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

[Peter Zijlstra]

On Thu, Apr 30, 2020 at 03:58:24PM +0100, Will Deacon wrote:
> On Fri, Mar 20, 2020 at 05:35:45PM +0800, Leo Yan wrote:

> > @@ -1164,5 +1165,21 @@ void arch_perf_update_userpage(struct perf_event *event,
> >  		userpg->time_mult >>= 1;
> >  	}
> >  	userpg->time_shift = (u16)shift;
> > -	userpg->time_offset = -now;
> > +
> > +	/*
> > +	 * Since arch timer is enabled ealier than sched clock registration,
> > +	 * compuate the delta (in nanosecond unit) between the arch timer
> > +	 * counter and sched clock, assign the delta to time_offset and
> > +	 * perf tool can use it for timestamp calculation.
> > +	 *
> > +	 * The formula for conversion arch timer cycle to ns is:
> > +	 *   quot = (cyc >> time_shift);
> > +	 *   rem  = cyc & ((1 << time_shift) - 1);
> > +	 *   ns   = quot * time_mult + ((rem * time_mult) >> time_shift);
> > +	 */
> > +	count = arch_timer_read_counter();
> > +	quot = count >> shift;
> > +	rem = count & ((1 << shift) - 1);
> > +	ns = quot * userpg->time_mult + ((rem * userpg->time_mult) >> shift);
> > +	userpg->time_offset = now - ns;
> 
> Hmm, reading the counter and calculating the delta feels horribly
> approximate to me. It would be much better if we could get hold of the
> initial epoch cycles from the point at which sched_clock was initialised
> using the counter. This represents the true cycle delta between the counter
> and what sched_clock uses for 0 ns.
> 
> Unfortunately, I can't see a straightforward way to grab that information.
> It looks like x86 pulls this directly from the TSC driver.

Yeah, and I'm thinking you should do the same. IIRC ARM uses this
kernel/time/sched_clock.c thing, and if I read that right, the struct
clock_data there has all the bits you need here.

So I'm thinking that you might want to add a helper function here to get
you the good stuff.