On Fri, Sep 17, 2021 at 11:11:49AM +0200, Peter Zijlstra wrote:

On Thu, Sep 16, 2021 at 10:07:07AM -0500, Bjorn Helgaas wrote:

quoted

This seems to be an ongoing issue, not just a point defect in a single
product, and I really hate the onesy-twosy nature of this.  Is there
really no way to detect this issue automatically or fix whatever Linux
bug makes us trip over this?  I am no clock expert, so I have
absolutely no idea whether this is possible.

X86 is gifted with the grant total of _0_ reliable clocks. Given no
accurate time, it is impossible to tell which one of them is broken
worst. Although I suppose we could attempt to synchronize against the
PMU or MPERF..

We could possibly disable the tsc watchdog for
X86_FEATURE_TSC_KNOWN_FREQ && X86_FEATURE_TSC_ADJUST I suppose.

And then have people with 'creative' BIOS get to keep the pieces.

Alternatively, we can change what the TSC watchdog does for
X86_FEATURE_TSC_ADJUST machines. Instead of checking time against HPET
it can check if TSC_ADJUST changes. That should make it more resillient
vs HPET time itself being off.