perf record: fix fallback to cpu-clock on ppc

PATCH Applied David Ahern 2 participants 6 messages May 8, 2012 – May 9, 2012 perf events

lore

Share & tools

lore

David Ahern

author

2012-05-08 04:40 UTC #0 lore raw

PATCH 1/1 perf record: fix fallback to cpu-clock on ppc

perf-record on PPC is not falling back to cpu-clock:

Old kernel, cannot exclude guest or host samples.

Error: sys_perf_event_open() syscall returned with 6 (No such device or address). /bin/dmesg may provide additional information.

Fatal: No CONFIG_PERF_EVENTS=y kernel support configured?

The problem is that until 2.6.37 (behavior changed with commit b0a873e) perf on PPC returns ENXIO when hw_perf_event_init() fails. With this patch we get the expected behavior:

Old kernel, cannot exclude guest or host samples.
The cycles event is not supported, trying to fall back to cpu-clock-ticks [ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.150 MB /tmp/perf.data (~6569 samples) ]

Signed-off-by David Ahern <[email protected]>

---
 tools/perf/builtin-record.c |    3 ++-
 1 files changed, 2 insertions(+), 1 deletions(-)

tools/perf/builtin-record.c +2 -1

index 42e2414..080c708 100644
@@ -244,7 +244,8 @@ try_again:
 			 * based cpu-clock-tick sw counter, which
 			 * is always available even if no PMU support:
 			 */
-			if (err == ENOENT && attr->type == PERF_TYPE_HARDWARE
+			if ((err == ENOENT || err == ENXIO)
+					&& attr->type == PERF_TYPE_HARDWARE
 					&& attr->config == PERF_COUNT_HW_CPU_CYCLES) {
 
 				if (verbose)

Signature · 1 line

1.7.5.4

Arnaldo Carvalho de Melo

2012-05-08 13:12 UTC #1 lore raw

Em Mon, May 07, 2012 at 10:40:04PM -0600, David Ahern escreveu:

Quoted 8 lines

perf-record on PPC is not falling back to cpu-clock:
Old kernel, cannot exclude guest or host samples.
Error: sys_perf_event_open() syscall returned with 6 (No such device or address). /bin/dmesg may provide additional information.
Fatal: No CONFIG_PERF_EVENTS=y kernel support configured?

Quoted 8 lines

The problem is that until 2.6.37 (behavior changed with commit b0a873e) perf on PPC returns ENXIO when hw_perf_event_init() fails. With this patch we get the expected behavior:
Old kernel, cannot exclude guest or host samples.
The cycles event is not supported, trying to fall back to cpu-clock-ticks [ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.150 MB /tmp/perf.data (~6569 samples) ]

Would be good to have this documented on the spot, i.e. just before that if.

Also, have you checked that 'perf stat' and 'perf top' works as well? We really need to move all this logic to a central place, namely perf_evlist__open().

For now checking if stat and top ara ok should be enough.

- Arnaldo

Quoted 21 lines

Signed-off-by David Ahern <[email protected]>

---
 tools/perf/builtin-record.c |    3 ++-
 1 files changed, 2 insertions(+), 1 deletions(-)

tools/perf/builtin-record.c +2 -1

index 42e2414..080c708 100644
@@ -244,7 +244,8 @@ try_again:
 			 * based cpu-clock-tick sw counter, which
 			 * is always available even if no PMU support:
 			 */
-			if (err == ENOENT && attr->type == PERF_TYPE_HARDWARE
+			if ((err == ENOENT || err == ENXIO)
+					&& attr->type == PERF_TYPE_HARDWARE
 					&& attr->config == PERF_COUNT_HW_CPU_CYCLES) {
 
 				if (verbose)

