[PATCH 0/2] Fix perf-tool crashes when kvm guest is running

[PATCH 0/2] Fix perf-tool crashes when kvm guest is running

[Joerg Roedel]

Hi,

here are the fixes I promised. These patches should fix the crashes
David and Jason have seen by making sure no sample is processed when
ther is no valid machine object for it. The second patch also changes
the perf_guest default back to false.

Regards,

	Joerg

[PATCH 1/2] perf-tool: Don't process samples with no valid machine object

[Joerg Roedel]

The perf sample processing code relies on a valid machine
object. Make sure that this path is only entered when such a
object exists.

Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
---
 tools/perf/builtin-top.c  |    2 +-
 tools/perf/util/session.c |    2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/tools/perf/builtin-top.c b/tools/perf/builtin-top.c
index dd162aa..7c81d1d 100644
--- a/tools/perf/builtin-top.c
+++ b/tools/perf/builtin-top.c
@@ -805,7 +805,7 @@ static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
 		}
 
 
-		if (event->header.type == PERF_RECORD_SAMPLE) {
+		if (event->header.type == PERF_RECORD_SAMPLE && machine != NULL) {
 			perf_event__process_sample(&top->tool, event, evsel,
 						   &sample, machine);
 		} else if (event->header.type < PERF_RECORD_MAX) {
diff --git a/tools/perf/util/session.c b/tools/perf/util/session.c
index b5ca2558..c9593c7 100644
--- a/tools/perf/util/session.c
+++ b/tools/perf/util/session.c
@@ -792,7 +792,7 @@ static int perf_session_deliver_event(struct perf_session *session,
 	switch (event->header.type) {
 	case PERF_RECORD_SAMPLE:
 		dump_sample(session, event, sample);
-		if (evsel == NULL) {
+		if (evsel == NULL || machine == NULL) {
 			++session->hists.stats.nr_unknown_id;
 			return -1;
 		}
-- 
1.7.5.4

Re: [PATCH 1/2] perf-tool: Don't process samples with no valid machine object

[Arnaldo Carvalho de Melo]

Em Thu, Feb 09, 2012 at 05:07:38PM +0100, Joerg Roedel escreveu:
> The perf sample processing code relies on a valid machine
> object. Make sure that this path is only entered when such a
> object exists.
> 
> Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
> ---
>  tools/perf/builtin-top.c  |    2 +-
>  tools/perf/util/session.c |    2 +-
>  2 files changed, 2 insertions(+), 2 deletions(-)
> 
> diff --git a/tools/perf/builtin-top.c b/tools/perf/builtin-top.c
> index dd162aa..7c81d1d 100644
> --- a/tools/perf/builtin-top.c
> +++ b/tools/perf/builtin-top.c
> @@ -805,7 +805,7 @@ static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
>  		}
>  
>  
> -		if (event->header.type == PERF_RECORD_SAMPLE) {
> +		if (event->header.type == PERF_RECORD_SAMPLE && machine != NULL) {

Shouldn't we warn the user, even if just once, on the status (last line
on the screen) line?

>  			perf_event__process_sample(&top->tool, event, evsel,
>  						   &sample, machine);
>  		} else if (event->header.type < PERF_RECORD_MAX) {
> diff --git a/tools/perf/util/session.c b/tools/perf/util/session.c
> index b5ca2558..c9593c7 100644
> --- a/tools/perf/util/session.c
> +++ b/tools/perf/util/session.c
> @@ -792,7 +792,7 @@ static int perf_session_deliver_event(struct perf_session *session,
>  	switch (event->header.type) {
>  	case PERF_RECORD_SAMPLE:
>  		dump_sample(session, event, sample);
> -		if (evsel == NULL) {
> +		if (evsel == NULL || machine == NULL) {
>  			++session->hists.stats.nr_unknown_id;
>  			return -1;
>  		}
> -- 
> 1.7.5.4
> 

Re: [PATCH 1/2] perf-tool: Don't process samples with no valid machine object

[Joerg Roedel]

On Thu, Feb 09, 2012 at 02:34:41PM -0200, Arnaldo Carvalho de Melo wrote:
> Em Thu, Feb 09, 2012 at 05:07:38PM +0100, Joerg Roedel escreveu:
> > -		if (event->header.type == PERF_RECORD_SAMPLE) {
> > +		if (event->header.type == PERF_RECORD_SAMPLE && machine != NULL) {
> 
> Shouldn't we warn the user, even if just once, on the status (last line
> on the screen) line?

Probably yes, what would be a good message? I guess something like

	"no machine object for sample"

is not helpful to the user. Maybe something like

	"Unresolvable sample(s) recorded"?

