perf stat: Do not show stats if workload fails
PATCH 1/1 perf stat: Do not show stats if workload fails
Currently perf-stat attempts to show counter stats even if the workload is bogus:
$ perf stat -- foo foo: No such file or directory
Performance counter stats for 'foo':
<not counted> task-clock
<not counted> context-switches
<not counted> cpu-migrations
<not counted> page-faults
<not counted> cycles
<not counted> stalled-cycles-frontend
<not counted> stalled-cycles-backend
<not counted> instructions
<not counted> branches
<not counted> branch-misses0.009769943 seconds time elapsed
It is impossible to differentiate all the failure modes, but it seems reasonable that if the workload handling fails, perf-stat should not try to print stats.
With this change:
$ perf stat -v -- foo Failed to start workload
--- tools/perf/builtin-stat.c | 11 +++++++++-- 1 file changed, 9 insertions(+), 2 deletions(-)
tools/perf/builtin-stat.c +9 -2
@@ -586,7 +586,11 @@ static int __run_perf_stat(int argc, const char **argv)clock_gettime (CLOCK_MONOTONIC , &ref_time);if (forks) { -perf_evlist__start_workload (evsel_list); +if (perf_evlist__start_workload (evsel_list) !=0 ) { +pr_err ("Failed to start workload\n" ); +return -1 ; + } +handle_initial_delay ();if (interval) {
@@ -1793,7 +1797,10 @@ int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused) run_idx + 1); status =run_perf_stat (argc, argv); -if (forever && status !=-1 ) { +if (status <0 ) +break ; + +if (forever) {print_stat (argc, argv);perf_stat__reset_stats (evsel_list); }
Signature · 1 line
1.8.3.4 (Apple Git-47)
Quoted 25 lines · David Ahern
Currently perf-stat attempts to show counter stats even if the workload is bogus:
$ perf stat -- foo foo: No such file or directoryPerformance counter stats for 'foo':
<not counted> task-clock <not counted> context-switches <not counted> cpu-migrations <not counted> page-faults <not counted> cycles <not counted> stalled-cycles-frontend <not counted> stalled-cycles-backend <not counted> instructions <not counted> branches <not counted> branch-misses0.009769943 seconds time elapsedIt is impossible to differentiate all the failure modes, but it seems reasonable that if the workload handling fails, perf-stat should not try to print stats.
Quoted 8 lines
With this change:
$ perf stat -v -- foo Failed to start workloadSigned-off-by David Ahern <[email protected]>Cc Ingo Molnar <[email protected]>Cc Stephane Eranian <[email protected]>
Nice!
Thanks,
Ingo
Em Fri, Dec 20, 2013 at 08:57:59AM +0100, Ingo Molnar escreveu:
Quoted 32 lines · David Ahern
Quoted 31 lines
Currently perf-stat attempts to show counter stats even if the workload is bogus:
$ perf stat -- foo foo: No such file or directoryPerformance counter stats for 'foo':
<not counted> task-clock <not counted> context-switches <not counted> cpu-migrations <not counted> page-faults <not counted> cycles <not counted> stalled-cycles-frontend <not counted> stalled-cycles-backend <not counted> instructions <not counted> branches <not counted> branch-misses0.009769943 seconds time elapsedIt is impossible to differentiate all the failure modes, but it seems reasonable that if the workload handling fails, perf-stat should not try to print stats.
With this change:
$ perf stat -v -- foo Failed to start workload
Quoted 7 lines
Quoted 3 lines
Signed-off-by David Ahern <[email protected]>Cc Ingo Molnar <[email protected]>Cc Stephane Eranian <[email protected]>Nice!
Acked-by Ingo Molnar <[email protected]>
Good intent, but...
[acme@ssdandy linux]$ perf stat usleep 1
Failed to start workload
[acme@ssdandy linux]$
Your patch does:
if (perf_evlist__start_workload(evsel_list) != 0)
But:
int perf_evlist__start_workload(struct perf_evlist *evlist)
{
if (evlist->workload.cork_fd > 0) {
char bf = 0;
int ret;
/*
* Remove the cork, let it rip!
