Device polling, with SMP?

Don Bowman don at sandvine.com
Thu Nov 20 13:15:32 PST 2003


From: Avleen Vig [mailto:lists-freebsd at silverwraith.com]
> On Thu, Nov 20, 2003 at 12:58:58PM -0500, Don Bowman wrote:
> > > I read Luigi's paper at info.iet.unipi.it/~luigi/polling/ 
> which at the
> > > end implies that DEVICE_POLLING on an SMP box might not make 
> > > sense - but
> > > right now for me it would make sense as both CPU's are locked:
> > >   One tries to handle interrupts
> > >   The other tries to manage the application
> > > 
> > > I could try forcing DEVICE_POLLING to compile as is 
> suggested in that
> > > URL but I wanted to see if anyone had tried this before.
> > > The interface is an FXP.
> > 
> > We use it on em. I just commented out the #error line that
> > says you can't do it.
> > device polling in idle doesn't work, and the user/system time
> > calculation isn't correct, but it works well otherwise.
> 
> This is pretty much what I wanted to confirm thanks!
> In which way is the system/user time incorrect? Always, or only under
> high load? what about it is incorrect? My skills are limited 
> but I might
> take a stab at fixing that.

Well, on -STABLE, there can be only one CPU active in the kernel.
Thus on a 2-way HTT system, the system thinks there are 4
processors, but there is no way to get 'system' cpu to exceed
25%. So in polling, the 'user frac' and 'kernel frac' are very
difficult to understand.

--don


More information about the freebsd-hackers mailing list