2019-04-01 14:21:16

by Thomas Richter

[permalink] [raw]
Subject: [PATCH] perf/report: [RFC] Handling OOM in perf report

A very strange -ENOMEM error which only when output
is displayed using the GTK GUI (I have copied the text from the GUI):

[root@m35lp76 perf]# ./perf report -i perf.data.error68-1

Processing events... [974K/3M]
Error:failed to process sample

0xf4198 [0x8]: failed to process type: 68

However when I use the same perf.data file with --stdio it works:

[root@m35lp76 perf]# ./perf report -i perf.data.error68-1 --stdio \
| head -12

# Total Lost Samples: 0
#
# Samples: 76K of event 'cycles'
# Event count (approx.): 99056160000
#
# Overhead Command Shared Object Symbol
# ........ ............... ................. .........
#
8.81% find [kernel.kallsyms] [k] ftrace_likely_update
8.74% swapper [kernel.kallsyms] [k] ftrace_likely_update
8.34% sshd [kernel.kallsyms] [k] ftrace_likely_update
2.19% kworker/u512:1- [kernel.kallsyms] [k] ftrace_likely_update

The sample precentage is a bit low.....

It always fails in the FINISHED_ROUND event, event number 68.

When happened is the following. Perf report calls a lot of functions and
down deep when a FINISHED_ROUND event is processed, these functions are
called:

perf_session__process_event()
+ perf_session__process_user_event()
+ process_finished_round()
+ ordered_events__flush()
+ __ordered_events__flush()
+ do_flush()
+ ordered_events__deliver_event()
+ perf_session__deliver_event()
+ machine__deliver_event()
+ perf_evlist__deliver_event()
+ process_sample_event()
+ hist_entry_iter_add() --> only called in GUI case!!!
+ hist_iter__report__callback()
+ symbol__inc_addr_sample()

Now this functions runs out of memory and
returns -ENOMEM. This is reported all the way up
until function

perf_session__process_event() returns to its caller, where -ENOMEM is
changed to -EINVAL and processing stops:

if ((skip = perf_session__process_event(session, event, head)) < 0) {
pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
head, event->header.size, event->header.type);
err = -EINVAL;
goto out_err;
}

This occured in the FINISHED_ROUND event when it has to process some
10000 entries and ran out of memory.

I understand that my perf.data file might just be too big, but I would
like to see some error message indicating OOM error instead of
processing failure of a unrelated event.

However this patch just does what the pr_debug() statement indicates,
the event is skipped and processing continues.
But at least the root cause is indicated and also shows up in the
GUI.

Signed-off-by: Thomas Richter <[email protected]>
---
tools/perf/builtin-report.c | 7 ++++++-
1 file changed, 6 insertions(+), 1 deletion(-)

diff --git a/tools/perf/builtin-report.c b/tools/perf/builtin-report.c
index 4054eb1f98ac..7a27b815f7a8 100644
--- a/tools/perf/builtin-report.c
+++ b/tools/perf/builtin-report.c
@@ -283,8 +283,13 @@ static int process_sample_event(struct perf_tool *tool,
al.map->dso->hit = 1;

ret = hist_entry_iter__add(&iter, &al, rep->max_stack, rep);
- if (ret < 0)
+ if (ret < 0) {
pr_debug("problem adding hist entry, skipping event\n");
+ if (ret == -ENOMEM) {
+ pr_err("Running out of memory\n");
+ ret = 0;
+ }
+ }
out_put:
addr_location__put(&al);
return ret;
--
2.16.4


2019-04-09 10:45:04

by Jiri Olsa

[permalink] [raw]
Subject: Re: [PATCH] perf/report: [RFC] Handling OOM in perf report

On Mon, Apr 01, 2019 at 04:20:00PM +0200, Thomas Richter wrote:

SNIP

> perf_session__process_event() returns to its caller, where -ENOMEM is
> changed to -EINVAL and processing stops:
>
> if ((skip = perf_session__process_event(session, event, head)) < 0) {
> pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
> head, event->header.size, event->header.type);
> err = -EINVAL;
> goto out_err;
> }
>
> This occured in the FINISHED_ROUND event when it has to process some
> 10000 entries and ran out of memory.
>
> I understand that my perf.data file might just be too big, but I would
> like to see some error message indicating OOM error instead of
> processing failure of a unrelated event.

you can limit the size of the report queue via ~/.perfconfig file:

