include/zfs/spa.h

   1 /*
   2  *  GRUB  --  GRand Unified Bootloader
   3  *  Copyright (C) 1999,2000,2001,2002,2003,2004  Free Software Foundation, Inc.
   4  *
   5  * SPDX-License-Identifier:     GPL-2.0+
   6  */
   7 /*
   8  * Copyright (c) 2008, 2011, Oracle and/or its affiliates. All rights reserved.
   9  */
  10
  11 #ifndef ZFS_SPA_HEADER
  12 #define ZFS_SPA_HEADER 1
  13
  14
  15 /*
  16  * General-purpose 32-bit and 64-bit bitfield encodings.
  17  */
  18 #define BF32_DECODE(x, low, len)        P2PHASE((x) >> (low), 1U << (len))
  19 #define BF64_DECODE(x, low, len)        P2PHASE((x) >> (low), 1ULL << (len))
  20 #define BF32_ENCODE(x, low, len)        (P2PHASE((x), 1U << (len)) << (low))
  21 #define BF64_ENCODE(x, low, len)        (P2PHASE((x), 1ULL << (len)) << (low))
  22
  23 #define BF32_GET(x, low, len)           BF32_DECODE(x, low, len)
  24 #define BF64_GET(x, low, len)           BF64_DECODE(x, low, len)
  25
  26 #define BF32_SET(x, low, len, val)                                              \
  27         ((x) ^= BF32_ENCODE((x >> low) ^ (val), low, len))
  28 #define BF64_SET(x, low, len, val)                                              \
  29         ((x) ^= BF64_ENCODE((x >> low) ^ (val), low, len))
  30
  31 #define BF32_GET_SB(x, low, len, shift, bias)           \
  32         ((BF32_GET(x, low, len) + (bias)) << (shift))
  33 #define BF64_GET_SB(x, low, len, shift, bias)           \
  34         ((BF64_GET(x, low, len) + (bias)) << (shift))
  35
  36 #define BF32_SET_SB(x, low, len, shift, bias, val)              \
  37         BF32_SET(x, low, len, ((val) >> (shift)) - (bias))
  38 #define BF64_SET_SB(x, low, len, shift, bias, val)              \
  39         BF64_SET(x, low, len, ((val) >> (shift)) - (bias))
  40
  41 /*
  42  * We currently support nine block sizes, from 512 bytes to 128K.
  43  * We could go higher, but the benefits are near-zero and the cost
  44  * of COWing a giant block to modify one byte would become excessive.
  45  */
  46 #define SPA_MINBLOCKSHIFT       9
  47 #define SPA_MAXBLOCKSHIFT       17
  48 #define SPA_MINBLOCKSIZE        (1ULL << SPA_MINBLOCKSHIFT)
  49 #define SPA_MAXBLOCKSIZE        (1ULL << SPA_MAXBLOCKSHIFT)
  50
  51 #define SPA_BLOCKSIZES          (SPA_MAXBLOCKSHIFT - SPA_MINBLOCKSHIFT + 1)
  52
  53 /*
  54  * Size of block to hold the configuration data (a packed nvlist)
  55  */
  56 #define SPA_CONFIG_BLOCKSIZE    (1 << 14)
  57
  58 /*
  59  * The DVA size encodings for LSIZE and PSIZE support blocks up to 32MB.
  60  * The ASIZE encoding should be at least 64 times larger (6 more bits)
  61  * to support up to 4-way RAID-Z mirror mode with worst-case gang block
  62  * overhead, three DVAs per bp, plus one more bit in case we do anything
  63  * else that expands the ASIZE.
  64  */
  65 #define SPA_LSIZEBITS           16      /* LSIZE up to 32M (2^16 * 512) */
  66 #define SPA_PSIZEBITS           16      /* PSIZE up to 32M (2^16 * 512) */
  67 #define SPA_ASIZEBITS           24      /* ASIZE up to 64 times larger  */
  68
  69 /*
  70  * All SPA data is represented by 128-bit data virtual addresses (DVAs).
  71  * The members of the dva_t should be considered opaque outside the SPA.
