xref: /netbsd-src/usr.sbin/fstyp/hammer2_disk.h (revision cef8759bd76c1b621f8eab8faa6f208faabc2e15)
1 /*        $NetBSD: hammer2_disk.h,v 1.1 2020/01/01 08:56:41 tkusumi Exp $      */
2 
3 /*
4  * Copyright (c) 2011-2019 The DragonFly Project.  All rights reserved.
5  *
6  * This code is derived from software contributed to The DragonFly Project
7  * by Matthew Dillon <dillon@dragonflybsd.org>
8  * by Venkatesh Srinivas <vsrinivas@dragonflybsd.org>
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  *
14  * 1. Redistributions of source code must retain the above copyright
15  *    notice, this list of conditions and the following disclaimer.
16  * 2. Redistributions in binary form must reproduce the above copyright
17  *    notice, this list of conditions and the following disclaimer in
18  *    the documentation and/or other materials provided with the
19  *    distribution.
20  * 3. Neither the name of The DragonFly Project nor the names of its
21  *    contributors may be used to endorse or promote products derived
22  *    from this software without specific, prior written permission.
23  *
24  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
25  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
26  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
27  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
28  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
29  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
30  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
31  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
32  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
33  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
34  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
35  * SUCH DAMAGE.
36  */
37 #include <sys/cdefs.h>
38 __KERNEL_RCSID(0, "$NetBSD: hammer2_disk.h,v 1.1 2020/01/01 08:56:41 tkusumi Exp $");
39 
40 #ifndef _VFS_HAMMER2_DISK_H_
41 #define _VFS_HAMMER2_DISK_H_
42 
43 #ifndef _SYS_UUID_H_
44 #include <sys/uuid.h>
45 #endif
46 #if 0
47 #ifndef _SYS_DMSG_H_
48 #include <sys/dmsg.h>
49 #endif
50 #endif
51 
52 /*
53  * The structures below represent the on-disk media structures for the HAMMER2
54  * filesystem.  Note that all fields for on-disk structures are naturally
55  * aligned.  The host endian format is typically used - compatibility is
56  * possible if the implementation detects reversed endian and adjusts accesses
57  * accordingly.
58  *
59  * HAMMER2 primarily revolves around the directory topology:  inodes,
60  * directory entries, and block tables.  Block device buffer cache buffers
61  * are always 64KB.  Logical file buffers are typically 16KB.  All data
62  * references utilize 64-bit byte offsets.
63  *
64  * Free block management is handled independently using blocks reserved by
65  * the media topology.
66  */
67 
68 /*
69  * The data at the end of a file or directory may be a fragment in order
70  * to optimize storage efficiency.  The minimum fragment size is 1KB.
71  * Since allocations are in powers of 2 fragments must also be sized in
72  * powers of 2 (1024, 2048, ... 65536).
73  *
74  * For the moment the maximum allocation size is HAMMER2_PBUFSIZE (64K),
75  * which is 2^16.  Larger extents may be supported in the future.  Smaller
76  * fragments might be supported in the future (down to 64 bytes is possible),
77  * but probably will not be.
78  *
79  * A full indirect block use supports 512 x 128-byte blockrefs in a 64KB
80  * buffer.  Indirect blocks down to 1KB are supported to keep small
81  * directories small.
82  *
83  * A maximally sized file (2^64-1 bytes) requires ~6 indirect block levels
84  * using 64KB indirect blocks (128 byte refs, 512 or radix 9 per indblk).
85  *
86  *	16(datablk) + 9 + 9 + 9 + 9 + 9 + 9 = ~70.
87  *	16(datablk) + 7 + 9 + 9 + 9 + 9 + 9 = ~68.  (smaller top level indblk)
88  *
89  * The actual depth depends on copies redundancy and whether the filesystem
90  * has chosen to use a smaller indirect block size at the top level or not.
91  */
92 #define HAMMER2_ALLOC_MIN	1024	/* minimum allocation size */
93 #define HAMMER2_RADIX_MIN	10	/* minimum allocation size 2^N */
94 #define HAMMER2_ALLOC_MAX	65536	/* maximum allocation size */
95 #define HAMMER2_RADIX_MAX	16	/* maximum allocation size 2^N */
96 #define HAMMER2_RADIX_KEY	64	/* number of bits in key */
97 
98 /*
99  * MINALLOCSIZE		- The minimum allocation size.  This can be smaller
100  *		  	  or larger than the minimum physical IO size.
101  *
102  *			  NOTE: Should not be larger than 1K since inodes
103  *				are 1K.
104  *
105  * MINIOSIZE		- The minimum IO size.  This must be less than
106  *			  or equal to HAMMER2_LBUFSIZE.
107  *
108  * HAMMER2_LBUFSIZE	- Nominal buffer size for I/O rollups.
109  *
110  * HAMMER2_PBUFSIZE	- Topological block size used by files for all
111  *			  blocks except the block straddling EOF.
112  *
113  * HAMMER2_SEGSIZE	- Allocation map segment size, typically 4MB
114  *			  (space represented by a level0 bitmap).
115  */
116 
117 #define HAMMER2_SEGSIZE		(1 << HAMMER2_FREEMAP_LEVEL0_RADIX)
118 #define HAMMER2_SEGRADIX	HAMMER2_FREEMAP_LEVEL0_RADIX
119 
120 #define HAMMER2_PBUFRADIX	16	/* physical buf (1<<16) bytes */
121 #define HAMMER2_PBUFSIZE	65536
122 #define HAMMER2_LBUFRADIX	14	/* logical buf (1<<14) bytes */
123 #define HAMMER2_LBUFSIZE	16384
124 
125 /*
126  * Generally speaking we want to use 16K and 64K I/Os
127  */
128 #define HAMMER2_MINIORADIX	HAMMER2_LBUFRADIX
129 #define HAMMER2_MINIOSIZE	HAMMER2_LBUFSIZE
130 
131 #define HAMMER2_IND_BYTES_MIN	4096
132 #define HAMMER2_IND_BYTES_NOM	HAMMER2_LBUFSIZE
133 #define HAMMER2_IND_BYTES_MAX	HAMMER2_PBUFSIZE
134 #define HAMMER2_IND_RADIX_MIN	12
135 #define HAMMER2_IND_RADIX_NOM	HAMMER2_LBUFRADIX
136 #define HAMMER2_IND_RADIX_MAX	HAMMER2_PBUFRADIX
137 #define HAMMER2_IND_COUNT_MIN	(HAMMER2_IND_BYTES_MIN / \
138 				 sizeof(hammer2_blockref_t))
139 #define HAMMER2_IND_COUNT_MAX	(HAMMER2_IND_BYTES_MAX / \
140 				 sizeof(hammer2_blockref_t))
141 
142 /*
143  * In HAMMER2, arrays of blockrefs are fully set-associative, meaning that
144  * any element can occur at any index and holes can be anywhere.  As a
145  * future optimization we will be able to flag that such arrays are sorted
146  * and thus optimize lookups, but for now we don't.
147  *
148  * Inodes embed either 512 bytes of direct data or an array of 4 blockrefs,
149  * resulting in highly efficient storage for files <= 512 bytes and for files
150  * <= 512KB.  Up to 4 directory entries can be referenced from a directory
151  * without requiring an indirect block.
152  *
153  * Indirect blocks are typically either 4KB (64 blockrefs / ~4MB represented),
154  * or 64KB (1024 blockrefs / ~64MB represented).
155  */
156 #define HAMMER2_SET_RADIX		2	/* radix 2 = 4 entries */
157 #define HAMMER2_SET_COUNT		(1 << HAMMER2_SET_RADIX)
158 #define HAMMER2_EMBEDDED_BYTES		512	/* inode blockset/dd size */
159 #define HAMMER2_EMBEDDED_RADIX		9
160 
161 #define HAMMER2_PBUFMASK	(HAMMER2_PBUFSIZE - 1)
162 #define HAMMER2_LBUFMASK	(HAMMER2_LBUFSIZE - 1)
163 #define HAMMER2_SEGMASK		(HAMMER2_SEGSIZE - 1)
164 
165 #define HAMMER2_LBUFMASK64	((hammer2_off_t)HAMMER2_LBUFMASK)
166 #define HAMMER2_PBUFSIZE64	((hammer2_off_t)HAMMER2_PBUFSIZE)
167 #define HAMMER2_PBUFMASK64	((hammer2_off_t)HAMMER2_PBUFMASK)
168 #define HAMMER2_SEGSIZE64	((hammer2_off_t)HAMMER2_SEGSIZE)
169 #define HAMMER2_SEGMASK64	((hammer2_off_t)HAMMER2_SEGMASK)
170 
171 #define HAMMER2_UUID_STRING	"5cbb9ad1-862d-11dc-a94d-01301bb8a9f5"
172 
173 /*
174  * A 4MB segment is reserved at the beginning of each 2GB zone.  This segment
175  * contains the volume header (or backup volume header), the free block
176  * table, and possibly other information in the future.  A 4MB segment for
177  * freemap is reserved at the beginning of every 1GB.
