Re: [PATCH] clocksource: Avoid selecting mult values that might overflow when adjusted

[John Stultz <johnstul@us.ibm.com>]

On Thu, 2011-11-03 at 10:01 -0400, John Stultz wrote:
> On Thu, 2011-11-03 at 14:26 +0100, Thomas Gleixner wrote:
> > On Thu, 3 Nov 2011, John Stultz wrote:
> > > On Thu, 2011-11-03 at 13:05 +0100, Thomas Gleixner wrote:
> > > > On Wed, 2 Nov 2011, John Stultz wrote:
> > > > >  
> > > > > +	WARN_ONCE(timekeeper.mult+adj >
> > > > > +			timekeeper.clock->mult + timekeeper.clock->maxadj,
> > > > > +			"Adjusting more then 11%%");
> > > > 
> > > > Shouldn't we rather limit the update instead of just warn and overflow ?
> > > 
> > > Well, I'm hesitant to commit to that, just yet. So I figured I'd start
> > > with the warning.
> > 
> > OTOH, we know right there that we might warp 32bit and confuse the
> > hell out of timekeeping, which is not a real good thing either.
> 
> Oh certainly, but two things:
> 1) The 11% max is not the actual overflow edge. Its just calculated as
> safe. The overflow could as far out as ~22%.
> 
> 2) This is the first case in however many years I've heard of of mult
> overflowing. So before we go changing the NTP code (which is really
> terribly complex, but has been working fairly well for awhile) I want to
> have some sense that the 11% max adjustment assumption is really
> correct.
> 
> But maybe I'm being too conservative? If we do limit the adjustment
> keeping the warning, I guess we'd know why things blew up on previously
> working machines. 

Oh, and the other bit is that not all clocksources have been converted
over to using clocksource_register_hz/khz, so some may be using very
small shift values, which could more easily hit large % mult adjustment
(due to the resulting coarseness of each integer change) that wouldn't
cause overflows.

thanks
-john