2 * Copyright (c) 2007 The DragonFly Project. All rights reserved.
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34 * $DragonFly: src/sys/vfs/hammer/hammer_disk.h,v 1.55 2008/11/13 02:18:43 dillon Exp $
37 #ifndef VFS_HAMMER_DISK_H_
38 #define VFS_HAMMER_DISK_H_
45 * The structures below represent the on-disk format for a HAMMER
46 * filesystem. Note that all fields for on-disk structures are naturally
47 * aligned. The host endian format is used - compatibility is possible
48 * if the implementation detects reversed endian and adjusts data accordingly.
50 * Most of HAMMER revolves around the concept of an object identifier. An
51 * obj_id is a 64 bit quantity which uniquely identifies a filesystem object
52 * FOR THE ENTIRE LIFE OF THE FILESYSTEM. This uniqueness allows backups
53 * and mirrors to retain varying amounts of filesystem history by removing
54 * any possibility of conflict through identifier reuse.
56 * A HAMMER filesystem may span multiple volumes.
58 * A HAMMER filesystem uses a 16K filesystem buffer size. All filesystem
59 * I/O is done in multiples of 16K.
61 * 64K X-bufs are used for blocks >= a file's 1MB mark.
63 * Per-volume storage limit: 52 bits 4096 TB
64 * Per-Zone storage limit: 60 bits 1 MTB
65 * Per-filesystem storage limit: 60 bits 1 MTB
67 #define HAMMER_BUFSIZE 16384
68 #define HAMMER_XBUFSIZE 65536
69 #define HAMMER_HBUFSIZE (HAMMER_BUFSIZE / 2)
70 #define HAMMER_XDEMARC (1024 * 1024)
71 #define HAMMER_BUFMASK (HAMMER_BUFSIZE - 1)
72 #define HAMMER_XBUFMASK (HAMMER_XBUFSIZE - 1)
74 #define HAMMER_BUFSIZE64 ((uint64_t)HAMMER_BUFSIZE)
75 #define HAMMER_BUFMASK64 ((uint64_t)HAMMER_BUFMASK)
77 #define HAMMER_XBUFSIZE64 ((uint64_t)HAMMER_XBUFSIZE)
78 #define HAMMER_XBUFMASK64 ((uint64_t)HAMMER_XBUFMASK)
80 #define HAMMER_OFF_ZONE_MASK 0xF000000000000000ULL /* zone portion */
81 #define HAMMER_OFF_VOL_MASK 0x0FF0000000000000ULL /* volume portion */
82 #define HAMMER_OFF_SHORT_MASK 0x000FFFFFFFFFFFFFULL /* offset portion */
83 #define HAMMER_OFF_LONG_MASK 0x0FFFFFFFFFFFFFFFULL /* offset portion */
85 #define HAMMER_OFF_BAD ((hammer_off_t)-1)
88 * The current limit of volumes that can make up a HAMMER FS
90 #define HAMMER_MAX_VOLUMES 256
93 * Hammer transaction ids are 64 bit unsigned integers and are usually
94 * synchronized with the time of day in nanoseconds.
96 * Hammer offsets are used for FIFO indexing and embed a cycle counter
97 * and volume number in addition to the offset. Most offsets are required
98 * to be 16 KB aligned.
100 typedef uint64_t hammer_tid_t;
101 typedef uint64_t hammer_off_t;
102 typedef uint32_t hammer_crc_t;
104 #define HAMMER_MIN_TID 0ULL /* unsigned */
105 #define HAMMER_MAX_TID 0xFFFFFFFFFFFFFFFFULL /* unsigned */
106 #define HAMMER_MIN_KEY -0x8000000000000000LL /* signed */
107 #define HAMMER_MAX_KEY 0x7FFFFFFFFFFFFFFFLL /* signed */
108 #define HAMMER_MIN_OBJID HAMMER_MIN_KEY /* signed */
109 #define HAMMER_MAX_OBJID HAMMER_MAX_KEY /* signed */
110 #define HAMMER_MIN_RECTYPE 0x0U /* unsigned */
111 #define HAMMER_MAX_RECTYPE 0xFFFFU /* unsigned */
112 #define HAMMER_MIN_OFFSET 0ULL /* unsigned */
113 #define HAMMER_MAX_OFFSET 0xFFFFFFFFFFFFFFFFULL /* unsigned */
116 * hammer_off_t has several different encodings. Note that not all zones
117 * encode a vol_no. Zone bits are not a part of filesystem capacity.
