Re: multiple writes of same block

[Steve Lord <lord@sgi.com>]

If someone is experiencing this can they please do the following:

	o add a call to BUG() after the printk in md, line 786 in 
	  drivers/block/md/raid5.c
	o build the kernel with kdb enabled
	o do an insmod xfsidbg and insmod kdbm_pg
	o run until the system crashes.
	o use the bt command in kdb to report the stack trace for
	  this process.
	o look at the second argument to add_stripe_bh, use the bh command
	  on it:
	   bh 0xaaaaaaaaaa

	o send us the output.

	It would also be useful to add another printk after line 786 to
	look something like this:

	printk(KERN_NOTICE "raid5: new bh at blk 0x%x len 0x%x, existing blk 
0x%x len 0x%x\n",
		bh->b_blocknr, bh->b_size,
		(*bhp)->b_blocknr, (*bhp)->b_size);

You can do the latter without all the kdb and panicing the system stuff,
it might be a good starting point.

Steve


> Using XFS (latest 2.4.9-xfs) over raid5 I got a message:
> 	raid5: multiple 1 requests for sector 24663824
> and the explanation I got from the raid list is:
> 
> > It means that while raid5 had an outstanding write request on a
> > particular sector, it received another write request for the same
> > sector.
> > 
> > It trys to do the right thing and write them both out in the order
> > that it received them, but it is a bit of a worry that any filesystem
> > would do this.  I'm guessing that Andrew is using XFS too.  Is that
> > right?
> > 
> > While raid5 tries to keep the requests in order, and I suspect other
> > drivers do to, I don't think that it is reasonable to assume that no
> > device driver will ever re-order two requests for the same sector.
> > If I were the author of the filesystem I would be worried.
> > 
> > NeilBrown
> 
> This truely sounds worrying. Is this OK to be this way or is this
> an issue that needs fixing?
> 
> --
> Eyal Lebedinsky (eyal@eyal.emu.id.au) <http://samba.anu.edu.au/eyal/>