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:
203 * 01 Possibly fragmented
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 four groups. Modifications rotate through
248 * the groups on a block by block basis (so all the blocks are not necessarily
249 * synchronized to the same group). Because the freemap is flushed
250 * independent of the main filesystem, the freemap only really needs two
251 * groups to operate efficiently.
256 #define HAMMER2_ZONE_VOLHDR 0 /* volume header or backup */
257 #define HAMMER2_ZONE_FREEMAP_A 1 /* freemap layer group A */
258 #define HAMMER2_ZONE_FREEMAP_B 5 /* freemap layer group B */
259 #define HAMMER2_ZONE_FREEMAP_C 9 /* freemap layer group C */
260 #define HAMMER2_ZONE_FREEMAP_D 13 /* freemap layer group D */
262 /* relative to FREEMAP_x */
263 #define HAMMER2_ZONEFM_LEVEL1 0 /* 2GB leafmap */
264 #define HAMMER2_ZONEFM_LEVEL2 1 /* 2TB indmap */
265 #define HAMMER2_ZONEFM_LEVEL3 2 /* 2PB indmap */
266 #define HAMMER2_ZONEFM_LEVEL4 3 /* 2EB indmap */
267 /* LEVEL5 is a set of 8 blockrefs in the volume header 16EB */
271 * Freemap radii. Please note that LEVEL 1 blockref array entries
272 * point to 256-byte sections of the bitmap representing 2MB of storage.
273 * Even though the chain structures represent only 256 bytes, they are
274 * mapped using larger 16K or 64K buffer cache buffers.
276 #define HAMMER2_FREEMAP_LEVEL5_RADIX 64 /* 16EB */
277 #define HAMMER2_FREEMAP_LEVEL4_RADIX 61 /* 2EB */
278 #define HAMMER2_FREEMAP_LEVEL3_RADIX 51 /* 2PB */
279 #define HAMMER2_FREEMAP_LEVEL2_RADIX 41 /* 2TB */
280 #define HAMMER2_FREEMAP_LEVEL1_RADIX 31 /* 2GB */
281 #define HAMMER2_FREEMAP_LEVEL0_RADIX 21 /* 2MB (entry in l-1 leaf) */
283 #define HAMMER2_FREEMAP_LEVELN_PSIZE 65536 /* physical bytes */
285 #define HAMMER2_FREEMAP_COUNT (int)(HAMMER2_FREEMAP_LEVELN_PSIZE / \
286 sizeof(hammer2_bmap_data_t))
287 #define HAMMER2_FREEMAP_BLOCK_RADIX 14
288 #define HAMMER2_FREEMAP_BLOCK_SIZE (1 << HAMMER2_FREEMAP_BLOCK_RADIX)
289 #define HAMMER2_FREEMAP_BLOCK_MASK (HAMMER2_FREEMAP_BLOCK_SIZE - 1)
292 * Two linear areas can be reserved after the initial 2MB segment in the base
293 * zone (the one starting at offset 0). These areas are NOT managed by the
294 * block allocator and do not fall under HAMMER2 crc checking rules based
295 * at the volume header (but can be self-CRCd internally, depending).
297 #define HAMMER2_BOOT_MIN_BYTES HAMMER2_VOLUME_ALIGN
298 #define HAMMER2_BOOT_NOM_BYTES (64*1024*1024)
299 #define HAMMER2_BOOT_MAX_BYTES (256*1024*1024)
301 #define HAMMER2_REDO_MIN_BYTES HAMMER2_VOLUME_ALIGN
302 #define HAMMER2_REDO_NOM_BYTES (256*1024*1024)
303 #define HAMMER2_REDO_MAX_BYTES (1024*1024*1024)
306 * Most HAMMER2 types are implemented as unsigned 64-bit integers.
307 * Transaction ids are monotonic.
309 * We utilize 32-bit iSCSI CRCs.
311 typedef uint64_t hammer2_tid_t;
312 typedef uint64_t hammer2_off_t;
313 typedef uint64_t hammer2_key_t;
314 typedef uint32_t hammer2_crc32_t;
317 * Miscellanious ranges (all are unsigned).