119 * zone 0: available, a big-block that contains the offset is unused
120 * zone 1 (z,v,o): raw volume relative (offset 0 is the volume header)
121 * zone 2 (z,v,o): raw buffer relative (offset 0 is the first buffer)
122 * zone 3 (z,o): undo fifo - actually zone-2 address, fixed phys array in vol hdr
123 * zone 4 (z,v,o): freemap - only real blockmap
124 * zone 8 (z,v,o): B-Tree - actually zone-2 address
125 * zone 9 (z,v,o): meta - actually zone-2 address
126 * zone 10 (z,v,o): large-data - actually zone-2 address
127 * zone 11 (z,v,o): small-data - actually zone-2 address
128 * zone 15: unavailable, usually the offset is beyond volume size
130 * layer1/layer2 direct map:
131 * Maximum HAMMER filesystem capacity from volume aspect
132 * 2^8(max volumes) * 2^52(max volume size) = 2^60 = 1EB
133 * <------------------------------------------------------------->
135 * <------><----------------------------------------------------->
136 * zzzzvvvvvvvvoooo oooooooooooooooo oooooooooooooooo oooooooooooooooo
137 * ----111111111111 1111112222222222 222222222ooooooo oooooooooooooooo
138 * <-----------------><------------------><---------------------->
139 * 18bits 19bits 23bits
140 * <------------------------------------------------------------->
141 * 2^18(layer1) * 2^19(layer2) * 2^23(big-block) = 2^60 = 1EB
142 * Maximum HAMMER filesystem capacity from blockmap aspect
145 #define HAMMER_ZONE_RAW_VOLUME 0x1000000000000000ULL
146 #define HAMMER_ZONE_RAW_BUFFER 0x2000000000000000ULL
147 #define HAMMER_ZONE_UNDO 0x3000000000000000ULL
148 #define HAMMER_ZONE_FREEMAP 0x4000000000000000ULL
149 #define HAMMER_ZONE_RESERVED05 0x5000000000000000ULL /* not used */
150 #define HAMMER_ZONE_RESERVED06 0x6000000000000000ULL /* not used */
151 #define HAMMER_ZONE_RESERVED07 0x7000000000000000ULL /* not used */
152 #define HAMMER_ZONE_BTREE 0x8000000000000000ULL
153 #define HAMMER_ZONE_META 0x9000000000000000ULL
154 #define HAMMER_ZONE_LARGE_DATA 0xA000000000000000ULL
155 #define HAMMER_ZONE_SMALL_DATA 0xB000000000000000ULL
156 #define HAMMER_ZONE_RESERVED0C 0xC000000000000000ULL /* not used */
157 #define HAMMER_ZONE_RESERVED0D 0xD000000000000000ULL /* not used */
158 #define HAMMER_ZONE_RESERVED0E 0xE000000000000000ULL /* not used */
159 #define HAMMER_ZONE_UNAVAIL 0xF000000000000000ULL
161 #define HAMMER_ZONE_RAW_VOLUME_INDEX 1
162 #define HAMMER_ZONE_RAW_BUFFER_INDEX 2
163 #define HAMMER_ZONE_UNDO_INDEX 3
164 #define HAMMER_ZONE_FREEMAP_INDEX 4
165 #define HAMMER_ZONE_BTREE_INDEX 8
166 #define HAMMER_ZONE_META_INDEX 9
167 #define HAMMER_ZONE_LARGE_DATA_INDEX 10
168 #define HAMMER_ZONE_SMALL_DATA_INDEX 11
169 #define HAMMER_ZONE_UNAVAIL_INDEX 15
171 #define HAMMER_MAX_ZONES 16
173 #define HAMMER_ZONE(offset) ((offset) & HAMMER_OFF_ZONE_MASK)
175 #define hammer_is_zone_raw_volume(offset) \
176 (HAMMER_ZONE(offset) == HAMMER_ZONE_RAW_VOLUME)
177 #define hammer_is_zone_raw_buffer(offset) \
178 (HAMMER_ZONE(offset) == HAMMER_ZONE_RAW_BUFFER)
179 #define hammer_is_zone_undo(offset) \
180 (HAMMER_ZONE(offset) == HAMMER_ZONE_UNDO)
181 #define hammer_is_zone_freemap(offset) \
182 (HAMMER_ZONE(offset) == HAMMER_ZONE_FREEMAP)
183 #define hammer_is_zone_btree(offset) \
184 (HAMMER_ZONE(offset) == HAMMER_ZONE_BTREE)
185 #define hammer_is_zone_meta(offset) \
186 (HAMMER_ZONE(offset) == HAMMER_ZONE_META)
187 #define hammer_is_zone_large_data(offset) \
188 (HAMMER_ZONE(offset) == HAMMER_ZONE_LARGE_DATA)
189 #define hammer_is_zone_small_data(offset) \
190 (HAMMER_ZONE(offset) == HAMMER_ZONE_SMALL_DATA)
191 #define hammer_is_zone_unavail(offset) \
192 (HAMMER_ZONE(offset) == HAMMER_ZONE_UNAVAIL)
195 * Test if the zone is directly mapped to zone-2 offset via freemap.
197 #define hammer_is_zone2_mapped_index(zone) \
198 ((zone) >= HAMMER_ZONE_BTREE_INDEX && \
199 (zone) < HAMMER_MAX_ZONES)
201 * Test if the zone is directly mapped to zone-2 offset. The word
202 * directly here means the zone is neither RAW_VOLUME nor UNDO zone.
204 #define hammer_is_direct_mapped_index(zone) \
205 (((zone) == HAMMER_ZONE_RAW_BUFFER_INDEX) || \
206 ((zone) == HAMMER_ZONE_FREEMAP_INDEX) || \
207 hammer_is_zone2_mapped_index(zone))
209 #define HAMMER_ZONE_ENCODE(zone, ham_off) \
210 (((hammer_off_t)(zone) << 60) | (ham_off))
211 #define HAMMER_ZONE_DECODE(ham_off) \
212 ((int)(((hammer_off_t)(ham_off) >> 60)))
214 #define HAMMER_VOL_ENCODE(vol_no) \
215 ((hammer_off_t)((vol_no) & 255) << 52)
216 #define HAMMER_VOL_DECODE(ham_off) \
217 ((int)(((hammer_off_t)(ham_off) >> 52) & 255))
219 #define HAMMER_OFF_SHORT_ENCODE(offset) \
220 ((hammer_off_t)(offset) & HAMMER_OFF_SHORT_MASK)
221 #define HAMMER_OFF_LONG_ENCODE(offset) \
222 ((hammer_off_t)(offset) & HAMMER_OFF_LONG_MASK)
224 #define HAMMER_ENCODE(zone, vol_no, offset) \
225 (((hammer_off_t)(zone) << 60) | \
226 HAMMER_VOL_ENCODE(vol_no) | \
227 HAMMER_OFF_SHORT_ENCODE(offset))
228 #define HAMMER_ENCODE_RAW_VOLUME(vol_no, offset) \
229 HAMMER_ENCODE(HAMMER_ZONE_RAW_VOLUME_INDEX, vol_no, offset)
230 #define HAMMER_ENCODE_RAW_BUFFER(vol_no, offset) \
231 HAMMER_ENCODE(HAMMER_ZONE_RAW_BUFFER_INDEX, vol_no, offset)
232 #define HAMMER_ENCODE_FREEMAP(vol_no, offset) \
233 HAMMER_ENCODE(HAMMER_ZONE_FREEMAP_INDEX, vol_no, offset)
236 * Translate a zone address to zone-X address.
