> From: Tomasz Duszynski [mailto:[email protected]]
> Sent: Friday, 24 October 2025 07.48
>
> Enable test to allow users to verify basic functionality. Due to
> varying
> configuration options across distributions and kernels user should
> ensure that all requirements are satisfied before starting test.
>
> Signed-off-by: Tomasz Duszynski <[email protected]>
> ---
With these changes, the test can never return TEST_FAILED, it returns
TEST_SKIPPED when something doesn't work.
IMO, the perf_allowed_quirk() check belongs in rte_pmu_init() instead.
Then, rte_pmu_init() could return a variety of error values depending on the
reason for the inability to init/start the PMU library.
Amongst the high level error values should be:
-EACCES (insufficient privileges)
-ENODEV (hardware has no PMU, or kernel is custom built without PMU)
The test should only return SKIPPED if rte_pmu_init() returns -ENODEV (no PMU
present to test); any other non-success should cause the test to return FAILED.
Also, the possible return values from rte_pmu_init() should be documented in
its function description.
> app/test/test_pmu.c | 60 ++++++++++++++++++++++++++++++++++++++++-----
> 1 file changed, 54 insertions(+), 6 deletions(-)
>
> diff --git a/app/test/test_pmu.c b/app/test/test_pmu.c
> index 10513bf9c9..7f450b3566 100644
> --- a/app/test/test_pmu.c
> +++ b/app/test/test_pmu.c
> @@ -2,10 +2,48 @@
> * Copyright(C) 2025 Marvell International Ltd.
> */
>
> +#include <stdbool.h>
> +#include <stdio.h>
> +#include <sys/types.h>
> +#include <unistd.h>
> +
> #include <rte_pmu.h>
>
> #include "test.h"
>
> +#define PERF_EVENT_PARANOID_PATH
> "/proc/sys/kernel/perf_event_paranoid"
> +
> +static bool perf_allowed_quirk(void)
> +{
> + int level, ret;
> + FILE *fp;
> +
> + fp = fopen(PERF_EVENT_PARANOID_PATH, "r");
> + if (!fp)
> + return false;
> +
> + ret = fscanf(fp, "%d", &level);
> + fclose(fp);
> + if (ret != 1)
> + return false;
> +
> + /* On vanilla Linux the default perf_event_paranoid level is 2,
> which allows non-privileged
> + * processes to access performance counters.
> + *
> + * Debian / Ubuntu and their derivatives apply patches that
> introduce
> + * additional paranoia levels:
> + *
> + * - Debian adds level 3, which restricts access to
> perf_event_open() for
> + * monitoring other processes, but still allows unprivileged
> self-monitoring.
> + * See: https://lore.kernel.org/all/1469630746-32279-1-git-
> [email protected]/
> + * - Ubuntu adds level 4 (which is also the default), completely
> disabling perf_event_open()
> + * for unprivileged users—effectively disabling self-
> monitoring.
> + *
> + * That said, check below should be sufficient to enable this
> test on most kernels.
> + */
> + return level < 4;
> +}
> +
> static int
> test_pmu_read(void)
> {
> @@ -24,8 +62,15 @@ test_pmu_read(void)
> return TEST_SKIPPED;
> }
>
> - if (rte_pmu_init() < 0)
> - return TEST_FAILED;
> + if ((getuid() != 0) && !perf_allowed_quirk()) {
> + printf("self-monitoring disabled\n");
> + return TEST_SKIPPED;
> + }
> +
> + if (rte_pmu_init() < 0) {
> + printf("PMU not initialized\n");
> + return TEST_SKIPPED;
> + }
>
> event = rte_pmu_add_event(name);
> while (tries--)
> @@ -33,7 +78,12 @@ test_pmu_read(void)
>
> rte_pmu_fini();
>
> - return val ? TEST_SUCCESS : TEST_FAILED;
> + /* rte_pmu_read() returns zero if it can't read perf counter.
> Thus series of zeros doesn't
> + * necessarily mean the counter is actually zero. It might just
> signal a problem with setup
> + * itself. So skip test to avoid testing failure and leave it to
> user to interpret this
> + * outcome.
> + */
> + return val ? TEST_SUCCESS : TEST_SKIPPED;
> }
>
> static struct unit_test_suite pmu_tests = {
> @@ -52,6 +102,4 @@ test_pmu(void)
> return unit_test_suite_runner(&pmu_tests);
> }
>
> -/* disabled because of reported failures, waiting for a fix
> - * REGISTER_FAST_TEST(pmu_autotest, true, true, test_pmu);
> - */
> +REGISTER_FAST_TEST(pmu_autotest, true, true, test_pmu);
> --
> 2.34.1
