The perf_event_open might fail due to various reasons, so blindly
reducing precise_ip level might not be the best way to deal with it.

It seems the kernel return -EOPNOTSUPP when PMU doesn't support the
given precise level.  Let's try again with the correct error code.

Signed-off-by: Namhyung Kim <namhyung@kernel.org>
---
 tools/perf/util/evsel.c | 10 ++++------
 1 file changed, 4 insertions(+), 6 deletions(-)

diff --git a/tools/perf/util/evsel.c b/tools/perf/util/evsel.c
index 8c4d70f7b2f5b880..0133c9ad3ce07a24 100644
--- a/tools/perf/util/evsel.c
+++ b/tools/perf/util/evsel.c
@@ -2565,9 +2565,6 @@ static int evsel__open_cpu(struct evsel *evsel, struct perf_cpu_map *cpus,
 	return 0;
 
 try_fallback:
-	if (evsel__precise_ip_fallback(evsel))
-		goto retry_open;
-
 	if (evsel__ignore_missing_thread(evsel, perf_cpu_map__nr(cpus),
 					 idx, threads, thread, err)) {
 		/* We just removed 1 thread, so lower the upper nthreads limit. */
@@ -2584,11 +2581,12 @@ static int evsel__open_cpu(struct evsel *evsel, struct perf_cpu_map *cpus,
 	if (err == -EMFILE && rlimit__increase_nofile(&set_rlimit))
 		goto retry_open;
 
-	if (err != -EINVAL || idx > 0 || thread > 0)
-		goto out_close;
+	if (err == -EOPNOTSUPP && evsel__precise_ip_fallback(evsel))
+		goto retry_open;
 
-	if (evsel__detect_missing_features(evsel))
+	if (err == -EINVAL && evsel__detect_missing_features(evsel))
 		goto fallback_missing_features;
+
 out_close:
 	if (err)
 		threads->err_thread = thread;
-- 
2.46.0.469.g59c65b2a67-goog

On Tue, Sep 3, 2024 at 11:41 PM Namhyung Kim <namhyung@kernel.org> wrote:
>
> The perf_event_open might fail due to various reasons, so blindly
> reducing precise_ip level might not be the best way to deal with it.
>
> It seems the kernel return -EOPNOTSUPP when PMU doesn't support the
> given precise level.  Let's try again with the correct error code.

We also have pmu's max_precise:
https://git.kernel.org/pub/scm/linux/kernel/git/perf/perf-tools-next.git/tree/tools/perf/util/pmu.h?h=perf-tools-next#n91
The reducing the precision approach was iirc taken for AMD who will
forward some precise events to IBS, but the max_precise on the cpu PMU
is 0. I think because of this, reducing the precision below
evsel->pmu->max_precise shouldn't be necessary and another fallback
may help better.

Thanks,
Ian

> Signed-off-by: Namhyung Kim <namhyung@kernel.org>
> ---
>  tools/perf/util/evsel.c | 10 ++++------
>  1 file changed, 4 insertions(+), 6 deletions(-)
>
> diff --git a/tools/perf/util/evsel.c b/tools/perf/util/evsel.c
> index 8c4d70f7b2f5b880..0133c9ad3ce07a24 100644
> --- a/tools/perf/util/evsel.c
> +++ b/tools/perf/util/evsel.c
> @@ -2565,9 +2565,6 @@ static int evsel__open_cpu(struct evsel *evsel, struct perf_cpu_map *cpus,
>         return 0;
>
>  try_fallback:
> -       if (evsel__precise_ip_fallback(evsel))
> -               goto retry_open;
> -
>         if (evsel__ignore_missing_thread(evsel, perf_cpu_map__nr(cpus),
>                                          idx, threads, thread, err)) {
>                 /* We just removed 1 thread, so lower the upper nthreads limit. */
> @@ -2584,11 +2581,12 @@ static int evsel__open_cpu(struct evsel *evsel, struct perf_cpu_map *cpus,
>         if (err == -EMFILE && rlimit__increase_nofile(&set_rlimit))
>                 goto retry_open;
>
> -       if (err != -EINVAL || idx > 0 || thread > 0)
> -               goto out_close;
> +       if (err == -EOPNOTSUPP && evsel__precise_ip_fallback(evsel))
> +               goto retry_open;
>
> -       if (evsel__detect_missing_features(evsel))
> +       if (err == -EINVAL && evsel__detect_missing_features(evsel))
>                 goto fallback_missing_features;
> +
>  out_close:
>         if (err)
>                 threads->err_thread = thread;
> --
> 2.46.0.469.g59c65b2a67-goog
>

On Wed, Sep 04, 2024 at 09:19:25AM -0700, Ian Rogers wrote:
> On Tue, Sep 3, 2024 at 11:41 PM Namhyung Kim <namhyung@kernel.org> wrote:
> >
> > The perf_event_open might fail due to various reasons, so blindly
> > reducing precise_ip level might not be the best way to deal with it.
> >
> > It seems the kernel return -EOPNOTSUPP when PMU doesn't support the
> > given precise level.  Let's try again with the correct error code.
> 
> We also have pmu's max_precise:
> https://git.kernel.org/pub/scm/linux/kernel/git/perf/perf-tools-next.git/tree/tools/perf/util/pmu.h?h=perf-tools-next#n91
> The reducing the precision approach was iirc taken for AMD who will
> forward some precise events to IBS, but the max_precise on the cpu PMU
> is 0. I think because of this, reducing the precision below
> evsel->pmu->max_precise shouldn't be necessary and another fallback
> may help better.

Internally IBS has max_precise of 2 and I think it should have that in
the sysfs.

But I found a problem with this code.  Now cycles:P would stop at 2
because after that it won't return EOPNOTSUPP.  Instead, it returns
EINVAL because of exclude_kernel and PERF_PMU_CAP_NO_EXCLUDE.

Maybe we need something like this.. :(

Thanks,
Namhyung


diff --git a/tools/perf/util/evsel.c b/tools/perf/util/evsel.c
index 0133c9ad3ce07a24..6157dc68044eb389 100644
--- a/tools/perf/util/evsel.c
+++ b/tools/perf/util/evsel.c
@@ -2587,6 +2587,13 @@ static int evsel__open_cpu(struct evsel *evsel, struct perf_cpu_map *cpus,
        if (err == -EINVAL && evsel__detect_missing_features(evsel))
                goto fallback_missing_features;
 
+       /* HACK: AMD IBS doesn't accept exclude_*, forwarding it back to core PMU */
+       if (err == -EINVAL && evsel->precise_max && evsel->core.attr.precise_ip &&
+                       evsel->core.attr.exclude_kernel) {
+               evsel->core.attr.precise_ip = 0;
+               goto fallback_missing_features;
+       }
+
 out_close:
        if (err)
                threads->err_thread = thread;