2020-01-31 05:27:35

by kajoljain

[permalink] [raw]
Subject: [PATCH v3] tools/perf/metricgroup: Fix printing event names of metric group with multiple events incase of overlapping events

Commit f01642e4912b ("perf metricgroup: Support multiple
events for metricgroup") introduced support for multiple events
in a metric group. But with the current upstream, metric events
names are not printed properly incase we try to run multiple
metric groups with overlapping event.

With current upstream version, incase of overlapping metric events
issue is, we always start our comparision logic from start.
So, the events which already matched with some metric group also
take part in comparision logic. Because of that when we have overlapping
events, we end up matching current metric group event with already matched
one.

For example, in skylake machine we have metric event CoreIPC and
Instructions. Both of them need 'inst_retired.any' event value.
As events in Instructions is subset of events in CoreIPC, they
endup in pointing to same 'inst_retired.any' value.

In skylake platform:

command:# ./perf stat -M CoreIPC,Instructions -C 0 sleep 1

Performance counter stats for 'CPU(s) 0':

1,254,992,790 inst_retired.any # 1254992790.0
Instructions
# 1.3 CoreIPC
977,172,805 cycles
1,254,992,756 inst_retired.any

1.000802596 seconds time elapsed

command:# sudo ./perf stat -M UPI,IPC sleep 1

Performance counter stats for 'sleep 1':

948,650 uops_retired.retire_slots
866,182 inst_retired.any # 0.7 IPC
866,182 inst_retired.any
1,175,671 cpu_clk_unhalted.thread

Patch fixes the issue by adding a new array 'evlist_used' to keep track of
events which already matched with some group by setting it true.
So, we skip all used events in list when we start comparision logic.
Patch also make some changes in comparision logic, incase we get a match
miss, we discard the whole match and start again with first event id in
metric event.

With this patch:
In skylake platform:

command:# ./perf stat -M CoreIPC,Instructions -C 0 sleep 1

Performance counter stats for 'CPU(s) 0':

3,348,415 inst_retired.any # 0.3 CoreIPC
11,779,026 cycles
3,348,381 inst_retired.any # 3348381.0
Instructions

1.001649056 seconds time elapsed

command:# ./perf stat -M UPI,IPC sleep 1

Performance counter stats for 'sleep 1':

1,023,148 uops_retired.retire_slots # 1.1 UPI
924,976 inst_retired.any
924,976 inst_retired.any # 0.6 IPC
1,489,414 cpu_clk_unhalted.thread

1.003064672 seconds time elapsed

Signed-off-by: Kajol Jain <[email protected]>
Suggested-by: Jiri Olsa <[email protected]>
Cc: Alexander Shishkin <[email protected]>
Cc: Andi Kleen <[email protected]>
Cc: Kan Liang <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Cc: Jin Yao <[email protected]>
Cc: Madhavan Srinivasan <[email protected]>
Cc: Anju T Sudhakar <[email protected]>
Cc: Ravi Bangoria <[email protected]>
---
tools/perf/util/metricgroup.c | 42 +++++++++++++++++++----------------
1 file changed, 23 insertions(+), 19 deletions(-)

---
Changelog:

v2 -> v3
- Add array in place of variable to keep track of matched events.
Because incase we miss match in previous approach, all events will
be rolled over in next condition. So, rather we add array and set
it incase that variable already match with some group.
- Suggested by Jiri Olsa

v1 -> v2
- Rather then adding static variable in metricgroup.c,
add a new variable in evlist itself with name 'evlist_iter'
---
diff --git a/tools/perf/util/metricgroup.c b/tools/perf/util/metricgroup.c
index 02aee946b6c1..e9d9fbaa2160 100644
--- a/tools/perf/util/metricgroup.c
+++ b/tools/perf/util/metricgroup.c
@@ -93,13 +93,16 @@ struct egroup {
static struct evsel *find_evsel_group(struct evlist *perf_evlist,
const char **ids,
int idnum,
- struct evsel **metric_events)
+ struct evsel **metric_events,
+ bool evlist_used[])
{
struct evsel *ev;
- int i = 0;
+ int i = 0, j = 0;
bool leader_found;

