perf trace: Fix perf trace -p <PID>

'perf trace -p <PID>' work on a syscall that is unaugmented, but doesn't
work on a syscall that's augmented (when it calls perf_event_output() in
BPF).

Let's take open() as an example. open() is augmented in perf trace.

Before:

  $ perf trace -e open -p 3792392
     ? (         ):  ... [continued]: open()) = -1 ENOENT (No such file or directory)
     ? (         ):  ... [continued]: open()) = -1 ENOENT (No such file or directory)

We can see there's no output.

After:

   $ perf trace -e open -p 3792392
      0.000 ( 0.123 ms): a.out/3792392 open(filename: "DINGZHEN", flags: WRONLY) = -1 ENOENT (No such file or directory)
   1000.398 ( 0.116 ms): a.out/3792392 open(filename: "DINGZHEN", flags: WRONLY) = -1 ENOENT (No such file or directory)

Reason:

bpf_perf_event_output() will fail when you specify a pid in 'perf trace' (EOPNOTSUPP).

When using 'perf trace -p 114', before perf_event_open(), we'll have PID
= 114, and CPU = -1.

This is bad for bpf-output event, because the ring buffer won't accept
output from BPF's perf_event_output(), making it fail. I'm still trying
to find out why.

If we open bpf-output for every cpu, instead of setting it to -1, like
this:

  PID = <PID>, CPU = 0
  PID = <PID>, CPU = 1
  PID = <PID>, CPU = 2
  PID = <PID>, CPU = 3

Everything works.

You can test it with this script (open.c):

  #include <unistd.h>
  #include <sys/syscall.h>

  int main()
  {
	int i1 = 1, i2 = 2, i3 = 3, i4 = 4;
	char s1[] = "DINGZHEN", s2[] = "XUEBAO";

	while (1) {
		syscall(SYS_open, s1, i1, i2);
		sleep(1);
	}

	return 0;
  }

save, compile:

  make open

perf trace:

  perf trace -e open <path-to-the-executable>

Signed-off-by: Howard Chu <howardchu95@gmail.com>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Ian Rogers <irogers@google.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Kan Liang <kan.liang@linux.intel.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Link: https://lore.kernel.org/r/20240815013626.935097-2-howardchu95@gmail.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>

diff --git a/tools/perf/util/evlist.c b/tools/perf/util/evlist.c
index ef58a7764318f371687bd96d30b35788c2441687..f14b7e6ff1dcc2cd3f1f6b1d8c152ab14b4e0b5b 100644 (file)
--- a/tools/perf/util/evlist.c
+++ b/tools/perf/util/evlist.c
@@ -1067,7 +1067,7 @@ int evlist__create_maps(struct evlist *evlist, struct target *target)
        if (!threads)
                return -1;
 
-       if (target__uses_dummy_map(target))
+       if (target__uses_dummy_map(target) && !evlist__has_bpf_output(evlist))
                cpus = perf_cpu_map__new_any_cpu();
        else
                cpus = perf_cpu_map__new(target->cpu_list);