Re: [PATCH 0/8] Virtio-over-PCIe on non-MIC

[Vincent Whitchurch]

On Thu, Jan 17, 2019 at 04:53:25PM +0100, Arnd Bergmann wrote:
> On Thu, Jan 17, 2019 at 4:19 PM Vincent Whitchurch
> <vincent.whitchurch@xxxxxxxx> wrote:
> > On Thu, Jan 17, 2019 at 01:39:27PM +0100, Arnd Bergmann wrote:
> > > Can you describe how you expect a VOP device over NTB or
> > > PCIe-endpoint would get created, configured and used?
> >
> > Assuming PCIe-endpoint:
> >
> > On the RC, a vop-host-backend driver (PCI driver) sets up some shared
> > memory area which the RC and the endpoint can use to communicate the
> > location of the MIC device descriptors and other information such as the
> > MSI address.  It implements vop callbacks to allow the vop framework to
> > obtain the address of the MIC descriptors and send/receive interrupts
> > to/from the guest.
> >
> > On the endpoint, the PCIe endpoint driver sets up (hardcoded) BARs and
> > memory regions as required to allow the endpoint and the root complex to
> > access each other's memory.
> >
> > On the endpoint, the vop-guest-backend, via the shared memory set up by
> > the vop-host-backend, obtains the address of the MIC device page and the
> > MSI address, and a method to receive vop interrupts from the host.  This
> > information is used to implement the vop callbacks allowing the vop
> > framework to access to the MIC device page and send/receive interrupts
> > from/to the host.
> 
> Ok, this seems fine so far. So the vop-host-backend is a regular PCI
> driver that implements the VOP protocol from the host side, and it
> can talk to either a MIC, or another guest-backend written for the PCI-EP
> framework to implement the same protocol, right?

Yes, but just to clarify:  the placement of the device page and the way
to communicate the location of the device page address and any other
information needed by the guest-backend are hardware-specific so there
is no generic vop-host-backend implementation which can talk to both a
MIC and to something else.

> > vop (despite its name) doesn't care about PCIe.  The vop-guest-backend
> > doesn't actually need to talk to the PCIe endpoint driver.  The
> > vop-guest-backend can be probed via any means, such as via a device tree
> > on the endpoint.
> >
> > On the RC, userspace opens the vop device and adds the virtio devices,
> > which end up in the MIC device page set up by the vop-host-backend.
> >
> > On the endpoint, when the vop framework (via the vop-guest-backend) sees
> > these devices, it registers devices on the virtio bus and the virtio
> > drivers are probed.
> 
> Ah, so the direction is fixed, and it's the opposite of what Christoph
> and I were expecting. This is probably something we need to discuss
> a bit. From what I understand, there is no technical requirement why
> it has to be this direction, right?

I don't think the vop framework itself has any such requirement.

The MIC uses it in this way (see Documentation/mic/mic_overview.txt) and
it also makes sense (to me, at least) if one wants to treat the endpoint
like one would treat a virtualized guest.

> What I mean is that the same vop framework could work with
> a PCI-EP driver implementing the vop-host-backend and
> a PCI driver implementing the vop-guest-backend? In order
> to do this, the PCI-EP configuration would need to pick whether
> it wants the EP to be the vop host or guest, but having more
> flexibility in it (letting each side add virtio devices) would be
> harder to do.

Correct, this is my understanding also.

> > On the RC, userspace implements the device end of the virtio
> > communication in userspace, using the MIC_VIRTIO_COPY_DESC ioctl.  I
> > also have patches to support vhost.
> 
> This is a part I don't understand yet. Does this mean that the
> normal operation is between a user space process on the vop-host
> talking to the kernel on the vop-guest?

Yes.  For example, the guest mounts a 9p filesystem with virtio-9p and
the 9p server is implemented in a userspace process on the host.  This
is again similar to virtualization.

> I'm a bit worried about the ioctl interface here, as this combines the
> configuration side with the actual data transfer, and that seems
> a bit inflexible.
>
> > > Is there always one master side that is responsible for creating
> > > virtio devices on it, with the slave side automatically attaching to
> > > them, or can either side create virtio devices?
> >
> > Only the master can create virtio devices.  The virtio drivers run on
> > the slave.
> 
> Ok.
> 
> > > Is there any limit on
> > > the number of virtio devices or queues within a VOP device?
> >
> > The virtio device information (mic_device_desc) is put into the MIC
> > device page whose size is limited by the ABI header in
> > include/uapi/linux/mic_ioctl.h (MIC_DP_SIZE, 4096 bytes).  So the number
> > of devices is limited by the limit of the number of device descriptors
> > that can fit in that size.  There is also a per-device limit on the
> > number of vrings (MIC_VRING_ENTRIES) and vring entries
> > (MIC_VRING_ENTRIES) in the ABI header.
> 
> Ok, so you can have multiple virtio devices (e.g. a virtio-net and
> virtio-console) but not an arbitrary number? I suppose we can always
> extend it later if that becomes a problem.

Yes.