*/
ret = write(evlist->workload.cork_fd, &bf, 1);
if (ret < 0)
perror("enable to write to pipe");
close(evlist->workload.cork_fd);
return ret;
}
return 0;
}Ret there is 1, so we need to change it to:
return ret != 1 ? -1 : 0;
Right?
- Arnaldo
Quoted 19 lines · Arnaldo Carvalho de Melo
int
perf_evlist__start_workload(struct perf_evlist *evlist){ if (evlist->workload.cork_fd > 0) { char bf = 0; int ret; /* * Remove the cork, let it rip! */ ret = write(evlist->workload.cork_fd, &bf, 1); if (ret < 0) perror("enable to write to pipe"); close(evlist->workload.cork_fd); return ret; } return 0; }
Quoted 5 lines
Ret there is 1, so we need to change it to:
return ret != 1 ? -1 : 0;Right?
Yes, nice catch.
David
Em Mon, Dec 23, 2013 at 07:29:54PM -0500, David Ahern escreveu:
Quoted 16 lines · Arnaldo Carvalho de Melo
Quoted 14 lines
int
perf_evlist__start_workload(struct perf_evlist *evlist){ if (evlist->workload.cork_fd > 0) { char bf = 0; /* Remove the cork, let it rip! */ int ret = write(evlist->workload.cork_fd, &bf, 1); if (ret < 0) perror("enable to write to pipe"); close(evlist->workload.cork_fd); return ret; } return 0; }
Quoted 7 lines
Quoted 5 lines
Ret there is 1, so we need to change it to:
return ret != 1 ? -1 : 0;Right?
Yes, nice catch.
But this doesn't really matters, if one uses the more usual:
if (function() < 0)
pattern for error checking, there would be no problem, and we would only be checking that we told perf_evlist__prepare_workload()'s fork branch to call execvp, not to check what happens then, i.e. if it finds the specified workload in the system's PATH, if there are no permission of file format problems, etc.
The thing to check is perf_evlist__{prepare,start}_workload notification errors using SIGUSR1, that we need to check for in the caller, and emit the message, no?
- Arnaldo
Em Tue, Dec 24, 2013 at 09:53:42AM -0300, Arnaldo Carvalho de Melo escreveu:
Quoted 3 lines
The thing to check is perf_evlist__{prepare,start}_workload notification errors using SIGUSR1, that we need to check for in the caller, and emit the message, no?
Something like this:
1. We tell perf_evlist__prepare_workload that we want a signal if execvp fails, it will be a SIGUSR1
2. We catch that signal in 'stat' and check that we got a signal, only problem so far with this signal maze is that we're getting a SIGCHLD while I was expecting a SIGUSR1... I.e. the "if (signr != -1) test really should be if (signr == SIGUSR1), but I'm getting a SIGCHLD there and the elves are tugging me away...
- Arnaldo
tools/perf/builtin-stat.c +5 -3
@@ -144,6 +144,7 @@ static struct timespec ref_time;static struct cpu_map *aggr_map;static int (*aggr_get_id)(struct cpu_map *m,int cpu); +static volatile int signr =-1 ;static volatile int done =0 ;struct perf_stat {
@@ -531,7 +532,7 @@ static int __run_perf_stat(int argc, const char **argv)if (forks) {if (perf_evlist__prepare_workload (evsel_list, &target, argv, - false, false) <0 ) { + false, true) <0 ) {perror ("failed to prepare workload" );return -1 ; }
@@ -598,6 +599,8 @@ static int __run_perf_stat(int argc, const char **argv)wait (&status);if (WIFSIGNALED (status))psignal (WTERMSIG (status), argv[0 ]); +else if (signr !=-1 ) +return -1 ; }else {handle_initial_delay ();while (!done) {
@@ -1335,8 +1338,6 @@ static void print_stat(int argc, const char **argv) } } -static volatile int signr =-1 ; -static void skip_signal (int signo) {if ((child_pid ==-1 ) || interval)
@@ -1785,6 +1786,7 @@ int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)signal (SIGCHLD , skip_signal);signal (SIGALRM , skip_signal);signal (SIGABRT , skip_signal); +signal (SIGUSR1 , skip_signal); status =0 ;for (run_idx =0 ; forever || run_idx < run_count; run_idx++) {
Quoted 6 lines · Arnaldo Carvalho de Melo
Em Tue, Dec 24, 2013 at 09:53:42AM -0300, Arnaldo Carvalho de Melo escreveu:
Quoted 3 lines
Quoted 3 lines
The thing to check is perf_evlist__{prepare,start}_workload notification errors using SIGUSR1, that we need to check for in the caller, and emit the message, no?
