Re: [PATCH 2/3] xfs: allow sparse inode records at the end of runt AGs

[Dave Chinner <david@fromorbit.com>]

On Thu, Oct 24, 2024 at 10:00:38AM -0700, Darrick J. Wong wrote:
> On Thu, Oct 24, 2024 at 01:51:04PM +1100, Dave Chinner wrote:
> > From: Dave Chinner <dchinner@redhat.com>
> > 
> > Due to the failure to correctly limit sparse inode chunk allocation
> > in runt AGs, we now have many production filesystems with sparse
> > inode chunks allocated across the end of the runt AG. xfs_repair
> > or a growfs is needed to fix this situation, neither of which are
> > particularly appealing.
> > 
> > The on disk layout from the metadump shows AG 12 as a runt that is
> > 1031 blocks in length and the last inode chunk allocated on disk at
> > agino 8192.
> 
> Does this problem also happen on non-runt AGs?

No. The highest agbno an inode chunk can be allocated at in a full
size AG is aligned by rounding down from sb_agblocks.  Hence
sb_agblocks can be unaligned and nothing will go wrong. The problem
is purely that the runt AG being shorter than sb_agblocks and so
this highest agbno allocation guard is set beyond the end of the
AG...

> If the only free space
> that could be turned into a sparse cluster is unaligned space at the
> end of AG 0, would you still get the same corruption error?

It will only happen if AG 0 is a runt AG, and then the same error
would occur. We don't currently allow single AG filesystems, nor
when they are set up  do we create them as a runt - the are always
full size. So current single AG filesystems made by mkfs won't have
this problem.

That said, the proposed single AG cloud image filesystems that set
AG 0 up as a runt (i.e. dblocks smaller than sb_agblocks) to allow
the AG 0 size to grow along with the size of the filesystem could
definitely have this problem. i.e. sb_dblocks needs to be inode
chunk aligned in this sort of setup, or these filesystems need to
be restricted to fixed kernels....

-Dave.

-- 
Dave Chinner
david@fromorbit.com