1 /* 2 * Copyright (c) 2011-2012 The DragonFly Project. All rights reserved. 3 * 4 * This code is derived from software contributed to The DragonFly Project 5 * by Matthew Dillon <dillon@dragonflybsd.org> 6 * by Venkatesh Srinivas <vsrinivas@dragonflybsd.org> 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 12 * 1. Redistributions of source code must retain the above copyright 13 * notice, this list of conditions and the following disclaimer. 14 * 2. Redistributions in binary form must reproduce the above copyright 15 * notice, this list of conditions and the following disclaimer in 16 * the documentation and/or other materials provided with the 17 * distribution. 18 * 3. Neither the name of The DragonFly Project nor the names of its 19 * contributors may be used to endorse or promote products derived 20 * from this software without specific, prior written permission. 21 * 22 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 23 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 24 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 25 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 26 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 27 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 28 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 29 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 30 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 31 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 32 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 33 * SUCH DAMAGE. 34 */ 35 #ifndef VFS_HAMMER2_DISK_H_ 36 #define VFS_HAMMER2_DISK_H_ 37 38 #ifndef _SYS_UUID_H_ 39 #include <sys/uuid.h> 40 #endif 41 #ifndef _SYS_DMSG_H_ 42 #include <sys/dmsg.h> 43 #endif 44 45 /* 46 * The structures below represent the on-disk media structures for the HAMMER2 47 * filesystem. Note that all fields for on-disk structures are naturally 48 * aligned. The host endian format is typically used - compatibility is 49 * possible if the implementation detects reversed endian and adjusts accesses 50 * accordingly. 51 * 52 * HAMMER2 primarily revolves around the directory topology: inodes, 53 * directory entries, and block tables. Block device buffer cache buffers 54 * are always 64KB. Logical file buffers are typically 16KB. All data 55 * references utilize 64-bit byte offsets. 56 * 57 * Free block management is handled independently using blocks reserved by 58 * the media topology. 59 */ 60 61 /* 62 * The data at the end of a file or directory may be a fragment in order 63 * to optimize storage efficiency. The minimum fragment size is 1KB. 64 * Since allocations are in powers of 2 fragments must also be sized in 65 * powers of 2 (1024, 2048, ... 65536). 66 * 67 * For the moment the maximum allocation size is HAMMER2_PBUFSIZE (64K), 68 * which is 2^16. Larger extents may be supported in the future. Smaller 69 * fragments might be supported in the future (down to 64 bytes is possible), 70 * but probably will not be. 71 * 72 * A full indirect block use supports 1024 x 64-byte blockrefs in a 64KB 73 * buffer. Indirect blocks down to 1KB are supported to keep small 74 * directories small. 75 * 76 * A maximally sized file (2^64-1 bytes) requires 5 indirect block levels. 77 * The hammer2_blockset in the volume header or file inode has another 8 78 * entries, giving us 66+3 = 69 bits of address space. However, some bits 79 * are taken up by (potentially) requests for redundant copies. HAMMER2 80 * currently supports up to 8 copies, which brings the address space down 81 * to 66 bits and gives us 2 bits of leeway. 82 */ 83 #define HAMMER2_MIN_ALLOC 1024 /* minimum allocation size */ 84 #define HAMMER2_MIN_RADIX 10 /* minimum allocation size 2^N */ 85 #define HAMMER2_MAX_ALLOC 65536 /* maximum allocation size */ 86 #define HAMMER2_MAX_RADIX 16 /* maximum allocation size 2^N */ 87 #define HAMMER2_KEY_RADIX 64 /* number of bits in key */ 88 89 /* 90 * MINALLOCSIZE - The minimum allocation size. This can be smaller 91 * or larger than the minimum physical IO size. 92 * 93 * NOTE: Should not be larger than 1K since inodes 94 * are 1K. 95 * 96 * MINIOSIZE - The minimum IO size. This must be less than 97 * or equal to HAMMER2_LBUFSIZE. 98 * 99 * HAMMER2_LBUFSIZE - Nominal buffer size for I/O rollups. 100 * 101 * HAMMER2_PBUFSIZE - Topological block size used by files for all 102 * blocks except the block straddling EOF. 103 * 104 * HAMMER2_SEGSIZE - Allocation map segment size, typically 2MB 105 * (space represented by a level0 bitmap). 