Re: [RFC PATCH-tip v4 02/10] locking/rwsem: Stop active read lock ASAP

[Christoph Hellwig]

On Mon, Oct 10, 2016 at 05:07:45PM +1100, Dave Chinner wrote:
> > > *However*, the DAX IO path locking in XFS  has changed in 4.9-rc1 to
> > > match the buffered IO single writer POSIX semantics - the test is a
> > > bad test based on the fact it exercised a path that is under heavy
> > > development and so can't be used as a regression test across
> > > multiple kernels.
> > 
> > That being said - I wonder if we should allow the shared lock on DAX
> > files IFF the user is specifying O_DIRECT in the open mode..
> 
> It should do - if it doesn't then we screwed up the IO path
> selection logic in XFS and we'll need to fix it.

Depends on your defintion of "we".  The DAX code has always abused the
direct I/O path, and that abuse is ingrained in the VFS path in a way that
we can't easily undo it in XFS, e.g. take a look at io_is_direct and
iocb_flags in include/linux/fs.h, which ensure that DAX I/O will always
appear as IOCB_DIRECT to the fs.  It will take some time to untagle
this, but it's on my todo list once the last file system (ext4)
untangles the DAX and direct I/O path.