XFS and 4k stacks/newer kernels.

XFS and 4k stacks/newer kernels.

[Mike Ashton]

Hi folks,

Firstly, I must apologise for the tremendously vague nature of this
"report".  I do hope it's of some use, and it's the best I can do at
the moment.

I've been using XFS for several years now on a raid6 partition on a
15-disk array on linux 2.6.19 without 4k stacks, and without incident.
And a very nice filesystem it is too.

I've just set up a new one based on exactly the same hardware but with
a 2.6.22.3 kernel (and thus with compulsory 4k stacks, as I understand
it?).  During a large rsync to populate the filesystem, the write to
local disk blocked in a fast device wait, but without any errors in
dmesg.  After giving it half an hour to wake up again, I hard rebooted
and the XFS superblock was unreadable; bye bye filesystem.

So I've rolled back to 2.6.19, unsurprisingly, but I wanted to let you
know that some issues with 4k stacks or other features of newer
kernels may not have been resolved yet.  Since there were no errors,
I've nothing specific to give you, which I appreciate gives you little
to go on.  It wouldn't be right to leave this anecdote unreported,
however.

Best of luck and thanks,
Mike.

Re: XFS and 4k stacks/newer kernels.

[Chris Wedgwood]

On Fri, Aug 24, 2007 at 11:38:47AM +0100, Mike Ashton wrote:

> I've just set up a new one based on exactly the same hardware but
> with a 2.6.22.3 kernel (and thus with compulsory 4k stacks, as I
> understand it?).

Not when I tested here.  Turn off 4K stacks.

> After giving it half an hour to wake up again, I hard rebooted and
> the XFS superblock was unreadable; bye bye filesystem.

Did xfs_repair fix the damage?

> So I've rolled back to 2.6.19, unsurprisingly, but I wanted to let
> you know that some issues with 4k stacks or other features of newer
> kernels may not have been resolved yet.

XFS on i386 with 4K stacks isn't safe.  People are working on making
it better but it's a bit of an uphill battle in some ways.

I've posted patches in the past to change of some the config logic to
avoid this but none of them ever got much more than a grunt ora snarky
anti-XFS comment, it's probably time to do it again.