Re: [PATCH v2 1/2] perf evlist: Ensure grouped events with same cpu map

From: Jin, Yao
Date: Wed May 27 2020 - 21:47:56 EST


Hi Jiri,

On 5/28/2020 12:28 AM, Jiri Olsa wrote:
On Wed, May 27, 2020 at 09:49:11PM +0800, Jin, Yao wrote:
Hi Jiri,

On 5/27/2020 6:28 PM, Jiri Olsa wrote:
On Wed, May 27, 2020 at 02:31:03PM +0800, Jin, Yao wrote:

SNIP

Thanks
Jin Yao

Issue is found!

It looks we can't set "pos->leader = pos" in either for_each_group_member()
or in for_each_group_evsel() because it may exit the iteration immediately.

evlist__for_each_entry(evlist, evsel) {
if (evsel->leader == evsel)
continue;

if (cpu_maps_matched(evsel->leader, evsel))
continue;

pr_warning("WARNING: event cpu maps are not fully matched, "
"disable group\n");

for_each_group_member(pos, evsel->leader) {
pos->leader = pos;
pos->core.nr_members = 0;
}

Let me use the example of '{cycles,unc_cbo_cache_lookup.any_i}' again.

In evlist:
cycles,
unc_cbo_cache_lookup.any_i,
unc_cbo_cache_lookup.any_i,
unc_cbo_cache_lookup.any_i,
unc_cbo_cache_lookup.any_i,

When we reach the for_each_group_member at first time, evsel is the first
unc_cbo_cache_lookup.any_i and evsel->leader is cycles. pos is same as the
evsel (the first unc_cbo_cache_lookup.any_i).

Once we execute "pos->leader = pos;", it's actually "evsel->leader = evsel".
So now evsel->leader is changed to the first unc_cbo_cache_lookup.any_i.

In next iteration, pos is the second unc_cbo_cache_lookup.any_i. pos->leader
is cycles but unfortunately evsel->leader has been changed to the first
unc_cbo_cache_lookup.any_i. So iteration stops immediately.

hum, AFAICS the iteration will not break but continue to next evsel and
pass the 'continue' for another group member.. what do I miss?

jirka


Let me use this example again.

cycles,
unc_cbo_cache_lookup.any_i,
unc_cbo_cache_lookup.any_i,
unc_cbo_cache_lookup.any_i,
unc_cbo_cache_lookup.any_i,

Yes, once for_each_group_member breaks (due to the issue in 'pos->leader =
pos'), evlist__for_each_entry will continue to the second
unc_cbo_cache_lookup.any_i. But now evsel->leader != evsel (evsel->leader is
"cycles"), so it will go to cpu_maps_matched.

But actually we don't need to go to cpu_maps_matched again.

for_each_group_member(pos, evsel->leader) {
pos->leader = pos;
pos->core.nr_members = 0;
}

If we solve the issue in above code, for_each_group_member doesn't break,
the leaders of all members in this group will be set to themselves.

if (evsel->leader == evsel)
continue;

I see.. the problem is in the for_each_group_member, how about
saving the leader into separate variable, like below

jirka


---
diff --git a/tools/perf/builtin-stat.c b/tools/perf/builtin-stat.c
index f789103d8306..a754cad3f5a0 100644
--- a/tools/perf/builtin-stat.c
+++ b/tools/perf/builtin-stat.c
@@ -189,6 +189,51 @@ static struct perf_stat_config stat_config = {
.big_num = true,
};
+static bool cpus_map_matched(struct evsel *a, struct evsel *b)
+{
+ if (!a->core.cpus && !b->core.cpus)
+ return true;
+
+ if (!a->core.cpus || !b->core.cpus)
+ return false;
+
+ if (a->core.cpus->nr != b->core.cpus->nr)
+ return false;
+
+ for (int i = 0; i < a->core.cpus->nr; i++) {
+ if (a->core.cpus->map[i] != b->core.cpus->map[i])
+ return false;
+ }
+
+ return true;
+}
+
+
+static void evlist__check_cpu_maps(struct evlist *evlist)
+{
+ struct evsel *evsel, *pos, *leader;
+
+ evlist__for_each_entry(evlist, evsel) {
+ char buf[1024];
+
+ leader = evsel->leader;
+ if (leader == evsel)
+ continue;
+ if (cpus_map_matched(leader, evsel))
+ continue;
+
+ evsel__group_desc(leader, buf, sizeof(buf));
+ WARN_ONCE(1, "WARNING: event cpu maps do not match, disabling group:\n");
+ pr_warning(" %s\n", buf);
+
+ for_each_group_evsel(pos, leader) {
+ pos->leader = pos;
+ pos->core.nr_members = 0;
+ }
+ evsel->leader->core.nr_members = 0;
+ }
+}
+
static inline void diff_timespec(struct timespec *r, struct timespec *a,
struct timespec *b)
{
@@ -1956,6 +2001,8 @@ int cmd_stat(int argc, const char **argv)
} else if (argc && !strncmp(argv[0], "rep", 3))
return __cmd_report(argc, argv);
+ evlist__check_cpu_maps(evsel_list);
+
interval = stat_config.interval;
timeout = stat_config.timeout;


This patch looks good. I guess you will post this patch, right? Thanks so much! :)

Thanks
Jin Yao