
 561
 562 #define BP_SET_REDACTED(bp) \
 563 {                                                       \
 564         BP_SET_EMBEDDED(bp, B_TRUE);                    \
 565         BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_REDACTED);   \
 566 }
 567 #define BP_IS_REDACTED(bp) \
 568         (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_REDACTED)
 569
 570 /* BP_IS_RAIDZ(bp) assumes no block compression */
 571 #define BP_IS_RAIDZ(bp)         (DVA_GET_ASIZE(&(bp)->blk_dva[0]) > \
 572                                 BP_GET_PSIZE(bp))
 573
 574 #define BP_ZERO(bp)                             \
 575 {                                               \
 576         (bp)->blk_dva[0].dva_word[0] = 0;       \
 577         (bp)->blk_dva[0].dva_word[1] = 0;       \
 578         (bp)->blk_dva[1].dva_word[0] = 0;       \
 579         (bp)->blk_dva[1].dva_word[1] = 0;       \
 580         (bp)->blk_dva[2].dva_word[0] = 0;       \
 581         (bp)->blk_dva[2].dva_word[1] = 0;       \
 582         (bp)->blk_prop = 0;                     \
 583         (bp)->blk_pad[0] = 0;                   \
 584         (bp)->blk_pad[1] = 0;                   \
 585         (bp)->blk_phys_birth = 0;               \
 586         (bp)->blk_birth = 0;                    \
 587         (bp)->blk_fill = 0;                     \
 588         ZIO_SET_CHECKSUM(&(bp)->blk_cksum, 0, 0, 0, 0); \
 589 }
 590
 591 #ifdef _ZFS_BIG_ENDIAN
 592 #define ZFS_HOST_BYTEORDER      (0ULL)
 593 #else
 594 #define ZFS_HOST_BYTEORDER      (1ULL)
 595 #endif
 596
 597 #define BP_SHOULD_BYTESWAP(bp)  (BP_GET_BYTEORDER(bp) != ZFS_HOST_BYTEORDER)
 598
 599 #define BP_SPRINTF_LEN  400
 600
 601 /*
 602  * This macro allows code sharing between zfs, libzpool, and mdb.
 603  * 'func' is either snprintf() or mdb_snprintf().
 604  * 'ws' (whitespace) can be ' ' for single-line format, '\n' for multi-line.
 605  */
 606
 607 #define SNPRINTF_BLKPTR(func, ws, buf, size, bp, type, checksum, compress) \
 608 {                                                                       \
 609         static const char *const copyname[] =                           \
 610             { "zero", "single", "double", "triple" };                   \
 611         int len = 0;                                                    \
 612         int copies = 0;                                                 \
 613         const char *crypt_type;                                         \
 614         if (bp != NULL) {                                               \
 615                 if (BP_IS_ENCRYPTED(bp)) {                              \
 616                         crypt_type = "encrypted";                       \
 617                         /* LINTED E_SUSPICIOUS_COMPARISON */            \
 618                 } else if (BP_IS_AUTHENTICATED(bp)) {                   \
 619                         crypt_type = "authenticated";                   \
 620                 } else if (BP_HAS_INDIRECT_MAC_CKSUM(bp)) {             \
 621                         crypt_type = "indirect-MAC";                    \
 622                 } else {                                                \
 623                         crypt_type = "unencrypted";                     \
 624                 }                                                       \
 625         }                                                               \
 626         if (bp == NULL) {                                               \
 627                 len += func(buf + len, size - len, "<NULL>");           \
 628         } else if (BP_IS_HOLE(bp)) {                                    \
 629                 len += func(buf + len, size - len,                      \
 630                     "HOLE [L%llu %s] "                                  \
 631                     "size=%llxL birth=%lluL",                           \
 632                     (u_longlong_t)BP_GET_LEVEL(bp),                     \
 633                     type,                                               \
 634                     (u_longlong_t)BP_GET_LSIZE(bp),                     \
 635                     (u_longlong_t)bp->blk_birth);                       \
 636         } else if (BP_IS_EMBEDDED(bp)) {                                \
 637                 len = func(buf + len, size - len,                       \
 638                     "EMBEDDED [L%llu %s] et=%u %s "                     \
 639                     "size=%llxL/%llxP birth=%lluL",                     \
 640                     (u_longlong_t)BP_GET_LEVEL(bp),                     \
 641                     type,                                               \
 642                     (int)BPE_GET_ETYPE(bp),                             \
 643                     compress,                                           \
 644                     (u_longlong_t)BPE_GET_LSIZE(bp),                    \
 645                     (u_longlong_t)BPE_GET_PSIZE(bp),                    \
 646                     (u_longlong_t)bp->blk_birth);                       \
 647         } else if (BP_IS_REDACTED(bp)) {                                \
 648                 len += func(buf + len, size - len,                      \
 649                     "REDACTED [L%llu %s] size=%llxL birth=%lluL",       \