[report]
queue-size=1M

above should use only 1M for the queue management data the data
for sample still get allocated thought.. but it could help

>
> However this patch just does what the pr_debug() statement indicates,
> the event is skipped and processing continues.
> But at least the root cause is indicated and also shows up in the
> GUI.
>
> Signed-off-by: Thomas Richter <[email protected]>
> ---
> tools/perf/builtin-report.c | 7 ++++++-
> 1 file changed, 6 insertions(+), 1 deletion(-)
>
> diff --git a/tools/perf/builtin-report.c b/tools/perf/builtin-report.c
> index 4054eb1f98ac..7a27b815f7a8 100644
> --- a/tools/perf/builtin-report.c
> +++ b/tools/perf/builtin-report.c
> @@ -283,8 +283,13 @@ static int process_sample_event(struct perf_tool *tool,
> al.map->dso->hit = 1;
>
> ret = hist_entry_iter__add(&iter, &al, rep->max_stack, rep);
> - if (ret < 0)
> + if (ret < 0) {
> pr_debug("problem adding hist entry, skipping event\n");
> + if (ret == -ENOMEM) {
> + pr_err("Running out of memory\n");
> + ret = 0;
> + }
> + }


I think we can propagate the error completely (like below),
and even warn about ENOMEM specificaly

but I think we should bail out in case of ENOMEM, because
the data are incomplete and misleading

jirka


---
diff --git a/tools/perf/util/session.c b/tools/perf/util/session.c
index b17f1c9bc965..eea247a26ad8 100644
--- a/tools/perf/util/session.c
+++ b/tools/perf/util/session.c
@@ -1933,7 +1933,7 @@ reader__process_events(struct reader *rd, struct perf_session *session,
pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
file_offset + head, event->header.size,
event->header.type);
- err = -EINVAL;
+ err = skip;
goto out;
}

2019-04-10 07:57:22

by Thomas Richter

[permalink] [raw]
Subject: Re: [PATCH] perf/report: [RFC] Handling OOM in perf report

On 4/9/19 12:42 PM, Jiri Olsa wrote:
> On Mon, Apr 01, 2019 at 04:20:00PM +0200, Thomas Richter wrote:
>
> SNIP
>
>> perf_session__process_event() returns to its caller, where -ENOMEM is
>> changed to -EINVAL and processing stops:
>>
>> if ((skip = perf_session__process_event(session, event, head)) < 0) {
>> pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
>> head, event->header.size, event->header.type);
>> err = -EINVAL;
>> goto out_err;
>> }
>>
>> This occured in the FINISHED_ROUND event when it has to process some
>> 10000 entries and ran out of memory.
>>
>> I understand that my perf.data file might just be too big, but I would
>> like to see some error message indicating OOM error instead of
>> processing failure of a unrelated event.
>
> you can limit the size of the report queue via ~/.perfconfig file:
>
> [report]
> queue-size=1M
>
> above should use only 1M for the queue management data the data
> for sample still get allocated thought.. but it could help
>
>>
>> However this patch just does what the pr_debug() statement indicates,
>> the event is skipped and processing continues.
>> But at least the root cause is indicated and also shows up in the
>> GUI.
>>
>> Signed-off-by: Thomas Richter <[email protected]>
>> ---
>> tools/perf/builtin-report.c | 7 ++++++-
>> 1 file changed, 6 insertions(+), 1 deletion(-)
>>
>> diff --git a/tools/perf/builtin-report.c b/tools/perf/builtin-report.c
>> index 4054eb1f98ac..7a27b815f7a8 100644
>> --- a/tools/perf/builtin-report.c
>> +++ b/tools/perf/builtin-report.c
>> @@ -283,8 +283,13 @@ static int process_sample_event(struct perf_tool *tool,
>> al.map->dso->hit = 1;
>>
>> ret = hist_entry_iter__add(&iter, &al, rep->max_stack, rep);
>> - if (ret < 0)
>> + if (ret < 0) {
>> pr_debug("problem adding hist entry, skipping event\n");
>> + if (ret == -ENOMEM) {
>> + pr_err("Running out of memory\n");
>> + ret = 0;
>> + }
>> + }
>
>
> I think we can propagate the error completely (like below),
> and even warn about ENOMEM specificaly
>
> but I think we should bail out in case of ENOMEM, because
> the data are incomplete and misleading
>
> jirka
>
>
> ---
> diff --git a/tools/perf/util/session.c b/tools/perf/util/session.c
> index b17f1c9bc965..eea247a26ad8 100644
> --- a/tools/perf/util/session.c
> +++ b/tools/perf/util/session.c
> @@ -1933,7 +1933,7 @@ reader__process_events(struct reader *rd, struct perf_session *session,
> pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
> file_offset + head, event->header.size,
> event->header.type);
> - err = -EINVAL;
> + err = skip;
> goto out;
> }
>
>

Above patch does not help, you simply return -ENOMEM instead of -EINVAL and processing
stops with no indication that perf ran out of memory. Bailing out in this case is ok.

