Re: [PATCH] drm/hyperv: Don't rely on screen_info.lfb_base for Gen1 VMs

[Thomas Zimmermann <tzimmermann@suse.de>]

[-- Attachment #1.1: Type: text/plain, Size: 5457 bytes --]

Hi

Am 13.09.22 um 17:15 schrieb Saurabh Singh Sengar:
> On Mon, Sep 12, 2022 at 09:03:53AM +0200, Thomas Zimmermann wrote:
>> Hi
>>
>> Am 11.09.22 um 18:21 schrieb Saurabh Singh Sengar:
>>> On Sat, Sep 10, 2022 at 08:11:24PM +0200, Thomas Zimmermann wrote:
>>>> Hi
>>>>
>>>> Am 09.09.22 um 16:43 schrieb Saurabh Sengar:
>>>>> hyperv_setup_vram tries to remove conflicting framebuffer based on
>>>>> 'screen_info'. As observed in past due to some bug or wrong setting
>>>>> in grub, the 'screen_info' fields may not be set for Gen1, and in such
>>>>> cases drm_aperture_remove_conflicting_framebuffers will not do anything
>>>>> useful.
>>>>> For Gen1 VMs, it should always be possible to get framebuffer
>>>>> conflict removed using PCI device instead.
>>>>>
>>>>> Fixes: a0ab5abced55 ("drm/hyperv : Removing the restruction of VRAM allocation with PCI bar size")
>>>>> Signed-off-by: Saurabh Sengar <ssengar@linux.microsoft.com>
>>>>> ---
>>>>>   drivers/gpu/drm/hyperv/hyperv_drm_drv.c | 24 ++++++++++++++++++++----
>>>>>   1 file changed, 20 insertions(+), 4 deletions(-)
>>>>>
>>>>> diff --git a/drivers/gpu/drm/hyperv/hyperv_drm_drv.c b/drivers/gpu/drm/hyperv/hyperv_drm_drv.c
>>>>> index 6d11e7938c83..b0cc974efa45 100644
>>>>> --- a/drivers/gpu/drm/hyperv/hyperv_drm_drv.c
>>>>> +++ b/drivers/gpu/drm/hyperv/hyperv_drm_drv.c
>>>>> @@ -73,12 +73,28 @@ static int hyperv_setup_vram(struct hyperv_drm_device *hv,
>>>>>   			     struct hv_device *hdev)
>>>>>   {
>>>>>   	struct drm_device *dev = &hv->dev;
>>>>> +	struct pci_dev *pdev;
>>>>>   	int ret;
>>>>> -	drm_aperture_remove_conflicting_framebuffers(screen_info.lfb_base,
>>>>> -						     screen_info.lfb_size,
>>>>> -						     false,
>>>>> -						     &hyperv_driver);
>>>>> +	if (efi_enabled(EFI_BOOT)) {
>>>>> +		drm_aperture_remove_conflicting_framebuffers(screen_info.lfb_base,
>>>>> +							     screen_info.lfb_size,
>>>>
>>>> Using screen_info here seems wrong in any case. You want to remove
>>>> the framebuffer devices that conflict with your driver, which might
>>>> be unrelated to screen_info. AFAICT the correct solution would
>>>> always retrieve the PCI device for removal (i.e., always do the else
>>>> branch).
>>>
>>> In a Gen2 VM, the Hyper-V frame buffer device is presented only as a VMbus device.
>>> It's not presented as a PCI device like it is in a Gen1 VM. This would have worked
>>> if we had the frame buffer device available as PCI device in Gen2 but unfortunately
>>> thats not the case here.
>>
>> Thanks for explaining. There is an instance of struct hv_device
>> passed to the probe function. I suspect you cannot get the
>> framebuffer range from this instance (e.g., via the device's
>> platform_data)?
>>
>> If you absolutely can't get the actual memory region from the
>> device, it's better to remove all framebuffers via
>> drm_aperture_remove_framebuffers() than to use screen_info.
>>
>> Best regards
>> Thomas
> 
> Thanks for your suggestion, and I thought of using drm_aperture_remove_framebuffers
> here, but this driver will be used in many different systems with many other graphics
> devices (GPU etc). Removing all the framebuffer is a bit blunt approach which may disturb
> the devices we are not intended to and which are even outside of the HyperV MMIO region.
> I feel this API use will be risky, and I would like to stick to the earlier method which
> is proven to be working for many years and we are sure it won't disturb anyone outside
> MMIO region.

But the problem with the current code is that you also remove the driver 
that covers the boot-up graphics, even if it's not even a hyperv device. 
That's why I said that you should try to detect if the hyperv device 
you're creating actually uses the screen_info when you create the device.

Set that information as device data and look it up in the hyperv-drm 
driver. If no such screen_info data is attached to the device, you don't 
have to remove conflicting framebuffers at all.

Best regards
Thomas

> 
> Regards,
> Saurabh
>>
>>>
>>> Regards,
>>> Saurabh
>>>
>>>>
>>>> Best regard
>>>> Thomas
>>>>
>>>>> +							     false,
>>>>> +							     &hyperv_driver);
>>>>> +	} else {
>>>>> +		pdev = pci_get_device(PCI_VENDOR_ID_MICROSOFT, PCI_DEVICE_ID_HYPERV_VIDEO, NULL);
>>>>> +		if (!pdev) {
>>>>> +			drm_err(dev, "Unable to find PCI Hyper-V video\n");
>>>>> +			return -ENODEV;
>>>>> +		}
>>>>> +
>>>>> +		ret = drm_aperture_remove_conflicting_pci_framebuffers(pdev, &hyperv_driver);
>>>>> +		pci_dev_put(pdev);
>>>>> +		if (ret) {
>>>>> +			drm_err(dev, "Not able to remove boot fb\n");
>>>>> +			return ret;
>>>>> +		}
>>>>> +	}
>>>>>   	hv->fb_size = (unsigned long)hv->mmio_megabytes * 1024 * 1024;
>>>>
>>>> -- 
>>>> Thomas Zimmermann
>>>> Graphics Driver Developer
>>>> SUSE Software Solutions Germany GmbH
>>>> Maxfeldstr. 5, 90409 Nürnberg, Germany
>>>> (HRB 36809, AG Nürnberg)
>>>> Geschäftsführer: Ivo Totev
>>>
>>>
>>>
>>
>> -- 
>> Thomas Zimmermann
>> Graphics Driver Developer
>> SUSE Software Solutions Germany GmbH
>> Maxfeldstr. 5, 90409 Nürnberg, Germany
>> (HRB 36809, AG Nürnberg)
>> Geschäftsführer: Ivo Totev
> 
> 
> 

-- 
Thomas Zimmermann
Graphics Driver Developer
SUSE Software Solutions Germany GmbH
Maxfeldstr. 5, 90409 Nürnberg, Germany
(HRB 36809, AG Nürnberg)
Geschäftsführer: Ivo Totev

[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 840 bytes --]