106 */ 107 108 #define HAMMER2_SEGSIZE (1 << HAMMER2_FREEMAP_LEVEL0_RADIX) 109 #define HAMMER2_SEGRADIX HAMMER2_FREEMAP_LEVEL0_RADIX 110 111 #define HAMMER2_PBUFRADIX 16 /* physical buf (1<<16) bytes */ 112 #define HAMMER2_PBUFSIZE 65536 113 #define HAMMER2_LBUFRADIX 14 /* logical buf (1<<14) bytes */ 114 #define HAMMER2_LBUFSIZE 16384 115 116 /* 117 * Generally speaking we want to use 16K and 64K I/Os 118 */ 119 #define HAMMER2_MINIORADIX HAMMER2_LBUFRADIX 120 #define HAMMER2_MINIOSIZE HAMMER2_LBUFSIZE 121 122 #define HAMMER2_IND_BYTES_MIN HAMMER2_LBUFSIZE 123 #define HAMMER2_IND_BYTES_MAX HAMMER2_PBUFSIZE 124 #define HAMMER2_IND_COUNT_MIN (HAMMER2_IND_BYTES_MIN / \ 125 sizeof(hammer2_blockref_t)) 126 #define HAMMER2_IND_COUNT_MAX (HAMMER2_IND_BYTES_MAX / \ 127 sizeof(hammer2_blockref_t)) 128 129 /* 130 * In HAMMER2, arrays of blockrefs are fully set-associative, meaning that 131 * any element can occur at any index and holes can be anywhere. As a 132 * future optimization we will be able to flag that such arrays are sorted 133 * and thus optimize lookups, but for now we don't. 134 * 135 * Inodes embed either 512 bytes of direct data or an array of 8 blockrefs, 136 * resulting in highly efficient storage for files <= 512 bytes and for files 137 * <= 512KB. Up to 8 directory entries can be referenced from a directory 138 * without requiring an indirect block. 139 * 140 * Indirect blocks are typically either 4KB (64 blockrefs / ~4MB represented), 141 * or 64KB (1024 blockrefs / ~64MB represented). 142 */ 143 #define HAMMER2_SET_COUNT 8 /* direct entries */ 144 #define HAMMER2_SET_RADIX 3 145 #define HAMMER2_EMBEDDED_BYTES 512 /* inode blockset/dd size */ 146 #define HAMMER2_EMBEDDED_RADIX 9 147 148 #define HAMMER2_PBUFMASK (HAMMER2_PBUFSIZE - 1) 149 #define HAMMER2_LBUFMASK (HAMMER2_LBUFSIZE - 1) 150 #define HAMMER2_SEGMASK (HAMMER2_SEGSIZE - 1) 151 152 #define HAMMER2_LBUFMASK64 ((hammer2_off_t)HAMMER2_LBUFMASK) 153 #define HAMMER2_PBUFSIZE64 ((hammer2_off_t)HAMMER2_PBUFSIZE) 154 #define HAMMER2_PBUFMASK64 ((hammer2_off_t)HAMMER2_PBUFMASK) 155 #define HAMMER2_SEGSIZE64 ((hammer2_off_t)HAMMER2_SEGSIZE) 156 #define HAMMER2_SEGMASK64 ((hammer2_off_t)HAMMER2_SEGMASK) 157 158 #define HAMMER2_UUID_STRING "5cbb9ad1-862d-11dc-a94d-01301bb8a9f5" 159 160 /* 161 * A HAMMER2 filesystem is always sized in multiples of 8MB. 162 * 163 * A 4MB segment is reserved at the beginning of each 2GB zone. This segment 164 * contains the volume header (or backup volume header), the free block 165 * table, and possibly other information in the future. 166 * 167 * 4MB = 64 x 64K blocks. Each 4MB segment is broken down as follows: 168 * 169 * +-----------------------+ 170 * | Volume Hdr | block 0 volume header & alternates 171 * +-----------------------+ (first four zones only) 172 * | FreeBlk Section A | block 1-4 173 * +-----------------------+ 174 * | FreeBlk Section B | block 5-8 175 * +-----------------------+ 176 * | FreeBlk Section C | block 9-12 177 * +-----------------------+ 178 * | FreeBlk Section D | block 13-16 179 * +-----------------------+ 180 * | | block 17...63 181 * | reserved | 182 * | | 183 * +-----------------------+ 184 * 185 * The first few 2GB zones contain volume headers and volume header backups. 186 * After that the volume header block# is reserved. 187 * 188 * Freemap (see the FREEMAP document) 189 * 190 * The freemap utilizes blocks #1-16 for now, see the FREEMAP document. 191 * The filesystems rotations through the sections to avoid disturbing the 192 * 'previous' version of the freemap during a flush. 193 * 194 * Each freemap section is 4 x 64K blocks and represents 2GB, 2TB, 2PB, 195 * and 2EB indirect map, plus the volume header has a set of 8 blockrefs 196 * for another 3 bits for a total of 64 bits of address space. The Level 0 197 * 64KB block representing 2GB of storage is a hammer2_bmap_data[1024]. 198 * Each element contains a 128x2 bit bitmap representing 16KB per chunk for 199 * 2MB of storage (x1024 elements = 2GB). 2 bits per chunk: 200 * 201 * 00 Free 202 * 01 Possibly fragmented 203 * 10 Possibly free 204 * 11 Allocated 205 * 206 * One important thing to note here is that the freemap resolution is 16KB, 207 * but the minimuim storage allocation size is 1KB. The hammer2 vfs keeps 208 * track of sub-allocations in memory (on umount or reboot obvious the whole 209 * 16KB will be considered allocated even if only 1KB is allocated). It is 210 * possible for fragmentation to build up over time. 211 * 212 * The Second thing to note is that due to the way snapshots and inode 213 * replication works, deleting a file cannot immediately free the related 214 * space. Instead, the freemap elements transition from 11->10. The bulk 215 * freeing code which does a complete scan is then responsible for 216 * transitioning the elements to 00 or back to 11 or to 01 for that matter. 217 * 218 * WARNING! ZONE_SEG and VOLUME_ALIGN must be a multiple of 1<<LEVEL0_RADIX 219 * (i.e. a multiple of 2MB). VOLUME_ALIGN must be >= ZONE_SEG. 220 */ 221 #define HAMMER2_VOLUME_ALIGN (8 * 1024 * 1024) 222 #define HAMMER2_VOLUME_ALIGN64 ((hammer2_off_t)HAMMER2_VOLUME_ALIGN) 223 #define HAMMER2_VOLUME_ALIGNMASK (HAMMER2_VOLUME_ALIGN - 1) 224 #define HAMMER2_VOLUME_ALIGNMASK64 ((hammer2_off_t)HAMMER2_VOLUME_ALIGNMASK) 225 226 #define HAMMER2_NEWFS_ALIGN (HAMMER2_VOLUME_ALIGN) 227 #define HAMMER2_NEWFS_ALIGN64 ((hammer2_off_t)HAMMER2_VOLUME_ALIGN) 228 #define HAMMER2_NEWFS_ALIGNMASK (HAMMER2_VOLUME_ALIGN - 1) 229 #define HAMMER2_NEWFS_ALIGNMASK64 ((hammer2_off_t)HAMMER2_NEWFS_ALIGNMASK) 230 231 #define HAMMER2_ZONE_BYTES64 (2LLU * 1024 * 1024 * 1024) 232 #define HAMMER2_ZONE_MASK64 (HAMMER2_ZONE_BYTES64 - 1) 233 #define HAMMER2_ZONE_SEG (4 * 1024 * 1024) 234 #define HAMMER2_ZONE_SEG64 ((hammer2_off_t)HAMMER2_ZONE_SEG) 235 #define HAMMER2_ZONE_BLOCKS_SEG (HAMMER2_ZONE_SEG / HAMMER2_PBUFSIZE) 236 237 /* 238 * 64 x 64KB blocks are reserved at the base of each 2GB zone. These blocks 239 * are used to store the volume header or volume header backups, allocation 240 * tree, and other information in the future. 