178  *
179  * 4MB = 64 x 64K blocks.  Each 4MB segment is broken down as follows:
180  *
181  * ==========
182  *  0 volume header (for the first four 2GB zones)
183  *  1 freemap00 level1 FREEMAP_LEAF (256 x 128B bitmap data per 1GB)
184  *  2           level2 FREEMAP_NODE (256 x 128B indirect block per 256GB)
185  *  3           level3 FREEMAP_NODE (256 x 128B indirect block per 64TB)
186  *  4           level4 FREEMAP_NODE (256 x 128B indirect block per 16PB)
187  *  5           level5 FREEMAP_NODE (256 x 128B indirect block per 4EB)
188  *  6 freemap01 level1 (rotation)
189  *  7           level2
190  *  8           level3
191  *  9           level4
192  * 10           level5
193  * 11 freemap02 level1 (rotation)
194  * 12           level2
195  * 13           level3
196  * 14           level4
197  * 15           level5
198  * 16 freemap03 level1 (rotation)
199  * 17           level2
200  * 18           level3
201  * 19           level4
202  * 20           level5
203  * 21 freemap04 level1 (rotation)
204  * 22           level2
205  * 23           level3
206  * 24           level4
207  * 25           level5
208  * 26 freemap05 level1 (rotation)
209  * 27           level2
210  * 28           level3
211  * 29           level4
212  * 30           level5
213  * 31 freemap06 level1 (rotation)
214  * 32           level2
215  * 33           level3
216  * 34           level4
217  * 35           level5
218  * 36 freemap07 level1 (rotation)
219  * 37           level2
220  * 38           level3
221  * 39           level4
222  * 40           level5
223  * 41 unused
224  * .. unused
225  * 63 unused
226  * ==========
227  *
228  * The first four 2GB zones contain volume headers and volume header backups.
229  * After that the volume header block# is reserved for future use.  Similarly,
230  * there are many blocks related to various Freemap levels which are not
231  * used in every segment and those are also reserved for future use.
232  * Note that each FREEMAP_LEAF or FREEMAP_NODE uses 32KB out of 64KB slot.
233  *
234  *			Freemap (see the FREEMAP document)
235  *
236  * The freemap utilizes blocks #1-40 in 8 sets of 5 blocks.  Each block in
237  * a set represents a level of depth in the freemap topology.  Eight sets
238  * exist to prevent live updates from disturbing the state of the freemap
239  * were a crash/reboot to occur.  That is, a live update is not committed
240  * until the update's flush reaches the volume root.  There are FOUR volume
241  * roots representing the last four synchronization points, so the freemap
242  * must be consistent no matter which volume root is chosen by the mount
243  * code.
244  *
245  * Each freemap set is 5 x 64K blocks and represents the 1GB, 256GB, 64TB,
246  * 16PB and 4EB indirect map.  The volume header itself has a set of 4 freemap
247  * blockrefs representing another 2 bits, giving us a total 64 bits of
248  * representable address space.
249  *
250  * The Level 0 64KB block represents 1GB of storage represented by 32KB
251  * (256 x struct hammer2_bmap_data).  Each structure represents 4MB of storage
252  * and has a 512 bit bitmap, using 2 bits to represent a 16KB chunk of
253  * storage.  These 2 bits represent the following states:
254  *
255  *	00	Free
256  *	01	(reserved) (Possibly partially allocated)
257  *	10	Possibly free
258  *	11	Allocated
259  *
260  * One important thing to note here is that the freemap resolution is 16KB,
261  * but the minimum storage allocation size is 1KB.  The hammer2 vfs keeps
262  * track of sub-allocations in memory, which means that on a unmount or reboot
263  * the entire 16KB of a partially allocated block will be considered fully
264  * allocated.  It is possible for fragmentation to build up over time, but
265  * defragmentation is fairly easy to accomplish since all modifications
266  * allocate a new block.
267  *
268  * The Second thing to note is that due to the way snapshots and inode
269  * replication works, deleting a file cannot immediately free the related
270  * space.  Furthermore, deletions often do not bother to traverse the
271  * block subhierarchy being deleted.  And to go even further, whole
272  * sub-directory trees can be deleted simply by deleting the directory inode
273  * at the top.  So even though we have a symbol to represent a 'possibly free'
274  * block (binary 10), only the bulk free scanning code can actually use it.
275  * Normal 'rm's or other deletions do not.
276  *
277  * WARNING!  ZONE_SEG and VOLUME_ALIGN must be a multiple of 1<<LEVEL0_RADIX
278  *	     (i.e. a multiple of 4MB).  VOLUME_ALIGN must be >= ZONE_SEG.
279  *
280  * In Summary:
281  *
282  * (1) Modifications to freemap blocks 'allocate' a new copy (aka use a block
283  *     from the next set).  The new copy is reused until a flush occurs at
284  *     which point the next modification will then rotate to the next set.
285  */
286 #define HAMMER2_VOLUME_ALIGN		(8 * 1024 * 1024)
287 #define HAMMER2_VOLUME_ALIGN64		((hammer2_off_t)HAMMER2_VOLUME_ALIGN)
288 #define HAMMER2_VOLUME_ALIGNMASK	(HAMMER2_VOLUME_ALIGN - 1)
289 #define HAMMER2_VOLUME_ALIGNMASK64     ((hammer2_off_t)HAMMER2_VOLUME_ALIGNMASK)
290 
291 #define HAMMER2_NEWFS_ALIGN		(HAMMER2_VOLUME_ALIGN)
292 #define HAMMER2_NEWFS_ALIGN64		((hammer2_off_t)HAMMER2_VOLUME_ALIGN)
293 #define HAMMER2_NEWFS_ALIGNMASK		(HAMMER2_VOLUME_ALIGN - 1)
294 #define HAMMER2_NEWFS_ALIGNMASK64	((hammer2_off_t)HAMMER2_NEWFS_ALIGNMASK)
295 
296 #define HAMMER2_ZONE_BYTES64		(2LLU * 1024 * 1024 * 1024)
297 #define HAMMER2_ZONE_MASK64		(HAMMER2_ZONE_BYTES64 - 1)
298 #define HAMMER2_ZONE_SEG		(4 * 1024 * 1024)
299 #define HAMMER2_ZONE_SEG64		((hammer2_off_t)HAMMER2_ZONE_SEG)
300 #define HAMMER2_ZONE_BLOCKS_SEG		(HAMMER2_ZONE_SEG / HAMMER2_PBUFSIZE)
301 
302 #define HAMMER2_ZONE_FREEMAP_INC	5	/* 5 deep */
303 
304 #define HAMMER2_ZONE_VOLHDR		0	/* volume header or backup */
305 #define HAMMER2_ZONE_FREEMAP_00		1	/* normal freemap rotation */
306 #define HAMMER2_ZONE_FREEMAP_01		6	/* normal freemap rotation */
307 #define HAMMER2_ZONE_FREEMAP_02		11	/* normal freemap rotation */
308 #define HAMMER2_ZONE_FREEMAP_03		16	/* normal freemap rotation */
309 #define HAMMER2_ZONE_FREEMAP_04		21	/* normal freemap rotation */
310 #define HAMMER2_ZONE_FREEMAP_05		26	/* normal freemap rotation */
311 #define HAMMER2_ZONE_FREEMAP_06		31	/* normal freemap rotation */
312 #define HAMMER2_ZONE_FREEMAP_07		36	/* normal freemap rotation */
313 #define HAMMER2_ZONE_FREEMAP_END	41	/* (non-inclusive) */
314 
315 #define HAMMER2_ZONE_UNUSED41		41
316 #define HAMMER2_ZONE_UNUSED42		42
317 #define HAMMER2_ZONE_UNUSED43		43
318 #define HAMMER2_ZONE_UNUSED44		44
319 #define HAMMER2_ZONE_UNUSED45		45
320 #define HAMMER2_ZONE_UNUSED46		46
321 #define HAMMER2_ZONE_UNUSED47		47
322 #define HAMMER2_ZONE_UNUSED48		48
323 #define HAMMER2_ZONE_UNUSED49		49
324 #define HAMMER2_ZONE_UNUSED50		50
325 #define HAMMER2_ZONE_UNUSED51		51
326 #define HAMMER2_ZONE_UNUSED52		52
327 #define HAMMER2_ZONE_UNUSED53		53
328 #define HAMMER2_ZONE_UNUSED54		54
329 #define HAMMER2_ZONE_UNUSED55		55
330 #define HAMMER2_ZONE_UNUSED56		56
331 #define HAMMER2_ZONE_UNUSED57		57
332 #define HAMMER2_ZONE_UNUSED58		58
333 #define HAMMER2_ZONE_UNUSED59		59
334 #define HAMMER2_ZONE_UNUSED60		60
335 #define HAMMER2_ZONE_UNUSED61		61
336 #define HAMMER2_ZONE_UNUSED62		62
337 #define HAMMER2_ZONE_UNUSED63		63
338 #define HAMMER2_ZONE_END		64	/* non-inclusive */
339 
340 #define HAMMER2_NFREEMAPS		8	/* FREEMAP_00 - FREEMAP_07 */
341 
342 						/* relative to FREEMAP_x */
343 #define HAMMER2_ZONEFM_LEVEL1		0	/* 1GB leafmap */
344 #define HAMMER2_ZONEFM_LEVEL2		1	/* 256GB indmap */
345 #define HAMMER2_ZONEFM_LEVEL3		2	/* 64TB indmap */
346 #define HAMMER2_ZONEFM_LEVEL4		3	/* 16PB indmap */
347 #define HAMMER2_ZONEFM_LEVEL5		4	/* 4EB indmap */
348 /* LEVEL6 is a set of 4 blockrefs in the volume header 16EB */
349 
350 /*
351  * Freemap radix.  Assumes a set-count of 4, 128-byte blockrefs,
352  * 32KB indirect block for freemap (LEVELN_PSIZE below).
