2 * Copyright (c) 1997, 1998, 1999
3 * Nan Yang Computer Services Limited. All rights reserved.
5 * Parts copyright (c) 1997, 1998 Cybernet Corporation, NetMAX project.
7 * Written by Greg Lehey
9 * This software is distributed under the so-called ``Berkeley
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
15 * 1. Redistributions of source code must retain the above copyright
16 * notice, this list of conditions and the following disclaimer.
17 * 2. Redistributions in binary form must reproduce the above copyright
18 * notice, this list of conditions and the following disclaimer in the
19 * documentation and/or other materials provided with the distribution.
20 * 3. All advertising materials mentioning features or use of this software
21 * must display the following acknowledgement:
22 * This product includes software developed by Nan Yang Computer
24 * 4. Neither the name of the Company nor the names of its contributors
25 * may be used to endorse or promote products derived from this software
26 * without specific prior written permission.
28 * This software is provided ``as is'', and any express or implied
29 * warranties, including, but not limited to, the implied warranties of
30 * merchantability and fitness for a particular purpose are disclaimed.
31 * In no event shall the company or contributors be liable for any
32 * direct, indirect, incidental, special, exemplary, or consequential
33 * damages (including, but not limited to, procurement of substitute
34 * goods or services; loss of use, data, or profits; or business
35 * interruption) however caused and on any theory of liability, whether
36 * in contract, strict liability, or tort (including negligence or
37 * otherwise) arising in any way out of the use of this software, even if
38 * advised of the possibility of such damage.
40 * $Id: vinumvar.h,v 1.24 2000/03/01 02:34:57 grog Exp grog $
41 * $FreeBSD: src/sys/dev/vinum/vinumvar.h,v 1.32.2.4 2001/05/28 05:56:27 grog Exp $
42 * $DragonFly: src/sys/dev/raid/vinum/vinumvar.h,v 1.6 2006/02/17 19:18:06 dillon Exp $
46 #include "vinumstate.h"
49 * Some configuration maxima. They're an enum because
50 * we can't define global constants. Sorry about that.
52 * These aren't as bad as they look: most of them are soft limits.
57 VINUM_HEADER = 512, /* size of header on disk */
58 MAXCONFIGLINE = 1024, /* maximum size of a single config line */
59 MINVINUMSLICE = 1048576, /* minimum size of a slice */
61 VINUM_CDEV_MAJOR = 91, /* major number for character device */
63 ROUND_ROBIN_READPOL = -1, /* round robin read policy */
65 /* type field in minor number */
66 VINUM_VOLUME_TYPE = 0,
70 VINUM_SUPERDEV_TYPE = 4, /* super device. */
71 VINUM_RAWPLEX_TYPE = 5, /* anonymous plex */
72 VINUM_RAWSD_TYPE = 6, /* anonymous subdisk */
74 /* Shifts for the individual fields in the device */
75 VINUM_TYPE_SHIFT = 28,
77 VINUM_PLEX_SHIFT = 16,
84 * Shifts for the second half of raw plex and
87 VINUM_RAWPLEX_SHIFT = 8, /* shift the second half this much */
88 VINUM_RAWPLEX_WIDTH = 12, /* width of second half */
92 MAXPLEX = 8, /* maximum number of plexes in a volume */
93 MAXSD = 256, /* maximum number of subdisks in a plex */
94 MAXDRIVENAME = 32, /* maximum length of a device name */
95 MAXSDNAME = 64, /* maximum length of a subdisk name */
96 MAXPLEXNAME = 64, /* maximum length of a plex name */
97 MAXVOLNAME = 64, /* maximum length of a volume name */
98 MAXNAME = 64, /* maximum length of any name */
102 * Define a minor device number.
103 * This is not used directly; instead, it's
104 * called by the other macros.
