xref: /dragonfly/sys/vfs/hammer/hammer_disk.h (revision b7367ef6)
1 /*
2  * Copyright (c) 2007 The DragonFly Project.  All rights reserved.
3  *
4  * This code is derived from software contributed to The DragonFly Project
5  * by Matthew Dillon <dillon@backplane.com>
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  *
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
16  *    distribution.
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.
20  *
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
32  * SUCH DAMAGE.
33  *
34  * $DragonFly: src/sys/vfs/hammer/hammer_disk.h,v 1.2 2007/10/16 18:16:42 dillon Exp $
35  */
36 
37 #ifndef _SYS_UUID_H_
38 #include <sys/uuid.h>
39 #endif
40 
41 /*
42  * The structures below represent the on-disk format for a HAMMER
43  * filesystem.  Note that all fields for on-disk structures are naturally
44  * aligned.  The host endian format is used - compatibility is possible
45  * if the implementation detects reversed endian and adjusts data accordingly.
46  *
47  * Most of HAMMER revolves around the concept of an object identifier.  An
48  * obj_id is a 64 bit quantity which uniquely identifies a filesystem object
49  * FOR THE ENTIRE LIFE OF THE FILESYSTEM.  This uniqueness allows backups
50  * and mirrors to retain varying amounts of filesystem history by removing
51  * any possibility of conflict through identifier reuse.
52  *
53  * A HAMMER filesystem may spam multiple volumes.
54  *
55  * A HAMMER filesystem uses a 16K filesystem buffer size.  All filesystem
56  * I/O is done in multiples of 16K.  Most buffer-sized headers such as those
57  * used by volumes, super-clusters, clusters, and basic filesystem buffers
58  * use fixed-sized A-lists which are heavily dependant on HAMMER_BUFSIZE.
59  */
60 #define HAMMER_BUFSIZE	16384
61 #define HAMMER_BUFMASK	(HAMMER_BUFSIZE - 1)
62 
63 /*
64  * Hammer transction ids are 64 bit unsigned integers and are usually
65  * synchronized with the time of day in nanoseconds.
66  */
67 typedef u_int64_t hammer_tid_t;
68 
69 /*
70  * Most HAMMER data structures are embedded in 16K filesystem buffers.
71  * All filesystem buffers except those designated as pure-data buffers
72  * contain this 128-byte header.
73  *
74  * This structure contains an embedded A-List used to manage space within
75  * the filesystem buffer.  It is not used by volume or cluster header
76  * buffers, or by pure-data buffers.  The granularity is variable and
77  * depends on the type of filesystem buffer.  BLKSIZE is just a minimum.
78  */
79 
80 #define HAMMER_FSBUF_HEAD_SIZE	128
81 #define HAMMER_FSBUF_MAXBLKS	256
82 #define HAMMER_FSBUF_BLKMASK	(HAMMER_FSBUF_MAXBLKS - 1)
83 #define HAMMER_FSBUF_METAELMS	HAMMER_ALIST_METAELMS_256_1LYR	/* 11 */
84 
85 struct hammer_fsbuf_head {
86 	u_int64_t buf_type;
87 	u_int32_t buf_crc;
88 	u_int32_t buf_reserved07;
89 	u_int32_t reserved[6];
90 	struct hammer_almeta buf_almeta[HAMMER_FSBUF_METAELMS];
91 };
92 
93 typedef struct hammer_fsbuf_head *hammer_fsbuf_head_t;
94 
95 /*
96  * Note: Pure-data buffers contain pure-data and have no buf_type.
97  * Piecemeal data buffers do have a header and use HAMMER_FSBUF_DATA.
98  */
99 #define HAMMER_FSBUF_VOLUME	0xC8414D4DC5523031ULL	/* HAMMER01 */
100 #define HAMMER_FSBUF_SUPERCL	0xC8414D52C3555052ULL	/* HAMRSUPR */
101 #define HAMMER_FSBUF_CLUSTER	0xC8414D52C34C5553ULL	/* HAMRCLUS */
102 #define HAMMER_FSBUF_RECORDS	0xC8414D52D2454353ULL	/* HAMRRECS */
103 #define HAMMER_FSBUF_BTREE	0xC8414D52C2545245ULL	/* HAMRBTRE */
104 #define HAMMER_FSBUF_DATA	0xC8414D52C4415441ULL	/* HAMRDATA */
105 
106 #define HAMMER_FSBUF_VOLUME_REV	0x313052C54D4D41C8ULL	/* (reverse endian) */
107 
108 /*
109  * The B-Tree structures need hammer_fsbuf_head.