evlist__for_each_entry (perf_evlist, ev) {
+ if (evlist_used[j++])
+ continue;
if (!strcmp(ev->name, ids[i])) {
if (!metric_events[i])
metric_events[i] = ev;
@@ -107,22 +110,17 @@ static struct evsel *find_evsel_group(struct evlist *perf_evlist,
if (i == idnum)
break;
} else {
- if (i + 1 == idnum) {
- /* Discard the whole match and start again */
- i = 0;
- memset(metric_events, 0,
- sizeof(struct evsel *) * idnum);
- continue;
- }
-
- if (!strcmp(ev->name, ids[i]))
- metric_events[i] = ev;
- else {
- /* Discard the whole match and start again */
- i = 0;
- memset(metric_events, 0,
- sizeof(struct evsel *) * idnum);
- continue;
+ /* Discard the whole match and start again */
+ i = 0;
+ memset(metric_events, 0,
+ sizeof(struct evsel *) * idnum);
+
+ if (!strcmp(ev->name, ids[i])) {
+ if (!metric_events[i])
+ metric_events[i] = ev;
+ i++;
+ if (i == idnum)
+ break;
}
}
}
@@ -144,7 +142,10 @@ static struct evsel *find_evsel_group(struct evlist *perf_evlist,
!strcmp(ev->name, metric_events[i]->name)) {
ev->metric_leader = metric_events[i];
}
+ j++;
}
+ ev = metric_events[i];
+ evlist_used[ev->idx] = true;
}

return metric_events[0];
@@ -160,6 +161,9 @@ static int metricgroup__setup_events(struct list_head *groups,
int ret = 0;
struct egroup *eg;
struct evsel *evsel;
+ bool evlist_used[perf_evlist->core.nr_entries];
+
+ memset(evlist_used, 0, perf_evlist->core.nr_entries);

list_for_each_entry (eg, groups, nd) {
struct evsel **metric_events;
@@ -170,7 +174,7 @@ static int metricgroup__setup_events(struct list_head *groups,
break;
}
evsel = find_evsel_group(perf_evlist, eg->ids, eg->idnum,
- metric_events);
+ metric_events, evlist_used);
if (!evsel) {
pr_debug("Cannot resolve %s: %s\n",
eg->metric_name, eg->metric_expr);
--
2.21.0


2020-02-06 18:46:59

by Jiri Olsa

[permalink] [raw]
Subject: Re: [PATCH v3] tools/perf/metricgroup: Fix printing event names of metric group with multiple events incase of overlapping events

On Fri, Jan 31, 2020 at 10:55:22AM +0530, Kajol Jain wrote:

SNIP

> ev->metric_leader = metric_events[i];
> }
> + j++;
> }
> + ev = metric_events[i];
> + evlist_used[ev->idx] = true;
> }
>
> return metric_events[0];
> @@ -160,6 +161,9 @@ static int metricgroup__setup_events(struct list_head *groups,
> int ret = 0;
> struct egroup *eg;
> struct evsel *evsel;
> + bool evlist_used[perf_evlist->core.nr_entries];
> +
> + memset(evlist_used, 0, perf_evlist->core.nr_entries);

I know I posted this in the previous email, but are we sure bool
is always 1 byte? would sizeod(evlist_used) be safer?

other than that it looks ok

Andi, you're ok with this?

thanks,
jirka

>
> list_for_each_entry (eg, groups, nd) {
> struct evsel **metric_events;
> @@ -170,7 +174,7 @@ static int metricgroup__setup_events(struct list_head *groups,
> break;
> }
> evsel = find_evsel_group(perf_evlist, eg->ids, eg->idnum,
> - metric_events);
> + metric_events, evlist_used);
> if (!evsel) {
> pr_debug("Cannot resolve %s: %s\n",
> eg->metric_name, eg->metric_expr);
> --
> 2.21.0
>

2020-02-06 19:00:31

by Joe Perches

[permalink] [raw]
Subject: Re: [PATCH v3] tools/perf/metricgroup: Fix printing event names of metric group with multiple events incase of overlapping events

On Thu, 2020-02-06 at 19:45 +0100, Jiri Olsa wrote:
> On Fri, Jan 31, 2020 at 10:55:22AM +0530, Kajol Jain wrote:
>
> SNIP
>
> > ev->metric_leader = metric_events[i];
> > }
> > + j++;
> > }
> > + ev = metric_events[i];
> > + evlist_used[ev->idx] = true;
> > }
> >
> > return metric_events[0];
> > @@ -160,6 +161,9 @@ static int metricgroup__setup_events(struct list_head *groups,
> > int ret = 0;
> > struct egroup *eg;
> > struct evsel *evsel;
> > + bool evlist_used[perf_evlist->core.nr_entries];
> > +
> > + memset(evlist_used, 0, perf_evlist->core.nr_entries);
>
> I know I posted this in the previous email, but are we sure bool
> is always 1 byte? would sizeod(evlist_used) be safer?
>
> other than that it looks ok
>
> Andi, you're ok with this?

stack declarations of variable length arrays are not
a good thing.

https://lwn.net/Articles/749089/

and

bool evlist_used[perf_evlist->core.nr_entries] = {};

would be better.


