Re: [PATCH] perf tool: report user-friendly error from timechart

[Arnaldo Carvalho de Melo <acme@ghostprotocols.net>]

Em Thu, Oct 03, 2013 at 07:35:18AM -0600, David Ahern escreveu:
> On 10/3/13 6:38 AM, Ingo Molnar wrote:
> >
> >* Ramkumar Ramachandra <artagnon@gmail.com> wrote:
> >
> >>+	/* Perform a quick sanity check */
> >>+	if (!is_valid_tracepoint("power:cpu_frequency")) {
> >>+		fprintf(stderr, "Error:\tNo permissions to read $debugfs/tracing/events/power/cpu_frequency\n");
> >>+		fprintf(stderr, "Hint:\tChange the permissions of debugfs: /sys/kernel/debug\n");
> >>+		fprintf(stderr, "\tThe directory will be present if your kernel was compiled with debugfs support.\n");
> >
> >Is missing permissions the only way how is_valid_tracepoint() can fail?
> >
> >What if debugfs has the right permissions but CONFIG_TRACEPOINTS is
> >disabled in the kernel?
> 
> There are a number of reasons that function can fail. The complete
> solution is to plumb various error numbers and on failure request a
> string for that failure. Take a look at util/target.[ch] as an
> example.
> 
> The comment applies to the perf-trace patch as well, but it gets
> more complicated to handle the error paths from perf_evsel__newtp
> when they dip into the tracepoint code

See the patch I posted, in that case we can use the old errno way, i.e.
do nothing and just look at it in the perf_evsel__newtp/
perf_evlist__add_newtp callers.

And is_valid_tracepoint() is a too big hammer, it traverses the whole
directory looking for a match instead of plain build the path and do an
access, its one of those things I need to ditch at some point. So far I
just try to do a perf_evlist__add_newtp and if it fails, look at errno.

- Arnaldo