 650                     (u_longlong_t)BP_GET_LEVEL(bp),                     \
 651                     type,                                               \
 652                     (u_longlong_t)BP_GET_LSIZE(bp),                     \
 653                     (u_longlong_t)bp->blk_birth);                       \
 654         } else {                                                        \
 655                 for (int d = 0; d < BP_GET_NDVAS(bp); d++) {            \
 656                         const dva_t *dva = &bp->blk_dva[d];             \
 657                         if (DVA_IS_VALID(dva))                          \
 658                                 copies++;                               \
 659                         len += func(buf + len, size - len,              \
 660                             "DVA[%d]=<%llu:%llx:%llx>%c", d,            \
 661                             (u_longlong_t)DVA_GET_VDEV(dva),            \
 662                             (u_longlong_t)DVA_GET_OFFSET(dva),          \
 663                             (u_longlong_t)DVA_GET_ASIZE(dva),           \
 664                             ws);                                        \
 665                 }                                                       \
 666                 if (BP_IS_ENCRYPTED(bp)) {                              \
 667                         len += func(buf + len, size - len,              \
 668                             "salt=%llx iv=%llx:%llx%c",                 \
 669                             (u_longlong_t)bp->blk_dva[2].dva_word[0],   \
 670                             (u_longlong_t)bp->blk_dva[2].dva_word[1],   \
 671                             (u_longlong_t)BP_GET_IV2(bp),               \
 672                             ws);                                        \
 673                 }                                                       \
 674                 if (BP_IS_GANG(bp) &&                                   \
 675                     DVA_GET_ASIZE(&bp->blk_dva[2]) <=                   \
 676                     DVA_GET_ASIZE(&bp->blk_dva[1]) / 2)                 \
 677                         copies--;                                       \
 678                 len += func(buf + len, size - len,                      \
 679                     "[L%llu %s] %s %s %s %s %s %s %s%c"                 \
 680                     "size=%llxL/%llxP birth=%lluL/%lluP fill=%llu%c"    \
 681                     "cksum=%llx:%llx:%llx:%llx",                        \
 682                     (u_longlong_t)BP_GET_LEVEL(bp),                     \
 683                     type,                                               \
 684                     checksum,                                           \
 685                     compress,                                           \
 686                     crypt_type,                                         \
 687                     BP_GET_BYTEORDER(bp) == 0 ? "BE" : "LE",            \
 688                     BP_IS_GANG(bp) ? "gang" : "contiguous",             \
 689                     BP_GET_DEDUP(bp) ? "dedup" : "unique",              \
 690                     copyname[copies],                                   \
 691                     ws,                                                 \
 692                     (u_longlong_t)BP_GET_LSIZE(bp),                     \
 693                     (u_longlong_t)BP_GET_PSIZE(bp),                     \
 694                     (u_longlong_t)bp->blk_birth,                        \
 695                     (u_longlong_t)BP_PHYSICAL_BIRTH(bp),                \
 696                     (u_longlong_t)BP_GET_FILL(bp),                      \
 697                     ws,                                                 \
 698                     (u_longlong_t)bp->blk_cksum.zc_word[0],             \
 699                     (u_longlong_t)bp->blk_cksum.zc_word[1],             \
 700                     (u_longlong_t)bp->blk_cksum.zc_word[2],             \
 701                     (u_longlong_t)bp->blk_cksum.zc_word[3]);            \
 702         }                                                               \
 703         ASSERT(len < size);                                             \
 704 }
 705
 706 #define BP_GET_BUFC_TYPE(bp)                                            \
 707         (BP_IS_METADATA(bp) ? ARC_BUFC_METADATA : ARC_BUFC_DATA)
 708
 709 typedef enum spa_import_type {
 710         SPA_IMPORT_EXISTING,
 711         SPA_IMPORT_ASSEMBLE
 712 } spa_import_type_t;
 713
 714 typedef enum spa_mode {
 715         SPA_MODE_UNINIT = 0,
 716         SPA_MODE_READ = 1,
 717         SPA_MODE_WRITE = 2,
 718 } spa_mode_t;
 719
 720 /*
 721  * Send TRIM commands in-line during normal pool operation while deleting.
 722  *      OFF: no
 723  *      ON: yes
 724  * NB: IN_FREEBSD_BASE is defined within the FreeBSD sources.
 725  */
 726 typedef enum {
 727         SPA_AUTOTRIM_OFF = 0,   /* default */
 728         SPA_AUTOTRIM_ON,
 729 #ifdef IN_FREEBSD_BASE
 730         SPA_AUTOTRIM_DEFAULT = SPA_AUTOTRIM_ON,
 731 #else
 732         SPA_AUTOTRIM_DEFAULT = SPA_AUTOTRIM_OFF,
 733 #endif
 734 } spa_autotrim_t;
 735
 736 /*
 737  * Reason TRIM command was issued, used internally for accounting purposes.
