[PATCH] trace: extend trace_clock to support arch_arm clock counter

[Will Deacon]

On Thu, Dec 15, 2016 at 06:46:09PM +0530, Srinivas Ramana wrote:
> On 12/12/2016 04:12 PM, Will Deacon wrote:
> >On Mon, Dec 12, 2016 at 10:31:52AM +0530, Srinivas Ramana wrote:
> >>On 12/06/2016 05:43 PM, Will Deacon wrote:
> >>>On Sun, Dec 04, 2016 at 02:06:23PM +0530, Srinivas Ramana wrote:
> >>>>On 12/02/2016 04:38 PM, Will Deacon wrote:
> >>>>>On Fri, Dec 02, 2016 at 01:44:55PM +0530, Srinivas Ramana wrote:
> >>>>>>Extend the trace_clock to support the arch timer cycle
> >>>>>>counter so that we can get the monotonic cycle count
> >>>>>>in the traces. This will help in correlating the traces with the
> >>>>>>timestamps/events in other subsystems in the soc which share
> >>>>>>this common counter for driving their timers.
> >>>>>
> >>>>>I'm not sure I follow this reasoning. What's wrong with nanoseconds? In
> >>>>>particular, the "perf" trace_clock hangs off sched_clock, which should
> >>>>>be backed by the architected counter anyway. What does the cycle counter in
> >>>>>isolation tell you, given that the frequency isn't architected?
> >>>>>
> >>>>>I think I'm missing something here.
> >>>>>
> >>>>
> >>>>Having cycle counter would help in the cases where we want to correlate the
> >>>>time with other subsystems which are outside cpu subsystem.
> >>>
> >>>Do you have an example of these subsystems? Can they be used to generate
> >>>trace data with mainline?
> >>
> >>Some of the subsystems i can list are Modem(on a mobilephone), GPU or video
> >>subsystem, or a DSP among others.
> >
> >Oh, you're talking about hardware subsystems. That makes this slightly more
> >compelling, but I don't think you want the virtual counter here, since
> >I assume those other subsystems don't take into account CNTVOFF (and I
> >don't really see how they could, it being a per-cpu thing). So, if you
> >want to expose the *physical* counter as a trace clock, I think that's
> >justifiable.
> >
> Yes, I meant HW subsystems. Sorry if I was not clear.
> In ARM64, it seems the access to physical counter is removed with commit
> "clocksource: arch_timer: Fix code to use physical timers when requested".
> Only ARM (32) is allowed to used physical counter in the current timer API.
> It seems only EL2 is supposed to access this. But yes, if there is an
> offset, it seems it would be difficult to get the exact value at EL0.
> However for systems where CNTVOFF is '0', this will work seamless. This
> clock would not be the default anyways and is optional. Local clock would
> continue to be the default for traces.

That still doesn't sound useful to userspace. I think we need to expose
the clock only in the cases where it's useful, so restricting it to the
physical counter is the right thing to do.

> >>>>local_clock or even the perf track_clock uses sched_clock which gets
> >>>>suspended during system suspend. Yes, they are backed up by the
> >>>>architected counter but they ignore the cycles spent in suspend.i
> >>>
> >>>Does mono_raw solve this (also hangs off the architected counter and is
> >>>supported in the vdso)?
> >>
> >>Doesn't seem like. Any of the existing clock sources are designed not show
> >>the jump, when there is a suspend and resume. Even though they run out of
> >>architected counter they just cane give exact correlation with the counter.
> >>Furthermore, during the initial kernel boot, these just run out of jiffies
> >>clock source. They also not account for the time spent in boot loaders.
> >
> >Hmm, there's a thing called CLOCK_BOOTTIME, but I don't think that helps
> >you when CNTVOFF comes into play.
> >
> CLOCK_BOOTTIME includes the time spent in suspend. But this also doesn't
> give exact counter value since power ON. So for the purpose of comparing
> with global counter, this would not help.
> 
> >>>>so, when comparing with monotonically increasing cycle counter, other
> >>>>clocks doesn't help. It seems X86 uses the TSC counter to help such cases.
> >>>
> >>>Does this mean we need a way to expose the frequency to userspace, too?
> >>
> >>Not really. The CNTFRQ_EL0 of timer subsystem holds the clock frequency of
> >>system timer and is available to EL0.
> >
> >Experience shows that CNTFRQ_EL0 is often unreliable, and the frequency
> >can be overridden by the device-tree. There are also systems where the
> >counter stops ticking across suspend. Whilst both of these can be considered
> >"broken", I suspect we want runtime buy-in from the arch-timer driver
> >before registering this trace_clock.
> 
> Agree. It doesnt seem like architecture mandates initializing this.
> For those systems where tick would stop, if not arch counter, i assume there
> is some counter which falls in 'always ON' domain without which they cant
> keep track of time.

We just need to avoid exposing this trace clock if the frequency was
provided by firmware.

Will