238 #define hammer_xlate_to_zoneX(zone, offset) \
239 HAMMER_ZONE_ENCODE((zone), (offset) & ~HAMMER_OFF_ZONE_MASK)
240 #define hammer_xlate_to_zone2(offset) \
241 hammer_xlate_to_zoneX(HAMMER_ZONE_RAW_BUFFER_INDEX, (offset))
243 #define hammer_data_zone(data_len) \
244 (((data_len) >= HAMMER_BUFSIZE) ? \
245 HAMMER_ZONE_LARGE_DATA : \
246 HAMMER_ZONE_SMALL_DATA)
247 #define hammer_data_zone_index(data_len) \
248 (((data_len) >= HAMMER_BUFSIZE) ? \
249 HAMMER_ZONE_LARGE_DATA_INDEX : \
250 HAMMER_ZONE_SMALL_DATA_INDEX)
253 * Big-Block backing store
255 * A blockmap is a two-level map which translates a blockmap-backed zone
256 * offset into a raw zone 2 offset. The layer 1 handles 18 bits and the
257 * layer 2 handles 19 bits. The 8M big-block size is 23 bits so two
258 * layers gives us 18+19+23 = 60 bits of address space.
260 * When using hinting for a blockmap lookup, the hint is lost when the
261 * scan leaves the HINTBLOCK, which is typically several BIGBLOCK's.
262 * HINTBLOCK is a heuristic.
264 #define HAMMER_HINTBLOCK_SIZE (HAMMER_BIGBLOCK_SIZE * 4)
265 #define HAMMER_HINTBLOCK_MASK64 ((uint64_t)HAMMER_HINTBLOCK_SIZE - 1)
266 #define HAMMER_BIGBLOCK_SIZE (8192 * 1024)
267 #define HAMMER_BIGBLOCK_SIZE64 ((uint64_t)HAMMER_BIGBLOCK_SIZE)
268 #define HAMMER_BIGBLOCK_MASK (HAMMER_BIGBLOCK_SIZE - 1)
269 #define HAMMER_BIGBLOCK_MASK64 ((uint64_t)HAMMER_BIGBLOCK_SIZE - 1)
270 #define HAMMER_BIGBLOCK_BITS 23
272 #define HAMMER_BIGBLOCK_OVERFILL (6144 * 1024)
274 #if (1 << HAMMER_BIGBLOCK_BITS) != HAMMER_BIGBLOCK_SIZE
275 #error "HAMMER_BIGBLOCK_BITS BROKEN"
278 #define HAMMER_BUFFERS_PER_BIGBLOCK \
279 (HAMMER_BIGBLOCK_SIZE / HAMMER_BUFSIZE)
280 #define HAMMER_BUFFERS_PER_BIGBLOCK_MASK \
281 (HAMMER_BUFFERS_PER_BIGBLOCK - 1)
282 #define HAMMER_BUFFERS_PER_BIGBLOCK_MASK64 \
283 ((hammer_off_t)HAMMER_BUFFERS_PER_BIGBLOCK_MASK)
286 * Maximum number of mirrors operating in master mode (multi-master
287 * clustering and mirroring). Note that HAMMER1 does not support
288 * multi-master clustering as of 2015.
290 #define HAMMER_MAX_MASTERS 16
293 * The blockmap is somewhat of a degenerate structure. HAMMER only actually
294 * uses it in its original incarnation to implement the freemap.
296 * zone:1 raw volume (no blockmap)
297 * zone:2 raw buffer (no blockmap)
298 * zone:3 undomap (direct layer2 array in volume header)
299 * zone:4 freemap (the only real blockmap)
300 * zone:8-15 zone id used to classify big-block only, address is actually
303 struct hammer_blockmap {
304 hammer_off_t phys_offset; /* zone-2 physical offset */
305 hammer_off_t first_offset; /* zone-X logical offset (zone 3) */
306 hammer_off_t next_offset; /* zone-X logical offset */
307 hammer_off_t alloc_offset; /* zone-X logical offset */
309 hammer_crc_t entry_crc;
312 typedef struct hammer_blockmap *hammer_blockmap_t;
314 #define HAMMER_BLOCKMAP_CRCSIZE \
315 offsetof(struct hammer_blockmap, entry_crc)
318 * The blockmap is a 2-layer entity made up of big-blocks. The first layer
319 * contains 262144 32-byte entries (18 bits), the second layer contains
320 * 524288 16-byte entries (19 bits), representing 8MB (23 bit) blockmaps.
321 * 18+19+23 = 60 bits. The top four bits are the zone id.
323 * Currently only the freemap utilizes both layers in all their glory.
324 * All primary data/meta-data zones actually encode a zone-2 address
325 * requiring no real blockmap translation.
327 * The freemap uses the upper 8 bits of layer-1 to identify the volume,
328 * thus any space allocated via the freemap can be directly translated
329 * to a zone:2 (or zone:8-15) address.
331 * zone-X blockmap offset: [zone:4][layer1:18][layer2:19][big-block:23]
335 * 32 bytes layer1 entry for 8MB big-block.
336 * A big-block can hold 2^23 / 2^5 = 2^18 layer1 entries,
337 * which equals bits assigned for layer1 in zone-2 address.
