Re: [PATCH 2/2] trace,x86: Add x86 irq vector entry/exit tracepoints

[Frederic Weisbecker <fweisbec@gmail.com>]

On Wed, Dec 14, 2011 at 10:09:11AM -0500, Seiji Aguchi wrote:
> Thank you for giving me a comment.
> Unfortunately, neither "perf record" nor "ftrace" works for me.
> 
> >What about using perf for that?
> >
> >Just run:
> >
> >	perf record -ag
> >	^C
> >	perf report
> >
> >And you should find in the callchains some informations about where your CPUs
> >are spending time.
> >
> >If you system is too slow for that 
> 
> When system is too slow, user command such as "perf record" may not work.
> 
> >but you're doing background tracing with
> >ftrace, you can use stacktrace with ftrace.
> 
> Actually, We're doing background tracing in our customer's system rather than kernel debugging. 
> Ftrace doesn't work for me because it checks the size of the stack at every function call.
> Our customers are seriously concerned about its overhead.
> 
> For reducing the overhead, I need tracepoints so we can hook minimal function calls.

Well ftrace is a whole subsystem that includes the function tracer and also an interface
for tracepoints in debugfs. I was rather suggesting the latter one. This is a good
choice for background tracing. And it supports stacktraces. If those generate too much
overhead perhaps you can tune the number of entries in the stacktrace, I don't remember
if we can do that currently but this can be an interesting feature.

What are you using currently for the background tracing?