Re: [9front] Speeding up snoopy(8)

[ori@eigenstate.org]

Quoth David Arroyo <droyo@aqwari.net>:
> On Fri, Oct 21, 2022, at 12:04, ori@eigenstate.org wrote:
> > Quoth David Arroyo <droyo@aqwari.net>:
> >> On Thu, Oct 20, 2022, at 12:03, Sigrid Solveig Haflínudóttir wrote:
> >> > It would be nice to have a way to get monotonic timestamps in userspace
> >> > without having to go into the kernel. For now, you can try nanosec and
> >> > see if anything changes.
> >> 
> >> I tried using nanosec, and that was sufficient for snoopy to keep up
> >> with a bit rate of 1.6 Gbps, with no overflows. Thank you for the tip!
> >
> > ...should we replace the nsec() function with it?
> >
> > we may be able to stash nsec (and monotonic nsec?) at last
> > context switch into tos, and interpolate using cycles.
> 
> I don't fully understand the tradeoffs in using nanosec() over nsec().
> For packet captures, I specifically *don't* want time adjustments during
> a capture, so nanosec() being monotonic is a plus.
> 
> It doesn't matter for snoopy currently, but is nanosec() thread-safe?
> It's keeping xstart in a static variable; would different threads have
> different xstarts, or could they step on each other if they call
> nanosec() at the same time?
> 
> David

it is not; as it stands, it's not a drop-in replacement for nsec();
doing it properly would mean extending the Tos struct, I think.