Re: [PATCH 1/2 v2] tracing/function-graph-tracer: various fixes and features

[Frederic Weisbecker <fweisbec@gmail.com>]

On Fri, Jan 23, 2009 at 12:00:37PM +0100, Ingo Molnar wrote:
> 
> * Frédéric Weisbecker <fweisbec@gmail.com> wrote:
> 
> > > Still needs a solution - if we do cross-CPU traces we want to have a 
> > > global trace clock with 'seemless' transition between CPUs.
> > 
> > So it doesn't only need a monotonic clock. It needs a global consistent 
> > clock like ktime for example? Unfortunately this one uses seq_locks and 
> > would add some drawbacks like verifying if the traced function doesn't 
> > hold the write seq_lock and it will bring some more ftrace recursion...
> 
> using ktime_get() is indeed out of question - GTOD callpaths are too 
> complex (and also too slow).
> 
> I'd not change anything in the current logic, but i was thinking of a new 
> trace_option, which can be set optionally. If that trace option is set 
> then this bit of ring_buffer_time_stamp():
> 
>         time = sched_clock() << DEBUG_SHIFT;
> 
> gets turned into:
> 
>         time = cpu_clock(cpu) << DEBUG_SHIFT;
> 
> This way we default to sched_clock(), but also gain some 'global' 
> properties if the trace_option is set.


Ok, yeah that's a good idea.

 
> Furthermore, another trace_option could introduce a third 'strongly 
> ordered' trace-clock variant, which would use cmpxchg and per cpu 
> timestamps, something like this:
> 
> atomic64_t curr_time;
> 
> DEFINE_PER_CPU(u64, prev_cpu_time);
> ...
> 
> retry:
> 	prev_cpu_time = per_cpu(prev_cpu_time, cpu);
> 	cpu_time = sched_clock();
> 	old_time = atomic64_read(&curr_time);
> 
> 	delta = cpu_time - prev_cpu_time;
> 	if (unlikely((s64)delta <= 0))
> 		delta = 1;
> 
> 	new_time = old_time + delta;
> 
> 	if (atomic64_cmpxchg(&curr_time, old_time, new_time) != new_time)
> 		goto repeat;
> 
>         time = new_time << DEBUG_SHIFT;
> 
> This would be a monotonic, global clock wrapped around sched_clock(). It 
> uses a cmpxchg to achieve it, but we have to use global ordering anyway. 
> 
> It would still be _much_ faster than any GTOD clocksource we have.
> 
> Hm?
> 

And that would be even more faster that cpu_clock().

But why implement both? Wouldn't the above be more faster while playing the same thing
than cpu_clock()