On Thu, 2024-06-20 at 17:19 +0100, David Woodhouse wrote:

quoted

quoted

+
+       /* Counter frequency, and error margin. Units of (second >> 64) */
+       uint64_t counter_period_frac_sec;

AFAIU this might limit the precision in case of high counter frequencies.
Could the unit be aligned to the expected frequency band of counters?

This field indicates the period of a single tick, in units of 1>>64 of
a second. That's about 5.4e-20 seconds, or 54 zeptoseconds? 

Can you walk me through a calculation where you believe that level of
precision is insufficient?

I guess the precision matters if the structure isn't updated for a long
period of time, and the delta between the current counter and the
snapshot is high? That's a *lot* of 54 zeptosecondses? But you really
would need a *lot* of them before you care? And if nobody's been
calibrating your counter for that long, surely you have bigger worries?

Am I missing something there?

Hm, that was a bit rushed at the end of the day; let's take a better look...

Let's take a hypothetical example of a 100GHz counter. That's two
orders of magnitude more than today's Arm arch counter.

The period of such a counter would be 10 picoseconds. 

(Let's ignore the question of how far light actually travels in that
time and how *realistic* that example is, for the moment.)

It turns out that at that rate, there *are* a lot of 54 zeptosecondses
of precision loss in the day. It could be half a millisecond a day, or
20µs an hour.

That particular example of 10 picoseconds is 184467440.7370955
(seconds>>64) which could be truncated to 184467440 — losing about 4PPB
(a third of a millisecond a day; 14µs an hour).

So yeah, I suppose a 'shift' field could make sense. It's easy enough
to consume on the guest side as it doesn't really perturb the 128-bit
multiplication very much; especially if we don't let it be negative.

And implementations *can* just set it to zero. It hurts nobody.

Or were you thinking of just using a fixed shift like (seconds>>80)
instead?