On Thu, May 29, 2014 at 4:36 PM, Minchan Kim <minchan@kernel.org> wrote:
>
> I did below hacky test to apply your idea and the result is overflow again.
> So, again it would second stack expansion. Otherwise, we should prevent
> swapout in direct reclaim.

So changing io_schedule() is bad, for the reasons I outlined elsewhere
(we use it for wait_for_page*() - see sleep_on_page().

It's the congestion waiting where the io_schedule() should be avoided.

So maybe test a patch something like the attached.

NOTE! This is absolutely TOTALLY UNTESTED! It might do horrible
horrible things. It seems to compile, but I have absolutely no reason
to believe that it would work. I didn't actually test that this moves
anything at all to kblockd. So think of it as a concept patch that
*might* work, but as Dave said, there might also be other things that
cause unplugging and need some tough love.

                   Linus