Re: [PATCH v1] drm/modes: Skip invalid cmdline mode

[Dmitry Osipenko]

12.07.2019 22:53, Maxime Ripard ÐÐÑÐÑ:
> On Fri, Jul 12, 2019 at 11:30:01AM +0300, Dmitry Osipenko wrote:
>> 12.07.2019 11:10, Maxime Ripard ÐÐÑÐÑ:
>>> On Thu, Jul 11, 2019 at 06:55:03PM +0300, Dmitry Osipenko wrote:
>>>> 11.07.2019 12:03, Maxime Ripard ÐÐÑÐÑ:
>>>>> On Wed, Jul 10, 2019 at 06:05:18PM +0300, Dmitry Osipenko wrote:
>>>>>> 10.07.2019 17:05, Maxime Ripard ÐÐÑÐÑ:
>>>>>>> On Wed, Jul 10, 2019 at 04:29:19PM +0300, Dmitry Osipenko wrote:
>>>>>>>> This works:
>>>>>>>>
>>>>>>>> diff --git a/drivers/gpu/drm/drm_client_modeset.c b/drivers/gpu/drm/drm_client_modeset.c
>>>>>>>> index 56d36779d213..e5a2f9c8f404 100644
>>>>>>>> --- a/drivers/gpu/drm/drm_client_modeset.c
>>>>>>>> +++ b/drivers/gpu/drm/drm_client_modeset.c
>>>>>>>> @@ -182,6 +182,8 @@ drm_connector_pick_cmdline_mode(struct drm_connector *connector)
>>>>>>>>         mode = drm_mode_create_from_cmdline_mode(connector->dev, cmdline_mode);
>>>>>>>>         if (mode)
>>>>>>>>                 list_add(&mode->head, &connector->modes);
>>>>>>>> +       else
>>>>>>>> +               cmdline_mode->specified = false;
>>>>>>>
>>>>>>> Hmmm, it's not clear to me why that wouldn't be the case.
>>>>>>>
>>>>>>> If we come back to the beginning of that function, we retrieve the
>>>>>>> cmdline_mode buffer from the connector pointer, that will probably
>>>>>>> have been parsed a first time using drm_mode_create_from_cmdline_mode
>>>>>>> in drm_helper_probe_add_cmdline_mode.
>>>>>>>
>>>>>>> Now, I'm guessing that the issue is that in
>>>>>>> drm_mode_parse_command_line_for_connector, if we have a named mode, we
>>>>>>> just copy the mode over and set mode->specified.
>>>>>>>
>>>>>>> And we then move over to do other checks, and that's probably what
>>>>>>> fails and returns, but our drm_cmdline_mode will have been modified.
>>>>>>>
>>>>>>> I'm not entirely sure how to deal with that though.
>>>>>>>
>>>>>>> I guess we could allocate a drm_cmdline_mode structure on the stack,
>>>>>>> fill that, and if successful copy over its content to the one in
>>>>>>> drm_connector. That would allow us to only change the content on
>>>>>>> success, which is what I would expect from such a function?
>>>>>>>
>>>>>>> How does that sound?
>>>>>>
>>>>>> I now see that there is DRM_MODE_TYPE_USERDEF flag that is assigned only
>>>>>> for the "cmdline" mode and drm_client_rotation() is the only place in
>>>>>> DRM code that cares about whether mode is from cmdline, hence looks like
>>>>>> it will be more correct to do the following:
>>>>>
>>>>> I'm still under the impression that we're dealing with workarounds of
>>>>> a more central issue, which is that we shouldn't return a partially
>>>>> modified drm_cmdline_mode.
>>>>>
>>>>> You said it yourself, the breakage is in the commit changing the
>>>>> command line parsing logic, while you're fixing here some code that
>>>>> was introduced later on.
>>>>
>>>> The problem stems from assumption that *any* named mode is valid. It
>>>> looks to me that the ultimate solution would be to move the mode's name
>>>> comparison into the [1], if that's possible.
>>>>
>>>> [1] drm_mode_parse_command_line_for_connector()
>>>
>>> Well, one could argue that video=tegrafb is invalid and should be
>>> rejected as well, but we haven't cleared that up.
>>
>> The video=tegrafb is invalid mode, there is nothing to argue here. And
>> the problem is that invalid modes and not rejected for the very beginning.
> 
> Yeah, I guess fb_get_options should also return an error in such a
> case, but I'm a bit worried about the side effects here.

At least the showstopper regression is fixed now. Everything else could
be worked out later on.

>>>>> Can you try the followintg patch?
>>>>> http://code.bulix.org/8cwk4c-794565?raw
>>>>
>>>> This doesn't help because the problem with the rotation_reflection is
>>>> that it's 0 if "rotation" not present in the cmdline and then ilog2(0)
>>>> returns -1. So the patch "drm/modes: Don't apply cmdline's rotation if
>>>> it wasn't specified" should be correct in any case.
>>>
>>> So we would have the same issue with rotate=0 then?
>>
>> No, we won't. Rotation mode is parsed into the DRM_MODE bitmask and
>> rotate=0 corresponds to DRM_MODE_ROTATE_0, which is BIT(0) as you may
>> notice. Hence rotation_reflection=0 is always an invalid value, meaning
>> that "rotate" option does not present in the cmdline. Please consult the
>> code, in particular see drm_mode_parse_cmdline_options() which was
>> written by yourself ;)
> 
> Sigh... You're right :)
> 
> Sorry for that, I'll reply to the other patch

Thank you very much.