[PATCH i-g-t] tests/intel/gem_watchdog: Reduced timeouts for worst case scenario

John.C.Harrison at Intel.com John.C.Harrison at Intel.com
Mon Feb 12 21:23:28 UTC 2024


From: John Harrison <John.C.Harrison at Intel.com>

The watchdog test reduces the watchdog timer from 20s to 1s and then
uses a 5s timeout waiting for the watchdog to do its stuff. This works
fine in general, but if an engine reset is required by a context that
is actually dead for real then a pre-emption timeout must be factored
in. For RCS/CCS engines, that timeout is 7.5 seconds by default. Thus,
the test timeout expires first and the test fails.

Normally, the system is not so dead when running this test as to
require an engine reset. A simple pre-emption works fine for the
spinner contexts that is uses. However, there is a hardware workaround
coming which prevents context switches when both RCS and CCS are busy.

So add an explicit override of the pre-emption timeout as well as the
watchdog timeout. That will allow the test to keep working after the
new w/a lands.

Signed-off-by: John Harrison <John.C.Harrison at Intel.com>
---
 tests/intel/gem_watchdog.c | 10 ++++++++++
 1 file changed, 10 insertions(+)

diff --git a/tests/intel/gem_watchdog.c b/tests/intel/gem_watchdog.c
index 1e4c350214c0..c9dd0deb51aa 100644
--- a/tests/intel/gem_watchdog.c
+++ b/tests/intel/gem_watchdog.c
@@ -577,6 +577,16 @@ igt_main
 
 		i915 = drm_reopen_driver(i915); /* Apply modparam. */
 		ctx = intel_ctx_create_all_physical(i915);
+
+		for_each_ctx_engine(i915, ctx, e) {
+			/*
+			 * Context termination by watchdog may require an engine reset. That only
+			 * occurs after a pre-emption attempt has expired. For RCS/CCS engines,
+			 * the pre-emption timeout is longer than this test is wanting to wait.
+			 * So reduce that timeout in addition to the watchdog timeout itself.
+			 */
+			gem_engine_property_printf(i915, e->name, "preempt_timeout_ms", "%d", 640);
+		}
 	}
 
 	igt_subtest_group {
-- 
2.43.0



More information about the igt-dev mailing list