110  */
111 #include "hammer_btree.h"
112 
113 /*
114  * HAMMER Volume header
115  *
116  * A HAMMER filesystem is built from any number of block devices,  Each block
117  * device contains a volume header followed by however many super-clusters
118  * and clusters fit into the volume.  Clusters cannot be migrated but the
119  * data they contain can, so HAMMER can use a truncated cluster for any
120  * extra space at the end of the volume.
121  *
122  * The volume containing the root cluster is designated as the master volume.
123  * The root cluster designation can be moved to any volume.
124  *
125  * The volume header takes up an entire 16K filesystem buffer and includes
126  * a one or two-layered A-list to manage the clusters making up the volume.
127  * A volume containing up to 32768 clusters (2TB) can be managed with a
128  * single-layered A-list.  A two-layer A-list is capable of managing up
129  * to 16384 super-clusters with each super-cluster containing 32768 clusters
130  * (32768 TB per volume total).  The number of volumes is limited to 32768
131  * but it only takes 512 to fill out a 64 bit address space so for all
132  * intents and purposes the filesystem has no limits.
133  *
134  * cluster addressing within a volume depends on whether a single or
135  * duel-layer A-list is used.  If a duel-layer A-list is used a 16K
136  * super-cluster buffer is needed for every 16384 clusters in the volume.
137  * However, because the A-list's hinting is grouped in multiples of 16
138  * we group 16 super-cluster buffers together (starting just after the
139  * volume header), followed by 16384x16 clusters, and repeat.
140  *
141  * NOTE: A 32768-element single-layer and 16384-element duel-layer A-list
142  * is the same size.
143  */
144 #define HAMMER_VOL_MAXCLUSTERS		32768	/* 1-layer */
145 #define HAMMER_VOL_MAXSUPERCLUSTERS	16384	/* 2-layer */
146 #define HAMMER_VOL_SUPERCLUSTER_GROUP	16
147 #define HAMMER_VOL_METAELMS_1LYR	HAMMER_ALIST_METAELMS_32K_1LYR
148 #define HAMMER_VOL_METAELMS_2LYR	HAMMER_ALIST_METAELMS_16K_2LYR
149 
150 struct hammer_volume_ondisk {
151 	struct hammer_fsbuf_head head;
152 	int64_t vol_beg;	/* byte offset of first cl/supercl in volume */
153 	int64_t vol_end;	/* byte offset of volume EOF */
154 	int64_t vol_locked;	/* reserved clusters are >= this offset */
155 
156 	uuid_t    vol_fsid;	/* identify filesystem */
157 	uuid_t    vol_fstype;	/* identify filesystem type */
158 	char	  vol_name[64];	/* Name of volume */
159 
160 	int32_t vol_no;		/* volume number within filesystem */
161 	int32_t vol_count;	/* number of volumes making up FS */
162 
163 	u_int32_t vol_version;	/* version control information */
164 	u_int32_t vol_reserved01;
165 	u_int32_t vol_flags;	/* volume flags */
166 	u_int32_t vol_rootvol;	/* which volume is the root volume? */
167 
168 	int32_t vol_clsize;	/* cluster size (same for all volumes) */
169 	int32_t vol_nclusters;
170 	u_int32_t vol_reserved06;
171 	u_int32_t vol_reserved07;
172 
173 	/*
174 	 * These fields are initialized and space is reserved in every
175 	 * volume making up a HAMMER filesytem, but only the master volume
176 	 * contains valid data.
177 	 */
178 	int32_t vol0_rootcluster;	/* root cluster no (index) in rootvol */
179 	u_int32_t vol0_reserved02;
180 	u_int32_t vol0_reserved03;
181 	hammer_tid_t vol0_nexttid;	/* next TID */
182 	u_int64_t vol0_recid;		/* fs-wide record id allocator */
183 
184 	char	reserved[1024];
185 
186 	/*
187 	 * Meta elements for the volume header's A-list, which is either a
188 	 * 1-layer A-list capable of managing 32768 clusters, or a 2-layer
189 	 * A-list capable of managing 16384 super-clusters (each of which
190 	 * can handle 32768 clusters).
