2 * Copyright (c) 2011-2012 The DragonFly Project. All rights reserved.
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>
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
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
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.
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
36 #ifndef _VFS_HAMMER2_DISK_H_
37 #define _VFS_HAMMER2_DISK_H_
47 * The structures below represent the on-disk media structures for the HAMMER2
48 * filesystem. Note that all fields for on-disk structures are naturally
49 * aligned. The host endian format is typically used - compatibility is
50 * possible if the implementation detects reversed endian and adjusts accesses
53 * HAMMER2 primarily revolves around the directory topology: inodes,
54 * directory entries, and block tables. Block device buffer cache buffers
55 * are always 64KB. Logical file buffers are typically 16KB. All data
56 * references utilize 64-bit byte offsets.
58 * Free block management is handled independently using blocks reserved by
63 * The data at the end of a file or directory may be a fragment in order
64 * to optimize storage efficiency. The minimum fragment size is 1KB.
65 * Since allocations are in powers of 2 fragments must also be sized in
66 * powers of 2 (1024, 2048, ... 65536).
68 * For the moment the maximum allocation size is HAMMER2_PBUFSIZE (64K),
69 * which is 2^16. Larger extents may be supported in the future. Smaller
70 * fragments might be supported in the future (down to 64 bytes is possible),
71 * but probably will not be.
73 * A full indirect block use supports 1024 x 64-byte blockrefs in a 64KB
74 * buffer. Indirect blocks down to 1KB are supported to keep small
77 * A maximally sized file (2^64-1 bytes) requires 5 indirect block levels.
78 * The hammer2_blockset in the volume header or file inode has another 8
79 * entries, giving us 66+3 = 69 bits of address space. However, some bits
80 * are taken up by (potentially) requests for redundant copies. HAMMER2
81 * currently supports up to 8 copies, which brings the address space down
82 * to 66 bits and gives us 2 bits of leeway.
84 #define HAMMER2_MIN_ALLOC 1024 /* minimum allocation size */
85 #define HAMMER2_MIN_RADIX 10 /* minimum allocation size 2^N */
86 #define HAMMER2_MAX_ALLOC 65536 /* maximum allocation size */
87 #define HAMMER2_MAX_RADIX 16 /* maximum allocation size 2^N */
88 #define HAMMER2_KEY_RADIX 64 /* number of bits in key */
91 * MINALLOCSIZE - The minimum allocation size. This can be smaller
92 * or larger than the minimum physical IO size.
94 * NOTE: Should not be larger than 1K since inodes
97 * MINIOSIZE - The minimum IO size. This must be less than
98 * or equal to HAMMER2_LBUFSIZE.
100 * HAMMER2_LBUFSIZE - Nominal buffer size for I/O rollups.
102 * HAMMER2_PBUFSIZE - Topological block size used by files for all
103 * blocks except the block straddling EOF.
105 * HAMMER2_SEGSIZE - Allocation map segment size, typically 2MB
106 * (space represented by a level0 bitmap).
109 #define HAMMER2_SEGSIZE (1 << HAMMER2_FREEMAP_LEVEL0_RADIX)
110 #define HAMMER2_SEGRADIX HAMMER2_FREEMAP_LEVEL0_RADIX
112 #define HAMMER2_PBUFRADIX 16 /* physical buf (1<<16) bytes */
113 #define HAMMER2_PBUFSIZE 65536
114 #define HAMMER2_LBUFRADIX 14 /* logical buf (1<<14) bytes */
115 #define HAMMER2_LBUFSIZE 16384
118 * Generally speaking we want to use 16K and 64K I/Os
120 #define HAMMER2_MINIORADIX HAMMER2_LBUFRADIX
121 #define HAMMER2_MINIOSIZE HAMMER2_LBUFSIZE
123 #define HAMMER2_IND_BYTES_MIN HAMMER2_LBUFSIZE
124 #define HAMMER2_IND_BYTES_MAX HAMMER2_PBUFSIZE
125 #define HAMMER2_IND_COUNT_MIN (HAMMER2_IND_BYTES_MIN / \
126 sizeof(hammer2_blockref_t))
127 #define HAMMER2_IND_COUNT_MAX (HAMMER2_IND_BYTES_MAX / \
128 sizeof(hammer2_blockref_t))
131 * In HAMMER2, arrays of blockrefs are fully set-associative, meaning that
132 * any element can occur at any index and holes can be anywhere. As a
133 * future optimization we will be able to flag that such arrays are sorted
134 * and thus optimize lookups, but for now we don't.
136 * Inodes embed either 512 bytes of direct data or an array of 8 blockrefs,
137 * resulting in highly efficient storage for files <= 512 bytes and for files
138 * <= 512KB. Up to 8 directory entries can be referenced from a directory
139 * without requiring an indirect block.
141 * Indirect blocks are typically either 4KB (64 blockrefs / ~4MB represented),
142 * or 64KB (1024 blockrefs / ~64MB represented).
144 #define HAMMER2_SET_COUNT 8 /* direct entries */
145 #define HAMMER2_SET_RADIX 3
146 #define HAMMER2_EMBEDDED_BYTES 512 /* inode blockset/dd size */
147 #define HAMMER2_EMBEDDED_RADIX 9
149 #define HAMMER2_PBUFMASK (HAMMER2_PBUFSIZE - 1)
150 #define HAMMER2_LBUFMASK (HAMMER2_LBUFSIZE - 1)
151 #define HAMMER2_SEGMASK (HAMMER2_SEGSIZE - 1)
153 #define HAMMER2_LBUFMASK64 ((hammer2_off_t)HAMMER2_LBUFMASK)
154 #define HAMMER2_PBUFSIZE64 ((hammer2_off_t)HAMMER2_PBUFSIZE)
155 #define HAMMER2_PBUFMASK64 ((hammer2_off_t)HAMMER2_PBUFMASK)
156 #define HAMMER2_SEGSIZE64 ((hammer2_off_t)HAMMER2_SEGSIZE)
157 #define HAMMER2_SEGMASK64 ((hammer2_off_t)HAMMER2_SEGMASK)
159 #define HAMMER2_UUID_STRING "5cbb9ad1-862d-11dc-a94d-01301bb8a9f5"
162 * A HAMMER2 filesystem is always sized in multiples of 8MB.
