[PATCH 0/4] xfs: fixes for XFS_DIFLAG2_DAX support

[Ross Zwisler]

On Mon, Feb 15, 2016 at 04:22:10PM +1100, Dave Chinner wrote:
> Hi folks,
> 
> This is a series to add the correct constraints to using the on-disk
> inode flag to enable DAX on per-file basis. The same constraints are
> placed on setting the flag on directories for inheritance purposes.
> 
> These constraints are:
> 	- the inode flag is limited to regular files or directory
> 	  inodes.
> 	- the S_DAX flag is only ever set on regular files
> 	- the flag can only ever be set on filesystems which have
> 	  blocksize == PAGE_SIZE (for now)
> 	- When the flag is set or cleared, the current mapping
> 	  contents are flushed and then invalidated so that the new
> 	  access mode starts with an empty mapping.
> 	- Setting or clearing the flag is atomic w.r.t. IO and
> 	  page faults.
> 
> I've tested these manually with xfs_io (patchset for supporting
> chattr +x/-x to be sent soon), and it all appears to work as
> expected. I'd like to push these for 4.5-rc6 so the initial kernel
> with support for this flag doesn't do silly things, so comments,
> testing and review woul dbe appreciated.
> 
> Cheers,
> 
> Dave.

This seems like what we want to be doing for raw block devices, when we add
back in support.  Default to DAX off so we have to opt-in, and when we get an
IOCTL to enable them block I/O and page faults, clear out all mappings and
radix tree entries, and then change the flag and turn back on I/O and page
faults.  We'll also need to add a check to prevent mounts on raw block devices
with S_DAX set.