I am fine with your patch, as long as it gives a reason why processing stopped.
In the GUI it shows on the bottom line the reason:

0xf4198 [0x8]: failed to process type: 68 [Cannot allocate memory]


diff --git a/tools/perf/util/session.c b/tools/perf/util/session.c
index b17f1c9bc965..e89716175588 100644
--- a/tools/perf/util/session.c
+++ b/tools/perf/util/session.c
@@ -1930,10 +1930,10 @@ reader__process_events(struct reader *rd, struct perf_session *session,

if (size < sizeof(struct perf_event_header) ||
(skip = rd->process(session, event, file_pos)) < 0) {
- pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
+ pr_err("%#" PRIx64 " [%#x]: failed to process type: %d [%s]\n",
file_offset + head, event->header.size,
- event->header.type);
- err = -EINVAL;
+ event->header.type, strerror(-skip));
+ err = skip;
goto out;
}

[root@m35lp76 perf]#
--
Thomas Richter, Dept 3252, IBM s390 Linux Development, Boeblingen, Germany
--
Vorsitzender des Aufsichtsrats: Matthias Hartmann
Geschäftsführung: Dirk Wittkopp
Sitz der Gesellschaft: Böblingen / Registergericht: Amtsgericht Stuttgart, HRB 243294

2019-04-15 08:29:08

by Jiri Olsa

[permalink] [raw]
Subject: Re: [PATCH] perf/report: [RFC] Handling OOM in perf report

On Wed, Apr 10, 2019 at 09:17:30AM +0200, Thomas-Mich Richter wrote:

SNIP

> >
> >
> > ---
> > diff --git a/tools/perf/util/session.c b/tools/perf/util/session.c
> > index b17f1c9bc965..eea247a26ad8 100644
> > --- a/tools/perf/util/session.c
> > +++ b/tools/perf/util/session.c
> > @@ -1933,7 +1933,7 @@ reader__process_events(struct reader *rd, struct perf_session *session,
> > pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
> > file_offset + head, event->header.size,
> > event->header.type);
> > - err = -EINVAL;
> > + err = skip;
> > goto out;
> > }
> >
> >
>
> Above patch does not help, you simply return -ENOMEM instead of -EINVAL and processing
> stops with no indication that perf ran out of memory. Bailing out in this case is ok.
>
> I am fine with your patch, as long as it gives a reason why processing stopped.
> In the GUI it shows on the bottom line the reason:
>
> 0xf4198 [0x8]: failed to process type: 68 [Cannot allocate memory]

looks good, please post the full patch

thanks,
jirka

>
>
> diff --git a/tools/perf/util/session.c b/tools/perf/util/session.c
> index b17f1c9bc965..e89716175588 100644
> --- a/tools/perf/util/session.c
> +++ b/tools/perf/util/session.c
> @@ -1930,10 +1930,10 @@ reader__process_events(struct reader *rd, struct perf_session *session,
>
> if (size < sizeof(struct perf_event_header) ||
> (skip = rd->process(session, event, file_pos)) < 0) {
> - pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
> + pr_err("%#" PRIx64 " [%#x]: failed to process type: %d [%s]\n",
> file_offset + head, event->header.size,
> - event->header.type);
> - err = -EINVAL;
> + event->header.type, strerror(-skip));
> + err = skip;
> goto out;
> }
>
> [root@m35lp76 perf]#
> --
> Thomas Richter, Dept 3252, IBM s390 Linux Development, Boeblingen, Germany
> --
> Vorsitzender des Aufsichtsrats: Matthias Hartmann
> Gesch?ftsf?hrung: Dirk Wittkopp
> Sitz der Gesellschaft: B?blingen / Registergericht: Amtsgericht Stuttgart, HRB 243294
>