Re: [PATCH 1/2] drm: Introduce crtc->mode_valid() callback

[Jose Abreu]

Hi Daniel,


On 03-05-2017 16:00, Daniel Vetter wrote:
> On Wed, May 03, 2017 at 03:16:13PM +0100, Jose Abreu wrote:
>> Hi Daniel,
>>
>>
>> On 03-05-2017 07:19, Daniel Vetter wrote:
>>> On Tue, May 2, 2017 at 11:29 AM, Jose Abreu <Jose.Abreu@xxxxxxxxxxxx> wrote:
>>>> On 02-05-2017 09:48, Daniel Vetter wrote:
>>>>> On Wed, Apr 26, 2017 at 11:48:34AM +0100, Jose Abreu wrote:
>>>>>> Some crtc's may have restrictions in the mode they can display. In
>>>>>> this patch a new callback (crtc->mode_valid()) is introduced that
>>>>>> is called at the same stage of connector->mode_valid() callback.
>>>>>>
>>>>>> This shall be implemented if the crtc has some sort of restriction
>>>>>> so that we don't probe modes that will fail in the commit() stage.
>>>>>> For example: A given crtc may be responsible to set a clock value.
>>>>>> If the clock can not produce all the values for the available
>>>>>> modes then this callback can be used to restrict the number of
>>>>>> probbed modes to only the ones that can be displayed.
>>>>>>
>>>>>> If the crtc does not implement the callback then the behaviour will
>>>>>> remain the same. Also, for a given set of crtcs that can be bound to
>>>>>> the connector, if at least one can display the mode then the mode
>>>>>> will be probbed.
>>>>>>
>>>>>> Signed-off-by: Jose Abreu <joabreu@xxxxxxxxxxxx>
>>>>>> Cc: Carlos Palminha <palminha@xxxxxxxxxxxx>
>>>>>> Cc: Alexey Brodkin <abrodkin@xxxxxxxxxxxx>
>>>>>> Cc: Ville Syrjälä <ville.syrjala@xxxxxxxxxxxxxxx>
>>>>>> Cc: Daniel Vetter <daniel.vetter@xxxxxxxx>
>>>>>> Cc: Dave Airlie <airlied@xxxxxxxx>
>>>>> Not sure this is useful, since you still have to duplicate the exact same
>>>>> check into your ->mode_fixup hook. That seems to make things even more
>>>>> confusing.
>>>> Yeah, in arcpgu I had to duplicate the code in ->atomic_check.
>>>>
>>>>> Also this doesn't update the various kerneldoc comments. For the existing
>>>>> hooks. Since this topic causes so much confusion, I don't think a
>>>>> half-solution will help, but has some good potential to make things worse.
>>>> I only documented the callback in drm_modeset_helper_vtables.h.
>>>>
>>>> Despite all of this, I think it doesn't makes sense delivering
>>>> modes to userspace which can never be used.
>>>>
>>>> This is really annoying in arcpgu. Imagine: I try to use mpv to
>>>> play a video, the full set of modes from EDID were probed so if I
>>>> just start mpv it will pick the native mode of the TV instead of
>>>> the one that is supported, so mpv will fail to play. I know the
>>>> value of clock which will work (so I know what mode shall be
>>>> used), but a normal user which is not aware of the HW will have
>>>> to cycle through the list of modes and try them all until it hits
>>>> one that works. Its really boring.
>>>>
>>>> For the modes that user specifies manually there is nothing we
>>>> can do, but we should not trick users into thinking that a given
>>>> mode is supported when it will always fail at commit.
>>> Yes, you are supposed to filter these out in ->mode_valid. But my
>>> stance is that only adding a half-baked support for a new callback to
>>> the core isn't going to make life easier for drivers, it will just add
>>> to the confusion. There's already piles of docs for both @mode_valid
>>> and @mode_fixup hooks explaining this, I don't want to make the
>>> documentation even more complex. And half-baked crtc checking is
>>> _much_ easier to implement in the driver directly (e.g. i915 checks
>>> for crtc constraints since forever, as do the other big x86 drivers).
>> But i915 crtc checks are done after handing the mode to
>> userspace, arcpgu also does that. We must let users specify
>> manually a mode but there is no point in returning modes in
>> get_connector which will always fail to commit. I get your point
>> and this can lead to code duplication, but I don't think it will
>> lead to confusion as long as it is well documented. And besides,
>> the callback is completely optional.
> Look closer, e.g. intel_dp_mode_valid calls
> intel_dp_downstream_max_dotclock which also looks at
> dev_priv->max_dotclkc_freq (which is the source dotclk limit, yeah it's a
> misnamed function).
>
> And the max dotclk is very much a crtc limit, not a port limit. Note that
> a bunch of other ports have port limits which are guaranteed to be lower
> than the crtc limit, hence the absence of the checks.
>
>>> So all taken together, if we add a ->mode_valid to crtcs, then imo we
>>> should do it right and actually make life easier for drivers. A good
>>> proof would be if your patch would allow us to drop a lot of the
>>> lenghty language from the @mode_valid hooks.
>> I completely agree that it should make life easier for drivers
>> but unfortunately I don't really see how :/
>>
>> So, in summary:
>>     Disadvantage 1: Code duplication
>>     Disadvantage 2: Confusing documentation can lead to callback
>> misuse
>>
>>     Advantage 1: User will get life simpler
> Ok, let me try to explain a bit in more detail what I think would be a
> real improvement:
> - Add ->mode_valid checks to all the places where we currently have
>   ->mode_fixup. That'd be crtc, encoder and bridges.
>
> - Pimp the probe helper code to go through all of the combinations,
>   filtering out those that aren't allowed by possible_* masks (essentially
>   do the same thing that userspace is supposed to do).
>
> - Call all these ->mode_valid checks from the atomic check functions (I
>   think we can forget about the legacy crtc helpers for old drivers). Do
>   this also for connector->mode_valid.
>
> Taken all together this gives us the guarantee that that any mode which
> fails the check in the probe path is guaranteed to never pass in an atomic
> commit. And since the probed mode list is what developers generally see,
> that's hopefully enough to make sure the filtering is correct.
>
> It is a bit more code than what you've typed here, but not a lot:
> - probe path needs to loop over all CRTCxEncoder combos (the
>   encoder->bridge routing is fixed) instead over just CRTCs.
> - Call ->mode_valid in all the places we already call ->mode_fixup. You
>   don't need a new loop over all connectors to be able to call
>   ->mode_valid since we already have that connector loop in
>   check_modesets().
>
> With that we should also be able to simplify the documentation and rip out
> all the warnings about how this is tricky.

This seems very nice! So we essentially can remove the validation
of modes in atomic_check as mode_valid will be called before, right?

Best regards,
Jose Miguel Abreu

> -Daniel