 738  */
 739 typedef enum trim_type {
 740         TRIM_TYPE_MANUAL = 0,
 741         TRIM_TYPE_AUTO = 1,
 742         TRIM_TYPE_SIMPLE = 2
 743 } trim_type_t;
 744
 745 /* state manipulation functions */
 746 extern int spa_open(const char *pool, spa_t **, const void *tag);
 747 extern int spa_open_rewind(const char *pool, spa_t **, const void *tag,
 748     nvlist_t *policy, nvlist_t **config);
 749 extern int spa_get_stats(const char *pool, nvlist_t **config, char *altroot,
 750     size_t buflen);
 751 extern int spa_create(const char *pool, nvlist_t *nvroot, nvlist_t *props,
 752     nvlist_t *zplprops, struct dsl_crypto_params *dcp);
 753 extern int spa_import(char *pool, nvlist_t *config, nvlist_t *props,
 754     uint64_t flags);
 755 extern nvlist_t *spa_tryimport(nvlist_t *tryconfig);
 756 extern int spa_destroy(const char *pool);
 757 extern int spa_checkpoint(const char *pool);
 758 extern int spa_checkpoint_discard(const char *pool);
 759 extern int spa_export(const char *pool, nvlist_t **oldconfig, boolean_t force,
 760     boolean_t hardforce);
 761 extern int spa_reset(const char *pool);
 762 extern void spa_async_request(spa_t *spa, int flag);
 763 extern void spa_async_unrequest(spa_t *spa, int flag);
 764 extern void spa_async_suspend(spa_t *spa);
 765 extern void spa_async_resume(spa_t *spa);
 766 extern int spa_async_tasks(spa_t *spa);
 767 extern spa_t *spa_inject_addref(char *pool);
 768 extern void spa_inject_delref(spa_t *spa);
 769 extern void spa_scan_stat_init(spa_t *spa);
 770 extern int spa_scan_get_stats(spa_t *spa, pool_scan_stat_t *ps);
 771 extern int bpobj_enqueue_alloc_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx);
 772 extern int bpobj_enqueue_free_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx);
 773
 774 #define SPA_ASYNC_CONFIG_UPDATE                 0x01
 775 #define SPA_ASYNC_REMOVE                        0x02
 776 #define SPA_ASYNC_PROBE                         0x04
 777 #define SPA_ASYNC_RESILVER_DONE                 0x08
 778 #define SPA_ASYNC_RESILVER                      0x10
 779 #define SPA_ASYNC_AUTOEXPAND                    0x20
 780 #define SPA_ASYNC_REMOVE_DONE                   0x40
 781 #define SPA_ASYNC_REMOVE_STOP                   0x80
 782 #define SPA_ASYNC_INITIALIZE_RESTART            0x100
 783 #define SPA_ASYNC_TRIM_RESTART                  0x200
 784 #define SPA_ASYNC_AUTOTRIM_RESTART              0x400
 785 #define SPA_ASYNC_L2CACHE_REBUILD               0x800
 786 #define SPA_ASYNC_L2CACHE_TRIM                  0x1000
 787 #define SPA_ASYNC_REBUILD_DONE                  0x2000
 788
 789 /* device manipulation */
 790 extern int spa_vdev_add(spa_t *spa, nvlist_t *nvroot);
 791 extern int spa_vdev_attach(spa_t *spa, uint64_t guid, nvlist_t *nvroot,
 792     int replacing, int rebuild);
 793 extern int spa_vdev_detach(spa_t *spa, uint64_t guid, uint64_t pguid,
 794     int replace_done);
 795 extern int spa_vdev_alloc(spa_t *spa, uint64_t guid);
 796 extern int spa_vdev_noalloc(spa_t *spa, uint64_t guid);
 797 extern boolean_t spa_vdev_remove_active(spa_t *spa);
 798 extern int spa_vdev_initialize(spa_t *spa, nvlist_t *nv, uint64_t cmd_type,
 799     nvlist_t *vdev_errlist);
 800 extern int spa_vdev_trim(spa_t *spa, nvlist_t *nv, uint64_t cmd_type,
 801     uint64_t rate, boolean_t partial, boolean_t secure, nvlist_t *vdev_errlist);
 802 extern int spa_vdev_setpath(spa_t *spa, uint64_t guid, const char *newpath);
 803 extern int spa_vdev_setfru(spa_t *spa, uint64_t guid, const char *newfru);
 804 extern int spa_vdev_split_mirror(spa_t *spa, const char *newname,
 805     nvlist_t *config, nvlist_t *props, boolean_t exp);
 806
 807 /* spare state (which is global across all pools) */
 808 extern void spa_spare_add(vdev_t *vd);
 809 extern void spa_spare_remove(vdev_t *vd);