191 	 */
192 	union {
193 		struct hammer_almeta	super[HAMMER_VOL_METAELMS_2LYR];
194 		struct hammer_almeta	normal[HAMMER_VOL_METAELMS_1LYR];
195 	} vol_almeta;
196 	u_int32_t	vol0_bitmap[1024];
197 };
198 
199 #define HAMMER_VOLF_VALID		0x0001	/* valid entry */
200 #define HAMMER_VOLF_OPEN		0x0002	/* volume is open */
201 #define HAMMER_VOLF_SUPERCL_ENABLE	0x0004	/* enable supercluster layer */
202 #define HAMMER_VOLF_SUPERCL_RESERVE	0x0008	/* supercluster layout */
203 
204 /*
205  * HAMMER Super-cluster header
206  *
207  * A super-cluster is used to increase the maximum size of a volume.
208  * HAMMER's volume header can manage up to 32768 direct clusters or
209  * 16384 super-clusters.  Each super-cluster (which is basically just
210  * a 16K filesystem buffer) can manage up to 32768 clusters.  So adding
211  * a super-cluster layer allows a HAMMER volume to be sized upwards of
212  * around 32768TB instead of 2TB.
213  *
214  * Any volume initially formatted to be over 32G reserves space for the layer
215  * but the layer is only enabled if the volume exceeds 2TB.
216  */
217 #define HAMMER_SUPERCL_METAELMS		HAMMER_ALIST_METAELMS_32K_1LYR
218 #define HAMMER_SCL_MAXCLUSTERS		HAMMER_VOL_MAXCLUSTERS
219 
220 struct hammer_supercl_ondisk {
221 	struct hammer_fsbuf_head head;
222 	uuid_t	vol_fsid;	/* identify filesystem - sanity check */
223 	uuid_t	vol_fstype;	/* identify filesystem type - sanity check */
224 	int32_t reserved[1024];
225 
226 	struct hammer_almeta	scl_meta[HAMMER_SUPERCL_METAELMS];
227 };
228 
229 /*
230  * HAMMER Cluster header
231  *
232  * A cluster is limited to 64MB and is made up of 4096 16K filesystem
233  * buffers.  The cluster header contains four A-lists to manage these
234  * buffers.
235  *
236  * master_alist - This is a non-layered A-list which manages pure-data
237  *		  allocations and allocations on behalf of other A-lists.
238  *
239  * btree_alist  - This is a layered A-list which manages filesystem buffers
240  *		  containing B-Tree nodes.
241  *
242  * record_alist - This is a layered A-list which manages filesystem buffers
243  *		  containing records.
244  *
245  * mdata_alist  - This is a layered A-list which manages filesystem buffers
246  *		  containing piecemeal record data.
247  *
248  * General storage management works like this:  All the A-lists except the
249  * master start in an all-allocated state.  Now lets say you wish to allocate
250  * a B-Tree node out the btree_alist.  If the allocation fails you allocate
251  * a pure data block out of master_alist and then free that  block in
252  * btree_alist, thereby assigning more space to the btree_alist, and then
253  * retry your allocation out of the btree_alist.  In the reverse direction,
254  * filesystem buffers can be garbage collected back to master_alist simply
255  * by doing whole-buffer allocations in btree_alist and then freeing the
256  * space in master_alist.  The whole-buffer-allocation approach to garbage
257  * collection works because A-list allocations are always power-of-2 sized
258  * and aligned.
259  */
260 #define HAMMER_CLU_MAXBUFFERS		4096
261 #define HAMMER_CLU_MASTER_METAELMS	HAMMER_ALIST_METAELMS_4K_1LYR
262 #define HAMMER_CLU_SLAVE_METAELMS	HAMMER_ALIST_METAELMS_4K_2LYR
263 #define HAMMER_CLU_MAXBYTES		(HAMMER_CLU_MAXBUFFERS * HAMMER_BUFSIZE)
264 
265 struct hammer_cluster_ondisk {
266 	struct hammer_fsbuf_head head;
267 	uuid_t	vol_fsid;	/* identify filesystem - sanity check */
268 	uuid_t	vol_fstype;	/* identify filesystem type - sanity check */
269 
270 	u_int64_t clu_gen;	/* identify generation number of cluster */
271 	u_int64_t clu_unused01;
272 
273 	hammer_tid_t clu_id;	/* unique cluster self identification */
274 	int32_t vol_no;		/* cluster contained in volume (sanity) */
275 	u_int32_t clu_flags;	/* cluster flags */
276 
277 	int32_t clu_start;	/* start of data (byte offset) */
278 	int32_t clu_limit;	/* end of data (byte offset) */
279 	int32_t clu_no;		/* cluster index in volume (sanity) */
280 	u_int32_t clu_reserved03;
281 
282 	u_int32_t clu_reserved04;
283 	u_int32_t clu_reserved05;
284 	u_int32_t clu_reserved06;
285 	u_int32_t clu_reserved07;
286 
287 	int32_t idx_data;	/* data append point (element no) */
288 	int32_t idx_index;	/* index append point (element no) */
289 	int32_t idx_record;	/* record prepend point (element no) */
290 	u_int32_t idx_reserved03;
291 
292 	/*
293 	 * Specify the range of information stored in this cluster as two
294 	 * btree elements.  These elements exist as separate records that
295 	 * point to us in the parent cluster's B-Tree.
