[Intel-gfx] [PATCH v2] drm/atomic: Handle vblank events in atomic ioctl correctly.

[Maarten Lankhorst]

Op 15-06-15 om 08:22 schreef Daniel Vetter:
> On Tue, Jun 02, 2015 at 09:00:46AM +0200, Maarten Lankhorst wrote:
>> All users of async updates seem to clear clear crtc_state->event
>> correctly, so move destroying vblank event to crtc_destroy_state.
>>
>> This is better than manually free'ing it in the atomic ioctl, since
>> this code seems to do it wrong.
>>
>> While we're at it handle -EDEADLK from atomic_commit correctly,
>> because the check might fail otherwise if it takes additional state.
> Please don't smash unrelated fixes together into the same patch.
>
>> Changes since v1:
>> - Fix freeing a NULL framebuffer, thanks for catching Daniel Stone.
>>
>> Signed-off-by: Maarten Lankhorst <maarten.lankhorst at linux.intel.com>
>> Reviewed-by: Daniel Stone <daniels at collabora.com>
>> ---
>>  drivers/gpu/drm/drm_atomic.c        | 45 +++++++++----------------------------
>>  drivers/gpu/drm/drm_atomic_helper.c | 15 +++++++++++++
>>  2 files changed, 26 insertions(+), 34 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/drm_atomic.c b/drivers/gpu/drm/drm_atomic.c
>> index c7e59b074e62..8add0dd8dc5d 100644
>> --- a/drivers/gpu/drm/drm_atomic.c
>> +++ b/drivers/gpu/drm/drm_atomic.c
>> @@ -1328,17 +1328,6 @@ out:
>>  	return e;
>>  }
>>  
>> -static void destroy_vblank_event(struct drm_device *dev,
>> -		struct drm_file *file_priv, struct drm_pending_vblank_event *e)
>> -{
>> -	unsigned long flags;
>> -
>> -	spin_lock_irqsave(&dev->event_lock, flags);
>> -	file_priv->event_space += sizeof e->event;
>> -	spin_unlock_irqrestore(&dev->event_lock, flags);
>> -	kfree(e);
>> -}
>> -
>>  static int atomic_set_prop(struct drm_atomic_state *state,
>>  		struct drm_mode_object *obj, struct drm_property *prop,
>>  		uint64_t prop_value)
>> @@ -1534,13 +1523,18 @@ retry:
>>  	if (arg->flags & DRM_MODE_ATOMIC_TEST_ONLY) {
>>  		ret = drm_atomic_check_only(state);
>>  		/* _check_only() does not free state, unlike _commit() */
>> -		drm_atomic_state_free(state);
>> +		if (!ret)
>> +			drm_atomic_state_free(state);
>>  	} else if (arg->flags & DRM_MODE_ATOMIC_NONBLOCK) {
>>  		ret = drm_atomic_async_commit(state);
>>  	} else {
>>  		ret = drm_atomic_commit(state);
>>  	}
>>  
>> +fail:
>> +	if (ret == -EDEADLK)
>> +		goto backoff;
>> +
>>  	/* if succeeded, fixup legacy plane crtc/fb ptrs before dropping
>>  	 * locks (ie. while it is still safe to deref plane->state).  We
>>  	 * need to do this here because the driver entry points cannot
>> @@ -1553,32 +1547,15 @@ retry:
>>  				drm_framebuffer_reference(new_fb);
>>  			plane->fb = new_fb;
>>  			plane->crtc = plane->state->crtc;
>> -		} else {
>> -			plane->old_fb = NULL;
>> -		}
>> -		if (plane->old_fb) {
>> -			drm_framebuffer_unreference(plane->old_fb);
>> -			plane->old_fb = NULL;
>> -		}
>> -	}
>> -
>> -	drm_modeset_drop_locks(&ctx);
>> -	drm_modeset_acquire_fini(&ctx);
>> -
>> -	return ret;
>> -
>> -fail:
>> -	if (ret == -EDEADLK)
>> -		goto backoff;
> We have a pre-existing bug here already of leaking plane->old_fb before
> dropping locks. I think the EDEADLK case needs to be after the cleanup
> loop to clear old_fb.
>
Ok, I'll send a fix for that first.

It's also bugged because old_fb is assigned before the plane lock is taken, resulting in possible memory corruption or leaks.