Re: [PATCH 2/6] xfs: also mark the buffer stale on verifier failure in xfs_buf_submit

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

On Fri, Jun 26, 2026 at 06:35:16AM +0200, Christoph Hellwig wrote:
> On Thu, Jun 25, 2026 at 11:00:02AM -0700, Darrick J. Wong wrote:
> > > -	if (bp->b_mount->m_log && xlog_is_shutdown(bp->b_mount->m_log))
> > > +	if (bp->b_mount->m_log && xlog_is_shutdown(bp->b_mount->m_log)) {
> > > +		xfs_buf_ioerror(bp, -EIO);
> > >  		goto ioerror;
> > 
> > IDGI.  If the filesystem was already shut down, we set EIO on the
> > buffer, stale it, and complete the ioend...
> > 
> > > +	}
> > >  
> > >  	if (bp->b_flags & XBF_WRITE)
> > >  		xfs_buf_wait_unpin(bp);
> > > @@ -1397,7 +1399,7 @@ xfs_buf_submit(
> > >  
> > >  	if ((bp->b_flags & XBF_WRITE) && !xfs_buf_verify_write(bp)) {
> > >  		xfs_force_shutdown(bp->b_mount, SHUTDOWN_CORRUPT_INCORE);
> > > -		goto end_io;
> > > +		goto ioerror;
> > 
> > ...but if bp itself causes the shutdown, we now leave b_error at 0, but
> > do the same stale-and-complete thing?  Shouldn't we set EIO on the
> > buffer in both cases?
> 
> ->verify_write is expected to set b_error when it fails (and yes,
> the API in that area is awful).

Would you mind leaving a comment here, then?

	if ((bp->b_flags & XBF_WRITE) && !xfs_buf_verify_write(bp)) {
		/* ->verify_write should have set b_error already */
		xfs_force_shutdown(...);
		goto out_ioerror;
	}

--D