164 * A 4MB segment is reserved at the beginning of each 2GB zone. This segment
165 * contains the volume header (or backup volume header), the free block
166 * table, and possibly other information in the future.
168 * 4MB = 64 x 64K blocks. Each 4MB segment is broken down as follows:
170 * +-----------------------+
171 * | Volume Hdr | block 0 volume header & alternates
172 * +-----------------------+ (first four zones only)
173 * | FreeBlk Section A | block 1-4
174 * +-----------------------+
175 * | FreeBlk Section B | block 5-8
176 * +-----------------------+
177 * | FreeBlk Section C | block 9-12
178 * +-----------------------+
179 * | FreeBlk Section D | block 13-16
180 * +-----------------------+
184 * +-----------------------+
186 * The first few 2GB zones contain volume headers and volume header backups.
187 * After that the volume header block# is reserved.
189 * Freemap (see the FREEMAP document)
191 * The freemap utilizes blocks #1-16 for now, see the FREEMAP document.
192 * The filesystems rotations through the sections to avoid disturbing the
193 * 'previous' version of the freemap during a flush.
195 * Each freemap section is 4 x 64K blocks and represents 2GB, 2TB, 2PB,
196 * and 2EB indirect map, plus the volume header has a set of 8 blockrefs
197 * for another 3 bits for a total of 64 bits of address space. The Level 0
198 * 64KB block representing 2GB of storage is a hammer2_bmap_data[1024].
199 * Each element contains a 128x2 bit bitmap representing 16KB per chunk for
200 * 2MB of storage (x1024 elements = 2GB). 2 bits per chunk:
207 * One important thing to note here is that the freemap resolution is 16KB,
208 * but the minimuim storage allocation size is 1KB. The hammer2 vfs keeps
209 * track of sub-allocations in memory (on umount or reboot obvious the whole
210 * 16KB will be considered allocated even if only 1KB is allocated). It is
211 * possible for fragmentation to build up over time.
213 * The Second thing to note is that due to the way snapshots and inode
214 * replication works, deleting a file cannot immediately free the related
215 * space. Instead, the freemap elements transition from 11->10. The bulk
216 * freeing code which does a complete scan is then responsible for
217 * transitioning the elements to 00 or back to 11 or to 01 for that matter.
219 * WARNING! ZONE_SEG and VOLUME_ALIGN must be a multiple of 1<<LEVEL0_RADIX
220 * (i.e. a multiple of 2MB). VOLUME_ALIGN must be >= ZONE_SEG.
222 #define HAMMER2_VOLUME_ALIGN (8 * 1024 * 1024)
223 #define HAMMER2_VOLUME_ALIGN64 ((hammer2_off_t)HAMMER2_VOLUME_ALIGN)
224 #define HAMMER2_VOLUME_ALIGNMASK (HAMMER2_VOLUME_ALIGN - 1)
225 #define HAMMER2_VOLUME_ALIGNMASK64 ((hammer2_off_t)HAMMER2_VOLUME_ALIGNMASK)
227 #define HAMMER2_NEWFS_ALIGN (HAMMER2_VOLUME_ALIGN)
228 #define HAMMER2_NEWFS_ALIGN64 ((hammer2_off_t)HAMMER2_VOLUME_ALIGN)
229 #define HAMMER2_NEWFS_ALIGNMASK (HAMMER2_VOLUME_ALIGN - 1)
230 #define HAMMER2_NEWFS_ALIGNMASK64 ((hammer2_off_t)HAMMER2_NEWFS_ALIGNMASK)
232 #define HAMMER2_ZONE_BYTES64 (2LLU * 1024 * 1024 * 1024)
233 #define HAMMER2_ZONE_MASK64 (HAMMER2_ZONE_BYTES64 - 1)
234 #define HAMMER2_ZONE_SEG (4 * 1024 * 1024)
235 #define HAMMER2_ZONE_SEG64 ((hammer2_off_t)HAMMER2_ZONE_SEG)
236 #define HAMMER2_ZONE_BLOCKS_SEG (HAMMER2_ZONE_SEG / HAMMER2_PBUFSIZE)
239 * 64 x 64KB blocks are reserved at the base of each 2GB zone. These blocks
240 * are used to store the volume header or volume header backups, allocation
241 * tree, and other information in the future.
243 * All specified blocks are not necessarily used in all 2GB zones. However,
244 * dead areas are reserved for future use and MUST NOT BE USED for other
247 * The freemap is arranged into 15 groups of 4x64KB each. The 4 sub-groups
248 * are labeled ZONEFM1..4 and representing HAMMER2_FREEMAP_LEVEL{1-4}_RADIX,
249 * for the up to 4 levels of radix tree representing the freemap. For
250 * simplicity we are reserving all four radix tree layers even though the
251 * higher layers do not require teh reservation at each 2GB mark. That
252 * space is reserved for future use.
254 * Freemap blocks are not allocated dynamically but instead rotate through
255 * one of 15 possible copies. We require 15 copies for several reasons:
257 * (1) For distinguishing freemap 'allocations' made by the current flush
258 * verses the concurrently running front-end (at flush_tid + 1). This
259 * theoretically requires two copies but the algorithm is greatly
260 * simplified if we use three.
262 * (2) There are up to 4 copies of the volume header (iterated on each flush),
263 * and if the mount code is forced to use an older copy due to corruption
264 * we must be sure that the state of the freemap AS-OF the earlier copy
267 * This means 3 copies x 4 flushes = 12 copies to be able to mount any
268 * of the four volume header backups after on boot or after a crash.
270 * (3) Freemap recovery on-mount eats a copy. We don't want freemap recovery
271 * to blow away the copy used by some other volume header in case H2
272 * crashes during the recovery. Total is now 13.
