Re: [PATCH RFC] x86/tsc: Make recalibration default on for TSC_KNOWN_FREQ cases

From: Feng Tang
Date: Sun Jun 04 2023 - 21:18:38 EST


On Fri, Jun 02, 2023 at 12:08:10PM -0700, Paul E. McKenney wrote:
> On Fri, Jun 02, 2023 at 11:36:54AM -0700, Dave Hansen wrote:
> > On 6/2/23 11:29, Paul E. McKenney wrote:
> > >>> One downside is, many VMs also has X86_FEATURE_TSC_KNOWN_FREQ set,
> > >>> and they will also do this recalibration.
> > >> It's also pointless for those SoCs which lack legacy hardware.
> > >>
> > >> So why do you force this on everyone?
> > > Just for the record, this patch could be helpful in allowing victims
> > > of TSC mis-synchronization to more easily provide a more complete bug
> > > report to the firmware people. There is of course no point if there is
> > > already a fix available.
> > >
> > > But it is not all that hard to work around not having this patch upstream.
> > > This can be hand-applied as needed, NTP drift rates can be pressed
> > > into service for those of us having atomic clocks near all our servers,
> > > or the firmware guys can be tasked with figuring it out.
> > >
> > > So this patch would be nice to have, but we could live without it.
> >
> > Is this the kind of thing we could relegate to a kernel unit test? Like
> > make the recalibration logic _available_, but don't have it affect the
> > rest of the system.
> >
> > I love patching my kernel as much as the next guy. But, you know what I
> > *don't* love? Explaining how to patch kernels to other people. ;)
>
> One could argue that we already have the TSC equivalent of a kernel unit
> test with the tsc=recalibrate kernel boot parameter.
>
> So, would it make sense to have something like tsc=recalibrate (already
> present) in the guise of something like hpet=recalibrate and/or
> pmtmr=recalibrate in order to allow people to opt into recalibrating
> whatever timer is marked unstable?

This kind of hint parsing should be in tsc.c, so some name like
'tsc_recal=hpet/pmtmr' ?

As hpet and pmtimer are the only 2 choices and hpet is the default
one, if people want to use pmtimer, they can use combined parameter
"tsc=recalibrate nohpet".

Also, thanks for sharing your thoughts form a victim's viewpoint
in the other mail! :)

Thanks,
Feng

> Thanx, Paul