Re: [PATCH] perf stat: Poll for monitored tasks being alive in fork mode

[Jiri Olsa]

On Sat, Jan 05, 2019 at 11:16:40AM +0800, Jin, Yao wrote:
> 
> 
> On 1/4/2019 8:54 PM, Jiri Olsa wrote:
> > On Fri, Jan 04, 2019 at 10:28:17AM +0800, Jin Yao wrote:
> > > Following test shows the stat keeps running even if no longer
> > > task to monitor (mgen exits at ~5s).
> > > 
> > > perf stat -e cycles -p `pgrep mgen` -I1000 -- sleep 10
> > >              time             counts unit events
> > >       1.000148916      1,308,365,864      cycles
> > >       2.000379171      1,297,269,875      cycles
> > >       3.000556719      1,297,187,078      cycles
> > >       4.000914241        761,261,827      cycles
> > >       5.001306091      <not counted>      cycles
> > >       6.001676881      <not counted>      cycles
> > >       7.002046336      <not counted>      cycles
> > >       8.002405651      <not counted>      cycles
> > >       9.002766625      <not counted>      cycles
> > >      10.001395827      <not counted>      cycles
> > > 
> > > We'd better finish stat immediately if there's no longer task to
> > > monitor.
> > > 
> > > After:
> > > 
> > > perf stat -e cycles -p `pgrep mgen` -I1000 -- sleep 10
> > >              time             counts unit events
> > >       1.000180062      1,236,592,661      cycles
> > >       2.000421539      1,223,733,572      cycles
> > >       3.000609910      1,297,047,663      cycles
> > >       4.000807545      1,297,215,816      cycles
> > >       5.001001578      1,297,208,032      cycles
> > >       6.001390345        582,343,659      cycles
> > > sleep: Terminated
> > > 
> > > Now the stat exits immediately when the monitored tasks ends.
> > > 
> > > Signed-off-by: Jin Yao <yao.jin@xxxxxxxxxxxxxxx>
> > > ---
> > >   tools/perf/builtin-stat.c | 7 +++++++
> > >   1 file changed, 7 insertions(+)
> > > 
> > > diff --git a/tools/perf/builtin-stat.c b/tools/perf/builtin-stat.c
> > > index 63a3afc..71f3bc8 100644
> > > --- a/tools/perf/builtin-stat.c
> > > +++ b/tools/perf/builtin-stat.c
> > > @@ -553,6 +553,13 @@ static int __run_perf_stat(int argc, const char **argv, int run_idx)
> > >   		if (interval || timeout) {
> > >   			while (!waitpid(child_pid, &status, WNOHANG)) {
> > > +				if (!is_target_alive(&target,
> > > +					evsel_list->threads) &&
> > > +					(child_pid != -1)) {
> > 
> > do we need that child_pid check? we just returned from waitpid
> > so we should be ok.. we just make the race window smaller
> > 
> > could we just do:
> > 
> > 				if (!is_target_alive(&target, evsel_list->threads)) {
> > 					kill(child_pid, SIGTERM);
> > 					break;
> > 				}
> > 
> 
> I think this code should be OK and I have tested yet. I have a question
> about the race condition, we really don't need a lock to protect the
> child_pid?
> 
> skip_signal()
> {
> 	/*
> 	 * render child_pid harmless
> 	 * won't send SIGTERM to a random
> 	 * process in case of race condition
> 	 * and fast PID recycling
> 	 */
> 	child_pid = -1;
> }
> 
> __run_perf_stat()
> {
> 	....
> 	kill(child_pid, SIGTERM);
> }
> 
> If child_pid is set by -1 in a small window between checking of child_pid
> and kill(), then kill(-1, SIGTERM) may happen. All processes except the kill
> process itself and init would receive SIGTERM.

ah right, -1 is special.. however that can still happen also
in the orginal patch.. how about we do something like below

jirka


---
diff --git a/tools/perf/builtin-stat.c b/tools/perf/builtin-stat.c
index acfd48db52dd..c322cb271180 100644
--- a/tools/perf/builtin-stat.c
+++ b/tools/perf/builtin-stat.c
@@ -583,6 +583,14 @@ static int __run_perf_stat(int argc, const char **argv, int run_idx)
 
 		if (interval || timeout) {
 			while (!waitpid(child_pid, &status, WNOHANG)) {
+				if (!is_target_alive(&target, evsel_list->threads)) {
+					int pid = child_pid;
+
+					if (pid != -1)
+						kill(pid, SIGTERM);
+					break;
+				}
+
 				nanosleep(&ts, NULL);
 				if (timeout)
 					break;