  72  */
  73 typedef struct dva {
  74         uint64_t        dva_word[2];
  75 } dva_t;
  76
  77 /*
  78  * Each block has a 256-bit checksum -- strong enough for cryptographic hashes.
  79  */
  80 typedef struct zio_cksum {
  81         uint64_t        zc_word[4];
  82 } zio_cksum_t;
  83
  84 /*
  85  * Each block is described by its DVAs, time of birth, checksum, etc.
  86  * The word-by-word, bit-by-bit layout of the blkptr is as follows:
  87  *
  88  *      64      56      48      40      32      24      16      8       0
  89  *      +-------+-------+-------+-------+-------+-------+-------+-------+
  90  * 0    |               vdev1           | GRID  |         ASIZE         |
  91  *      +-------+-------+-------+-------+-------+-------+-------+-------+
  92  * 1    |G|                      offset1                                |
  93  *      +-------+-------+-------+-------+-------+-------+-------+-------+
  94  * 2    |               vdev2           | GRID  |         ASIZE         |
  95  *      +-------+-------+-------+-------+-------+-------+-------+-------+
  96  * 3    |G|                      offset2                                |
  97  *      +-------+-------+-------+-------+-------+-------+-------+-------+
  98  * 4    |               vdev3           | GRID  |         ASIZE         |
  99  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 100  * 5    |G|                      offset3                                |
 101  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 102  * 6    |BDX|lvl| type  | cksum | comp  |         PSIZE |         LSIZE |
 103  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 104  * 7    |                       padding                                 |
 105  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 106  * 8    |                       padding                                 |
 107  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 108  * 9    |                       physical birth txg                      |
 109  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 110  * a    |                       logical birth txg                       |
 111  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 112  * b    |                       fill count                              |
 113  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 114  * c    |                       checksum[0]                             |
 115  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 116  * d    |                       checksum[1]                             |
 117  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 118  * e    |                       checksum[2]                             |
 119  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 120  * f    |                       checksum[3]                             |
 121  *      +-------+-------+-------+-------+-------+-------+-------+-------+
 122  *
 123  * Legend:
 124  *
 125  * vdev         virtual device ID
 126  * offset       offset into virtual device
 127  * LSIZE        logical size
 128  * PSIZE        physical size (after compression)
 129  * ASIZE        allocated size (including RAID-Z parity and gang block headers)
 130  * GRID         RAID-Z layout information (reserved for future use)
 131  * cksum        checksum function
 132  * comp         compression function
 133  * G            gang block indicator
 134  * B            byteorder (endianness)
 135  * D            dedup
 136  * X            unused
 137  * lvl          level of indirection
 138  * type         DMU object type
 139  * phys birth   txg of block allocation; zero if same as logical birth txg
 140  * log. birth   transaction group in which the block was logically born
 141  * fill count   number of non-zero blocks under this bp
 142  * checksum[4]  256-bit checksum of the data this bp describes
 143  */
 144 #define SPA_BLKPTRSHIFT 7               /* blkptr_t is 128 bytes        */
 145 #define SPA_DVAS_PER_BP 3               /* Number of DVAs in a bp       */
 146
 147 typedef struct blkptr {
 148         dva_t           blk_dva[SPA_DVAS_PER_BP]; /* Data Virtual Addresses */
 149         uint64_t        blk_prop;       /* size, compression, type, etc         */
 150         uint64_t        blk_pad[2];     /* Extra space for the future           */
 151         uint64_t        blk_phys_birth; /* txg when block was allocated         */
 152         uint64_t        blk_birth;      /* transaction group at birth           */
 153         uint64_t        blk_fill;       /* fill count                           */
 154         zio_cksum_t     blk_cksum;      /* 256-bit checksum                     */
 155 } blkptr_t;
 156
 157 /*
 158  * Macros to get and set fields in a bp or DVA.