296 	 *
297 	 * Note that clu_btree_end is range-inclusive, not range-exclusive.
298 	 * i.e. 0-1023 instead of 0,1024.
299 	 */
300 	struct hammer_base_elm clu_btree_beg;
301 	struct hammer_base_elm clu_btree_end;
302 
303 	/*
304 	 * The cluster's B-Tree root can change as a side effect of insertion
305 	 * and deletion operations so store an offset instead of embedding
306 	 * the root node.
307 	 */
308 	int32_t		clu_btree_root;
309 	int32_t		clu_btree_parent_vol_no;
310 	int32_t		clu_btree_parent_clu_no;
311 	hammer_tid_t	clu_btree_parent_clu_id;
312 
313 	u_int64_t synchronized_rec_id;
314 
315 	struct hammer_almeta	clu_master_meta[HAMMER_CLU_MASTER_METAELMS];
316 	struct hammer_almeta	clu_btree_meta[HAMMER_CLU_SLAVE_METAELMS];
317 	struct hammer_almeta	clu_record_meta[HAMMER_CLU_SLAVE_METAELMS];
318 	struct hammer_almeta	clu_mdata_meta[HAMMER_CLU_SLAVE_METAELMS];
319 };
320 
321 /*
322  * HAMMER records are 96 byte entities encoded into 16K filesystem buffers.
323  * Each record has a 64 byte header and a 32 byte extension.  170 records
324  * fit into each buffer.  Storage is managed by the buffer's A-List.
325  *
326  * Each record may have an explicit data reference to a block of data up
327  * to 2^31-1 bytes in size within the current cluster.  Note that multiple
328  * records may share the same or overlapping data references.
329  */
330 
331 /*
332  * All HAMMER records have a common 64-byte base and a 32-byte extension.
333  *
334  * Many HAMMER record types reference out-of-band data within the cluster.
335  * This data can also be stored in-band in the record itself if it is small
336  * enough.  Either way, (data_offset, data_len) points to it.
337  *
338  * Key comparison order:  obj_id, rec_type, key, create_tid
339  */
340 struct hammer_base_record {
341 	int64_t	obj_id;		/* 00 object record is associated with */
342 	int64_t key;		/* 08 indexing key (offset or namekey) */
343 
344 	hammer_tid_t create_tid;/* 10 transaction id for record creation */
345 	hammer_tid_t delete_tid;/* 18 transaction id for record update/delete */
346 
347 	u_int16_t rec_type;	/* 20 type of record */
348 	u_int16_t obj_type;	/* 22 type of object (if inode) */
349 	u_int32_t data_offset;	/* 24 intra-cluster data reference */
350 				/*    An offset of 0 indicates zero-fill */
351 	int32_t data_len;	/* 28 size of data (remainder zero-fill) */
352 	u_int32_t data_crc;	/* 2C data sanity check */
353 	u_int64_t rec_id;	/* 30 record id (iterator for recovery) */
354 	u_int64_t reserved07;	/* 38 */
355 				/* 40 */
356 };
357 
358 /*
359  * Record types are fairly straightforward.  The B-Tree includes the record
360  * type in its index sort.
361  *
362  * In particular please note that it is possible to create a pseudo-
363  * filesystem within a HAMMER filesystem by creating a special object
364  * type within a directory.  Pseudo-filesystems are used as replication
365  * targets and even though they are built within a HAMMER filesystem they
366  * get their own obj_id space (and thus can serve as a replication target)
367  * and look like a mount point to the system.
