[dpdk-dev] [PATCH 00/28] vhost: add virtio-vhost-user transport

[Maxime Coquelin]

Hi Nikos,

On 6/19/19 5:14 PM, Nikos Dragazis wrote:
> Hi everyone,
> 
> this patch series introduces the concept of the virtio-vhost-user
> transport. This is actually a revised version of an earlier RFC
> implementation that has been proposed by Stefan Hajnoczi [1]. Though
> this is a great feature, it seems to have been stalled, so I’d like to
> restart the conversation on this and hopefully get it merged with your
> help. Let me give you an overview.

Thanks for taking over the series!

I think you are already aware of that, but it arrives too late to
consider it for v19.08, as the proposal deadline is over by almost 3
weeks.

That said, it is good that you sent it early, so that we can work to
make it in for v19.11.

> The virtio-vhost-user transport is a vhost-user transport implementation
> that is based on the virtio-vhost-user device. Its key difference with
> the existing transport is that it allows deploying vhost-user targets
> inside dedicated Storage Appliance VMs instead of host user space. In
> other words, it allows having guests that act as vhost-user backends for
> other guests.
> 
> The virtio-vhost-user device implements the vhost-user control plane
> (master-slave communication) as follows:
> 
> 1. it parses the vhost-user messages from the vhost-user unix domain
>     socket and forwards them to the slave guest through virtqueues
> 
> 2. it maps the vhost memory regions in QEMU’s process address space and
>     exposes them to the slave guest as a RAM-backed PCI MMIO region
> 
> 3. it hooks up doorbells to the callfds. The slave guest can use these
>     doorbells to interrupt the master guest driver
> 
> The device code has not yet been merged into upstream QEMU, but this is
> definitely the end goal.

Could you provide a pointer to the QEMU series, and instructions to test
this new device?

> The current state is that we are awaiting for
> the approval of the virtio spec.

Ditto, a link to the spec patches would be useful.

> I have Cced Darek from the SPDK community who has helped me a lot by
> reviewing this series. Note that any device type could be implemented
> over this new transport. So, adding the virtio-vhost-user transport in
> DPDK would allow using it from SPDK as well.
> 
> Getting into the code internals, this patch series makes the following
> changes:
> 
> 1. introduce a generic interface for the transport-specific operations.
>     Each of the two available transports, the pre-existing AF_UNIX
>     transport and the virtio-vhost-user transport, is going to implement
>     this interface. The AF_UNIX-specific code has been extracted from the
>     core vhost-user code and is now part of the AF_UNIX transport
>     implementation in trans_af_unix.c.
> 
> 2. introduce the virtio-vhost-user transport. The virtio-vhost-user
>     transport requires a driver for the virtio-vhost-user devices. The
>     driver along with the transport implementation have been packed into
>     a separate library in `drivers/virtio_vhost_user/`. The necessary
>     virtio-pci code has been copied from `drivers/net/virtio/`. Some
>     additional changes have been made so that the driver can utilize the
>     additional resources of the virtio-vhost-user device.
> 
> 3. update librte_vhost public API to enable choosing transport for each
>     new vhost device. Extend the vhost net driver and vhost-scsi example
>     application to export this new API to the end user.
> 
> The primary changes I did to Stefan’s RFC implementation are the
> following:
> 
> 1. moved postcopy live migration code into trans_af_unix.c. Postcopy
>     live migration relies on the userfault fd mechanism, which cannot be
>     supported by virtio-vhost-user.
> 
> 2. moved setup of the log memory region into trans_af_unix.c. Setting up
>     the log memory region involves mapping/unmapping guest memory. This
>     is an AF_UNIX transport-specific operation.
> 
> 3. introduced a vhost transport operation for
>     process_slave_message_reply()
> 
> 4. moved the virtio-vhost-user transport/driver into a separate library
>     in `drivers/virtio_vhost_user/`. This required making vhost.h and
>     vhost_user.h part of librte_vhost public API and exporting some
>     private symbols via the version script. This looks better to me that
>     just moving the entire librte_vhost into `drivers/`. I am not sure if
>     this is the most appropriate solution. I am looking forward to your
>     suggestions on this.

I'm not sure this is the right place to put it.

> 5. made use of the virtio PCI capabilities for the additional device
>     resources (doorbells, shared memory). This required changes in
>     virtio_pci.c and trans_virtio_vhost_user.c.
> 
> 6. [minor] changed some commit headlines to comply with
>     check-git-log.sh.
> 
> Please, have a look and let me know about your thoughts. Any
> reviews/pointers/suggestions are welcome.

Maxime