drm: reference count event->completion
authorDaniel Vetter <daniel.vetter@ffwll.ch>
Wed, 21 Dec 2016 10:23:30 +0000 (11:23 +0100)
committerDaniel Vetter <daniel.vetter@ffwll.ch>
Wed, 4 Jan 2017 10:03:06 +0000 (11:03 +0100)
When writing the generic nonblocking commit code I assumed that
through clever lifetime management I can assure that the completion
(stored in drm_crtc_commit) only gets freed after it is completed. And
that worked.

I also wanted to make nonblocking helpers resilient against driver
bugs, by having timeouts everywhere. And that worked too.

Unfortunately taking boths things together results in oopses :( Well,
at least sometimes: What seems to happen is that the drm event hangs
around forever stuck in limbo land. The nonblocking helpers eventually
time out, move on and release it. Now the bug I tested all this
against is drivers that just entirely fail to deliver the vblank
events like they should, and in those cases the event is simply
leaked. But what seems to happen, at least sometimes, on i915 is that
the event is set up correctly, but somohow the vblank fails to fire in
time. Which means the event isn't leaked, it's still there waiting for
eventually a vblank to fire. That tends to happen when re-enabling the
pipe, and then the trap springs and the kernel oopses.

The correct fix here is simply to refcount the crtc commit to make
sure that the event sticks around even for drivers which only
sometimes fail to deliver vblanks for some arbitrary reasons. Since
crtc commits are already refcounted that's easy to do.

References: https://bugs.freedesktop.org/show_bug.cgi?id=96781
Cc: Jim Rees <rees@umich.edu>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
Cc: Jani Nikula <jani.nikula@linux.intel.com>
Reviewed-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20161221102331.31033-1-daniel.vetter@ffwll.ch
drivers/gpu/drm/drm_atomic_helper.c
drivers/gpu/drm/drm_fops.c
include/drm/drmP.h

index 5e52244600420df91a5e6e60fcc94713769f953f..9460f3a4c1c2ba2348927d10946a7920956a0e53 100644 (file)
@@ -1355,6 +1355,15 @@ static int stall_checks(struct drm_crtc *crtc, bool nonblock)
        return ret < 0 ? ret : 0;
 }
 
+void release_crtc_commit(struct completion *completion)
+{
+       struct drm_crtc_commit *commit = container_of(completion,
+                                                     typeof(*commit),
+                                                     flip_done);
+
+       drm_crtc_commit_put(commit);
+}
+
 /**
  * drm_atomic_helper_setup_commit - setup possibly nonblocking commit
  * @state: new modeset state to be committed
@@ -1447,6 +1456,8 @@ int drm_atomic_helper_setup_commit(struct drm_atomic_state *state,
                }
 
                crtc_state->event->base.completion = &commit->flip_done;
+               crtc_state->event->base.completion_release = release_crtc_commit;
+               drm_crtc_commit_get(commit);
        }
 
        return 0;
index 48e106557c92c99a563a66792310b1b3506c9050..e22645375e6074ed572ac5b9144da209289140b9 100644 (file)
@@ -689,8 +689,8 @@ void drm_send_event_locked(struct drm_device *dev, struct drm_pending_event *e)
        assert_spin_locked(&dev->event_lock);
 
        if (e->completion) {
-               /* ->completion might disappear as soon as it signalled. */
                complete_all(e->completion);
+               e->completion_release(e->completion);
                e->completion = NULL;
        }
 
index 8671ddd256fc5a1f276e5505ba35ae445364c238..ffa8b6d6d3679c9cad742b8728ad6be9fa5f3168 100644 (file)
@@ -360,6 +360,7 @@ struct drm_ioctl_desc {
 /* Event queued up for userspace to read */
 struct drm_pending_event {
        struct completion *completion;
+       void (*completion_release)(struct completion *completion);
        struct drm_event *event;
        struct dma_fence *fence;
        struct list_head link;