Re: [PATCH v2] perf evsel: Don't set sample_regs_intr/sample_regs_user for dummy event

[Jin, Yao]

Hi Jiri,

On 7/20/2020 5:17 PM, Jiri Olsa wrote:

On Mon, Jul 20, 2020 at 09:00:13AM +0800, Jin Yao wrote:

Since commit 0a892c1c9472 ("perf record: Add dummy event during system wide 
synthesis"),
a dummy event is added to capture mmaps.

But if we run perf-record as,

  # perf record -e cycles:p -IXMM0 -a -- sleep 1
  Error:
  dummy:HG: PMU Hardware doesn't support sampling/overflow-interrupts. Try 
'perf stat'

The issue is, if we enable the extended regs (-IXMM0), but the
pmu->capabilities is not set with PERF_PMU_CAP_EXTENDED_REGS, the kernel
will return -EOPNOTSUPP error.

See following code:

/* in kernel/events/core.c */
static int perf_try_init_event(struct pmu *pmu, struct perf_event *event)

{
         ....
         if (!(pmu->capabilities & PERF_PMU_CAP_EXTENDED_REGS) &&
             has_extended_regs(event))
                 ret = -EOPNOTSUPP;
         ....
}

For software dummy event, the PMU should not be set with
PERF_PMU_CAP_EXTENDED_REGS. But unfortunately now, the dummy
event has possibility to be set with PERF_REG_EXTENDED_MASK bit.

In evsel__config, /* tools/perf/util/evsel.c */

if (opts->sample_intr_regs) {
         attr->sample_regs_intr = opts->sample_intr_regs;
}

If we use -IXMM0, the attr>sample_regs_intr will be set with
PERF_REG_EXTENDED_MASK bit.

It doesn't make sense to set attr->sample_regs_intr for a
software dummy event.

This patch adds dummy event checking before setting
attr->sample_regs_intr and attr->sample_regs_user.

After:
   # ./perf record -e cycles:p -IXMM0 -a -- sleep 1
   [ perf record: Woken up 1 times to write data ]
   [ perf record: Captured and wrote 0.413 MB perf.data (45 samples) ]

  v2:
  ---
  Rebase to perf/core

Fixes: 0a892c1c9472 ("perf record: Add dummy event during system wide 
synthesis")
Signed-off-by: Jin Yao <yao....@linux.intel.com>
---
  tools/perf/util/evsel.c | 6 ++++--
  1 file changed, 4 insertions(+), 2 deletions(-)

diff --git a/tools/perf/util/evsel.c b/tools/perf/util/evsel.c
index 9aa51a65593d..11794d3b7879 100644
--- a/tools/perf/util/evsel.c
+++ b/tools/perf/util/evsel.c
@@ -1014,12 +1014,14 @@ void evsel__config(struct evsel *evsel, struct 
record_opts *opts,
        if (callchain && callchain->enabled && !evsel->no_aux_samples)
                evsel__config_callchain(evsel, opts, callchain);

- if (opts->sample_intr_regs && !evsel->no_aux_samples) {

+       if (opts->sample_intr_regs && !evsel->no_aux_samples &&
+           !evsel__is_dummy_event(evsel)) {

hum, I thought it'd look something like this:

   if (opts->sample_intr_regs && (!evsel->no_aux_samples || 
!evsel__is_dummy_event(evsel))

but I'm not sure how no_aux_samples flag works exactly.. so it might be
correct.. just making sure ;-)

cc-ing Adrian

jirka

no_aux_samples is set to false by default and it's only set to true by pt, 
right?

So most of the time, !evsel->no_aux_samples is always true.

if (opts->sample_intr_regs && (!evsel->no_aux_samples || 
!evsel__is_dummy_event(evsel)) {
        attr->sample_regs_intr = opts->sample_intr_regs;
        evsel__set_sample_bit(evsel, REGS_INTR);
}

So even if the evsel is dummy event, the condition check is true. :(

Or maybe I misunderstand anything?

Thanks
Jin Yao

                attr->sample_regs_intr = opts->sample_intr_regs;
                evsel__set_sample_bit(evsel, REGS_INTR);
        }

- if (opts->sample_user_regs && !evsel->no_aux_samples) {

+       if (opts->sample_user_regs && !evsel->no_aux_samples &&
+           !evsel__is_dummy_event(evsel)) {
                attr->sample_regs_user |= opts->sample_user_regs;
                evsel__set_sample_bit(evsel, REGS_USER);
        }
--
2.17.1