274 * (4) And I want some breathing room to ensure that complex flushes do not
275 * cause problems. Also note that bulk block freeing itself must be
276 * careful so even on a live system, post-mount, the four volume header
277 * backups effectively represent short-lived snapshots. And I only
278 * have room for 15 copies so it works out.
280 * Preferably I would like to improve the algorithm to only use 2 copies per
281 * volume header (which would be a total of 2 x 4 = 8 + 1 for freemap recovery
282 * + 1 for breathing room = 10 total instead of 15). For now we use 15.
284 #define HAMMER2_ZONE_VOLHDR 0 /* volume header or backup */
285 #define HAMMER2_ZONE_FREEMAP_00 1
286 #define HAMMER2_ZONE_FREEMAP_01 5
287 #define HAMMER2_ZONE_FREEMAP_02 9
288 #define HAMMER2_ZONE_FREEMAP_03 13
289 #define HAMMER2_ZONE_FREEMAP_04 17
290 #define HAMMER2_ZONE_FREEMAP_05 21
291 #define HAMMER2_ZONE_FREEMAP_06 25
292 #define HAMMER2_ZONE_FREEMAP_07 29
293 #define HAMMER2_ZONE_FREEMAP_08 33
294 #define HAMMER2_ZONE_FREEMAP_09 37
295 #define HAMMER2_ZONE_FREEMAP_10 41
296 #define HAMMER2_ZONE_FREEMAP_11 45
297 #define HAMMER2_ZONE_FREEMAP_12 49
298 #define HAMMER2_ZONE_FREEMAP_13 53
299 #define HAMMER2_ZONE_FREEMAP_14 57
300 #define HAMMER2_ZONE_FREEMAP_END 61 /* (non-inclusive) */
301 #define HAMMER2_ZONE_UNUSED62 62
302 #define HAMMER2_ZONE_UNUSED63 63
304 #define HAMMER2_ZONE_FREEMAP_COPIES 15
305 /* relative to FREEMAP_x */
306 #define HAMMER2_ZONEFM_LEVEL1 0 /* 2GB leafmap */
307 #define HAMMER2_ZONEFM_LEVEL2 1 /* 2TB indmap */
308 #define HAMMER2_ZONEFM_LEVEL3 2 /* 2PB indmap */
309 #define HAMMER2_ZONEFM_LEVEL4 3 /* 2EB indmap */
310 /* LEVEL5 is a set of 8 blockrefs in the volume header 16EB */
314 * Freemap radii. Please note that LEVEL 1 blockref array entries
315 * point to 256-byte sections of the bitmap representing 2MB of storage.
316 * Even though the chain structures represent only 256 bytes, they are
317 * mapped using larger 16K or 64K buffer cache buffers.
319 #define HAMMER2_FREEMAP_LEVEL5_RADIX 64 /* 16EB */
320 #define HAMMER2_FREEMAP_LEVEL4_RADIX 61 /* 2EB */
321 #define HAMMER2_FREEMAP_LEVEL3_RADIX 51 /* 2PB */
322 #define HAMMER2_FREEMAP_LEVEL2_RADIX 41 /* 2TB */
323 #define HAMMER2_FREEMAP_LEVEL1_RADIX 31 /* 2GB */
324 #define HAMMER2_FREEMAP_LEVEL0_RADIX 21 /* 2MB (entry in l-1 leaf) */
326 #define HAMMER2_FREEMAP_LEVELN_PSIZE 65536 /* physical bytes */
328 #define HAMMER2_FREEMAP_COUNT (int)(HAMMER2_FREEMAP_LEVELN_PSIZE / \
329 sizeof(hammer2_bmap_data_t))
330 #define HAMMER2_FREEMAP_BLOCK_RADIX 14
331 #define HAMMER2_FREEMAP_BLOCK_SIZE (1 << HAMMER2_FREEMAP_BLOCK_RADIX)
332 #define HAMMER2_FREEMAP_BLOCK_MASK (HAMMER2_FREEMAP_BLOCK_SIZE - 1)
335 * Two linear areas can be reserved after the initial 2MB segment in the base
336 * zone (the one starting at offset 0). These areas are NOT managed by the
337 * block allocator and do not fall under HAMMER2 crc checking rules based
338 * at the volume header (but can be self-CRCd internally, depending).
340 #define HAMMER2_BOOT_MIN_BYTES HAMMER2_VOLUME_ALIGN
341 #define HAMMER2_BOOT_NOM_BYTES (64*1024*1024)
342 #define HAMMER2_BOOT_MAX_BYTES (256*1024*1024)
344 #define HAMMER2_REDO_MIN_BYTES HAMMER2_VOLUME_ALIGN
345 #define HAMMER2_REDO_NOM_BYTES (256*1024*1024)
346 #define HAMMER2_REDO_MAX_BYTES (1024*1024*1024)
349 * Most HAMMER2 types are implemented as unsigned 64-bit integers.
350 * Transaction ids are monotonic.
352 * We utilize 32-bit iSCSI CRCs.
354 typedef uint64_t hammer2_tid_t;
355 typedef uint64_t hammer2_off_t;
356 typedef uint64_t hammer2_key_t;
357 typedef uint32_t hammer2_crc32_t;
360 * Miscellanious ranges (all are unsigned).
362 #define HAMMER2_MIN_TID 1ULL
363 #define HAMMER2_MAX_TID 0xFFFFFFFFFFFFFFFFULL
364 #define HAMMER2_MIN_KEY 0ULL
365 #define HAMMER2_MAX_KEY 0xFFFFFFFFFFFFFFFFULL
366 #define HAMMER2_MIN_OFFSET 0ULL
367 #define HAMMER2_MAX_OFFSET 0xFFFFFFFFFFFFFFFFULL
370 * HAMMER2 data offset special cases and masking.
372 * All HAMMER2 data offsets have to be broken down into a 64K buffer base
373 * offset (HAMMER2_OFF_MASK_HI) and a 64K buffer index (HAMMER2_OFF_MASK_LO).
