Virtualization - Apr 2021 - [RFC] vsock: add multiple transports support for dgram

On Mon, Apr 12, 2021 at 7:04 AM Stefano Garzarella <sgarzare at
redhat.com> wrote:
>
> Hi Jiang,
> thanks for re-starting the multi-transport support for dgram!
No problem.
> On Wed, Apr 07, 2021 at 11:25:36AM -0700, Jiang Wang . wrote:
> >On Wed, Apr 7, 2021 at 2:51 AM Jorgen Hansen <jhansen at
vmware.com> wrote:
> >>
> >>
> >> > On 6 Apr 2021, at 20:31, Jiang Wang <jiang.wang at
bytedance.com> wrote:
> >> >
> >> > From: "jiang.wang" <jiang.wang at
bytedance.com>
> >> >
> >> > Currently, only VMCI supports dgram sockets. To supported
> >> > nested VM use case, this patch removes transport_dgram and
> >> > uses transport_g2h and transport_h2g for dgram too.
>
> I agree on this part, I think that's the direction to go.
> transport_dgram was added as a shortcut.
Got it.
> >>
> >> Could you provide some background for introducing this change -
are you
> >> looking at introducing datagrams for a different transport? VMCI
datagrams
> >> already support the nested use case,
> >
> >Yes, I am trying to introduce datagram for virtio transport. I wrote a
> >spec patch for
> >virtio dgram support and also a code patch, but the code patch is still
WIP.
> >When I wrote this commit message, I was thinking nested VM is the same
as
> >multiple transport support. But now, I realize they are different.
> >Nested VMs may use
> >the same virtualization layer(KVM on KVM), or different virtualization
layers
> >(KVM on ESXi). Thanks for letting me know that VMCI already supported
nested
> >use cases. I think you mean VMCI on VMCI, right?
> >
> >> but if we need to support multiple datagram
> >> transports we need to rework how we administer port assignment for
datagrams.
> >> One specific issue is that the vmci transport won?t receive any
datagrams for a
> >> port unless the datagram socket has already been assigned the vmci
transport
> >> and the port bound to the underlying VMCI device (see below for
more details).
> >>
> >I see.
> >
> >> > The transport is assgined when sending every packet and
> >> > receiving every packet on dgram sockets.
> >>
> >> Is the intent that the same datagram socket can be used for
sending packets both
> >> In the host to guest, and the guest to directions?
> >
> >Nope. One datagram socket will only send packets to one direction,
either to the
> >host or to the guest. My above description is wrong. When sending
packets, the
> >transport is assigned with the first packet (with auto_bind).
>
> I'm not sure this is right.
> The auto_bind on the first packet should only assign a local port to the
> socket, but does not affect the transport to be used.
>
> A user could send one packet to the nested guest and another to the host
> using the same socket, or am I wrong?
OK. I think you are right.
> >
> >The problem is when receiving packets. The listener can bind to the
> >VMADDR_CID_ANY
> >address. Then it is unclear which transport we should use. For stream
> >sockets, there will be a new socket for each connection, and transport
> >can be decided
> >at that time. For datagram sockets, I am not sure how to handle that.
>
> yes, this I think is the main problem, but maybe the sender one is even
> more complicated.
>
> Maybe we should remove the 1:1 association we have now between vsk and
> transport.
Yes, I thought about that too. One idea is to define two transports in vsk.
For sending pkt, we can pick the right transport when we get the packet, like
in virtio_transport_send_pkt_info(). For receiving pkts, we have to check
and call both
transports dequeue callbacks if the local cid is CID_ANY.
> At least for DGRAM, for connected sockets I think the association makes
> sense.
Yeah. For a connected socket, we will only use one transport.
> >For VMCI, does the same transport can be used for both receiving from
> >host and from
> >the guest?
>
> Yes, they're registered at different times, but it's the same
transport.
>
> >
> >For virtio, the h2g and g2h transports are different,, so we have to
> >choose
> >one of them. My original thought is to wait until the first packet
> >arrives.
> >
> >Another idea is that we always bind to host addr and use h2g
> >transport because I think that might
> >be more common. If a listener wants to recv packets from the host, then
> >it
> >should bind to the guest addr instead of CID_ANY.
>
> Yes, I remember we discussed this idea, this would simplify the
> receiving, but there is still the issue of a user wanting to receive
> packets from both the nested guest and the host.
OK. Agree.
> >Any other suggestions?
> >
>
> I think one solution could be to remove the 1:1 association between
> DGRAM socket and transport.
>
> IIUC VMCI creates a skb for each received packet and queues it through
> sk_receive_skb() directly in the struct sock.
>
> Then the .dgram_dequeue() callback dequeues them using
> skb_recv_datagram().
>
> We can move these parts in the vsock core, and create some helpers to
> allow the transports to enqueue received DGRAM packets in the same way
> (and with the same format) directly in the struct sock.
>
I agree to use skbs (and move them to vscok core). We have another use case
which will need to use skb. But I am not sure how this helps with multiple
transport cases. Each transport has a dgram_dequeue callback. So we still
need to let vsk have multiple transports somehow. Could you elaborate how
using skb help with multiple transport support? Will that be similar to what I
mentioned above? Thanks.