353  *
354  * Leaf entry represents 4MB of storage broken down into a 512-bit
355  * bitmap, 2-bits per entry.  So course bitmap item represents 16KB.
356  */
357 #if HAMMER2_SET_COUNT != 4
358 #error "hammer2_disk.h - freemap assumes SET_COUNT is 4"
359 #endif
360 #define HAMMER2_FREEMAP_LEVEL6_RADIX	64	/* 16EB (end) */
361 #define HAMMER2_FREEMAP_LEVEL5_RADIX	62	/* 4EB */
362 #define HAMMER2_FREEMAP_LEVEL4_RADIX	54	/* 16PB */
363 #define HAMMER2_FREEMAP_LEVEL3_RADIX	46	/* 64TB */
364 #define HAMMER2_FREEMAP_LEVEL2_RADIX	38	/* 256GB */
365 #define HAMMER2_FREEMAP_LEVEL1_RADIX	30	/* 1GB */
366 #define HAMMER2_FREEMAP_LEVEL0_RADIX	22	/* 4MB (128by in l-1 leaf) */
367 
368 #define HAMMER2_FREEMAP_LEVELN_PSIZE	32768	/* physical bytes */
369 
370 #define HAMMER2_FREEMAP_LEVEL5_SIZE	((hammer2_off_t)1 <<		\
371 					 HAMMER2_FREEMAP_LEVEL5_RADIX)
372 #define HAMMER2_FREEMAP_LEVEL4_SIZE	((hammer2_off_t)1 <<		\
373 					 HAMMER2_FREEMAP_LEVEL4_RADIX)
374 #define HAMMER2_FREEMAP_LEVEL3_SIZE	((hammer2_off_t)1 <<		\
375 					 HAMMER2_FREEMAP_LEVEL3_RADIX)
376 #define HAMMER2_FREEMAP_LEVEL2_SIZE	((hammer2_off_t)1 <<		\
377 					 HAMMER2_FREEMAP_LEVEL2_RADIX)
378 #define HAMMER2_FREEMAP_LEVEL1_SIZE	((hammer2_off_t)1 <<		\
379 					 HAMMER2_FREEMAP_LEVEL1_RADIX)
380 #define HAMMER2_FREEMAP_LEVEL0_SIZE	((hammer2_off_t)1 <<		\
381 					 HAMMER2_FREEMAP_LEVEL0_RADIX)
382 
383 #define HAMMER2_FREEMAP_LEVEL5_MASK	(HAMMER2_FREEMAP_LEVEL5_SIZE - 1)
384 #define HAMMER2_FREEMAP_LEVEL4_MASK	(HAMMER2_FREEMAP_LEVEL4_SIZE - 1)
385 #define HAMMER2_FREEMAP_LEVEL3_MASK	(HAMMER2_FREEMAP_LEVEL3_SIZE - 1)
386 #define HAMMER2_FREEMAP_LEVEL2_MASK	(HAMMER2_FREEMAP_LEVEL2_SIZE - 1)
387 #define HAMMER2_FREEMAP_LEVEL1_MASK	(HAMMER2_FREEMAP_LEVEL1_SIZE - 1)
388 #define HAMMER2_FREEMAP_LEVEL0_MASK	(HAMMER2_FREEMAP_LEVEL0_SIZE - 1)
389 
390 #define HAMMER2_FREEMAP_COUNT		(int)(HAMMER2_FREEMAP_LEVELN_PSIZE / \
391 					 sizeof(hammer2_bmap_data_t))
392 
393 /*
394  * XXX I made a mistake and made the reserved area begin at each LEVEL1 zone,
395  *     which is on a 1GB demark.  This will eat a little more space but for
396  *     now we retain compatibility and make FMZONEBASE every 1GB
397  */
398 #define H2FMZONEBASE(key)	((key) & ~HAMMER2_FREEMAP_LEVEL1_MASK)
399 #define H2FMBASE(key, radix)	rounddown2(key, (hammer2_off_t)1 << (radix))
400 
401 /*
402  * 16KB bitmap granularity (x2 bits per entry).
403  */
404 #define HAMMER2_FREEMAP_BLOCK_RADIX	14
405 #define HAMMER2_FREEMAP_BLOCK_SIZE	(1 << HAMMER2_FREEMAP_BLOCK_RADIX)
406 #define HAMMER2_FREEMAP_BLOCK_MASK	(HAMMER2_FREEMAP_BLOCK_SIZE - 1)
407 
408 /*
409  * bitmap[] structure.  2 bits per HAMMER2_FREEMAP_BLOCK_SIZE.
410  *
411  * 8 x 64-bit elements, 2 bits per block.
412  * 32 blocks (radix 5) per element.
413  * representing INDEX_SIZE bytes worth of storage per element.
414  */
415 
416 typedef uint64_t			hammer2_bitmap_t;
417 
418 #define HAMMER2_BMAP_ALLONES		((hammer2_bitmap_t)-1)
419 #define HAMMER2_BMAP_ELEMENTS		8
420 #define HAMMER2_BMAP_BITS_PER_ELEMENT	64
421 #define HAMMER2_BMAP_INDEX_RADIX	5	/* 32 blocks per element */
422 #define HAMMER2_BMAP_BLOCKS_PER_ELEMENT	(1 << HAMMER2_BMAP_INDEX_RADIX)
423 
424 #define HAMMER2_BMAP_INDEX_SIZE		(HAMMER2_FREEMAP_BLOCK_SIZE * \
425 					 HAMMER2_BMAP_BLOCKS_PER_ELEMENT)
426 #define HAMMER2_BMAP_INDEX_MASK		(HAMMER2_BMAP_INDEX_SIZE - 1)
427 
428 #define HAMMER2_BMAP_SIZE		(HAMMER2_BMAP_INDEX_SIZE * \
429 					 HAMMER2_BMAP_ELEMENTS)
430 #define HAMMER2_BMAP_MASK		(HAMMER2_BMAP_SIZE - 1)
431 
432 /*
433  * Two linear areas can be reserved after the initial 4MB segment in the base
434  * zone (the one starting at offset 0).  These areas are NOT managed by the
435  * block allocator and do not fall under HAMMER2 crc checking rules based
436  * at the volume header (but can be self-CRCd internally, depending).
437  */
438 #define HAMMER2_BOOT_MIN_BYTES		HAMMER2_VOLUME_ALIGN
439 #define HAMMER2_BOOT_NOM_BYTES		(64*1024*1024)
440 #define HAMMER2_BOOT_MAX_BYTES		(256*1024*1024)
441 
442 #define HAMMER2_REDO_MIN_BYTES		HAMMER2_VOLUME_ALIGN
443 #define HAMMER2_REDO_NOM_BYTES		(256*1024*1024)
444 #define HAMMER2_REDO_MAX_BYTES		(1024*1024*1024)
445 
446 /*
447  * Most HAMMER2 types are implemented as unsigned 64-bit integers.
448  * Transaction ids are monotonic.
449  *
450  * We utilize 32-bit iSCSI CRCs.
451  */
452 typedef uint64_t hammer2_tid_t;
453 typedef uint64_t hammer2_off_t;
454 typedef uint64_t hammer2_key_t;
455 typedef uint32_t hammer2_crc32_t;
456 
457 /*
458  * Miscellanious ranges (all are unsigned).
459  */
460 #define HAMMER2_TID_MIN		1ULL
461 #define HAMMER2_TID_MAX		0xFFFFFFFFFFFFFFFFULL
462 #define HAMMER2_KEY_MIN		0ULL
463 #define HAMMER2_KEY_MAX		0xFFFFFFFFFFFFFFFFULL
464 #define HAMMER2_OFFSET_MIN	0ULL
465 #define HAMMER2_OFFSET_MAX	0xFFFFFFFFFFFFFFFFULL
466 
467 /*
468  * HAMMER2 data offset special cases and masking.
469  *
470  * All HAMMER2 data offsets have to be broken down into a 64K buffer base
471  * offset (HAMMER2_OFF_MASK_HI) and a 64K buffer index (HAMMER2_OFF_MASK_LO).
472  *
473  * Indexes into physical buffers are always 64-byte aligned.  The low 6 bits
474  * of the data offset field specifies how large the data chunk being pointed
475  * to as a power of 2.  The theoretical minimum radix is thus 6 (The space
476  * needed in the low bits of the data offset field).  However, the practical
477  * minimum allocation chunk size is 1KB (a radix of 10), so HAMMER2 sets
478  * HAMMER2_RADIX_MIN to 10.  The maximum radix is currently 16 (64KB), but
479  * we fully intend to support larger extents in the future.
