Message-ID: <20190715134115-mutt-send-email-mst@kernel.org>
Date:   Mon, 15 Jul 2019 13:50:28 -0400
From:   "Michael S. Tsirkin" <mst@...hat.com>
To:     Stefano Garzarella <sgarzare@...hat.com>
Cc:     Jason Wang <jasowang@...hat.com>,
        Stefan Hajnoczi <stefanha@...hat.com>,
        virtualization@...ts.linux-foundation.org, netdev@...r.kernel.org
Subject: Re: [RFC] virtio-net: share receive_*() and add_recvbuf_*() with
 virtio-vsock

On Mon, Jul 15, 2019 at 09:44:16AM +0200, Stefano Garzarella wrote:
> On Fri, Jul 12, 2019 at 06:14:39PM +0800, Jason Wang wrote:
> > 
> > On 2019/7/12 下午6:00, Stefano Garzarella wrote:
> > > On Thu, Jul 11, 2019 at 03:52:21PM -0400, Michael S. Tsirkin wrote:
> > > > On Thu, Jul 11, 2019 at 01:41:34PM +0200, Stefano Garzarella wrote:
> > > > > On Thu, Jul 11, 2019 at 03:37:00PM +0800, Jason Wang wrote:
> > > > > > On 2019/7/10 下午11:37, Stefano Garzarella wrote:
> > > > > > > Hi,
> > > > > > > as Jason suggested some months ago, I looked better at the virtio-net driver to
> > > > > > > understand if we can reuse some parts also in the virtio-vsock driver, since we
> > > > > > > have similar challenges (mergeable buffers, page allocation, small
> > > > > > > packets, etc.).
> > > > > > > 
> > > > > > > Initially, I would add the skbuff in the virtio-vsock in order to re-use
> > > > > > > receive_*() functions.
> > > > > > 
> > > > > > Yes, that will be a good step.
> > > > > > 
> > > > > Okay, I'll go on this way.
> > > > > 
> > > > > > > Then I would move receive_[small, big, mergeable]() and
> > > > > > > add_recvbuf_[small, big, mergeable]() outside of virtio-net driver, in order to
> > > > > > > call them also from virtio-vsock. I need to do some refactoring (e.g. leave the
> > > > > > > XDP part on the virtio-net driver), but I think it is feasible.
> > > > > > > 
> > > > > > > The idea is to create a virtio-skb.[h,c] where put these functions and a new
> > > > > > > object where stores some attributes needed (e.g. hdr_len ) and status (e.g.
> > > > > > > some fields of struct receive_queue).
> > > > > > 
> > > > > > My understanding is we could be more ambitious here. Do you see any blocker
> > > > > > for reusing virtio-net directly? It's better to reuse not only the functions
> > > > > > but also the logic like NAPI to avoid re-inventing something buggy and
> > > > > > duplicated.
> > > > > > 
> > > > > These are my concerns:
> > > > > - virtio-vsock is not a "net_device", so a lot of code related to
> > > > >    ethtool, net devices (MAC address, MTU, speed, VLAN, XDP, offloading) will be
> > > > >    not used by virtio-vsock.
> > 
> > 
> > Linux support device other than ethernet, so it should not be a problem.
> > 
> > 
> > > > > 
> > > > > - virtio-vsock has a different header. We can consider it as part of
> > > > >    virtio_net payload, but it precludes the compatibility with old hosts. This
> > > > >    was one of the major doubts that made me think about using only the
> > > > >    send/recv skbuff functions, that it shouldn't break the compatibility.
> > 
> > 
> > We can extend the current vnet header helper for it to work for vsock.
> 
> Okay, I'll do it.
> 
> > 
> > 
> > > > > 
> > > > > > > This is an idea of virtio-skb.h that
> > > > > > > I have in mind:
> > > > > > >       struct virtskb;
> > > > > > 
> > > > > > What fields do you want to store in virtskb? It looks to be exist sk_buff is
> > > > > > flexible enough to us?
> > > > > My idea is to store queues information, like struct receive_queue or
> > > > > struct send_queue, and some device attributes (e.g. hdr_len ).
> > 
> > 
> > If you reuse skb or virtnet_info, there is not necessary.
> > 
> 
> Okay.
> 
> > 
> > > > > 
> > > > > > 
> > > > > > >       struct sk_buff *virtskb_receive_small(struct virtskb *vs, ...);
> > > > > > >       struct sk_buff *virtskb_receive_big(struct virtskb *vs, ...);
> > > > > > >       struct sk_buff *virtskb_receive_mergeable(struct virtskb *vs, ...);
> > > > > > > 
> > > > > > >       int virtskb_add_recvbuf_small(struct virtskb*vs, ...);
> > > > > > >       int virtskb_add_recvbuf_big(struct virtskb *vs, ...);
> > > > > > >       int virtskb_add_recvbuf_mergeable(struct virtskb *vs, ...);
> > > > > > > 
> > > > > > > For the Guest->Host path it should be easier, so maybe I can add a
> > > > > > > "virtskb_send(struct virtskb *vs, struct sk_buff *skb)" with a part of the code
> > > > > > > of xmit_skb().
> > > > > > 
> > > > > > I may miss something, but I don't see any thing that prevents us from using
> > > > > > xmit_skb() directly.
> > > > > > 
> > > > > Yes, but my initial idea was to make it more parametric and not related to the
> > > > > virtio_net_hdr, so the 'hdr_len' could be a parameter and the
> > > > > 'num_buffers' should be handled by the caller.
> > > > > 
> > > > > > > Let me know if you have in mind better names or if I should put these function
> > > > > > > in another place.
> > > > > > > 
> > > > > > > I would like to leave the control part completely separate, so, for example,
> > > > > > > the two drivers will negotiate the features independently and they will call
> > > > > > > the right virtskb_receive_*() function based on the negotiation.
> > > > > > 
> > > > > > If it's one the issue of negotiation, we can simply change the
> > > > > > virtnet_probe() to deal with different devices.
> > > > > > 
> > > > > > 
> > > > > > > I already started to work on it, but before to do more steps and send an RFC
> > > > > > > patch, I would like to hear your opinion.
> > > > > > > Do you think that makes sense?
> > > > > > > Do you see any issue or a better solution?
> > > > > > 
> > > > > > I still think we need to seek a way of adding some codes on virtio-net.c
> > > > > > directly if there's no huge different in the processing of TX/RX. That would
> > > > > > save us a lot time.
> > > > > After the reading of the buffers from the virtqueue I think the process
> > > > > is slightly different, because virtio-net will interface with the network
> > > > > stack, while virtio-vsock will interface with the vsock-core (socket).
> > > > > So the virtio-vsock implements the following:
> > > > > - control flow mechanism to avoid to loose packets, informing the peer
> > > > >    about the amount of memory available in the receive queue using some
> > > > >    fields in the virtio_vsock_hdr
> > > > > - de-multiplexing parsing the virtio_vsock_hdr and choosing the right
> > > > >    socket depending on the port
> > > > > - socket state handling
> > 
> > 
> > I think it's just a branch, for ethernet, go for networking stack. otherwise
> > go for vsock core?
> > 
> 
> Yes, that should work.
> 
> So, I should refactor the functions that can be called also from the vsock
> core, in order to remove "struct net_device *dev" parameter.
> Maybe creating some wrappers for the network stack.
> 
> Otherwise I should create a fake net_device for vsock_core.
> 
> What do you suggest?

