Re: LVM on md0: raid0_make_request bug: can't convert block across chunks or bigger than 64k

[Kevin Corry <kevcorry@us.ibm.com>]

On Monday 20 October 2003 19:09, Neil Brown wrote:
> On Monday October 20, karl.vogel@seagha.com wrote:
> > On Mon, 2003-10-20 at 23:06, Kevin Corry wrote:
> > > On Monday 20 October 2003 15:35, Karl Vogel wrote:
> > > > I'm getting the following kernel messages on V2.6.0-test8-mm1 (I've
> > > > also tried plain -test7 and some kernels before that) when copying
> > > > moderately sized files from a raid-0/LVM volume:
> > > >
> > > > --- snip ---
> > > > raid0_make_request bug: can't convert block across chunks or bigger
> > > > than 64k 24081064 64
> > > > raid0_make_request bug: can't convert block across chunks or bigger
> > > > than 64k 24080656 64
> > > > raid0_make_request bug: can't convert block across chunks or bigger
> > > > than 64k 24080784 64
> > > > raid0_make_request bug: can't convert block across chunks or bigger
> > > > than 64k 24080928 64
> > >
> > > Looks like this was just recently fixed on the linux-raid list.
> > >
> > > http://marc.theaimsgroup.com/?l=linux-raid&m=106661294929434
> >
> > Applied the patch on 2.6.0-test8-mm1 but it made no difference.
>
> no, thats a completely different problem.

Sorry for the confusion. Karl's error messages looked just like ones that were 
reported on the evms-devel list, which were supposedly fixed by the above 
patch.

> The problem is that dm is not honouring the merge_bvec_fn that
> raid0 has set.
>
> This patch might fix it, but I'm not very familiar with the dm code,
> so I make no promises.
>
>  ----------- Diffstat output ------------
>  ./drivers/md/dm-table.c |    5 +++++
>  1 files changed, 5 insertions(+)
>
> diff ./drivers/md/dm-table.c~current~ ./drivers/md/dm-table.c
> --- ./drivers/md/dm-table.c~current~	2003-10-21 10:05:29.000000000 +1000
> +++ ./drivers/md/dm-table.c	2003-10-21 10:06:27.000000000 +1000
> @@ -489,6 +489,11 @@ int dm_get_device(struct dm_target *ti,
>  		rs->max_sectors =
>  			min_not_zero(rs->max_sectors, q->max_sectors);
>
> +		if (q->merge_bvec_fn)
> +			rs->max_sectors =
> +				min_not_zero(rs->max_sectors, PAGE_SIZE>>9);
> +
> +
>  		rs->max_phys_segments =
>  			min_not_zero(rs->max_phys_segments,
>  				     q->max_phys_segments);

This will probably work, as long as raid0 can split a one-page request that 
spans a chunk boundary. I'll be interested to see if this solves Karl's 
problem.

Joe Thornber was telling me about an idea he had to solve the 
Device-Mapper-on-top-of-MD problem using the dm-io code in his latest 
Device-Mapper patchset 
(http://people.sistina.com/~thornber/patches/2.6/2.6.0-test6/2.6.0-t6-mm1-dm2.tar.bz2). 
I believe he's away for this week, but I'll ask him about it again the next 
time I talk to him.

-- 
Kevin Corry
kevcorry@us.ibm.com
http://evms.sourceforge.net/