480  *
481  * WARNING! A radix of 0 (such as when data_off is all 0's) is a special
482  *	    case which means no data associated with the blockref, and
483  *	    not the '1 byte' it would otherwise calculate to.
484  */
485 #define HAMMER2_OFF_BAD		((hammer2_off_t)-1)
486 #define HAMMER2_OFF_MASK	0xFFFFFFFFFFFFFFC0ULL
487 #define HAMMER2_OFF_MASK_LO	(HAMMER2_OFF_MASK & HAMMER2_PBUFMASK64)
488 #define HAMMER2_OFF_MASK_HI	(~HAMMER2_PBUFMASK64)
489 #define HAMMER2_OFF_MASK_RADIX	0x000000000000003FULL
490 #define HAMMER2_MAX_COPIES	6
491 
492 /*
493  * HAMMER2 directory support and pre-defined keys
494  */
495 #define HAMMER2_DIRHASH_VISIBLE	0x8000000000000000ULL
496 #define HAMMER2_DIRHASH_USERMSK	0x7FFFFFFFFFFFFFFFULL
497 #define HAMMER2_DIRHASH_LOMASK	0x0000000000007FFFULL
498 #define HAMMER2_DIRHASH_HIMASK	0xFFFFFFFFFFFF0000ULL
499 #define HAMMER2_DIRHASH_FORCED	0x0000000000008000ULL	/* bit forced on */
500 
501 #define HAMMER2_SROOT_KEY	0x0000000000000000ULL	/* volume to sroot */
502 #define HAMMER2_BOOT_KEY	0xd9b36ce135528000ULL	/* sroot to BOOT PFS */
503 
504 /************************************************************************
505  *				DMSG SUPPORT				*
506  ************************************************************************
507  * LNK_VOLCONF
508  *
509  * All HAMMER2 directories directly under the super-root on your local
510  * media can be mounted separately, even if they share the same physical
511  * device.
512  *
513  * When you do a HAMMER2 mount you are effectively tying into a HAMMER2
514  * cluster via local media.  The local media does not have to participate
515  * in the cluster, other than to provide the hammer2_volconf[] array and
516  * root inode for the mount.
517  *
518  * This is important: The mount device path you specify serves to bootstrap
519  * your entry into the cluster, but your mount will make active connections
520  * to ALL copy elements in the hammer2_volconf[] array which match the
521  * PFSID of the directory in the super-root that you specified.  The local
522  * media path does not have to be mentioned in this array but becomes part
523  * of the cluster based on its type and access rights.  ALL ELEMENTS ARE
524  * TREATED ACCORDING TO TYPE NO MATTER WHICH ONE YOU MOUNT FROM.
525  *
526  * The actual cluster may be far larger than the elements you list in the
527  * hammer2_volconf[] array.  You list only the elements you wish to
528  * directly connect to and you are able to access the rest of the cluster
529  * indirectly through those connections.
530  *
531  * WARNING!  This structure must be exactly 128 bytes long for its config
532  *	     array to fit in the volume header.
533  */
534 struct hammer2_volconf {
535 	uint8_t	copyid;		/* 00	 copyid 0-255 (must match slot) */
536 	uint8_t inprog;		/* 01	 operation in progress, or 0 */
537 	uint8_t chain_to;	/* 02	 operation chaining to, or 0 */
538 	uint8_t chain_from;	/* 03	 operation chaining from, or 0 */
539 	uint16_t flags;		/* 04-05 flags field */
540 	uint8_t error;		/* 06	 last operational error */
541 	uint8_t priority;	/* 07	 priority and round-robin flag */
542 	uint8_t remote_pfs_type;/* 08	 probed direct remote PFS type */
543 	uint8_t reserved08[23];	/* 09-1F */
544 	uuid_t	pfs_clid;	/* 20-2F copy target must match this uuid */
545 	uint8_t label[16];	/* 30-3F import/export label */
546 	uint8_t path[64];	/* 40-7F target specification string or key */
547 } __packed;
548 
549 typedef struct hammer2_volconf hammer2_volconf_t;
550 
551 #define DMSG_VOLF_ENABLED	0x0001
552 #define DMSG_VOLF_INPROG	0x0002
553 #define DMSG_VOLF_CONN_RR	0x80	/* round-robin at same priority */
554 #define DMSG_VOLF_CONN_EF	0x40	/* media errors flagged */
555 #define DMSG_VOLF_CONN_PRI	0x0F	/* select priority 0-15 (15=best) */
556 
557 #if 0
558 struct dmsg_lnk_hammer2_volconf {
559 	dmsg_hdr_t		head;
560 	hammer2_volconf_t	copy;	/* copy spec */
561 	int32_t			index;
562 	int32_t			unused01;
563 	uuid_t			mediaid;
564 	int64_t			reserved02[32];
565 } __packed;
566 #endif
567 
568 typedef struct dmsg_lnk_hammer2_volconf dmsg_lnk_hammer2_volconf_t;
569 
570 #define DMSG_LNK_HAMMER2_VOLCONF DMSG_LNK(DMSG_LNK_CMD_HAMMER2_VOLCONF, \
571 					  dmsg_lnk_hammer2_volconf)
572 
573 #define H2_LNK_VOLCONF(msg)	((dmsg_lnk_hammer2_volconf_t *)(msg)->any.buf)
574 
575 /*
576  * HAMMER2 directory entry header (embedded in blockref)  exactly 16 bytes
577  */
578 struct hammer2_dirent_head {
579 	hammer2_tid_t		inum;		/* inode number */
580 	uint16_t		namlen;		/* name length */
581 	uint8_t			type;		/* OBJTYPE_*	*/
582 	uint8_t			unused0B;
583 	uint8_t			unused0C[4];
584 } __packed;
585 
586 typedef struct hammer2_dirent_head hammer2_dirent_head_t;
587 
588 /*
589  * The media block reference structure.  This forms the core of the HAMMER2
590  * media topology recursion.  This 128-byte data structure is embedded in the
591  * volume header, in inodes (which are also directory entries), and in
592  * indirect blocks.
593  *
594  * A blockref references a single media item, which typically can be a
595  * directory entry (aka inode), indirect block, or data block.
596  *
597  * The primary feature a blockref represents is the ability to validate
598  * the entire tree underneath it via its check code.  Any modification to
599  * anything propagates up the blockref tree all the way to the root, replacing
600  * the related blocks and compounding the generated check code.
601  *
602  * The check code can be a simple 32-bit iscsi code, a 64-bit crc, or as
603  * complex as a 512 bit cryptographic hash.  I originally used a 64-byte
604  * blockref but later expanded it to 128 bytes to be able to support the
605  * larger check code as well as to embed statistics for quota operation.
606  *
607  * Simple check codes are not sufficient for unverified dedup.  Even with
608  * a maximally-sized check code unverified dedup should only be used in
609  * in subdirectory trees where you do not need 100% data integrity.
610  *
611  * Unverified dedup is deduping based on meta-data only without verifying
612  * that the data blocks are actually identical.  Verified dedup guarantees
613  * integrity but is a far more I/O-expensive operation.
614  *
615  * --
616  *
617  * mirror_tid - per cluster node modified (propagated upward by flush)
618  * modify_tid - clc record modified (not propagated).
619  * update_tid - clc record updated (propagated upward on verification)
620  *
621  * CLC - Stands for 'Cluster Level Change', identifiers which are identical
622  *	 within the topology across all cluster nodes (when fully
623  *	 synchronized).
624  *
625  * NOTE: The range of keys represented by the blockref is (key) to
626  *	 ((key) + (1LL << keybits) - 1).  HAMMER2 usually populates
627  *	 blocks bottom-up, inserting a new root when radix expansion
628  *	 is required.
629  *
630  * leaf_count  - Helps manage leaf collapse calculations when indirect
631  *		 blocks become mostly empty.  This value caps out at
632  *		 HAMMER2_BLOCKREF_LEAF_MAX (65535).
633  *
634  *		 Used by the chain code to determine when to pull leafs up
635  *		 from nearly empty indirect blocks.  For the purposes of this
636  *		 calculation, BREF_TYPE_INODE is considered a leaf, along
637  *		 with DIRENT and DATA.
638  *
639  *				    RESERVED FIELDS
640  *
641  * A number of blockref fields are reserved and should generally be set to
642  * 0 for future compatibility.
643  *
644  *				FUTURE BLOCKREF EXPANSION
645  *
646  * CONTENT ADDRESSABLE INDEXING (future) - Using a 256 or 512-bit check code.