241 * 242 * All specified blocks are not necessarily used in all 2GB zones. However, 243 * dead areas are reserved for future use and MUST NOT BE USED for other 244 * purposes. 245 * 246 * The freemap is arranged into four groups. Modifications rotate through 247 * the groups on a block by block basis (so all the blocks are not necessarily 248 * synchronized to the same group). Because the freemap is flushed 249 * independent of the main filesystem, the freemap only really needs two 250 * groups to operate efficiently. 251 * 252 * 253 * 254 */ 255 #define HAMMER2_ZONE_VOLHDR 0 /* volume header or backup */ 256 #define HAMMER2_ZONE_FREEMAP_A 1 /* freemap layer group A */ 257 #define HAMMER2_ZONE_FREEMAP_B 5 /* freemap layer group B */ 258 #define HAMMER2_ZONE_FREEMAP_C 9 /* freemap layer group C */ 259 #define HAMMER2_ZONE_FREEMAP_D 13 /* freemap layer group D */ 260 261 /* relative to FREEMAP_x */ 262 #define HAMMER2_ZONEFM_LEVEL1 0 /* 2GB leafmap */ 263 #define HAMMER2_ZONEFM_LEVEL2 1 /* 2TB indmap */ 264 #define HAMMER2_ZONEFM_LEVEL3 2 /* 2PB indmap */ 265 #define HAMMER2_ZONEFM_LEVEL4 3 /* 2EB indmap */ 266 /* LEVEL5 is a set of 8 blockrefs in the volume header 16EB */ 267 268 269 /* 270 * Freemap radii. Please note that LEVEL 1 blockref array entries 271 * point to 256-byte sections of the bitmap representing 2MB of storage. 272 * Even though the chain structures represent only 256 bytes, they are 273 * mapped using larger 16K or 64K buffer cache buffers. 274 */ 275 #define HAMMER2_FREEMAP_LEVEL5_RADIX 64 /* 16EB */ 276 #define HAMMER2_FREEMAP_LEVEL4_RADIX 61 /* 2EB */ 277 #define HAMMER2_FREEMAP_LEVEL3_RADIX 51 /* 2PB */ 278 #define HAMMER2_FREEMAP_LEVEL2_RADIX 41 /* 2TB */ 279 #define HAMMER2_FREEMAP_LEVEL1_RADIX 31 /* 2GB */ 280 #define HAMMER2_FREEMAP_LEVEL0_RADIX 21 /* 2MB (entry in l-1 leaf) */ 281 282 #define HAMMER2_FREEMAP_LEVELN_PSIZE 65536 /* physical bytes */ 283 284 #define HAMMER2_FREEMAP_COUNT (int)(HAMMER2_FREEMAP_LEVELN_PSIZE / \ 285 sizeof(hammer2_bmap_data_t)) 286 #define HAMMER2_FREEMAP_BLOCK_RADIX 14 287 #define HAMMER2_FREEMAP_BLOCK_SIZE (1 << HAMMER2_FREEMAP_BLOCK_RADIX) 288 #define HAMMER2_FREEMAP_BLOCK_MASK (HAMMER2_FREEMAP_BLOCK_SIZE - 1) 289 290 /* 291 * Two linear areas can be reserved after the initial 2MB segment in the base 292 * zone (the one starting at offset 0). These areas are NOT managed by the 293 * block allocator and do not fall under HAMMER2 crc checking rules based 294 * at the volume header (but can be self-CRCd internally, depending). 295 */ 296 #define HAMMER2_BOOT_MIN_BYTES HAMMER2_VOLUME_ALIGN 297 #define HAMMER2_BOOT_NOM_BYTES (64*1024*1024) 298 #define HAMMER2_BOOT_MAX_BYTES (256*1024*1024) 299 300 #define HAMMER2_REDO_MIN_BYTES HAMMER2_VOLUME_ALIGN 301 #define HAMMER2_REDO_NOM_BYTES (256*1024*1024) 302 #define HAMMER2_REDO_MAX_BYTES (1024*1024*1024) 303 304 /* 305 * Most HAMMER2 types are implemented as unsigned 64-bit integers. 306 * Transaction ids are monotonic. 307 * 308 * We utilize 32-bit iSCSI CRCs. 309 */ 310 typedef uint64_t hammer2_tid_t; 311 typedef uint64_t hammer2_off_t; 312 typedef uint64_t hammer2_key_t; 313 typedef uint32_t hammer2_crc32_t; 314 315 /* 316 * Miscellanious ranges (all are unsigned). 317 */ 318 #define HAMMER2_MIN_TID 1ULL 319 #define HAMMER2_MAX_TID 0xFFFFFFFFFFFFFFFFULL 320 #define HAMMER2_MIN_KEY 0ULL 321 #define HAMMER2_MAX_KEY 0xFFFFFFFFFFFFFFFFULL 322 #define HAMMER2_MIN_OFFSET 0ULL 323 #define HAMMER2_MAX_OFFSET 0xFFFFFFFFFFFFFFFFULL 324 325 /* 326 * HAMMER2 data offset special cases and masking. 327 * 328 * All HAMMER2 data offsets have to be broken down into a 64K buffer base 329 * offset (HAMMER2_OFF_MASK_HI) and a 64K buffer index (HAMMER2_OFF_MASK_LO). 330 * 331 * Indexes into physical buffers are always 64-byte aligned. The low 6 bits 332 * of the data offset field specifies how large the data chunk being pointed 333 * to as a power of 2. The theoretical minimum radix is thus 6 (The space 334 * needed in the low bits of the data offset field). However, the practical 335 * minimum allocation chunk size is 1KB (a radix of 10), so HAMMER2 sets 336 * HAMMER2_MIN_RADIX to 10. The maximum radix is currently 16 (64KB), but 337 * we fully intend to support larger extents in the future. 