368  */
369 #define HAMMER_RECTYPE_UNKNOWN		0
370 #define HAMMER_RECTYPE_INODE		1	/* inode in obj_id space */
371 #define HAMMER_RECTYPE_PSEUDO_INODE	2	/* pseudo filesysem */
372 #define HAMMER_RECTYPE_DATA_CREATE	0x10
373 #define HAMMER_RECTYPE_DATA_ZEROFILL	0x11
374 #define HAMMER_RECTYPE_DATA_DELETE	0x12
375 #define HAMMER_RECTYPE_DATA_UPDATE	0x13
376 #define HAMMER_RECTYPE_DIR_CREATE	0x20
377 #define HAMMER_RECTYPE_DIR_DELETE	0x22
378 #define HAMMER_RECTYPE_DIR_UPDATE	0x23
379 #define HAMMER_RECTYPE_DB_CREATE	0x30
380 #define HAMMER_RECTYPE_DB_DELETE	0x32
381 #define HAMMER_RECTYPE_DB_UPDATE	0x33
382 #define HAMMER_RECTYPE_EXT_CREATE	0x40	/* ext attributes */
383 #define HAMMER_RECTYPE_EXT_DELETE	0x42
384 #define HAMMER_RECTYPE_EXT_UPDATE	0x43
385 
386 #define HAMMER_OBJTYPE_DIRECTORY	1
387 #define HAMMER_OBJTYPE_REGFILE		2
388 #define HAMMER_OBJTYPE_DBFILE		3
389 #define HAMMER_OBJTYPE_FIFO		4
390 #define HAMMER_OBJTYPE_CDEV		5
391 #define HAMMER_OBJTYPE_BDEV		6
392 #define HAMMER_OBJTYPE_SOFTLINK		7
393 #define HAMMER_OBJTYPE_PSEUDOFS		8	/* pseudo filesystem obj */
394 
395 #define HAMMER_OBJTYPE_CLUSTER_BEG	0x10
396 #define HAMMER_OBJTYPE_CLUSTER_END	0x11
397 
398 /*
399  * Generic full-sized record
400  */
401 struct hammer_generic_record {
402 	struct hammer_base_record base;
403 	char filler[32];
404 };
405 
406 /*
407  * A HAMMER inode record.
408  *
409  * This forms the basis for a filesystem object.  obj_id is the inode number,
410  * key1 represents the pseudo filesystem id for security partitioning
411  * (preventing cross-links and/or restricting a NFS export and specifying the
412  * security policy), and key2 represents the data retention policy id.
413  *
414  * Inode numbers are 64 bit quantities which uniquely identify a filesystem
415  * object for the ENTIRE life of the filesystem, even after the object has
416  * been deleted.  For all intents and purposes inode numbers are simply
417  * allocated by incrementing a sequence space.
418  *
419  * There is an important distinction between the data stored in the inode
420  * record and the record's data reference.  The record references a
421  * hammer_inode_data structure but the filesystem object size and hard link
422  * count is stored in the inode record itself.  This allows multiple inodes
423  * to share the same hammer_inode_data structure.  This is possible because
424  * any modifications will lay out new data.  The HAMMER implementation need
425  * not use the data-sharing ability when laying down new records.
426  *
427  * A HAMMER inode is subject to the same historical storage requirements
428  * as any other record.  In particular any change in filesystem or hard link
429  * count will lay down a new inode record when the filesystem is synced to
430  * disk.  This can lead to a lot of junk records which get cleaned up by
431  * the data retention policy.
432  *
433  * The ino_atime and ino_mtime fields are a special case.  Modifications to
434  * these fields do NOT lay down a new record by default, though the values
435  * are effectively frozen for snapshots which access historical versions
436  * of the inode record due to other operations.  This means that atime will
437  * not necessarily be accurate in snapshots, backups, or mirrors.  mtime
438  * will be accurate in backups and mirrors since it can be regenerated from
439  * the mirroring stream.
440  *
441  * Because nlinks is historically retained the hardlink count will be
442  * accurate when accessing a HAMMER filesystem snapshot.
443  */
444 struct hammer_inode_record {
445 	struct hammer_base_record base;
446 	u_int64_t ino_atime;	/* last access time (not historical) */
447 	u_int64_t ino_mtime;	/* last modified time (not historical) */
448 	u_int64_t ino_size;	/* filesystem object size */
449 	u_int64_t ino_nlinks;	/* hard links */
450 };
451 
452 /*
453  * Data records specify the entire contents of a regular file object,
454  * including attributes.  Small amounts of data can theoretically be
455  * embedded in the record itself but the use of this ability verses using
456  * an out-of-band data reference depends on the implementation.