319 #define HAMMER2_MIN_TID 1ULL
320 #define HAMMER2_MAX_TID 0xFFFFFFFFFFFFFFFFULL
321 #define HAMMER2_MIN_KEY 0ULL
322 #define HAMMER2_MAX_KEY 0xFFFFFFFFFFFFFFFFULL
323 #define HAMMER2_MIN_OFFSET 0ULL
324 #define HAMMER2_MAX_OFFSET 0xFFFFFFFFFFFFFFFFULL
327 * HAMMER2 data offset special cases and masking.
329 * All HAMMER2 data offsets have to be broken down into a 64K buffer base
330 * offset (HAMMER2_OFF_MASK_HI) and a 64K buffer index (HAMMER2_OFF_MASK_LO).
332 * Indexes into physical buffers are always 64-byte aligned. The low 6 bits
333 * of the data offset field specifies how large the data chunk being pointed
334 * to as a power of 2. The theoretical minimum radix is thus 6 (The space
335 * needed in the low bits of the data offset field). However, the practical
336 * minimum allocation chunk size is 1KB (a radix of 10), so HAMMER2 sets
337 * HAMMER2_MIN_RADIX to 10. The maximum radix is currently 16 (64KB), but
338 * we fully intend to support larger extents in the future.
340 #define HAMMER2_OFF_BAD ((hammer2_off_t)-1)
341 #define HAMMER2_OFF_MASK 0xFFFFFFFFFFFFFFC0ULL
342 #define HAMMER2_OFF_MASK_LO (HAMMER2_OFF_MASK & HAMMER2_PBUFMASK64)
343 #define HAMMER2_OFF_MASK_HI (~HAMMER2_PBUFMASK64)
344 #define HAMMER2_OFF_MASK_RADIX 0x000000000000003FULL
345 #define HAMMER2_MAX_COPIES 6
348 * HAMMER2 directory support and pre-defined keys
350 #define HAMMER2_DIRHASH_VISIBLE 0x8000000000000000ULL
351 #define HAMMER2_DIRHASH_USERMSK 0x7FFFFFFFFFFFFFFFULL
352 #define HAMMER2_DIRHASH_LOMASK 0x0000000000007FFFULL
353 #define HAMMER2_DIRHASH_HIMASK 0xFFFFFFFFFFFF0000ULL
354 #define HAMMER2_DIRHASH_FORCED 0x0000000000008000ULL /* bit forced on */
356 #define HAMMER2_SROOT_KEY 0x0000000000000000ULL /* volume to sroot */
359 * The media block reference structure. This forms the core of the HAMMER2
360 * media topology recursion. This 64-byte data structure is embedded in the
361 * volume header, in inodes (which are also directory entries), and in
364 * A blockref references a single media item, which typically can be a
365 * directory entry (aka inode), indirect block, or data block.
367 * The primary feature a blockref represents is the ability to validate
368 * the entire tree underneath it via its check code. Any modification to
369 * anything propagates up the blockref tree all the way to the root, replacing
370 * the related blocks. Propagations can shortcut to the volume root to
371 * implement the 'fast syncing' feature but this only delays the eventual
374 * The check code can be a simple 32-bit iscsi code, a 64-bit crc,
375 * or as complex as a 192 bit cryptographic hash. 192 bits is the maximum
376 * supported check code size, which is not sufficient for unverified dedup
377 * UNLESS one doesn't mind once-in-a-blue-moon data corruption (such as when
378 * farming web data). HAMMER2 has an unverified dedup feature for just this
383 * NOTE: The range of keys represented by the blockref is (key) to
384 * ((key) + (1LL << keybits) - 1). HAMMER2 usually populates
385 * blocks bottom-up, inserting a new root when radix expansion
388 struct hammer2_blockref { /* MUST BE EXACTLY 64 BYTES */
389 uint8_t type; /* type of underlying item */
390 uint8_t methods; /* check method & compression method */
391 uint8_t copyid; /* specify which copy this is */
392 uint8_t keybits; /* #of keybits masked off 0=leaf */
393 uint8_t vradix; /* virtual data/meta-data size */
394 uint8_t flags; /* blockref flags */
397 hammer2_key_t key; /* key specification */
398 hammer2_tid_t mirror_tid; /* propagate for mirror scan */
399 hammer2_tid_t modify_tid; /* modifications sans propagation */
400 hammer2_off_t data_off; /* low 6 bits is phys size (radix)*/
401 union { /* check info */
416 * Freemap hints are embedded in addition to the icrc32.
418 * bigmask - Radixes available for allocation (0-31).
