perf evsel: Move ignore_missing_thread() to fallback code
authorRiccardo Mancini <rickyman7@gmail.com>
Sat, 21 Aug 2021 09:19:27 +0000 (11:19 +0200)
committerArnaldo Carvalho de Melo <acme@redhat.com>
Tue, 31 Aug 2021 19:44:30 +0000 (16:44 -0300)
This patch moves ignore_missing_thread outside the perf_event_open loop.

Doing so, we need to move the retry_open flag a few places higher, with
minimal impact. Furthermore, thread need not be decreased since it won't
get increased by the for loop (since we're jumping back inside), but we
need to check that the nthreads decrease didn't put thread out of range.

The goal is to have fallbacks handled in one place only, since in the
future parallel code, these would be handled separately.

Signed-off-by: Riccardo Mancini <rickyman7@gmail.com>
Cc: Ian Rogers <irogers@google.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lore.kernel.org/lkml/4eca51443c786baaf6811b7cd8e73aafd97f7606.1629490974.git.rickyman7@gmail.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
tools/perf/util/evsel.c
tools/perf/util/evsel.h

index f370b88..f691f9e 100644 (file)
@@ -1656,7 +1656,7 @@ static int update_fds(struct evsel *evsel,
        return 0;
 }
 
-static bool ignore_missing_thread(struct evsel *evsel,
+bool evsel__ignore_missing_thread(struct evsel *evsel,
                                  int nr_cpus, int cpu,
                                  struct perf_thread_map *threads,
                                  int thread, int err)
@@ -1993,12 +1993,15 @@ fallback_missing_features:
 
                for (thread = 0; thread < nthreads; thread++) {
                        int fd, group_fd;
+retry_open:
+                       if (thread >= nthreads)
+                               break;
 
                        if (!evsel->cgrp && !evsel->core.system_wide)
                                pid = perf_thread_map__pid(threads, thread);
 
                        group_fd = get_group_fd(evsel, cpu, thread);
-retry_open:
+
                        test_attr__ready();
 
                        fd = perf_event_open(evsel, pid, cpus->map[cpu],
@@ -2016,20 +2019,6 @@ retry_open:
                        if (fd < 0) {
                                err = -errno;
 
-                               if (ignore_missing_thread(evsel, cpus->nr, cpu, threads, thread, err)) {
-                                       /*
-                                        * We just removed 1 thread, so take a step
-                                        * back on thread index and lower the upper
-                                        * nthreads limit.
-                                        */
-                                       nthreads--;
-                                       thread--;
-
-                                       /* ... and pretend like nothing have happened. */
-                                       err = 0;
-                                       continue;
-                               }
-
                                pr_debug2_peo("\nsys_perf_event_open failed, error %d\n",
                                          err);
                                goto try_fallback;
@@ -2069,6 +2058,14 @@ retry_open:
        return 0;
 
 try_fallback:
+       if (evsel__ignore_missing_thread(evsel, cpus->nr, cpu, threads, thread, err)) {
+               /* We just removed 1 thread, so lower the upper nthreads limit. */
+               nthreads--;
+
+               /* ... and pretend like nothing have happened. */
+               err = 0;
+               goto retry_open;
+       }
        /*
         * perf stat needs between 5 and 22 fds per CPU. When we run out
         * of them try to increase the limits.
index 47916db..b749bee 100644 (file)
@@ -294,6 +294,11 @@ bool evsel__detect_missing_features(struct evsel *evsel);
 enum rlimit_action { NO_CHANGE, SET_TO_MAX, INCREASED_MAX };
 bool evsel__increase_rlimit(enum rlimit_action *set_rlimit);
 
+bool evsel__ignore_missing_thread(struct evsel *evsel,
+                                 int nr_cpus, int cpu,
+                                 struct perf_thread_map *threads,
+                                 int thread, int err);
+
 struct perf_sample;
 
 void *evsel__rawptr(struct evsel *evsel, struct perf_sample *sample, const char *name);