On 2011-04-12 18:58, hch@infradead.org wrote:

On Tue, Apr 12, 2011 at 11:31:17PM +1000, Dave Chinner wrote:

quoted

I don't think so. e.g. in the XFS allocation path we do btree block
readahead, then go do the real work. The real work can end up with a
deeper stack before blocking on locks or completions unrelated to
the readahead, leading to schedule() being called and an unplug
being issued at that point.  You might think it contrived, but if
you can't provide a guarantee that it can't happen then I have to
assume it will happen.

In addition to the stack issue, which is a killer to this also has
latency implications.  Before we could submit a synchronous metadata
read request inside readpage or writepage and kick it off to the disk
immediately, while now it won't get submitted until we block the next
time, i.e. have done some more work that could have been used for
doing I/O in the background.  With the kblockd offload not only have
we spent more time but at the point where we finally kick it we
also need another context switch.  It seem like we really need to
go through the filesystems and explicitly flush the plugging queue
for such cases.  In fact a bio flag marking things as synchronous
metadata reads would help, but then again we need to clean up our
existing bio flags first..

I think it would be a good idea to audit the SYNC cases, and if feasible
let that retain the 'immediate kick off' logic. If not, have some way to
signal that at least. Essentially allow some fine grained control of
what goes into the plug and what does not.

-- 
Jens Axboe