Re: [PATCH 2/3] ftrace: add tracepoint for xtime

[john stultz <johnstul@us.ibm.com>]

On Wed, 2009-09-16 at 16:32 -0400, Steven Rostedt wrote:
> On Wed, 2009-09-16 at 12:58 -0700, john stultz wrote:
> > On Wed, Sep 16, 2009 at 12:56 PM, john stultz <johnstul@us.ibm.com> wrote:
> > > On Tue, Aug 25, 2009 at 1:14 AM, Zhaolei <zhaolei@cn.fujitsu.com> wrote:
> > >> From: Xiao Guangrong <xiaoguangrong@cn.fujitsu.com>
> > >>  /* Structure holding internal timekeeping values. */
> > >>  struct timekeeper {
> > >>        /* Current clocksource used for timekeeping. */
> > >> @@ -338,6 +341,8 @@ int do_settimeofday(struct timespec *tv)
> > >>
> > >>        update_vsyscall(&xtime, timekeeper.clock);
> > >>
> > >> +       trace_gtod_update_xtime(&xtime, &wall_to_monotonic);
> > >> +
> > >
> > > So the only thing to watch out on here is that xtime doesn't hold the
> > > current time, but the accumulated time. There is some unaccumulated
> > > time still kept in the clocksource structure.
> > >
> > > You probably want (assuming you only need tick granularity time) to
> > > use current_kernel_time().
> > >
> > > As an aside, is there a reason you have to have update callbacks and
> > > duplicate the time storage instead of using the existing interfaces?
> > > (ie: Is this due to locking or something else?)
> > 
> > Doh. Sorry, you're actually tracing the timekeeping code. Not using
> > this to assist tracing. Got this confused.
> > 
> > So yea, I think this should be ok, plus or minus determining if you
> > really want xtime or xtime_cache.
> 
> Well this may be a real concern. It's not about tracing timekeeping
> (although it adds that functionality). His second patch (I didn't Cc you
> on that one) hooks to these tracepoints to update time accordingly.
> 
> What is being done is a way to have a "wall time" value being added to
> the ring buffer. But this needs to be very carefully done, because the
> all tracers use this, including the function tracer in NMI code. So the
> clock source can not take locks or do anything fancy.
> 
> What the idea is, is to have a semi clock that is read with some kind of
> fast increment, and then at clock ticks, this clock is synced up.

Hmm.. Yea, if that's the case, then I'm not a big fan of this approach.

It sounds like what's really needed is a lock-free variant of
current_kernel_time() or something close to the CLOCK_MONOTONIC_COARSE
functionality currently queued.

Doing it without locks might have some downsides, and I guess that's the
point of the callback method (updates happen at prescribed times and
likely under locks that the trace code understands so it avoids races
and deadlocks).

-john