Re: xfs_clear_incompat_log_features considered harmful?

["Darrick J. Wong" <djwong@kernel.org>]

On Sun, Feb 04, 2024 at 10:45:26PM -0800, Christoph Hellwig wrote:
> On Mon, Feb 05, 2024 at 12:09:23PM +1100, Dave Chinner wrote:
> > The issue arises if the host tries to mount the guest VM image to
> > configure the clone of a golden image prior to first start. If there
> > are log incompat fields set in the golden image that was generated
> > by a newer kernel/OS image builder then the provisioning
> > host cannot mount the filesystem even though the log is clean and
> > recovery is unnecessary to mount the filesystem.
> 
> Well, even with the current code base in Darrick's queue a mount alone
> won't upgrade features, you need to do an explicit exchrange or online
> repair operation.  And I think we should basically never do log or
> other format incompatible changes without an explicit user action.

Should I add a flags bit to the ioctls so that programs can force them
on if the process has CAP_SYS_ADMIN?  Or would you rather a mount option
"-o allow_log_upgrades=1" so that's totally under control of whoever
writes fstab?

The first option probably turns into an "and now everyone sets this"
thing; the second one clutters up the mount options.

> The only exception would be if the feature is so old that we finally
> want to get rid of the old implementation, in which case we can think
> of automatically doing the upgrade with a big fat warning.

Heh, we're probably going to have to do that with bigtime come 2035.

> > Hence on unmount we really want the journal contents based log
> > incompat bits cleared because there is nothing incompatible in the
> > log and so there is no reason to prevent older kernels from
> > mounting the filesytsem.
> 
> Doing the clearing at unmount time only (and maybe freeze if someone
> really cares) sounds perfectly fine.

Ugh, I only want to do this at umount time if I can get away with it.
Freeze is already hard enough to grok.

--D