Re: [PATCH] perf tools: Move precise_ip detection into perf_evsel__open

[Jiri Olsa]

On Fri, Mar 15, 2019 at 11:35:04AM -0300, Arnaldo Carvalho de Melo wrote:

SNIP

> > +static void display_attr(struct perf_event_attr *attr)
> > +{
> > +	if (verbose >= 2) {
> > +		fprintf(stderr, "%.60s\n", graph_dotted_line);
> > +		fprintf(stderr, "perf_event_attr:\n");
> > +		perf_event_attr__fprintf(stderr, attr, __open_attr__fprintf, NULL);
> > +		fprintf(stderr, "%.60s\n", graph_dotted_line);
> > +	}
> > +}
> > +
> > +static int perf_event_open(struct perf_evsel *evsel,
> > +			   pid_t pid, int cpu, int group_fd,
> > +			   unsigned long flags)
> 
> 
> The patch is ok, but I think the naming of this function is too generic,
> so I'm renaming it to:
> 
> static int perf_evsel__open_adjust_precise_ip(struct perf_evsel *evsel,
> 					      pid_t pid, int cpu, int group_fd,
> 					      unsigned long flags)
> 
> Ok?

ok

> 
> The perf_evsel__open() code is already complex with that fallback
> mechanism, this is just one more way of fallbacking when asking the
> kernel for something that may fail.
> 
> In fact what happens if the precise_ip that is being asked _is_
> supported but sys_perf_event_open() fails because some other
> perf_event_attr attribute that is set is not supported? 

it's outside the scope of standard feature fallback code,
so we will try it for any possible fallback variant, so:

we will try all possible precise_ip (3,2,1,0) and they will
all fail because of the unsupported attribute - so we will
restore the precise_ip back and continue in standard fallback
code that will eventualy switch that attribute off

> 
> I see, it gets it back restored to what the user asked so that the
> standard fallback is tried, ok, I'll apply with just the rename for this
> function,

thanks,
jirka