Quoting Chuansheng Liu (2019-11-20 06:29:12) > Reference: > https://bugs.freedesktop.org/show_bug.cgi?id=112126 > > The issue we hit is the GPU keeps very high load after running > the subtest min-max-config-loaded. > > Some background of the issue: > Currently the rps is not fully enabled yet on TGL, and running > the subtest min-max-config-loaded will hit below assertion: > == > (i915_pm_rps:1261) CRITICAL: Test assertion failure function loaded_check, file ../tests/i915/i915_pm_rps.c:505: > (i915_pm_rps:1261) CRITICAL: Failed assertion: freqs[MAX] <= freqs[CUR] > (i915_pm_rps:1261) CRITICAL: Last errno: 2, No such file or directory > == > > with igt stress test, we find the GT keeps busy after running > this subtest, it is due to the igt_spin_end() is not called > randomly. > > The root cause analysis is: > When the main process i915_pm_rps for running the subtest > min-max-config-loaded hits the assertion, the main process will > try to send signal SIGTERM to the child process loader_helper > which is created by main process for starting GT load, then the > main process itself will exit. > > The SIGTERM handler for loader_helper is the default one, which > will cause the loader_helper exits directly. That is unsafe, we > always expect the igt_spin_end() is called before loader_helper > process exits, which is used to stop the load of GT. > > Furthermore, in normal scenario, before main process exits, > it will send SIGUSR1 to child process for stopping GT loading > in safe way. > > So here we install the proper handler for signal SIGTERM in the > similar way. Without this patch, the GT may keep busy after > running this subtest. Enabling rps should be tracked on the > other side. > > Signed-off-by: Chuansheng Liu <chuansheng.liu@xxxxxxxxx> > --- > tests/i915/i915_pm_rps.c | 1 + > 1 file changed, 1 insertion(+) > > diff --git a/tests/i915/i915_pm_rps.c b/tests/i915/i915_pm_rps.c > index ef627c0b..8c71c1a1 100644 > --- a/tests/i915/i915_pm_rps.c > +++ b/tests/i915/i915_pm_rps.c > @@ -252,6 +252,7 @@ static void load_helper_run(enum load load) > > signal(SIGUSR1, load_helper_signal_handler); > signal(SIGUSR2, load_helper_signal_handler); > + signal(SIGTERM, load_helper_signal_handler); I don't see any behaviour changes to igt to cause it to send SIGTERM on exit_subtest. But you might as well just s/SIGUSR2/SIGTERM/ for clearer and common intentions. -Chris _______________________________________________ Intel-gfx mailing list Intel-gfx@xxxxxxxxxxxxxxxxxxxxx https://lists.freedesktop.org/mailman/listinfo/intel-gfx
