On Tue, Oct 25, 2016 at 09:13:29AM -0800, Kent Overstreet wrote:

Also - what are you doing about the race between shooting down the range in the
pagecache and dirty pages being readded? The existing direct IO code falls back
to buffered IO for that, but your code doesn't appear to - I seem to recall that
XFS has its own locking for this, are you just relying on that for now?  It'd be
really nice to get some generic locking for this, anything that relies on
pagecache invalidation is sketchy as hell in other filesystems.

Yes, XFS always had a shared/exclusive lock for I/O operations,
which is taken exclusive for buffered writes and those corner cases
of direct writes that needs exclusіon (e.g. sub-fs block size I/O).

This prevents new dirty pages from being added while direct I/O is
in progress.  There is nothing to prevent direct reads, though - that's
why both the old common code, the old XFS code and this new code do
a second invalidation after the write is done.

Now that the VFS i_mutex has been replaced with i_rwsem we can apply
this scheme to common code as well by taking i_rwsem shared for
direct I/O reads.