d/freebsd-nq
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Lock stripes of striped and RAID-5 plexes before writing them. This

Browse Source 
avoids a race condition where multiple RAID-5 subdisks are being
revived at the same time.  The locks should also prevent conflicts
with user requests on concatenated and striped plexes, but this needs
more work.

Tidy up some comments.
This commit is contained in:
Greg Lehey 1999-08-24 02:28:37 +00:00
parent 7b0c9dde22
commit 0911b330a0
1 changed files with 36 additions and 25 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

61
sys/dev/vinum/vinumrevive.c
View File

@ -44,16 +44,14 @@
#include <dev/vinum/request.h>
/*
* revive a block of a subdisk. Return an error
* Revive a block of a subdisk. Return an error
* indication. EAGAIN means successful copy, but
* that more blocks remain to be copied. EINVAL means
* that the subdisk isn't associated with a plex (which
* means a programming error if we get here at all;
* FIXME)
* XXX We should specify a block size here. At the moment,
* just take a default value. FIXME
* that more blocks remain to be copied. EINVAL
* means that the subdisk isn't associated with a
* plex (which means a programming error if we get
* here at all; FIXME).
*/
int
int
revive_block(int sdno)
{
struct sd *sd;
@ -67,9 +65,11 @@ revive_block(int sdno)
int psd; /* parity subdisk number */
int stripe; /* stripe number */
int isparity = 0; /* set if this is the parity stripe */
struct rangelock *lock; /* for locking */
plexblkno = 0; /* to keep the compiler happy */
sd = &SD[sdno];
lock = NULL;
if (sd->plexno < 0) /* no plex? */
return EINVAL;
plex = &PLEX[sd->plexno]; /* point to plex */
@ -100,7 +100,7 @@ revive_block(int sdno)
/*
* Amount to transfer: block size, unless it
* would overlap the end
* would overlap the end
*/
bp->b_bufsize = size;
bp->b_bcount = bp->b_bufsize;
@ -120,6 +120,7 @@ revive_block(int sdno)
plexblkno = sd->plexoffset /* base */
+ (sd->revived - stripeoffset) * plex->subdisks /* offset to beginning of stripe */
+ sd->revived % plex->stripesize; /* offset from beginning of stripe */
lock = lockrange(plexblkno << DEV_BSHIFT, bp, plex); /* lock it */
break;
case plex_raid5:
@ -130,13 +131,16 @@ revive_block(int sdno)
stripe = (sd->revived / plex->stripesize); /* stripe number */
psd = plex->subdisks - 1 - stripe % plex->subdisks; /* parity subdisk for this stripe */
isparity = plex->sdnos[psd] == sdno; /* note if it's the parity subdisk */
/*
* Now adjust for the strangenesses
* in RAID-5 striping
* Now adjust for the strangenesses
* in RAID-5 striping.
*/
if (sd->plexsdno > psd) /* beyond the parity stripe, */
plexblkno -= plex->stripesize; /* one stripe less */
lock = lockrange(plexblkno << DEV_BSHIFT, bp, plex); /* lock it */
break;
case plex_disorg: /* to keep the compiler happy */
}
@ -149,16 +153,17 @@ revive_block(int sdno)
tbuf = (int *) Malloc(size);
isize = size / (sizeof(int)); /* number of ints in the buffer */
/*
* We have calculated plexblkno assuming it
* was a data block. Go back to the beginning
* of the band
* of the band.
*/
plexblkno -= plex->stripesize * sd->plexsdno;
/*
* Read each subdisk in turn, except for
* this one, and xor them together
* Read each subdisk in turn, except for this
* one, and xor them together.
*/
parity_buf = bp->b_data; /* save the buffer getblk gave us */
bzero(parity_buf, size); /* start with nothing */
@ -167,8 +172,9 @@ revive_block(int sdno)
if (mysdno != sdno) { /* not our subdisk */
if (vol != NULL) /* it's part of a volume, */
/*
* First, read the data from the volume. We don't
* care which plex, that's the driver's job
* First, read the data from the volume.
* We don't care which plex, that's the
* driver's job.
*/
bp->b_dev = VINUMBDEV(plex->volno, 0, 0, VINUM_VOLUME_TYPE); /* create the device number */
else /* it's an unattached plex */
@ -182,16 +188,16 @@ revive_block(int sdno)
biowait(bp);
if (bp->b_flags & B_ERROR) /* can't read, */
/*
* If we have a read error, there's nothing
* we can do. By this time, the daemon has
* already run out of magic
* If we have a read error, there's
* nothing we can do. By this time, the
* daemon has already run out of magic.
*/
break;
/*
* To save time, we do the XOR wordwise. This
* requires sectors to be a multiple of the
* length of an int, which is currently always
* the case
* To save time, we do the XOR wordwise.
* This requires sectors to be a multiple
* of the length of an int, which is
* currently always the case.
*/
for (i = 0; i < isize; i++)
((int *) parity_buf)[i] ^= tbuf[i]; /* xor in the buffer */
@ -204,8 +210,8 @@ revive_block(int sdno)
bp->b_blkno = plexblkno; /* start here */
if (vol != NULL) /* it's part of a volume, */
/*
* First, read the data from the volume. We don't
* care which plex, that's bre's job
* First, read the data from the volume. We
* don't care which plex, that's bre's job.
*/
bp->b_dev = VINUMBDEV(plex->volno, 0, 0, VINUM_VOLUME_TYPE); /* create the device number */
else /* it's an unattached plex */
@ -245,6 +251,8 @@ revive_block(int sdno)
error = 0; /* we're done */
}
}
if (lock) /* we took a lock, */
unlockrange(sd->plexno, lock); /* give it back */
while (sd->waitlist) { /* we have waiting requests */
#if VINUMDEBUG
struct request *rq = sd->waitlist;
@ -268,3 +276,6 @@ revive_block(int sdno)
brelse(bp); /* is this kosher? */
return error;
}
/* Local Variables: */
/* fill-column: 50 */
/* End: */