419 * Heuristical (may be permissive but not
420 * restrictive). Typically only radix values
421 * 10-16 are used (i.e. (1<<10) through (1<<16)).
423 * avail - Total available space remaining, in bytes
427 uint32_t bigmask; /* available radixes */
428 uint64_t avail; /* total available bytes */
429 uint64_t unused; /* unused must be 0 */
436 hammer2_tid_t sync_tid;
441 typedef struct hammer2_blockref hammer2_blockref_t;
444 #define HAMMER2_BREF_SYNC1 0x01 /* modification synchronized */
445 #define HAMMER2_BREF_SYNC2 0x02 /* modification committed */
446 #define HAMMER2_BREF_DESYNCCHLD 0x04 /* desynchronize children */
447 #define HAMMER2_BREF_DELETED 0x80 /* indicates a deletion */
450 #define HAMMER2_BLOCKREF_BYTES 64 /* blockref struct in bytes */
453 * On-media and off-media blockref types.
455 #define HAMMER2_BREF_TYPE_EMPTY 0
456 #define HAMMER2_BREF_TYPE_INODE 1
457 #define HAMMER2_BREF_TYPE_INDIRECT 2
458 #define HAMMER2_BREF_TYPE_DATA 3
459 #define HAMMER2_BREF_TYPE_UNUSED04 4
460 #define HAMMER2_BREF_TYPE_FREEMAP_NODE 5
461 #define HAMMER2_BREF_TYPE_FREEMAP_LEAF 6
462 #define HAMMER2_BREF_TYPE_FREEMAP 254 /* pseudo-type */
463 #define HAMMER2_BREF_TYPE_VOLUME 255 /* pseudo-type */
465 #define HAMMER2_ENC_CHECK(n) ((n) << 4)
466 #define HAMMER2_DEC_CHECK(n) (((n) >> 4) & 15)
468 #define HAMMER2_CHECK_NONE 0
469 #define HAMMER2_CHECK_ISCSI32 1
470 #define HAMMER2_CHECK_CRC64 2
471 #define HAMMER2_CHECK_SHA192 3
472 #define HAMMER2_CHECK_FREEMAP 4
474 #define HAMMER2_ENC_COMP(n) (n)
475 #define HAMMER2_ENC_LEVEL(n) ((n) << 4)
476 #define HAMMER2_DEC_COMP(n) ((n) & 15)
477 #define HAMMER2_DEC_LEVEL(n) (((n) >> 4) & 15)
479 #define HAMMER2_COMP_NONE 0
480 #define HAMMER2_COMP_AUTOZERO 1
481 #define HAMMER2_COMP_LZ4 2
482 #define HAMMER2_COMP_ZLIB 3
484 #define HAMMER2_COMP_NEWFS_DEFAULT HAMMER2_COMP_LZ4
485 #define HAMMER2_COMP_STRINGS { "none", "autozero", "lz4", "zlib" }
486 #define HAMMER2_COMP_STRINGS_COUNT 4
490 * HAMMER2 block references are collected into sets of 8 blockrefs. These
491 * sets are fully associative, meaning the elements making up a set are
492 * not sorted in any way and may contain duplicate entries, holes, or
493 * entries which shortcut multiple levels of indirection. Sets are used
496 * (1) When redundancy is desired a set may contain several duplicate
497 * entries pointing to different copies of the same data. Up to 8 copies
498 * are supported but the set structure becomes a bit inefficient once
501 * (2) The blockrefs in a set can shortcut multiple levels of indirections
502 * within the bounds imposed by the parent of set.
504 * When a set fills up another level of indirection is inserted, moving
505 * some or all of the set's contents into indirect blocks placed under the
506 * set. This is a top-down approach in that indirect blocks are not created
507 * until the set actually becomes full (that is, the entries in the set can
508 * shortcut the indirect blocks when the set is not full). Depending on how
509 * things are filled multiple indirect blocks will eventually be created.
511 * Indirect blocks are typically 4KB (64 entres) or 64KB (1024 entries) and
512 * are also treated as fully set-associative.
