[PATCH 1/2] drm/rockchip: vop: Do check if an update is pending during disable

[Tomeu Vizoso]

On 23 May 2016 at 08:32, Mark yao <mark.yao at rock-chips.com> wrote:
> On 2016年05月05日 17:34, Tomeu Vizoso wrote:
>>
>> On 20 April 2016 at 16:23, Tomeu Vizoso <tomeu.vizoso at collabora.com>
>> wrote:
>>>
>>> On 11 April 2016 at 03:15, Mark yao <mark.yao at rock-chips.com> wrote:
>>>>
>>>> On 2016年04月08日 18:54, Tomeu Vizoso wrote:
>>>>>
>>>>> On 8 April 2016 at 03:07, Mark yao <mark.yao at rock-chips.com> wrote:
>>>>>>
>>>>>> On 2016年04月06日 18:14, Tomeu Vizoso wrote:
>>>>>>
>>>>>> When a plane is being disabled but it's still enabled, do check if the
>>>>>> previous update has been completed by reading yrgb_mst back.
>>>>>>
>>>>>> Otherwise, pending pageflips would remain pending after a CRTC is
>>>>>> disabled.
>>>>>>
>>>>>> Signed-off-by: Tomeu Vizoso <tomeu.vizoso at collabora.com>
>>>>>> ---
>>>>>>    drivers/gpu/drm/rockchip/rockchip_drm_vop.c | 5 +++--
>>>>>>    1 file changed, 3 insertions(+), 2 deletions(-)
>>>>>>
>>>>>> diff --git a/drivers/gpu/drm/rockchip/rockchip_drm_vop.c
>>>>>> b/drivers/gpu/drm/rockchip/rockchip_drm_vop.c
>>>>>> index a9b1e8b5ac85..f46b1fd1887b 100644
>>>>>> --- a/drivers/gpu/drm/rockchip/rockchip_drm_vop.c
>>>>>> +++ b/drivers/gpu/drm/rockchip/rockchip_drm_vop.c
>>>>>> @@ -1064,8 +1064,9 @@ static bool vop_win_pending_is_complete(struct
>>>>>> vop_win
>>>>>> *vop_win)
>>>>>>     struct vop_plane_state *state = to_vop_plane_state(plane->state);
>>>>>>     dma_addr_t yrgb_mst;
>>>>>>
>>>>>> - if (!state->enable)
>>>>>> - return VOP_WIN_GET(vop_win->vop, vop_win->data, enable) == 0;
>>>>>> + if (!state->enable &&
>>>>>> +    VOP_WIN_GET(vop_win->vop, vop_win->data, enable) == 0)
>>>>>> + return true;
>>>>>>
>>>>>>
>>>>>> It is wrong, the patch would cause a bug.
>>>>>>
>>>>>> when state->enable is 0, check yrgb_mst == state->yrgb_mst always be
>>>>>> true,
>>>>>> because state->yrgb_mst not update on plane disabled funtion, that
>>>>>> would
>>>>>> cause iommu crash.
>>>>>
>>>>> Sorry, but I don't understand where's the bug and what could cause
>>>>> that crash. What the existing code was doing is saying that a pageflip
>>>>> event is still pending if we have told the plane to disable but for
>>>>> some reason it hasn't yet.
>>>>>
>>>>> With this modification, if we read back that it's already disabled, we
>>>>> return true as before. But if we read back that it isn't disabled yet,
>>>>> then we still check the fb pointers and compare them.
>>>>>
>>>>> The iommu mapping is removed when the _CRTC_ is disabled, and what
>>>>> this series does is to wait for the pending pageflip to finish before
>>>>> conitnuing with CRTC disablement.
>>>>
>>>>
>>>> the iommu mapping will unmap after plane disabled, we need sure that the
>>>> plane really disabled before unmap, if not, the unmap may call before
>>>> plane
>>>> really disable, vop may access unmap address, then would get iommu page
>>>> fault.
>>>
>>> Sorry, but I still don't see the error condition that you are
>>> describing. AFAICS, the unmap will happen after the last pageflip has
>>> completed.
>>>
>>>>>> About pending pageflips would remain pending, can you  describe more
>>>>>> info
>>>>>> about it? I think those pending pageflips should be ignore when CRTC
>>>>>> is
>>>>>> disabled.
>>>>>
>>>>> Well, right now in rockchip-drm pending pageflips won't be ignored
>>>>> when a CRTC is disabled, but will be delivered when it's re-enabled.
>>>>>
>>>>> If they would be to be ignored (understanding that as dropped), that
>>>>> would require modifications to clients so they keep track of which fbs
>>>>> were used in a particular crtc and destroy them when the crtc is
>>>>> disabled, but that would be incorrect when using the i915 DRM driver
>>>>> (I also assume others do the same). Given that the pageflip ioctl
>>>>> isn't driver-specific, I think there cannot be such a difference in
>>>>> behavior between drivers.
>>>>>
>>>>> With the current behavior (pending pageflip events being delayed until
>>>>> the CRTC is enabled again), compositors and other clients will be
>>>>> holding on to the fb in the pending pageflip until an arbitrary point
>>>>> in the future that may not ever come. To me that sounds like a serious
>>>>> modification of the assumptions on fb lifecycle that might not be
>>>>> warranted.
>>>>>
>>>>> So in summary, even if I haven't found any explicit documentation on
>>>>> this, I think the ABI is that any pending pageflips are to be
>>>>> delivered when that CRTC is being disabled and not later.
>>>>
>>>>
>>>> on drivers/gpu/drm/rockchip/rockchip_drm_fb.c
>>>>
>>>> drm_atomic_helper_commit_planes(dev, state, true);
>>>> rockchip_atomic_wait_for_complete(dev, state);
>>>>
>>>> We set active_only = true, I think planes can only update when crtc is
>>>> active. and rockchip_atomic_wait_for_complete only wait when crtc is
>>>> active.
>>>
>>> That's fine, but if a pageflip is pending when the client requests the
>>> CRTC to be disabled, we need to wait for the event to be emitted
>>> before we actually disable the HW.
>>
>> Hi Mark,
>>
>> could you tell me if you agree that there's a bug that needs to be
>> solved, and if so, what do you think we should do about it?
>
> Hi Tomeu
>
> Sorry for reply late.
> I don't agree the changes:
>
> - if (!state->enable)
> - return VOP_WIN_GET(vop_win->vop, vop_win->data, enable) == 0;
> + if (!state->enable &&
> +    VOP_WIN_GET(vop_win->vop, vop_win->data, enable) == 0)
> + return true;
>
> This changes actually would lead a bug.
> when we disable a plane, the vop_win_pending_is_complete would always return
> true, That is not allowed, would cause fb free too early.

Ok, maybe I need to ask you first what the original block of code
intended to achieve. The TRM I have is very terse and I don't find any
explanation there. The battery of tests I have pass just fine without
it.

> Does this patch is needed for "[PATCH 2/2] drm/rockchip: vop: Wait for
> pending events when disabling a CRTC"

Yes, this function is currently returning false when the pageflip has
been completed but the plan has been already disabled.

If you have any different idea of how to fix this bug, please share.

Thanks,

Tomeu