Re: [3/8] x86/tsc: Store and check TSC ADJUST MSR

[Henning Schild]

On Mon, 30 Jan 2017 11:20:25 +0100
Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote:

> Henning,
> 
> On Fri, 27 Jan 2017, Henning Schild wrote:
> > 
> > did you by any chance look into TSC synchronization by adjusting the
> > absolute value (MSR_IA32_TSC) as well? As far as i have seen Linux
> > did that a long time ago and eventually it was stopped because it
> > caused more harm than good.  
> 
> I was involved in both developing the TSC sync patches and ripping
> them out again.
> 
> The problem with writing TSC directly is that you really _CANNOT_
> reliably handle run-time differences and SMI/NMI induced deltas. With
> the adjust MRS it's a halfways sane thing to do, except for the
> brokeness of that botch job vs. the TSC deadline timer.
> 
> > The ADJUST MSR offers an easy way to synchronize, still taking care
> > of all the special cases resulted in an 8-patch series. Synching
> > without that using the absolute value is likely much harder, but
> > that series might be a good foundation.  
> 
> I'm not even thinking about bringing the pure TSC based sync back.
> 
> > The big question is whether we can rely on all future CPUs to 
> > support that MSR. Do "new MSRs" disappear again at some point? If we
> > can not rely on the ADJUST MSR, now might be a good time to revisit
> > the idea of synching the absolute values.  
> 
> There is nothing you can ever be sure about, but I doubt that the
> ADJUST MSR is going to vanish.

That sounds very much like i expected. But assuming the MSR has come to
stay, the problem should be solved for recent kernels and Intel-CPUs.

The AMD-Manual from 12/16 does not mention that MSR. I do not have
access to an AMD machine. But i can only assume that bigger machines
also suffer from async TSCs and basically all fall back to HPET.

> > I remember having read somewhere that this series might get
> > backported to longterm kernels, what is the status on that?  
> 
> No idea.
> 
> Thanks,

Thanks a lot!
Henning