Re: [discuss] Re: 2.6.19-rc5: known regressions (v3)

[Andi Kleen <ak@suse.de>]

On Wed, Nov 15, 2006 at 12:21:18PM -0800, Andrew Morton wrote:
> Andi Kleen <ak@suse.de> wrote:
> 
> > 
> > > The fact is, it used to work, and the kernel changed interfaces, so now it 
> > > doesn't. 
> > 
> > No, it didn't work. oprofile may have done something, but it 
> > just silently killed the NMI watchdog in the process.
> > That was never acceptable.
> 
> But people could get profiles out.  I know, I've seen them!

Just the nmi watchdog was gone then.

> 
> > Now we do proper accounting of NMI sources and also proper allocation
> > of performance counters.
> > 
> >  
> > > Yes, "oprofile" should be fixed to not depend on that, but the kernel 
> > > shouldn't change the interfaces, and we should add back the zero entry.
> > 
> > That would break the nmi watchdog again.
> > 
> > Anyways, there is a sysctl to disable the nmi watchdog if someone
> > is desperate.
> > 
> > But I think it is clearly oprofile who did wrong here and needs
> > to be fixed.
> > 
> 
> Is it correct to say that oprofile-on-2.6.18 works, and that
> oprofile-on-2.6.19-rc5 does not?
> 
> Or is there some sort of workaround for this, or does 2.6.19-rc5 only fail

echo 0 > /proc/sys/kernel/nmi_watchdog before the oprofile module is loaded.
With builtin oprofile probably nmi_watchdog=0

> in some particular scenarios?

On x86-64 and on newer i386 machines (based on DMI year)


> 
> If it's really true that oprofile is simply busted then that's a serious
> problem and we should find some way of unbusting it.  If that means just
> adding a dummy "0" entry which always returns zero or something like that,
> then fine.

That could be probably done.

> But we can't just go and bust it.

It just did something unbelievable broken before. I would say it busted
itself.

-Andi