106 #define VINUMMINOR(v,p,s,t) ( (v << VINUM_VOL_SHIFT) \
107 | (p << VINUM_PLEX_SHIFT) \
108 | (s << VINUM_SD_SHIFT) \
109 | (t << VINUM_TYPE_SHIFT) )
111 /* Create device minor numbers */
115 #define VINUMDEV(v,p,s,t) \
116 make_adhoc_dev (&vinum_cdevsw, VINUMMINOR (v, p, s, t))
118 #define VINUM_PLEX(p) \
119 make_adhoc_dev (&vinum_cdevsw, \
120 (VINUM_RAWPLEX_TYPE << VINUM_TYPE_SHIFT) \
122 | ((p & ~0xff) << 8) )
124 #define VINUM_SD(s) \
125 make_adhoc_dev (&vinum_cdevsw, \
126 (VINUM_RAWSD_TYPE << VINUM_TYPE_SHIFT) \
128 | ((s & ~0xff) << 8) )
132 #define VINUMDEV(v,p,s,t) \
133 makedev(VINUM_CDEV_MAJOR, VINUMMINOR (v, p, s, t))
135 #define VINUM_PLEX(p) \
136 makedev(VINUM_CDEV_MAJOR, \
137 (VINUM_RAWPLEX_TYPE << VINUM_TYPE_SHIFT) \
139 | ((p & ~0xff) << 8) )
141 #define VINUM_SD(s) \
142 makedev(VINUM_CDEV_MAJOR, \
143 (VINUM_RAWSD_TYPE << VINUM_TYPE_SHIFT) \
145 | ((s & ~0xff) << 8) )
149 /* Create a bit mask for x bits */
150 #define MASK(x) ((1 << (x)) - 1)
152 /* Create a raw block device minor number */
153 #define VINUMRMINOR(d,t) ( ((d & MASK (VINUM_VOL_WIDTH)) << VINUM_VOL_SHIFT) \
154 | ((d & ~MASK (VINUM_VOL_WIDTH)) \
155 << (VINUM_PLEX_SHIFT + VINUM_VOL_WIDTH)) \
156 | (t << VINUM_TYPE_SHIFT) )
158 /* extract device type */
159 #define DEVTYPE(x) ((minor (x) >> VINUM_TYPE_SHIFT) & 7)
162 * This mess is used to catch people who compile
163 * a debug vinum(8) and non-debug kernel module,
164 * or the other way round.
168 #define VINUM_SUPERDEV VINUMMINOR (1, 0, 0, VINUM_SUPERDEV_TYPE) /* superdevice number */
169 #define VINUM_WRONGSUPERDEV VINUMMINOR (2, 0, 0, VINUM_SUPERDEV_TYPE) /* non-debug superdevice number */
171 #define VINUM_SUPERDEV VINUMMINOR (2, 0, 0, VINUM_SUPERDEV_TYPE) /* superdevice number */
172 #define VINUM_WRONGSUPERDEV VINUMMINOR (1, 0, 0, VINUM_SUPERDEV_TYPE) /* debug superdevice number */
175 #define VINUM_DAEMON_DEV VINUMMINOR (0, 0, 0, VINUM_SUPERDEV_TYPE) /* daemon superdevice number */
178 * the number of object entries to cater for initially, and also the
179 * value by which they are incremented. It doesn't take long
180 * to extend them, so theoretically we could start with 1 of each, but
181 * it's untidy to allocate such small areas. These values are
182 * probably too small.
188 INITIAL_SUBDISKS = 16,
189 INITIAL_SUBDISKS_IN_PLEX = 4, /* number of subdisks to allocate to a plex */
190 INITIAL_SUBDISKS_IN_DRIVE = 4, /* number of subdisks to allocate to a drive */
191 INITIAL_DRIVE_FREELIST = 16, /* number of entries in drive freelist */
192 PLEX_REGION_TABLE_SIZE = 8, /* number of entries in plex region tables */
193 PLEX_LOCKS = 256, /* number of locks to allocate to a plex */
194 MAX_REVIVE_BLOCKSIZE = MAXPHYS, /* maximum revive block size */
195 DEFAULT_REVIVE_BLOCKSIZE = 65536, /* default revive block size */
196 VINUMHOSTNAMELEN = 32, /* host name field in label */
202 * 31 30 28 27 20 19 18 16 15 8 7 0
203 * |-----------------------------------------------------------------------------------------------|
204 * |X | Type | Subdisk number | X| Plex | Major number | volume number |
205 * |-----------------------------------------------------------------------------------------------|
209 * The fields in the minor number are interpreted as follows:
211 * Volume: Only type and volume number are relevant
212 * Plex in volume: type, plex number in volume and volume number are relevant
213 * raw plex: type, plex number is made of bits 27-16 and 7-0
214 * raw subdisk: type, subdisk number is made of bits 27-16 and 7-0
217 /* This doesn't get used. Consider removing it. */
220 * CARE. These fields assume a big-endian word. On a
221 * little-endian system, they're the wrong way around
223 unsigned volume:8; /* up to 256 volumes */
224 unsigned major:8; /* this is where the major number fits */
225 unsigned plex:3; /* up to 8 plexes per volume */
226 unsigned unused:1; /* up for grabs */
227 unsigned sd:8; /* up to 256 subdisks per plex */
228 unsigned type:3; /* type of object */
238 unsigned signbit:1; /* to make 32 bits */
241 #define VINUM_DIR "/dev/vinum"
244 * These definitions help catch
245 * userland/kernel mismatches.
