On Wed, Mar 17, 2021 at 6:05 PM Stefan Hajnoczi <stefanha@redhat.com> wrote:

> On Fri, Dec 18, 2020 at 05:39:34PM +0800, Jiachen Zhang wrote:
> > Thanks for the suggestions. Actually, we choose to save all state
> > information to QEMU because a virtiofsd has the same lifecycle as its
> > QEMU master. However, saving things to a file do avoid communication with
> > QEMU, and we no longer need to increase the complexity of vhost-user
> > protocol. The suggestion to save fds to the systemd is also very
> reasonable
> > if we don't consider the lifecycle issues, we will try it.
>
> Hi,
> We recently discussed crash recovery in the virtio-fs bi-weekly call and
> I read some of this email thread because it's a topic I'm interested in.
>
> I agree with Marc-André that storing file descriptors does not need to
> be in the vhost-user protocol. The lifetime of a vhost-user device
> backend is not controlled by the VMM since the device backend is
> launched separately. Therefore it's reasonable for the component that
> launched the device backend to also have the responsibility of cleaning
> up the vhost-user device backend.
>
> Using the sd_notify(3) interface is a neat idea. It's supported natively
> by systemd but you can also implement a compatible interface in your own
> software. This way the vhost-user device backend can be launched using
> systemd or your own software.
>
> That said, if people find it more convenient to store fds using the
> vhost-user protocol, then I think that is enough justification to add a
> generic message to the vhost-user protocol. The important thing is to
> make the message generic so it solves all crash recovery use cases. The
> inflight fd messages were too specific and now we're having to think
> about adding more messages again.
>
> Stefan
>


Hi, Stefan,

I agreed with you that a virtiofsd must be launched by a software like
systemd. So we are planning to define more generic persist/restore
interfaces (callbacks). Then anyone can implement their own persist/restore
callbacks to store states to proper places.  And I think in the next
version we will implement default callbacks for the interfaces. Instead of
vhost-user messages, systemd's sd_notify(3) will be the default method for
storing fds, and several tmpfs files can be the default place to store the
shm regions.

Jiachen