Re: [PATCH] perf parse-events: Avoid two scenarios involving the reordering of topdown events

From: Yanfei Xu
Date: Fri Apr 12 2024 - 06:23:05 EST

Next message: Ingo Molnar: "Re: [PATCH] sched: Add missing memory barrier in switch_mm_cid"
Previous message: Martin K. Petersen: "Re: linux-next: build failure after merge of the scsi-mkp tree"
In reply to: Ian Rogers: "Re: [PATCH] perf parse-events: Avoid two scenarios involving the reordering of topdown events"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]

Hi Ian,

On 4/11/2024 11:14 PM, Ian Rogers wrote:

On Thu, Apr 11, 2024 at 7:53 AM Yanfei Xu <yanfei.xu@xxxxxxxxx> wrote:

We found that even an events group with slots but without topdown events
will still reroder to place the slots first. It's unnecessary, and may
break tools expecting the command line order to match the printed order.
The issue was previously fixed [1], but was later discarded since [2].

The requirements for topdown events is a gift that keeps on giving :-( Please:
1) provide the tool that is broken,
2) provide a reason why the tool is depending on text output that may
vary over time (intended audience for it is humans) instead of say the
json output that is intended for consumption by tools,
3) provide a test.

We noticed this issue from the regroup WARNNING which didn't trigger in
past regular testing. After digging into the code change I noticed that
there was also a fix for this kind of unnecessary reorder with saying
tools may expecting the original command line order. From my side, I
don't have the tool.
You can refer to the link[1]. Actually it was from you. :)

[1] https://lore.kernel.org/lkml/20220321223344.1034479-1-irogers@xxxxxxxxxx/#/

Add an extra check of evsel leader, variable must_be_in_group, to ensure
the slots event is only moved if the group has non-slots topdown events.

Without the patch:

$ perf stat -e '{cpu/cpu-cycles/,slots}' sleep 1
WARNING: events were regrouped to match PMUs

Performance counter stats for 'sleep 1':

2,663,256 slots:u
443,876 cpu/cpu-cycles/u

1.001079566 seconds time elapsed

0.001054000 seconds user
0.000000000 seconds sys

With the patch:

$ perf stat -e '{cpu/cpu-cycles/,slots}' sleep 1

Performance counter stats for 'sleep 1':

469,039 cpu/cpu-cycles/u
2,814,234 slots:u

1.001148306 seconds time elapsed

0.001123000 seconds user
0.000000000 seconds sys

In cases where both slots and topdown events are present, moving the
slots event to be the first event is necessary. However there is no
requirement to move the topdown events to be adjacent to slots event.
So keep the original order of the topdown events is expected. Further
more, if a group doesn't have slots event, the topdown events will be
unexpectedly moved to the head of the group.

Remove the movements regarding topdown events in arch_evlist__cmp()

Without the patch:

$ perf stat -e '{slots,cpu/cpu-cycles/,cpu/topdown-bad-spec/}' sleep 1
WARNING: events were regrouped to match PMUs

Performance counter stats for 'sleep 1':

2,681,460 slots:u
336,496 cpu/topdown-bad-spec/u
446,910 cpu/cpu-cycles/u

1.001210088 seconds time elapsed

0.001160000 seconds user
0.000000000 seconds sys

With the patch:

$ perf stat -e '{slots,cpu/cpu-cycles/,cpu/topdown-bad-spec/}' sleep 1

Performance counter stats for 'sleep 1':

2,715,486 slots:u
452,581 cpu/cpu-cycles/u
340,766 cpu/topdown-bad-spec/u

1.001116709 seconds time elapsed

0.001111000 seconds user
0.000000000 seconds sys

[1] https://lore.kernel.org/lkml/20220321223344.1034479-1-irogers@xxxxxxxxxx/#/
[2] https://lore.kernel.org/lkml/20230302041211.852330-10-irogers@xxxxxxxxxx/#/

Fixes: 347c2f0a0988 ("perf parse-events: Sort and group parsed events")
Signed-off-by: Yanfei Xu <yanfei.xu@xxxxxxxxx>
---
tools/perf/arch/x86/util/evlist.c | 13 +++++--------
tools/perf/arch/x86/util/evsel.c | 6 ++++++
tools/perf/util/evsel.h | 2 ++
tools/perf/util/parse-events.c | 9 ++++++---
4 files changed, 19 insertions(+), 11 deletions(-)

diff --git a/tools/perf/arch/x86/util/evlist.c b/tools/perf/arch/x86/util/evlist.c
index b1ce0c52d88d..eed0a74c561a 100644
--- a/tools/perf/arch/x86/util/evlist.c
+++ b/tools/perf/arch/x86/util/evlist.c
@@ -75,17 +75,14 @@ int arch_evlist__add_default_attrs(struct evlist *evlist,

int arch_evlist__cmp(const struct evsel *lhs, const struct evsel *rhs)
{
+ struct evsel *leader;
+
if (topdown_sys_has_perf_metrics() &&
(arch_evsel__must_be_in_group(lhs) || arch_evsel__must_be_in_group(rhs))) {
+ leader = evsel__leader(rhs);

Usually the leader doesn't make sense at this point. For example, a
metric following event parsing may group
{cycles,slots,topdown-fe-bound} and cycles will be the leader. The
leader's properties say nothing about the ordering of slots and
topdown-fe-bound.