248 #define VINUM_WRONGSUPERDEV_NAME VINUM_DIR"/control" /* normal super device */
249 #define VINUM_SUPERDEV_NAME VINUM_DIR"/Control" /* debug super device */
251 #define VINUM_WRONGSUPERDEV_NAME VINUM_DIR"/Control" /* debug super device */
252 #define VINUM_SUPERDEV_NAME VINUM_DIR"/control" /* normal super device */
254 #define VINUM_DAEMON_DEV_NAME VINUM_DIR"/controld" /* super device for daemon only */
257 * Flags for all objects. Most of them only apply to
258 * specific objects, but we have space for all in any
262 VF_LOCKED = 1, /* somebody has locked access to this object */
263 VF_LOCKING = 2, /* we want access to this object */
264 VF_OPEN = 4, /* object has openers */
265 VF_WRITETHROUGH = 8, /* volume: write through */
266 VF_INITED = 0x10, /* unit has been initialized */
267 VF_WLABEL = 0x20, /* label area is writable */
268 VF_LABELLING = 0x40, /* unit is currently being labelled */
269 VF_WANTED = 0x80, /* someone is waiting to obtain a lock */
270 VF_RAW = 0x100, /* raw volume (no file system) */
271 VF_LOADED = 0x200, /* module is loaded */
272 VF_CONFIGURING = 0x400, /* somebody is changing the config */
273 VF_WILL_CONFIGURE = 0x800, /* somebody wants to change the config */
274 VF_CONFIG_INCOMPLETE = 0x1000, /* haven't finished changing the config */
275 VF_CONFIG_SETUPSTATE = 0x2000, /* set a volume up if all plexes are empty */
276 VF_READING_CONFIG = 0x4000, /* we're reading config database from disk */
277 VF_FORCECONFIG = 0x8000, /* configure drives even with different names */
278 VF_NEWBORN = 0x10000, /* for objects: we've just created it */
279 VF_CONFIGURED = 0x20000, /* for drives: we read the config */
280 VF_STOPPING = 0x40000, /* for vinum_conf: stop on last close */
281 VF_DAEMONOPEN = 0x80000, /* the daemon has us open (only superdev) */
282 VF_CREATED = 0x100000, /* for volumes: freshly created, more then new */
283 VF_HOTSPARE = 0x200000, /* for drives: use as hot spare */
284 VF_RETRYERRORS = 0x400000, /* don't down subdisks on I/O errors */
287 /* Global configuration information for the vinum subsystem */
289 /* Pointers to vinum structures */
293 struct volume *volume;
295 /* the number allocated */
296 int drives_allocated;
297 int subdisks_allocated;
298 int plexes_allocated;
299 int volumes_allocated;
301 /* and the number currently in use */
309 #define VINUM_MAXACTIVE 30000 /* maximum number of active requests */
310 int active; /* current number of requests outstanding */
311 int maxactive; /* maximum number of requests ever outstanding */
313 struct request *lastrq;
318 /* Use these defines to simplify code */
319 #define DRIVE vinum_conf.drive
320 #define SD vinum_conf.sd
321 #define PLEX vinum_conf.plex
322 #define VOL vinum_conf.volume
323 #define VFLAGS vinum_conf.flags
328 * Vinum drives start with this structure:
331 * |--------------------------------------|
332 * | PDP-11 memorial boot block | 0
333 * |--------------------------------------|
334 * | Disk label, maybe | 1
335 * |--------------------------------------|
336 * | Slice definition (vinum_hdr) | 8
337 * |--------------------------------------|
339 * | Configuration info, first copy | 9
341 * |--------------------------------------|
343 * | Configuration info, second copy | 9 + size of config
345 * |--------------------------------------|
348 /* Sizes and offsets of our information */
350 VINUM_LABEL_OFFSET = 4096, /* offset of vinum label */
351 VINUMHEADERLEN = 512, /* size of vinum label */
352 VINUM_CONFIG_OFFSET = 4608, /* offset of first config copy */
353 MAXCONFIG = 65536, /* and size of config copy */
354 DATASTART = (MAXCONFIG * 2 + VINUM_CONFIG_OFFSET) / DEV_BSIZE /* this is where the data starts */
358 * hostname is 256 bytes long, but we don't need to shlep
359 * multiple copies in vinum. We use the host name just