375 * Indexes into physical buffers are always 64-byte aligned. The low 6 bits
376 * of the data offset field specifies how large the data chunk being pointed
377 * to as a power of 2. The theoretical minimum radix is thus 6 (The space
378 * needed in the low bits of the data offset field). However, the practical
379 * minimum allocation chunk size is 1KB (a radix of 10), so HAMMER2 sets
380 * HAMMER2_MIN_RADIX to 10. The maximum radix is currently 16 (64KB), but
381 * we fully intend to support larger extents in the future.
383 #define HAMMER2_OFF_BAD ((hammer2_off_t)-1)
384 #define HAMMER2_OFF_MASK 0xFFFFFFFFFFFFFFC0ULL
385 #define HAMMER2_OFF_MASK_LO (HAMMER2_OFF_MASK & HAMMER2_PBUFMASK64)
386 #define HAMMER2_OFF_MASK_HI (~HAMMER2_PBUFMASK64)
387 #define HAMMER2_OFF_MASK_RADIX 0x000000000000003FULL
388 #define HAMMER2_MAX_COPIES 6
391 * HAMMER2 directory support and pre-defined keys
393 #define HAMMER2_DIRHASH_VISIBLE 0x8000000000000000ULL
394 #define HAMMER2_DIRHASH_USERMSK 0x7FFFFFFFFFFFFFFFULL
395 #define HAMMER2_DIRHASH_LOMASK 0x0000000000007FFFULL
396 #define HAMMER2_DIRHASH_HIMASK 0xFFFFFFFFFFFF0000ULL
397 #define HAMMER2_DIRHASH_FORCED 0x0000000000008000ULL /* bit forced on */
399 #define HAMMER2_SROOT_KEY 0x0000000000000000ULL /* volume to sroot */
402 * The media block reference structure. This forms the core of the HAMMER2
403 * media topology recursion. This 64-byte data structure is embedded in the
404 * volume header, in inodes (which are also directory entries), and in
407 * A blockref references a single media item, which typically can be a
408 * directory entry (aka inode), indirect block, or data block.
410 * The primary feature a blockref represents is the ability to validate
411 * the entire tree underneath it via its check code. Any modification to
412 * anything propagates up the blockref tree all the way to the root, replacing
413 * the related blocks. Propagations can shortcut to the volume root to
414 * implement the 'fast syncing' feature but this only delays the eventual
417 * The check code can be a simple 32-bit iscsi code, a 64-bit crc,
418 * or as complex as a 192 bit cryptographic hash. 192 bits is the maximum
419 * supported check code size, which is not sufficient for unverified dedup
420 * UNLESS one doesn't mind once-in-a-blue-moon data corruption (such as when
421 * farming web data). HAMMER2 has an unverified dedup feature for just this
426 * NOTE: The range of keys represented by the blockref is (key) to
427 * ((key) + (1LL << keybits) - 1). HAMMER2 usually populates
428 * blocks bottom-up, inserting a new root when radix expansion
431 struct hammer2_blockref { /* MUST BE EXACTLY 64 BYTES */
432 uint8_t type; /* type of underlying item */
433 uint8_t methods; /* check method & compression method */
434 uint8_t copyid; /* specify which copy this is */
435 uint8_t keybits; /* #of keybits masked off 0=leaf */
436 uint8_t vradix; /* virtual data/meta-data size */
437 uint8_t flags; /* blockref flags */
440 hammer2_key_t key; /* key specification */
441 hammer2_tid_t mirror_tid; /* propagate for mirror scan */
442 hammer2_tid_t modify_tid; /* modifications sans propagation */
443 hammer2_off_t data_off; /* low 6 bits is phys size (radix)*/
444 union { /* check info */
459 * Freemap hints are embedded in addition to the icrc32.
461 * bigmask - Radixes available for allocation (0-31).
462 * Heuristical (may be permissive but not
463 * restrictive). Typically only radix values
464 * 10-16 are used (i.e. (1<<10) through (1<<16)).
466 * avail - Total available space remaining, in bytes
470 uint32_t bigmask; /* available radixes */
471 uint64_t avail; /* total available bytes */
472 uint64_t unused; /* unused must be 0 */
479 hammer2_tid_t sync_tid;
484 typedef struct hammer2_blockref hammer2_blockref_t;
487 #define HAMMER2_BREF_SYNC1 0x01 /* modification synchronized */
488 #define HAMMER2_BREF_SYNC2 0x02 /* modification committed */
489 #define HAMMER2_BREF_DESYNCCHLD 0x04 /* desynchronize children */
490 #define HAMMER2_BREF_DELETED 0x80 /* indicates a deletion */
493 #define HAMMER2_BLOCKREF_BYTES 64 /* blockref struct in bytes */
496 * On-media and off-media blockref types.
498 #define HAMMER2_BREF_TYPE_EMPTY 0
499 #define HAMMER2_BREF_TYPE_INODE 1
500 #define HAMMER2_BREF_TYPE_INDIRECT 2
501 #define HAMMER2_BREF_TYPE_DATA 3
502 #define HAMMER2_BREF_TYPE_UNUSED04 4
503 #define HAMMER2_BREF_TYPE_FREEMAP_NODE 5
504 #define HAMMER2_BREF_TYPE_FREEMAP_LEAF 6
505 #define HAMMER2_BREF_TYPE_FREEMAP 254 /* pseudo-type */
506 #define HAMMER2_BREF_TYPE_VOLUME 255 /* pseudo-type */
508 #define HAMMER2_ENC_CHECK(n) ((n) << 4)
509 #define HAMMER2_DEC_CHECK(n) (((n) >> 4) & 15)
511 #define HAMMER2_CHECK_NONE 0
512 #define HAMMER2_CHECK_ISCSI32 1
513 #define HAMMER2_CHECK_CRC64 2
514 #define HAMMER2_CHECK_SHA192 3
515 #define HAMMER2_CHECK_FREEMAP 4
517 #define HAMMER2_ENC_COMP(n) (n)
518 #define HAMMER2_ENC_LEVEL(n) ((n) << 4)
519 #define HAMMER2_DEC_COMP(n) ((n) & 15)
520 #define HAMMER2_DEC_LEVEL(n) (((n) >> 4) & 15)
522 #define HAMMER2_COMP_NONE 0
523 #define HAMMER2_COMP_AUTOZERO 1
524 #define HAMMER2_COMP_LZ4 2
525 #define HAMMER2_COMP_ZLIB 3
527 #define HAMMER2_COMP_NEWFS_DEFAULT HAMMER2_COMP_LZ4
528 #define HAMMER2_COMP_STRINGS { "none", "autozero", "lz4", "zlib" }
529 #define HAMMER2_COMP_STRINGS_COUNT 4
533 * HAMMER2 block references are collected into sets of 8 blockrefs. These
534 * sets are fully associative, meaning the elements making up a set are
535 * not sorted in any way and may contain duplicate entries, holes, or
536 * entries which shortcut multiple levels of indirection. Sets are used
539 * (1) When redundancy is desired a set may contain several duplicate
540 * entries pointing to different copies of the same data. Up to 8 copies
541 * are supported but the set structure becomes a bit inefficient once
544 * (2) The blockrefs in a set can shortcut multiple levels of indirections
545 * within the bounds imposed by the parent of set.