647  */
648 struct hammer2_blockref {		/* MUST BE EXACTLY 64 BYTES */
649 	uint8_t		type;		/* type of underlying item */
650 	uint8_t		methods;	/* check method & compression method */
651 	uint8_t		copyid;		/* specify which copy this is */
652 	uint8_t		keybits;	/* #of keybits masked off 0=leaf */
653 	uint8_t		vradix;		/* virtual data/meta-data size */
654 	uint8_t		flags;		/* blockref flags */
655 	uint16_t	leaf_count;	/* leaf aggregation count */
656 	hammer2_key_t	key;		/* key specification */
657 	hammer2_tid_t	mirror_tid;	/* media flush topology & freemap */
658 	hammer2_tid_t	modify_tid;	/* clc modify (not propagated) */
659 	hammer2_off_t	data_off;	/* low 6 bits is phys size (radix)*/
660 	hammer2_tid_t	update_tid;	/* clc modify (propagated upward) */
661 	union {
662 		char	buf[16];
663 
664 		/*
665 		 * Directory entry header (BREF_TYPE_DIRENT)
666 		 *
667 		 * NOTE: check.buf contains filename if <= 64 bytes.  Longer
668 		 *	 filenames are stored in a data reference of size
669 		 *	 HAMMER2_ALLOC_MIN (at least 256, typically 1024).
670 		 *
671 		 * NOTE: inode structure may contain a copy of a recently
672 		 *	 associated filename, for recovery purposes.
673 		 *
674 		 * NOTE: Superroot entries are INODEs, not DIRENTs.  Code
675 		 *	 allows both cases.
676 		 */
677 		hammer2_dirent_head_t dirent;
678 
679 		/*
680 		 * Statistics aggregation (BREF_TYPE_INODE, BREF_TYPE_INDIRECT)
681 		 */
682 		struct {
683 			hammer2_key_t	data_count;
684 			hammer2_key_t	inode_count;
685 		} stats;
686 	} embed;
687 	union {				/* check info */
688 		char	buf[64];
689 		struct {
690 			uint32_t value;
691 			uint32_t reserved[15];
692 		} iscsi32;
693 		struct {
694 			uint64_t value;
695 			uint64_t reserved[7];
696 		} xxhash64;
697 		struct {
698 			char data[24];
699 			char reserved[40];
700 		} sha192;
701 		struct {
702 			char data[32];
703 			char reserved[32];
704 		} sha256;
705 		struct {
706 			char data[64];
707 		} sha512;
708 
709 		/*
710 		 * Freemap hints are embedded in addition to the icrc32.
711 		 *
712 		 * bigmask - Radixes available for allocation (0-31).
713 		 *	     Heuristical (may be permissive but not
714 		 *	     restrictive).  Typically only radix values
715 		 *	     10-16 are used (i.e. (1<<10) through (1<<16)).
716 		 *
717 		 * avail   - Total available space remaining, in bytes
718 		 */
719 		struct {
720 			uint32_t icrc32;
721 			uint32_t bigmask;	/* available radixes */
722 			uint64_t avail;		/* total available bytes */
723 			char reserved[48];
724 		} freemap;
725 	} check;
726 } __packed;
727 
728 typedef struct hammer2_blockref hammer2_blockref_t;
729 
730 #define HAMMER2_BLOCKREF_BYTES		128	/* blockref struct in bytes */
731 #define HAMMER2_BLOCKREF_RADIX		7
732 
733 #define HAMMER2_BLOCKREF_LEAF_MAX	65535
734 
735 /*
736  * On-media and off-media blockref types.
737  *
738  * types >= 128 are pseudo values that should never be present on-media.
739  */
740 #define HAMMER2_BREF_TYPE_EMPTY		0
741 #define HAMMER2_BREF_TYPE_INODE		1
742 #define HAMMER2_BREF_TYPE_INDIRECT	2
743 #define HAMMER2_BREF_TYPE_DATA		3
744 #define HAMMER2_BREF_TYPE_DIRENT	4
745 #define HAMMER2_BREF_TYPE_FREEMAP_NODE	5
746 #define HAMMER2_BREF_TYPE_FREEMAP_LEAF	6
747 #define HAMMER2_BREF_TYPE_FREEMAP	254	/* pseudo-type */
748 #define HAMMER2_BREF_TYPE_VOLUME	255	/* pseudo-type */
749 
750 #define HAMMER2_BREF_FLAG_PFSROOT	0x01	/* see also related opflag */
751 #define HAMMER2_BREF_FLAG_ZERO		0x02
752 #define HAMMER2_BREF_FLAG_EMERG_MIP	0x04	/* emerg modified-in-place */
753 
754 /*
755  * Encode/decode check mode and compression mode for
756  * bref.methods.  The compression level is not encoded in
757  * bref.methods.
758  */
759 #define HAMMER2_ENC_CHECK(n)		(((n) & 15) << 4)
760 #define HAMMER2_DEC_CHECK(n)		(((n) >> 4) & 15)
761 #define HAMMER2_ENC_COMP(n)		((n) & 15)
762 #define HAMMER2_DEC_COMP(n)		((n) & 15)
763 
764 #define HAMMER2_CHECK_NONE		0
765 #define HAMMER2_CHECK_DISABLED		1
766 #define HAMMER2_CHECK_ISCSI32		2
767 #define HAMMER2_CHECK_XXHASH64		3
768 #define HAMMER2_CHECK_SHA192		4
769 #define HAMMER2_CHECK_FREEMAP		5
770 
771 #define HAMMER2_CHECK_DEFAULT		HAMMER2_CHECK_XXHASH64
772 
773 /* user-specifiable check modes only */
774 #define HAMMER2_CHECK_STRINGS		{ "none", "disabled", "crc32", \
775 					  "xxhash64", "sha192" }
776 #define HAMMER2_CHECK_STRINGS_COUNT	5
777 
778 /*
779  * Encode/decode check or compression algorithm request in
780  * ipdata->meta.check_algo and ipdata->meta.comp_algo.
781  */
782 #define HAMMER2_ENC_ALGO(n)		(n)
783 #define HAMMER2_DEC_ALGO(n)		((n) & 15)
784 #define HAMMER2_ENC_LEVEL(n)		((n) << 4)
785 #define HAMMER2_DEC_LEVEL(n)		(((n) >> 4) & 15)
786 
787 #define HAMMER2_COMP_NONE		0
788 #define HAMMER2_COMP_AUTOZERO		1
789 #define HAMMER2_COMP_LZ4		2
790 #define HAMMER2_COMP_ZLIB		3
791 
792 #define HAMMER2_COMP_NEWFS_DEFAULT	HAMMER2_COMP_LZ4
793 #define HAMMER2_COMP_STRINGS		{ "none", "autozero", "lz4", "zlib" }
794 #define HAMMER2_COMP_STRINGS_COUNT	4
795 
796 /*
797  * Passed to hammer2_chain_create(), causes methods to be inherited from
798  * parent.
799  */
800 #define HAMMER2_METH_DEFAULT		-1
801 
802 /*
803  * HAMMER2 block references are collected into sets of 4 blockrefs.  These
804  * sets are fully associative, meaning the elements making up a set are
805  * not sorted in any way and may contain duplicate entries, holes, or
806  * entries which shortcut multiple levels of indirection.  Sets are used
807  * in various ways:
808  *
809  * (1) When redundancy is desired a set may contain several duplicate
810  *     entries pointing to different copies of the same data.  Up to 4 copies
811  *     are supported.
812  *
813  * (2) The blockrefs in a set can shortcut multiple levels of indirections
814  *     within the bounds imposed by the parent of set.
815  *
816  * When a set fills up another level of indirection is inserted, moving
817  * some or all of the set's contents into indirect blocks placed under the
818  * set.  This is a top-down approach in that indirect blocks are not created
819  * until the set actually becomes full (that is, the entries in the set can
820  * shortcut the indirect blocks when the set is not full).  Depending on how
821  * things are filled multiple indirect blocks will eventually be created.
822  *
823  * Indirect blocks are typically 4KB (64 entres) or 64KB (1024 entries) and
824  * are also treated as fully set-associative.
825  */
826 struct hammer2_blockset {
827 	hammer2_blockref_t	blockref[HAMMER2_SET_COUNT];
828 };
829 
830 typedef struct hammer2_blockset hammer2_blockset_t;
831 
832 /*
833  * Catch programmer snafus
834  */
835 #if (1 << HAMMER2_SET_RADIX) != HAMMER2_SET_COUNT
836 #error "hammer2 direct radix is incorrect"
837 #endif
838 #if (1 << HAMMER2_PBUFRADIX) != HAMMER2_PBUFSIZE
839 #error "HAMMER2_PBUFRADIX and HAMMER2_PBUFSIZE are inconsistent"
840 #endif
841 #if (1 << HAMMER2_RADIX_MIN) != HAMMER2_ALLOC_MIN
842 #error "HAMMER2_RADIX_MIN and HAMMER2_ALLOC_MIN are inconsistent"
843 #endif
844 
845 /*
846  * hammer2_bmap_data - A freemap entry in the LEVEL1 block.
847  *
848  * Each 128-byte entry contains the bitmap and meta-data required to manage
849  * a LEVEL0 (4MB) block of storage.  The storage is managed in 256 x 16KB
850  * chunks.
851  *
852  * A smaller allocation granularity is supported via a linear iterator and/or
853  * must otherwise be tracked in ram.
854  *
855  * (data structure must be 128 bytes exactly)
856  *
857  * linear  - A BYTE linear allocation offset used for sub-16KB allocations
858  *	     only.  May contain values between 0 and 4MB.  Must be ignored
859  *	     if 16KB-aligned (i.e. force bitmap scan), otherwise may be
860  *	     used to sub-allocate within the 16KB block (which is already
861  *	     marked as allocated in the bitmap).