514 struct hammer2_blockset {
515 hammer2_blockref_t blockref[HAMMER2_SET_COUNT];
518 typedef struct hammer2_blockset hammer2_blockset_t;
521 * Catch programmer snafus
523 #if (1 << HAMMER2_SET_RADIX) != HAMMER2_SET_COUNT
524 #error "hammer2 direct radix is incorrect"
526 #if (1 << HAMMER2_PBUFRADIX) != HAMMER2_PBUFSIZE
527 #error "HAMMER2_PBUFRADIX and HAMMER2_PBUFSIZE are inconsistent"
529 #if (1 << HAMMER2_MIN_RADIX) != HAMMER2_MIN_ALLOC
530 #error "HAMMER2_MIN_RADIX and HAMMER2_MIN_ALLOC are inconsistent"
534 * hammer2_bmap_data - A freemap entry in the LEVEL1 block.
536 * Each 64-byte entry contains the bitmap and meta-data required to manage
537 * a LEVEL0 (2MB) block of storage. The storage is managed in 128 x 16KB
538 * chunks. Smaller allocation granularity is supported via a linear iterator
539 * and/or must otherwise be tracked in ram.
541 * (data structure must be 64 bytes exactly)
543 * linear - A BYTE linear allocation offset used for sub-16KB allocations
544 * only. May contain values between 0 and 2MB. Must be ignored
545 * if 16KB-aligned (i.e. force bitmap scan), otherwise may be
546 * used to sub-allocate within the 16KB block (which is already
547 * marked as allocated in the bitmap).
549 * Sub-allocations need only be 1KB-aligned and do not have to be
550 * size-aligned, and 16KB or larger allocations do not update this
551 * field, resulting in pretty good packing.
553 * Please note that file data granularity may be limited by
554 * other issues such as buffer cache direct-mapping and the
555 * desire to support sector sizes up to 16KB (so H2 only issues
556 * I/O's in multiples of 16KB anyway).
558 * class - Clustering class. Cleared to 0 only if the entire leaf becomes
559 * free. Used to cluster device buffers so all elements must have
560 * the same device block size, but may mix logical sizes.
562 * Typically integrated with the blockref type in the upper 8 bits
563 * to localize inodes and indrect blocks, improving bulk free scans
564 * and directory scans.
566 * bitmap - Two bits per 16KB allocation block arranged in arrays of
567 * 32-bit elements, 128x2 bits representing ~2MB worth of media
568 * storage. Bit patterns are as follows:
571 * 01 Armed for bulk free scan
575 struct hammer2_bmap_data {
576 int32_t linear; /* 00 linear sub-granular allocation offset */
577 uint16_t class; /* 04-05 clustering class ((type<<8)|radix) */
578 uint8_t reserved06; /* 06 */
579 uint8_t reserved07; /* 07 */
580 uint32_t reserved08; /* 08 */
581 uint32_t reserved0C; /* 0C */
582 uint32_t reserved10; /* 10 */
583 uint32_t reserved14; /* 14 */
584 uint32_t reserved18; /* 18 */
585 uint32_t avail; /* 1C */
586 uint32_t bitmap[8]; /* 20-3F 256 bits manages 2MB/16KB/2-bits */
589 typedef struct hammer2_bmap_data hammer2_bmap_data_t;
592 * In HAMMER2 inodes ARE directory entries, with a special exception for
593 * hardlinks. The inode number is stored in the inode rather than being
594 * based on the location of the inode (since the location moves every time
595 * the inode or anything underneath the inode is modified).
597 * The inode is 1024 bytes, made up of 256 bytes of meta-data, 256 bytes
598 * for the filename, and 512 bytes worth of direct file data OR an embedded
601 * Directories represent one inode per blockref. Inodes are not laid out
602 * as a file but instead are represented by the related blockrefs. The
603 * blockrefs, in turn, are indexed by the 64-bit directory hash key. Remember
604 * that blocksets are fully associative, so a certain degree efficiency is
605 * achieved just from that.
607 * Up to 512 bytes of direct data can be embedded in an inode, and since
608 * inodes are essentially directory entries this also means that small data
609 * files end up simply being laid out linearly in the directory, resulting
610 * in fewer seeks and highly optimal access.
612 * The compression mode can be changed at any time in the inode and is
613 * recorded on a blockref-by-blockref basis.
615 * Hardlinks are supported via the inode map. Essentially the way a hardlink
616 * works is that all individual directory entries representing the same file
617 * are special cased and specify the same inode number. The actual file
618 * is placed in the nearest parent directory that is parent to all instances
619 * of the hardlink. If all hardlinks to a file are in the same directory
620 * the actual file will also be placed in that directory. This file uses
621 * the inode number as the directory entry key and is invisible to normal
622 * directory scans. Real directory entry keys are differentiated from the
623 * inode number key via bit 63. Access to the hardlink silently looks up
624 * the real file and forwards all operations to that file. Removal of the
625 * last hardlink also removes the real file.
