Re: [PATCH] arm64: perf_event: Fix time_offset for arch timer

[Marc Zyngier]

On 2020-05-01 16:14, Leo Yan wrote:
> Hi all,
>
> On Thu, Apr 30, 2020 at 05:18:15PM +0100, Will Deacon wrote:
> > On Thu, Apr 30, 2020 at 06:04:36PM +0200, Peter Zijlstra wrote:
> > > On Thu, Apr 30, 2020 at 04:29:23PM +0100, Marc Zyngier wrote:
> > >
> > > > I wonder if we could/should make __sched_clock_offset available even when
> > > > CONFIG_HAVE_UNSTABLE_SCHED_CLOCK isn't defined. It feels like it would
> > > > help with this particular can or worm...
> > >
> > > Errrgh. __sched_clock_offset is only needed on x86 because we transition
> > > from one clock device to another on boot. It really shouldn't exist on
> > > anything sane.
> >
> > I think we still transition from jiffies on arm64, because we don't 
> > register
> > with sched_clock until the timer driver probes. Marc, is that right?
> >
> > > Let me try and understand your particular problem better.
> >
> > I think the long and short of it is that userspace needs a way to 
> > convert
> > the raw counter cycles into a ns value that can be compared against 
> > values
> > coming out of sched_clock. To do this accurately, I think it needs the
> > cycles value at the point when sched_clock was initialised.
>
> Will's understanding is exactly what I want to resolve in this patch.
>
> The background info is for the ARM SPE [1] decoding with perf tool, if
> the timestamp is enabled, it uses the generic timer's counter as
> timestamp source.  SPE trace data only contains the raw counter cycles,
> as Will mentioned, the perf tool needs to convert it to a coordinate
> value with sched_clock.  This is why this patch tries to calculate the
> offset between the raw counter's ns value and sched_clock, eventually
> this offset value will be used by SPE's decoding code in Perf tool to
> calibrate a 'correct' timestamp.
>
> Based on your suggestions, I will use __sched_clock_offset to resolve
> the accuracy issue in patch v2.  (I noticed Peter suggested to use a
> new API for wrapping clock_data structure, IIUC, __sched_clock_offset
> is more straightforward for this case).

I think Peter's point was *not* to use __sched_clock_offset which
appears to be only there for the purpose of an x86 workaround (and not
available to other architecture), but to instead expose the relevant
field (epoch_cyc) to the perf subsystem.

I feel it makes sense to make this a generic API, and see whether x86
can move over to it rather than the other way around.

Thanks,

        M.
--
Jazz is not dead. It just smells funny...