2020-02-06 20:12:31

by Andi Kleen

[permalink] [raw]
Subject: Re: [PATCH v3] tools/perf/metricgroup: Fix printing event names of metric group with multiple events incase of overlapping events

On Thu, Feb 06, 2020 at 07:45:10PM +0100, Jiri Olsa wrote:
> On Fri, Jan 31, 2020 at 10:55:22AM +0530, Kajol Jain wrote:
>
> SNIP
>
> > ev->metric_leader = metric_events[i];
> > }
> > + j++;
> > }
> > + ev = metric_events[i];
> > + evlist_used[ev->idx] = true;
> > }
> >
> > return metric_events[0];
> > @@ -160,6 +161,9 @@ static int metricgroup__setup_events(struct list_head *groups,
> > int ret = 0;
> > struct egroup *eg;
> > struct evsel *evsel;
> > + bool evlist_used[perf_evlist->core.nr_entries];
> > +
> > + memset(evlist_used, 0, perf_evlist->core.nr_entries);
>
> I know I posted this in the previous email, but are we sure bool
> is always 1 byte? would sizeod(evlist_used) be safer?
>
> other than that it looks ok
>
> Andi, you're ok with this?

Yes.

-Andi

2020-02-10 12:13:05

by Jiri Olsa

[permalink] [raw]
Subject: Re: [PATCH v3] tools/perf/metricgroup: Fix printing event names of metric group with multiple events incase of overlapping events

On Thu, Feb 06, 2020 at 10:58:12AM -0800, Joe Perches wrote:
> On Thu, 2020-02-06 at 19:45 +0100, Jiri Olsa wrote:
> > On Fri, Jan 31, 2020 at 10:55:22AM +0530, Kajol Jain wrote:
> >
> > SNIP
> >
> > > ev->metric_leader = metric_events[i];
> > > }
> > > + j++;
> > > }
> > > + ev = metric_events[i];
> > > + evlist_used[ev->idx] = true;
> > > }
> > >
> > > return metric_events[0];
> > > @@ -160,6 +161,9 @@ static int metricgroup__setup_events(struct list_head *groups,
> > > int ret = 0;
> > > struct egroup *eg;
> > > struct evsel *evsel;
> > > + bool evlist_used[perf_evlist->core.nr_entries];
> > > +
> > > + memset(evlist_used, 0, perf_evlist->core.nr_entries);
> >
> > I know I posted this in the previous email, but are we sure bool
> > is always 1 byte? would sizeod(evlist_used) be safer?
> >
> > other than that it looks ok
> >
> > Andi, you're ok with this?
>
> stack declarations of variable length arrays are not
> a good thing.
>
> https://lwn.net/Articles/749089/
>
> and
>
> bool evlist_used[perf_evlist->core.nr_entries] = {};

hum, I think we already have few of them in perf ;-)
thanks for the link

right, that initialization is of course much better, thanks

jirka

2020-02-10 18:01:28

by Andi Kleen

[permalink] [raw]
Subject: Re: [PATCH v3] tools/perf/metricgroup: Fix printing event names of metric group with multiple events incase of overlapping events

> > stack declarations of variable length arrays are not
> > a good thing.
> >
> > https://lwn.net/Articles/749089/
> >
> > and
> >
> > bool evlist_used[perf_evlist->core.nr_entries] = {};
>
> hum, I think we already have few of them in perf ;-)

For user space they don't really matter as long as the size is
not totally out of bound, it has a fairly large stack compared
to the kernel, and also is less security sensitive.

