Re: [virtio-comment] Re: [PATCH v18] virtio-net: support inner header hash

[Heng Qi <hengqi@linux.alibaba.com>]

å 2023/6/22 äå8:32, Parav Pandit åé:
> > From: Michael S. Tsirkin <mst@redhat.com>
> > Sent: Thursday, June 22, 2023 2:23 AM
> >
> > On Wed, Jun 21, 2023 at 08:52:04PM +0000, Parav Pandit wrote:
> > > > From: Michael S. Tsirkin <mst@redhat.com>
> > > > Sent: Wednesday, June 21, 2023 4:38 PM
> > > > > > And the field is RO so no memory cost to exposing it in all VFs.
> > > > > Two structures do not bring the asymmetry.
> > > > > Accessing current and enabled fields via two different mechanism
> > > > > is bringing
> > > > the asymmetry.
> > > >
> > > > I guess it's a matter of taste, but it is clearly more consistent
> > > > with other hash things, to which it's very similar.
> > > This is consistent with new commands we define including notification
> > coalescing whose GET is not coming config space.
> >
> > But there GET just reports the current state. Not the read only capability. So
> > there would be cost per VF to keep it in config space.
> > This one is RO no cost per VF. Let's make it convenient?
> And each VF can have different value hence requires per VF storage in the device.
>   
> > > > Nah, config space is too convenient when we can live with its
> > > > limitations. I don't thin kwe prefer not to keep growing it.
> > > > For some things such as this one it's perfect.
> > > Fields are different between different devices.
> > Not sure what's the implication?
> Implication is device needs to store this in always available on-chip memory which is not good.
>
> > > > For example, for migration driver might want to validate that two
> > > > devices have same capability. doing it without dma is nicer.
> > > A migration driver for real world scenario, will almost have to use the dma for
> > amount of data it needs to exchange.
> >
> > Not migration itself, provisioning.
> Provisioning driver usually do not attach to the member device directly.
> This requires device reset, followed by reaching _DRIVER stage, querying features etc and config area.
> And unbinding it and second reset by member driver. Ugh.
>
> Provisioning driver also needs to get the state or capabilities even when member driver is already attached.
> So config space is not much a gain either.
>   
> > > > Another example, future admin transport will have ability to
> > > > provision devices by supplying their config space.
> > > > This will include this capability automatically, if instead we hide
> > > > it in a command we need to do extra custom work.
> > > >
> > > > > So we do not prefer to keep growing the config space anymore,
> > > > > hence GET is the right approach to me.
> > > > Heh I know you hate config space. Let it go, stop wasting time
> > > > arguing about the same thing on every turn and instead help define
> > > > admin transport to solve it
> > > This was discussed many times, a driver to have a direct (non-intercepted by
> > owner device) channel to device.
> > > If you mean this non-intercepted channel as admin transport, fine.
> > we can do that, sure.
> >
> > > If you mean this is intercepted and it is going over admin cmd, then it is of no
> > use for all future interfaces.
> > > We discussed this in thread with you and Jason.
> > > I provided concrete example with size and device provisioning math too and
> > other example of multi-physical address VQ.
> > > So transporting register by register over some admin transport is sub-optimal.
> > Not register by register, we can send all of config space as long as it's RO. This
> > field is.
> It is RO in context of one member device, but every VF can have different value.
> The device will never know if one will use new cmdvq to access or some old driver will use without it.
> And hence, it always needs to provision it on onchip memory for backward compatibility.

Yes, I think we also have to consider upcoming
ÂÂÂ 1. device counters (e.g. supported_device_counter),
ÂÂÂ 2. receive flow filters (e.g. supported_flow_types, 
supported_max_entries),
ÂÂÂ 3. header splits (e.g. supported_split_types) etc.
Continuous expansion of the configuration space needs to be careful.

> Instead of decision point being RO vs RW,
> any new fields via cmdvq and existing fields stays in cfg space, give predictable behavior to size the member devices in the system.
> Once the cmdvq is available, we can get rid of GET command used in this version for new future features.
> Till that arrives, GET command is the efficient way.

Yes, I agree.

Thanks.