vsock fails when guest sends large packets to host

[r0ci]

The vsock implementation fails when guests send large enough packets to the host (> 32KB on x86_64 linux guests).

I think it is caused by this change in the kernel: https://lkml.org/lkml/2025/7/17/487. They're allocating non linear skbs, so packets that are larger than one skb frag are getting split across multiple descriptors. On x86_64 this is 32KB (PAGE_SIZE << PAGE_ALLOC_COSTLY_ORDER).

The current implementation compares packet length and the descriptor's size, sees it doesn't fit and returns BufDescTooSmall:

libkrun/src/devices/src/virtio/vsock/packet.rs

Lines 248 to 252 in c9885c6

	// The data buffer should be large enough to fit the size of the data, as described by
	// the header descriptor.
	if buf_desc.len < pkt.len() {
	return Err(VsockError::BufDescTooSmall);
	}

Reproducing

# on host
nc -lp 8000

# start container
podman run --runtime=krun --network=pasta:--map-host-loopback=10.10.10.10 --rm -it python:latest

# on container repl
import socket
import time

with socket.create_connection(("10.10.10.10", 8000)) as s:
    sizes = [
        1024 * 32 - 1,
        32 * 1024,
        32 * 1024 + 1,
        64 * 1024,
        96 * 1024,
        96 * 1024 + 1,
    ]

    for size in sizes:
        print(f"Sending {size}")
        d = b"a" * size
        s.sendall(d)
        print("Sent")
        time.sleep(1)

You'll see output like:

sending 32767
sent
sending 32768
sent
sending 32769
[2026-02-07T23:21:56Z ERROR devices::virtio::vsock::device] error reading TX packet: BufDescTooSmall
sent
sending 65536
[2026-02-07T23:21:57Z ERROR devices::virtio::vsock::device] error reading TX packet: BufDescTooSmall
sent
sending 98304
[2026-02-07T23:21:58Z ERROR devices::virtio::vsock::device] error reading TX packet: BufDescTooSmall
sent
sending 98305
<-- hangs here until I interrupt

host kernel: 6.18.9-2
guest kernel: 6.12.68
libkrun 1.17.1

[slp]

Thanks a lot for the detailed issue report. @mtjhrc I think we can reuse the TxQueueConsumer you're introducing for virtio-net here. Would you mind taking a look at this once #530 lands?

[mtjhrc]

I think we can reuse the TxQueueConsumer you're introducing for virtio-net here

Yeah, I was hoping that it would be useful vsock too. I'll take a look.

[stefano-garzarella]

FYI in vhost-devices and vm-virtio we have the same issue, but we are going to reuse the Reader/Writer shipped by virtio-queue.
@luigix25 is looking at rust-vmm/vm-virtio#216 so would be nice to share the code as much as possible (if possible of course ;-)

[mtjhrc]

We have the Reader/Writer too (it's currently vendored though, ugh...). When I start working on this I'll take a look how the rust-vmm parsing and ours differ (and most importantly see how we actually want to consume the parsed packets).

The TxQueueConsumer/ RxQueueProducer are designed to allow processing all the the descriptor chains in a virtio-queue together in one operation - e.g. if you want to use something like a non-blocking sendmmsg with every descriptor in the queue at once (we want to do that in virtio-net).

[ihexon]

I encountered the same problem:

[2026-03-10T07:41:02.370309Z ERROR devices::virtio::vsock::device] error reading TX packet: BufDescTooSmall

This issue seems to be triggered quite frequently under heavy network load.
Are there any ways to alleviate this problem?

[slp]

@ihexon This PR in libkrunfw (containers/libkrunfw#113) should get TSI/vsock into a working state until we update our implementation here.

Please give it a try if you have the opportunity.