339 struct hammer_blockmap_layer1 {
340 hammer_off_t blocks_free; /* big-blocks free */
341 hammer_off_t phys_offset; /* UNAVAIL or zone-2 */
342 hammer_off_t reserved01;
343 hammer_crc_t layer2_crc; /* xor'd crc's of HAMMER_BLOCKSIZE */
345 hammer_crc_t layer1_crc; /* MUST BE LAST FIELD OF STRUCTURE*/
348 typedef struct hammer_blockmap_layer1 *hammer_blockmap_layer1_t;
350 #define HAMMER_LAYER1_CRCSIZE \
351 offsetof(struct hammer_blockmap_layer1, layer1_crc)
354 * 16 bytes layer2 entry for 8MB big-blocks.
355 * A big-block can hold 2^23 / 2^4 = 2^19 layer2 entries,
356 * which equals bits assigned for layer2 in zone-2 address.
358 * NOTE: bytes_free is signed and can legally go negative if/when data
359 * de-dup occurs. This field will never go higher than
360 * HAMMER_BIGBLOCK_SIZE. If exactly HAMMER_BIGBLOCK_SIZE
361 * the big-block is completely free.
363 struct hammer_blockmap_layer2 {
364 uint8_t zone; /* typed allocation zone */
367 uint32_t append_off; /* allocatable space index */
368 int32_t bytes_free; /* bytes free within this big-block */
369 hammer_crc_t entry_crc;
372 typedef struct hammer_blockmap_layer2 *hammer_blockmap_layer2_t;
374 #define HAMMER_LAYER2_CRCSIZE \
375 offsetof(struct hammer_blockmap_layer2, entry_crc)
377 #define HAMMER_BLOCKMAP_UNAVAIL ((hammer_off_t)-1LL)
379 #define HAMMER_BLOCKMAP_RADIX1 /* 2^18 = 262144 */ \
380 ((int)(HAMMER_BIGBLOCK_SIZE / sizeof(struct hammer_blockmap_layer1)))
381 #define HAMMER_BLOCKMAP_RADIX2 /* 2^19 = 524288 */ \
382 ((int)(HAMMER_BIGBLOCK_SIZE / sizeof(struct hammer_blockmap_layer2)))
384 #define HAMMER_BLOCKMAP_LAYER1 /* 2^(18+19+23) = 1EB */ \
385 (HAMMER_BLOCKMAP_RADIX1 * HAMMER_BLOCKMAP_LAYER2)
386 #define HAMMER_BLOCKMAP_LAYER2 /* 2^(19+23) = 4TB */ \
387 (HAMMER_BLOCKMAP_RADIX2 * HAMMER_BIGBLOCK_SIZE64)
389 #define HAMMER_BLOCKMAP_LAYER1_MASK (HAMMER_BLOCKMAP_LAYER1 - 1)
390 #define HAMMER_BLOCKMAP_LAYER2_MASK (HAMMER_BLOCKMAP_LAYER2 - 1)
393 * Index within layer1 or layer2 big-block for the entry representing
394 * a zone-2 physical offset.
396 #define HAMMER_BLOCKMAP_LAYER1_INDEX(zone2_offset) \
397 ((int)(((zone2_offset) & HAMMER_BLOCKMAP_LAYER1_MASK) / \
398 HAMMER_BLOCKMAP_LAYER2))
400 #define HAMMER_BLOCKMAP_LAYER2_INDEX(zone2_offset) \
401 ((int)(((zone2_offset) & HAMMER_BLOCKMAP_LAYER2_MASK) / \
402 HAMMER_BIGBLOCK_SIZE64))
405 * Byte offset within layer1 or layer2 big-block for the entry representing
406 * a zone-2 physical offset. Multiply the index by sizeof(blockmap_layer).
408 #define HAMMER_BLOCKMAP_LAYER1_OFFSET(zone2_offset) \
409 (HAMMER_BLOCKMAP_LAYER1_INDEX(zone2_offset) * \
410 sizeof(struct hammer_blockmap_layer1))
412 #define HAMMER_BLOCKMAP_LAYER2_OFFSET(zone2_offset) \
413 (HAMMER_BLOCKMAP_LAYER2_INDEX(zone2_offset) * \
414 sizeof(struct hammer_blockmap_layer2))
417 * HAMMER UNDO parameters. The UNDO fifo is mapped directly in the volume
418 * header with an array of layer2 structures. A maximum of (128x8MB) = 1GB
419 * may be reserved. The size of the undo fifo is usually set a newfs time
420 * but can be adjusted if the filesystem is taken offline.
422 #define HAMMER_UNDO_LAYER2 128 /* max layer2 undo mapping entries */
425 * All on-disk HAMMER structures which make up elements of the UNDO FIFO
426 * contain a hammer_fifo_head and hammer_fifo_tail structure. This structure
427 * contains all the information required to validate the fifo element
428 * and to scan the fifo in either direction. The head is typically embedded
429 * in higher level hammer on-disk structures while the tail is typically
430 * out-of-band. hdr_size is the size of the whole mess, including the tail.
432 * All undo structures are guaranteed to not cross a 16K filesystem
433 * buffer boundary. Most undo structures are fairly small. Data spaces
434 * are not immediately reused by HAMMER so file data is not usually recorded
435 * as part of an UNDO.
437 * PAD elements are allowed to take up only 8 bytes of space as a special
438 * case, containing only hdr_signature, hdr_type, and hdr_size fields,
439 * and with the tail overloaded onto the head structure for 8 bytes total.
441 * Every undo record has a sequence number. This number is unrelated to
442 * transaction ids and instead collects the undo transactions associated
443 * with a single atomic operation. A larger transactional operation, such
444 * as a remove(), may consist of several smaller atomic operations
445 * representing raw meta-data operations.
447 * HAMMER VERSION 4 CHANGES
449 * In HAMMER version 4 the undo structure alignment is reduced from 16384
450 * to 512 bytes in order to ensure that each 512 byte sector begins with
451 * a header. The reserved01 field in the header is now a 32 bit sequence
452 * number. This allows the recovery code to detect missing sectors
453 * without relying on the 32-bit crc and to definitively identify the current
454 * undo sequence space without having to rely on information from the volume
455 * header. In addition, new REDO entries in the undo space are used to
456 * record write, write/extend, and transaction id updates.
