Re: [PATCH v4 00/14] forcealign for xfs

[Dave Chinner <david@fromorbit.com>]

On Tue, Sep 10, 2024 at 06:03:12PM +0530, Ritesh Harjani wrote:
> >> Is it something to just prevent users from destroying their own data
> >> by not allowing a rw mount from an older kernel where users could do
> >> unaligned writes to files marked for atomic writes?
> >> Or is there any other reasoning to prevent XFS filesystem from becoming
> >> inconsistent if an older kernel does a rw mount here.
> >
> > The older kernel does not know what the unknown inode flag means
> > (i.e. atomic writes) and so, by definition, we cannot allow it to
> > modify metadata or file data because it may not modify it in the
> > correct way for that flag being set on the inode.
> >
> > Kernels that don't understand feature flags need to treat the
> > filesystem as read-only, no matter how trivial the feature addition
> > might seem.
> >
> 
> 1. Will it require a fresh formatting of filesystem with mkfs.xfs for
> enabling atomic writes (/forcealign) on XFS?

Initially, yes.

>   a. Is that because reflink is not support with atomic writes
>   (/forcealign) today?

It's much more complex than that.

e.g. How does force-align and COW interact, especially w.r.t.
sub-alloc unit overwrites, cowextsz based preallocation and
unwritten extents in the COW fork?

> As I understand for setting forcealign attr on any inode it checks for
> whether xfs_has_forcealign(mp). That means forcealign can _only_ be
> enabled during mkfs time and it also needs reflink to be disabled with
> -m reflink=0. Right?

forcealign doesn't need to be completely turned off when reflink is
enabled and/or vice versa. Both can co-exist in the filesytsem at
the same time, but the current implementation does not allow
forcealign and reflink to be used on the same inode at the same
time.

It was decided that the best way to handle the lack of reflink
support initially was to make the two feature bits incompatible at
mount time. Hence we currently have to make a non-reflink filesystem
to test forcealign based functionality.

However, doing it this way means that when we fix the implementation
to support reflink and forcealign together, we just remove the mount
time check and all existing reflink filesystems can be immediately
upgraded to support forcealign.

OTOH, we can't do this with atomic writes. Atomic writes require
some mkfs help because they require explicit physical alignment of
the filesystem to the underlying storage. Hence we'll eventually end
up with atomic writes needing to be enabled at mkfs time, but force
align will be an upgradeable feature flag.

-Dave.
-- 
Dave Chinner
david@fromorbit.com