338 */ 339 #define HAMMER2_OFF_BAD ((hammer2_off_t)-1) 340 #define HAMMER2_OFF_MASK 0xFFFFFFFFFFFFFFC0ULL 341 #define HAMMER2_OFF_MASK_LO (HAMMER2_OFF_MASK & HAMMER2_PBUFMASK64) 342 #define HAMMER2_OFF_MASK_HI (~HAMMER2_PBUFMASK64) 343 #define HAMMER2_OFF_MASK_RADIX 0x000000000000003FULL 344 #define HAMMER2_MAX_COPIES 6 345 346 /* 347 * HAMMER2 directory support and pre-defined keys 348 */ 349 #define HAMMER2_DIRHASH_VISIBLE 0x8000000000000000ULL 350 #define HAMMER2_DIRHASH_USERMSK 0x7FFFFFFFFFFFFFFFULL 351 #define HAMMER2_DIRHASH_LOMASK 0x0000000000007FFFULL 352 #define HAMMER2_DIRHASH_HIMASK 0xFFFFFFFFFFFF0000ULL 353 #define HAMMER2_DIRHASH_FORCED 0x0000000000008000ULL /* bit forced on */ 354 355 #define HAMMER2_SROOT_KEY 0x0000000000000000ULL /* volume to sroot */ 356 357 /* 358 * The media block reference structure. This forms the core of the HAMMER2 359 * media topology recursion. This 64-byte data structure is embedded in the 360 * volume header, in inodes (which are also directory entries), and in 361 * indirect blocks. 362 * 363 * A blockref references a single media item, which typically can be a 364 * directory entry (aka inode), indirect block, or data block. 365 * 366 * The primary feature a blockref represents is the ability to validate 367 * the entire tree underneath it via its check code. Any modification to 368 * anything propagates up the blockref tree all the way to the root, replacing 369 * the related blocks. Propagations can shortcut to the volume root to 370 * implement the 'fast syncing' feature but this only delays the eventual 371 * propagation. 372 * 373 * The check code can be a simple 32-bit iscsi code, a 64-bit crc, 374 * or as complex as a 192 bit cryptographic hash. 192 bits is the maximum 375 * supported check code size, which is not sufficient for unverified dedup 376 * UNLESS one doesn't mind once-in-a-blue-moon data corruption (such as when 377 * farming web data). HAMMER2 has an unverified dedup feature for just this 378 * purpose. 379 * 380 * -- 381 * 382 * NOTE: The range of keys represented by the blockref is (key) to 383 * ((key) + (1LL << keybits) - 1). HAMMER2 usually populates 384 * blocks bottom-up, inserting a new root when radix expansion 385 * is required. 386 */ 387 struct hammer2_blockref { /* MUST BE EXACTLY 64 BYTES */ 388 uint8_t type; /* type of underlying item */ 389 uint8_t methods; /* check method & compression method */ 390 uint8_t copyid; /* specify which copy this is */ 391 uint8_t keybits; /* #of keybits masked off 0=leaf */ 392 uint8_t vradix; /* virtual data/meta-data size */ 393 uint8_t flags; /* blockref flags */ 394 uint8_t reserved06; 395 uint8_t reserved07; 396 hammer2_key_t key; /* key specification */ 397 hammer2_tid_t mirror_tid; /* propagate for mirror scan */ 398 hammer2_tid_t modify_tid; /* modifications sans propagation */ 399 hammer2_off_t data_off; /* low 6 bits is phys size (radix)*/ 400 union { /* check info */ 401 char buf[24]; 402 struct { 403 uint32_t value; 404 uint32_t unused[5]; 405 } iscsi32; 406 struct { 407 uint64_t value; 408 uint64_t unused[2]; 409 } crc64; 410 struct { 411 char data[24]; 412 } sha192; 413 414 /* 415 * Freemap hints are embedded in addition to the icrc32. 416 * 417 * bigmask - Radixes available for allocation (0-31). 418 * Heuristical (may be permissive but not 419 * restrictive). Typically only radix values 420 * 10-16 are used (i.e. (1<<10) through (1<<16)). 421 * 422 * avail - Total available space remaining, in bytes 423 */ 424 struct { 425 uint32_t icrc32; 426 uint32_t bigmask; /* available radixes */ 427 uint64_t avail; /* total available bytes */ 428 uint64_t unused; /* unused must be 0 */ 429 } freemap; 430 } check; 431 }; 432 433 typedef struct hammer2_blockref hammer2_blockref_t; 434 435 #if 0 436 #define HAMMER2_BREF_SYNC1 0x01 /* modification synchronized */ 437 #define HAMMER2_BREF_SYNC2 0x02 /* modification committed */ 438 #define HAMMER2_BREF_DESYNCCHLD 0x04 /* desynchronize children */ 439 #define HAMMER2_BREF_DELETED 0x80 /* indicates a deletion */ 440 #endif 441 442 #define HAMMER2_BLOCKREF_BYTES 64 /* blockref struct in bytes */ 443 444 /* 445 * On-media and off-media blockref types. 446 */ 447 #define HAMMER2_BREF_TYPE_EMPTY 0 448 #define HAMMER2_BREF_TYPE_INODE 1 449 #define HAMMER2_BREF_TYPE_INDIRECT 2 450 #define HAMMER2_BREF_TYPE_DATA 3 451 #define HAMMER2_BREF_TYPE_UNUSED04 4 452 #define HAMMER2_BREF_TYPE_FREEMAP_NODE 5 453 #define HAMMER2_BREF_TYPE_FREEMAP_LEAF 6 454 #define HAMMER2_BREF_TYPE_FREEMAP 254 /* pseudo-type */ 455 #define HAMMER2_BREF_TYPE_VOLUME 255 /* pseudo-type */ 456 457 #define HAMMER2_ENC_CHECK(n) ((n) << 4) 458 #define HAMMER2_DEC_CHECK(n) (((n) >> 4) & 15) 459 460 #define HAMMER2_CHECK_NONE 0 461 #define HAMMER2_CHECK_ISCSI32 1 462 #define HAMMER2_CHECK_CRC64 2 463 #define HAMMER2_CHECK_SHA192 3 464 #define HAMMER2_CHECK_FREEMAP 4 465 466 #define HAMMER2_ENC_COMP(n) (n) 467 #define HAMMER2_DEC_COMP(n) ((n) & 15) 468 469 #define HAMMER2_COMP_NONE 0 470 #define HAMMER2_COMP_AUTOZERO 1 471 472 473 /* 474 * HAMMER2 block references are collected into sets of 8 blockrefs. These 475 * sets are fully associative, meaning the elements making up a set are 476 * not sorted in any way and may contain duplicate entries, holes, or 477 * entries which shortcut multiple levels of indirection. Sets are used 478 * in various ways: 479 * 480 * (1) When redundancy is desired a set may contain several duplicate 481 * entries pointing to different copies of the same data. Up to 8 copies 482 * are supported but the set structure becomes a bit inefficient once 483 * you go over 4. 