458 * The grand result is:
460 * (1) The volume header no longer needs to be synchronized for most
461 * flush and fsync operations.
463 * (2) Most fsync operations need only lay down REDO records
465 * (3) Data overwrite for nohistory operations covered by REDO records
466 * can be supported (instead of rolling a new block allocation),
467 * by rolling UNDO for the prior contents of the data.
469 * HAMMER VERSION 5 CHANGES
471 * Hammer version 5 contains a minor adjustment making layer2's bytes_free
472 * field signed, allowing dedup to push it into the negative domain.
474 #define HAMMER_HEAD_ALIGN 8
475 #define HAMMER_HEAD_ALIGN_MASK (HAMMER_HEAD_ALIGN - 1)
476 #define HAMMER_HEAD_DOALIGN(bytes) \
477 (((bytes) + HAMMER_HEAD_ALIGN_MASK) & ~HAMMER_HEAD_ALIGN_MASK)
479 #define HAMMER_UNDO_ALIGN 512
480 #define HAMMER_UNDO_ALIGN64 ((uint64_t)512)
481 #define HAMMER_UNDO_MASK (HAMMER_UNDO_ALIGN - 1)
482 #define HAMMER_UNDO_MASK64 (HAMMER_UNDO_ALIGN64 - 1)
484 struct hammer_fifo_head {
485 uint16_t hdr_signature;
487 uint32_t hdr_size; /* Aligned size of the whole mess */
488 uint32_t hdr_seq; /* Sequence number */
489 hammer_crc_t hdr_crc; /* XOR crc up to field w/ crc after field */
492 #define HAMMER_FIFO_HEAD_CRCOFF offsetof(struct hammer_fifo_head, hdr_crc)
494 struct hammer_fifo_tail {
495 uint16_t tail_signature;
497 uint32_t tail_size; /* aligned size of the whole mess */
500 typedef struct hammer_fifo_head *hammer_fifo_head_t;
501 typedef struct hammer_fifo_tail *hammer_fifo_tail_t;
506 * NOTE: 0x8000U part of HAMMER_HEAD_TYPE_PAD can be removed if the HAMMER
507 * version ever gets bumped again. It exists only to keep compatibility with
510 #define HAMMER_HEAD_TYPE_PAD (0x0040U | 0x8000U)
511 #define HAMMER_HEAD_TYPE_DUMMY 0x0041U /* dummy entry w/seqno */
512 #define HAMMER_HEAD_TYPE_UNDO 0x0043U /* random UNDO information */
513 #define HAMMER_HEAD_TYPE_REDO 0x0044U /* data REDO / fast fsync */
515 #define HAMMER_HEAD_SIGNATURE 0xC84EU
516 #define HAMMER_TAIL_SIGNATURE 0xC74FU
519 * Misc FIFO structures.
521 * UNDO - Raw meta-data media updates.
523 struct hammer_fifo_undo {
524 struct hammer_fifo_head head;
525 hammer_off_t undo_offset; /* zone-1,2 offset */
526 int32_t undo_data_bytes;
527 int32_t undo_reserved01;
528 /* followed by data */
532 * REDO (HAMMER version 4+) - Logical file writes/truncates.
534 * REDOs contain information which will be duplicated in a later meta-data
535 * update, allowing fast write()+fsync() operations. REDOs can be ignored
536 * without harming filesystem integrity but must be processed if fsync()
537 * semantics are desired.
539 * Unlike UNDOs which are processed backwards within the recovery span,
540 * REDOs must be processed forwards starting further back (starting outside
541 * the recovery span).
543 * WRITE - Write logical file (with payload). Executed both
544 * out-of-span and in-span. Out-of-span WRITEs may be
545 * filtered out by TERMs.
547 * TRUNC - Truncate logical file (no payload). Executed both
548 * out-of-span and in-span. Out-of-span WRITEs may be
549 * filtered out by TERMs.
551 * TERM_* - Indicates meta-data was committed (if out-of-span) or
552 * will be rolled-back (in-span). Any out-of-span TERMs
553 * matching earlier WRITEs remove those WRITEs from
554 * consideration as they might conflict with a later data
555 * commit (which is not being rolled-back).
557 * SYNC - The earliest in-span SYNC (the last one when scanning
558 * backwards) tells the recovery code how far out-of-span
559 * it must go to run REDOs.
561 * NOTE: WRITEs do not always have matching TERMs even under
562 * perfect conditions because truncations might remove the
563 * buffers from consideration. I/O problems can also remove
564 * buffers from consideration.
566 * TRUNCSs do not always have matching TERMs because several
567 * truncations may be aggregated together into a single TERM.
569 struct hammer_fifo_redo {
570 struct hammer_fifo_head head;
571 int64_t redo_objid; /* file being written */
572 hammer_off_t redo_offset; /* logical offset in file */
573 int32_t redo_data_bytes;
575 uint32_t redo_localization;
576 uint32_t redo_reserved;
577 uint64_t redo_mtime; /* set mtime */
580 #define HAMMER_REDO_WRITE 0x00000001
581 #define HAMMER_REDO_TRUNC 0x00000002
582 #define HAMMER_REDO_TERM_WRITE 0x00000004
583 #define HAMMER_REDO_TERM_TRUNC 0x00000008
584 #define HAMMER_REDO_SYNC 0x00000010
586 union hammer_fifo_any {
587 struct hammer_fifo_head head;
588 struct hammer_fifo_undo undo;
589 struct hammer_fifo_redo redo;
592 typedef struct hammer_fifo_redo *hammer_fifo_redo_t;
593 typedef struct hammer_fifo_undo *hammer_fifo_undo_t;
594 typedef union hammer_fifo_any *hammer_fifo_any_t;
597 * Volume header types
599 #define HAMMER_FSBUF_VOLUME 0xC8414D4DC5523031ULL /* HAMMER01 */
600 #define HAMMER_FSBUF_VOLUME_REV 0x313052C54D4D41C8ULL /* (reverse endian) */
603 * HAMMER Volume header
605 * A HAMMER filesystem can be built from 1-256 block devices, each block
606 * device contains a volume header followed by however many buffers fit
609 * One of the volumes making up a HAMMER filesystem is the root volume.
