Re: [RFC][PATCH 2/3] perf: Take a hot regs snapshot for trace events

[Ingo Molnar <mingo@elte.hu>]

* Steven Rostedt <rostedt@goodmis.org> wrote:

> On Thu, 2010-03-04 at 12:25 +0100, Ingo Molnar wrote:
> > * Peter Zijlstra <peterz@infradead.org> wrote:
> > 
> > > On Wed, 2010-03-03 at 12:07 -0500, Steven Rostedt wrote:
> > > > oops, my bad :-), I thought this was in the x86 arch directory. For the
> > > > University, I was helping them with adding trace points for page faults
> > > > when I came across this in arch/x86/mm/fault.c:
> > > > 
> > > >         perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, 0, regs, address);
> > > > 
> > > > 
> > > > This is what I actually was wondering about. Why is it a "perf only" trace 
> > > > point instead of a TRACE_EVENT()?
> > > 
> > > Because I wanted to make perf usable without having to rely on funny 
> > > tracepoints. That is, I am less worried about committing software counters 
> > > to ABI than I am about TRACE_EVENT(), which still gives me a terribly 
> > > uncomfortable feeling.
> > 
> > I'd still like a much less error-prone and work-intense way of doing it.
> > 
> > I'd suggest we simply add a TRACE_EVENT_ABI() for such cases, where we 
> > really want to expose a tracepoint to tooling, programmatically. Maybe 
> > even change the usage sites to trace_foo_ABI(), to make it really clear 
> > and to make people aware of the consequences.
> 
> Would this still be available as a normal trace event?

Yeah, of course. It would not result in any usage or flexibility restriction.

( In the future we might want to add some sort of automated signature thing to 
  make sure that an event that has been declared an 'ABI' isnt changed - at 
  least as far as the record format goes. )

> > 
> > > Also, building with all CONFIG_TRACE_*=n will still yield a usable perf, 
> > > which is something the embedded people might fancy, all that TRACE stuff 
> > > adds lots of code.
> > 
> > Not a real issue i suspect when you do lock profiling ...
> > 
> > Or if it is, some debloating might be in order - and the detaching of event 
> > enumeration and ftrace TRACE_EVENT infrastructure from other ftrace bits. (i 
> > suggested an '/eventfs' special filesystem before, for nicely layed out 
> > hierarchy of ftrace/perf events.)
> 
> Actually, we already have a way to decouple it.
> 
> include/trace/define_trace.h is the file that just adds the tracepoint
> that is needed.
> 
> include/trace/ftrace.h is the file that does the magic and adds the code
> for callbacks and tracing.
> 
> The perf hooks probably should not have gone in that file and been put
> into a include/trace/perf.h file, and then in define_trace.h we would
> add:
> 
>  #ifdef CONFIG_EVENT_TRACING
>  #include <trace/ftrace.h>
>  #endif
> 
> +#ifdef CONFIG_PERF_EVENTS
> +#include <trace/perf.h>
> +#endif
> 
> This should be done anyway. But it would also let you decouple ftrace trace 
> events from perf trace events but still let the two use the same trace 
> points.

I think the main thing would be to have a decoupled /eventfs - basically 
/debug/tracing/events/ moved to "/eventfs" or maybe to "/proc/events/". This 
would make them available more widely, and in a standardized way.

Thanks,

	Ingo