Re: [PATCH] drm_fb_helper: fix regression in pixclock check

[Jason Wessel <jason.wessel@windriver.com>]

Clemens Ladisch wrote:
> Jason Wessel wrote:
>   
>> Commit 5349ef3127c77075ff70b2014f17ae0fbcaaf199 changed pixclock to be
>> initialized to zero instead of -1.
>>     
>
> Zero meaning "there is no pixel clock value".  When I wrote the patch,
> the FB (helper) code did not use this value at all, so zero was the only
> value that could possibly be used.
>
>   
>> The validation routine always returns -EINVAL for a valid pixclock
>>     
>
> It declares that the only valid pixclock value is "none".  This was
> required to make the userspace API work, since the GET ioctls return
> this value, and all programs expect to be able to PUT this value back.
>
>   
>> which prevents atomic kernel mode setting from working correctly.
>> ...
>> @@ -602,7 +602,7 @@ int drm_fb_helper_check_var(struct fb_var_screeninfo *var,
>>  
>> -	if (var->pixclock != 0)
>> +	if (!var->pixclock)
>>  		return -EINVAL;
>>     
>
> This change breaks the userspace API again, unless the other parts of
> the code have been changed to always return a valid pixclock value when
> reading the current mode settings.
>
> If we want to accept both zero and non-zero values for pixclock, this
> check should be dropped altogether.
>
> In drm_fb_helper_set_par(), there is a corresponding check that must be
> kept in sync with this one.
>
>   

Fair enough.  We'll make a special case for when the kernel debugger is
active.

Looks to me like it worked by luck during the 2.6.32 kernel cycle for
the atomic kernel mode setting.  I have changed it in the next version
of this patch to always return -EINVAL while the kernel debugger is
active.  This restores the atomic kernel mode setting back to the
working state, until we can find a better solution.

Thanks,
Jason.