Signature · 1 line

1.7.5.4

David Ahern

author

2012-05-08 14:10 UTC #2 lore raw

Quoted 18 lines · Arnaldo Carvalho de Melo

Em Mon, May 07, 2012 at 10:40:04PM -0600, David Ahern escreveu:
Quoted 8 lines
perf-record on PPC is not falling back to cpu-clock:
Old kernel, cannot exclude guest or host samples.
Error: sys_perf_event_open() syscall returned with 6 (No such device or address). /bin/dmesg may provide additional information.
Fatal: No CONFIG_PERF_EVENTS=y kernel support configured?
Quoted 8 lines
The problem is that until 2.6.37 (behavior changed with commit b0a873e) perf on PPC returns ENXIO when hw_perf_event_init() fails. With this patch we get the expected behavior:
Old kernel, cannot exclude guest or host samples.
The cycles event is not supported, trying to fall back to cpu-clock-ticks [ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.150 MB /tmp/perf.data (~6569 samples) ]

Quoted 8 lines

Would be good to have this documented on the spot, i.e. just before that if.
Also, have you checked that 'perf stat' and 'perf top' works as well? We really need to move all this logic to a central place, namely perf_evlist__open().
For now checking if stat and top ara ok should be enough.

perf-top works because it drops back to cpu-clock for any perf_event_open failure; perf-stat does not. Will fix it and add a comment about PPC and ENXIO errno to both.

David

Arnaldo Carvalho de Melo

Applied

2012-05-09 14:56 UTC #3 lore raw

In reply to David Ahern

Em Mon, May 07, 2012 at 10:40:04PM -0600, David Ahern escreveu:

Quoted 8 lines

perf-record on PPC is not falling back to cpu-clock:
Old kernel, cannot exclude guest or host samples.
Error: sys_perf_event_open() syscall returned with 6 (No such device or address). /bin/dmesg may provide additional information.
Fatal: No CONFIG_PERF_EVENTS=y kernel support configured?

Quoted 8 lines

The problem is that until 2.6.37 (behavior changed with commit b0a873e) perf on PPC returns ENXIO when hw_perf_event_init() fails. With this patch we get the expected behavior:
Old kernel, cannot exclude guest or host samples.
The cycles event is not supported, trying to fall back to cpu-clock-ticks [ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.150 MB /tmp/perf.data (~6569 samples) ]

Thanks, applied to my perf/core branch.

Can you please provide one for perf/urgent?

- Arnaldo

David Ahern

author

2012-05-09 16:22 UTC #4 lore raw

Quoted 13 lines · Arnaldo Carvalho de Melo

Em Mon, May 07, 2012 at 10:40:04PM -0600, David Ahern escreveu:
Quoted 12 lines
perf-record on PPC is not falling back to cpu-clock:
Old kernel, cannot exclude guest or host samples.
Error: sys_perf_event_open() syscall returned with 6 (No such device or address). /bin/dmesg may provide additional information.
Fatal: No CONFIG_PERF_EVENTS=y kernel support configured?
The problem is that until 2.6.37 (behavior changed with commit b0a873e) perf on PPC returns ENXIO when hw_perf_event_init() fails. With this patch we get the expected behavior:

Quoted 8 lines

Quoted 4 lines
Old kernel, cannot exclude guest or host samples.
The cycles event is not supported, trying to fall back to cpu-clock-ticks [ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.150 MB /tmp/perf.data (~6569 samples) ]
Thanks, applied to my perf/core branch.
Can you please provide one for perf/urgent?

Not needed. Problem on perf/core was introduced by 5a7ed29c7572d00a75e8c4529e30c5ac2ef82271

The perf-stat version is needed and applies cleanly (though with an offset) to perf/urgent.

David

Arnaldo Carvalho de Melo

2012-05-09 17:24 UTC #5 lore raw

Em Wed, May 09, 2012 at 10:22:45AM -0600, David Ahern escreveu:

Quoted 6 lines · Arnaldo Carvalho de Melo

Quoted 4 lines
Em Mon, May 07, 2012 at 10:40:04PM -0600, David Ahern escreveu:
Quoted 1 line
perf-record on PPC is not falling back to cpu-clock:
Thanks, applied to my perf/core branch.

Quoted 7 lines

Quoted 1 line
Can you please provide one for perf/urgent?
Not needed. Problem on perf/core was introduced by 5a7ed29c7572d00a75e8c4529e30c5ac2ef82271
The perf-stat version is needed and applies cleanly (though with an offset) to perf/urgent.

Ok, git cherry-pick worked on it, pushing perf/urgent out with it.

- Arnaldo