On Sat, Mar 19, 2011 at 09:45:00AM -0400, Christoph Hellwig wrote:
> On Fri, Mar 18, 2011 at 03:06:48PM +1100, Dave Chinner wrote:
> > It gets used by a second caller in the next patch that uses a
> > timeout of zero. The idea of adding a delay to a normal push is to
> > rate limit the number of times we do work so we always work on
> > batches rather a few items at a time in multiple executions of the
> > work.
> > I'll see if it's simpler to just do this work directly in teh
> > callers, though.
> I don't think hiding this delay (uncommented) in the workqueue use is
> a good idea.
FWIW, we already have an implicit delay for frequent callers when
the AIL is busy - the uninterruptible sleep for sleeps of <= 20ms.
That was implemented specifically to rate-limit wakeups while the
xfsaild was busy pushing. This is essentially a different
implementation of the same mechanism.
> xlog_grant_push_ail has all the logics about when to push
> the AIL, so any batching should be grouped with that logic, and
> documented there. It in fact already has some comments static that
> a min/max watermark scheme would be useful.
Yes, it does, but that's a much bigger change that has some
potentially nasty problems like ensuring the watermarks are always a
sane distance apart which is difficult to do on small logs were a
single transaction reservation can easily be larger than 10% of the
log. Hence watermarks are a much harder change to validate and tune
compared to a simple push wakeup rate-limit...