 810 extern boolean_t spa_spare_exists(uint64_t guid, uint64_t *pool, int *refcnt);
 811 extern void spa_spare_activate(vdev_t *vd);
 812
 813 /* L2ARC state (which is global across all pools) */
 814 extern void spa_l2cache_add(vdev_t *vd);
 815 extern void spa_l2cache_remove(vdev_t *vd);
 816 extern boolean_t spa_l2cache_exists(uint64_t guid, uint64_t *pool);
 817 extern void spa_l2cache_activate(vdev_t *vd);
 818 extern void spa_l2cache_drop(spa_t *spa);
 819
 820 /* scanning */
 821 extern int spa_scan(spa_t *spa, pool_scan_func_t func);
 822 extern int spa_scan_stop(spa_t *spa);
 823 extern int spa_scrub_pause_resume(spa_t *spa, pool_scrub_cmd_t flag);
 824
 825 /* spa syncing */
 826 extern void spa_sync(spa_t *spa, uint64_t txg); /* only for DMU use */
 827 extern void spa_sync_allpools(void);
 828
 829 extern uint_t zfs_sync_pass_deferred_free;
 830
 831 /* spa namespace global mutex */
 832 extern kmutex_t spa_namespace_lock;
 833
 834 /*
 835  * SPA configuration functions in spa_config.c
 836  */
 837
 838 #define SPA_CONFIG_UPDATE_POOL  0
 839 #define SPA_CONFIG_UPDATE_VDEVS 1
 840
 841 extern void spa_write_cachefile(spa_t *, boolean_t, boolean_t, boolean_t);
 842 extern void spa_config_load(void);
 843 extern nvlist_t *spa_all_configs(uint64_t *);
 844 extern void spa_config_set(spa_t *spa, nvlist_t *config);
 845 extern nvlist_t *spa_config_generate(spa_t *spa, vdev_t *vd, uint64_t txg,
 846     int getstats);
 847 extern void spa_config_update(spa_t *spa, int what);
 848 extern int spa_config_parse(spa_t *spa, vdev_t **vdp, nvlist_t *nv,
 849     vdev_t *parent, uint_t id, int atype);
 850
 851
 852 /*
 853  * Miscellaneous SPA routines in spa_misc.c
 854  */
 855
 856 /* Namespace manipulation */
 857 extern spa_t *spa_lookup(const char *name);
 858 extern spa_t *spa_add(const char *name, nvlist_t *config, const char *altroot);
 859 extern void spa_remove(spa_t *spa);
 860 extern spa_t *spa_next(spa_t *prev);
 861
 862 /* Refcount functions */
 863 extern void spa_open_ref(spa_t *spa, const void *tag);
 864 extern void spa_close(spa_t *spa, const void *tag);
 865 extern void spa_async_close(spa_t *spa, const void *tag);
 866 extern boolean_t spa_refcount_zero(spa_t *spa);
 867
 868 #define SCL_NONE        0x00
 869 #define SCL_CONFIG      0x01
 870 #define SCL_STATE       0x02
 871 #define SCL_L2ARC       0x04            /* hack until L2ARC 2.0 */
 872 #define SCL_ALLOC       0x08
 873 #define SCL_ZIO         0x10
 874 #define SCL_FREE        0x20
 875 #define SCL_VDEV        0x40
 876 #define SCL_LOCKS       7
 877 #define SCL_ALL         ((1 << SCL_LOCKS) - 1)
 878 #define SCL_STATE_ALL   (SCL_STATE | SCL_L2ARC | SCL_ZIO)
 879
 880 /* Historical pool statistics */
 881 typedef struct spa_history_kstat {
 882         kmutex_t                lock;
 883         uint64_t                count;
 884         uint64_t                size;
 885         kstat_t                 *kstat;
 886         void                    *priv;
 887         list_t                  list;
 888 } spa_history_kstat_t;
 889
 890 typedef struct spa_history_list {
 891         uint64_t                size;
 892         procfs_list_t           procfs_list;
 893 } spa_history_list_t;
 894
 895 typedef struct spa_stats {
 896         spa_history_list_t      read_history;
 897         spa_history_list_t      txg_history;
 898         spa_history_kstat_t     tx_assign_histogram;
 899         spa_history_list_t      mmp_history;
 900         spa_history_kstat_t     state;          /* pool state */
 901         spa_history_kstat_t     guid;           /* pool guid */
 902         spa_history_kstat_t     iostats;
 903 } spa_stats_t;
 904
 905 typedef enum txg_state {
 906         TXG_STATE_BIRTH         = 0,
 907         TXG_STATE_OPEN          = 1,
 908         TXG_STATE_QUIESCED      = 2,
 909         TXG_STATE_WAIT_FOR_SYNC = 3,