Something like this:
Quoted 8 lines
1. We tell
perf_evlist__prepare_workloadthat we want a signal if execvp fails, it will be a SIGUSR12. We catch that signal in 'stat' and check that we got a signal, only problem so far with this signal maze is that we're getting a SIGCHLD while I was expecting a SIGUSR1... I.e. the "if (signr != -1) test really should be if (signr == SIGUSR1), but I'm getting a SIGCHLD there and the elves are tugging me away...
Did the elves release you?
There are all kinds of failure paths with the workload functions. In the end I was focusing on perf-stat actually checking the rc on the start_workload function. If it fails, then write() failed and something happened to the workload process. In that case don't show the counters.
Handling the other error paths with appropriate messages will take additional effort - as you are finding. ;-)
David
Em Thu, Dec 26, 2013 at 09:11:22AM -0500, David Ahern escreveu:
Quoted 18 lines · Arnaldo Carvalho de Melo
Quoted 6 lines
Em Tue, Dec 24, 2013 at 09:53:42AM -0300, Arnaldo Carvalho de Melo escreveu:
Quoted 3 lines
Quoted 3 lines
The thing to check is perf_evlist__{prepare,start}_workload notification errors using SIGUSR1, that we need to check for in the caller, and emit the message, no?
Something like this:
Quoted 8 lines
1. We tell
perf_evlist__prepare_workloadthat we want a signal if execvp fails, it will be a SIGUSR12. We catch that signal in 'stat' and check that we got a signal, only problem so far with this signal maze is that we're getting a SIGCHLD while I was expecting a SIGUSR1... I.e. the "if (signr != -1) test really should be if (signr == SIGUSR1), but I'm getting a SIGCHLD there and the elves are tugging me away...
Did the elves release you?
Quoted 8 lines
There are all kinds of failure paths with the workload functions. In the end I was focusing on perf-stat actually checking the rc on the
start_workloadfunction. If it fails, thenwrite()failed and something happened to the workload process. In that case don't show the counters.Handling the other error paths with appropriate messages will take additional effort - as you are finding. ;-)
Right, but I can't apply that patch, as it makes 'perf stat whatever-workload' to fail, as I realized when doing a demo to someone interested in using perf ;-\
So for now I'm not applying that one.
Ah, at this point elves are everywhere, dammit! ;-)
- Arnaldo
Quoted 5 lines · Arnaldo Carvalho de Melo
Right, but I can't apply that patch, as it makes 'perf stat whatever-workload' to fail, as I realized when doing a demo to someone interested in using perf ;-\
So for now I'm not applying that one.
right, so you want one with < 0 check or wait for something else? I was not expecting to find it in your perf/core branch, yet there it is.
Quoted 2 lines
Ah, at this point elves are everywhere, dammit! ;-)
An elf put it there?
David
Em Thu, Dec 26, 2013 at 09:18:17AM -0500, David Ahern escreveu:
Quoted 10 lines · Arnaldo Carvalho de Melo
Quoted 5 lines
Right, but I can't apply that patch, as it makes 'perf stat whatever-workload' to fail, as I realized when doing a demo to someone interested in using perf ;-\
So for now I'm not applying that one.
right, so you want one with < 0 check or wait for something else? I was not expecting to find it in your perf/core branch, yet there it is.
I'll remove it from there, but try it, IIRC there will be some other problem :-\
I'd have to reread the messages I sent, but from what I recall the return from perf_evlist__start_workload() will _always_ be valid, i.e. what you're testing there is just if the parent wrote a byte to a pipe to signal the waiting child to call exec, and that _will_ work, the exec()? perhaps not, you'd have to setup the signal error reporting mechanism, etc.
