On Fri, Nov 21, 2014 at 02:06:57AM -0800, Christoph Hellwig wrote:

On Tue, Nov 18, 2014 at 11:22:35PM -0800, Omar Sandoval wrote:

quoted

Here's a nice little bit of insanity I put together in that direction --
consider it a discussion point more than a patch. It does two things:

- Uses an ITER_BVEC iov_iter to do direct_IO for swap_readpage. This makes
  swap_readpage a synchronous operation, but I think that's the best we can do
  with the existing interface.

Note that ->read_iter for direct-io supports async I/O in general.  By
resurrecting some of the older attempts to do in-kernel aio this could
be made async easily.

quoted

- Unless I'm missing something, there don't appear to be any instances of
  ITER_BVEC | READ in the kernel, so the dio path doesn't know not to dirty
  pages it gets that way. Dave Kleikamp and Ming Lei each previously submitted
  patches doing this as part of adding an aio_kernel interface. (The NFS direct
  I/O implementation doesn't know how to deal with these either, so this patch
  actually breaks the only existing user of this code path, but in the interest
  of keeping the patch short, I didn't try to fix it :)

Right, we'd need to look into.  Bonus points of allowing this as a zero
copy read.


Btw, in the long run I would much prefer killing of the current horrible
swap using bmap path in favor of an enhanced direct I/O path.

Looks like I just raced on your email and sent a v2 of my patch series before
seeing this response :) I'll take a closer look at this tomorrow, thanks for
getting back to me.

-- 
Omar