 910         TXG_STATE_SYNCED        = 4,
 911         TXG_STATE_COMMITTED     = 5,
 912 } txg_state_t;
 913
 914 typedef struct txg_stat {
 915         vdev_stat_t             vs1;
 916         vdev_stat_t             vs2;
 917         uint64_t                txg;
 918         uint64_t                ndirty;
 919 } txg_stat_t;
 920
 921 /* Assorted pool IO kstats */
 922 typedef struct spa_iostats {
 923         kstat_named_t   trim_extents_written;
 924         kstat_named_t   trim_bytes_written;
 925         kstat_named_t   trim_extents_skipped;
 926         kstat_named_t   trim_bytes_skipped;
 927         kstat_named_t   trim_extents_failed;
 928         kstat_named_t   trim_bytes_failed;
 929         kstat_named_t   autotrim_extents_written;
 930         kstat_named_t   autotrim_bytes_written;
 931         kstat_named_t   autotrim_extents_skipped;
 932         kstat_named_t   autotrim_bytes_skipped;
 933         kstat_named_t   autotrim_extents_failed;
 934         kstat_named_t   autotrim_bytes_failed;
 935         kstat_named_t   simple_trim_extents_written;
 936         kstat_named_t   simple_trim_bytes_written;
 937         kstat_named_t   simple_trim_extents_skipped;
 938         kstat_named_t   simple_trim_bytes_skipped;
 939         kstat_named_t   simple_trim_extents_failed;
 940         kstat_named_t   simple_trim_bytes_failed;
 941 } spa_iostats_t;
 942
 943 extern void spa_stats_init(spa_t *spa);
 944 extern void spa_stats_destroy(spa_t *spa);
 945 extern void spa_read_history_add(spa_t *spa, const zbookmark_phys_t *zb,
 946     uint32_t aflags);
 947 extern void spa_txg_history_add(spa_t *spa, uint64_t txg, hrtime_t birth_time);
 948 extern int spa_txg_history_set(spa_t *spa,  uint64_t txg,
 949     txg_state_t completed_state, hrtime_t completed_time);
 950 extern txg_stat_t *spa_txg_history_init_io(spa_t *, uint64_t,
 951     struct dsl_pool *);
 952 extern void spa_txg_history_fini_io(spa_t *, txg_stat_t *);
 953 extern void spa_tx_assign_add_nsecs(spa_t *spa, uint64_t nsecs);
 954 extern int spa_mmp_history_set_skip(spa_t *spa, uint64_t mmp_kstat_id);
 955 extern int spa_mmp_history_set(spa_t *spa, uint64_t mmp_kstat_id, int io_error,
 956     hrtime_t duration);
 957 extern void spa_mmp_history_add(spa_t *spa, uint64_t txg, uint64_t timestamp,
 958     uint64_t mmp_delay, vdev_t *vd, int label, uint64_t mmp_kstat_id,
 959     int error);
 960 extern void spa_iostats_trim_add(spa_t *spa, trim_type_t type,
 961     uint64_t extents_written, uint64_t bytes_written,
 962     uint64_t extents_skipped, uint64_t bytes_skipped,
 963     uint64_t extents_failed, uint64_t bytes_failed);
 964 extern void spa_import_progress_add(spa_t *spa);
 965 extern void spa_import_progress_remove(uint64_t spa_guid);
 966 extern int spa_import_progress_set_mmp_check(uint64_t pool_guid,
 967     uint64_t mmp_sec_remaining);
 968 extern int spa_import_progress_set_max_txg(uint64_t pool_guid,
 969     uint64_t max_txg);
 970 extern int spa_import_progress_set_state(uint64_t pool_guid,
 971     spa_load_state_t spa_load_state);
 972
 973 /* Pool configuration locks */
 974 extern int spa_config_tryenter(spa_t *spa, int locks, const void *tag,
 975     krw_t rw);
 976 extern void spa_config_enter(spa_t *spa, int locks, const void *tag, krw_t rw);
 977 extern void spa_config_exit(spa_t *spa, int locks, const void *tag);
 978 extern int spa_config_held(spa_t *spa, int locks, krw_t rw);
 979
 980 /* Pool vdev add/remove lock */
 981 extern uint64_t spa_vdev_enter(spa_t *spa);
 982 extern uint64_t spa_vdev_detach_enter(spa_t *spa, uint64_t guid);
 983 extern uint64_t spa_vdev_config_enter(spa_t *spa);
 984 extern void spa_vdev_config_exit(spa_t *spa, vdev_t *vd, uint64_t txg,
 985     int error, const char *tag);
 986 extern int spa_vdev_exit(spa_t *spa, vdev_t *vd, uint64_t txg, int error);