610 * The root volume is always volume #0 which is the first block device path
611 * specified by newfs_hammer(8). All HAMMER volumes have a volume header,
612 * however the root volume may be the only volume that has valid values for
613 * some fields in the header.
615 * Special field notes:
617 * vol_bot_beg - offset of boot area (mem_beg - bot_beg bytes)
618 * vol_mem_beg - offset of memory log (buf_beg - mem_beg bytes)
619 * vol_buf_beg - offset of the first buffer in volume
620 * vol_buf_end - offset of volume EOF (on buffer boundary)
622 * The memory log area allows a kernel to cache new records and data
623 * in memory without allocating space in the actual filesystem to hold
624 * the records and data. In the event that a filesystem becomes full,
625 * any records remaining in memory can be flushed to the memory log
626 * area. This allows the kernel to immediately return success.
628 * The buffer offset is a physical offset of zone-2 offset. The lower
629 * 52 bits of the zone-2 offset is added to the buffer offset of each
630 * volume to generate an actual I/O offset within the block device.
632 * NOTE: boot area and memory log are currently not used.
636 * These macros are only used by userspace when userspace commands either
637 * initialize or add a new HAMMER volume.
639 #define HAMMER_BOOT_MINBYTES (32*1024)
640 #define HAMMER_BOOT_NOMBYTES (64LL*1024*1024)
641 #define HAMMER_BOOT_MAXBYTES (256LL*1024*1024)
643 #define HAMMER_MEM_MINBYTES (256*1024)
644 #define HAMMER_MEM_NOMBYTES (1LL*1024*1024*1024)
645 #define HAMMER_MEM_MAXBYTES (64LL*1024*1024*1024)
647 struct hammer_volume_ondisk {
648 uint64_t vol_signature; /* HAMMER_FSBUF_VOLUME for a valid header */
651 * These are relative to block device offset, not zone offsets.
653 int64_t vol_bot_beg; /* offset of boot area */
654 int64_t vol_mem_beg; /* offset of memory log */
655 int64_t vol_buf_beg; /* offset of the first buffer in volume */
656 int64_t vol_buf_end; /* offset of volume EOF (on buffer boundary) */
657 int64_t vol_reserved00;
659 uuid_t vol_fsid; /* identify filesystem */
660 uuid_t vol_fstype; /* identify filesystem type */
661 char vol_label[64]; /* filesystem label */
663 int32_t vol_no; /* volume number within filesystem */
664 int32_t vol_count; /* number of volumes making up filesystem */
666 uint32_t vol_version; /* version control information */
667 hammer_crc_t vol_crc; /* header crc */
668 uint32_t vol_flags; /* volume flags */
669 uint32_t vol_rootvol; /* the root volume number (must be 0) */
671 int32_t vol_reserved04;
672 int32_t vol_reserved05;
673 uint32_t vol_reserved06;
674 uint32_t vol_reserved07;
676 int64_t vol_reserved08;
677 int64_t vol_reserved09;
680 * These fields are initialized and space is reserved in every
681 * volume making up a HAMMER filesytem, but only the root volume
682 * contains valid data. Note that vol0_stat_bigblocks does not
683 * include big-blocks for freemap and undomap initially allocated
684 * by newfs_hammer(8).
686 int64_t vol0_stat_bigblocks; /* total big-blocks when fs is empty */
687 int64_t vol0_stat_freebigblocks;/* number of free big-blocks */
688 int64_t vol0_reserved11;
689 int64_t vol0_stat_inodes; /* for statfs only */
690 int64_t vol0_reserved10;
691 hammer_off_t vol0_btree_root; /* B-Tree root offset in zone-8 */
692 hammer_tid_t vol0_next_tid; /* highest partially synchronized TID */
693 hammer_off_t vol0_unused03;
696 * Blockmaps for zones. Not all zones use a blockmap. Note that
697 * the entire root blockmap is cached in the hammer_mount structure.
699 struct hammer_blockmap vol0_blockmap[HAMMER_MAX_ZONES];
702 * Array of zone-2 addresses for undo FIFO.
704 hammer_off_t vol0_undo_array[HAMMER_UNDO_LAYER2];
707 typedef struct hammer_volume_ondisk *hammer_volume_ondisk_t;
709 #define HAMMER_ROOT_VOLNO 0
711 #define HAMMER_VOLF_NEEDFLUSH 0x0004 /* volume needs flush */
713 #define HAMMER_VOL_CRCSIZE1 \
714 offsetof(struct hammer_volume_ondisk, vol_crc)
715 #define HAMMER_VOL_CRCSIZE2 \
716 (sizeof(struct hammer_volume_ondisk) - HAMMER_VOL_CRCSIZE1 - \
717 sizeof(hammer_crc_t))
719 #define HAMMER_VOL_VERSION_MIN 1 /* minimum supported version */
720 #define HAMMER_VOL_VERSION_DEFAULT 6 /* newfs default version */
721 #define HAMMER_VOL_VERSION_WIP 7 /* version >= this is WIP */
722 #define HAMMER_VOL_VERSION_MAX 6 /* maximum supported version */
724 #define HAMMER_VOL_VERSION_ONE 1
725 #define HAMMER_VOL_VERSION_TWO 2 /* new dirent layout (2.3+) */
726 #define HAMMER_VOL_VERSION_THREE 3 /* new snapshot layout (2.5+) */
727 #define HAMMER_VOL_VERSION_FOUR 4 /* new undo/flush (2.5+) */
728 #define HAMMER_VOL_VERSION_FIVE 5 /* dedup (2.9+) */
729 #define HAMMER_VOL_VERSION_SIX 6 /* DIRHASH_ALG1 */
732 * Translate a zone-2 address to physical address
734 #define hammer_xlate_to_phys(volume, zone2_offset) \
735 ((volume)->vol_buf_beg + \
736 ((zone2_offset) & HAMMER_OFF_SHORT_MASK))
739 * Record types are fairly straightforward. The B-Tree includes the record
740 * type in its index sort.
