Re: [PATCH v2 1/5] drm/rockchip: fix fb references in async update

["Michel Dänzer" <michel@daenzer.net>]

On 2019-03-14 6:51 p.m., Helen Koike wrote:
> On 3/14/19 6:15 AM, Michel Dänzer wrote:
>> On 2019-03-13 7:08 p.m., Helen Koike wrote:
>>> On 3/13/19 6:58 AM, Michel Dänzer wrote:
>>>> On 2019-03-13 4:42 a.m., Tomasz Figa wrote:
>>>>> On Wed, Mar 13, 2019 at 12:52 AM Boris Brezillon
>>>>> <boris.brezillon@collabora.com> wrote:
>>>>>> On Tue, 12 Mar 2019 12:34:45 -0300
>>>>>> Helen Koike <helen.koike@collabora.com> wrote:
>>>>>>> On 3/12/19 3:34 AM, Boris Brezillon wrote:
>>>>>>>> On Mon, 11 Mar 2019 23:21:59 -0300
>>>>>>>> Helen Koike <helen.koike@collabora.com> wrote:
>>>>>>>>
>>>>>>>>> --- a/drivers/gpu/drm/rockchip/rockchip_drm_vop.c
>>>>>>>>> +++ b/drivers/gpu/drm/rockchip/rockchip_drm_vop.c
>>>>>>>>> @@ -912,30 +912,31 @@ static void vop_plane_atomic_async_update(struct drm_plane *plane,
>>>>>>>>>                                      struct drm_plane_state *new_state)
>>>>>>>>>  {
>>>>>>>>>    struct vop *vop = to_vop(plane->state->crtc);
>>>>>>>>> -  struct drm_plane_state *plane_state;
>>>>>>>>> +  struct drm_framebuffer *old_fb = plane->state->fb;
>>>>>>>>>
>>>>>>>>> -  plane_state = plane->funcs->atomic_duplicate_state(plane);
>>>>>>>>> -  plane_state->crtc_x = new_state->crtc_x;
>>>>>>>>> -  plane_state->crtc_y = new_state->crtc_y;
>>>>>>>>> -  plane_state->crtc_h = new_state->crtc_h;
>>>>>>>>> -  plane_state->crtc_w = new_state->crtc_w;
>>>>>>>>> -  plane_state->src_x = new_state->src_x;
>>>>>>>>> -  plane_state->src_y = new_state->src_y;
>>>>>>>>> -  plane_state->src_h = new_state->src_h;
>>>>>>>>> -  plane_state->src_w = new_state->src_w;
>>>>>>>>> -
>>>>>>>>> -  if (plane_state->fb != new_state->fb)
>>>>>>>>> -          drm_atomic_set_fb_for_plane(plane_state, new_state->fb);
>>>>>>>>> -
>>>>>>>>> -  swap(plane_state, plane->state);
>>>>>>>>> -
>>>>>>>>> -  if (plane->state->fb && plane->state->fb != new_state->fb) {
>>>>>>>>> +  /*
>>>>>>>>> +   * A scanout can still be occurring, so we can't drop the reference to
>>>>>>>>> +   * the old framebuffer. To solve this we get a reference to old_fb and
>>>>>>>>> +   * set a worker to release it later.
>>>>>>>>
>>>>>>>> Hm, doesn't look like an async update to me if we have to wait for the
>>>>>>>> next VBLANK to happen to get the new content on the screen. Maybe we
>>>>>>>> should reject async updates when old_fb != new_fb in the rk
>>>>>>>> ->async_check() hook.
>>>>>>>
>>>>>>> Unless I am misunderstanding this, we don't wait here, we just grab a
>>>>>>> reference to the fb in case it is being still used by the hw, so it
>>>>>>> doesn't get released prematurely.
>>>>>>
>>>>>> I was just reacting to the comment that says the new FB should stay
>>>>>> around until the next VBLANK event happens. If the FB must stay around
>>>>>> that probably means the HW is still using, which made me wonder if this
>>>>>> HW actually supports async update (where async means "update now and
>>>>>> don't care about about tearing"). Or maybe it takes some time to switch
>>>>>> to the new FB and waiting for the next VBLANK to release the old FB was
>>>>>> an easy solution to not wait for the flip to actually happen in
>>>>>> ->async_update() (which is kind of a combination of async+non-blocking).
>>>>>
>>>>> The hardware switches framebuffers on vblank, so whatever framebuffer
>>>>> is currently being scanned out from needs to stay there until the
>>>>> hardware switches to the new one in shadow registers. If that doesn't
>>>>> happen, you get IOMMU faults and the display controller stops working
>>>>> since we don't have any fault handling currently, just printing a
>>>>> message.
>>>>
>>>> Sounds like your hardware doesn't actually support async flips. It's
>>>> probably better for the driver not to pretend otherwise.
>>>
>>> I think wee need to clarify the meaning of the async_update callback
>>> (and we should clarify it in the docs).
>>>
>>> The way I understand what the async_update callback should do is: don't
>>> block (i.e. don't wait for the next vblank),
>>
>> Note that those are two separate things. "Async flips" are about "don't
>> wait for vblank", not about "don't block".
>>
>>
>>> and update the hw state at some point with the latest state from the
>>> last call to async_update.
>>>
>>> Which means that: any driver can implement the async_update callback,
>>> independently if it supports changing its state right away or not.
>>> If hw supports, async_update can change the hw state right away, if not,
>>> then changes will be applied in the next vblank (it can even amend the
>>> pending commit if there is one).
>>> With this, we can remove all the legacy cursor code to use the
>>> async_update callback, since async_update can be called 100 times before
>>> the next vblank, and the latest state will be set to the hw without
>>> waiting 100 vblanks.
>>>
>>> Please, let me know if this is your understanding as well. If not, then
>>> we need to remodel things.
>>
>> While this may make sense for cursor updates, I don't think it does for
>> async flips. If the flip only actually takes effect during the next
>> vblank, it doesn't really fit the definition and userspace expectation
>> of an async flip. It's better to clearly communicate to userspace that
>> the hardware cannot do async flips, than to pretend it can and fake
>> them. Userspace has to deal with this anyway, since async flips weren't
>> always supported in general.
> 
> What do you think if we separate two concepts here:
> 
> - amend mode: works like cursor updates, i.e, update the hw state at
> some point with the latest state from the last call to async_update. No
> special hardware support is required.
> 
> - async update: update hw state immediately. This depends if the hw
> supports it or not.
> 
> Every async update is an amend, but the opposite is not necessarily true.
> 
> What do you think if we rename the current async_update to amend_update,
> and we add a parameter "force_async" to it? (or maybe
> force_immediate_update?)
> Then amend_check with force_async=1 would fail if the hardware doesn't
> support it (we could also add flags in the capabilities to inform
> userspace the expected behaviour of things and if the hw supports
> force_sync).
> 
> Like this, we can implement the cursors using the amend_update (which is
> now called async_update), and async_flips with amend_update with
> force_async=1.

Might force_async make sense for cursor updates as well? I thought some
hardware supported HW cursor updates outside of vblank, but I'm not sure.

Without force_async, are cursor updates always applied to the hardware
on the next vblank, even if the pending commit is delayed further (e.g.
because a fence it depends on doesn't signal before vblank)? If cursor
updates can be delayed beyond the next vblank, that can result in bad
user experience.


-- 
Earthling Michel Dänzer               |              https://www.amd.com
Libre software enthusiast             |             Mesa and X developer