360 * to identify this system, and 32 bytes should be ample
365 char sysname[VINUMHOSTNAMELEN]; /* system name at time of creation */
366 char name[MAXDRIVENAME]; /* our name of the drive */
367 struct timeval date_of_birth; /* the time it was created */
368 struct timeval last_update; /* and the time of last update */
370 * total size in bytes of the drive. This value
371 * includes the headers.
377 uint64_t magic; /* we're long on magic numbers */
378 #define VINUM_MAGIC 22322600044678729LL /* should be this */
379 #define VINUM_NOMAGIC 22322600044678990LL /* becomes this after obliteration */
381 * Size in bytes of each copy of the
382 * configuration info. This must be a multiple
383 * of the sector size.
386 struct vinum_label label; /* unique label */
389 /* Information returned from read_drive_label */
390 enum drive_label_info {
391 DL_CANT_OPEN, /* invalid partition */
392 DL_NOT_OURS, /* valid partition, but no vinum label */
393 DL_DELETED_LABEL, /* valid partition, deleted label found */
394 DL_WRONG_DRIVE, /* drive name doesn't match */
395 DL_OURS /* valid partition and label found */
398 /*** Drive definitions ***/
400 * A drive corresponds to a disk slice. We use a different term to show
401 * the difference in usage: it doesn't have to be a slice, and could
402 * theoretically be a complete, unpartitioned disk
406 char devicename[MAXDRIVENAME]; /* name of the slice it's on */
407 enum drivestate state; /* current state */
408 int flags; /* flags */
409 int subdisks_allocated; /* number of entries in sd */
410 int subdisks_used; /* and the number used */
411 int blocksize; /* size of fs blocks */
412 int pid; /* of locker */
413 u_int64_t sectors_available; /* number of sectors still available */
415 int lasterror; /* last error on drive */
416 int driveno; /* index of drive in vinum_conf */
417 int opencount; /* number of up subdisks */
418 u_int64_t reads; /* number of reads on this drive */
419 u_int64_t writes; /* number of writes on this drive */
420 u_int64_t bytes_read; /* number of bytes read */
421 u_int64_t bytes_written; /* number of bytes written */
422 struct vinum_label label; /* and the label information */
423 #define DRIVE_MAXACTIVE 30000 /* maximum number of active requests */
424 int active; /* current number of requests outstanding */
425 int maxactive; /* maximum number of requests ever outstanding */
426 int freelist_size; /* number of entries alloced in free list */
427 int freelist_entries; /* number of entries used in free list */
428 struct drive_freelist { /* sorted list of free space on drive */
429 u_int64_t offset; /* offset of entry */
430 u_int64_t sectors; /* and length in sectors */
432 struct partinfo partinfo; /* partition information */
433 /* XXX kludge until we get this struct cleaned up */
435 dev_t dev; /* device information */
437 char dev [sizeof (int *)];
440 char lockfilename[16]; /* name of file from which we were locked */
441 int lockline; /* and the line number */
445 /*** Subdisk definitions ***/
448 char name[MAXSDNAME]; /* name of subdisk */
449 enum sdstate state; /* state */
451 int lasterror; /* last error occurred */
452 /* offsets in blocks */
453 int64_t driveoffset; /* offset on drive */
455 * plexoffset is the offset from the beginning
456 * of the plex to the very first part of the
457 * subdisk, in sectors. For striped, RAID-4 and
458 * RAID-5 plexes, only the first stripe is
459 * located at this offset
461 int64_t plexoffset; /* offset in plex */
462 u_int64_t sectors; /* and length in sectors */
463 int plexno; /* index of plex, if it belongs */
464 int driveno; /* index of the drive on which it is located */
465 int sdno; /* our index in vinum_conf */
466 int plexsdno; /* and our number in our plex */
467 /* (undefined if no plex) */
468 u_int64_t reads; /* number of reads on this subdisk */
