Re: [Virtio-fs] [PATCH v3 0/5] vhost-user: Back-end state migration

[Hanna Czenczek]

On 26.09.23 21:20, Stefan Hajnoczi wrote:

On Tue, Sep 26, 2023, 09:33 Hanna Czenczek <hreitz@redhat.com> wrote:

On 25.09.23 22:48, Stefan Hajnoczi wrote:
> On Fri, Sep 15, 2023 at 12:25:25PM +0200, Hanna Czenczek wrote:
>> RFC:
>> https://lists.nongnu.org/archive/html/qemu-devel/2023-03/msg04263.html
>>
>> v1:
>> https://lists.nongnu.org/archive/html/qemu-devel/2023-04/msg01575.html
>>
>> v2:
>> https://lists.nongnu.org/archive/html/qemu-devel/2023-07/msg02604.html
>>
>> Hi,
>>
>> I’ve decided not to work on vhost-user SUSPEND/RESUME for now – it is
>> not technically required for virtio-fs migration, which is the actual
>> priority for me now.  While we do want to have SUSPEND/RESUME at some
>> point, the only practically existing reason for it is to be able to
>> implement vhost-level resetting in virtiofsd, but that is not related to
>> migration.
> QEMU sends VHOST_USER_SET_STATUS 0 in vhost_dev_stop(). Are you assuming
> that virtiofs back-ends do not reset the device upon receiving this
> message?

Absolutely.  vhost_dev_stop() is not in the migration-specific path, but
is called whenever the vCPUs are stopped, which indeed happens to be
part of migration, but is also used in other cases, like QMP stop.  We
have identified that it is wrong that we reset the back-end just because
the vCPUs are stopped (e.g. on migration), but it is what we do right
now when the VM is paused (e.g. through QMP stop).

Therefore, stateful back-ends cannot implement reset lest stop/cont
breaks the device.  I don’t think anybody really cares whether a
vhost-user back-end actually resets its internal state (if there is any)
when the guest driver asks for a reset on the virtio level, as long as
the guest driver is then able to initialize the device afterwards.

Some devices send configuration change notifications. For example, virtio-net speed and duplex changes.

Imagine a network boot ROM runs and the firmware resets the virtio-net device when transferring control to the loaded kernel. Before the kernel driver initializes the device again, the vhost-user-net back-end reports a speed or duplex change and sends a Configuration Change Notification to the guest. The guest receives a spurious interrupt because the vhost-user-net device wasn't actually reset.

    I don’t see how this relates to my argument that no stateful
    back-end can implement a full reset because doing so would break
    stop/cont.
    
    If vhost-user-net were stateful (which it isn’t, AFAIK), it could
    choose to implement a work-around such that it would stop sending
    notifications on reset, but not reset its internal state.  Then,
    when qemu restores vring state in vhost_dev_start(), it would resume
    sending notifications.  But again, I fail to see how this is not
    already an issue for stop/cont.

I'm concerned that ignoring reset matters (admittedly in corner cases) and think that stateful device functionality shouldn't be added to the vhost-user protocol without a solution for reset.

    I disagree.  We have a stateful device already, whether we add
    functionality acknowledging this to the protocol or not.  The
    problem exists.  It is independent of migration.  If there’s a
    problem because of this with migration, there’s a problem with
    stop/cont, too, that must already have been worked around.

This patch series changes the vhost-user protocol, which is used by many different devices, not just virtiofs. The solution should work for vhost-user devices of any type and not be based on what we can get away with when running the current QEMU + virtiofsd.

    My argument was generic.  Any existing stateful device
    implementation must implement reset in such a way that it won’t
    break stop/cont, i.e., it must not reset its internal state.

I do
think people care that stop/cont works, so it follows to me that no
stateful back-end will reset its internal state when receiving a
virtio/vhost reset.  If they do, stop/cont breaks, which is a
user-visible bug that needs to be addressed – either properly by
implementing SUSPEND/RESUME in both qemu and the affected back-ends, or
by using a similar work-around to virtiofsd, which is to ignore reset
commands.

Properly, please.

    You misunderstand me.  I’m not presenting the choice I have.  I’m
    presenting the choices existing implementations *have had until this
    point*.  *None* chose to do it properly.  I don’t know of stateful
    implementations besides virtiofsd, but virtiofsd chose to be content
    with not implementing reset and thus having things “just work”.
    
    The work-arounds must exist already.

It’s hard for me to imagine that people don’t care that stop/cont breaks
some vhost-user back-end, but suddenly would start to care that
migration doesn’t work – especially given that first of all someone will
need to manually implement any migration support in that back-end even
with this series, which means that really, the only back-end we are
talking about here is our virtiofsd.  To this day I’m not even aware of
any other back-end that has internal state.

Another one I can think of is vhost-user-gpu.

    I sure hope stop/cont works for them.

Why did you give up on implementing SUSPEND/RESUME?

    Because I think it’s unnecessary for implementing migration, and
    migration is what’s on my priority list.  None of these issues are
    new, they have always existed with stop/cont, work-arounds must be
    in place to make stop/cont work, and because I don’t see the
    difference in how stop/cont is used outside of migration and how it
    is used during migration, I assume those work-arounds must work for
    migration as well.
    
    Implementing virtio-fs migration is a chain of dependencies.  We
    need at least the specification to be in qemu before we can start
    sending merge requests to the vhost Rust crates to implement support
    there.  We need that support there before we can make the changes to
    virtiofsd.
    
    Adding SUSPEND/RESUME adds another hard dependency to the whole
    discussion (this would have to go on before this series), which has
    proven absolutely clearly in the past months that it is a very
    complex finnicky isue that would take a ton of time still.  And I
    can’t justify that for myself, given that I don’t see any
    practically existing problem.
    
    PS: As far as I remember, vhost-user doesn’t even have a working
    reset today.  vhost_dev_stop() calls vhost_reset_status(), which is
    a no-op unless the back-end supports SET_STATUS.  The only back-end
    implementation we found (while discussing SUSPEND/RESUME) to support
    SET_STATUS was dpdk, but while it logs SET_STATUS 0 as a reset, it
    doesn’t do a reset, i.e. doesn’t call reset_device(), which it would
    do on RESET_OWNER.
    
    Hanna