On Fri, 2021-06-25 at 10:55 +0800, Jason Wang wrote:
> 在 2021/6/24 下午6:42, David Woodhouse 写道:
> > On Thu, 2021-06-24 at 14:12 +0800, Jason Wang wrote:
> > > 在 2021/6/24 上午12:12, David Woodhouse 写道:
> > > > We *should* eventually expand this test case to attach an AF_PACKET
> > > > device to the vhost-net, instead of using a tun device as the back end.
> > > > (Although I don't really see *why* vhost is limited to AF_PACKET. Why
> > > > *can't* I attach anything else, like an AF_UNIX socket, to vhost-net?)
> > > 
> > > It's just because nobody wrote the code. And we're lacking the real use
> > > case.
> > 
> > Hm, what code?
> 
> 
> The codes to support AF_UNIX.
> 
> 
> >   For AF_PACKET I haven't actually spotted that there *is* any.
> 
> 
> Vhost_net has this support for more than 10 years. It's hard to say 
> there's no user for that.
> 

I wasn't saying I hadn't spotted the use case. I hadn't spotted the
*code* which is in af_packet to support vhost. But...

> > As I've been refactoring the interaction between vhost and tun/tap, and
> > fixing it up for different vhdr lengths, PI, and (just now) frowning in
> > horror at the concept that tun and vhost can have *different*
> > endiannesses, I hadn't spotted that there was anything special on the
> > packet socket.
> 
> Vnet header support.

... I have no idea how I failed to spot that. OK, so AF_PACKET sockets
can *optionally* support the case where *they* provide the
virtio_net_hdr — instead of vhost doing it, or there being none.

But any other sockets would work for the "vhost does it" or the "no
vhdr" case.

... and I need to fix my 'get sock_hlen from the underlying tun/tap
device' patch to *not* assume that sock_hlen is zero for a raw socket;
it needs to check the PACKET_VNET_HDR sockopt. And *that* was broken
for the VERSION_1|MRG_RXBUF case before I came along, wasn't it?
Because vhost would have assumed sock_hlen to be 12 bytes, while in
AF_PACKET it's always only 10?

> >   For that case, sock_hlen is just zero and we
> > send/receive plain packets... or so I thought? Did I miss something?
> 
> 
> With vnet header, it can have GSO and csum offload.
> 
> 
> > 
> > As far as I was aware, that ought to have worked with any datagram
> > socket. I was pondering not just AF_UNIX but also UDP (since that's my
> > main transport for VPN data, at least in the case where I care about
> > performance).
> 
> 
> My understanding is that vhost_net designed for accelerating virtio 
> datapath which is mainly used for VM (L2 traffic). So all kinds of TAPs 
> (tuntap,macvtap or packet socket) are the main users. If you check git 
> history, vhost can only be enabled without KVM until sometime last year. 
> So I confess it can serve as a more general use case, and we had already 
> has some discussions. But it's hard to say it's worth to do that since 
> it became a re-invention of io_uring?

Yeah, ultimately I'm not sure that's worth exploring. As I said, I was
looking for something that works on *current* kernels. Which means no
io_uring on the underlying tun socket, and no vhost on UDP. If I want
to go and implement *both* ring protocols in userspace and make use of
each of them on the socket that they do support, I can do that. Yay! :)

If I'm going to require new kernels, then I should just work on the
"ideal" data path which doesn't really involve userspace at all. But we
should probably take that discussion to a separate thread.