Re: [RFC] sched_clock: Track monotonic raw clock

[Pawel Moll]

On Tue, 2014-07-22 at 20:39 +0100, Richard Cochran wrote:
> So the only reasonable way, IMHO, is to simply provide the needed
> information in the traces, and then add some user space code to find
> the relationship between the perf clock and the system clock. The
> great advantage of this approach is that you don't have to create a
> perfect servo from day one. The time stamps can be re-analyzed at any
> time after the fact.

Ok, so I had yet another (already lost count ;-) look into this...

I came up with a conclusion that the simplest way it can be done is to
have a periodically repeated events marked with both "kernel timestamp"
and "user timestamp". And a scheduler switch is such a periodic event.
Moreover, it seems quite natural: a user process is scheduled in (take a
snapshot of sched clock and monotonic raw at this moment, let's call
them t_s_in and t_m_in), the process "generates" the "user timestamp" by
clock_gettime(CLOCK_MONOTONIC_RAW, t_m_x), the process is scheduled out
(again, take sched and monotonic timestamps then, t_s_out and t_m_out).

The, after extracting all the t_?_? from the trace stream, a translation
(we've know the monotonic raw value t_m_x, we're looking for its sched
clock equivalent t_s_x) should simply get down to:

t_s_x = t_s_in + ((t_m_x - t_m_1) * (t_s_out - t_s_in)) / (t_m_out - t_m_in)

I quickly made an experiment, by adding a customised variant of
sched_switch trace event (note: I'm *not* proposing it to be merged)

8<-------------