[Intel-gfx] [PATCH v8 07/17] drm/i915: Protect workers against disappearing connectors

[Sean Paul]

From: Sean Paul <seanpaul at chromium.org>

This patch adds some protection against connectors being destroyed
before the HDCP workers are finished.

For check_work, we do a synchronous cancel after the connector is
unregistered which will ensure that it is finished before destruction.

In the case of prop_work, we can't do a synchronous wait since it needs
to take connection_mutex which could cause deadlock. Instead, we'll take
a reference on the connector when scheduling prop_work and give it up
once we're done.

Reviewed-by: Ramalingam C <ramalingam.c at intel.com>
Signed-off-by: Sean Paul <seanpaul at chromium.org>
Link: https://patchwork.freedesktop.org/patch/msgid/20191212190230.188505-8-sean@poorly.run #v2
Link: https://patchwork.freedesktop.org/patch/msgid/20200117193103.156821-8-sean@poorly.run #v3
Link: https://patchwork.freedesktop.org/patch/msgid/20200218220242.107265-8-sean@poorly.run #v4
Link: https://patchwork.freedesktop.org/patch/msgid/20200305201236.152307-8-sean@poorly.run #v5
Link: https://patchwork.freedesktop.org/patch/msgid/20200429195502.39919-8-sean@poorly.run #v6
Link: https://patchwork.freedesktop.org/patch/msgid/20200623155907.22961-8-sean@poorly.run #v7

Changes in v2:
-Added to the set
Changes in v3:
-Change the WARN_ON condition in intel_hdcp_cleanup to allow for
 initializing connectors as well
Changes in v4:
-None
Changes in v5:
-Change WARN_ON to drm_WARN_ON
Changes in v6:
-None
Changes in v7:
-None
Changes in v8:
-None
---
 drivers/gpu/drm/i915/display/intel_hdcp.c | 44 ++++++++++++++++++++---
 1 file changed, 39 insertions(+), 5 deletions(-)

diff --git a/drivers/gpu/drm/i915/display/intel_hdcp.c b/drivers/gpu/drm/i915/display/intel_hdcp.c
index ab2e2f9d0020..fe9377a6e4d5 100644
--- a/drivers/gpu/drm/i915/display/intel_hdcp.c
+++ b/drivers/gpu/drm/i915/display/intel_hdcp.c
@@ -887,8 +887,10 @@ static void intel_hdcp_update_value(struct intel_connector *connector,
 		return;
 
 	hdcp->value = value;
-	if (update_property)
+	if (update_property) {
+		drm_connector_get(&connector->base);
 		schedule_work(&hdcp->prop_work);
+	}
 }
 
 /* Implements Part 3 of the HDCP authorization procedure */
@@ -980,6 +982,8 @@ static void intel_hdcp_prop_work(struct work_struct *work)
 
 	mutex_unlock(&hdcp->mutex);
 	drm_modeset_unlock(&dev_priv->drm.mode_config.connection_mutex);
+
+	drm_connector_put(&connector->base);
 }
 
 bool is_hdcp_supported(struct drm_i915_private *dev_priv, enum port port)
@@ -1859,6 +1863,9 @@ static void intel_hdcp_check_work(struct work_struct *work)
 					       check_work);
 	struct intel_connector *connector = intel_hdcp_to_connector(hdcp);
 
+	if (drm_connector_is_unregistered(&connector->base))
+		return;
+
 	if (!intel_hdcp2_check_link(connector))
 		schedule_delayed_work(&hdcp->check_work,
 				      DRM_HDCP2_CHECK_PERIOD_MS);
@@ -2185,12 +2192,39 @@ void intel_hdcp_component_fini(struct drm_i915_private *dev_priv)
 
 void intel_hdcp_cleanup(struct intel_connector *connector)
 {
-	if (!connector->hdcp.shim)
+	struct intel_hdcp *hdcp = &connector->hdcp;
+
+	if (!hdcp->shim)
 		return;
 
-	mutex_lock(&connector->hdcp.mutex);
-	kfree(connector->hdcp.port_data.streams);
-	mutex_unlock(&connector->hdcp.mutex);
+	/*
+	 * If the connector is registered, it's possible userspace could kick
+	 * off another HDCP enable, which would re-spawn the workers.
+	 */
+	drm_WARN_ON(connector->base.dev,
+		connector->base.registration_state == DRM_CONNECTOR_REGISTERED);
+
+	/*
+	 * Now that the connector is not registered, check_work won't be run,
+	 * but cancel any outstanding instances of it
+	 */
+	cancel_delayed_work_sync(&hdcp->check_work);
+
+	/*
+	 * We don't cancel prop_work in the same way as check_work since it
+	 * requires connection_mutex which could be held while calling this
+	 * function. Instead, we rely on the connector references grabbed before
+	 * scheduling prop_work to ensure the connector is alive when prop_work
+	 * is run. So if we're in the destroy path (which is where this
+	 * function should be called), we're "guaranteed" that prop_work is not
+	 * active (tl;dr This Should Never Happen).
+	 */
+	drm_WARN_ON(connector->base.dev, work_pending(&hdcp->prop_work));
+
+	mutex_lock(&hdcp->mutex);
+	kfree(hdcp->port_data.streams);
+	hdcp->shim = NULL;
+	mutex_unlock(&hdcp->mutex);
 }
 
 void intel_hdcp_atomic_check(struct drm_connector *connector,
-- 
Sean Paul, Software Engineer, Google / Chromium OS