perf record: Support weak groups
authorAndi Kleen <ak@linux.intel.com>
Mon, 1 Oct 2018 19:59:27 +0000 (12:59 -0700)
committerArnaldo Carvalho de Melo <acme@redhat.com>
Mon, 5 Nov 2018 17:37:10 +0000 (14:37 -0300)
Implement a weak group fallback for 'perf record', similar to the
existing 'perf stat' support.  This allows to use groups that might be
longer than the available counters without failing.

Before:

  $ perf record  -e '{cycles,cache-misses,cache-references,cpu_clk_unhalted.thread,cycles,cycles,cycles}' -a sleep 1
  Error:
  The sys_perf_event_open() syscall returned with 22 (Invalid argument) for event (cycles).
  /bin/dmesg | grep -i perf may provide additional information.

After:

  $ ./perf record  -e '{cycles,cache-misses,cache-references,cpu_clk_unhalted.thread,cycles,cycles,cycles}:W' -a sleep 1
  WARNING: No sample_id_all support, falling back to unordered processing
  [ perf record: Woken up 3 times to write data ]
  [ perf record: Captured and wrote 8.136 MB perf.data (134069 samples) ]

Signed-off-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Link: http://lkml.kernel.org/r/20181001195927.14211-2-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
tools/perf/Documentation/perf-list.txt
tools/perf/builtin-record.c

index 236b9b9..667c14e 100644 (file)
@@ -55,7 +55,6 @@ counted. The following modifiers exist:
  S - read sample value (PERF_SAMPLE_READ)
  D - pin the event to the PMU
  W - group is weak and will fallback to non-group if not schedulable,
-     only supported in 'perf stat' for now.
 
 The 'p' modifier can be used for specifying how precise the instruction
 address should be. The 'p' modifier can be specified multiple times:
index 10cf889..488779b 100644 (file)
@@ -391,7 +391,12 @@ try_again:
                                        ui__warning("%s\n", msg);
                                goto try_again;
                        }
-
+                       if ((errno == EINVAL || errno == EBADF) &&
+                           pos->leader != pos &&
+                           pos->weak_group) {
+                               pos = perf_evlist__reset_weak_group(evlist, pos);
+                               goto try_again;
+                       }
                        rc = -errno;
                        perf_evsel__open_strerror(pos, &opts->target,
                                                  errno, msg, sizeof(msg));