Re: [Qemu-devel] [PATCH v3 2/2] net: introduce command to query rx-filter information

[Amos Kong]

On Thu, Jul 04, 2013 at 08:28:59AM +0200, Markus Armbruster wrote:
> Amos Kong <address@hidden> writes:
> 
> > On Tue, Jul 02, 2013 at 03:27:12PM +0200, Markus Armbruster wrote:
> >> Amos Kong <address@hidden> writes:
> >> 
> >> > On Tue, Jul 02, 2013 at 11:05:56AM +0200, Markus Armbruster wrote:
> >> >> Amos Kong <address@hidden> writes:
> >> [...]
> >> >> >> This interface is abstract in the sense that it applies to all NICs. 
> >> >> >>  At
> >> >> >> this time, it's implemented only virtio-net implements it.  I'm
> >> >> >> habitually wary of abstractions based on just one concrete instance,
> >> >> >> which makes me ask:
> >> >> >> 
> >> >> >> 1. Ignorant question first: could the feature make sense for other 
> >> >> >> NICs,
> >> >> >> too, or is it specific to virtio-net?
> >> >> >
> >> >> > We will not. 
> >> >> >
> >> >> > It's ugly to check if nic is virtio-net nic in net/net.c, so I
> >> >> > register the query function to NetClientInfo. Traversal the net
> >> >> > client list in net/net.c, and execute query of each virtio-net
> >> >> > instance in virtio-net.c
> >> >> 
> >> >> Implementing the feature as an optional callback is fine.
> >> >> 
> >> >> Let me rephrase my question: could this feature be implemented for other
> >> >> NICs?  I'm *not* asking you to do that, just whether it would be
> >> >> possible.
> >> >> 
> >> >> I'm asking because my review of the QAPI schema depends on the answer.
> >> >> 
> >> >> >> 2. If the former, are you reasonably sure this object will do for 
> >> >> >> other
> >> >> >> NICs?
> >> >> >
> >> >> > No.
> >> >> 
> >> >> I'm not sure I understand you.  Do you mean to say that the feature
> >> >> could be implemented for other NICs, but RxFilterInfo would probably not
> >> >> fit for them?
> >> >
> >> > We will not implement the feature to other NICs, no request.
> >> >
> >> > We notify the management of virtio-net rx-filter change, because
> >> > we want to sync the the rx-filter change to macvtap device.
> >> 
> >> I understand there are no plans to implement this feature for other
> >> NICs.  But I'm not asking whether we *want* to implement it for other
> >> NICs, I'm asking whether we *could*.
> >  
> > In theory, we can.
> >
> >> Or rephrased yet another way: what exactly makes this feature applicable
> >> to virtio-net only?
> >
> > Macvtap can only be used by virtio-net, not other emulated nic.
> > It's meaningless for management to know the rx-filter change of
> > non-virtio-net NICs.
> 
> I'm having trouble squaring "in theory, we can" with "meaningless".  So
> I'm rephrasing my question yet again.
> 
> Do NICs other than virtio-net have rx-filters?
 
Yes.

Talked with Jason, upstream kernel fixed some bugs, then we can also
use e1000 with macvtap. In this case, we should also implement a
.query_rx_filter function for e1000. We can do it by another patchset.

> If yes, what have these NIC rx-filters in common, and how do they
> differ?
> 
> Why would anybody want to query rx-filters?  Use cases, please.

It's a way to tell management the rx-filter setup in guest nic.
Management query the rx-filter setup of guest, then change the
setup of macvtap device, macvtap uses same rx-filter setup as
virtual nic.

> Why is querying rx-filters "meaningless" for anything but virtio-net?
> The dictionary explains "meaningless" as "having no meaning; of no
> value".  Thus, for the query to be meaningless, the answer must carry no
> information, or at least none of value.  Is querying rx-filters really
> meaningless?  Or is it just something we don't need right now, and can't
> see being needed in the future?

-- 
                        Amos.