Re: [PATCHv2 11/11] writeback: prevent unnecessary bdi threads wakeups

[Dave Chinner <david@fromorbit.com>]

On Wed, Jul 21, 2010 at 02:45:41PM +0300, Artem Bityutskiy wrote:
> On Wed, 2010-07-21 at 12:31 +0300, Artem Bityutskiy wrote:
> >  	spin_unlock(&inode_lock);
> > +
> > +	if (wakeup_bdi) {
> > +		spin_lock(&bdi->wb_lock);
> > +		if (!bdi->wb.task)
> > +			wake_up_process(default_backing_dev_info.wb.task);
> > +		else
> > +			wake_up_process(bdi->wb.task);
> > +		spin_unlock(&bdi->wb_lock);
> > +	}
> >  }
> 
> Dave,
> 
> I do not know whether this stuff will end up in upstream, I did not get
> any feed back from Jens so far. But if it will, I'd like to let you know
> that the code quoted above is similar to the 'bdi_queue_work()'
> function. And the purpose is very similar. But you added a
> 'trace_writeback_nothread()' call to 'bdi_queue_work()', and I think a
> similar call has to be here.

Yes, that seems like a sane thing to do ;)

> Can I call 'trace_writeback_nothread()'? I guess not. Should I create
> another trace point? Any hints/instructions?

The bdi_queue_work() tracepoints expect a work structure to be
passed in, so you can't use them (or that class of event) if you
don't have a struct wb_writeback_work.

For __mark_inode_dirty(), I'd add two new tracepoints like:

DEFINE_WRITEBACK_EVENT(writeback_wakeup);
DEFINE_WRITEBACK_EVENT(writeback_wakeup_nothread);

and place them as:

	if (wakeup_bdi) {
		trace_writeback_wakeup(bdi)
		spin_lock(&bdi->wb_lock);
		if (!bdi->wb.task) {{
			trace_writeback_wakeup_nothread(bdi);
			wake_up_process(default_backing_dev_info.wb.task);
		} else
			wake_up_process(bdi->wb.task);
		spin_unlock(&bdi->wb_lock);
	}

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com