484 * 485 * (2) The blockrefs in a set can shortcut multiple levels of indirections 486 * within the bounds imposed by the parent of set. 487 * 488 * When a set fills up another level of indirection is inserted, moving 489 * some or all of the set's contents into indirect blocks placed under the 490 * set. This is a top-down approach in that indirect blocks are not created 491 * until the set actually becomes full (that is, the entries in the set can 492 * shortcut the indirect blocks when the set is not full). Depending on how 493 * things are filled multiple indirect blocks will eventually be created. 494 * 495 * Indirect blocks are typically 4KB (64 entres) or 64KB (1024 entries) and 496 * are also treated as fully set-associative. 497 */ 498 struct hammer2_blockset { 499 hammer2_blockref_t blockref[HAMMER2_SET_COUNT]; 500 }; 501 502 typedef struct hammer2_blockset hammer2_blockset_t; 503 504 /* 505 * Catch programmer snafus 506 */ 507 #if (1 << HAMMER2_SET_RADIX) != HAMMER2_SET_COUNT 508 #error "hammer2 direct radix is incorrect" 509 #endif 510 #if (1 << HAMMER2_PBUFRADIX) != HAMMER2_PBUFSIZE 511 #error "HAMMER2_PBUFRADIX and HAMMER2_PBUFSIZE are inconsistent" 512 #endif 513 #if (1 << HAMMER2_MIN_RADIX) != HAMMER2_MIN_ALLOC 514 #error "HAMMER2_MIN_RADIX and HAMMER2_MIN_ALLOC are inconsistent" 515 #endif 516 517 /* 518 * hammer2_bmap_data - A freemap entry in the LEVEL1 block. 519 * 520 * Each 64-byte entry contains the bitmap and meta-data required to manage 521 * a LEVEL0 (2MB) block of storage. The storage is managed in 128 x 16KB 522 * chunks. Smaller allocation granularity is supported via a linear iterator 523 * and/or must otherwise be tracked in ram. 524 * 525 * (data structure must be 64 bytes exactly) 526 * 527 * linear - A BYTE linear allocation offset used for sub-16KB allocations 528 * only. May contain values between 0 and 2MB. Must be ignored 529 * if 16KB-aligned (i.e. force bitmap scan), otherwise may be 530 * used to sub-allocate within the 16KB block (which is already 531 * marked as allocated in the bitmap). 532 * 533 * Sub-allocations need only be 1KB-aligned and do not have to be 534 * size-aligned, and 16KB or larger allocations do not update this 535 * field, resulting in pretty good packing. 536 * 537 * Please note that file data granularity may be limited by 538 * other issues such as buffer cache direct-mapping and the 539 * desire to support sector sizes up to 16KB (so H2 only issues 540 * I/O's in multiples of 16KB anyway). 541 * 542 * class - Clustering class. Cleared to 0 only if the entire leaf becomes 543 * free. Used to cluster device buffers so all elements must have 544 * the same device block size, but may mix logical sizes. 545 * 546 * Typically integrated with the blockref type in the upper 8 bits 547 * to localize inodes and indrect blocks, improving bulk free scans 548 * and directory scans. 549 * 550 * bitmap - Two bits per 16KB allocation block arranged in arrays of 551 * 32-bit elements, 128x2 bits representing ~2MB worth of media 552 * storage. Bit patterns are as follows: 553 * 554 * 00 Unallocated 555 * 01 Armed for bulk free scan 556 * 10 Possibly free 557 * 11 Allocated 558 */ 559 struct hammer2_bmap_data { 560 int32_t linear; /* 00 linear sub-granular allocation offset */ 561 uint16_t class; /* 04-05 clustering class ((type<<8)|radix) */ 562 uint8_t reserved06; /* 06 */ 563 uint8_t reserved07; /* 07 */ 564 uint32_t reserved08; /* 08 */ 565 uint32_t reserved0C; /* 0C */ 566 uint32_t reserved10; /* 10 */ 567 uint32_t reserved14; /* 14 */ 568 uint32_t reserved18; /* 18 */ 569 uint32_t avail; /* 1C */ 570 uint32_t bitmap[8]; /* 20-3F 256 bits manages 2MB/16KB/2-bits */ 571 }; 572 573 typedef struct hammer2_bmap_data hammer2_bmap_data_t; 574 575 /* 576 * In HAMMER2 inodes ARE directory entries, with a special exception for 577 * hardlinks. The inode number is stored in the inode rather than being 578 * based on the location of the inode (since the location moves every time 579 * the inode or anything underneath the inode is modified). 580 * 581 * The inode is 1024 bytes, made up of 256 bytes of meta-data, 256 bytes 582 * for the filename, and 512 bytes worth of direct file data OR an embedded 583 * blockset. 584 * 585 * Directories represent one inode per blockref. Inodes are not laid out 586 * as a file but instead are represented by the related blockrefs. The 587 * blockrefs, in turn, are indexed by the 64-bit directory hash key. Remember 588 * that blocksets are fully associative, so a certain degree efficiency is 589 * achieved just from that. 590 * 591 * Up to 512 bytes of direct data can be embedded in an inode, and since 592 * inodes are essentially directory entries this also means that small data 593 * files end up simply being laid out linearly in the directory, resulting 594 * in fewer seeks and highly optimal access. 595 * 596 * The compression mode can be changed at any time in the inode and is 597 * recorded on a blockref-by-blockref basis. 