862  *
863  *	     Sub-allocations need only be 1KB-aligned and do not have to be
864  *	     size-aligned, and 16KB or larger allocations do not update this
865  *	     field, resulting in pretty good packing.
866  *
867  *	     Please note that file data granularity may be limited by
868  *	     other issues such as buffer cache direct-mapping and the
869  *	     desire to support sector sizes up to 16KB (so H2 only issues
870  *	     I/O's in multiples of 16KB anyway).
871  *
872  * class   - Clustering class.  Cleared to 0 only if the entire leaf becomes
873  *	     free.  Used to cluster device buffers so all elements must have
874  *	     the same device block size, but may mix logical sizes.
875  *
876  *	     Typically integrated with the blockref type in the upper 8 bits
877  *	     to localize inodes and indrect blocks, improving bulk free scans
878  *	     and directory scans.
879  *
880  * bitmap  - Two bits per 16KB allocation block arranged in arrays of
881  *	     64-bit elements, 256x2 bits representing ~4MB worth of media
882  *	     storage.  Bit patterns are as follows:
883  *
884  *	     00	Unallocated
885  *	     01 (reserved)
886  *	     10 Possibly free
887  *           11 Allocated
888  */
889 struct hammer2_bmap_data {
890 	int32_t linear;		/* 00 linear sub-granular allocation offset */
891 	uint16_t class;		/* 04-05 clustering class ((type<<8)|radix) */
892 	uint8_t reserved06;	/* 06 */
893 	uint8_t reserved07;	/* 07 */
894 	uint32_t reserved08;	/* 08 */
895 	uint32_t reserved0C;	/* 0C */
896 	uint32_t reserved10;	/* 10 */
897 	uint32_t reserved14;	/* 14 */
898 	uint32_t reserved18;	/* 18 */
899 	uint32_t avail;		/* 1C */
900 	uint32_t reserved20[8];	/* 20-3F 256 bits manages 128K/1KB/2-bits */
901 				/* 40-7F 512 bits manages 4MB of storage */
902 	hammer2_bitmap_t bitmapq[HAMMER2_BMAP_ELEMENTS];
903 } __packed;
904 
905 typedef struct hammer2_bmap_data hammer2_bmap_data_t;
906 
907 /*
908  * XXX "Inodes ARE directory entries" is no longer the case.  Hardlinks are
909  * dirents which refer to the same inode#, which is how filesystems usually
910  * implement hardlink.  The following comments need to be updated.
911  *
912  * In HAMMER2 inodes ARE directory entries, with a special exception for
913  * hardlinks.  The inode number is stored in the inode rather than being
914  * based on the location of the inode (since the location moves every time
915  * the inode or anything underneath the inode is modified).
916  *
917  * The inode is 1024 bytes, made up of 256 bytes of meta-data, 256 bytes
918  * for the filename, and 512 bytes worth of direct file data OR an embedded
919  * blockset.  The in-memory hammer2_inode structure contains only the mostly-
920  * node-independent meta-data portion (some flags are node-specific and will
921  * not be synchronized).  The rest of the inode is node-specific and chain I/O
922  * is required to obtain it.
923  *
924  * Directories represent one inode per blockref.  Inodes are not laid out
925  * as a file but instead are represented by the related blockrefs.  The
926  * blockrefs, in turn, are indexed by the 64-bit directory hash key.  Remember
927  * that blocksets are fully associative, so a certain degree efficiency is
928  * achieved just from that.
929  *
930  * Up to 512 bytes of direct data can be embedded in an inode, and since
931  * inodes are essentially directory entries this also means that small data
932  * files end up simply being laid out linearly in the directory, resulting
933  * in fewer seeks and highly optimal access.
934  *
935  * The compression mode can be changed at any time in the inode and is
936  * recorded on a blockref-by-blockref basis.
937  *
938  * Hardlinks are supported via the inode map.  Essentially the way a hardlink
939  * works is that all individual directory entries representing the same file
940  * are special cased and specify the same inode number.  The actual file
941  * is placed in the nearest parent directory that is parent to all instances
942  * of the hardlink.  If all hardlinks to a file are in the same directory
943  * the actual file will also be placed in that directory.  This file uses
944  * the inode number as the directory entry key and is invisible to normal
945  * directory scans.  Real directory entry keys are differentiated from the
946  * inode number key via bit 63.  Access to the hardlink silently looks up
947  * the real file and forwards all operations to that file.  Removal of the
948  * last hardlink also removes the real file.
949  */
950 #define HAMMER2_INODE_BYTES		1024	/* (asserted by code) */
951 #define HAMMER2_INODE_MAXNAME		256	/* maximum name in bytes */
952 #define HAMMER2_INODE_VERSION_ONE	1
953 
954 #define HAMMER2_INODE_START		1024	/* dynamically allocated */
955 
956 struct hammer2_inode_meta {
957 	uint16_t	version;	/* 0000 inode data version */
958 	uint8_t		reserved02;	/* 0002 */
959 	uint8_t		pfs_subtype;	/* 0003 pfs sub-type */
960 
961 	/*
962 	 * core inode attributes, inode type, misc flags
963 	 */
964 	uint32_t	uflags;		/* 0004 chflags */
965 	uint32_t	rmajor;		/* 0008 available for device nodes */
966 	uint32_t	rminor;		/* 000C available for device nodes */
967 	uint64_t	ctime;		/* 0010 inode change time */
968 	uint64_t	mtime;		/* 0018 modified time */
969 	uint64_t	atime;		/* 0020 access time (unsupported) */
970 	uint64_t	btime;		/* 0028 birth time */
971 	uuid_t		uid;		/* 0030 uid / degenerate unix uid */
972 	uuid_t		gid;		/* 0040 gid / degenerate unix gid */
973 
974 	uint8_t		type;		/* 0050 object type */
975 	uint8_t		op_flags;	/* 0051 operational flags */
976 	uint16_t	cap_flags;	/* 0052 capability flags */
977 	uint32_t	mode;		/* 0054 unix modes (typ low 16 bits) */
978 
979 	/*
980 	 * inode size, identification, localized recursive configuration
981 	 * for compression and backup copies.
982 	 *
983 	 * NOTE: Nominal parent inode number (iparent) is only applicable
984 	 *	 for directories but can also help for files during
985 	 *	 catastrophic recovery.
986 	 */
987 	hammer2_tid_t	inum;		/* 0058 inode number */
988 	hammer2_off_t	size;		/* 0060 size of file */
989 	uint64_t	nlinks;		/* 0068 hard links (typ only dirs) */
990 	hammer2_tid_t	iparent;	/* 0070 nominal parent inum */
991 	hammer2_key_t	name_key;	/* 0078 full filename key */
992 	uint16_t	name_len;	/* 0080 filename length */
993 	uint8_t		ncopies;	/* 0082 ncopies to local media */
994 	uint8_t		comp_algo;	/* 0083 compression request & algo */
995 
996 	/*
997 	 * These fields are currently only applicable to PFSROOTs.
998 	 *
999 	 * NOTE: We can't use {volume_data->fsid, pfs_clid} to uniquely
1000 	 *	 identify an instance of a PFS in the cluster because
1001 	 *	 a mount may contain more than one copy of the PFS as
1002 	 *	 a separate node.  {pfs_clid, pfs_fsid} must be used for
1003 	 *	 registration in the cluster.
1004 	 */
1005 	uint8_t		target_type;	/* 0084 hardlink target type */
1006 	uint8_t		check_algo;	/* 0085 check code request & algo */
1007 	uint8_t		pfs_nmasters;	/* 0086 (if PFSROOT) if multi-master */
1008 	uint8_t		pfs_type;	/* 0087 (if PFSROOT) node type */
1009 	hammer2_tid_t	pfs_inum;	/* 0088 (if PFSROOT) inum allocator */
1010 	uuid_t		pfs_clid;	/* 0090 (if PFSROOT) cluster uuid */
1011 	uuid_t		pfs_fsid;	/* 00A0 (if PFSROOT) unique uuid */
1012 
1013 	/*
1014 	 * Quotas and aggregate sub-tree inode and data counters.  Note that
1015 	 * quotas are not replicated downward, they are explicitly set by
1016 	 * the sysop and in-memory structures keep track of inheritance.
1017 	 */
1018 	hammer2_key_t	data_quota;	/* 00B0 subtree quota in bytes */
1019 	hammer2_key_t	unusedB8;	/* 00B8 subtree byte count */
1020 	hammer2_key_t	inode_quota;	/* 00C0 subtree quota inode count */
1021 	hammer2_key_t	unusedC8;	/* 00C8 subtree inode count */
1022 
1023 	/*
1024 	 * The last snapshot tid is tested against modify_tid to determine
1025 	 * when a copy must be made of a data block whos check mode has been
1026 	 * disabled (a disabled check mode allows data blocks to be updated
1027 	 * in place instead of copy-on-write).
1028 	 */
1029 	hammer2_tid_t	pfs_lsnap_tid;	/* 00D0 last snapshot tid */
1030 	hammer2_tid_t	reservedD8;	/* 00D8 (avail) */
1031 
1032 	/*
1033 	 * Tracks (possibly degenerate) free areas covering all sub-tree
1034 	 * allocations under inode, not counting the inode itself.
