[PATCH v9 9/9] PCI: Add a new 'boot_display' attribute

[Mario Limonciello]

On 7/21/25 6:00 PM, Bjorn Helgaas wrote:
> On Fri, Jul 18, 2025 at 12:44:11PM -0500, Mario Limonciello wrote:
>> On 7/18/2025 12:36 PM, Bjorn Helgaas wrote:
>>> On Fri, Jul 18, 2025 at 12:29:05PM -0500, Mario Limonciello wrote:
>>>> On 7/18/2025 12:25 PM, Bjorn Helgaas wrote:
>>>>> On Thu, Jul 17, 2025 at 12:38:12PM -0500, Mario Limonciello wrote:
>>>>>> From: Mario Limonciello <mario.limonciello at amd.com>
>>>>>>
>>>>>> On systems with multiple GPUs there can be uncertainty which GPU is the
>>>>>> primary one used to drive the display at bootup. In some desktop
>>>>>> environments this can lead to increased power consumption because
>>>>>> secondary GPUs may be used for rendering and never go to a low power
>>>>>> state. In order to disambiguate this add a new sysfs attribute
>>>>>> 'boot_display' that uses the output of video_is_primary_device() to
>>>>>> populate whether a PCI device was used for driving the display.
>>>>>
>>>>>> +What:		/sys/bus/pci/devices/.../boot_display
>>>>>> +Date:		October 2025
>>>>>> +Contact:	Linux PCI developers <linux-pci at vger.kernel.org>
>>>>>> +Description:
>>>>>> +		This file indicates that displays connected to the device were
>>>>>> +		used to display the boot sequence.  If a display connected to
>>>>>> +		the device was used to display the boot sequence the file will
>>>>>> +		be present and contain "1".
>>>>>
>>>>>>     int __must_check pci_create_sysfs_dev_files(struct pci_dev *pdev)
>>>>>>     {
>>>>>> +	int retval;
>>>>>> +
>>>>>>     	if (!sysfs_initialized)
>>>>>>     		return -EACCES;
>>>>>> +	retval = pci_create_boot_display_file(pdev);
>>>>>
>>>>> In addition to Mani's question about whether /sys/bus/pci/ is
>>>>> the right place for this (which is a very good question), it's
>>>>> also been pointed out to me that we've been trying to get rid
>>>>> of pci_create_sysfs_dev_files() for years.
>>>>>
>>>>> If it's possible to make this a static attribute that would be
>>>>> much, much cleaner.
>>>>
>>>> Right - I tried to do this, but the problem is at the time the
>>>> PCI device is created the information needed to make the
>>>> judgement isn't ready.  The options end up being:
>>>> * a sysfs file for every display device with 0/1
>>>> * a sysfs file that is not accurate until later in the boot
>>>
>>> What's missing?  The specifics might be helpful if someone has
>>> another crack at getting rid of pci_create_sysfs_dev_files() in
>>> the future.
>>
>> The underlying SCREEN_INFO code tries to walk through all the PCI
>> devices in a loop, but at the time all the devices are walked the
>> memory regions associated with the device weren't populated.
> 
> Which loop are you referring to that walks through all the PCI
> devices?  I see this:
> 
>    efifb_set_system
>      for_each_pci_dev(dev)
> 
> but that only looks at VGA devices and IIUC you also want to look at
> non-VGA GPUs.
> 
> I don't see a loop in *this* series, where the screen_info path looks
> like this:
> 
>    pci_create_boot_display_file
>      video_is_primary_device
>        screen_info_pci_dev      # added by "fbcon: Use screen info to find primary device"
>          screen_info_resources
>          __screen_info_pci_dev
> 
> and we're basically matching the screen_info base/address with BAR
> values.
> 
> The usual problem is that BARs may not have been assigned by the time
> pci_device_add() -> device_add() creates the static attributes.
> 
> So we call pci_assign_unassigned_root_bus_resources() to assign all
> the BARs.  Then we call pci_create_sysfs_dev_files(), where
> pci_create_resource_files() creates a "resource%d" file for each BAR.
> 
> But since we're trying to find the GPU that was used by BIOS, I assume
> its BARs were programmed by BIOS and we shouldn't have to wait until
> after pci_assign_unassigned_root_bus_resources().
> 
> Bjorn

Yes it was screen_info_pci_dev() and __screen_info_pci_dev().  The 
resources weren't ready on the first call into __screen_info_pci_dev().

That's why the attribute needed to be created later.  But the sysfs 
group update or using DRM both avoid this problem and are totally fine 
alternatives.