469 u_int64_t writes; /* number of writes on this subdisk */
470 u_int64_t bytes_read; /* number of bytes read */
471 u_int64_t bytes_written; /* number of bytes written */
472 /* revive parameters */
473 u_int64_t revived; /* block number of current revive request */
474 int revive_blocksize; /* revive block size (bytes) */
475 int revive_interval; /* and time to wait between transfers */
476 pid_t reviver; /* PID of reviving process */
477 /* init parameters */
478 u_int64_t initialized; /* block number of current init request */
479 int init_blocksize; /* init block size (bytes) */
480 int init_interval; /* and time to wait between transfers */
481 struct request *waitlist; /* list of requests waiting on revive op */
484 /*** Plex definitions ***/
486 /* kinds of plex organization */
488 plex_disorg, /* disorganized */
489 plex_concat, /* concatenated plex */
490 plex_striped, /* striped plex */
491 plex_raid4, /* RAID4 plex */
492 plex_raid5 /* RAID5 plex */
495 /* Recognize plex organizations */
496 #define isstriped(p) (p->organization >= plex_striped) /* RAID 1, 4 or 5 */
497 #define isparity(p) (p->organization >= plex_raid4) /* RAID 4 or 5 */
500 char name[MAXPLEXNAME]; /* name of plex */
501 enum plexorg organization; /* Plex organization */
502 enum plexstate state; /* and current state */
503 u_int64_t length; /* total length of plex (sectors) */
505 int stripesize; /* size of stripe or raid band, in sectors */
506 int subdisks; /* number of associated subdisks */
507 int subdisks_allocated; /* number of subdisks allocated space for */
508 int *sdnos; /* list of component subdisks */
509 int plexno; /* index of plex in vinum_conf */
510 int volno; /* index of volume */
511 int volplexno; /* number of plex in volume */
513 u_int64_t reads; /* number of reads on this plex */
514 u_int64_t writes; /* number of writes on this plex */
515 u_int64_t bytes_read; /* number of bytes read */
516 u_int64_t bytes_written; /* number of bytes written */
517 u_int64_t recovered_reads; /* number of recovered read operations */
518 u_int64_t degraded_writes; /* number of degraded writes */
519 u_int64_t parityless_writes; /* number of parityless writes */
520 u_int64_t multiblock; /* requests that needed more than one block */
521 u_int64_t multistripe; /* requests that needed more than one stripe */
522 int sddowncount; /* number of subdisks down */
523 /* Lock information */
524 int usedlocks; /* number currently in use */
525 int lockwaits; /* and number of waits for locks */
526 off_t checkblock; /* block number for parity op */
527 struct rangelock *lock; /* ranges of locked addresses */
530 /*** Volume definitions ***/
532 /* Address range definitions, for locking volumes */
534 daddr_t stripe; /* address + 1 of the range being locked */
535 struct buf *bp; /* user's buffer pointer */
539 char name[MAXVOLNAME]; /* name of volume */
540 enum volumestate state; /* current state */
541 int plexes; /* number of plexes */
542 int preferred_plex; /* plex to read from, -1 for round-robin */
544 * index of plex used for last read, for
548 int volno; /* volume number */
549 int flags; /* status and configuration flags */
550 int openflags; /* flags supplied to last open(2) */
551 u_int64_t size; /* size of volume */
552 int blocksize; /* logical block size */
553 int active; /* number of outstanding requests active */
554 int subops; /* and the number of suboperations */
556 u_int64_t bytes_read; /* number of bytes read */
557 u_int64_t bytes_written; /* number of bytes written */
558 u_int64_t reads; /* number of reads on this volume */
559 u_int64_t writes; /* number of writes on this volume */
560 u_int64_t recovered_reads; /* reads recovered from another plex */
562 * Unlike subdisks in the plex, space for the
563 * plex pointers is static.