547 * When a set fills up another level of indirection is inserted, moving
548 * some or all of the set's contents into indirect blocks placed under the
549 * set. This is a top-down approach in that indirect blocks are not created
550 * until the set actually becomes full (that is, the entries in the set can
551 * shortcut the indirect blocks when the set is not full). Depending on how
552 * things are filled multiple indirect blocks will eventually be created.
554 * Indirect blocks are typically 4KB (64 entres) or 64KB (1024 entries) and
555 * are also treated as fully set-associative.
557 struct hammer2_blockset {
558 hammer2_blockref_t blockref[HAMMER2_SET_COUNT];
561 typedef struct hammer2_blockset hammer2_blockset_t;
564 * Catch programmer snafus
566 #if (1 << HAMMER2_SET_RADIX) != HAMMER2_SET_COUNT
567 #error "hammer2 direct radix is incorrect"
569 #if (1 << HAMMER2_PBUFRADIX) != HAMMER2_PBUFSIZE
570 #error "HAMMER2_PBUFRADIX and HAMMER2_PBUFSIZE are inconsistent"
572 #if (1 << HAMMER2_MIN_RADIX) != HAMMER2_MIN_ALLOC
573 #error "HAMMER2_MIN_RADIX and HAMMER2_MIN_ALLOC are inconsistent"
577 * hammer2_bmap_data - A freemap entry in the LEVEL1 block.
579 * Each 64-byte entry contains the bitmap and meta-data required to manage
580 * a LEVEL0 (2MB) block of storage. The storage is managed in 128 x 16KB
581 * chunks. Smaller allocation granularity is supported via a linear iterator
582 * and/or must otherwise be tracked in ram.
584 * (data structure must be 64 bytes exactly)
586 * linear - A BYTE linear allocation offset used for sub-16KB allocations
587 * only. May contain values between 0 and 2MB. Must be ignored
588 * if 16KB-aligned (i.e. force bitmap scan), otherwise may be
589 * used to sub-allocate within the 16KB block (which is already
590 * marked as allocated in the bitmap).
592 * Sub-allocations need only be 1KB-aligned and do not have to be
593 * size-aligned, and 16KB or larger allocations do not update this
594 * field, resulting in pretty good packing.
596 * Please note that file data granularity may be limited by
597 * other issues such as buffer cache direct-mapping and the
598 * desire to support sector sizes up to 16KB (so H2 only issues
599 * I/O's in multiples of 16KB anyway).
601 * class - Clustering class. Cleared to 0 only if the entire leaf becomes
602 * free. Used to cluster device buffers so all elements must have
603 * the same device block size, but may mix logical sizes.
605 * Typically integrated with the blockref type in the upper 8 bits
606 * to localize inodes and indrect blocks, improving bulk free scans
607 * and directory scans.
609 * bitmap - Two bits per 16KB allocation block arranged in arrays of
610 * 32-bit elements, 128x2 bits representing ~2MB worth of media
611 * storage. Bit patterns are as follows:
618 struct hammer2_bmap_data {
619 int32_t linear; /* 00 linear sub-granular allocation offset */
620 uint16_t class; /* 04-05 clustering class ((type<<8)|radix) */
621 uint8_t reserved06; /* 06 */
622 uint8_t reserved07; /* 07 */
623 uint32_t reserved08; /* 08 */
624 uint32_t reserved0C; /* 0C */
625 uint32_t reserved10; /* 10 */
626 uint32_t reserved14; /* 14 */
627 uint32_t reserved18; /* 18 */
628 uint32_t avail; /* 1C */
629 uint32_t bitmap[8]; /* 20-3F 256 bits manages 2MB/16KB/2-bits */
632 typedef struct hammer2_bmap_data hammer2_bmap_data_t;
635 * In HAMMER2 inodes ARE directory entries, with a special exception for
636 * hardlinks. The inode number is stored in the inode rather than being
637 * based on the location of the inode (since the location moves every time
638 * the inode or anything underneath the inode is modified).
640 * The inode is 1024 bytes, made up of 256 bytes of meta-data, 256 bytes
641 * for the filename, and 512 bytes worth of direct file data OR an embedded
644 * Directories represent one inode per blockref. Inodes are not laid out
645 * as a file but instead are represented by the related blockrefs. The
646 * blockrefs, in turn, are indexed by the 64-bit directory hash key. Remember
647 * that blocksets are fully associative, so a certain degree efficiency is
648 * achieved just from that.
650 * Up to 512 bytes of direct data can be embedded in an inode, and since
651 * inodes are essentially directory entries this also means that small data
652 * files end up simply being laid out linearly in the directory, resulting
653 * in fewer seeks and highly optimal access.
655 * The compression mode can be changed at any time in the inode and is
656 * recorded on a blockref-by-blockref basis.