Perhaps this should be somehow done by perf_evlist__start_workload, so that what it reports is the result of the exec in the child, and not merely if it managed to tell it to try to exec...
- Arnaldo
Quoted 4 lines
Quoted 2 lines
Ah, at this point elves are everywhere, dammit! ;-)
An elf put it there?
Right, dwarves may be involved, didn't had the time to figure that out...
Quoted 1 line
David
Em Thu, Dec 26, 2013 at 09:18:17AM -0500, David Ahern escreveu:
Quoted 10 lines · Arnaldo Carvalho de Melo
Quoted 5 lines
Right, but I can't apply that patch, as it makes 'perf stat whatever-workload' to fail, as I realized when doing a demo to someone interested in using perf ;-\
So for now I'm not applying that one.
right, so you want one with < 0 check or wait for something else? I was not expecting to find it in your perf/core branch, yet there it is.
Shouldn't be anymore,
- Arnaldo
Quoted 6 lines
Quoted 2 lines
Ah, at this point elves are everywhere, dammit! ;-)
An elf put it there?
David
Em Thu, Dec 26, 2013 at 09:11:22AM -0500, David Ahern escreveu:
Quoted 18 lines · Arnaldo Carvalho de Melo
Quoted 6 lines
Em Tue, Dec 24, 2013 at 09:53:42AM -0300, Arnaldo Carvalho de Melo escreveu:
Quoted 3 lines
Quoted 3 lines
The thing to check is perf_evlist__{prepare,start}_workload notification errors using SIGUSR1, that we need to check for in the caller, and emit the message, no?
Something like this:
Quoted 8 lines
1. We tell
perf_evlist__prepare_workloadthat we want a signal if execvp fails, it will be a SIGUSR12. We catch that signal in 'stat' and check that we got a signal, only problem so far with this signal maze is that we're getting a SIGCHLD while I was expecting a SIGUSR1... I.e. the "if (signr != -1) test really should be if (signr == SIGUSR1), but I'm getting a SIGCHLD there and the elves are tugging me away...
Did the elves release you?
Quoted 8 lines
There are all kinds of failure paths with the workload functions. In the end I was focusing on perf-stat actually checking the rc on the
start_workloadfunction. If it fails, thenwrite()failed and something happened to the workload process. In that case don't show the counters.Handling the other error paths with appropriate messages will take additional effort - as you are finding. ;-)
So, please try the attached patch, then apply this one on top, still needs work for 'record' and 'trace', that now don't print anything when a workload fails. I'll work on passing a pointer to evlist to the sigaction, then all this will be hidden away inside perf_evlist__prepare_workload, etc.
tools/perf/builtin-stat.c +15 -6
@@ -58,6 +58,7 @@#include "util/thread.h" #include "util/thread_map.h" +#include <signal.h> #include <stdlib.h> #include <sys/prctl.h> #include <locale.h>
@@ -509,16 +510,17 @@ static void handle_initial_delay(void) } } -static volatile bool workload_exec_failed; +static volatile int workload_exec_errno;/* * perf_evlist__prepare_workload will send a SIGUSR1 * if the fork fails, since we asked by setting its * want_signal to true. */ -static void workload_exec_failed_signal (int signo__maybe_unused ) +static void workload_exec_failed_signal (int signo__maybe_unused ,siginfo_t *info, +void *ucontext__maybe_unused ) { - workload_exec_failed = true; + workload_exec_errno = info->si_value.sival_int; }static int __run_perf_stat (int argc,const char **argv)
