[Bug 256781] EC2 Nitro: TSC-low timecounter lags

https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=256781

--- Comment #40 from dwmw2@infradead.org ---
Apologies for not confirming sooner. Yes, this was due to a bug in the PIT
implementation. The PIT counter counts down, and the interrupt was supposed to
fire when it reaches zero and wraps back to its start value (often 65535).

In some circumstances the timer interrupt could be delivered slightly too soon,
and the guest could read a counter value which hasn't quite reached zero yet —
which the guest would take to mean that a whole extra countdown has occurred in
the interim! This would overestimate the amount of actual time taken, as
measured by the PIT, and thus underestimate the frequency of the TSC which was
being calibrated.

I hereby rescind the advice to use the KVM clock. Use the TSC instead. It's
much more reliable. The KVM clock has... problems:
https://lore.kernel.org/all/20240522001817.619072-1-dwmw2@infradead.org/

The TSC on modern hardware has none of the issues which made the KVM clock make
sense in the first place. And now the KVM clock is the buggy one.

-- 
You are receiving this mail because:
You are the assignee for the bug.