658 * Hardlinks are supported via the inode map. Essentially the way a hardlink
659 * works is that all individual directory entries representing the same file
660 * are special cased and specify the same inode number. The actual file
661 * is placed in the nearest parent directory that is parent to all instances
662 * of the hardlink. If all hardlinks to a file are in the same directory
663 * the actual file will also be placed in that directory. This file uses
664 * the inode number as the directory entry key and is invisible to normal
665 * directory scans. Real directory entry keys are differentiated from the
666 * inode number key via bit 63. Access to the hardlink silently looks up
667 * the real file and forwards all operations to that file. Removal of the
668 * last hardlink also removes the real file.
670 * (attr_tid) is only updated when the inode's specific attributes or regular
671 * file size has changed, and affects path lookups and stat. (attr_tid)
672 * represents a special cache coherency lock under the inode. The inode
673 * blockref's modify_tid will always cover it.
675 * (dirent_tid) is only updated when an entry under a directory inode has
676 * been created, deleted, renamed, or had its attributes change, and affects
677 * directory lookups and scans. (dirent_tid) represents another special cache
678 * coherency lock under the inode. The inode blockref's modify_tid will
681 #define HAMMER2_INODE_BYTES 1024 /* (asserted by code) */
682 #define HAMMER2_INODE_MAXNAME 256 /* maximum name in bytes */
683 #define HAMMER2_INODE_VERSION_ONE 1
685 #define HAMMER2_INODE_HIDDENDIR 16 /* special inode */
686 #define HAMMER2_INODE_START 1024 /* dynamically allocated */
688 struct hammer2_inode_data {
689 uint16_t version; /* 0000 inode data version */
690 uint16_t reserved02; /* 0002 */
693 * core inode attributes, inode type, misc flags
695 uint32_t uflags; /* 0004 chflags */
696 uint32_t rmajor; /* 0008 available for device nodes */
697 uint32_t rminor; /* 000C available for device nodes */
698 uint64_t ctime; /* 0010 inode change time */
699 uint64_t mtime; /* 0018 modified time */
700 uint64_t atime; /* 0020 access time (unsupported) */
701 uint64_t btime; /* 0028 birth time */
702 uuid_t uid; /* 0030 uid / degenerate unix uid */
703 uuid_t gid; /* 0040 gid / degenerate unix gid */
705 uint8_t type; /* 0050 object type */
706 uint8_t op_flags; /* 0051 operational flags */
707 uint16_t cap_flags; /* 0052 capability flags */
708 uint32_t mode; /* 0054 unix modes (typ low 16 bits) */
711 * inode size, identification, localized recursive configuration
712 * for compression and backup copies.
714 hammer2_tid_t inum; /* 0058 inode number */
715 hammer2_off_t size; /* 0060 size of file */
716 uint64_t nlinks; /* 0068 hard links (typ only dirs) */
717 hammer2_tid_t iparent; /* 0070 parent inum (recovery only) */
718 hammer2_key_t name_key; /* 0078 full filename key */
719 uint16_t name_len; /* 0080 filename length */
720 uint8_t ncopies; /* 0082 ncopies to local media */
721 uint8_t comp_algo; /* 0083 compression request & algo */
724 * These fields are currently only applicable to PFSROOTs.
726 * NOTE: We can't use {volume_data->fsid, pfs_clid} to uniquely
727 * identify an instance of a PFS in the cluster because
728 * a mount may contain more than one copy of the PFS as
729 * a separate node. {pfs_clid, pfs_fsid} must be used for
730 * registration in the cluster.
732 uint8_t target_type; /* 0084 hardlink target type */
733 uint8_t reserved85; /* 0085 */
734 uint8_t reserved86; /* 0086 */
735 uint8_t pfs_type; /* 0087 (if PFSROOT) node type */
736 uint64_t pfs_inum; /* 0088 (if PFSROOT) inum allocator */
737 uuid_t pfs_clid; /* 0090 (if PFSROOT) cluster uuid */
738 uuid_t pfs_fsid; /* 00A0 (if PFSROOT) unique uuid */
741 * Quotas and cumulative sub-tree counters.
743 hammer2_key_t data_quota; /* 00B0 subtree quota in bytes */
744 hammer2_key_t data_count; /* 00B8 subtree byte count */
745 hammer2_key_t inode_quota; /* 00C0 subtree quota inode count */
746 hammer2_key_t inode_count; /* 00C8 subtree inode count */
747 hammer2_tid_t attr_tid; /* 00D0 attributes changed */
748 hammer2_tid_t dirent_tid; /* 00D8 directory/attr changed */
751 * Tracks (possibly degenerate) free areas covering all sub-tree
752 * allocations under inode, not counting the inode itself.
753 * 0/0 indicates empty entry. fully set-associative.
755 hammer2_off_t freezones[4]; /* 00E0/E8/F0/F8 base|radix */
757 unsigned char filename[HAMMER2_INODE_MAXNAME];
758 /* 0100-01FF (256 char, unterminated) */
759 union { /* 0200-03FF (64x8 = 512 bytes) */
760 struct hammer2_blockset blockset;
761 char data[HAMMER2_EMBEDDED_BYTES];
765 typedef struct hammer2_inode_data hammer2_inode_data_t;
767 #define HAMMER2_OPFLAG_DIRECTDATA 0x01
768 #define HAMMER2_OPFLAG_PFSROOT 0x02
769 #define HAMMER2_OPFLAG_COPYIDS 0x04 /* copyids override parent */
771 #define HAMMER2_OBJTYPE_UNKNOWN 0
772 #define HAMMER2_OBJTYPE_DIRECTORY 1
773 #define HAMMER2_OBJTYPE_REGFILE 2
774 #define HAMMER2_OBJTYPE_FIFO 4
775 #define HAMMER2_OBJTYPE_CDEV 5
776 #define HAMMER2_OBJTYPE_BDEV 6
777 #define HAMMER2_OBJTYPE_SOFTLINK 7
778 #define HAMMER2_OBJTYPE_HARDLINK 8 /* dummy entry for hardlink */
779 #define HAMMER2_OBJTYPE_SOCKET 9
780 #define HAMMER2_OBJTYPE_WHITEOUT 10
782 #define HAMMER2_COPYID_NONE 0
783 #define HAMMER2_COPYID_LOCAL ((uint8_t)-1)
786 * PEER types identify connections and help cluster controller filter
787 * out unwanted SPANs.