-Andi

2020-02-11 12:17:05

by kajoljain

[permalink] [raw]
Subject: Re: [PATCH v3] tools/perf/metricgroup: Fix printing event names of metric group with multiple events incase of overlapping events



On 2/10/20 5:41 PM, Jiri Olsa wrote:
> On Thu, Feb 06, 2020 at 10:58:12AM -0800, Joe Perches wrote:
>> On Thu, 2020-02-06 at 19:45 +0100, Jiri Olsa wrote:
>>> On Fri, Jan 31, 2020 at 10:55:22AM +0530, Kajol Jain wrote:
>>>
>>> SNIP
>>>
>>>> ev->metric_leader = metric_events[i];
>>>> }
>>>> + j++;
>>>> }
>>>> + ev = metric_events[i];
>>>> + evlist_used[ev->idx] = true;
>>>> }
>>>>
>>>> return metric_events[0];
>>>> @@ -160,6 +161,9 @@ static int metricgroup__setup_events(struct list_head *groups,
>>>> int ret = 0;
>>>> struct egroup *eg;
>>>> struct evsel *evsel;
>>>> + bool evlist_used[perf_evlist->core.nr_entries];
>>>> +
>>>> + memset(evlist_used, 0, perf_evlist->core.nr_entries);
>>>
>>> I know I posted this in the previous email, but are we sure bool
>>> is always 1 byte? would sizeod(evlist_used) be safer?


Hi jiri,
Yes you are right. We should use 'evlist_used' size itself.

>>>
>>> other than that it looks ok
>>>
>>> Andi, you're ok with this?
>>
>> stack declarations of variable length arrays are not
>> a good thing.
>>
>> https://lwn.net/Articles/749089/
>>
>> and
>>
>> bool evlist_used[perf_evlist->core.nr_entries] = {};


I am planning to use calloc and free that memory later in function 'metricgroup__setup_events'.
Something like this.


+ bool *evlist_used;
+
+ evlist_used = (bool *)calloc(perf_evlist->core.nr_entries,
+ sizeof(bool));
+ if (!evlist_used) {
+ ret = -ENOMEM;
+ break;
+ }

Please let me know if its looking fine.

Thanks,
Kajol

>
> hum, I think we already have few of them in perf ;-)
> thanks for the link
>
> right, that initialization is of course much better, thanks
>
> jirka
>

2020-02-11 13:57:08

by Jiri Olsa

[permalink] [raw]
Subject: Re: [PATCH v3] tools/perf/metricgroup: Fix printing event names of metric group with multiple events incase of overlapping events

On Tue, Feb 11, 2020 at 04:50:41PM +0530, kajoljain wrote:
>
>
> On 2/10/20 5:41 PM, Jiri Olsa wrote:
> > On Thu, Feb 06, 2020 at 10:58:12AM -0800, Joe Perches wrote:
> >> On Thu, 2020-02-06 at 19:45 +0100, Jiri Olsa wrote:
> >>> On Fri, Jan 31, 2020 at 10:55:22AM +0530, Kajol Jain wrote:
> >>>
> >>> SNIP
> >>>
> >>>> ev->metric_leader = metric_events[i];
> >>>> }
> >>>> + j++;
> >>>> }
> >>>> + ev = metric_events[i];
> >>>> + evlist_used[ev->idx] = true;
> >>>> }
> >>>>
> >>>> return metric_events[0];
> >>>> @@ -160,6 +161,9 @@ static int metricgroup__setup_events(struct list_head *groups,
> >>>> int ret = 0;
> >>>> struct egroup *eg;
> >>>> struct evsel *evsel;
> >>>> + bool evlist_used[perf_evlist->core.nr_entries];
> >>>> +
> >>>> + memset(evlist_used, 0, perf_evlist->core.nr_entries);
> >>>
> >>> I know I posted this in the previous email, but are we sure bool
> >>> is always 1 byte? would sizeod(evlist_used) be safer?
>
>
> Hi jiri,
> Yes you are right. We should use 'evlist_used' size itself.
>
> >>>
> >>> other than that it looks ok
> >>>
> >>> Andi, you're ok with this?
> >>
> >> stack declarations of variable length arrays are not
> >> a good thing.
> >>
> >> https://lwn.net/Articles/749089/
> >>
> >> and
> >>
> >> bool evlist_used[perf_evlist->core.nr_entries] = {};
>
>
> I am planning to use calloc and free that memory later in function 'metricgroup__setup_events'.
> Something like this.
>
>
> + bool *evlist_used;
> +
> + evlist_used = (bool *)calloc(perf_evlist->core.nr_entries,
> + sizeof(bool));
> + if (!evlist_used) {
> + ret = -ENOMEM;
> + break;
> + }
>
> Please let me know if its looking fine.

I'm also ok with the array on the stack, but I don't mind
this change as well

thanks,
jirka