Re: [edk2-devel] Does EDK2 ArmVirtPkg has support for a virtio-mmio-blk device

["Laszlo Ersek" <lersek@redhat.com>]

On 02/09/21 03:54, Ying Fang wrote:

> I now realize that we emulate the virtio-blk-device over mmio, and we
> only emulate virtio-1.0 spec.
> As mentioned in (1c) , EDK2 only supports virtio-0.95 spec for now, so
> this maybe a big problem.
> Since it may not handshake ok if we only emulate virtio-1.0.

Yes.

First, the MMIO transport (as I remember from checking it last time,
which was quite some time ago) is very different between 0.9.5 and 1.0.

Second, device initialization steps differ:
- between 0.9.5 MMIO and 0.9.5 PCI,
- between 0.9.5 and 1.0, regardless of transport.

This means that the device driver code has *some* specifics (=
abstraction leaks) that relate to the underlying transport (MMIO vs.
PCI, and 0.9.5 vs. 1.0). See:

  OvmfPkg/VirtioBlkDxe/VirtioBlk.c

    752   //
    753   // Set Page Size - MMIO VirtIo Specific
    754   //
    755   Status = Dev->VirtIo->SetPageSize (Dev->VirtIo, EFI_PAGE_SIZE);

    822   //
    823   // In virtio-1.0, feature negotiation is expected to complete before queue
    824   // discovery, and the device can also reject the selected set of features.
    825   //
    826   if (Dev->VirtIo->Revision >= VIRTIO_SPEC_REVISION (1, 0, 0)) {
    827     Status = Virtio10WriteFeatures (Dev->VirtIo, Features, &NextDevStat);

    867   //
    868   // Additional steps for MMIO: align the queue appropriately, and set the
    869   // size. If anything fails from here on, we must unmap the ring resources.
    870   //
    871   Status = Dev->VirtIo->SetQueueNum (Dev->VirtIo, QueueSize);

    894   //
    895   // step 5 -- Report understood features.
    896   //
    897   if (Dev->VirtIo->Revision < VIRTIO_SPEC_REVISION (1, 0, 0)) {
    898     Features &= ~(UINT64)(VIRTIO_F_VERSION_1 | VIRTIO_F_IOMMU_PLATFORM);
    899     Status = Dev->VirtIo->SetGuestFeatures (Dev->VirtIo, Features);

We tried to make these "abstraction leaks" as small as possible; for
example the MMIO specific operations (SetPageSize, SetQueueNum) are
performed unconditionally, and it's only the PCI transport backends that
simply ignore those actions (after performing some sanity checks).
However, the different order of initialization steps couldn't really be
hidden (I wasn't comfortable with simply regression-testing the new 1.0
order against 0.9.5 transports of QEMU, so we kept both init orders).

Virtio MMIO was always classified as "temporary" and "legacy", needed
only until PCI support would be brought about on ARM. So given the
increased complexity of Virtio MMIO in the 1.0 spec, I always believed
that designing and implementing the latter in OVMF would be a waste of
effort.


> I will try to emulate the virtio-0.95 later to see if it is the root
> cause.

Yes, please either do that, or please add a PCI host.

Given that you do get a BLK0: alias in the UEFI shell, the
initialization of the device might even *appear* to complete, to OVMF;
however, the actual virtio transfers likely fail.


> BTW, I found it really hard to read and understand the EDK2 code for
> me, there is a long way to go.

Yes. Edk2 uses PPIs and protocols [*] and library classes / library
instances and sometimes callback registrations for composability, and so
edk2 is really difficult to read in comparison to other projects, where
you can just follow function calls.

In edk2, you have to grep the source code for GUIDs, to understand what
calls what. It was one of the hardest things for me as well, when
starting with edk2.

[*] Basically a GUID-identified structure of function pointers, and some
data fields.

Thanks
Laszlo