565 int plex[MAXPLEX]; /* index of plexes */
566 struct disklabel label; /* for DIOCGPART */
570 * Table expansion. Expand table, which contains oldcount
571 * entries of type element, by increment entries, and change
572 * oldcount accordingly
574 #define EXPAND(table, element, oldcount, increment) \
576 expand_table ((void **) &table, \
577 oldcount * sizeof (element), \
578 (oldcount + increment) * sizeof (element) ); \
579 oldcount += increment; \
582 /* Information on vinum's memory usage */
584 int mallocs; /* number of malloced blocks */
585 int total_malloced; /* total amount malloced */
586 int highwater; /* maximum number of mallocs */
587 struct mc *malloced; /* pointer to kernel table */
590 #define MCFILENAMELEN 16
597 char file[MCFILENAMELEN];
601 * These enums are used by the state transition
602 * routines. They're in bit map format:
604 * Bit 0: Other plexes in the volume are down
605 * Bit 1: Other plexes in the volume are up
606 * Bit 2: The current plex is up
607 * Maybe they should be local to
611 volplex_onlyusdown = 0, /* 0: we're the only plex, and we're down */
612 volplex_alldown, /* 1: another plex is down, and so are we */
613 volplex_otherup, /* 2: another plex is up */
614 volplex_otherupdown, /* 3: other plexes are up and down */
615 volplex_onlyus, /* 4: we're up and alone */
616 volplex_onlyusup, /* 5: only we are up, others are down */
617 volplex_allup, /* 6: all plexes are up */
618 volplex_someup /* 7: some plexes are up, including us */
621 /* state map for plex */
624 sd_downstate = 2, /* SD is down */
625 sd_crashedstate = 4, /* SD is crashed */
626 sd_obsoletestate = 8, /* SD is obsolete */
627 sd_stalestate = 16, /* SD is stale */
628 sd_rebornstate = 32, /* SD is reborn */
629 sd_upstate = 64, /* SD is up */
630 sd_initstate = 128, /* SD is initializing */
631 sd_initializedstate = 256, /* SD is initialized */
632 sd_otherstate = 512, /* SD is in some other state */
636 * This is really just a parameter to pass to
637 * set_<foo>_state, but since it needs to be known
638 * in the external definitions, we need to define
642 setstate_none = 0, /* no flags */
643 setstate_force = 1, /* force the state change */
644 setstate_configuring = 2, /* we're currently configuring, don't save */
647 /* Operations for parityops to perform. */
651 rebuildandcheckparity, /* rebuildparity with the -v option */
655 /* Debugging stuff */
657 DEBUG_ADDRESSES = 1, /* show buffer information during requests */
658 DEBUG_NUMOUTPUT = 2, /* show the value of vp->v_numoutput */
659 DEBUG_RESID = 4, /* go into debugger in complete_rqe */
660 DEBUG_LASTREQS = 8, /* keep a circular buffer of last requests */
661 DEBUG_REVIVECONFLICT = 16, /* print info about revive conflicts */
662 DEBUG_EOFINFO = 32, /* print info about EOF detection */
663 DEBUG_MEMFREE = 64, /* keep info about Frees */
664 DEBUG_BIGDRIVE = 128, /* pretend our drives are 100 times the size */
665 DEBUG_REMOTEGDB = 256, /* go into remote gdb */
666 DEBUG_WARNINGS = 512, /* log various relatively harmless warnings */
671 #define longjmp LongJmp /* test our longjmps */
675 /* Local Variables: */
676 /* fill-column: 50 */