[PATCH] drm/i915: PIPE_CONTROL TLB invalidation fix

From: Ben Widawsky <ben@bwidawsk.net>
To: intel-gfx@lists.freedesktop.org
Cc: Ben Widawsky <ben@bwidawsk.net>
Subject: [PATCH] drm/i915: PIPE_CONTROL TLB invalidation fix
Date: Thu, 26 Jul 2012 16:48:43 -0700	[thread overview]
Message-ID: <1343346523-25255-1-git-send-email-ben@bwidawsk.net> (raw)

The IVB simulator really doesn't like a TLB invalidate with no post-sync
operation, in fact it blows up in an assertion failure. The
documentation states that we must issue the TLB invalidate with a CS
stall: "Also Requires stall bit ([20] of DW1) set." This patch doesn't
comply with the docs, but we're able to satisfy the simulator with this
very small change, and I think simulator has historically trumped docs.

Note, I don't think this belongs in stable as our TLB invalidation
should be correct since we use the global invalidation per batch. Using
TLB invalidation is itself only a requirement of HW contexts.

Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
---
 drivers/gpu/drm/i915/intel_ringbuffer.c | 10 +++++++---
 1 file changed, 7 insertions(+), 3 deletions(-)

diff --git a/drivers/gpu/drm/i915/intel_ringbuffer.c b/drivers/gpu/drm/i915/intel_ringbuffer.c
index c58f1b9..339712a 100644
--- a/drivers/gpu/drm/i915/intel_ringbuffer.c
+++ b/drivers/gpu/drm/i915/intel_ringbuffer.c
@@ -198,8 +198,12 @@ intel_emit_post_sync_nonzero_flush(struct intel_ring_buffer *ring)
 	if (ret)
 		return ret;
 
+	/* NOTE: we want the TLB invalidate for render ring flush, but it must
+	 * be sent with a non-zero post sync op. So it can be stuffed in here
+	 * for convenience.
+	 */
 	intel_ring_emit(ring, GFX_OP_PIPE_CONTROL(5));
-	intel_ring_emit(ring, PIPE_CONTROL_QW_WRITE);
+	intel_ring_emit(ring, PIPE_CONTROL_QW_WRITE | PIPE_CONTROL_TLB_INVALIDATE);
 	intel_ring_emit(ring, scratch_addr | PIPE_CONTROL_GLOBAL_GTT); /* address */
 	intel_ring_emit(ring, 0);
 	intel_ring_emit(ring, 0);
@@ -225,10 +229,10 @@ gen6_render_ring_flush(struct intel_ring_buffer *ring,
 
 	/* Just flush everything.  Experiments have shown that reducing the
 	 * number of bits based on the write domains has little performance
-	 * impact.
+	 * impact. Note to copy/pasters: the TLB invlidate we want is tucked
+	 * into intel_emit_post_sync_nonzero_flush().
 	 */
 	flags |= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH;
-	flags |= PIPE_CONTROL_TLB_INVALIDATE;
 	flags |= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE;
 	flags |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE;
 	flags |= PIPE_CONTROL_DEPTH_CACHE_FLUSH;
-- 
1.7.11.3