598 * 599 * Hardlinks are supported via the inode map. Essentially the way a hardlink 600 * works is that all individual directory entries representing the same file 601 * are special cased and specify the same inode number. The actual file 602 * is placed in the nearest parent directory that is parent to all instances 603 * of the hardlink. If all hardlinks to a file are in the same directory 604 * the actual file will also be placed in that directory. This file uses 605 * the inode number as the directory entry key and is invisible to normal 606 * directory scans. Real directory entry keys are differentiated from the 607 * inode number key via bit 63. Access to the hardlink silently looks up 608 * the real file and forwards all operations to that file. Removal of the 609 * last hardlink also removes the real file. 610 * 611 * (attr_tid) is only updated when the inode's specific attributes or regular 612 * file size has changed, and affects path lookups and stat. (attr_tid) 613 * represents a special cache coherency lock under the inode. The inode 614 * blockref's modify_tid will always cover it. 615 * 616 * (dirent_tid) is only updated when an entry under a directory inode has 617 * been created, deleted, renamed, or had its attributes change, and affects 618 * directory lookups and scans. (dirent_tid) represents another special cache 619 * coherency lock under the inode. The inode blockref's modify_tid will 620 * always cover it. 621 */ 622 #define HAMMER2_INODE_BYTES 1024 /* (asserted by code) */ 623 #define HAMMER2_INODE_MAXNAME 256 /* maximum name in bytes */ 624 #define HAMMER2_INODE_VERSION_ONE 1 625 626 struct hammer2_inode_data { 627 uint16_t version; /* 0000 inode data version */ 628 uint16_t reserved02; /* 0002 */ 629 630 /* 631 * core inode attributes, inode type, misc flags 632 */ 633 uint32_t uflags; /* 0004 chflags */ 634 uint32_t rmajor; /* 0008 available for device nodes */ 635 uint32_t rminor; /* 000C available for device nodes */ 636 uint64_t ctime; /* 0010 inode change time */ 637 uint64_t mtime; /* 0018 modified time */ 638 uint64_t atime; /* 0020 access time (unsupported) */ 639 uint64_t btime; /* 0028 birth time */ 640 uuid_t uid; /* 0030 uid / degenerate unix uid */ 641 uuid_t gid; /* 0040 gid / degenerate unix gid */ 642 643 uint8_t type; /* 0050 object type */ 644 uint8_t op_flags; /* 0051 operational flags */ 645 uint16_t cap_flags; /* 0052 capability flags */ 646 uint32_t mode; /* 0054 unix modes (typ low 16 bits) */ 647 648 /* 649 * inode size, identification, localized recursive configuration 650 * for compression and backup copies. 651 */ 652 hammer2_tid_t inum; /* 0058 inode number */ 653 hammer2_off_t size; /* 0060 size of file */ 654 uint64_t nlinks; /* 0068 hard links (typ only dirs) */ 655 hammer2_tid_t iparent; /* 0070 parent inum (recovery only) */ 656 hammer2_key_t name_key; /* 0078 full filename key */ 657 uint16_t name_len; /* 0080 filename length */ 658 uint8_t ncopies; /* 0082 ncopies to local media */ 659 uint8_t comp_algo; /* 0083 compression request & algo */ 660 661 /* 662 * These fields are currently only applicable to PFSROOTs. 663 * 664 * NOTE: We can't use {volume_data->fsid, pfs_clid} to uniquely 665 * identify an instance of a PFS in the cluster because 666 * a mount may contain more than one copy of the PFS as 667 * a separate node. {pfs_clid, pfs_fsid} must be used for 668 * registration in the cluster. 669 */ 670 uint8_t target_type; /* 0084 hardlink target type */ 671 uint8_t reserved85; /* 0085 */ 672 uint8_t reserved86; /* 0086 */ 673 uint8_t pfs_type; /* 0087 (if PFSROOT) node type */ 674 uint64_t pfs_inum; /* 0088 (if PFSROOT) inum allocator */ 675 uuid_t pfs_clid; /* 0090 (if PFSROOT) cluster uuid */ 676 uuid_t pfs_fsid; /* 00A0 (if PFSROOT) unique uuid */ 677 678 /* 679 * Quotas and cumulative sub-tree counters. 680 */ 681 hammer2_key_t data_quota; /* 00B0 subtree quota in bytes */ 682 hammer2_key_t data_count; /* 00B8 subtree byte count */ 683 hammer2_key_t inode_quota; /* 00C0 subtree quota inode count */ 684 hammer2_key_t inode_count; /* 00C8 subtree inode count */ 685 hammer2_tid_t attr_tid; /* 00D0 attributes changed */ 686 hammer2_tid_t dirent_tid; /* 00D8 directory/attr changed */ 687 688 /* 689 * Tracks (possibly degenerate) free areas covering all sub-tree 690 * allocations under inode, not counting the inode itself. 691 * 0/0 indicates empty entry. fully set-associative. 692 */ 693 hammer2_off_t freezones[4]; /* 00E0/E8/F0/F8 base|radix */ 694 695 unsigned char filename[HAMMER2_INODE_MAXNAME]; 696 /* 0100-01FF (256 char, unterminated) */ 697 union { /* 0200-03FF (64x8 = 512 bytes) */ 698 struct hammer2_blockset blockset; 699 char data[HAMMER2_EMBEDDED_BYTES]; 700 } u; 701 }; 702 703 typedef struct hammer2_inode_data hammer2_inode_data_t; 704 705 #define HAMMER2_OPFLAG_DIRECTDATA 0x01 706 #define HAMMER2_OPFLAG_PFSROOT 0x02 707 #define HAMMER2_OPFLAG_COPYIDS 0x04 /* copyids override parent */ 708 709 #define HAMMER2_OBJTYPE_UNKNOWN 0 710 #define HAMMER2_OBJTYPE_DIRECTORY 1 711 #define HAMMER2_OBJTYPE_REGFILE 2 712 #define HAMMER2_OBJTYPE_FIFO 4 713 #define HAMMER2_OBJTYPE_CDEV 5 714 #define HAMMER2_OBJTYPE_BDEV 6 715 #define HAMMER2_OBJTYPE_SOFTLINK 7 716 #define HAMMER2_OBJTYPE_HARDLINK 8 /* dummy entry for hardlink */ 717 #define HAMMER2_OBJTYPE_SOCKET 9 718 #define HAMMER2_OBJTYPE_WHITEOUT 10 719 720 #define HAMMER2_COPYID_NONE 0 721 #define HAMMER2_COPYID_LOCAL ((uint8_t)-1) 722 723 /* 724 * PEER types identify connections and help cluster controller filter 725 * out unwanted SPANs. 