Neither.

I think what Jason was saying all along is this:

virtio net doesn't actually lose packets, at least most
of the time. And it actually most of the time
passes all packets to host. So it's possible to use a virtio net
device (possibly with a feature flag that says "does not lose packets,
all packets go to host") and build vsock on top.

and all of this is nice, but don't expect anything easy,
or any quick results.

Also, in a sense it's a missed opportunity: we could cut out a lot
of fat and see just how fast can a protocol that is completely
new and separate from networking stack go.
Instead vsock implementation carries so much baggage from both
networking stack - such as softirq processing - and itself such as
workqueues, global state and crude locking - to the point where
it's actually slower than TCP.

> > 
> > > > > 
> > > > > We can use the virtio-net as transport, but we should add a lot of
> > > > > code to skip "net device" stuff when it is used by the virtio-vsock.
> > 
> > 
> > This could be another choice, but consider it was not transparent to the
> > admin and require new features, we may seek a transparent solution here.
> > 
> > 
> > > > > This could break something in virtio-net, for this reason, I thought to reuse
> > > > > only the send/recv functions starting from the idea to split the virtio-net
> > > > > driver in two parts:
> > > > > a. one with all stuff related to the network stack
> > > > > b. one with the stuff needed to communicate with the host
> > > > > 
> > > > > And use skbuff to communicate between parts. In this way, virtio-vsock
> > > > > can use only the b part.
> > > > > 
> > > > > Maybe we can do this split in a better way, but I'm not sure it is
> > > > > simple.
> > > > > 
> > > > > Thanks,
> > > > > Stefano
> > > > Frankly, skb is a huge structure which adds a lot of
> > > > overhead. I am not sure that using it is such a great idea
> > > > if building a device that does not have to interface
> > > > with the networking stack.
> > 
> > 
> > I believe vsock is mainly used for stream performance not for PPS. So the
> > impact should be minimal. We can use other metadata, just need branch in
> > recv_xxx().
> > 
> 
> Yes, I think stream performance is the case.
> 
> > 
> > > Thanks for the advice!
> > > 
> > > > So I agree with Jason in theory. To clarify, he is basically saying
> > > > current implementation is all wrong, it should be a protocol and we
> > > > should teach networking stack that there are reliable net devices that
> > > > handle just this protocol. We could add a flag in virtio net that
> > > > will say it's such a device.
> > > > 
> > > > Whether it's doable, I don't know, and it's definitely not simple - in
> > > > particular you will have to also re-implement existing devices in these
> > > > terms, and not just virtio - vmware vsock too.
> > 
> > 
> > Merging vsock protocol to exist networking stack could be a long term goal,
> > I believe for the first phase, we can seek to use virtio-net first.
> >
> 
> Yes, I agree.
> 
> > 
> > > > 
> > > > If you want to do a POC you can add a new address family,
> > > > that's easier.
> > > Very interesting!
> > > I agree with you. In this way we can completely split the protocol
> > > logic, from the device.
> > > 
> > > As you said, it will not simple to do, but can be an opportunity to learn
> > > better the Linux networking stack!
> > > I'll try to do a PoC with AF_VSOCK2 that will use the virtio-net.
> > 
> > 
> > I suggest to do this step by step:
> > 
> > 1) use virtio-net but keep some protocol logic
> > 
> > 2) separate protocol logic and merge it to exist Linux networking stack
> 
> Make sense, thanks for the suggestions, I'll try to do these steps!
> 
> Thanks,
> Stefano


An alternative is look at sources of overhead in vsock and get rid of
them, or rewrite it from scratch focusing on performance.


-- 
MST

Powered by blists - more mailing lists