|
| 1 | +# Virtio-vsock DGRAM Implementation on Windows |
| 2 | + |
| 3 | +## Overview |
| 4 | + |
| 5 | +This document describes the implementation of DGRAM (datagram/connectionless) support for virtio-vsock on Windows, completing the P2 feature set for the Windows backend. |
| 6 | + |
| 7 | +## Background |
| 8 | + |
| 9 | +Virtio-vsock supports two socket types: |
| 10 | +- **STREAM (type 1)**: Connection-oriented, reliable, ordered (like TCP) |
| 11 | +- **DGRAM (type 3)**: Connectionless, unreliable, unordered (like UDP) |
| 12 | + |
| 13 | +Prior to this implementation, the Windows backend only supported STREAM sockets via TCP and Named Pipes. DGRAM support enables connectionless communication scenarios. |
| 14 | + |
| 15 | +## Architecture |
| 16 | + |
| 17 | +### Data Structures |
| 18 | + |
| 19 | +```rust |
| 20 | +pub struct Vsock { |
| 21 | + // ... existing fields ... |
| 22 | + streams: HashMap<u32, StreamState>, // STREAM sockets |
| 23 | + dgram_sockets: HashMap<u32, UdpSocket>, // DGRAM sockets (NEW) |
| 24 | + // ... other fields ... |
| 25 | +} |
| 26 | +``` |
| 27 | + |
| 28 | +### Key Components |
| 29 | + |
| 30 | +1. **DGRAM Socket Management** |
| 31 | + - `dgram_sockets: HashMap<u32, UdpSocket>` maps guest port → UDP socket |
| 32 | + - Sockets are created on-demand when first DGRAM packet is sent |
| 33 | + - Each socket is bound to `0.0.0.0:0` (any local address/port) |
| 34 | + |
| 35 | +2. **TX Path (Guest → Host)** |
| 36 | + - Guest sends DGRAM packet via `VSOCK_OP_RW` with `VSOCK_TYPE_DGRAM` |
| 37 | + - VMM creates UDP socket if not exists |
| 38 | + - VMM sends datagram to mapped host port via `UdpSocket::send_to()` |
| 39 | + |
| 40 | +3. **RX Path (Host → Guest)** |
| 41 | + - `harvest_dgram_reads()` polls all DGRAM sockets |
| 42 | + - Receives datagrams via `UdpSocket::recv_from()` |
| 43 | + - Constructs vsock header with `VSOCK_TYPE_DGRAM` |
| 44 | + - Queues packet to guest RX queue |
| 45 | + |
| 46 | +## Implementation Details |
| 47 | + |
| 48 | +### Feature Advertisement |
| 49 | + |
| 50 | +```rust |
| 51 | +const AVAIL_FEATURES: u64 = (1 << VIRTIO_F_VERSION_1 as u64) |
| 52 | + | (1 << VIRTIO_F_IN_ORDER as u64) |
| 53 | + | (1 << VIRTIO_VSOCK_F_DGRAM as u64); // Bit 3 |
| 54 | +``` |
| 55 | + |
| 56 | +### TX Processing (VSOCK_OP_RW) |
| 57 | + |
| 58 | +```rust |
| 59 | +if pkt_type == VSOCK_TYPE_DGRAM { |
| 60 | + // Create socket on first use |
| 61 | + if !self.dgram_sockets.contains_key(&src_port) { |
| 62 | + let socket = UdpSocket::bind("0.0.0.0:0")?; |
| 63 | + socket.set_nonblocking(true)?; |
| 64 | + self.dgram_sockets.insert(src_port, socket); |
| 65 | + } |
| 66 | + |
| 67 | + // Send datagram to host |
| 68 | + if let Some(socket) = self.dgram_sockets.get(&src_port) { |
| 69 | + if let Some(addr) = self.host_socket_addr(dst_port) { |
| 70 | + socket.send_to(&payload, addr)?; |
| 71 | + } |
| 72 | + } |
| 73 | +} |
| 74 | +``` |
| 75 | + |
| 76 | +### RX Processing (harvest_dgram_reads) |
| 77 | + |
| 78 | +```rust |
| 79 | +fn harvest_dgram_reads(&mut self) { |
| 80 | + for (guest_port, socket) in &self.dgram_sockets { |
| 81 | + let mut rx_buf = [0u8; 4096]; |
| 82 | + match socket.recv_from(&mut rx_buf) { |
| 83 | + Ok((n, peer_addr)) => { |
| 84 | + // Construct vsock header |
| 85 | + let mut hdr = [0u8; 44]; |
| 86 | + Self::set_u64(&mut hdr, 0, VSOCK_HOST_CID); |
| 87 | + Self::set_u64(&mut hdr, 8, self.cid); |
| 88 | + Self::set_u32(&mut hdr, 16, peer_addr.port() as u32); |
| 89 | + Self::set_u32(&mut hdr, 20, guest_port); |
| 90 | + Self::set_u32(&mut hdr, 24, n as u32); |
