RFT: Do we detect WA_TAIL?

Commit Message

Chris Wilson Aug. 15, 2018, 11:13 a.m. UTC

---
 drivers/gpu/drm/i915/intel_lrc.c | 8 +++++---
 1 file changed, 5 insertions(+), 3 deletions(-)

Comments

Chris Wilson Aug. 15, 2018, 12:23 p.m. UTC | #1

Quoting Chris Wilson (2018-08-15 12:13:51)

The answer is no. So either our wait-for-ack is the ultimate panacea, or
we just don't have the right pattern to trigger the bug.
-Chris

Chris Wilson Aug. 15, 2018, 12:24 p.m. UTC | #2

Quoting Chris Wilson (2018-08-15 13:23:12)
> Quoting Chris Wilson (2018-08-15 12:13:51)
> 
> The answer is no. So either our wait-for-ack is the ultimate panacea, or
> we just don't have the right pattern to trigger the bug.

Fwiw, the small batch sizes for gem_concurrent_blit were good at hitting
it on bdw/skl.
-Chris

Patch

diff --git a/drivers/gpu/drm/i915/intel_lrc.c b/drivers/gpu/drm/i915/intel_lrc.c
index 3f90c74038ef..fff2fbb6bac5 100644
--- a/drivers/gpu/drm/i915/intel_lrc.c
+++ b/drivers/gpu/drm/i915/intel_lrc.c
@@ -161,7 +161,7 @@ 
 
 /* Typical size of the average request (2 pipecontrols and a MI_BB) */
 #define EXECLISTS_REQUEST_SIZE 64 /* bytes */
-#define WA_TAIL_DWORDS 2
+#define WA_TAIL_DWORDS 0
 #define WA_TAIL_BYTES (sizeof(u32) * WA_TAIL_DWORDS)
 
 static int execlists_context_deferred_alloc(struct i915_gem_context *ctx,
@@ -2195,8 +2195,10 @@  static int gen8_emit_flush_render(struct i915_request *request,
 static void gen8_emit_wa_tail(struct i915_request *request, u32 *cs)
 {
 	/* Ensure there's always at least one preemption point per-request. */
-	*cs++ = MI_ARB_CHECK;
-	*cs++ = MI_NOOP;
+	if (WA_TAIL_DWORDS) {
+		*cs++ = MI_ARB_CHECK;
+		*cs++ = MI_NOOP;
+	}
 	request->wa_tail = intel_ring_offset(request, cs);
 }