789 #define HAMMER2_PEER_NONE DMSG_PEER_NONE
790 #define HAMMER2_PEER_CLUSTER DMSG_PEER_CLUSTER
791 #define HAMMER2_PEER_BLOCK DMSG_PEER_BLOCK
792 #define HAMMER2_PEER_HAMMER2 DMSG_PEER_HAMMER2
794 #define HAMMER2_COPYID_COUNT DMSG_COPYID_COUNT
797 * PFS types identify a PFS on media and in LNK_SPAN messages.
799 #define HAMMER2_PFSTYPE_NONE DMSG_PFSTYPE_NONE
800 #define HAMMER2_PFSTYPE_ADMIN DMSG_PFSTYPE_ADMIN
801 #define HAMMER2_PFSTYPE_CLIENT DMSG_PFSTYPE_CLIENT
802 #define HAMMER2_PFSTYPE_CACHE DMSG_PFSTYPE_CACHE
803 #define HAMMER2_PFSTYPE_COPY DMSG_PFSTYPE_COPY
804 #define HAMMER2_PFSTYPE_SLAVE DMSG_PFSTYPE_SLAVE
805 #define HAMMER2_PFSTYPE_SOFT_SLAVE DMSG_PFSTYPE_SOFT_SLAVE
806 #define HAMMER2_PFSTYPE_SOFT_MASTER DMSG_PFSTYPE_SOFT_MASTER
807 #define HAMMER2_PFSTYPE_MASTER DMSG_PFSTYPE_MASTER
808 #define HAMMER2_PFSTYPE_SNAPSHOT DMSG_PFSTYPE_SNAPSHOT
809 #define HAMMER2_PFSTYPE_MAX DMSG_PFSTYPE_MAX
818 * Flags (8 bits) - blockref, for freemap only
820 * Note that the minimum chunk size is 1KB so we could theoretically have
821 * 10 bits here, but we might have some future extension that allows a
822 * chunk size down to 256 bytes and if so we will need bits 8 and 9.
824 #define HAMMER2_AVF_SELMASK 0x03 /* select group */
825 #define HAMMER2_AVF_ALL_ALLOC 0x04 /* indicate all allocated */
826 #define HAMMER2_AVF_ALL_FREE 0x08 /* indicate all free */
827 #define HAMMER2_AVF_RESERVED10 0x10
828 #define HAMMER2_AVF_RESERVED20 0x20
829 #define HAMMER2_AVF_RESERVED40 0x40
830 #define HAMMER2_AVF_RESERVED80 0x80
831 #define HAMMER2_AVF_AVMASK32 ((uint32_t)0xFFFFFF00LU)
832 #define HAMMER2_AVF_AVMASK64 ((uint64_t)0xFFFFFFFFFFFFFF00LLU)
834 #define HAMMER2_AV_SELECT_A 0x00
835 #define HAMMER2_AV_SELECT_B 0x01
836 #define HAMMER2_AV_SELECT_C 0x02
837 #define HAMMER2_AV_SELECT_D 0x03
840 * The volume header eats a 64K block. There is currently an issue where
841 * we want to try to fit all nominal filesystem updates in a 512-byte section
842 * but it may be a lost cause due to the need for a blockset.
844 * All information is stored in host byte order. The volume header's magic
845 * number may be checked to determine the byte order. If you wish to mount
846 * between machines w/ different endian modes you'll need filesystem code
847 * which acts on the media data consistently (either all one way or all the
848 * other). Our code currently does not do that.
850 * A read-write mount may have to recover missing allocations by doing an
851 * incremental mirror scan looking for modifications made after alloc_tid.
852 * If alloc_tid == last_tid then no recovery operation is needed. Recovery
853 * operations are usually very, very fast.
855 * Read-only mounts do not need to do any recovery, access to the filesystem
856 * topology is always consistent after a crash (is always consistent, period).
857 * However, there may be shortcutted blockref updates present from deep in
858 * the tree which are stored in the volumeh eader and must be tracked on
861 * NOTE: The copyinfo[] array contains the configuration for both the
862 * cluster connections and any local media copies. The volume
863 * header will be replicated for each local media copy.
865 * The mount command may specify multiple medias or just one and
866 * allow HAMMER2 to pick up the others when it checks the copyinfo[]
869 * NOTE: root_blockref points to the super-root directory, not the root
870 * directory. The root directory will be a subdirectory under the
873 * The super-root directory contains all root directories and all
874 * snapshots (readonly or writable). It is possible to do a
875 * null-mount of the super-root using special path constructions
876 * relative to your mounted root.
878 * NOTE: HAMMER2 allows any subdirectory tree to be managed as if it were
879 * a PFS, including mirroring and storage quota operations, and this is
880 * prefered over creating discrete PFSs in the super-root. Instead
881 * the super-root is most typically used to create writable snapshots,
882 * alternative roots, and so forth. The super-root is also used by
883 * the automatic snapshotting mechanism.
885 #define HAMMER2_VOLUME_ID_HBO 0x48414d3205172011LLU
886 #define HAMMER2_VOLUME_ID_ABO 0x11201705324d4148LLU
888 struct hammer2_volume_data {
890 * sector #0 - 512 bytes
892 uint64_t magic; /* 0000 Signature */
893 hammer2_off_t boot_beg; /* 0008 Boot area (future) */
894 hammer2_off_t boot_end; /* 0010 (size = end - beg) */
895 hammer2_off_t aux_beg; /* 0018 Aux area (future) */
896 hammer2_off_t aux_end; /* 0020 (size = end - beg) */
897 hammer2_off_t volu_size; /* 0028 Volume size, bytes */
899 uint32_t version; /* 0030 */
900 uint32_t flags; /* 0034 */
901 uint8_t copyid; /* 0038 copyid of phys vol */
902 uint8_t freemap_version; /* 0039 freemap algorithm */
903 uint8_t peer_type; /* 003A HAMMER2_PEER_xxx */
904 uint8_t reserved003B; /* 003B */
905 uint32_t reserved003C; /* 003C */
907 uuid_t fsid; /* 0040 */
908 uuid_t fstype; /* 0050 */
911 * allocator_size is precalculated at newfs time and does not include
912 * reserved blocks, boot, or redo areas.
