Re: [PATCH v9 05/15] xfs: ignore HW which cannot atomic write a single block

[Darrick J. Wong]

On Thu, May 01, 2025 at 09:22:16AM -0700, Darrick J. Wong wrote:
> On Wed, Apr 30, 2025 at 02:59:06PM +0200, Christoph Hellwig wrote:
> > On Tue, Apr 29, 2025 at 07:44:46AM -0700, Darrick J. Wong wrote:
> > > > So this can't be merged into xfs_setsize_buftarg as suggeted last round
> > > > instead of needing yet another per-device call into the buftarg code?
> > > 
> > > Oh, heh, I forgot that xfs_setsize_buftarg is called a second time by
> > > xfs_setup_devices at the end of fill_super.
> > 
> > That's actually the real call.  The first is just a dummy to have
> > bt_meta_sectorsize/bt_meta_sectormask initialized because if we didn't
> > do that some assert in the block layer triggered.  We should probably
> > remove that call and open code the two assignments..
> > 
> > > I don't like the idea of merging the hw atomic write detection into
> > > xfs_setsize_buftarg itself because (a) it gets called for the data
> > > device before we've read the fs blocksize so the validation is
> > > meaningless and (b) that makes xfs_setsize_buftarg's purpose less
> > > cohesive.
> > 
> > As explained last round this came up I'd of course rename it if
> > we did that.  But I can do that later.
> 
> <nod> Would you be willing to review this patch as it is now and either
> you or me can just tack a new cleanup patch on the end?  I tried writing
> a patch to clean this up, but ran into questions:
> 
> At first I thought that the xfs_setsize_buftarg call in
> xfs_alloc_buftarg could be replaced by open-coding the bt_meta_sector*
> assignment, checking that bdev_validate_blocksize is ok, and dropping
> the sync_blockdev.
> 
> Once we get to xfs_setup_devices, we can call xfs_setsize_buftarg on the
> three buftargs, and xfs_setsize_buftarg will configure the atomic writes
> geometry.
> 
> But then as I was reading the patch, it occurred to me that at least for
> the data device, we actually /do/ want that sync_blockdev call so that
> any dirty pagecache for the superblock actually get written to disk.
> Maybe that can go at the end of xfs_open_devices?  But would it be
> preferable to sync all the devices prior to trying to read the primary
> sb?  I don't think there's a need, but maybe someone else has a
> different viewpoint?

Eh, since John posted a V10 I'll just tack my new patches on the end of
that so everyone can look at them.

--D