Regards,

Jiang
> What do you think?
>
> Thanks,
> Stefano
>

On 12 Apr 2021, at 20:53, Jiang Wang . <jiang.wang at
bytedance.com<mailto:jiang.wang at bytedance.com>> wrote:

On Mon, Apr 12, 2021 at 7:04 AM Stefano Garzarella <sgarzare at
redhat.com<mailto:sgarzare at redhat.com>> wrote:

Hi Jiang,
thanks for re-starting the multi-transport support for dgram!

No problem.

On Wed, Apr 07, 2021 at 11:25:36AM -0700, Jiang Wang . wrote:
On Wed, Apr 7, 2021 at 2:51 AM Jorgen Hansen <jhansen at
vmware.com<mailto:jhansen at vmware.com>> wrote:


On 6 Apr 2021, at 20:31, Jiang Wang <jiang.wang at
bytedance.com<mailto:jiang.wang at bytedance.com>> wrote:

From: "jiang.wang<http://jiang.wang>" <jiang.wang at
bytedance.com<mailto:jiang.wang at bytedance.com>>

Currently, only VMCI supports dgram sockets. To supported
nested VM use case, this patch removes transport_dgram and
uses transport_g2h and transport_h2g for dgram too.

I agree on this part, I think that's the direction to go.
transport_dgram was added as a shortcut.

Got it.


Could you provide some background for introducing this change - are you
looking at introducing datagrams for a different transport? VMCI datagrams
already support the nested use case,

Yes, I am trying to introduce datagram for virtio transport. I wrote a
spec patch for
virtio dgram support and also a code patch, but the code patch is still WIP.
When I wrote this commit message, I was thinking nested VM is the same as
multiple transport support. But now, I realize they are different.
Nested VMs may use
the same virtualization layer(KVM on KVM), or different virtualization layers
(KVM on ESXi). Thanks for letting me know that VMCI already supported nested
use cases. I think you mean VMCI on VMCI, right?

but if we need to support multiple datagram
transports we need to rework how we administer port assignment for datagrams.
One specific issue is that the vmci transport won?t receive any datagrams for a
port unless the datagram socket has already been assigned the vmci transport
and the port bound to the underlying VMCI device (see below for more details).

I see.

The transport is assgined when sending every packet and
receiving every packet on dgram sockets.

Is the intent that the same datagram socket can be used for sending packets both
In the host to guest, and the guest to directions?

Nope. One datagram socket will only send packets to one direction, either to the
host or to the guest. My above description is wrong. When sending packets, the
transport is assigned with the first packet (with auto_bind).

I'm not sure this is right.
The auto_bind on the first packet should only assign a local port to the
socket, but does not affect the transport to be used.

A user could send one packet to the nested guest and another to the host
using the same socket, or am I wrong?

OK. I think you are right.


