Noticed two problems here:
* We're not dropping our runtime PM refs after getting an error
* We're not backing off when pm_runtime_get() indicates that there's
already a resume in progress (-EINPROGRESS) (after which any delayed
fbcon events will get handled anyway)
So, let's fix those.
Signed-off-by: Lyude Paul <lyude@xxxxxxxxxx>
Fixes: 7fec8f5379fb ("drm/nouveau/drm/nouveau: Fix deadlock with fb_helper with async RPM requests")
Cc: stable@xxxxxxxxxxxxxxx
Cc: Ben Skeggs <bskeggs@xxxxxxxxxx>
Cc: dri-devel@xxxxxxxxxxxxxxxxxxxxx
Cc: nouveau@xxxxxxxxxxxxxxxxxxxxx
Cc: <stable@xxxxxxxxxxxxxxx> # v4.19+
---
drivers/gpu/drm/nouveau/nouveau_fbcon.c | 9 +++++----
1 file changed, 5 insertions(+), 4 deletions(-)
diff --git a/drivers/gpu/drm/nouveau/nouveau_fbcon.c b/drivers/gpu/drm/nouveau/nouveau_fbcon.c
index 4d9f3b5ae72d2..b936bf1c14dec 100644
--- a/drivers/gpu/drm/nouveau/nouveau_fbcon.c
+++ b/drivers/gpu/drm/nouveau/nouveau_fbcon.c
@@ -503,10 +503,7 @@ nouveau_fbcon_output_poll_changed(struct drm_device *dev)
ret = pm_runtime_get(dev->dev);
if (ret == 1 || ret == -EACCES) {
drm_fb_helper_hotplug_event(&fbcon->helper);
-
- pm_runtime_mark_last_busy(dev->dev);
- pm_runtime_put_autosuspend(dev->dev);
- } else if (ret == 0) {
+ } else if (ret == 0 || ret == -EINPROGRESS) {
/* If the GPU was already in the process of suspending before
* this event happened, then we can't block here as we'll
* deadlock the runtime pmops since they wait for us to
@@ -516,11 +513,15 @@ nouveau_fbcon_output_poll_changed(struct drm_device *dev)
NV_DEBUG(drm, "fbcon HPD event deferred until runtime resume\n");
fbcon->hotplug_waiting = true;
pm_runtime_put_noidle(drm->dev->dev);
+ goto out;
} else {
DRM_WARN("fbcon HPD event lost due to RPM failure: %d\n",
ret);
}
+ pm_runtime_mark_last_busy(dev->dev);
+ pm_runtime_put_autosuspend(dev->dev);
+out:
mutex_unlock(&fbcon->hotplug_lock);
}
--
2.26.2