1035 	 * 0/0 indicates empty entry.  fully set-associative.
1036 	 *
1037 	 * (not yet implemented)
1038 	 */
1039 	uint64_t	decrypt_check;	/* 00E0 decryption validator */
1040 	hammer2_off_t	reservedE0[3];	/* 00E8/F0/F8 */
1041 } __packed;
1042 
1043 typedef struct hammer2_inode_meta hammer2_inode_meta_t;
1044 
1045 struct hammer2_inode_data {
1046 	hammer2_inode_meta_t	meta;	/* 0000-00FF */
1047 	unsigned char	filename[HAMMER2_INODE_MAXNAME];
1048 					/* 0100-01FF (256 char, unterminated) */
1049 	union {				/* 0200-03FF (64x8 = 512 bytes) */
1050 		hammer2_blockset_t blockset;
1051 		char data[HAMMER2_EMBEDDED_BYTES];
1052 	} u;
1053 } __packed;
1054 
1055 typedef struct hammer2_inode_data hammer2_inode_data_t;
1056 
1057 #define HAMMER2_OPFLAG_DIRECTDATA	0x01
1058 #define HAMMER2_OPFLAG_PFSROOT		0x02	/* (see also bref flag) */
1059 #define HAMMER2_OPFLAG_COPYIDS		0x04	/* copyids override parent */
1060 
1061 #define HAMMER2_OBJTYPE_UNKNOWN		0
1062 #define HAMMER2_OBJTYPE_DIRECTORY	1
1063 #define HAMMER2_OBJTYPE_REGFILE		2
1064 #define HAMMER2_OBJTYPE_FIFO		4
1065 #define HAMMER2_OBJTYPE_CDEV		5
1066 #define HAMMER2_OBJTYPE_BDEV		6
1067 #define HAMMER2_OBJTYPE_SOFTLINK	7
1068 #define HAMMER2_OBJTYPE_UNUSED08	8
1069 #define HAMMER2_OBJTYPE_SOCKET		9
1070 #define HAMMER2_OBJTYPE_WHITEOUT	10
1071 
1072 #define HAMMER2_COPYID_NONE		0
1073 #define HAMMER2_COPYID_LOCAL		((uint8_t)-1)
1074 
1075 #define HAMMER2_COPYID_COUNT		256
1076 
1077 /*
1078  * PFS types identify the role of a PFS within a cluster.  The PFS types
1079  * is stored on media and in LNK_SPAN messages and used in other places.
1080  *
1081  * The low 4 bits specify the current active type while the high 4 bits
1082  * specify the transition target if the PFS is being upgraded or downgraded,
1083  * If the upper 4 bits are not zero it may effect how a PFS is used during
1084  * the transition.
1085  *
1086  * Generally speaking, downgrading a MASTER to a SLAVE cannot complete until
1087  * at least all MASTERs have updated their pfs_nmasters field.  And upgrading
1088  * a SLAVE to a MASTER cannot complete until the new prospective master has
1089  * been fully synchronized (though theoretically full synchronization is
1090  * not required if a (new) quorum of other masters are fully synchronized).
1091  *
1092  * It generally does not matter which PFS element you actually mount, you
1093  * are mounting 'the cluster'.  So, for example, a network mount will mount
1094  * a DUMMY PFS type on a memory filesystem.  However, there are two exceptions.
1095  * In order to gain the benefits of a SOFT_MASTER or SOFT_SLAVE, those PFSs
1096  * must be directly mounted.
1097  */
1098 #define HAMMER2_PFSTYPE_NONE		0x00
1099 #define HAMMER2_PFSTYPE_CACHE		0x01
1100 #define HAMMER2_PFSTYPE_UNUSED02	0x02
1101 #define HAMMER2_PFSTYPE_SLAVE		0x03
1102 #define HAMMER2_PFSTYPE_SOFT_SLAVE	0x04
1103 #define HAMMER2_PFSTYPE_SOFT_MASTER	0x05
1104 #define HAMMER2_PFSTYPE_MASTER		0x06
1105 #define HAMMER2_PFSTYPE_UNUSED07	0x07
1106 #define HAMMER2_PFSTYPE_SUPROOT		0x08
1107 #define HAMMER2_PFSTYPE_DUMMY		0x09
1108 #define HAMMER2_PFSTYPE_MAX		16
1109 
1110 #define HAMMER2_PFSTRAN_NONE		0x00	/* no transition in progress */
1111 #define HAMMER2_PFSTRAN_CACHE		0x10
1112 #define HAMMER2_PFSTRAN_UNMUSED20	0x20
1113 #define HAMMER2_PFSTRAN_SLAVE		0x30
1114 #define HAMMER2_PFSTRAN_SOFT_SLAVE	0x40
1115 #define HAMMER2_PFSTRAN_SOFT_MASTER	0x50
1116 #define HAMMER2_PFSTRAN_MASTER		0x60
1117 #define HAMMER2_PFSTRAN_UNUSED70	0x70
1118 #define HAMMER2_PFSTRAN_SUPROOT		0x80
1119 #define HAMMER2_PFSTRAN_DUMMY		0x90
1120 
1121 #define HAMMER2_PFS_DEC(n)		((n) & 0x0F)
1122 #define HAMMER2_PFS_DEC_TRANSITION(n)	(((n) >> 4) & 0x0F)
1123 #define HAMMER2_PFS_ENC_TRANSITION(n)	(((n) & 0x0F) << 4)
1124 
1125 #define HAMMER2_PFSSUBTYPE_NONE		0
1126 #define HAMMER2_PFSSUBTYPE_SNAPSHOT	1	/* manual/managed snapshot */
1127 #define HAMMER2_PFSSUBTYPE_AUTOSNAP	2	/* automatic snapshot */
1128 
1129 /*
1130  * PFS mode of operation is a bitmask.  This is typically not stored
1131  * on-media, but defined here because the field may be used in dmsgs.
1132  */
1133 #define HAMMER2_PFSMODE_QUORUM		0x01
1134 #define HAMMER2_PFSMODE_RW		0x02
1135 
1136 /*
1137  *				Allocation Table
1138  *
1139  */
1140 
1141 
1142 /*
1143  * Flags (8 bits) - blockref, for freemap only
1144  *
1145  * Note that the minimum chunk size is 1KB so we could theoretically have
1146  * 10 bits here, but we might have some future extension that allows a
1147  * chunk size down to 256 bytes and if so we will need bits 8 and 9.
1148  */
1149 #define HAMMER2_AVF_SELMASK		0x03	/* select group */
1150 #define HAMMER2_AVF_ALL_ALLOC		0x04	/* indicate all allocated */
1151 #define HAMMER2_AVF_ALL_FREE		0x08	/* indicate all free */
1152 #define HAMMER2_AVF_RESERVED10		0x10
1153 #define HAMMER2_AVF_RESERVED20		0x20
1154 #define HAMMER2_AVF_RESERVED40		0x40
1155 #define HAMMER2_AVF_RESERVED80		0x80
1156 #define HAMMER2_AVF_AVMASK32		((uint32_t)0xFFFFFF00LU)
1157 #define HAMMER2_AVF_AVMASK64		((uint64_t)0xFFFFFFFFFFFFFF00LLU)
1158 
1159 #define HAMMER2_AV_SELECT_A		0x00
1160 #define HAMMER2_AV_SELECT_B		0x01
1161 #define HAMMER2_AV_SELECT_C		0x02
1162 #define HAMMER2_AV_SELECT_D		0x03
1163 
1164 /*
1165  * The volume header eats a 64K block.  There is currently an issue where
1166  * we want to try to fit all nominal filesystem updates in a 512-byte section
1167  * but it may be a lost cause due to the need for a blockset.
1168  *
1169  * All information is stored in host byte order.  The volume header's magic
1170  * number may be checked to determine the byte order.  If you wish to mount
1171  * between machines w/ different endian modes you'll need filesystem code
1172  * which acts on the media data consistently (either all one way or all the
1173  * other).  Our code currently does not do that.
1174  *
1175  * A read-write mount may have to recover missing allocations by doing an
1176  * incremental mirror scan looking for modifications made after alloc_tid.
1177  * If alloc_tid == last_tid then no recovery operation is needed.  Recovery
1178  * operations are usually very, very fast.
1179  *
1180  * Read-only mounts do not need to do any recovery, access to the filesystem
1181  * topology is always consistent after a crash (is always consistent, period).
1182  * However, there may be shortcutted blockref updates present from deep in
1183  * the tree which are stored in the volumeh eader and must be tracked on
1184  * the fly.
1185  *
1186  * NOTE: The copyinfo[] array contains the configuration for both the
1187  *	 cluster connections and any local media copies.  The volume
1188  *	 header will be replicated for each local media copy.
1189  *
1190  *	 The mount command may specify multiple medias or just one and
1191  *	 allow HAMMER2 to pick up the others when it checks the copyinfo[]
1192  *	 array on mount.
1193  *
1194  * NOTE: root_blockref points to the super-root directory, not the root
1195  *	 directory.  The root directory will be a subdirectory under the
1196  *	 super-root.
