Re: [PATCH net-next V2 1/2] virtio-net: convert rx mode setting to use workqueue

[Jason Wang]

On Mon, May 15, 2023 at 6:17 PM Michael S. Tsirkin <mst@xxxxxxxxxx> wrote:
>
> On Mon, May 15, 2023 at 01:13:33PM +0800, Jason Wang wrote:
> > On Mon, May 15, 2023 at 12:45 PM Michael S. Tsirkin <mst@xxxxxxxxxx> wrote:
> > >
> > > On Mon, May 15, 2023 at 09:05:54AM +0800, Jason Wang wrote:
> > > > On Wed, May 10, 2023 at 1:33 PM Michael S. Tsirkin <mst@xxxxxxxxxx> wrote:
> > > > >
> > > > > On Mon, Apr 17, 2023 at 11:40:58AM +0800, Jason Wang wrote:
> > > > > > On Fri, Apr 14, 2023 at 3:21 PM Michael S. Tsirkin <mst@xxxxxxxxxx> wrote:
> > > > > > >
> > > > > > > On Fri, Apr 14, 2023 at 01:04:15PM +0800, Jason Wang wrote:
> > > > > > > > Forget to cc netdev, adding.
> > > > > > > >
> > > > > > > > On Fri, Apr 14, 2023 at 12:25 AM Michael S. Tsirkin <mst@xxxxxxxxxx> wrote:
> > > > > > > > >
> > > > > > > > > On Thu, Apr 13, 2023 at 02:40:26PM +0800, Jason Wang wrote:
> > > > > > > > > > This patch convert rx mode setting to be done in a workqueue, this is
> > > > > > > > > > a must for allow to sleep when waiting for the cvq command to
> > > > > > > > > > response since current code is executed under addr spin lock.
> > > > > > > > > >
> > > > > > > > > > Signed-off-by: Jason Wang <jasowang@xxxxxxxxxx>
> > > > > > > > >
> > > > > > > > > I don't like this frankly. This means that setting RX mode which would
> > > > > > > > > previously be reliable, now becomes unreliable.
> > > > > > > >
> > > > > > > > It is "unreliable" by design:
> > > > > > > >
> > > > > > > >       void                    (*ndo_set_rx_mode)(struct net_device *dev);
> > > > > > > >
> > > > > > > > > - first of all configuration is no longer immediate
> > > > > > > >
> > > > > > > > Is immediate a hard requirement? I can see a workqueue is used at least:
> > > > > > > >
> > > > > > > > mlx5e, ipoib, efx, ...
> > > > > > > >
> > > > > > > > >   and there is no way for driver to find out when
> > > > > > > > >   it actually took effect
> > > > > > > >
> > > > > > > > But we know rx mode is best effort e.g it doesn't support vhost and we
> > > > > > > > survive from this for years.
> > > > > > > >
> > > > > > > > > - second, if device fails command, this is also not
> > > > > > > > >   propagated to driver, again no way for driver to find out
> > > > > > > > >
> > > > > > > > > VDUSE needs to be fixed to do tricks to fix this
> > > > > > > > > without breaking normal drivers.
> > > > > > > >
> > > > > > > > It's not specific to VDUSE. For example, when using virtio-net in the
> > > > > > > > UP environment with any software cvq (like mlx5 via vDPA or cma
> > > > > > > > transport).
> > > > > > > >
> > > > > > > > Thanks
> > > > > > >
> > > > > > > Hmm. Can we differentiate between these use-cases?
> > > > > >
> > > > > > It doesn't look easy since we are drivers for virtio bus. Underlayer
> > > > > > details were hidden from virtio-net.
> > > > > >
> > > > > > Or do you have any ideas on this?
> > > > > >
> > > > > > Thanks
> > > > >
> > > > > I don't know, pass some kind of flag in struct virtqueue?
> > > > >         "bool slow; /* This vq can be very slow sometimes. Don't wait for it! */"
> > > > >
> > > > > ?
> > > > >
> > > >
> > > > So if it's slow, sleep, otherwise poll?
> > > >
> > > > I feel setting this flag might be tricky, since the driver doesn't
> > > > know whether or not it's really slow. E.g smartNIC vendor may allow
> > > > virtio-net emulation over PCI.
> > > >
> > > > Thanks
> > >
> > > driver will have the choice, depending on whether
> > > vq is deterministic or not.
> >
> > Ok, but the problem is, such booleans are only useful for virtio ring
> > codes. But in this case, virtio-net knows what to do for cvq. So I'm
> > not sure who the user is.
> >
> > Thanks
>
> Circling back, what exactly does the architecture you are trying
> to fix look like? Who is going to introduce unbounded latency?
> The hypervisor?

Hypervisor is one of the possible reason, we have many more:

Hardware device that provides virtio-pci emulation.
Userspace devices like VDUSE.

> If so do we not maybe want a new feature bit
> that documents this? Hypervisor then can detect old guests
> that spin and decide what to do, e.g. prioritise cvq more,
> or fail FEATURES_OK.

We suffer from this for bare metal as well.

But a question is what's wrong with the approach that is used in this
patch? I've answered that set_rx_mode is not reliable, so it should be
fine to use workqueue. Except for this, any other thing that worries
you?

Thanks

>
> > >
> > >
> > > > > --
> > > > > MST
> > > > >
> > >
>