This is the mail archive of the
systemtap@sourceware.org
mailing list for the systemtap project.
Re: Unified tracing buffer
On Tue, 23 Sep 2008, Martin Bligh wrote:
> >> > Yup, explains your TSC observation. Nothing we can do about. Broken by
> >> > system design :( Welcome in the wonderful world of Inhell/BIOS/ACPI !
> >>
> >> We have linux patches that sync the TSC on exit_idle. I'll see if I can get
> >> Michael to send them out.
> >
> > Are you sure that they sync it precicely enough that there is no user
> > space observable way of time going backwards between cores ?
>
> I think the tolerance is about 500 cycles. If that's not sufficient, I guess
> we'll have to either live with some slight misordering (which people have
> pointed out is kind of inevitable anyway) on these broken machines?
> It was sufficient for what we were using it for, but maybe not for everyone.
Well, I dont care about the trace reordering at all. I care about user
space visible time going backwards issues observed via the
gettimeofday vsyscall. 500 cycles should be fine, I doubt that we can
migrate in less than that :)
I guess you try this only for machines where the TSC runs with
constant frequency, right ?
Looking forward to your patches.
Thanks,
tglx