Re: Flash Back -- kernel 2.1.111

[Vojtech Pavlik]

On Sun, Feb 24, 2002 at 04:59:37PM -0600, Troy Benjegerdes wrote:
> On Sun, Feb 24, 2002 at 11:51:13PM +0100, Vojtech Pavlik wrote:
> > On Sun, Feb 24, 2002 at 02:44:23PM -0800, David S. Miller wrote:
> > 
> > >    From: Vojtech Pavlik <[e-mail]>
> > >    Date: Sun, 24 Feb 2002 23:39:37 +0100
> > >    
> > >    > > happens if you plug in a 66MHz non-capable card to the 50 MHz bus.
> > >    > 
> > >    > The bus speed drops to 33MHz.
> > >    
> > >    Interesting. I'd expect 25 myself ... then we'll definitely need two
> > >    clock values in struct pci_bus - because the hi-speed one isn't always a
> > >    double the low one - as shown by your example.
> > > 
> > > You only need one, the current active one.
> > >
> > > If you think that hot-plug is an issue, the arch dependant could would
> > > need to recalculate the "current bus speed" and all would be fine.
> > > 
> > > So why do we need two values?
> > 
> > Oh, you're right. We indeed need only one.
> > 
> > Hmm, now hotplug changing the PCI clock would be quite a lot of fun -
> > all running drivers will need to know about the change, because some may
> > need to recompute the timings they have programmed into the chips ...
> > 
> > Because virtually disconnecting and reconnecting all the cards for which
> > the timings have changed will probably not be an option.
> 
> Personally, I think hot-plugging a 33mhz pci device into a 66mhz pci bus 
> with other active devices on it is either user error or designer error 
> (should have had a bridge), and a 'virtual disconnect and reconnect' is 
> reasonable. 
> 
> You're going to kill (or at least stop) any transactions going on on the
> bus while you're physically hot-plugging anway..

I'd guess most hotpluggable PCIs will have a bridge per slot ...
hopefully.

-- 
Vojtech Pavlik
SuSE Labs