Re: [PATCH v9] virtio-net: support inner header hash

["Michael S. Tsirkin" <mst@redhat.com>]

On Tue, Feb 28, 2023 at 11:04:26AM +0800, Jason Wang wrote:
> On Tue, Feb 28, 2023 at 1:49 AM Michael S. Tsirkin <mst@redhat.com> wrote:
> >
> > On Mon, Feb 27, 2023 at 04:35:09PM +0800, Jason Wang wrote:
> > > On Mon, Feb 27, 2023 at 3:39 PM Michael S. Tsirkin <mst@redhat.com> wrote:
> > > >
> > > > On Mon, Feb 27, 2023 at 12:07:17PM +0800, Jason Wang wrote:
> > > > > Btw, this kind of 1:1 hash features seems not scalable and flexible.
> > > > > It requires an endless extension on bits/fields. Modern NICs allow the
> > > > > user to customize the hash calculation, for virtio-net we can allow to
> > > > > use eBPF program to classify the packets. It seems to be more flexible
> > > > > and scalable and there's almost no maintain burden in the spec (only
> > > > > bytecode is required, no need any fancy features/interactions like
> > > > > maps), easy to be migrated etc.
> > > > >
> > > > > Prototype is also easy, tun/tap had an eBPF classifier for years.
> > > > >
> > > > > Thanks
> > > >
> > > > Yea BPF offload would be great to have. We have been discussing it for
> > > > years though - security issues keep blocking it. *Maybe* it's finally
> > > > going to be there but I'm not going to block this work waiting for BPF
> > > > offload.  And easily migrated is what BPF is not.
> > >
> > > Just to make sure we're at the same page. I meant to find a way to
> > > allow the driver/user to fully customize what it wants to
> > > hash/classify. Similar technologies which is based on private solution
> > > has been used by some vendors, which allow user to customize the
> > > classifier[1]
> > >
> > > ePBF looks like a good open-source solution candidate for this (there
> > > could be others). But there could be many kinds of eBPF programs that
> > > could be offloaded. One famous one is XDP which requires many features
> > > other than the bytecode/VM like map access, tailcall. Starting from
> > > such a complicated type is hard. Instead, we can start from a simple
> > > type, that is the eBPF classifier. All it needs is to pass the
> > > bytecode to the device, the device can choose to run it or compile it
> > > to what it can understand for classifying. We don't need maps, tail
> > > calls and other features.
> >
> > Until people start asking exactly for maps because they want
> > state for their classifier?
> 
> Yes, but let's compare the eBPF without maps with the static feature
> proposed here. It is much more scalable and flexible.

I looked for some examples of RSS using BPF and only found this:
https://github.com/Netronome/bpf-samples/blob/master/programmable_rss/rss_user.c
seems to use maps.


> > And it makes sense - if you want
> > e.g. load balancing you need stats which needs maps.
> 
> Yes, but we know it's possible to have that (through the XDP offload).

Not without a lot more work to make xdp offload happen.

> This is impossible with the approach proposed here.
> 
> >
> > > We don't need to worry about the security
> > > because of its simplicity: the eBPF program is only in charge of doing
> > > classification, no other interactions with the driver and packet
> > > modification is prohibited. The feature is limited only to the
> > > VM/bytecode abstraction itself.
> > >
> > > What's more, it's a good first step to achieve full eBPF offloading in
> > > the future.
> > >
> > > Thanks
> > >
> > > [1] https://www.intel.com/content/www/us/en/architecture-and-technology/ethernet/dynamic-device-personalization-brief.html
> >
> > Dave seems to have nacked this approach, no?
> 
> I may miss something but looking at kernel commit, there are few
> patches to support that:
> 
> E.g
> 
> commit c7648810961682b9388be2dd041df06915647445
> Author: Tony Nguyen <anthony.l.nguyen@intel.com>
> Date:   Mon Sep 9 06:47:44 2019 -0700
> 
>     ice: Implement Dynamic Device Personalization (DDP) download
> 
> And it has been used by DPDK drivers.
> 
> Thanks
> 
> >
> > > >
> > > > --
> > > > MST
> > > >
> >