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

[Michael S. Tsirkin]

On Tue, May 16, 2023 at 12:17:50PM +0800, Jason Wang wrote:
> On Tue, May 16, 2023 at 12:13 PM Michael S. Tsirkin <mst@xxxxxxxxxx> wrote:
> >
> > On Tue, May 16, 2023 at 10:44:45AM +0800, Jason Wang wrote:
> > > 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.
> >
> > So let's start by addressing VDUSE maybe?
> 
> It's reported by at least one hardware vendor as well. I remember it
> was Alvaro who reported this first in the past.
> 
> >
> > > > 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
> >
> > It's not reliable for other drivers but has been reliable for virtio.
> > I worry some software relied on this.
> 
> It's probably fine since some device like vhost doesn't support this
> at all and we manage to survive for several years.

vhost is often connected to a clever learning backend
such as a bridge which will DTRT without guest configuring
anything at all though, this could be why it works.



> > You are making good points though ... could we get some
> > maintainer's feedback on this?
> 
> That would be helpful. Jakub, any input on this?
> 
> Thanks
> 
> >
> > > >
> > > > > >
> > > > > >
> > > > > > > > --
> > > > > > > > MST
> > > > > > > >
> > > > > >
> > > >
> >