Re: [dm-devel] hunt for 2.6.37 dm-crypt+ext4 corruption? (was: Re: dm-crypt barrier support is effective)

[Jon Nelson <jnelson jamponi net>]

On Sat, Dec 11, 2010 at 7:40 PM, Ted Ts'o <tytso mit edu> wrote:
> Yes, indeed.  Is this in the virtualized environment or on real
> hardware at this point?  And how many CPU's do you have configured in
> your virtualized environment, and how memory memory?  Is having a
> certain number of CPU's critical for reproducing the problem?  Is
> constricting the amount of memory important?

Originally, I observed the behavior on really real hardware.

Since then, I have been able to reproduce it in VirtualBox and
qemu-kvm, with openSUSE 11.3 and KUbuntu. All of the more recent tests
have been with qemu-kvm.

I have one CPU configured in the environment, 512MB of memory.
I have not done any memory-constriction tests whatsoever.

> It'll be a lot easier if I can reproduce it locally, which is why I'm
> asking all of these questions.

On Sat, Dec 11, 2010 at 8:34 PM, Ted Ts'o <tytso mit edu> wrote:
> One experiment --- can you try this with the file system mounted with
> data=writeback, and see if the problem reproduces in that journalling
> mode?

That test is running now, first with encryption. I will report if it
shows problems. If it does, I will wait until I have been able to see
that a few times, and move to a no-encryption test. Typically, I have
to run quite a few more iterations of that test before problems show
up (if they will at all).

> I want to rule out (if possible) journal_submit_inode_data_buffers()
> racing with mpage_da_submit_io().  I don't think that's the issue, but
> I'd prefer to do the experiment to make sure.  So if you can use a
> kernel and system configuration which triggers the problem, and then
> try changing the mount options to include data=writeback, and then
> rerun the test, and let me know if the problem still reproduces, I'd
> be really grateful.


-- 
Jon