726 */ 727 #define HAMMER2_PEER_NONE DMSG_PEER_NONE 728 #define HAMMER2_PEER_CLUSTER DMSG_PEER_CLUSTER 729 #define HAMMER2_PEER_BLOCK DMSG_PEER_BLOCK 730 #define HAMMER2_PEER_HAMMER2 DMSG_PEER_HAMMER2 731 732 #define HAMMER2_COPYID_COUNT DMSG_COPYID_COUNT 733 734 /* 735 * PFS types identify a PFS on media and in LNK_SPAN messages. 736 */ 737 #define HAMMER2_PFSTYPE_NONE DMSG_PFSTYPE_NONE 738 #define HAMMER2_PFSTYPE_ADMIN DMSG_PFSTYPE_ADMIN 739 #define HAMMER2_PFSTYPE_CLIENT DMSG_PFSTYPE_CLIENT 740 #define HAMMER2_PFSTYPE_CACHE DMSG_PFSTYPE_CACHE 741 #define HAMMER2_PFSTYPE_COPY DMSG_PFSTYPE_COPY 742 #define HAMMER2_PFSTYPE_SLAVE DMSG_PFSTYPE_SLAVE 743 #define HAMMER2_PFSTYPE_SOFT_SLAVE DMSG_PFSTYPE_SOFT_SLAVE 744 #define HAMMER2_PFSTYPE_SOFT_MASTER DMSG_PFSTYPE_SOFT_MASTER 745 #define HAMMER2_PFSTYPE_MASTER DMSG_PFSTYPE_MASTER 746 #define HAMMER2_PFSTYPE_SNAPSHOT DMSG_PFSTYPE_SNAPSHOT 747 #define HAMMER2_PFSTYPE_MAX DMSG_PFSTYPE_MAX 748 749 /* 750 * Allocation Table 751 * 752 */ 753 754 755 /* 756 * Flags (8 bits) - blockref, for freemap only 757 * 758 * Note that the minimum chunk size is 1KB so we could theoretically have 759 * 10 bits here, but we might have some future extension that allows a 760 * chunk size down to 256 bytes and if so we will need bits 8 and 9. 761 */ 762 #define HAMMER2_AVF_SELMASK 0x03 /* select group */ 763 #define HAMMER2_AVF_ALL_ALLOC 0x04 /* indicate all allocated */ 764 #define HAMMER2_AVF_ALL_FREE 0x08 /* indicate all free */ 765 #define HAMMER2_AVF_RESERVED10 0x10 766 #define HAMMER2_AVF_RESERVED20 0x20 767 #define HAMMER2_AVF_RESERVED40 0x40 768 #define HAMMER2_AVF_RESERVED80 0x80 769 #define HAMMER2_AVF_AVMASK32 ((uint32_t)0xFFFFFF00LU) 770 #define HAMMER2_AVF_AVMASK64 ((uint64_t)0xFFFFFFFFFFFFFF00LLU) 771 772 #define HAMMER2_AV_SELECT_A 0x00 773 #define HAMMER2_AV_SELECT_B 0x01 774 #define HAMMER2_AV_SELECT_C 0x02 775 #define HAMMER2_AV_SELECT_D 0x03 776 777 /* 778 * The volume header eats a 64K block. There is currently an issue where 779 * we want to try to fit all nominal filesystem updates in a 512-byte section 780 * but it may be a lost cause due to the need for a blockset. 781 * 782 * All information is stored in host byte order. The volume header's magic 783 * number may be checked to determine the byte order. If you wish to mount 784 * between machines w/ different endian modes you'll need filesystem code 785 * which acts on the media data consistently (either all one way or all the 786 * other). Our code currently does not do that. 787 * 788 * A read-write mount may have to recover missing allocations by doing an 789 * incremental mirror scan looking for modifications made after alloc_tid. 790 * If alloc_tid == last_tid then no recovery operation is needed. Recovery 791 * operations are usually very, very fast. 792 * 793 * Read-only mounts do not need to do any recovery, access to the filesystem 794 * topology is always consistent after a crash (is always consistent, period). 795 * However, there may be shortcutted blockref updates present from deep in 796 * the tree which are stored in the volumeh eader and must be tracked on 797 * the fly. 798 * 799 * NOTE: The copyinfo[] array contains the configuration for both the 800 * cluster connections and any local media copies. The volume 801 * header will be replicated for each local media copy. 802 * 803 * The mount command may specify multiple medias or just one and 804 * allow HAMMER2 to pick up the others when it checks the copyinfo[] 805 * array on mount. 806 * 807 * NOTE: root_blockref points to the super-root directory, not the root 808 * directory. The root directory will be a subdirectory under the 809 * super-root. 810 * 811 * The super-root directory contains all root directories and all 812 * snapshots (readonly or writable). It is possible to do a 813 * null-mount of the super-root using special path constructions 814 * relative to your mounted root. 815 * 816 * NOTE: HAMMER2 allows any subdirectory tree to be managed as if it were 817 * a PFS, including mirroring and storage quota operations, and this is 818 * prefered over creating discrete PFSs in the super-root. Instead 819 * the super-root is most typically used to create writable snapshots, 820 * alternative roots, and so forth. The super-root is also used by 821 * the automatic snapshotting mechanism. 