Re: [Qemu-devel] [PATCH RFC v2 3/5] net/filter: Introduce a helper to add a filter to the netdev

[Hailiang Zhang]

On 2016/2/1 17:18, Jason Wang wrote:
> On 02/01/2016 04:21 PM, Hailiang Zhang wrote:
> > > > > > > Instead of this, I wonder maybe it's better to:
> > > > > > >
> > > > > > > - store the default filter property into a pointer to string
> > > > > >
> > > > > > Do you mean, pass a string parameter which stores the filter property
> > > > > > instead of
> > > > > > assemble it in this helper ?
> > > > >
> > > > > Yes. E.g just a global string which could be changed by any subsystem.
> > > > > E.g colo may change it to
> > > > > "filter-buffer,interval=0,status=disable". But
> > > > > filter ids need to be generated automatically.
> > > >
> > > > Got it. Then we don't need the global default_netfilter_type[] in
> > > > patch 5,
> > > > Just use this global string instead ?
> > > >
> > > > > > > - colo code may change the pointer to "filter-buffer,status=disable"
> > > > > >
> > > > > > > Then, there's no need for lots of codes above:
> > > > > > > - no need a "is_default" parameter in netdev_add_filter which
> > > > > > > does not
> > > > > > > scale consider we may want to have more property in the future
> > > > > > > - no need to hacking like "qemu_filter_opts"
> > > > > >
> > > > > > Yes, we can use qemu_find_opts("object") instead of it.
> > > > > >
> > > > > > > - no need to have a special flag like "is_default"
> > > > > >
> > > > > > But we have to distinguish the default filter from the common
> > > > > > filter, use the name (id) to distinguish it ?
> > > > >
> > > > > What's the reason that you want to distinguish default filters from
> > > > > others?
> > > >
> > > > The default filters will be used by COLO or MC, (In COLO, we will
> > > > use it
> > > > to control packets buffering/releasing).
> > > > For COLO, we don't want to control (use) other filters that added by
> > > > users.
> > >
> > > I think Jason's point is that COLO is a manager, you can add the filter
> > > to netdev when doing COLO, so the only difference between COLO's default
> >
> > Er, then we came back to the original question, 'is it necessary to
> > add each netdev
> > a default filter ?'
>
> The question could be extended to:
>
> 1) Do we need a default filter? I think the answer is yes, but of course
> COLO can work even without this.

Yes, after colo-proxy is realized, we can switch to colo-proxy
(It should have the capability of buffer and release packets directly).
But for now, we want to merge COLO prototype without colo-proxy, the COLO
prototype should have the basic capability. Just like Remus or
Micro-checkpointing. It is based on the default buffer-filter to control net
packets.

> 2) Do we want to implement COLO on top of default filter? If yes, as you
> suggest, we may record the ids of the default filter and do what ever we

Yes, we need it.

> what. If not, COLO need codes to go through each netdev and add filter
> itself (hotplug is not supported). Or you want management to do this,
> then even hotplug could be supported.

We also want to support hotplug during VM is in COLO state in the future.
(For this point, I'm not quite sure if this usage case is really exist.)

Thanks,
Hailiang

> Any thoughts?
>
> > If we add the a filter to netdev when doing COLO, it will be added
> > dynamically,
> > Here we want to add each netdev a default filter while launch QEMU
> > (no matter if this VM will go into COLO or not),
> > just to support hot-add NIC for VM while in COLO lifetime.
>
> Yes.
>
>
>
>
> .