627 * (attr_tid) is only updated when the inode's specific attributes or regular
628 * file size has changed, and affects path lookups and stat. (attr_tid)
629 * represents a special cache coherency lock under the inode. The inode
630 * blockref's modify_tid will always cover it.
632 * (dirent_tid) is only updated when an entry under a directory inode has
633 * been created, deleted, renamed, or had its attributes change, and affects
634 * directory lookups and scans. (dirent_tid) represents another special cache
635 * coherency lock under the inode. The inode blockref's modify_tid will
638 #define HAMMER2_INODE_BYTES 1024 /* (asserted by code) */
639 #define HAMMER2_INODE_MAXNAME 256 /* maximum name in bytes */
640 #define HAMMER2_INODE_VERSION_ONE 1
642 struct hammer2_inode_data {
643 uint16_t version; /* 0000 inode data version */
644 uint16_t reserved02; /* 0002 */
647 * core inode attributes, inode type, misc flags
649 uint32_t uflags; /* 0004 chflags */
650 uint32_t rmajor; /* 0008 available for device nodes */
651 uint32_t rminor; /* 000C available for device nodes */
652 uint64_t ctime; /* 0010 inode change time */
653 uint64_t mtime; /* 0018 modified time */
654 uint64_t atime; /* 0020 access time (unsupported) */
655 uint64_t btime; /* 0028 birth time */
656 uuid_t uid; /* 0030 uid / degenerate unix uid */
657 uuid_t gid; /* 0040 gid / degenerate unix gid */
659 uint8_t type; /* 0050 object type */
660 uint8_t op_flags; /* 0051 operational flags */
661 uint16_t cap_flags; /* 0052 capability flags */
662 uint32_t mode; /* 0054 unix modes (typ low 16 bits) */
665 * inode size, identification, localized recursive configuration
666 * for compression and backup copies.
668 hammer2_tid_t inum; /* 0058 inode number */
669 hammer2_off_t size; /* 0060 size of file */
670 uint64_t nlinks; /* 0068 hard links (typ only dirs) */
671 hammer2_tid_t iparent; /* 0070 parent inum (recovery only) */
672 hammer2_key_t name_key; /* 0078 full filename key */
673 uint16_t name_len; /* 0080 filename length */
674 uint8_t ncopies; /* 0082 ncopies to local media */
675 uint8_t comp_algo; /* 0083 compression request & algo */
678 * These fields are currently only applicable to PFSROOTs.
680 * NOTE: We can't use {volume_data->fsid, pfs_clid} to uniquely
681 * identify an instance of a PFS in the cluster because
682 * a mount may contain more than one copy of the PFS as
683 * a separate node. {pfs_clid, pfs_fsid} must be used for
684 * registration in the cluster.
686 uint8_t target_type; /* 0084 hardlink target type */
687 uint8_t reserved85; /* 0085 */
688 uint8_t reserved86; /* 0086 */
689 uint8_t pfs_type; /* 0087 (if PFSROOT) node type */
690 uint64_t pfs_inum; /* 0088 (if PFSROOT) inum allocator */
691 uuid_t pfs_clid; /* 0090 (if PFSROOT) cluster uuid */
692 uuid_t pfs_fsid; /* 00A0 (if PFSROOT) unique uuid */
695 * Quotas and cumulative sub-tree counters.
697 hammer2_key_t data_quota; /* 00B0 subtree quota in bytes */
698 hammer2_key_t data_count; /* 00B8 subtree byte count */
699 hammer2_key_t inode_quota; /* 00C0 subtree quota inode count */
700 hammer2_key_t inode_count; /* 00C8 subtree inode count */
701 hammer2_tid_t attr_tid; /* 00D0 attributes changed */
702 hammer2_tid_t dirent_tid; /* 00D8 directory/attr changed */
705 * Tracks (possibly degenerate) free areas covering all sub-tree
706 * allocations under inode, not counting the inode itself.
707 * 0/0 indicates empty entry. fully set-associative.