 987
 988 /* Pool vdev state change lock */
 989 extern void spa_vdev_state_enter(spa_t *spa, int oplock);
 990 extern int spa_vdev_state_exit(spa_t *spa, vdev_t *vd, int error);
 991
 992 /* Log state */
 993 typedef enum spa_log_state {
 994         SPA_LOG_UNKNOWN = 0,    /* unknown log state */
 995         SPA_LOG_MISSING,        /* missing log(s) */
 996         SPA_LOG_CLEAR,          /* clear the log(s) */
 997         SPA_LOG_GOOD,           /* log(s) are good */
 998 } spa_log_state_t;
 999
1000 extern spa_log_state_t spa_get_log_state(spa_t *spa);
1001 extern void spa_set_log_state(spa_t *spa, spa_log_state_t state);
1002 extern int spa_reset_logs(spa_t *spa);
1003
1004 /* Log claim callback */
1005 extern void spa_claim_notify(zio_t *zio);
1006 extern void spa_deadman(void *);
1007
1008 /* Accessor functions */
1009 extern boolean_t spa_shutting_down(spa_t *spa);
1010 extern struct dsl_pool *spa_get_dsl(spa_t *spa);
1011 extern boolean_t spa_is_initializing(spa_t *spa);
1012 extern boolean_t spa_indirect_vdevs_loaded(spa_t *spa);
1013 extern blkptr_t *spa_get_rootblkptr(spa_t *spa);
1014 extern void spa_set_rootblkptr(spa_t *spa, const blkptr_t *bp);
1015 extern void spa_altroot(spa_t *, char *, size_t);
1016 extern uint32_t spa_sync_pass(spa_t *spa);
1017 extern char *spa_name(spa_t *spa);
1018 extern uint64_t spa_guid(spa_t *spa);
1019 extern uint64_t spa_load_guid(spa_t *spa);
1020 extern uint64_t spa_last_synced_txg(spa_t *spa);
1021 extern uint64_t spa_first_txg(spa_t *spa);
1022 extern uint64_t spa_syncing_txg(spa_t *spa);
1023 extern uint64_t spa_final_dirty_txg(spa_t *spa);
1024 extern uint64_t spa_version(spa_t *spa);
1025 extern pool_state_t spa_state(spa_t *spa);
1026 extern spa_load_state_t spa_load_state(spa_t *spa);
1027 extern uint64_t spa_freeze_txg(spa_t *spa);
1028 extern uint64_t spa_get_worst_case_asize(spa_t *spa, uint64_t lsize);
1029 extern uint64_t spa_get_dspace(spa_t *spa);
1030 extern uint64_t spa_get_checkpoint_space(spa_t *spa);
1031 extern uint64_t spa_get_slop_space(spa_t *spa);
1032 extern void spa_update_dspace(spa_t *spa);
1033 extern uint64_t spa_version(spa_t *spa);
1034 extern boolean_t spa_deflate(spa_t *spa);
1035 extern metaslab_class_t *spa_normal_class(spa_t *spa);
1036 extern metaslab_class_t *spa_log_class(spa_t *spa);
1037 extern metaslab_class_t *spa_embedded_log_class(spa_t *spa);
1038 extern metaslab_class_t *spa_special_class(spa_t *spa);
1039 extern metaslab_class_t *spa_dedup_class(spa_t *spa);
1040 extern metaslab_class_t *spa_preferred_class(spa_t *spa, uint64_t size,
1041     dmu_object_type_t objtype, uint_t level, uint_t special_smallblk);
1042
1043 extern void spa_evicting_os_register(spa_t *, objset_t *os);
1044 extern void spa_evicting_os_deregister(spa_t *, objset_t *os);
1045 extern void spa_evicting_os_wait(spa_t *spa);
1046 extern int spa_max_replication(spa_t *spa);
1047 extern int spa_prev_software_version(spa_t *spa);
1048 extern uint64_t spa_get_failmode(spa_t *spa);
1049 extern uint64_t spa_get_deadman_failmode(spa_t *spa);
1050 extern void spa_set_deadman_failmode(spa_t *spa, const char *failmode);
1051 extern boolean_t spa_suspended(spa_t *spa);
1052 extern uint64_t spa_bootfs(spa_t *spa);
1053 extern uint64_t spa_delegation(spa_t *spa);
1054 extern objset_t *spa_meta_objset(spa_t *spa);
1055 extern space_map_t *spa_syncing_log_sm(spa_t *spa);
1056 extern uint64_t spa_deadman_synctime(spa_t *spa);
1057 extern uint64_t spa_deadman_ziotime(spa_t *spa);
1058 extern uint64_t spa_dirty_data(spa_t *spa);
1059 extern spa_autotrim_t spa_get_autotrim(spa_t *spa);
1060
1061 /* Miscellaneous support routines */
1062 extern void spa_load_failed(spa_t *spa, const char *fmt, ...)
1063     __attribute__((format(printf, 2, 3)));
1064 extern void spa_load_note(spa_t *spa, const char *fmt, ...)
