Re: [PATCH 0/4] Virtio non-discoverable devices

[Daniil Egranov <daniil.egranov@arm.com>]

Hello Ard,

On 03/08/2018 02:29 AM, Ard Biesheuvel wrote:
> Hello Daniil,
> 
> Could you please use a text based email client? Gmail does not
> consider the indentation as threading, so the context below is
> unintelligible
> 

I forced Thunderbird to plain text. I assumed it should follow a 
previous email composition style. At least i did not have problem with 
it so far. I hope it's formated correctly now.

> On 8 March 2018 at 08:21, Daniil Egranov <daniil.egranov@arm.com> wrote:
>> Hello Ard,
>>
>> Thanks for reply. Please see my comments below.
>>
>> On 03/07/2018 02:03 AM, Ard Biesheuvel wrote:
>>
>> (+ Laszlo)
>>
>> Hello Daniil,
>>
>> On 7 March 2018 at 01:36, Daniil Egranov <daniil.egranov@arm.com> wrote:
>>
>> This is an attempt to add MMIO Virtio devices into the
>> non-discoverable device registration procedure and allow
>> Virtio PCI drivers to recognize and program such devices
>> correctly.
>>
>> Why? The purpose of the non-discoverable device layer is to make
>> non-PCI controllers that can be driven by PCI class drivers appear as
>> PCI devices. We have started using the base non-discoverable device
>> protocol for other devices as well, but the PCI wrapper is really only
>> intended for PCI class drivers.
>>
>>
>> I am looking for a proper way to handle multiple MMIO Virtio devices on a
>> single platform. As both PCI and MMIO types of Virtio device programmed in
>> about the same way, non-discoverable devices approach was looking valid.
>>
>> I understand you point. Correct me if I am wrong but all non-discoverable
>> devices are MMIO devices so if there is PCI version of the device exists,
>> PCI wrapper can be used. The Virtio PCI class devices are using
>> VirtioPciDeviceDxe driver. Is this driver not fitting to the category of PCI
>> class drivers?
>>
> 
> That is not the point. The Intel guys have decided that the AHCI, XHCI
> and other drivers (whose specs do no mandate PCI) are implemented as
> PCI drivers only, which means that they are essentially combining two
> layers of the driver stack (the PCI part and the _HCI part)
> 
> Splitting all of those drivers into PCI drivers that produce _HCI
> protocols and _HCI drivers that produce the USB host, SATA host, etc
> protocols is not going to be accepted by upstream EDK2, so instead, we
> decided to turn the PCI 'emulation' that was duplicated across many
> platforms into a proper abstraction.
> 
> In the virtio case, we don't have that problem. We have MMIO and PCI
> support using drivers that use the proper abstractions, and so
> presenting MMIO devices as PCI devices is really a step backwards.
> 

I see, thanks for details.

> What are you trying to achieve that the current code won't let you?
> 

I have a situation when a platform has multiple Virtio MMIO devices. The 
initial way to handle them was installing a device path protocol and 
calling VirtioMmioDeviceLib for each device as part of the platform DXE 
driver. The non-discoverable way was hiding all these operations and was 
looking like more clean approach.
In case of multiple Virtio MMIO devices, what will be a proper way to 
handle them?

Thanks,
Daniil