Re: DM-CRYPT: Scale to multiple CPUs v3 on 2.6.37-rc* ?

[Chris Mason <chris.mason@oracle.com>]

Excerpts from Mike Snitzer's message of 2010-11-08 09:58:09 -0500:
> On Sun, Nov 07 2010 at  6:05pm -0500,
> Andi Kleen <andi@firstfloor.org> wrote:
> 
> > On Sun, Nov 07, 2010 at 10:39:23PM +0100, Milan Broz wrote:
> > > On 11/07/2010 08:45 PM, Andi Kleen wrote:
> > > >> I read about barrier-problems and data getting to the partition when
> > > >> using dm-crypt and several layers so I don't know if that could be
> > > >> related
> > > > 
> > > > Barriers seem to be totally broken on dm-crypt currently.
> > > 
> > > Can you explain it?
> > 
> > e.g. the btrfs mailing list is full of corruption reports
> > on dm-crypt and most of the symptoms point to broken barriers.
> 
> [cc'ing linux-btrfs, hopefully in the future dm-devel will get cc'd when
> concerns about DM come up on linux-btrfs (or other lists)]
> 
> I spoke with Josef Bacik and these corruption reports are apparently
> against older kernels (e.g. <= 2.6.33).  I say <= 2.6.33 because:

We've consistently seen reports about corruptions on power hits with
dm-crypt.  The logs didn't have any messages about barriers failing, but
the corruptions were still there.  The most likely cause is that
barriers just aren't getting through somehow.

> 
> https://btrfs.wiki.kernel.org/index.php/Gotchas states:
> "btrfs volumes on top of dm-crypt block devices (and possibly LVM)
> require write-caching to be turned off on the underlying HDD. Failing to
> do so, in the event of a power failure, may result in corruption not yet
> handled by btrfs code. (2.6.33)"
> 
> But Josef was not aware of any reports with kernels newer than 2.6.32
> (F12).
> 
> Josef also noted that until last week btrfs wouldn't retry another
> mirror in the face of some corruption, the fix is here:
> http://git.kernel.org/?p=linux/kernel/git/mason/btrfs-unstable.git;a=commit;h=cb44921a09221
> 
> This obviously doesn't fix any source of corruption but it makes btrfs
> more resilient when it encounters the corruption.

Right.

> 
> > > Barriers/flush change should work, if it is broken, it is not only dm-crypt.
> > > (dm-crypt simply relies on dm-core implementation, when barrier/flush
> > > request come to dmcrypt, all previous IO must be already finished).
> > 
> > Possibly, at least it doesn't seem to work.
> 
> Can you please be more specific?  What test(s)?  What kernel(s)?
> 
> Any pointers to previous (and preferably: recent) reports would be
> appreciated.
> 
> The DM barrier code has seen considerable change recently (via flush+fua
> changes in 2.6.37).  Those changes have been tested quite a bit
> (including ext4 consistency after a crash).
> 
> But even prior to those flush+fua changes DM's support for barriers
> (Linux >= 2.6.31) was held to be robust.  No known (at least no
> reported) issues with DM's barrier support.

I think it would be best to move forward with just hammering on the
dm-crypt barriers:

http://oss.oracle.com/~mason/barrier-test

This script is the best I've found so far to reliably trigger
corruptions with barriers off.  I'd start with ext3 + barriers off just
to prove it corrupts things, then move to ext3 + barriers on.

-chris