The problem is when receiving packets. The listener can bind to the
VMADDR_CID_ANY
address. Then it is unclear which transport we should use. For stream
sockets, there will be a new socket for each connection, and transport
can be decided
at that time. For datagram sockets, I am not sure how to handle that.

yes, this I think is the main problem, but maybe the sender one is even
more complicated.

Maybe we should remove the 1:1 association we have now between vsk and
transport.

Yes, I thought about that too. One idea is to define two transports in vsk.
For sending pkt, we can pick the right transport when we get the packet, like
in virtio_transport_send_pkt_info(). For receiving pkts, we have to check
and call both
transports dequeue callbacks if the local cid is CID_ANY.

At least for DGRAM, for connected sockets I think the association makes
sense.

Yeah. For a connected socket, we will only use one transport.

For VMCI, does the same transport can be used for both receiving from
host and from
the guest?

Yes, they're registered at different times, but it's the same transport.


For virtio, the h2g and g2h transports are different,, so we have to
choose
one of them. My original thought is to wait until the first packet
arrives.

Another idea is that we always bind to host addr and use h2g
transport because I think that might
be more common. If a listener wants to recv packets from the host, then
it
should bind to the guest addr instead of CID_ANY.

Yes, I remember we discussed this idea, this would simplify the
receiving, but there is still the issue of a user wanting to receive
packets from both the nested guest and the host.

OK. Agree.

Any other suggestions?


I think one solution could be to remove the 1:1 association between
DGRAM socket and transport.

IIUC VMCI creates a skb for each received packet and queues it through
sk_receive_skb() directly in the struct sock.

Then the .dgram_dequeue() callback dequeues them using
skb_recv_datagram().

We can move these parts in the vsock core, and create some helpers to
allow the transports to enqueue received DGRAM packets in the same way
(and with the same format) directly in the struct sock.


I agree to use skbs (and move them to vscok core). We have another use case
which will need to use skb. But I am not sure how this helps with multiple
transport cases. Each transport has a dgram_dequeue callback. So we still
need to let vsk have multiple transports somehow. Could you elaborate how
using skb help with multiple transport support? Will that be similar to what I
mentioned above? Thanks.

Moving away from the 1:1 association between DGRAM socket and transports sounds
like the right approach to me. A dgram socket bound to CID_ANY would be able to
use either h2g or g2h on a per dgram basis. If the socket is bound to a specific
CID -
either host or the guest CID, it should only use either the h2g for host CID or
g2h
for the guest CID. This would match the logic for the stream sockets.

I like the idea of removing the dgram_dequeue callback from the transports and
instead
having a call that allow the transports to enqueue received dgrams into the
socket
receive queue as skbs. This is what the VMCI transport does today. Then the
vsock_dgram_recvmsg function will provide functionality similar to what
vmci_transport_dgram_dequeue does today. The current datagram format used was
created specifically for VMCI datagrams, but the header just contains source and
dest
CID and port, so we should be able to use it as is.

For sends from CID_ANY, the same logic as for streams in vsock_assign_transport
can
be applied on each send - but without locking the dgram socket to a specific
transport.

So the above is mostly restating what Stefano proposed, so this was a verbose
way
of agreeing with that.

With respect to binding a dgram socket to a port, we could introduce a bound
list for
dgram sockets just like we have for streams. However, for VMCI, the port space
is shared with other VMCI datagram clients (at the VMCI device level), so if a
dgram socket can potentially use the vmci transport, it should reserve the port
with the VMCI transport before assigning it to the socket. So similar to how
__vsock_bind_stream checks if an port is already bound/in use, the dgram socket
would have an additional call to potential transports to reserve the port. If
the
port cannot be reserved with the transport, move on to the next port in the case
of VMADDR_PORT_ANY, or return EADDRINUSE otherwise. Once reserved,
It will ensure that VMCI can deliver datagrams to the specified port. A reserved
port should be released when the socket is removed from the bound list.

Thanks,
Jorgen
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.linuxfoundation.org/pipermail/virtualization/attachments/20210413/99c04ae4/attachment-0001.html>