fix(vsock): preserve fragmented stream uploads

[sazlin]

Context

I am employing microsandbox to sandbox an OMP agent with an OpenAI Codex subscription. I don't want the subscription credentials visible inside the vm, so I am using a vsock channel to broker/proxy OMP --> OpenAI communication through the host. So: OMP (sandboxed guest) --vsock--> Broker/Proxy (host) --internet--> OpenAI. This patch fixes libkrun’s vsock handling so complete OMP requests can reach the host-side broker without being truncated or stalled.

Problem

Large guest-to-host stream writes can arrive as a virtqueue descriptor chain. VsockPacket::from_tx_virtq_head currently requires the entire declared payload in the first descriptor, logs BufDescTooSmall, and drops the remaining HTTP request. Separately, the custom stream advertises an 8 MiB receive window but returns credit only after 4 MiB; ordinary guest socket transmit buffers can stall before reaching that threshold.

This was reproduced through microsandbox 0.6.17 and 0.6.18 with an HTTP POST: 32 KiB succeeds, 65,535 bytes triggers BufDescTooSmall, and larger uploads stall.

Change

• Preserve the zero-copy path for a single payload descriptor.
• Gather fragmented TX payloads into a bounded (64 KiB maximum) allocation.
• Reject writable or truncated descriptor chains.
• Use a 64 KiB custom-stream receive window, returning credit after 32 KiB.

Verification

• New regressions fail against the previous parser/credit window and pass with this change.
• 11 scoped virtio::vsock tests pass.
• A patched microsandbox CLI delivered byte-identical HTTP requests of 32 KiB, 65,535 B, 64 KiB, 256 KiB, 1 MiB, and 16 MiB both directly and through a guest TCP-to-vsock relay.
• Chunked 1 MiB and four concurrent 1 MiB relay requests pass.

Confidence Score: 5/5

The PR appears safe to merge based on the changes since the previous review.

_{Reviews (6) · Last reviewed commit: "fix(vsock): retry credit safely after RX..."}

[sazlin]

Cloud Hypervisor had the same issue: cloud-hypervisor/cloud-hypervisor#7672

[toksdotdev]

hey @sazlin, thanks for flagging and for the fix.

could we use a PacketBuffer enum with Borrowed { ptr, len } and Owned(Vec<u8>) variants, similar to Cloud Hypervisor’s fix? that would replace buf, buf_size, and owned_buf with one field, avoiding a separate pointer and length for the owned buffer. behavior would stay the same, but ownership would be clearer.

[toksdotdev]

could we also add a CI step for the vsock tests? the default cargo test doesn't select the devices crate, and the existing devices step filters for virtio::net, so these new regressions won't run.

cargo test --locked --manifest-path src/devices/Cargo.toml virtio::vsock

[toksdotdev]

thanks for the fix. made a tiny adjustment to use concrete enum types. asides that, lgtm!