709 hammer2_off_t freezones[4]; /* 00E0/E8/F0/F8 base|radix */
711 unsigned char filename[HAMMER2_INODE_MAXNAME];
712 /* 0100-01FF (256 char, unterminated) */
713 union { /* 0200-03FF (64x8 = 512 bytes) */
714 struct hammer2_blockset blockset;
715 char data[HAMMER2_EMBEDDED_BYTES];
719 typedef struct hammer2_inode_data hammer2_inode_data_t;
721 #define HAMMER2_OPFLAG_DIRECTDATA 0x01
722 #define HAMMER2_OPFLAG_PFSROOT 0x02
723 #define HAMMER2_OPFLAG_COPYIDS 0x04 /* copyids override parent */
725 #define HAMMER2_OBJTYPE_UNKNOWN 0
726 #define HAMMER2_OBJTYPE_DIRECTORY 1
727 #define HAMMER2_OBJTYPE_REGFILE 2
728 #define HAMMER2_OBJTYPE_FIFO 4
729 #define HAMMER2_OBJTYPE_CDEV 5
730 #define HAMMER2_OBJTYPE_BDEV 6
731 #define HAMMER2_OBJTYPE_SOFTLINK 7
732 #define HAMMER2_OBJTYPE_HARDLINK 8 /* dummy entry for hardlink */
733 #define HAMMER2_OBJTYPE_SOCKET 9
734 #define HAMMER2_OBJTYPE_WHITEOUT 10
736 #define HAMMER2_COPYID_NONE 0
737 #define HAMMER2_COPYID_LOCAL ((uint8_t)-1)
740 * PEER types identify connections and help cluster controller filter
741 * out unwanted SPANs.
743 #define HAMMER2_PEER_NONE DMSG_PEER_NONE
744 #define HAMMER2_PEER_CLUSTER DMSG_PEER_CLUSTER
745 #define HAMMER2_PEER_BLOCK DMSG_PEER_BLOCK
746 #define HAMMER2_PEER_HAMMER2 DMSG_PEER_HAMMER2
748 #define HAMMER2_COPYID_COUNT DMSG_COPYID_COUNT
751 * PFS types identify a PFS on media and in LNK_SPAN messages.
753 #define HAMMER2_PFSTYPE_NONE DMSG_PFSTYPE_NONE
754 #define HAMMER2_PFSTYPE_ADMIN DMSG_PFSTYPE_ADMIN
755 #define HAMMER2_PFSTYPE_CLIENT DMSG_PFSTYPE_CLIENT
756 #define HAMMER2_PFSTYPE_CACHE DMSG_PFSTYPE_CACHE
757 #define HAMMER2_PFSTYPE_COPY DMSG_PFSTYPE_COPY
758 #define HAMMER2_PFSTYPE_SLAVE DMSG_PFSTYPE_SLAVE
759 #define HAMMER2_PFSTYPE_SOFT_SLAVE DMSG_PFSTYPE_SOFT_SLAVE
760 #define HAMMER2_PFSTYPE_SOFT_MASTER DMSG_PFSTYPE_SOFT_MASTER
761 #define HAMMER2_PFSTYPE_MASTER DMSG_PFSTYPE_MASTER
762 #define HAMMER2_PFSTYPE_SNAPSHOT DMSG_PFSTYPE_SNAPSHOT
763 #define HAMMER2_PFSTYPE_MAX DMSG_PFSTYPE_MAX
772 * Flags (8 bits) - blockref, for freemap only
774 * Note that the minimum chunk size is 1KB so we could theoretically have
775 * 10 bits here, but we might have some future extension that allows a
776 * chunk size down to 256 bytes and if so we will need bits 8 and 9.
778 #define HAMMER2_AVF_SELMASK 0x03 /* select group */
779 #define HAMMER2_AVF_ALL_ALLOC 0x04 /* indicate all allocated */
780 #define HAMMER2_AVF_ALL_FREE 0x08 /* indicate all free */
781 #define HAMMER2_AVF_RESERVED10 0x10
782 #define HAMMER2_AVF_RESERVED20 0x20
783 #define HAMMER2_AVF_RESERVED40 0x40
784 #define HAMMER2_AVF_RESERVED80 0x80
785 #define HAMMER2_AVF_AVMASK32 ((uint32_t)0xFFFFFF00LU)
786 #define HAMMER2_AVF_AVMASK64 ((uint64_t)0xFFFFFFFFFFFFFF00LLU)
788 #define HAMMER2_AV_SELECT_A 0x00
789 #define HAMMER2_AV_SELECT_B 0x01
790 #define HAMMER2_AV_SELECT_C 0x02
791 #define HAMMER2_AV_SELECT_D 0x03
794 * The volume header eats a 64K block. There is currently an issue where
795 * we want to try to fit all nominal filesystem updates in a 512-byte section
796 * but it may be a lost cause due to the need for a blockset.
