Re: [PATCH v10 3/4] fbcon: Use screen info to find primary device

[Mario Limonciello <superm1@kernel.org>]

On 10/28/25 11:50 AM, Thomas Zimmermann wrote:
> Hi
> 
> Am 28.10.25 um 14:15 schrieb Mario Limonciello:
>> On 10/28/25 5:16 AM, Aaron Erhardt wrote:
>>>
>>> On Mon, Aug 11, 2025 at 11:26:05AM -0500, Mario Limonciello (AMD) wrote:
>>>> On systems with non VGA GPUs fbcon can't find the primary GPU because
>>>> video_is_primary_device() only checks the VGA arbiter.
>>>>
>>>> Add a screen info check to video_is_primary_device() so that callers
>>>> can get accurate data on such systems.
>>>
>>> I have a question regarding this change. To me, the function name
>>> video_is_primary_device() implies that there is only one primary GPU.
>>> I would also expect that the 'boot_display' attribute added later in
>>> the patch series based on this function is only set for one GPU, but
>>> that is not necessarily the case. Since I'm working on a user-space
>>> program that reads the 'boot_display' attribute, I need to know what
>>> behavior is intended in order to do a correct implementation.
>>>
>>>>
>>>> Reviewed-by: Thomas Zimmermann <tzimmermann@suse.de>
>>>> Suggested-by: Thomas Zimmermann <tzimmermann@suse.de>
>>>> Suggested-by: Bjorn Helgaas <bhelgaas@google.com>
>>>> Signed-off-by: Mario Limonciello (AMD) <superm1@kernel.org>
>>>> ---
>>>> v10:
>>>>   * Rebase on 6.17-rc1
>>>>   * Squash 'fbcon: Stop using screen_info_pci_dev()'
>>>> ---
>>>>   arch/x86/video/video-common.c | 25 ++++++++++++++++++++++++-
>>>>   1 file changed, 24 insertions(+), 1 deletion(-)
>>>>
>>>> diff --git a/arch/x86/video/video-common.c b/arch/x86/video/video- 
>>>> common.c
>>>> index 81fc97a2a837a..e0aeee99bc99e 100644
>>>> --- a/arch/x86/video/video-common.c
>>>> +++ b/arch/x86/video/video-common.c
>>>> @@ -9,6 +9,7 @@
>>>>     #include <linux/module.h>
>>>>   #include <linux/pci.h>
>>>> +#include <linux/screen_info.h>
>>>>   #include <linux/vgaarb.h>
>>>>     #include <asm/video.h>
>>>> @@ -27,6 +28,11 @@ EXPORT_SYMBOL(pgprot_framebuffer);
>>>>     bool video_is_primary_device(struct device *dev)
>>>>   {
>>>> +#ifdef CONFIG_SCREEN_INFO
>>>> +    struct screen_info *si = &screen_info;
>>>> +    struct resource res[SCREEN_INFO_MAX_RESOURCES];
>>>> +    ssize_t i, numres;
>>>> +#endif
>>>>       struct pci_dev *pdev;
>>>>         if (!dev_is_pci(dev))
>>>> @@ -34,7 +40,24 @@ bool video_is_primary_device(struct device *dev)
>>>>         pdev = to_pci_dev(dev);
>>>>   -    return (pdev == vga_default_device());
>>>> +    if (!pci_is_display(pdev))
>>>> +        return false;
>>>> +
>>>> +    if (pdev == vga_default_device())
>>>> +        return true;
>>>
>>> This can mark a VGA device as primary GPU.
> 
> Is the value returned from vga_default_device() eq to NULL?
> 
>>>
>>>> +
>>>> +#ifdef CONFIG_SCREEN_INFO
>>>> +    numres = screen_info_resources(si, res, ARRAY_SIZE(res));
>>>> +    for (i = 0; i < numres; ++i) {
>>>> +        if (!(res[i].flags & IORESOURCE_MEM))
>>>> +            continue;
>>>> +
>>>> +        if (pci_find_resource(pdev, &res[i]))
>>>> +            return true;
>>>> +    }
>>>> +#endif
>>>
>>> And then the new code can also choose a primary GPU.
> 
> Maybe we should drop this block or move it to [1]? At [1] it would only 
> run if the more sophisticated vgaarb has been disabled.
> 
> The vgaarb now selects CONFIG_SCREEN_INFO and already tests for 
> overlapping resources. There's nothing here that vgaarb shouldn't 
> already do. Yet, I don't understand how only one of the can be true at a 
> time.
> 
> [1] https://elixir.bootlin.com/linux/v6.18-rc3/source/include/linux/ 
> vgaarb.h#L55

There is a hunk at the end of vga_is_boot_device() which I think is 
causing this issue.  We have one VGA device which is NOT the boot 
display but the code figures if it found nothing this must be right.

	/*
	 * Vgadev has neither IO nor MEM enabled.  If we haven't found any
	 * other VGA devices, it is the best candidate so far.
	 */
	if (!boot_vga)
		return true;

I feel the right solution is to drop this now.

> 
>>>
>>>> +
>>>> +    return false;
>>>>   }
>>>>   EXPORT_SYMBOL(video_is_primary_device);
>>>
>>> In particular, I have hardware that has this exact configuration where
>>> two GPUs are marked as primary and have a 'boot_display' attribute: the
>>> first one through vga_default_device(), the second one through the new
>>> detection method.
>>>
>>> Is this intended?
>>>
>>> Kind regards
>>> Aaron
>>>
>>
>> I wouldn't have expected a case like this and I think it means there 
>> is a logic error.
>>
>> Can you please file a kernel bugzilla with details about your system 
>> and CC me?
>>
>> dmesg and lspci -vvnn please
>>
>> Also; please clarify which GPU shows something when booting up.
> 
> Agreed.
> 
> Best regards
> Thomas
> 
>