Yes, the leader at this point is temporary. Actually what we want here
is to get a information of the group the event located at to know if
group contains topdown events.

/* Ensure the topdown slots comes first. */
- if (strcasestr(lhs->name, "slots") && !strcasestr(lhs->name, "uops_retired.slots"))
- return -1;
- if (strcasestr(rhs->name, "slots") && !strcasestr(rhs->name, "uops_retired.slots"))
- return 1;
- /* Followed by topdown events. */
- if (strcasestr(lhs->name, "topdown") && !strcasestr(rhs->name, "topdown"))
- return -1;
- if (!strcasestr(lhs->name, "topdown") && strcasestr(rhs->name, "topdown"))
+ if (strcasestr(rhs->name, "slots") && !strcasestr(rhs->name, "uops_retired.slots")
+ && leader->must_be_in_group)

Why isn't this symmetric? As in the lhs has a test that will return -1.

The full code of this function is as below, the last "return" can keep
the -1 case.

int arch_evlist__cmp(const struct evsel *lhs, const struct evsel *rhs)
{
struct evsel *leader;

if (topdown_sys_has_perf_metrics() &&
(arch_evsel__must_be_in_group(lhs) || arch_evsel__must_be_in_group(rhs))) {
leader = evsel__leader(rhs);
/* Ensure the topdown slots comes first. */
if (strcasestr(rhs->name, "slots") && !strcasestr(rhs->name, "uops_retired.slots")
&& leader->must_be_in_group)
return 1;
}

/* Default ordering by insertion index. */
return lhs->core.idx - rhs->core.idx;
}

return 1;
}

diff --git a/tools/perf/arch/x86/util/evsel.c b/tools/perf/arch/x86/util/evsel.c
index 090d0f371891..16f42fcfbe0b 100644
--- a/tools/perf/arch/x86/util/evsel.c
+++ b/tools/perf/arch/x86/util/evsel.c
@@ -44,6 +44,12 @@ bool arch_evsel__must_be_in_group(const struct evsel *evsel)
strcasestr(evsel->name, "uops_retired.slots"))
return false;

+ if (strcasestr(evsel->name, "topdown")) {
+ struct evsel *leader = evsel__leader(evsel);
+
+ leader->must_be_in_group = true;
+ }
+
return strcasestr(evsel->name, "topdown") || strcasestr(evsel->name, "slots");
}

diff --git a/tools/perf/util/evsel.h b/tools/perf/util/evsel.h
index 517cff431de2..a7ab266bc915 100644
--- a/tools/perf/util/evsel.h
+++ b/tools/perf/util/evsel.h
@@ -98,6 +98,8 @@ struct evsel {
bool bpf_counter;
bool use_config_name;
bool skippable;
+ /* any evsels with the flag set must be in a group */
+ bool must_be_in_group;

This duplicates arch_evsel__must_be_in_group but one may be true when
the other isn't. This doesn't make sense to me.

Indeed, this name might be not much accurate. The good point is that it
can keep all checks of "topdown" and "slots" being completed in
arch_evsel__must_be_in_group(). How about rename it as
group_has_topdown? or do you have any suggestion?

I think this patch needs a rethink and the behavior with it looks
broken and more complex than what already exists. I also think the

I was thinking to check the existence of topdown event in
arch_evlist__cmp(). It could make the fix more consistent. However the
evlist at comparison stage of sorting is broken, not traversable. So the
check is moved in front of sorting list.

problem is one that is more to do with the tool using the output than
the reordering.

In my opinion, moving slots when topdown is present is neccesary. But
rerodering in above cases is unneccesary, just for looking more
comfortable? why not users do that by themself if need when inputting,
instead of perf do that only for topdown.

Thanks,
Yanfei

Thanks,
Ian

int bpf_fd;
struct bpf_object *bpf_obj;
struct list_head config_terms;
diff --git a/tools/perf/util/parse-events.c b/tools/perf/util/parse-events.c
index 6f8b0fa17689..37950056a661 100644
--- a/tools/perf/util/parse-events.c
+++ b/tools/perf/util/parse-events.c
@@ -2052,9 +2052,12 @@ static int parse_events__sort_events_and_fix_groups(struct list_head *list)
*/
pos->core.idx = idx++;

- /* Remember an index to sort all forced grouped events together to. */
- if (force_grouped_idx == -1 && pos == pos_leader && pos->core.nr_members < 2 &&
- arch_evsel__must_be_in_group(pos))
+ /*
+ * Remember an index to sort all forced grouped events together to,
+ * and check each evsel for setting must_be_in_group of its leader.
+ */
+ if (arch_evsel__must_be_in_group(pos) && force_grouped_idx == -1 &&
+ pos == pos_leader && pos->core.nr_members < 2)
force_grouped_idx = pos->core.idx;
}

--
2.40.1

Next message: Ingo Molnar: "Re: [PATCH] sched: Add missing memory barrier in switch_mm_cid"
Previous message: Martin K. Petersen: "Re: linux-next: build failure after merge of the scsi-mkp tree"
In reply to: Ian Rogers: "Re: [PATCH] perf parse-events: Avoid two scenarios involving the reordering of topdown events"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]