798 * All information is stored in host byte order. The volume header's magic
799 * number may be checked to determine the byte order. If you wish to mount
800 * between machines w/ different endian modes you'll need filesystem code
801 * which acts on the media data consistently (either all one way or all the
802 * other). Our code currently does not do that.
804 * A read-write mount may have to recover missing allocations by doing an
805 * incremental mirror scan looking for modifications made after alloc_tid.
806 * If alloc_tid == last_tid then no recovery operation is needed. Recovery
807 * operations are usually very, very fast.
809 * Read-only mounts do not need to do any recovery, access to the filesystem
810 * topology is always consistent after a crash (is always consistent, period).
811 * However, there may be shortcutted blockref updates present from deep in
812 * the tree which are stored in the volumeh eader and must be tracked on
815 * NOTE: The copyinfo[] array contains the configuration for both the
816 * cluster connections and any local media copies. The volume
817 * header will be replicated for each local media copy.
819 * The mount command may specify multiple medias or just one and
820 * allow HAMMER2 to pick up the others when it checks the copyinfo[]
823 * NOTE: root_blockref points to the super-root directory, not the root
824 * directory. The root directory will be a subdirectory under the
827 * The super-root directory contains all root directories and all
828 * snapshots (readonly or writable). It is possible to do a
829 * null-mount of the super-root using special path constructions
830 * relative to your mounted root.
832 * NOTE: HAMMER2 allows any subdirectory tree to be managed as if it were
833 * a PFS, including mirroring and storage quota operations, and this is
834 * prefered over creating discrete PFSs in the super-root. Instead
835 * the super-root is most typically used to create writable snapshots,
836 * alternative roots, and so forth. The super-root is also used by
837 * the automatic snapshotting mechanism.
839 #define HAMMER2_VOLUME_ID_HBO 0x48414d3205172011LLU
840 #define HAMMER2_VOLUME_ID_ABO 0x11201705324d4148LLU
842 struct hammer2_volume_data {
844 * sector #0 - 512 bytes
846 uint64_t magic; /* 0000 Signature */
847 hammer2_off_t boot_beg; /* 0008 Boot area (future) */
848 hammer2_off_t boot_end; /* 0010 (size = end - beg) */
849 hammer2_off_t aux_beg; /* 0018 Aux area (future) */
850 hammer2_off_t aux_end; /* 0020 (size = end - beg) */
851 hammer2_off_t volu_size; /* 0028 Volume size, bytes */
853 uint32_t version; /* 0030 */
854 uint32_t flags; /* 0034 */
855 uint8_t copyid; /* 0038 copyid of phys vol */
856 uint8_t freemap_version; /* 0039 freemap algorithm */
857 uint8_t peer_type; /* 003A HAMMER2_PEER_xxx */
858 uint8_t reserved003B; /* 003B */
859 uint32_t reserved003C; /* 003C */
861 uuid_t fsid; /* 0040 */
862 uuid_t fstype; /* 0050 */
865 * allocator_size is precalculated at newfs time and does not include
866 * reserved blocks, boot, or redo areas.
868 * Initial non-reserved-area allocations do not use the freemap
869 * but instead adjust alloc_iterator. Dynamic allocations take
870 * over starting at (allocator_beg). This makes newfs_hammer2's
871 * job a lot easier and can also serve as a testing jig.
873 hammer2_off_t allocator_size; /* 0060 Total data space */
874 hammer2_off_t allocator_free; /* 0068 Free space */
875 hammer2_off_t allocator_beg; /* 0070 Initial allocations */
876 hammer2_tid_t mirror_tid; /* 0078 committed tid (vol) */
877 hammer2_tid_t alloc_tid; /* 0080 Alloctable modify tid */
878 hammer2_tid_t inode_tid; /* 0088 Inode allocator tid */
879 hammer2_tid_t freemap_tid; /* 0090 committed tid (fmap) */
880 hammer2_tid_t reserved0090[6]; /* 0098-00C7 */
883 * Copyids are allocated dynamically from the copyexists bitmap.
