drm/i915/ringbuffer: Reload PDs harder on byt/bcs

Baytrail takes a little more convincing that it needs to actually reload
its Page Directoy (ppGTT) before the context switch, so repeat it until
it gets the message. Once again the arbitrary values here are
empirically derived.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=107861
Testcase: igt/gem_exec_parallel/fds
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20180910130808.10809-1-chris@chris-wilson.co.uk

diff --git a/drivers/gpu/drm/i915/intel_ringbuffer.c b/drivers/gpu/drm/i915/intel_ringbuffer.c
index 472939f..d0ef50b 100644 (file)
--- a/drivers/gpu/drm/i915/intel_ringbuffer.c
+++ b/drivers/gpu/drm/i915/intel_ringbuffer.c
@@ -1677,9 +1677,26 @@ static int switch_context(struct i915_request *rq)
        GEM_BUG_ON(HAS_EXECLISTS(rq->i915));
 
        if (ppgtt) {
-               ret = load_pd_dir(rq, ppgtt);
-               if (ret)
-                       goto err;
+               int loops;
+
+               /*
+                * Baytail takes a little more convincing that it really needs
+                * to reload the PD between contexts. It is not just a little
+                * longer, as adding more stalls after the load_pd_dir (i.e.
+                * adding a long loop around flush_pd_dir) is not as effective
+                * as reloading the PD umpteen times. 32 is derived from
+                * experimentation (gem_exec_parallel/fds) and has no good
+                * explanation.
+                */
+               loops = 1;
+               if (engine->id == BCS && IS_VALLEYVIEW(engine->i915))
+                       loops = 32;
+
+               do {
+                       ret = load_pd_dir(rq, ppgtt);
+                       if (ret)
+                               goto err;
+               } while (--loops);
 
                if (intel_engine_flag(engine) & ppgtt->pd_dirty_rings) {
                        unwind_mm = intel_engine_flag(engine);