1065     __attribute__((format(printf, 2, 3)));
1066 extern void spa_activate_mos_feature(spa_t *spa, const char *feature,
1067     dmu_tx_t *tx);
1068 extern void spa_deactivate_mos_feature(spa_t *spa, const char *feature);
1069 extern spa_t *spa_by_guid(uint64_t pool_guid, uint64_t device_guid);
1070 extern boolean_t spa_guid_exists(uint64_t pool_guid, uint64_t device_guid);
1071 extern char *spa_strdup(const char *);
1072 extern void spa_strfree(char *);
1073 extern uint64_t spa_generate_guid(spa_t *spa);
1074 extern void snprintf_blkptr(char *buf, size_t buflen, const blkptr_t *bp);
1075 extern void spa_freeze(spa_t *spa);
1076 extern int spa_change_guid(spa_t *spa);
1077 extern void spa_upgrade(spa_t *spa, uint64_t version);
1078 extern void spa_evict_all(void);
1079 extern vdev_t *spa_lookup_by_guid(spa_t *spa, uint64_t guid,
1080     boolean_t l2cache);
1081 extern boolean_t spa_has_l2cache(spa_t *, uint64_t guid);
1082 extern boolean_t spa_has_spare(spa_t *, uint64_t guid);
1083 extern uint64_t dva_get_dsize_sync(spa_t *spa, const dva_t *dva);
1084 extern uint64_t bp_get_dsize_sync(spa_t *spa, const blkptr_t *bp);
1085 extern uint64_t bp_get_dsize(spa_t *spa, const blkptr_t *bp);
1086 extern boolean_t spa_has_slogs(spa_t *spa);
1087 extern boolean_t spa_is_root(spa_t *spa);
1088 extern boolean_t spa_writeable(spa_t *spa);
1089 extern boolean_t spa_has_pending_synctask(spa_t *spa);
1090 extern int spa_maxblocksize(spa_t *spa);
1091 extern int spa_maxdnodesize(spa_t *spa);
1092 extern boolean_t spa_has_checkpoint(spa_t *spa);
1093 extern boolean_t spa_importing_readonly_checkpoint(spa_t *spa);
1094 extern boolean_t spa_suspend_async_destroy(spa_t *spa);
1095 extern uint64_t spa_min_claim_txg(spa_t *spa);
1096 extern boolean_t zfs_dva_valid(spa_t *spa, const dva_t *dva,
1097     const blkptr_t *bp);
1098 typedef void (*spa_remap_cb_t)(uint64_t vdev, uint64_t offset, uint64_t size,
1099     void *arg);
1100 extern boolean_t spa_remap_blkptr(spa_t *spa, blkptr_t *bp,
1101     spa_remap_cb_t callback, void *arg);
1102 extern uint64_t spa_get_last_removal_txg(spa_t *spa);
1103 extern boolean_t spa_trust_config(spa_t *spa);
1104 extern uint64_t spa_missing_tvds_allowed(spa_t *spa);
1105 extern void spa_set_missing_tvds(spa_t *spa, uint64_t missing);
1106 extern boolean_t spa_top_vdevs_spacemap_addressable(spa_t *spa);
1107 extern uint64_t spa_total_metaslabs(spa_t *spa);
1108 extern boolean_t spa_multihost(spa_t *spa);
1109 extern uint32_t spa_get_hostid(spa_t *spa);
1110 extern void spa_activate_allocation_classes(spa_t *, dmu_tx_t *);
1111 extern boolean_t spa_livelist_delete_check(spa_t *spa);
1112
1113 extern spa_mode_t spa_mode(spa_t *spa);
1114 extern uint64_t zfs_strtonum(const char *str, char **nptr);
1115
1116 extern char *spa_his_ievent_table[];
1117
1118 extern void spa_history_create_obj(spa_t *spa, dmu_tx_t *tx);
1119 extern int spa_history_get(spa_t *spa, uint64_t *offset, uint64_t *len_read,
1120     char *his_buf);
1121 extern int spa_history_log(spa_t *spa, const char *his_buf);
1122 extern int spa_history_log_nvl(spa_t *spa, nvlist_t *nvl);
1123 extern void spa_history_log_version(spa_t *spa, const char *operation,
1124     dmu_tx_t *tx);
1125 extern void spa_history_log_internal(spa_t *spa, const char *operation,
1126     dmu_tx_t *tx, const char *fmt, ...) __printflike(4, 5);
1127 extern void spa_history_log_internal_ds(struct dsl_dataset *ds, const char *op,
1128     dmu_tx_t *tx, const char *fmt, ...)  __printflike(4, 5);
1129 extern void spa_history_log_internal_dd(dsl_dir_t *dd, const char *operation,
1130     dmu_tx_t *tx, const char *fmt, ...) __printflike(4, 5);
1131
1132 extern const char *spa_state_to_name(spa_t *spa);
1133
1134 /* error handling */
1135 struct zbookmark_phys;
1136 extern void spa_log_error(spa_t *spa, const zbookmark_phys_t *zb);
1137 extern void spa_remove_error(spa_t *spa, zbookmark_phys_t *zb);
1138 extern int zfs_ereport_post(const char *clazz, spa_t *spa, vdev_t *vd,
1139     const zbookmark_phys_t *zb, zio_t *zio, uint64_t state);