742 #define HAMMER_RECTYPE_UNKNOWN 0x0000
743 #define HAMMER_RECTYPE_LOWEST 0x0001 /* lowest record type avail */
744 #define HAMMER_RECTYPE_INODE 0x0001 /* inode in obj_id space */
745 #define HAMMER_RECTYPE_DATA 0x0010
746 #define HAMMER_RECTYPE_DIRENTRY 0x0011
747 #define HAMMER_RECTYPE_DB 0x0012
748 #define HAMMER_RECTYPE_EXT 0x0013 /* ext attributes */
749 #define HAMMER_RECTYPE_FIX 0x0014 /* fixed attribute */
750 #define HAMMER_RECTYPE_PFS 0x0015 /* PFS management */
751 #define HAMMER_RECTYPE_SNAPSHOT 0x0016 /* Snapshot management */
752 #define HAMMER_RECTYPE_CONFIG 0x0017 /* hammer cleanup config */
753 #define HAMMER_RECTYPE_MAX 0xFFFF
755 #define HAMMER_RECTYPE_ENTRY_START (HAMMER_RECTYPE_INODE + 1)
756 #define HAMMER_RECTYPE_CLEAN_START HAMMER_RECTYPE_EXT
758 #define HAMMER_FIXKEY_SYMLINK 1
760 #define HAMMER_OBJTYPE_UNKNOWN 0 /* never exists on-disk as unknown */
761 #define HAMMER_OBJTYPE_DIRECTORY 1
762 #define HAMMER_OBJTYPE_REGFILE 2
763 #define HAMMER_OBJTYPE_DBFILE 3
764 #define HAMMER_OBJTYPE_FIFO 4
765 #define HAMMER_OBJTYPE_CDEV 5
766 #define HAMMER_OBJTYPE_BDEV 6
767 #define HAMMER_OBJTYPE_SOFTLINK 7
768 #define HAMMER_OBJTYPE_PSEUDOFS 8 /* pseudo filesystem obj */
769 #define HAMMER_OBJTYPE_SOCKET 9
772 * HAMMER inode attribute data
774 * The data reference for a HAMMER inode points to this structure. Any
775 * modifications to the contents of this structure will result in a
776 * replacement operation.
778 * parent_obj_id is only valid for directories (which cannot be hard-linked),
779 * and specifies the parent directory obj_id. This field will also be set
780 * for non-directory inodes as a recovery aid, but can wind up holding
781 * stale information. However, since object id's are not reused, the worse
782 * that happens is that the recovery code is unable to use it.
783 * A parent_obj_id of 0 means it's a root inode of root or non-root PFS.
785 * NOTE: Future note on directory hardlinks. We can implement a record type
786 * which allows us to point to multiple parent directories.
788 struct hammer_inode_data {
789 uint16_t version; /* inode data version */
790 uint16_t mode; /* basic unix permissions */
791 uint32_t uflags; /* chflags */
792 uint32_t rmajor; /* used by device nodes */
793 uint32_t rminor; /* used by device nodes */
795 int64_t parent_obj_id; /* parent directory obj_id */
800 uint8_t cap_flags; /* capability support flags (extension) */
802 uint32_t reserved02; /* RESERVED FOR POSSIBLE FUTURE BIRTHTIME */
803 uint64_t nlinks; /* hard links */
804 uint64_t size; /* filesystem object size */
806 char symlink[24]; /* HAMMER_INODE_BASESYMLEN */
808 uint64_t mtime; /* mtime must be second-to-last */
809 uint64_t atime; /* atime must be last */
813 * Neither mtime nor atime upates are CRCd by the B-Tree element.
814 * mtime updates have UNDO, atime updates do not.
816 #define HAMMER_INODE_CRCSIZE \
817 offsetof(struct hammer_inode_data, mtime)
819 #define HAMMER_INODE_DATA_VERSION 1
820 #define HAMMER_OBJID_ROOT 1 /* root inodes # */
821 #define HAMMER_INODE_BASESYMLEN 24 /* see ext.symlink */
824 * Capability & implementation flags.
826 * HAMMER_INODE_CAP_DIR_LOCAL_INO - Use inode B-Tree localization
827 * for directory entries. Also see HAMMER_DIR_INODE_LOCALIZATION().
829 #define HAMMER_INODE_CAP_DIRHASH_MASK 0x03 /* directory: hash algorithm */
830 #define HAMMER_INODE_CAP_DIRHASH_ALG0 0x00
831 #define HAMMER_INODE_CAP_DIRHASH_ALG1 0x01
832 #define HAMMER_INODE_CAP_DIRHASH_ALG2 0x02
833 #define HAMMER_INODE_CAP_DIRHASH_ALG3 0x03
834 #define HAMMER_INODE_CAP_DIR_LOCAL_INO 0x04 /* use inode localization */
837 * A HAMMER directory entry associates a HAMMER filesystem object with a
838 * namespace. It is possible to hook into a pseudo-filesystem (with its
839 * own inode numbering space) in the filesystem by setting the high
840 * 16 bits of the localization field. The low 16 bits must be 0 and
841 * are reserved for future use.