1197  *
1198  *	 The super-root directory contains all root directories and all
1199  *	 snapshots (readonly or writable).  It is possible to do a
1200  *	 null-mount of the super-root using special path constructions
1201  *	 relative to your mounted root.
1202  *
1203  * NOTE: HAMMER2 allows any subdirectory tree to be managed as if it were
1204  *	 a PFS, including mirroring and storage quota operations, and this is
1205  *	 prefered over creating discrete PFSs in the super-root.  Instead
1206  *	 the super-root is most typically used to create writable snapshots,
1207  *	 alternative roots, and so forth.  The super-root is also used by
1208  *	 the automatic snapshotting mechanism.
1209  */
1210 #define HAMMER2_VOLUME_ID_HBO	0x48414d3205172011LLU
1211 #define HAMMER2_VOLUME_ID_ABO	0x11201705324d4148LLU
1212 
1213 struct hammer2_volume_data {
1214 	/*
1215 	 * sector #0 - 512 bytes
1216 	 */
1217 	uint64_t	magic;			/* 0000 Signature */
1218 	hammer2_off_t	boot_beg;		/* 0008 Boot area (future) */
1219 	hammer2_off_t	boot_end;		/* 0010 (size = end - beg) */
1220 	hammer2_off_t	aux_beg;		/* 0018 Aux area (future) */
1221 	hammer2_off_t	aux_end;		/* 0020 (size = end - beg) */
1222 	hammer2_off_t	volu_size;		/* 0028 Volume size, bytes */
1223 
1224 	uint32_t	version;		/* 0030 */
1225 	uint32_t	flags;			/* 0034 */
1226 	uint8_t		copyid;			/* 0038 copyid of phys vol */
1227 	uint8_t		freemap_version;	/* 0039 freemap algorithm */
1228 	uint8_t		peer_type;		/* 003A HAMMER2_PEER_xxx */
1229 	uint8_t		reserved003B;		/* 003B */
1230 	uint32_t	reserved003C;		/* 003C */
1231 
1232 	uuid_t		fsid;			/* 0040 */
1233 	uuid_t		fstype;			/* 0050 */
1234 
1235 	/*
1236 	 * allocator_size is precalculated at newfs time and does not include
1237 	 * reserved blocks, boot, or redo areas.
1238 	 *
1239 	 * Initial non-reserved-area allocations do not use the freemap
1240 	 * but instead adjust alloc_iterator.  Dynamic allocations take
1241 	 * over starting at (allocator_beg).  This makes newfs_hammer2's
1242 	 * job a lot easier and can also serve as a testing jig.
1243 	 */
1244 	hammer2_off_t	allocator_size;		/* 0060 Total data space */
1245 	hammer2_off_t   allocator_free;		/* 0068	Free space */
1246 	hammer2_off_t	allocator_beg;		/* 0070 Initial allocations */
1247 
1248 	/*
1249 	 * mirror_tid reflects the highest committed change for this
1250 	 * block device regardless of whether it is to the super-root
1251 	 * or to a PFS or whatever.
1252 	 *
1253 	 * freemap_tid reflects the highest committed freemap change for
1254 	 * this block device.
1255 	 */
1256 	hammer2_tid_t	mirror_tid;		/* 0078 committed tid (vol) */
1257 	hammer2_tid_t	reserved0080;		/* 0080 */
1258 	hammer2_tid_t	reserved0088;		/* 0088 */
1259 	hammer2_tid_t	freemap_tid;		/* 0090 committed tid (fmap) */
1260 	hammer2_tid_t	bulkfree_tid;		/* 0098 bulkfree incremental */
1261 	hammer2_tid_t	reserved00A0[5];	/* 00A0-00C7 */
1262 
1263 	/*
1264 	 * Copyids are allocated dynamically from the copyexists bitmap.
1265 	 * An id from the active copies set (up to 8, see copyinfo later on)
1266 	 * may still exist after the copy set has been removed from the
1267 	 * volume header and its bit will remain active in the bitmap and
1268 	 * cannot be reused until it is 100% removed from the hierarchy.
1269 	 */
1270 	uint32_t	copyexists[8];		/* 00C8-00E7 copy exists bmap */
1271 	char		reserved0140[248];	/* 00E8-01DF */
1272 
1273 	/*
1274 	 * 32 bit CRC array at the end of the first 512 byte sector.
1275 	 *
1276 	 * icrc_sects[7] - First 512-4 bytes of volume header (including all
1277 	 *		   the other icrc's except this one).
1278 	 *
1279 	 * icrc_sects[6] - Sector 1 (512 bytes) of volume header, which is
1280 	 *		   the blockset for the root.
1281 	 *
1282 	 * icrc_sects[5] - Sector 2
1283 	 * icrc_sects[4] - Sector 3
1284 	 * icrc_sects[3] - Sector 4 (the freemap blockset)
1285 	 */
1286 	hammer2_crc32_t	icrc_sects[8];		/* 01E0-01FF */
1287 
1288 	/*
1289 	 * sector #1 - 512 bytes
1290 	 *
1291 	 * The entire sector is used by a blockset.
1292 	 */
1293 	hammer2_blockset_t sroot_blockset;	/* 0200-03FF Superroot dir */
1294 
1295 	/*
1296 	 * sector #2-7
1297 	 */
1298 	char	sector2[512];			/* 0400-05FF reserved */
1299 	char	sector3[512];			/* 0600-07FF reserved */
1300 	hammer2_blockset_t freemap_blockset;	/* 0800-09FF freemap  */
1301 	char	sector5[512];			/* 0A00-0BFF reserved */
1302 	char	sector6[512];			/* 0C00-0DFF reserved */
1303 	char	sector7[512];			/* 0E00-0FFF reserved */
1304 
1305 	/*
1306 	 * sector #8-71	- 32768 bytes
1307 	 *
1308 	 * Contains the configuration for up to 256 copyinfo targets.  These
1309 	 * specify local and remote copies operating as masters or slaves.
1310 	 * copyid's 0 and 255 are reserved (0 indicates an empty slot and 255
1311 	 * indicates the local media).
1312 	 *
1313 	 * Each inode contains a set of up to 8 copyids, either inherited
1314 	 * from its parent or explicitly specified in the inode, which
1315 	 * indexes into this array.
1316 	 */
1317 						/* 1000-8FFF copyinfo config */
1318 	hammer2_volconf_t copyinfo[HAMMER2_COPYID_COUNT];
1319 
1320 	/*
1321 	 * Remaining sections are reserved for future use.
1322 	 */
1323 	char		reserved0400[0x6FFC];	/* 9000-FFFB reserved */
1324 
1325 	/*
1326 	 * icrc on entire volume header
1327 	 */
1328 	hammer2_crc32_t	icrc_volheader;		/* FFFC-FFFF full volume icrc*/
1329 } __packed;
1330 
1331 typedef struct hammer2_volume_data hammer2_volume_data_t;
1332 
1333 /*
1334  * Various parts of the volume header have their own iCRCs.
1335  *
1336  * The first 512 bytes has its own iCRC stored at the end of the 512 bytes
1337  * and not included the icrc calculation.
1338  *
1339  * The second 512 bytes also has its own iCRC but it is stored in the first
1340  * 512 bytes so it covers the entire second 512 bytes.
1341  *
1342  * The whole volume block (64KB) has an iCRC covering all but the last 4 bytes,
1343  * which is where the iCRC for the whole volume is stored.  This is currently
1344  * a catch-all for anything not individually iCRCd.
1345  */
1346 #define HAMMER2_VOL_ICRC_SECT0		7
1347 #define HAMMER2_VOL_ICRC_SECT1		6
1348 
1349 #define HAMMER2_VOLUME_BYTES		65536
1350 
1351 #define HAMMER2_VOLUME_ICRC0_OFF	0
1352 #define HAMMER2_VOLUME_ICRC1_OFF	512
1353 #define HAMMER2_VOLUME_ICRCVH_OFF	0
1354 
1355 #define HAMMER2_VOLUME_ICRC0_SIZE	(512 - 4)
1356 #define HAMMER2_VOLUME_ICRC1_SIZE	(512)
1357 #define HAMMER2_VOLUME_ICRCVH_SIZE	(65536 - 4)
1358 
1359 #define HAMMER2_VOL_VERSION_MIN		1
1360 #define HAMMER2_VOL_VERSION_DEFAULT	1
1361 #define HAMMER2_VOL_VERSION_WIP 	2
1362 
1363 #define HAMMER2_NUM_VOLHDRS		4
1364 
1365 union hammer2_media_data {
1366 	hammer2_volume_data_t	voldata;
1367         hammer2_inode_data_t    ipdata;
1368 	hammer2_blockset_t	blkset;
1369 	hammer2_blockref_t	npdata[HAMMER2_IND_COUNT_MAX];
1370 	hammer2_bmap_data_t	bmdata[HAMMER2_FREEMAP_COUNT];
1371 	char			buf[HAMMER2_PBUFSIZE];
1372 } __packed;
1373 
1374 typedef union hammer2_media_data hammer2_media_data_t;
1375 
1376 #endif /* !_VFS_HAMMER2_DISK_H_ */
1377