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

From: Jin, Yao
Date: Sun Jul 05 2020 - 20:55:50 EST


Hi Ian,

On 7/6/2020 8:47 AM, Ian Rogers wrote:
On Fri, Jul 3, 2020 at 5:31 PM Jin, Yao <yao.jin@xxxxxxxxxxxxxxx> wrote:

Hi Jiri,

On 7/3/2020 7:00 PM, Jiri Olsa wrote:
On Fri, Jul 03, 2020 at 08:42:15AM +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'


Sorry for the breakage caused by modifying the dummy event. Could we
add a test to cover the issue? Perhaps in tools/perf/tests/shell/.
Trying to reproduce with a register on my skylakex on a 5.6.14 kernel
with:

$ perf record -e cycles:p -IAX -a -- sleep 1

succeeds.

Thanks,
Ian


-IAX should be no problem. The issue only occurs on the platform with extended regs supports, such as ICL. So I don't know if it's suitable to add it to perf test suite.

Thanks
Jin Yao

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 pieces.

/* 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 be not set with
PERF_PMU_CAP_EXTENDED_REGS. But unfortunately in current code, 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.

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) ]

LGTM, Adrian (cc-ed) just added another check to the same place,
but it looks like both of them should be there:

https://lore.kernel.org/lkml/20200630133935.11150-2-adrian.hunter@xxxxxxxxx/

jirka


Thanks Jiri! Yes, it looks like both of checks should be added here.

So do I post v2 (just rebase) once Adrian's patch gets merged?

Thanks
Jin Yao


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

diff --git a/tools/perf/util/evsel.c b/tools/perf/util/evsel.c
index 96e5171dce41..df3315543e86 100644
--- a/tools/perf/util/evsel.c
+++ b/tools/perf/util/evsel.c
@@ -1020,12 +1020,12 @@ 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) {
+ if (opts->sample_intr_regs && !is_dummy_event(evsel)) {
attr->sample_regs_intr = opts->sample_intr_regs;
evsel__set_sample_bit(evsel, REGS_INTR);
}

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