[Gluster-users] KVM guest I/O errors with xfs backed gluster volumes

[Bharata B Rao]

My below mail didn't make it to the list, hence resending...

On Tue, Nov 5, 2013 at 8:04 PM, Bharata B Rao <bharata at linux.vnet.ibm.com>wrote:

> On Wed, Oct 30, 2013 at 11:26:48PM +0530, Bharata B Rao wrote:
> > On Tue, Oct 29, 2013 at 1:21 PM, Anand Avati <avati at gluster.org> wrote:
> >
> > > Looks like what is happening is that qemu performs ioctls() on the
> backend
> > > to query logical_block_size (for direct IO alignment). That works on
> XFS,
> > > but fails on FUSE (hence qemu ends up performing IO with default 512
> > > alignment rather than 4k).
> > >
> > > Looks like this might be something we can enhance gluster driver in
> qemu.
> > > Note that glusterfs does not have an ioctl() FOP, but we could probably
> > > wire up a virtual xattr call for this purpose.
> > >
> > > Copying Bharata to check if he has other solutions in mind.
> > >
> >
> > I see alignment issues and subsequent QEMU failure (pread() failing with
> > EINVAL) when I use a file from XFS mount point (with sectsz=4k) as a
> virtio
> > disk with cache=none QEMU option. However this failure isn't seen when I
> > have sectsz=512. And all this is w/o gluster. So there seems to be some
> > alignment issues even w/o gluster, I will debug more and get back.
>
> I gather that QEMU block layer and SeaBIOS don't yet support 4k sectors.
> So this is not a QEMU-GlusterFS specific issue.
>
> You could either not use cache=none option which results in O_DIRECT
> or use the below something like below which explicitly sets the sector size
> and min io size for the guest.
>
> -drive file=/mnt/xfs.img,if=none,cache=none,format=raw,id=mydisk -device
> virtio-blk,drive=mydisk,logical_block_size=4096,physical_block_size=4096,min_io_size=4096
>
> Ref: https://bugzilla.redhat.com/show_bug.cgi?id=997839
>
> Regards,
> Bharata.
>
>


-- 
http://raobharata.wordpress.com/
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://supercolony.gluster.org/pipermail/gluster-users/attachments/20131106/02664a98/attachment.html>