On 21.06.24 10:45, David Woodhouse wrote: > On Thu, 2024-06-20 at 17:19 +0100, David Woodhouse wrote: >> >>> >>>> + >>>> + /* 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? The 'shift' field should be fine.
