Re: Reproducable OOPS with MD RAID-5 on 2.6.0-test11

["Kevin P. Fleming" <kpfleming@backtobasicsmgmt.com>]

Nathan Scott wrote:

> One thing that might be of interest - XFS does tend to pass
> variable size requests down to the block layer, and this has
> tripped up md and other drivers in 2.4 in the distant past.
> 
> Log IO is typically 512 byte aligned (as opposed to block or
> page size aligned), as are IOs into several of XFS' metadata
> structures.

Hey, thanks for the pointer! I think we're getting somewhere now. Here's 
a recap of the tested combinations:

XFS on raw disk: OK
XFS on LVM2 on single disk: OK
XFS on LVM2 on RAID-5: fails
ext2 on LVM2 on RAID-5: OK

I just tested XFS on LVM2 on RAID-5 using "-l sunit=8" while creating 
the filesystem to force log writes be block-sized and block-aligned; 
this seems to work :-) I have not been able to force a failure using my 
test script, although ATM the system is still running a RAID-5 resync of 
the array, but that should only make the problem more likely, not less.

So, this does appear to be an md/dm stacking problem, that is exposed by 
XFS sending non-block-sized and/or non-block-aligned IOs.