843 * Directory entries are indexed with a 128 bit namekey rather then an
844 * offset. A portion of the namekey is an iterator/randomizer to deal
847 * NOTE: leaf.base.obj_type from the related B-Tree leaf entry holds
848 * the filesystem object type of obj_id, e.g. a den_type equivalent.
849 * It is not stored in hammer_direntry_data.
851 * NOTE: name field / the filename data reference is NOT terminated with \0.
853 struct hammer_direntry_data {
854 int64_t obj_id; /* object being referenced */
855 uint32_t localization; /* identify pseudo-filesystem */
857 char name[16]; /* name (extended) */
860 #define HAMMER_ENTRY_NAME_OFF offsetof(struct hammer_direntry_data, name[0])
861 #define HAMMER_ENTRY_SIZE(nlen) offsetof(struct hammer_direntry_data, name[nlen])
864 * Symlink data which does not fit in the inode is stored in a separate
867 struct hammer_symlink_data {
868 char name[16]; /* name (extended) */
871 #define HAMMER_SYMLINK_NAME_OFF offsetof(struct hammer_symlink_data, name[0])
874 * The root inode for the primary filesystem and root inode for any
875 * pseudo-fs may be tagged with an optional data structure using
876 * HAMMER_RECTYPE_PFS and localization id. This structure allows
877 * the node to be used as a mirroring master or slave.
879 * When operating as a slave CD's into the node automatically become read-only
880 * and as-of sync_end_tid.
882 * When operating as a master the read PFSD info sets sync_end_tid to
883 * the most recently flushed TID.
885 * sync_low_tid is not yet used but will represent the highest pruning
886 * end-point, after which full history is available.
888 * We need to pack this structure making it equally sized on both 32-bit and
889 * 64-bit machines as it is part of struct hammer_ioc_mrecord_pfs which is
890 * send over the wire in hammer mirror operations. Only on 64-bit machines
891 * the size of this struct differ when packed or not. This leads us to the
892 * situation where old 64-bit systems (using the non-packed structure),
893 * which were never able to mirror to/from 32-bit systems, are now no longer
894 * able to mirror to/from newer 64-bit systems (using the packed structure).
896 struct hammer_pseudofs_data {
897 hammer_tid_t sync_low_tid; /* full history beyond this point */
898 hammer_tid_t sync_beg_tid; /* earliest tid w/ full history avail */
899 hammer_tid_t sync_end_tid; /* current synchronizatoin point */
900 uint64_t sync_beg_ts; /* real-time of last completed sync */
901 uint64_t sync_end_ts; /* initiation of current sync cycle */
902 uuid_t shared_uuid; /* shared uuid (match required) */
903 uuid_t unique_uuid; /* unique uuid of this master/slave */
904 int32_t reserved01; /* reserved for future master_id */
905 int32_t mirror_flags; /* misc flags */
906 char label[64]; /* filesystem space label */
907 char snapshots[64]; /* softlink dir for pruning */
908 int32_t reserved02; /* was prune_{time,freq} */
909 int32_t reserved03; /* was reblock_{time,freq} */
910 int32_t reserved04; /* was snapshot_freq */
911 int32_t prune_min; /* do not prune recent history */
912 int32_t prune_max; /* do not retain history beyond here */
913 int32_t reserved[16];
916 typedef struct hammer_pseudofs_data *hammer_pseudofs_data_t;
918 #define HAMMER_PFSD_SLAVE 0x00000001
919 #define HAMMER_PFSD_DELETED 0x80000000
921 #define hammer_is_pfs_slave(pfsd) \
922 (((pfsd)->mirror_flags & HAMMER_PFSD_SLAVE) != 0)
923 #define hammer_is_pfs_master(pfsd) \
924 (!hammer_is_pfs_slave(pfsd))
925 #define hammer_is_pfs_deleted(pfsd) \
926 (((pfsd)->mirror_flags & HAMMER_PFSD_DELETED) != 0)
928 #define HAMMER_MAX_PFS 65536
929 #define HAMMER_MAX_PFSID (HAMMER_MAX_PFS - 1)
930 #define HAMMER_ROOT_PFSID 0
933 * Snapshot meta-data { Objid = HAMMER_OBJID_ROOT, Key = tid, rectype = SNAPSHOT }.
935 * Snapshot records replace the old <fs>/snapshots/<softlink> methodology. Snapshot
936 * records are mirrored but may be independantly managed once they are laid down on
939 * NOTE: The b-tree key is signed, the tid is not, so callers must still sort the
942 * NOTE: Reserved fields must be zero (as usual)
944 struct hammer_snapshot_data {
945 hammer_tid_t tid; /* the snapshot TID itself (== key) */
946 uint64_t ts; /* real-time when snapshot was made */
949 char label[64]; /* user-supplied description */
950 uint64_t reserved03[4];
954 * Config meta-data { ObjId = HAMMER_OBJID_ROOT, Key = 0, rectype = CONFIG }.
956 * Used to store the hammer cleanup config. This data is not mirrored.
958 struct hammer_config_data {
963 * Rollup various structures embedded as record data
965 union hammer_data_ondisk {
966 struct hammer_direntry_data entry;
967 struct hammer_inode_data inode;
968 struct hammer_symlink_data symlink;
969 struct hammer_pseudofs_data pfsd;
970 struct hammer_snapshot_data snap;
971 struct hammer_config_data config;
974 typedef union hammer_data_ondisk *hammer_data_ondisk_t;
977 * Ondisk layout of B-Tree related structures
979 #include "hammer_btree.h"
981 #define HAMMER_DIR_INODE_LOCALIZATION(ino_data) \
982 (((ino_data)->cap_flags & HAMMER_INODE_CAP_DIR_LOCAL_INO) ? \
983 HAMMER_LOCALIZE_INODE : \
984 HAMMER_LOCALIZE_MISC)
986 #endif /* !VFS_HAMMER_DISK_H_ */