Hi,
thanks for this RFC!
On Fri, May 17, 2024 at 10:46:02PM GMT, Xuewei Niu wrote:
# Motivition
Vsock is a lightweight and widely used data exchange mechanism between host
and guest. Kata Containers, a secure container runtime, leverages the
capability to exchange control data between the shim and the kata-agent.
The Linux kernel only supports one vsock device for virtio-vsock transport,
resulting in the following limitations:
* Poor performance isolation: All vsock connections share the same
virtqueue.
This might be fixed if we implement multi-queue in virtio-vsock.
* Cannot enable more than one backend: Virtio-vsock, vhost-vsock, and
vhost-user-vsock cannot be enabled simultaneously on the transport.
We’d like to transfer networking data, such as TSI (Transparent Socket
Impersonation), over vsock via the vhost-user protocol to reduce overhead.
However, by default, the vsock device is occupied by the kata-agent.
# Usages
Principle: **Supporting virtio-vsock multi-devices while also being
compatible with existing ones.**
## Connection from Guest to Host
There are two valuable questions to take about:
1. How to be compatible with the existing usages?
2. How do we specify a virtio-vsock device?
### Question 1
Before we delve into question 1, I'd like to provide a piece of pseudocode
as an example of one of the existing use cases from the guest's
perspective.
Assuming there is one virtio-vsock device with CID 4. One of existing
usages to connect to host is shown as following.
```
fd = socket(AF_VSOCK);
connect(fd, 2, 1234);
n = write(fd, buffer);
```
The result is that a connection is established from the guest (4, ?) to the
host (2, 1234), where "?" denotes a random port.
In the context of multi-devices, there are more than two devices. If the
users don’t specify one CID explicitly, the kernel becomes confused about
which device to use. The new implementation should be compatible with the
old one.
We expanded the virtio-vsock specification to address this issue. The
specification now includes a new field called "order".
```
struct virtio_vsock_config {
__le64 guest_cid;
__le64 order;
} _attribute_((packed));
```
In the phase of virtio-vsock driver probing, the guest kernel reads
from
VMM to get the order of each device. **We stipulate that the device with the
smallest order is regarded as the default device**(this mechanism functions
as a 'default gateway' in networking).
Assuming there are three virtio-vsock devices: device1 (CID=3), device2
(CID=4), and device3 (CID=5). The arrangement of the list is as follows
from the perspective of the guest kernel:
```
virtio_vsock_list =
virtio_vsock { cid: 4, order: 0 } -> virtio_vsock { cid: 3, order: 1 } ->
virtio_vsock { cid: 5, order: 10 }
```
At this time, the guest kernel realizes that the device2 (CID=4) is the
default device. Execute the same code as before.
```
fd = socket(AF_VSOCK);
connect(fd, 2, 1234);
n = write(fd, buffer);
```
A connection will be established from the guest (4, ?) to the host (2, 1234).
It seems that only the one with order 0 is used here though, so what is
the ordering for?
Wouldn't it suffice to simply indicate the default device (e.g., like
the default gateway for networking)?
### Question 2
Now, the user wants to specify a device instead of the default one. An
explicit binding operation is required to be performed.
Use the device (CID=3), where “-1” represents any port, the kernel will
We have a macro: VMADDR_PORT_ANY (which is -1)
search an available port automatically.
```
fd = socket(AF_VSOCK);
bind(fd, 3, -1);
connect(fd, 2, 1234);)
n = write(fd, buffer);
```
Use the device (CID=4).
```
fd = socket(AF_VSOCK);
bind(fd, 4, -1);
connect(fd, 2, 1234);
n = write(fd, buffer);
```
## Connection from Host to Guest
Connection from host to guest is quite similar to the existing usages. The
device’s CID is specified by the bind operation.
Listen at the device (CID=3)’s port 10000.
```
fd = socket(AF_VSOCK);
bind(fd, 3, 10000);
listen(fd);
new_fd = accept(fd, &host_cid, &host_port);
n = write(fd, buffer);
```
Listen at the device (CID=4)’s port 10000.
```
fd = socket(AF_VSOCK);
bind(fd, 4, 10000);
listen(fd);
new_fd = accept(fd, &host_cid, &host_port);
n = write(fd, buffer);
```
# Use Cases
We've completed a POC with Kata Containers, Ztunnel, which is a
purpose-built per-node proxy for Istio ambient mesh, and TSI. Please refer
to the following link for more details.
Link: https://bit.ly/4bdPJbU
Thank you for this RFC, I left several comments in the patches, we still
have some work to do, but I think it is something we can support :-)
Here I summarize the things that I think we need to fix:
1. Avoid adding transport-specific things in af_vsock.c
We need to have a generic API to allow other transports to implement
the same functionality.
2. We need to add negotiation of a new feature in virtio/vhost transports
We need to enable or disable support depending on whether the
feature is negotiated, since guest and host may not support it.
3. Re-work the patch order for bisectability (more detail on patches 3/4)
4. Do we really need the order or just a default device?
5. Check if we can add some tests in tools/testing/vsock
6. When we agree on the RFC, we should discuss the spec changes in the
virtio ML before sending a non-RFC series on Linux
These are the main things, but I left other comments in the patches.
Thanks,
Stefano
