Re: BUG: ext3 corruption in domU

[Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>]

On Thu, May 23, 2013 at 02:19:50PM -0400, Anthony Sheetz wrote:
> On Wed, May 22, 2013 at 4:10 PM, Konrad Rzeszutek Wilk
> <konrad.wilk@oracle.com> wrote:
> > On Mon, Apr 22, 2013 at 01:26:34PM +0100, Ian Campbell wrote:
> >> Konrad is on vacation this week, so it'll probably be next week before
> >> this gets looked at by him.
> >
> > And I finally got to this email in my 'vacation-mbox'
> >>
> >> Ian.
> >>
> >> On Mon, 2013-04-22 at 13:22 +0100, Anthony Sheetz wrote:
> >> > I realize folks are pretty busy, but we're still interested in getting
> >> > this problem solved, and I want to be sure it's not lost in the
> >> > shuffle.
> >> > Any chance of getting some attention for it?
> >> >
> >> > On Wed, Apr 17, 2013 at 9:00 AM, Ian Campbell <Ian.Campbell@citrix.com> wrote:
> >> > > On Tue, 2013-04-16 at 18:39 +0100, Anthony Sheetz wrote:
> >> > >> (re-sending, first message seems to have gotten lost)
> >> > >>
> >> > >> I was referred here by Ian Campbell ijc@hellion.org.uk from bugs.debian.org.
> >> > >
> >> > > I'm here too (different hat ;-)), thanks for posting it here. I've added
> >> > > some people who know about the block stuff to the CC.
> >> > >
> >> > > Guys, my suspicion is that the issue is that barriers issued by ext3
> >> > > inside the guest aren't making it all the way down the
> >> > > ext3->blkfront->blkback->lvm->dm-crypt->disk chain leading the
> >> > > filesystem to eventually corrupt itself.
> >> > >
> >> > > The issue seems to relate to the use of dm-crypt since
> >> > > ext3->blkfront->blkback->lvm->disk is reported work fine.
> >> > >
> >> > > However there is no problem with the local dom0 ext3 root filesystem
> >> > > which is also in the same lvm VG on the crypt device (i.e.
> >> > > ext3->lvm->dm-crypt->disk), so its not purely a dm-crypt issue. I figure
> >> > > something is up at the blkfront->back link which causes the barriers
> >> > > which blkback is injecting into the block subsystem either don't make it
> >> > > to the dm-crypt layer or do not DTRT once they arrive.
> >> > >
> >> > > I'm not really sure with how to proceed (or how to ask Anthony to
> >> > > proceed) with verifying any part of that hypothesis though.
> >> > >
> >> > > ISTR issues with old vs new style barriers or barriers with no data in
> >> > > them or something, could this be related to that? (or am I thinking of
> >> > > DISCARD?)
> >
> > You are using two different kernel versions. The 2.6.32 domU is only using
> > WRITE_BARRIERs, while in the 3.2 kernels that have been completly eliminated.
> > The mechanism they use is called 'WRITE_FLUSH'. The 3.2 kernel has a patch:
> > ommit 29bde093787f3bdf7b9b4270ada6be7c8076e36b
> > Author: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
> > Date:   Mon Oct 10 00:42:22 2011 -0400
> >
> >     xen/blkback: Support 'feature-barrier' aka old-style BARRIER requests.
> >
> >
> > which emulates the barrier request by draining all of the oustanding I/Os and then
> > sending the WRITE_FLUSH.
> >
> > But it looks like you are hitting an issue here. Just to make sure
> > that is the case, what happens if you use the _same_ kernel in both dom0 and
> > domU? Does it work then?
> >
> 
> First, thank you so much for getting back to me, it's really appreciated.
> At this point I've forgotten if I did this with Wheezy on Wheezy, and
> what the result was.
> I'll have to test using the 3.2 kernel on the domU Debian Squeeze and
> get back to you. I should be able to do that early next week.

Thank you. Also when you do this test, could you also provide the 'xenstore-ls'
output from dom0? And the 'dmesg' output from the guest (or at least
the 'xl console <guest> | tee /tmp/log' ? That would give me and idea if
the frontend/backend have the right negotiation parameters.

Have a good weekend!