On Wed, Nov 24, 2010 at 10:31:57PM +0800, Peter Zijlstra wrote:
> On Wed, 2010-11-24 at 22:21 +0800, Wu Fengguang wrote:
> >
> > Hmm, but why not avoid locking at all? With per-cpu bandwidth vars,
> > each CPU will see slightly different bandwidth, but that should be
> > close enough and not a big problem.
>
> I don't think so, on a large enough machine some cpus might hardly ever
> use a particular BDI and hence get very stale data.
Good point!
> Also, it increases the memory footprint of the whole solution.
Yeah, maybe not a good trade off.
> > > +void bdi_update_write_bandwidth(struct backing_dev_info *bdi)
> > > +{
> > > + unsigned long time_now, write_now;
> > > + long time_delta, write_delta;
> > > + long bw;
> > > +
> > > + if (!spin_try_lock(&bdi->bw_lock))
> > > + return;
> >
> > spin_try_lock is good, however is still global state and risks
> > cacheline bouncing..
>
> If there are many concurrent writers to the BDI I don't think this is
> going to be the top sore spot, once it is we can think of something
> else.
When there are lots of concurrent writers, we'll target at ~100ms
pause time, hence the update frequency will be lowered accordingly.
Thanks,
Fengguang
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxxx For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom policy in Canada: sign http://dissolvethecrtc.ca/
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>