457  */
458 struct hammer_data_record {
459 	struct hammer_base_record base;
460 	char filler[32];
461 };
462 
463 /*
464  * A directory entry specifies the HAMMER filesystem object id, a copy of
465  * the file type, and file name (either embedded or as out-of-band data).
466  * If the file name is short enough to fit into den_name[] (including a
467  * terminating nul) then it will be embedded in the record, otherwise it
468  * is stored out-of-band.  The base record's data reference always points
469  * to the nul-terminated filename regardless.
470  *
471  * Directory entries are indexed with a 128 bit namekey rather then an
472  * offset.  A portion of the namekey is an iterator or randomizer to deal
473  * with collisions.
474  */
475 struct hammer_entry_record {
476 	struct hammer_base_record base;
477 	u_int64_t obj_id;		/* object being referenced */
478 	u_int64_t reserved01;
479 	u_int8_t  den_type;		/* cached file type */
480 	char	  den_name[15];		/* short file names fit in record */
481 };
482 
483 /*
484  * Hammer rollup record
485  */
486 union hammer_record_ondisk {
487 	struct hammer_base_record	base;
488 	struct hammer_generic_record	generic;
489 	struct hammer_inode_record	inode;
490 	struct hammer_data_record	data;
491 	struct hammer_entry_record	entry;
492 };
493 
494 typedef union hammer_record_ondisk *hammer_record_ondisk_t;
495 
496 /*
497  * Filesystem buffer for records
498  */
499 #define HAMMER_RECORD_NODES	\
500 	((HAMMER_BUFSIZE - sizeof(struct hammer_fsbuf_head)) / \
501 	sizeof(union hammer_record_ondisk))
502 
503 struct hammer_fsbuf_recs {
504 	struct hammer_fsbuf_head	head;
505 	char				unused[32];
506 	union hammer_record_ondisk	recs[HAMMER_RECORD_NODES];
507 };
508 
509 /*
510  * Filesystem buffer for piecemeal data.  Note that this does not apply
511  * to dedicated pure-data buffers as such buffers do not have a header.
512  */
513 
514 #define HAMMER_DATA_SIZE	(HAMMER_BUFSIZE - sizeof(struct hammer_fsbuf_head))
515 #define HAMMER_DATA_BLKSIZE	64
516 #define HAMMER_DATA_BLKMASK	(HAMMER_DATA_BLKSIZE-1)
517 #define HAMMER_DATA_NODES	(HAMMER_DATA_SIZE / HAMMER_DATA_BLKSIZE)
518 
519 struct hammer_fsbuf_data {
520 	struct hammer_fsbuf_head head;
521 	u_int8_t		data[HAMMER_DATA_NODES][HAMMER_DATA_BLKSIZE];
522 };
523 
524 /*
525  * Filesystem buffer rollup
526  */
527 union hammer_fsbuf_ondisk {
528 	struct hammer_fsbuf_head	head;
529 	struct hammer_fsbuf_btree	btree;
530 	struct hammer_fsbuf_recs	record;
531 	struct hammer_fsbuf_data	data;
532 };
533 
534 typedef union hammer_fsbuf_ondisk *hammer_fsbuf_ondisk_t;
535 
536 /*
537  * HAMMER UNIX Attribute data
538  *
539  * The data reference in a HAMMER inode record points to this structure.  Any
540  * modifications to the contents of this structure will result in a record
541  * replacement operation.
542  *
543  * state_sum allows a filesystem object to be validated to a degree by
544  * generating a checksum of all of its pieces (in no particular order) and
545  * checking it against this field.
546  */
547 struct hammer_inode_data {
548 	u_int16_t version;	/* inode data version */
549 	u_int16_t mode;		/* basic unix permissions */
550 	u_int32_t uflags;	/* chflags */
551 	u_int64_t reserved01;
552 	u_int64_t reserved02;
553 	u_int64_t state_sum;	/* cumulative checksum */
554 	uuid_t	uid;
555 	uuid_t	gid;
556 };
557 
558 #define HAMMER_INODE_DATA_VERSION	1
559 
560 /*
561  * Rollup various structures embedded as record data
562  */
563 union hammer_data {
564 	struct hammer_inode_data inode;
565 };
566 
567