822 */ 823 #define HAMMER2_VOLUME_ID_HBO 0x48414d3205172011LLU 824 #define HAMMER2_VOLUME_ID_ABO 0x11201705324d4148LLU 825 826 struct hammer2_volume_data { 827 /* 828 * sector #0 - 512 bytes 829 */ 830 uint64_t magic; /* 0000 Signature */ 831 hammer2_off_t boot_beg; /* 0008 Boot area (future) */ 832 hammer2_off_t boot_end; /* 0010 (size = end - beg) */ 833 hammer2_off_t aux_beg; /* 0018 Aux area (future) */ 834 hammer2_off_t aux_end; /* 0020 (size = end - beg) */ 835 hammer2_off_t volu_size; /* 0028 Volume size, bytes */ 836 837 uint32_t version; /* 0030 */ 838 uint32_t flags; /* 0034 */ 839 uint8_t copyid; /* 0038 copyid of phys vol */ 840 uint8_t freemap_version; /* 0039 freemap algorithm */ 841 uint8_t peer_type; /* 003A HAMMER2_PEER_xxx */ 842 uint8_t reserved003B; /* 003B */ 843 uint32_t reserved003C; /* 003C */ 844 845 uuid_t fsid; /* 0040 */ 846 uuid_t fstype; /* 0050 */ 847 848 /* 849 * allocator_size is precalculated at newfs time and does not include 850 * reserved blocks, boot, or redo areas. 851 * 852 * Initial non-reserved-area allocations do not use the freemap 853 * but instead adjust alloc_iterator. Dynamic allocations take 854 * over starting at (allocator_beg). This makes newfs_hammer2's 855 * job a lot easier and can also serve as a testing jig. 856 */ 857 hammer2_off_t allocator_size; /* 0060 Total data space */ 858 hammer2_off_t allocator_free; /* 0068 Free space */ 859 hammer2_off_t allocator_beg; /* 0070 Initial allocations */ 860 hammer2_tid_t mirror_tid; /* 0078 best committed tid */ 861 hammer2_tid_t alloc_tid; /* 0080 Alloctable modify tid */ 862 hammer2_blockref_t reserved0088; /* 0088-00C7 */ 863 864 /* 865 * Copyids are allocated dynamically from the copyexists bitmap. 866 * An id from the active copies set (up to 8, see copyinfo later on) 867 * may still exist after the copy set has been removed from the 868 * volume header and its bit will remain active in the bitmap and 869 * cannot be reused until it is 100% removed from the hierarchy. 870 */ 871 uint32_t copyexists[8]; /* 00C8-00E7 copy exists bmap */ 872 char reserved0140[248]; /* 00E8-01DF */ 873 874 /* 875 * 32 bit CRC array at the end of the first 512 byte sector. 876 * 877 * icrc_sects[7] - First 512-4 bytes of volume header (including all 878 * the other icrc's except this one). 879 * 880 * icrc_sects[6] - Sector 1 (512 bytes) of volume header, which is 881 * the blockset for the root. 882 * 883 * icrc_sects[5] - Sector 2 884 * icrc_sects[4] - Sector 3 885 * icrc_sects[3] - Sector 4 (the freemap blockset) 886 */ 887 hammer2_crc32_t icrc_sects[8]; /* 01E0-01FF */ 888 889 /* 890 * sector #1 - 512 bytes 891 * 892 * The entire sector is used by a blockset. 893 */ 894 hammer2_blockset_t sroot_blockset; /* 0200-03FF Superroot dir */ 895 896 /* 897 * sector #2-7 898 */ 899 char sector2[512]; /* 0400-05FF reserved */ 900 char sector3[512]; /* 0600-07FF reserved */ 901 hammer2_blockset_t freemap_blockset; /* 0800-09FF freemap */ 902 char sector5[512]; /* 0A00-0BFF reserved */ 903 char sector6[512]; /* 0C00-0DFF reserved */ 904 char sector7[512]; /* 0E00-0FFF reserved */ 905 906 /* 907 * sector #8-71 - 32768 bytes 908 * 909 * Contains the configuration for up to 256 copyinfo targets. These 910 * specify local and remote copies operating as masters or slaves. 911 * copyid's 0 and 255 are reserved (0 indicates an empty slot and 255 912 * indicates the local media). 913 * 914 * Each inode contains a set of up to 8 copyids, either inherited 915 * from its parent or explicitly specified in the inode, which 916 * indexes into this array. 917 */ 918 /* 1000-8FFF copyinfo config */ 919 dmsg_vol_data_t copyinfo[HAMMER2_COPYID_COUNT]; 920 921 /* 922 * Remaining sections are reserved for future use. 923 */ 924 char reserved0400[0x6FFC]; /* 9000-FFFB reserved */ 925 926 /* 927 * icrc on entire volume header 928 */ 929 hammer2_crc32_t icrc_volheader; /* FFFC-FFFF full volume icrc*/ 930 }; 931 932 typedef struct hammer2_volume_data hammer2_volume_data_t; 933 934 /* 935 * Various parts of the volume header have their own iCRCs. 936 * 937 * The first 512 bytes has its own iCRC stored at the end of the 512 bytes 938 * and not included the icrc calculation. 939 * 940 * The second 512 bytes also has its own iCRC but it is stored in the first 941 * 512 bytes so it covers the entire second 512 bytes. 942 * 943 * The whole volume block (64KB) has an iCRC covering all but the last 4 bytes, 944 * which is where the iCRC for the whole volume is stored. This is currently 945 * a catch-all for anything not individually iCRCd. 946 */ 947 #define HAMMER2_VOL_ICRC_SECT0 7 948 #define HAMMER2_VOL_ICRC_SECT1 6 949 950 #define HAMMER2_VOLUME_BYTES 65536 951 952 #define HAMMER2_VOLUME_ICRC0_OFF 0 953 #define HAMMER2_VOLUME_ICRC1_OFF 512 954 #define HAMMER2_VOLUME_ICRCVH_OFF 0 955 956 #define HAMMER2_VOLUME_ICRC0_SIZE (512 - 4) 957 #define HAMMER2_VOLUME_ICRC1_SIZE (512) 958 #define HAMMER2_VOLUME_ICRCVH_SIZE (65536 - 4) 959 960 #define HAMMER2_VOL_VERSION_MIN 1 961 #define HAMMER2_VOL_VERSION_DEFAULT 1 962 #define HAMMER2_VOL_VERSION_WIP 2 963 964 #define HAMMER2_NUM_VOLHDRS 4 965 966 union hammer2_media_data { 967 hammer2_volume_data_t voldata; 968 hammer2_inode_data_t ipdata; 969 hammer2_blockref_t npdata[HAMMER2_IND_COUNT_MAX]; 970 hammer2_bmap_data_t bmdata[HAMMER2_FREEMAP_COUNT]; 971 char buf[HAMMER2_PBUFSIZE]; 972 }; 973 974 typedef union hammer2_media_data hammer2_media_data_t; 975 976 #endif 977