 159  */
 160 #define DVA_GET_ASIZE(dva)                                                                                      \
 161         BF64_GET_SB((dva)->dva_word[0], 0, 24, SPA_MINBLOCKSHIFT, 0)
 162 #define DVA_SET_ASIZE(dva, x)                                                                           \
 163         BF64_SET_SB((dva)->dva_word[0], 0, 24, SPA_MINBLOCKSHIFT, 0, x)
 164
 165 #define DVA_GET_GRID(dva)       BF64_GET((dva)->dva_word[0], 24, 8)
 166 #define DVA_SET_GRID(dva, x)    BF64_SET((dva)->dva_word[0], 24, 8, x)
 167
 168 #define DVA_GET_VDEV(dva)       BF64_GET((dva)->dva_word[0], 32, 32)
 169 #define DVA_SET_VDEV(dva, x)    BF64_SET((dva)->dva_word[0], 32, 32, x)
 170
 171 #define DVA_GET_GANG(dva)       BF64_GET((dva)->dva_word[1], 63, 1)
 172 #define DVA_SET_GANG(dva, x)    BF64_SET((dva)->dva_word[1], 63, 1, x)
 173
 174 #define BP_GET_LSIZE(bp)                                                                                \
 175         BF64_GET_SB((bp)->blk_prop, 0, 16, SPA_MINBLOCKSHIFT, 1)
 176 #define BP_SET_LSIZE(bp, x)                                                                             \
 177         BF64_SET_SB((bp)->blk_prop, 0, 16, SPA_MINBLOCKSHIFT, 1, x)
 178
 179 #define BP_GET_COMPRESS(bp)             BF64_GET((bp)->blk_prop, 32, 8)
 180 #define BP_SET_COMPRESS(bp, x)          BF64_SET((bp)->blk_prop, 32, 8, x)
 181
 182 #define BP_GET_CHECKSUM(bp)             BF64_GET((bp)->blk_prop, 40, 8)
 183 #define BP_SET_CHECKSUM(bp, x)          BF64_SET((bp)->blk_prop, 40, 8, x)
 184
 185 #define BP_GET_TYPE(bp)                 BF64_GET((bp)->blk_prop, 48, 8)
 186 #define BP_SET_TYPE(bp, x)              BF64_SET((bp)->blk_prop, 48, 8, x)
 187
 188 #define BP_GET_LEVEL(bp)                BF64_GET((bp)->blk_prop, 56, 5)
 189 #define BP_SET_LEVEL(bp, x)             BF64_SET((bp)->blk_prop, 56, 5, x)
 190
 191 #define BP_GET_PROP_BIT_61(bp)          BF64_GET((bp)->blk_prop, 61, 1)
 192 #define BP_SET_PROP_BIT_61(bp, x)       BF64_SET((bp)->blk_prop, 61, 1, x)
 193
 194 #define BP_GET_DEDUP(bp)                BF64_GET((bp)->blk_prop, 62, 1)
 195 #define BP_SET_DEDUP(bp, x)             BF64_SET((bp)->blk_prop, 62, 1, x)
 196
 197 #define BP_GET_BYTEORDER(bp)            (0 - BF64_GET((bp)->blk_prop, 63, 1))
 198 #define BP_SET_BYTEORDER(bp, x)         BF64_SET((bp)->blk_prop, 63, 1, x)
 199
 200 #define BP_PHYSICAL_BIRTH(bp)                                                                           \
 201         ((bp)->blk_phys_birth ? (bp)->blk_phys_birth : (bp)->blk_birth)
 202
 203 #define BP_SET_BIRTH(bp, logical, physical)                                                             \
 204         {                                                                                                                                       \
 205                 (bp)->blk_birth = (logical);                                                                    \
 206                 (bp)->blk_phys_birth = ((logical) == (physical) ? 0 : (physical)); \
 207         }
 208
 209 #define BP_GET_ASIZE(bp)                                                                                                \
 210         (DVA_GET_ASIZE(&(bp)->blk_dva[0]) + DVA_GET_ASIZE(&(bp)->blk_dva[1]) + \
 211          DVA_GET_ASIZE(&(bp)->blk_dva[2]))
 212
 213 #define BP_GET_UCSIZE(bp)                                                                                               \
 214         ((BP_GET_LEVEL(bp) > 0 || dmu_ot[BP_GET_TYPE(bp)].ot_metadata) ?        \
 215          BP_GET_PSIZE(bp) : BP_GET_LSIZE(bp));
 216
 217 #define BP_GET_NDVAS(bp)                                                \
 218         (!!DVA_GET_ASIZE(&(bp)->blk_dva[0]) +           \