914 * Initial non-reserved-area allocations do not use the freemap
915 * but instead adjust alloc_iterator. Dynamic allocations take
916 * over starting at (allocator_beg). This makes newfs_hammer2's
917 * job a lot easier and can also serve as a testing jig.
919 hammer2_off_t allocator_size; /* 0060 Total data space */
920 hammer2_off_t allocator_free; /* 0068 Free space */
921 hammer2_off_t allocator_beg; /* 0070 Initial allocations */
922 hammer2_tid_t mirror_tid; /* 0078 committed tid (vol) */
923 hammer2_tid_t alloc_tid; /* 0080 Alloctable modify tid */
924 hammer2_tid_t inode_tid; /* 0088 Inode allocator tid */
925 hammer2_tid_t freemap_tid; /* 0090 committed tid (fmap) */
926 hammer2_tid_t bulkfree_tid; /* 0098 bulkfree incremental */
927 hammer2_tid_t reserved00A0[5]; /* 00A0-00C7 */
930 * Copyids are allocated dynamically from the copyexists bitmap.
931 * An id from the active copies set (up to 8, see copyinfo later on)
932 * may still exist after the copy set has been removed from the
933 * volume header and its bit will remain active in the bitmap and
934 * cannot be reused until it is 100% removed from the hierarchy.
936 uint32_t copyexists[8]; /* 00C8-00E7 copy exists bmap */
937 char reserved0140[248]; /* 00E8-01DF */
940 * 32 bit CRC array at the end of the first 512 byte sector.
942 * icrc_sects[7] - First 512-4 bytes of volume header (including all
943 * the other icrc's except this one).
945 * icrc_sects[6] - Sector 1 (512 bytes) of volume header, which is
946 * the blockset for the root.
948 * icrc_sects[5] - Sector 2
949 * icrc_sects[4] - Sector 3
950 * icrc_sects[3] - Sector 4 (the freemap blockset)
952 hammer2_crc32_t icrc_sects[8]; /* 01E0-01FF */
955 * sector #1 - 512 bytes
957 * The entire sector is used by a blockset.
959 hammer2_blockset_t sroot_blockset; /* 0200-03FF Superroot dir */
964 char sector2[512]; /* 0400-05FF reserved */
965 char sector3[512]; /* 0600-07FF reserved */
966 hammer2_blockset_t freemap_blockset; /* 0800-09FF freemap */
967 char sector5[512]; /* 0A00-0BFF reserved */
968 char sector6[512]; /* 0C00-0DFF reserved */
969 char sector7[512]; /* 0E00-0FFF reserved */
972 * sector #8-71 - 32768 bytes
974 * Contains the configuration for up to 256 copyinfo targets. These
975 * specify local and remote copies operating as masters or slaves.
976 * copyid's 0 and 255 are reserved (0 indicates an empty slot and 255
977 * indicates the local media).
979 * Each inode contains a set of up to 8 copyids, either inherited
980 * from its parent or explicitly specified in the inode, which
981 * indexes into this array.
983 /* 1000-8FFF copyinfo config */
984 dmsg_vol_data_t copyinfo[HAMMER2_COPYID_COUNT];
987 * Remaining sections are reserved for future use.
989 char reserved0400[0x6FFC]; /* 9000-FFFB reserved */
992 * icrc on entire volume header
994 hammer2_crc32_t icrc_volheader; /* FFFC-FFFF full volume icrc*/
997 typedef struct hammer2_volume_data hammer2_volume_data_t;
1000 * Various parts of the volume header have their own iCRCs.
1002 * The first 512 bytes has its own iCRC stored at the end of the 512 bytes
1003 * and not included the icrc calculation.
1005 * The second 512 bytes also has its own iCRC but it is stored in the first
1006 * 512 bytes so it covers the entire second 512 bytes.
1008 * The whole volume block (64KB) has an iCRC covering all but the last 4 bytes,
1009 * which is where the iCRC for the whole volume is stored. This is currently
1010 * a catch-all for anything not individually iCRCd.
1012 #define HAMMER2_VOL_ICRC_SECT0 7
1013 #define HAMMER2_VOL_ICRC_SECT1 6
1015 #define HAMMER2_VOLUME_BYTES 65536
1017 #define HAMMER2_VOLUME_ICRC0_OFF 0
1018 #define HAMMER2_VOLUME_ICRC1_OFF 512
1019 #define HAMMER2_VOLUME_ICRCVH_OFF 0
1021 #define HAMMER2_VOLUME_ICRC0_SIZE (512 - 4)
1022 #define HAMMER2_VOLUME_ICRC1_SIZE (512)
1023 #define HAMMER2_VOLUME_ICRCVH_SIZE (65536 - 4)
1025 #define HAMMER2_VOL_VERSION_MIN 1
1026 #define HAMMER2_VOL_VERSION_DEFAULT 1
1027 #define HAMMER2_VOL_VERSION_WIP 2
1029 #define HAMMER2_NUM_VOLHDRS 4
1031 union hammer2_media_data {
1032 hammer2_volume_data_t voldata;
1033 hammer2_inode_data_t ipdata;
1034 hammer2_blockref_t npdata[HAMMER2_IND_COUNT_MAX];
1035 hammer2_bmap_data_t bmdata[HAMMER2_FREEMAP_COUNT];
1036 char buf[HAMMER2_PBUFSIZE];
1039 typedef union hammer2_media_data hammer2_media_data_t;
1041 #endif /* !_VFS_HAMMER2_DISK_H_ */