RE:(2) [PATCH] xfs: mark the inode preallocated for pNFS write layouts past EOF

[Daejun Park <daejun7.park@samsung.com>]

On Mon, Oct 12, 2026 at 02:41:18PM +0200, Christoph Hellwig wrote:
> > +		if (write && !(ip->i_diflags & XFS_DIFLAG_PREALLOC) &&
> > +		    imap.br_startoff + imap.br_blockcount >
> > +		    XFS_B_TO_FSB(mp, XFS_ISIZE(ip))) {
> > +			error = xfs_fs_map_update_inode(ip);
> > +			if (!error)
> > +				error = xfs_log_force_inode(ip);
> > +			if (error)
> > +				goto out_unlock;
> > +		}
>
> Let's avoid the code duplication here.  Add a boolean flag that
> we need the update, set it both for the new allocation in the if
> le, and based on the condition here, and then do the update
> in a separate conditional below.
>
> And mayne the xfs_log_force_inode should move into
> xfs_fs_map_update_inode while we're at it.

Will do both in v2. xfs_fs_map_blocks() sets a flag for the allocation
and for a write mapping that ends past EOF over blocks that are already
there, and calls xfs_fs_map_update_inode() once after the if/else.
xfs_fs_map_update_inode() forces the log after committing the
transaction, so the separate xfs_log_force_inode() calls go away.

The behaviour does not change: the update still happens under the
IOLOCK, before the blocks are handed out. On nfsd-testing with the same
reproducer, the inode is marked at LAYOUTGET, blockgc keeps the blocks,
and the LAYOUTCOMMIT after 360 seconds succeeds with the data in place,
as with v1.

I will wait a few days for other comments on v1 and then send v2.