Re: [PATCH v10] i2c: virtio: add a virtio i2c frontend driver

[Arnd Bergmann <arnd@arndb.de>]

On Mon, Jun 28, 2021 at 10:39 AM Wolfram Sang <wsa@kernel.org> wrote:
>
> sorry for the long delay. I am not familiar with VFIO, so I had to dive
> into the topic a little first. I am still not seeing through it
> completely, so I have very high-level questions first.

You probably know this already, but just in case for clarification
these are two different things:

VFIO: kernel feature to make raw (usually PCI) devices available
           to user space drivers and virtual machines from a kernel
           running on bare metal.

virtio: transport protocol for implementing arbitrary paravirtualized
          drivers in (usually) a virtual machine guest without giving the
          guest access to hardware registers.

Both can be used for letting a KVM guest talk to the outside world,
but usually you have one or the other, not both.

> > The device specification can be found on
> > https://lists.oasis-open.org/archives/virtio-comment/202101/msg00008.html.
>
> I think we need to start here:
>
> ===
>
> If ``length of \field{read_buf}''=0 and ``length of \field{write_buf}''>0,
> the request is called write request.
>
> If ``length of \field{read_buf}''>0 and ``length of \field{write_buf}''=0,
> the request is called read request.
>
> If ``length of \field{read_buf}''>0 and ``length of \field{write_buf}''>0,
> the request is called write-read request. It means an I2C write segment followed
> by a read segment. Usually, the write segment provides the number of an I2C
> controlled device register to be read.
>
> ===
>
> I2C transactions can have an arbitrary number of messages which can
> arbitrarily be read or write. As I understand the above, only one read,
> write or read-write transaction is supported. If that is the case, it
> would be not very much I2C but more SMBus. If my assumptions are true,
> we first need to decide if you want to go the I2C way or SMBus subset.

This has come up in previous reviews already. I think it comes down
to the requirement that the virtio i2c protocol should allow passthrough
access to any client devices connected to a physical i2c bus on the host,
and this should ideally be independent of whether the host driver
exposes I2C_RDWR or I2C_SMBUS ioctl interface, or both.

This can be done either by having both interface types in the transport,
or picking one of the two, and translating to the host interface type
in software.

As far as I understand me (please clarify), implementing only the smbus
subset would mean that we cannot communicate with all client devices,
while implementing both would add more complexity than the lower-level
protocol.

> ===
>
> The case when ``length of \field{write_buf}''=0, and at the same time,
> ``length of \field{read_buf}''=0 doesn't make any sense.
>
> ===
>
> Oh, it does. That's a legal transfer, both in SMBus and I2C. It is used
> to e.g. discover devices. I think it should be supported, even though
> not all bus master drivers on the host can support it. Is it possible?
>
> Also, as I read it, a whole bus is para-virtualized to the guest, or?
> Wouldn't it be better to allow just specific devices on a bus? Again, I
> am kinda new to this, so I may have overlooked things.

Do you mean just allowing a single device per bus (as opposed to
having multiple devices as on a real bus), or just allowing
a particular set of client devices that can be identified using
virtio specific configuration (as opposed to relying on device
tree or similar for probing). Both of these are valid questions that
have been discussed before, but that could be revisited.

          Arnd