 219          !!DVA_GET_ASIZE(&(bp)->blk_dva[1]) +           \
 220          !!DVA_GET_ASIZE(&(bp)->blk_dva[2]))
 221
 222 #define BP_COUNT_GANG(bp)                                               \
 223         (DVA_GET_GANG(&(bp)->blk_dva[0]) +                      \
 224          DVA_GET_GANG(&(bp)->blk_dva[1]) +                      \
 225          DVA_GET_GANG(&(bp)->blk_dva[2]))
 226
 227 #define DVA_EQUAL(dva1, dva2)                                           \
 228         ((dva1)->dva_word[1] == (dva2)->dva_word[1] &&  \
 229          (dva1)->dva_word[0] == (dva2)->dva_word[0])
 230
 231 #define BP_EQUAL(bp1, bp2)                                                                      \
 232         (BP_PHYSICAL_BIRTH(bp1) == BP_PHYSICAL_BIRTH(bp2) &&    \
 233          DVA_EQUAL(&(bp1)->blk_dva[0], &(bp2)->blk_dva[0]) &&   \
 234          DVA_EQUAL(&(bp1)->blk_dva[1], &(bp2)->blk_dva[1]) &&   \
 235          DVA_EQUAL(&(bp1)->blk_dva[2], &(bp2)->blk_dva[2]))
 236
 237 #define ZIO_CHECKSUM_EQUAL(zc1, zc2)                            \
 238         (0 == (((zc1).zc_word[0] - (zc2).zc_word[0]) |  \
 239                    ((zc1).zc_word[1] - (zc2).zc_word[1]) |      \
 240                    ((zc1).zc_word[2] - (zc2).zc_word[2]) |      \
 241                    ((zc1).zc_word[3] - (zc2).zc_word[3])))
 242
 243 #define DVA_IS_VALID(dva)       (DVA_GET_ASIZE(dva) != 0)
 244
 245 #define ZIO_SET_CHECKSUM(zcp, w0, w1, w2, w3)   \
 246         {                                                                                       \
 247                 (zcp)->zc_word[0] = w0;                                 \
 248                 (zcp)->zc_word[1] = w1;                                 \
 249                 (zcp)->zc_word[2] = w2;                                 \
 250                 (zcp)->zc_word[3] = w3;                                 \
 251         }
 252
 253 #define BP_IDENTITY(bp)         (&(bp)->blk_dva[0])
 254 #define BP_IS_GANG(bp)          DVA_GET_GANG(BP_IDENTITY(bp))
 255 #define BP_IS_HOLE(bp)          ((bp)->blk_birth == 0)
 256
 257 /* BP_IS_RAIDZ(bp) assumes no block compression */
 258 #define BP_IS_RAIDZ(bp)         (DVA_GET_ASIZE(&(bp)->blk_dva[0]) > \
 259                                                          BP_GET_PSIZE(bp))
 260
 261 #define BP_ZERO(bp)                                                                             \
 262         {                                                                                                       \
 263                 (bp)->blk_dva[0].dva_word[0] = 0;                               \
 264                 (bp)->blk_dva[0].dva_word[1] = 0;                               \
 265                 (bp)->blk_dva[1].dva_word[0] = 0;                               \
 266                 (bp)->blk_dva[1].dva_word[1] = 0;                               \
 267                 (bp)->blk_dva[2].dva_word[0] = 0;                               \
 268                 (bp)->blk_dva[2].dva_word[1] = 0;                               \
 269                 (bp)->blk_prop = 0;                                                             \
 270                 (bp)->blk_pad[0] = 0;                                                   \
 271                 (bp)->blk_pad[1] = 0;                                                   \
 272                 (bp)->blk_phys_birth = 0;                                               \
 273                 (bp)->blk_birth = 0;                                                    \
 274                 (bp)->blk_fill = 0;                                                             \
 275                 ZIO_SET_CHECKSUM(&(bp)->blk_cksum, 0, 0, 0, 0); \
 276         }
 277
 278 #define BP_SPRINTF_LEN  320
 279
 280 #endif  /* ! ZFS_SPA_HEADER */