[PATCH 10/13] tests/intel/xe_oa: Rewrite enable-disable test

Umesh Nerlige Ramappa umesh.nerlige.ramappa at intel.com
Tue Feb 25 22:25:10 UTC 2025


On Mon, Feb 24, 2025 at 06:54:42PM -0800, Dixit, Ashutosh wrote:
>On Mon, 24 Feb 2025 16:30:19 -0800, Dixit, Ashutosh wrote:
>>
>> On Fri, 14 Feb 2025 17:06:25 -0800, Umesh Nerlige Ramappa wrote:
>> >
>> > Keep it simple and just check if enable/disable is working correctly
>> > using mmio.
>> >
>> > Signed-off-by: Umesh Nerlige Ramappa <umesh.nerlige.ramappa at intel.com>
>> > ---
>> >  tests/intel/xe_oa.c | 129 +++++---------------------------------------
>> >  1 file changed, 13 insertions(+), 116 deletions(-)
>> >
>> > diff --git a/tests/intel/xe_oa.c b/tests/intel/xe_oa.c
>> > index ee87b7338..c6c1c2358 100644
>> > --- a/tests/intel/xe_oa.c
>> > +++ b/tests/intel/xe_oa.c
>> > @@ -2534,21 +2534,13 @@ test_non_zero_reason(const struct drm_xe_engine_class_instance *hwe, size_t oa_b
>> >  static void
>> >  test_enable_disable(const struct drm_xe_engine_class_instance *hwe)
>> >  {
>> > -	/* ~5 micro second period */
>> > -	int oa_exponent = max_oa_exponent_for_period_lte(5000);
>> > -	uint64_t oa_period = oa_exponent_to_ns(oa_exponent);
>> >	struct intel_xe_perf_metric_set *test_set = metric_set(hwe);
>> > -	uint64_t fmt = test_set->perf_oa_format;
>> >	uint64_t properties[] = {
>> >		DRM_XE_OA_PROPERTY_OA_UNIT_ID, 0,
>> > -
>> > -		/* Include OA reports in samples */
>> >		DRM_XE_OA_PROPERTY_SAMPLE_OA, true,
>> > -
>> > -		/* OA unit configuration */
>> >		DRM_XE_OA_PROPERTY_OA_METRIC_SET, test_set->perf_oa_metrics_set,
>> > -		DRM_XE_OA_PROPERTY_OA_FORMAT, __ff(fmt),
>> > -		DRM_XE_OA_PROPERTY_OA_PERIOD_EXPONENT, oa_exponent,
>> > +		DRM_XE_OA_PROPERTY_OA_FORMAT, __ff(test_set->perf_oa_format),
>> > +		DRM_XE_OA_PROPERTY_OA_PERIOD_EXPONENT, oa_exponent_default,
>> >		DRM_XE_OA_PROPERTY_OA_DISABLED, true,
>> >		DRM_XE_OA_PROPERTY_OA_ENGINE_INSTANCE, hwe->engine_instance,
>> >	};
>> > @@ -2556,123 +2548,28 @@ test_enable_disable(const struct drm_xe_engine_class_instance *hwe)
>> >		.num_properties = ARRAY_SIZE(properties) / 2,
>> >		.properties_ptr = to_user_pointer(properties),
>> >	};
>> > -	size_t report_size = get_oa_format(fmt).size;
>> > -	int buf_size = 65536 * report_size;
>> > -	uint8_t *buf = malloc(buf_size);
>> > -	int n_full_oa_reports = default_oa_buffer_size / report_size;
>> > -	uint64_t fill_duration = n_full_oa_reports * oa_period;
>> > -	uint32_t *last_periodic_report = malloc(report_size);
>> > -
>> > -	load_helper_init();
>> > -	load_helper_run(HIGH);
>> > +	u32 oacontrol;
>> >
>> >	stream_fd = __perf_open(drm_fd, &param, true /* prevent_pm */);
>> >          set_fd_flags(stream_fd, O_CLOEXEC);
>> >
>> > -	for (int i = 0; i < 5; i++) {
>> > -		int len;
>> > -		uint32_t n_periodic_reports;
>> > -		uint64_t first_timestamp = 0, last_timestamp = 0;
>> > -		u32 oa_status;
>> > -
>> > -		/* Giving enough time for an overflow might help catch whether
>> > -		 * the OA unit has been enabled even if the driver might at
>> > -		 * least avoid copying reports while disabled.
>> > -		 */
>> > -		nanosleep(&(struct timespec){ .tv_sec = 0,
>> > -					      .tv_nsec = fill_duration * 1.25 },
>> > -			  NULL);
>> > -
>> > -		while ((len = read(stream_fd, buf, buf_size)) == -1 &&
>> > -		       (errno == EINTR || errno == EIO))
>> > -			;
>> > -
>> > -		igt_assert_eq(len, -1);
>> > -		igt_assert_eq(errno, EINVAL);
>> > -
>> > -		do_ioctl(stream_fd, DRM_XE_OBSERVATION_IOCTL_ENABLE, 0);
>> > +#define OAG_OACONTROL	(0xdaf4)
>> > +#define  OAG_OACONTROL_OA_COUNTER_ENABLE	(1 << 0)
>> >
>> > -		nanosleep(&(struct timespec){ .tv_sec = 0,
>> > -					      .tv_nsec = fill_duration / 2 },
>> > -			NULL);
>> > -
>> > -		n_periodic_reports = 0;
>> > -
>> > -		/* Because of the race condition between notification of new
>> > -		 * reports and reports landing in memory, we need to rely on
>> > -		 * timestamps to figure whether we've read enough of them.
>> > -		 */
>> > -		while (((last_timestamp - first_timestamp) * oa_period) < (fill_duration / 2)) {
>> > -
>> > -			while ((len = read(stream_fd, buf, buf_size)) == -1 && errno == EINTR)
>> > -				;
>> > -			if (errno == EIO) {
>> > -				oa_status = get_stream_status(stream_fd);
>> > -				igt_debug("oa_status %#x\n", oa_status);
>> > -				igt_assert(!(oa_status & DRM_XE_OASTATUS_BUFFER_OVERFLOW));
>> > -				continue;
>> > -			}
>> > -			igt_assert_neq(len, -1);
>> > -
>> > -			for (int offset = 0; offset < len; offset += report_size) {
>> > -				uint32_t *report = (void *) (buf + offset);
