Re: [PATCHv5] perf tools: enable close-on-exec flag on perf file descriptor

[Jiri Olsa <jolsa@redhat.com>]

On Tue, Jun 03, 2014 at 10:57:15AM +0200, Yann Droneaud wrote:
> Le lundi 02 juin 2014 à 21:23 +0200, Jiri Olsa a écrit :
> > On Mon, Jun 02, 2014 at 12:56:34PM +0200, Yann Droneaud wrote:
> > 
> > SNIP
> > 
> > > 
> > > Hi,
> > > 
> > > Quite the same patch from v4. I'm interested in some
> > > feedback so that I could improve the patch if needed.
> > > 
> > > Regards.
> > > 
> > > Changes from v4 [1]:
> > > - rebase on next-20140530 and update commit message.
> > 
> > hi,
> > I wasn't following on this one before.. so please shut me up
> > if my comments were already discussed ;-)
> > 
> 
> You had a chance to review it before :)
> See http://lkml.kernel.org/r/20140106142220.GB1183@krava.brq.redhat.com

yep, but then I drift away ;-)

SNIP

> > >  
> > >  #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
> > > @@ -988,7 +989,7 @@ static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
> > >  			      struct thread_map *threads)
> > >  {
> > >  	int cpu, thread;
> > > -	unsigned long flags = 0;
> > > +	unsigned long flags = PERF_FLAG_FD_CLOEXEC;
> > >  	int pid = -1, err;
> > >  	enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
> > >  
> > > @@ -997,11 +998,13 @@ static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
> > >  		return -ENOMEM;
> > >  
> > >  	if (evsel->cgrp) {
> > > -		flags = PERF_FLAG_PID_CGROUP;
> > > +		flags |= PERF_FLAG_PID_CGROUP;
> > >  		pid = evsel->cgrp->fd;
> > >  	}
> > >  
> > >  fallback_missing_features:
> > > +	if (perf_missing_features.cloexec)
> > > +		flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
> > >  	if (perf_missing_features.mmap2)
> > >  		evsel->attr.mmap2 = 0;
> > >  	if (perf_missing_features.exclude_guest)
> > > @@ -1070,7 +1073,10 @@ try_fallback:
> > >  	if (err != -EINVAL || cpu > 0 || thread > 0)
> > >  		goto out_close;
> > >  
> > > -	if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
> > > +	if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
> > > +		perf_missing_features.cloexec = true;
> > > +		goto fallback_missing_features;
> > 
> > I think it does not fit in here, because we check latest perf_event_attr
> > bits and removing one after another, to see if miracle happens.
> > (also looks like we miss exclude_callchain_(kernel|user) bits)
> > 
> > Wouldn't it be better just to use perf_event_open_cloexec_flag() function
> > as you did in the rest of the code? I think we dont need 2 detection codes
> > for this.
> > 
> 
> I was asked by ACME to use here the missing feature detection pattern
> and to treat the flag just as a new feature like perf_event_attr new
> attributes.
> 
> See http://lkml.kernel.org/r/20140106144347.GA13500@ghostprotocols.net
>     http://lkml.kernel.org/r/20140106211444.GD2810@ghostprotocols.net

ook, I probably missed that.. anyway I still feel like the perf_event_open_cloexec_flag
function should be enough.. I'll leave this one to Arnadlo

thanks,
jirka