@@ -596,13 +598,17 @@ static int __run_perf_stat(int argc, const char **argv)clock_gettime (CLOCK_MONOTONIC , &ref_time);if (forks) { +struct sigaction act = { + .sa_flags =SA_SIGINFO , + .sa_sigaction = workload_exec_failed_signal, + };/* * perf_evlist__prepare_workload will, after we call * perf_evlist__start_Workload, send a SIGUSR1 if the exec call * fails, that we will catch in workload_signal to flip -* workload_exec_failed. +* workload_exec_errno. */ -signal (SIGUSR1 , workload_exec_failed_signal); +sigaction (SIGUSR1 , &act,NULL );perf_evlist__start_workload (evsel_list);handle_initial_delay ();
@@ -615,8 +621,11 @@ static int __run_perf_stat(int argc, const char **argv) }wait (&status); -if (workload_exec_failed) +if (workload_exec_errno) { +const char *emsg =strerror_r (workload_exec_errno, msg,sizeof (msg)); +pr_err ("Workload failed: %s\n" , emsg);return -1 ; + }if (WIFSIGNALED (status))psignal (WTERMSIG (status), argv[0 ]);
tools/perf/util/evlist.c +8 -3
@@ -1073,9 +1073,14 @@ int perf_evlist__prepare_workload(struct perf_evlist *evlist, struct target *tarexecvp (argv[0 ], (char **)argv); -perror (argv[0 ]); -if (want_signal) -kill (getppid (),SIGUSR1 ); +if (want_signal) { +union sigval val; + + val.sival_int = errno; +if (sigqueue (getppid (),SIGUSR1 , val)) +perror (argv[0 ]); + }else +perror (argv[0 ]);exit (-1 ); }
Quoted 1 line
From
b015481108106b7ef42d46a5096b572b1bd71b50Mon Sep 17 00:00:00 2001
From: Arnaldo Carvalho de Melo <[email protected]>
Date: Sat, 28 Dec 2013 15:45:08 -0300
Subject: [PATCH] perf stat: Don't show counter information when workload fails
When starting a workload 'stat' wasn't using prepare_workload evlist method's signal based exec() error reporting mechanism.
Use it so that the we don't report 'not counted' counters.
Before:
[acme@zoo linux]$ perf stat dfadsfa
dfadsfa: No such file or directory
Performance counter stats for 'dfadsfa':
<not counted> task-clock
<not counted> context-switches
<not counted> cpu-migrations
<not counted> page-faults
<not counted> cycles
<not counted> stalled-cycles-frontend
<not supported> stalled-cycles-backend
<not counted> instructions
<not counted> branches
<not counted> branch-misses0.001831462 seconds time elapsed
[acme@zoo linux]$
After:
[acme@zoo linux]$ perf stat dfadsfa
dfadsfa: No such file or directory
[acme@zoo linux]$
--- tools/perf/builtin-stat.c | 26 +++++++++++++++++++++++++- 1 file changed, 25 insertions(+), 1 deletion(-)
tools/perf/builtin-stat.c +25 -1
@@ -509,6 +509,18 @@ static void handle_initial_delay(void) } } +static volatile bool workload_exec_failed; + +/* +* perf_evlist__prepare_workload will send a SIGUSR1 +* if the fork fails, since we asked by setting its +* want_signal to true. +*/ +static void workload_exec_failed_signal (int signo__maybe_unused ) +{ + workload_exec_failed = true; +} +static int __run_perf_stat (int argc,const char **argv) {char msg[512 ];
@@ -529,7 +541,7 @@ static int __run_perf_stat(int argc, const char **argv)if (forks) {if (perf_evlist__prepare_workload (evsel_list, &target, argv, - false, false) <0 ) { + false, true) <0 ) {perror ("failed to prepare workload" );return -1 ; }
@@ -584,6 +596,14 @@ static int __run_perf_stat(int argc, const char **argv)clock_gettime (CLOCK_MONOTONIC , &ref_time);if (forks) { +/* +* perf_evlist__prepare_workload will, after we call +* perf_evlist__start_Workload, send a SIGUSR1 if the exec call +* fails, that we will catch in workload_signal to flip +* workload_exec_failed. +*/ +signal (SIGUSR1 , workload_exec_failed_signal); +perf_evlist__start_workload (evsel_list);handle_initial_delay ();
@@ -594,6 +614,10 @@ static int __run_perf_stat(int argc, const char **argv) } }wait (&status); + +if (workload_exec_failed) +return -1 ; +if (WIFSIGNALED (status))psignal (WTERMSIG (status), argv[0 ]); }else {
Signature · 1 line
1.8.5.rc2