Re: [PATCH] drm: mxsfb: check framebuffer pitch

[Stefan Agner <stefan@agner.ch>]

On 2020-09-08 14:33, Laurent Pinchart wrote:
> On Tue, Sep 08, 2020 at 02:29:02PM +0200, Daniel Vetter wrote:
>> On Tue, Sep 8, 2020 at 2:07 PM Stefan Agner <stefan@agner.ch> wrote:
>> > On 2020-09-08 10:48, Daniel Vetter wrote:
>> >> On Tue, Sep 08, 2020 at 11:18:25AM +0300, Tomi Valkeinen wrote:
>> >>> On 08/09/2020 10:55, Stefan Agner wrote:
>> >>>> On 2020-09-07 20:18, Daniel Vetter wrote:
>> >>>>> On Mon, Sep 07, 2020 at 07:17:12PM +0300, Laurent Pinchart wrote:
>> >>>>>> Hi Stefan,
>> >>>>>>
>> >>>>>> Thank you for the patch.
>> >>>>>>
>> >>>>>> On Mon, Sep 07, 2020 at 06:03:43PM +0200, Stefan Agner wrote:
>> >>>>>>> The lcdif IP does not support a framebuffer pitch (stride) other than
>> >>>>>>> the CRTC width. Check for equality and reject the state otherwise.
>> >>>>>>>
>> >>>>>>> This prevents a distorted picture when using 640x800 and running the
>> >>>>>>> Mesa graphics stack. Mesa tires to use a cache aligned stride, which
>> >>>>>>
>> >>>>>> s/tires/tries/
>> >>>>>>
>> >>>>>>> leads at that particular resolution to width != stride. Currently
>> >>>>>>> Mesa has no fallback behavior, but rejecting this configuration allows
>> >>>>>>> userspace to handle the issue correctly.
>> >>>>>>
>> >>>>>> I'm increasingly impressed by how featureful this IP core is :-)
>> >>>>>>
>> >>>>>>> Signed-off-by: Stefan Agner <stefan@agner.ch>
>> >>>>>>> ---
>> >>>>>>>  drivers/gpu/drm/mxsfb/mxsfb_kms.c | 22 ++++++++++++++++++----
>> >>>>>>>  1 file changed, 18 insertions(+), 4 deletions(-)
>> >>>>>>>
>> >>>>>>> diff --git a/drivers/gpu/drm/mxsfb/mxsfb_kms.c b/drivers/gpu/drm/mxsfb/mxsfb_kms.c
>> >>>>>>> index b721b8b262ce..79aa14027f91 100644
>> >>>>>>> --- a/drivers/gpu/drm/mxsfb/mxsfb_kms.c
>> >>>>>>> +++ b/drivers/gpu/drm/mxsfb/mxsfb_kms.c
>> >>>>>>> @@ -403,14 +403,28 @@ static int mxsfb_plane_atomic_check(struct drm_plane *plane,
>> >>>>>>>  {
>> >>>>>>>         struct mxsfb_drm_private *mxsfb = to_mxsfb_drm_private(plane->dev);
>> >>>>>>>         struct drm_crtc_state *crtc_state;
>> >>>>>>> +       unsigned int pitch;
>> >>>>>>> +       int ret;
>> >>>>>>>
>> >>>>>>>         crtc_state = drm_atomic_get_new_crtc_state(plane_state->state,
>> >>>>>>>                                                    &mxsfb->crtc);
>> >>>>>>>
>> >>>>>>> -       return drm_atomic_helper_check_plane_state(plane_state, crtc_state,
>> >>>>>>> -                                                  DRM_PLANE_HELPER_NO_SCALING,
>> >>>>>>> -                                                  DRM_PLANE_HELPER_NO_SCALING,
>> >>>>>>> -                                                  false, true);
>> >>>>>>> +       ret = drm_atomic_helper_check_plane_state(plane_state, crtc_state,
>> >>>>>>> +                                                 DRM_PLANE_HELPER_NO_SCALING,
>> >>>>>>> +                                                 DRM_PLANE_HELPER_NO_SCALING,
>> >>>>>>> +                                                 false, true);
>> >>>>>>> +       if (ret || !plane_state->visible)
>> >>>>>>
>> >>>>>> Would it be more explict to check for !plane_state->fb ? Otherwise I'll
>> >>>>>> have to verify that !fb always implies !visible :-)
>> >>>>>>
>> >>>>>>> +               return ret;
>> >>>>>>> +
>> >>>>>>> +       pitch = crtc_state->mode.hdisplay *
>> >>>>>>> +               plane_state->fb->format->cpp[0];
>> >>>>>>
>> >>>>>> This holds on a single line.
>> >>>>>>
>> >>>>>>> +       if (plane_state->fb->pitches[0] != pitch) {
>> >>>>>>> +               dev_err(plane->dev->dev,
>> >>>>>>> +                       "Invalid pitch: fb and crtc widths must be the same");
>> >>>>>>
>> >>>>>> I'd turn this into a dev_dbg(), printing error messages to the kernel
>> >>>>>> log in response to user-triggered conditions is a bit too verbose and
>> >>>>>> could flood the log.
>> >>>>>>
>> >>>>>> Wouldn't it be best to catch this issue when creating the framebuffer ?
>> >>>>>
>> >>>>> Yeah this should be verified at addfb time. We try to validate as early as
>> >>>>> possible.
>> >>>>> -Daniel
>> >>>>>
>> >>>>
>> >>>> Sounds sensible. From what I can tell fb_create is the proper callback
>> >>>> to implement this at addfb time. Will give this a try.
>> >>>>
>> >>>> FWIW, I got the idea from drivers/gpu/drm/tilcdc/tilcdc_plane.c. Maybe
>> >>>> should be moved to addfb there too?
>> >>>
>> >>> But you don't know the crtc width when creating the framebuffer.
>> >>
>> >> Hm right this is a different check. What we could check in fb_create for
>> >> both is that the logical fb size matches exactly the pitch. That's not
>> >> sufficient criteria, but it will at least catch some of them already.
>> >>
>> >> But yeah we'd need both here.
>> >
>> > After validating width of framebuffer against pitch, the only thing we
>> > need to check here is that the width matches. From what I can tell,
>> > least for mxsfb, this should be covered by
>> > drm_atomic_helper_check_plane_state's can_position parameter set to
>> > false.
>>
>> This only checks against the src rectangle of the crtc state, there's
>> nothing forcing that the size of the fb matches the src rectangle
>> exactly. I guess we could maybe add that as another parameter for hw
>> like yours or tilcdc. Naming is a bit tricky, maybe
>> require_matching_fb or src_must_match_fb or something like that.
> 
> Can we turn those parameters into flags ? false, true, false is hard to
> read.
> 

Since it must match, in this case, it would be false, true, true,
obviously ;-)

I guess this would mean to convert the two existing boolean parameters
to flags first, and then introduce a new flag handling fb size vs. CRTC
src.

Hm, this gets all a bit more involved. It is actually not the issue at
hand (in my case the fb width does match the CRTC). Not sure if that
case is actually a problem in real world? I can give this a shot still,
if preferred. But I would do it independently of the framebuffer pitch
validation.

--
Stefan

>> > So I think in my case I can get away by only checking the framebuffer.
>>
>> You still need both I think.