1140 extern boolean_t zfs_ereport_is_valid(const char *clazz, spa_t *spa, vdev_t *vd,
1141     zio_t *zio);
1142 extern void zfs_ereport_taskq_fini(void);
1143 extern void zfs_ereport_clear(spa_t *spa, vdev_t *vd);
1144 extern nvlist_t *zfs_event_create(spa_t *spa, vdev_t *vd, const char *type,
1145     const char *name, nvlist_t *aux);
1146 extern void zfs_post_remove(spa_t *spa, vdev_t *vd);
1147 extern void zfs_post_state_change(spa_t *spa, vdev_t *vd, uint64_t laststate);
1148 extern void zfs_post_autoreplace(spa_t *spa, vdev_t *vd);
1149 extern uint64_t spa_get_errlog_size(spa_t *spa);
1150 extern int spa_get_errlog(spa_t *spa, void *uaddr, uint64_t *count);
1151 extern void spa_errlog_rotate(spa_t *spa);
1152 extern void spa_errlog_drain(spa_t *spa);
1153 extern void spa_errlog_sync(spa_t *spa, uint64_t txg);
1154 extern void spa_get_errlists(spa_t *spa, avl_tree_t *last, avl_tree_t *scrub);
1155 extern void spa_delete_dataset_errlog(spa_t *spa, uint64_t ds, dmu_tx_t *tx);
1156 extern void spa_swap_errlog(spa_t *spa, uint64_t new_head_ds,
1157     uint64_t old_head_ds, dmu_tx_t *tx);
1158 extern void sync_error_list(spa_t *spa, avl_tree_t *t, uint64_t *obj,
1159     dmu_tx_t *tx);
1160 extern void spa_upgrade_errlog(spa_t *spa, dmu_tx_t *tx);
1161
1162 /* vdev cache */
1163 extern void vdev_cache_stat_init(void);
1164 extern void vdev_cache_stat_fini(void);
1165
1166 /* vdev mirror */
1167 extern void vdev_mirror_stat_init(void);
1168 extern void vdev_mirror_stat_fini(void);
1169
1170 /* Initialization and termination */
1171 extern void spa_init(spa_mode_t mode);
1172 extern void spa_fini(void);
1173 extern void spa_boot_init(void);
1174
1175 /* properties */
1176 extern int spa_prop_set(spa_t *spa, nvlist_t *nvp);
1177 extern int spa_prop_get(spa_t *spa, nvlist_t **nvp);
1178 extern void spa_prop_clear_bootfs(spa_t *spa, uint64_t obj, dmu_tx_t *tx);
1179 extern void spa_configfile_set(spa_t *, nvlist_t *, boolean_t);
1180
1181 /* asynchronous event notification */
1182 extern void spa_event_notify(spa_t *spa, vdev_t *vdev, nvlist_t *hist_nvl,
1183     const char *name);
1184 extern void zfs_ereport_zvol_post(const char *subclass, const char *name,
1185     const char *device_name, const char *raw_name);
1186
1187 /* waiting for pool activities to complete */
1188 extern int spa_wait(const char *pool, zpool_wait_activity_t activity,
1189     boolean_t *waited);
1190 extern int spa_wait_tag(const char *name, zpool_wait_activity_t activity,
1191     uint64_t tag, boolean_t *waited);
1192 extern void spa_notify_waiters(spa_t *spa);
1193 extern void spa_wake_waiters(spa_t *spa);
1194
1195 extern void spa_import_os(spa_t *spa);
1196 extern void spa_export_os(spa_t *spa);
1197 extern void spa_activate_os(spa_t *spa);
1198 extern void spa_deactivate_os(spa_t *spa);
1199
1200 /* module param call functions */
1201 int param_set_deadman_ziotime(ZFS_MODULE_PARAM_ARGS);
1202 int param_set_deadman_synctime(ZFS_MODULE_PARAM_ARGS);
1203 int param_set_slop_shift(ZFS_MODULE_PARAM_ARGS);
1204 int param_set_deadman_failmode(ZFS_MODULE_PARAM_ARGS);
1205
1206 #ifdef ZFS_DEBUG
1207 #define dprintf_bp(bp, fmt, ...) do {                           \
1208         if (zfs_flags & ZFS_DEBUG_DPRINTF) {                    \
1209         char *__blkbuf = kmem_alloc(BP_SPRINTF_LEN, KM_SLEEP);  \
1210         snprintf_blkptr(__blkbuf, BP_SPRINTF_LEN, (bp));        \
1211         dprintf(fmt " %s\n", __VA_ARGS__, __blkbuf);            \
1212         kmem_free(__blkbuf, BP_SPRINTF_LEN);                    \
1213         } \
1214 } while (0)
1215 #else
1216 #define dprintf_bp(bp, fmt, ...)
1217 #endif
1218
1219 extern spa_mode_t spa_mode_global;
1220 extern int zfs_deadman_enabled;
1221 extern unsigned long zfs_deadman_synctime_ms;
1222 extern unsigned long zfs_deadman_ziotime_ms;
1223 extern unsigned long zfs_deadman_checktime_ms;
1224
1225 extern kmem_cache_t *zio_buf_cache[];
1226 extern kmem_cache_t *zio_data_buf_cache[];
1227
1228 #ifdef  __cplusplus
1229 }
1230 #endif
1231
1232 #endif  /* _SYS_SPA_H */