| 91 | + Self::set_u16(&mut hdr, 28, VSOCK_TYPE_DGRAM); |
| 92 | + Self::set_u16(&mut hdr, 30, VSOCK_OP_RW); |
| 93 | + |
| 94 | + self.queue_response(&hdr, VSOCK_OP_RW, rx_buf[..n].to_vec()); |
| 95 | + } |
| 96 | + Err(e) if e.kind() == io::ErrorKind::WouldBlock => {} |
| 97 | + Err(_) => {} |
| 98 | + } |
| 99 | + } |
| 100 | +} |
| 101 | +``` |
| 102 | + |
| 103 | +## Differences from STREAM |
| 104 | + |
| 105 | +| Aspect | STREAM | DGRAM | |
| 106 | +|--------|--------|-------| |
| 107 | +| Connection | Requires REQUEST/RESPONSE handshake | No handshake | |
| 108 | +| State | Maintains StreamState per connection | Stateless (socket per port) | |
| 109 | +| Flow Control | Credit-based (buf_alloc, fwd_cnt, tx_cnt) | None | |
| 110 | +| Backend | TCP or Named Pipe | UDP | |
| 111 | +| Reliability | Guaranteed delivery, ordered | Best-effort, may be lost/reordered | |
| 112 | +| Operations | REQUEST, RESPONSE, RW, CREDIT_UPDATE, SHUTDOWN, RST | RW only | |
| 113 | + |
| 114 | +## Testing |
| 115 | + |
| 116 | +### Smoke Test |
| 117 | + |
| 118 | +```rust |
| 119 | +#[test] |
| 120 | +fn test_whpx_vsock_dgram_feature() { |
| 121 | + let vsock = Vsock::new(3, None, None, Default::default()).unwrap(); |
| 122 | + |
| 123 | + // Verify DGRAM feature is advertised |
| 124 | + let features = vsock.avail_features(); |
| 125 | + assert_ne!(features & (1 << 3), 0, "VIRTIO_VSOCK_F_DGRAM not advertised"); |
| 126 | +} |
| 127 | +``` |
| 128 | + |
| 129 | +### Test Results |
| 130 | + |
| 131 | +``` |
| 132 | +running 54 tests |
| 133 | +test windows::vstate::tests::test_whpx_vsock_dgram_feature ... ok |
| 134 | +test windows::vstate::tests::test_whpx_vsock_init_smoke ... ok |
| 135 | +test windows::vstate::tests::test_whpx_vsock_tx_smoke ... ok |
| 136 | +test result: ok. 44 passed; 0 failed; 10 ignored; 0 measured |
| 137 | +``` |
| 138 | + |
| 139 | +## Limitations |
| 140 | + |
| 141 | +1. **Port Mapping Heuristic**: RX path uses peer UDP port as guest dst_port. This may not match the original guest port if NAT is involved. |
| 142 | + |
| 143 | +2. **No Reverse Mapping**: The implementation doesn't maintain a reverse mapping from host UDP ports to guest ports, which could cause issues in complex scenarios. |
| 144 | + |
| 145 | +3. **UDP Only**: DGRAM support is limited to UDP. Named Pipe DGRAM is not implemented (Windows Named Pipes don't support datagram mode). |
| 146 | + |
| 147 | +4. **No Fragmentation**: Large datagrams (>4096 bytes) are not supported. UDP fragmentation is handled by the network stack. |
| 148 | + |
| 149 | +## Future Improvements |
| 150 | + |
| 151 | +1. **Port Mapping Table**: Maintain bidirectional mapping between guest ports and host UDP ports for accurate RX routing. |
| 152 | + |
| 153 | +2. **Socket Cleanup**: Implement timeout-based cleanup for idle DGRAM sockets to prevent resource leaks. |
| 154 | + |
| 155 | +3. **Error Handling**: Improve error handling for socket creation and I/O failures. |
| 156 | + |
| 157 | +4. **Metrics**: Add counters for DGRAM packets sent/received, errors, etc. |
| 158 | + |
| 159 | +## References |
| 160 | + |
| 161 | +- [Virtio Specification - vsock Device](https://docs.oasis-open.org/virtio/virtio/v1.2/csd01/virtio-v1.2-csd01.html#x1-4050008) |
| 162 | +- [Linux vsock DGRAM implementation](https://github.com/torvalds/linux/blob/master/net/vmw_vsock/af_vsock.c) |
| 163 | +- Windows UDP Socket API: `std::net::UdpSocket` |
| 164 | + |
| 165 | +--- |
| 166 | + |
| 167 | +*Implementation Date: 2026-03-05* |
| 168 | +*Commit: e7700cc* |
0 commit comments