Or something completly different?


	Joerg

-- 
AMD Operating System Research Center

Advanced Micro Devices GmbH Einsteinring 24 85609 Dornach
General Managers: Alberto Bozzo
Registration: Dornach, Landkr. Muenchen; Registerger. Muenchen, HRB Nr. 43632

Re: [PATCH 1/2] perf-tool: Don't process samples with no valid machine object

[Arnaldo Carvalho de Melo]

Em Thu, Feb 09, 2012 at 06:13:34PM +0100, Joerg Roedel escreveu:
> On Thu, Feb 09, 2012 at 02:34:41PM -0200, Arnaldo Carvalho de Melo wrote:
> > Em Thu, Feb 09, 2012 at 05:07:38PM +0100, Joerg Roedel escreveu:
> > > -		if (event->header.type == PERF_RECORD_SAMPLE) {
> > > +		if (event->header.type == PERF_RECORD_SAMPLE && machine != NULL) {
> > 
> > Shouldn't we warn the user, even if just once, on the status (last line
> > on the screen) line?
> 
> Probably yes, what would be a good message? I guess something like
> 
> 	"no machine object for sample"
> 
> is not helpful to the user. Maybe something like
> 
> 	"Unresolvable sample(s) recorded"?
> 
> Or something completly different?

This is not completely standardized or harminized across the sources,
but we have things like:

	if (verbose)
		error("Failed to resolve callchain.  Skipping\n");

---

        if (perf_event__preprocess_sample(event, machine, &al, sample,
                                          NULL) < 0) {
                error("problem processing %d event, skipping it.\n",
                        event->header.type);
                return;
        }

---

       if (!ip_callchain__valid(sample->callchain, event)) {
                pr_debug("call-chain problem with event, skipping it.\n");
                ++session->hists.stats.nr_invalid_chains;

---


	What has been done more recently is like the above, account the
number of different problems and then, at the end do like:

        if (session->hists.stats.nr_unknown_id != 0) {
                ui__warning("%u samples with id not present in the header\n",
                            session->hists.stats.nr_unknown_id);
        }


	But this is for perf report or other tools that process all
events and then provide some post processed results, just before
presenting these results.

	I guess we could do that too for 'perf top' like tools, that
continuously show results while collecting more. Perhaps call
ui__warning() if some threshold happens, i.e. if way too many errors of
some kind are happening and then at the end provide a summary of errors
found.

- Arnaldo

[PATCH 2/2] perf-tool: Change perf_guest default back to false

[Joerg Roedel]

Setting perf_guest to true by default makes no sense because
the perf subcommands can not setup guest symbol information
and thus not process and guest samples. The only exception
is perf-kvm which changes the perf_guest value on its own.
So change the default for perf_guest back to false.

Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
---
 tools/perf/util/util.c |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)

diff --git a/tools/perf/util/util.c b/tools/perf/util/util.c
index 8131410..fb25d13 100644
--- a/tools/perf/util/util.c
+++ b/tools/perf/util/util.c
@@ -6,7 +6,7 @@
  * XXX We need to find a better place for these things...
  */
 bool perf_host  = true;
-bool perf_guest = true;
+bool perf_guest = false;
 
 void event_attr_init(struct perf_event_attr *attr)
 {
-- 
1.7.5.4

[PATCH 0/2 v2] Fix perf-tool crashes when kvm guest is running

[Joerg Roedel]

Hi,

here are the fixes I promised. These patches should fix the crashes
David and Jason have seen by making sure no sample is processed when
ther is no valid machine object for it. The second patch also changes
the perf_guest default back to false.

v1->v2:

* Introduced counter for unprocessable samples
* Added ui warnings when unprocessable samples are found

Regards,

        Joerg

[PATCH 2/2] perf-tool: Change perf_guest default back to false

[Joerg Roedel]

Setting perf_guest to true by default makes no sense because
the perf subcommands can not setup guest symbol information
and thus not process and guest samples. The only exception
is perf-kvm which changes the perf_guest value on its own.
So change the default for perf_guest back to false.

Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
---
 tools/perf/util/util.c |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)

diff --git a/tools/perf/util/util.c b/tools/perf/util/util.c
index 8131410..fb25d13 100644
--- a/tools/perf/util/util.c
+++ b/tools/perf/util/util.c
@@ -6,7 +6,7 @@
  * XXX We need to find a better place for these things...
  */
 bool perf_host  = true;
-bool perf_guest = true;
+bool perf_guest = false;
 
 void event_attr_init(struct perf_event_attr *attr)
 {
-- 
1.7.5.4