884 * An id from the active copies set (up to 8, see copyinfo later on)
885 * may still exist after the copy set has been removed from the
886 * volume header and its bit will remain active in the bitmap and
887 * cannot be reused until it is 100% removed from the hierarchy.
889 uint32_t copyexists[8]; /* 00C8-00E7 copy exists bmap */
890 char reserved0140[248]; /* 00E8-01DF */
893 * 32 bit CRC array at the end of the first 512 byte sector.
895 * icrc_sects[7] - First 512-4 bytes of volume header (including all
896 * the other icrc's except this one).
898 * icrc_sects[6] - Sector 1 (512 bytes) of volume header, which is
899 * the blockset for the root.
901 * icrc_sects[5] - Sector 2
902 * icrc_sects[4] - Sector 3
903 * icrc_sects[3] - Sector 4 (the freemap blockset)
905 hammer2_crc32_t icrc_sects[8]; /* 01E0-01FF */
908 * sector #1 - 512 bytes
910 * The entire sector is used by a blockset.
912 hammer2_blockset_t sroot_blockset; /* 0200-03FF Superroot dir */
917 char sector2[512]; /* 0400-05FF reserved */
918 char sector3[512]; /* 0600-07FF reserved */
919 hammer2_blockset_t freemap_blockset; /* 0800-09FF freemap */
920 char sector5[512]; /* 0A00-0BFF reserved */
921 char sector6[512]; /* 0C00-0DFF reserved */
922 char sector7[512]; /* 0E00-0FFF reserved */
925 * sector #8-71 - 32768 bytes
927 * Contains the configuration for up to 256 copyinfo targets. These
928 * specify local and remote copies operating as masters or slaves.
929 * copyid's 0 and 255 are reserved (0 indicates an empty slot and 255
930 * indicates the local media).
932 * Each inode contains a set of up to 8 copyids, either inherited
933 * from its parent or explicitly specified in the inode, which
934 * indexes into this array.
936 /* 1000-8FFF copyinfo config */
937 dmsg_vol_data_t copyinfo[HAMMER2_COPYID_COUNT];
940 * Remaining sections are reserved for future use.
942 char reserved0400[0x6FFC]; /* 9000-FFFB reserved */
945 * icrc on entire volume header
947 hammer2_crc32_t icrc_volheader; /* FFFC-FFFF full volume icrc*/
950 typedef struct hammer2_volume_data hammer2_volume_data_t;
953 * Various parts of the volume header have their own iCRCs.
955 * The first 512 bytes has its own iCRC stored at the end of the 512 bytes
956 * and not included the icrc calculation.
958 * The second 512 bytes also has its own iCRC but it is stored in the first
959 * 512 bytes so it covers the entire second 512 bytes.
961 * The whole volume block (64KB) has an iCRC covering all but the last 4 bytes,
962 * which is where the iCRC for the whole volume is stored. This is currently
963 * a catch-all for anything not individually iCRCd.
965 #define HAMMER2_VOL_ICRC_SECT0 7
966 #define HAMMER2_VOL_ICRC_SECT1 6
968 #define HAMMER2_VOLUME_BYTES 65536
970 #define HAMMER2_VOLUME_ICRC0_OFF 0
971 #define HAMMER2_VOLUME_ICRC1_OFF 512
972 #define HAMMER2_VOLUME_ICRCVH_OFF 0
974 #define HAMMER2_VOLUME_ICRC0_SIZE (512 - 4)
975 #define HAMMER2_VOLUME_ICRC1_SIZE (512)
976 #define HAMMER2_VOLUME_ICRCVH_SIZE (65536 - 4)
978 #define HAMMER2_VOL_VERSION_MIN 1
979 #define HAMMER2_VOL_VERSION_DEFAULT 1
980 #define HAMMER2_VOL_VERSION_WIP 2
982 #define HAMMER2_NUM_VOLHDRS 4
984 union hammer2_media_data {
985 hammer2_volume_data_t voldata;
986 hammer2_inode_data_t ipdata;
987 hammer2_blockref_t npdata[HAMMER2_IND_COUNT_MAX];
988 hammer2_bmap_data_t bmdata[HAMMER2_FREEMAP_COUNT];
989 char buf[HAMMER2_PBUFSIZE];
992 typedef union hammer2_media_data hammer2_media_data_t;
994 #endif /* !_VFS_HAMMER2_DISK_H_ */