[PATCH 11/30] drm/i915: Flush the execlist ports if idle

[Chris Wilson]

When doing a GPU reset, the CSB register will be trashed and we will
lose any context-switch notifications that happened since the tasklet
was disabled. If we find that all requests on this engine were
completed, we want to make sure that the ELSP tracker is similarly empty
so that we do not feed back in the completed requests upon recovering
from the reset.

Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
Cc: Mika Kuoppala <mika.kuoppala at linux.intel.com>
---
 drivers/gpu/drm/i915/intel_lrc.c | 36 ++++++++++++++++++++++++++----------
 1 file changed, 26 insertions(+), 10 deletions(-)

diff --git a/drivers/gpu/drm/i915/intel_lrc.c b/drivers/gpu/drm/i915/intel_lrc.c
index 7c3dce27e504..25787bcaa103 100644
--- a/drivers/gpu/drm/i915/intel_lrc.c
+++ b/drivers/gpu/drm/i915/intel_lrc.c
@@ -1327,6 +1327,31 @@ static void reset_common_ring(struct intel_engine_cs *engine,
 {
 	struct execlist_port *port = engine->execlist_port;
 	struct intel_context *ce;
+	unsigned int n;
+
+	/*
+	 * Catch up with any missed context-switch interrupts.
+	 *
+	 * Ideally we would just read the remaining CSB entries now that we
+	 * know the gpu is idle. However, the CSB registers are sometimes^W
+	 * often trashed across a GPU reset! Instead we have to rely on
+	 * guessing the missed context-switch events by looking at what
+	 * requests were completed.
+	 */
+	if (!request) {
+		for (n = 0; n < ARRAY_SIZE(engine->execlist_port); n++)
+			i915_gem_request_put(port_request(&port[n]));
+		memset(engine->execlist_port, 0, sizeof(engine->execlist_port));
+		return;
+	}
+
+	if (request->ctx != port_request(port)->ctx) {
+		i915_gem_request_put(port_request(port));
+		port[0] = port[1];
+		memset(&port[1], 0, sizeof(port[1]));
+	}
+
+	GEM_BUG_ON(request->ctx != port_request(port)->ctx);
 
 	/* If the request was innocent, we leave the request in the ELSP
 	 * and will try to replay it on restarting. The context image may
@@ -1338,7 +1363,7 @@ static void reset_common_ring(struct intel_engine_cs *engine,
 	 * and have to at least restore the RING register in the context
 	 * image back to the expected values to skip over the guilty request.
 	 */
-	if (!request || request->fence.error != -EIO)
+	if (request->fence.error != -EIO)
 		return;
 
 	/* We want a simple context + ring to execute the breadcrumb update.
@@ -1360,15 +1385,6 @@ static void reset_common_ring(struct intel_engine_cs *engine,
 	request->ring->head = request->postfix;
 	intel_ring_update_space(request->ring);
 
-	/* Catch up with any missed context-switch interrupts */
-	if (request->ctx != port_request(port)->ctx) {
-		i915_gem_request_put(port_request(port));
-		port[0] = port[1];
-		memset(&port[1], 0, sizeof(port[1]));
-	}
-
-	GEM_BUG_ON(request->ctx != port_request(port)->ctx);
-
 	/* Reset WaIdleLiteRestore:bdw,skl as well */
 	request->tail =
 		intel_ring_wrap(request->ring,
-- 
2.13.2