>> > -
>> > -				if (first_timestamp == 0)
>> > -					first_timestamp = oa_timestamp(report, fmt);
>> > -				last_timestamp = oa_timestamp(report, fmt);
>> > -
>> > -				igt_debug(" > report ts=%"PRIx64""
>> > -					  " ts_delta_last_periodic=%s%"PRIu64""
>> > -					  " is_timer=%i ctx_id=0x%8x\n",
>> > -					  oa_timestamp(report, fmt),
>> > -					  oa_report_is_periodic(report) ? " " : "*",
>> > -					  n_periodic_reports > 0 ?  oa_timestamp_delta(report, last_periodic_report, fmt) : 0,
>> > -					  oa_report_is_periodic(report),
>> > -					  oa_report_get_ctx_id(report));
>> > -
>> > -				if (oa_report_is_periodic(report)) {
>> > -					memcpy(last_periodic_report, report, report_size);
>> > -
>> > -					/* We want to measure only the periodic reports,
>> > -					 * ctx-switch might inflate the content of the
>> > -					 * buffer and skew or measurement.
>> > -					 */
>> > -					n_periodic_reports++;
>> > -				}
>> > -			}
>> > -		}
>> > -
>> > -		do_ioctl(stream_fd, DRM_XE_OBSERVATION_IOCTL_DISABLE, 0);
>> > -
>> > -		igt_debug("first ts = %"PRIu64", last ts = %"PRIu64"\n", first_timestamp, last_timestamp);
>> > -
>> > -		igt_debug("%f < %zu < %f\n",
>> > -			  report_size * n_full_oa_reports * 0.45,
>> > -			  n_periodic_reports * report_size,
>> > -			  report_size * n_full_oa_reports * 0.55);
>> > -
>> > -		igt_assert((n_periodic_reports * report_size) >
>> > -			   (report_size * n_full_oa_reports * 0.45));
>> > -		igt_assert((n_periodic_reports * report_size) <
>> > -			   report_size * n_full_oa_reports * 0.55);
>> > +	intel_register_access_init(&mmio_data,
>> > +				   igt_device_get_pci_device(drm_fd), 0);
>> >
>> > +	oacontrol = intel_register_read(&mmio_data, OAG_OACONTROL);
>> > +	igt_assert_eq(oacontrol & OAG_OACONTROL_OA_COUNTER_ENABLE, 0);
>> >
>> > -		/* It's considered an error to read a stream while it's disabled
>> > -		 * since it would block indefinitely...
>> > -		 */
>> > -		len = read(stream_fd, buf, buf_size);
>> > +	do_ioctl(stream_fd, DRM_XE_OBSERVATION_IOCTL_ENABLE, 0);
>> >
>> > -		igt_assert_eq(len, -1);
>> > -		igt_assert_eq(errno, EINVAL);
>> > -	}
>> > +	oacontrol = intel_register_read(&mmio_data, OAG_OACONTROL);
>> >
>> > -	free(last_periodic_report);
>> > -	free(buf);
>> > +	igt_assert_eq(oacontrol & OAG_OACONTROL_OA_COUNTER_ENABLE, 1);
>> >
>> > +	intel_register_access_fini(&mmio_data);
>> >	__perf_close(stream_fd);
>> > -
>> > -	load_helper_stop();
>> > -	load_helper_fini();
>>
>> The sync tests actually change the config on an open OA stream. However
>> they do it without disable-enable. We could add disable-enable there, but
>> still those tests only check for the syncs to signal. And the syncs signal
>> after a delay after the OA config is applied, not when disable/enable
>> completes, so those tests wouldn't check for enable/disable.
>>
>> At least here we have some basic testing of enable disable. So this is:
>>
>> Reviewed-by: Ashutosh Dixit <ashutosh.dixit at intel.com>
>
>Actually, I want to question the "wisdom" of bypassing the uapi and
>verifying something by this sideband method or direct register reads,
>clever though that method is.

Since this use case is privileged, I thought I might as well check stuff 
using mmio compared to cooking up complex logic to do so. Also the logic 
needs to be simple enough to run quickly on all platforms.

>
>So for example, another way to simplify this test is:
>
>1. Disable the stream and check that read() returns -EINVAL
>2. Re-enable the stream and read() a single report
>
>So why not do just that?

Sure, I did not think of that. Looks like EINVAL is also returned in 
read if stream->sample is not set, but since our test already sets 
sample, we should be good with that. I can change this test to do that 
since anyways, this is only testing basic stuff now (see below).

>
>Fine we verify the register bit here, but the uapi is more than the
>register bit. Verifying the uapi as indicated above verifies that
>re-enabling the stream actually returns valid data, which is more than the
>register bit and of which the register bit is a subset.

There's more to actually verify in this test, but the uapi itself does 
not allow that - i.e. check that enable would have reset the OA buffer 
tail/head/status etc. Sometimes users use it to restart the capture.

For now, I have kept it to bare minimum. We can add additional coverage 
in this as we test more stuff.

Thanks,
Umesh

>
>So not sure.
>
>Thanks.
>--
>Ashutosh


More information about the igt-dev mailing list