dsg/cloud-hypervisor
2
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

virtio-devices: vsock: support half-closed host unix sockets

Browse source 
When the host-side unix socket peer does shutdown(SHUT_WR), read()
returns 0 on the stream. Previously this was treated as a full
connection closure, sending VSOCK_OP_SHUTDOWN with both SHUTDOWN_RCV
and SHUTDOWN_SEND flags and arming the kill timer. This prevented the
guest from sending any further data back through the still-writable
socket.

Change LocalClosed to carry a bool indicating whether receiving from
the guest is also shut down. On stream EOF, only set SHUTDOWN_SEND
(host won't send anymore) and transition to LocalClosed(false),
keeping the connection alive for guest-to-host data flow. The
connection fully closes when the guest also sends SHUTDOWN with
SHUTDOWN_SEND, or sends RST.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
Davíð Steinn Geirsson 2026-03-23 23:34:11 +00:00
parent 2f77d61b52
commit c5682801d6
2 changed files with 66 additions and 25 deletions
Download patch file Download diff file Expand all files Collapse all files

85
virtio-devices/src/vsock/csm/connection.rs
View file

@ -215,15 +215,12 @@ where
match self.stream.read(&mut buf[..max_len]) { match self.stream.read(&mut buf[..max_len]) {
Ok(read_cnt) => { Ok(read_cnt) => {
if read_cnt == 0 { if read_cnt == 0 {
// A 0-length read means the host stream was closed down. In that case, // A 0-length read means the host stream's read side was closed (e.g.
// we'll ask our peer to shut down the connection. We can neither send nor // the host peer did shutdown(SHUT_WR)). We can no longer send data to
// receive any more data. // the guest, but the guest may still send data to us (to be written
self.state = ConnState::LocalClosed; // to the host stream).
self.expiry = Some( self.state = ConnState::LocalClosed(false);
Instant::now() + Duration::from_millis(defs::CONN_SHUTDOWN_TIMEOUT_MS),
);
pkt.set_op(uapi::VSOCK_OP_SHUTDOWN) pkt.set_op(uapi::VSOCK_OP_SHUTDOWN)
.set_flag(uapi::VSOCK_FLAGS_SHUTDOWN_RCV)
.set_flag(uapi::VSOCK_FLAGS_SHUTDOWN_SEND); .set_flag(uapi::VSOCK_FLAGS_SHUTDOWN_SEND);
} else { } else {
// On a successful data read, we fill in the packet with the RW op, and // On a successful data read, we fill in the packet with the RW op, and
@ -288,7 +285,9 @@ where
// Most frequent case: this is an established connection that needs to forward some // Most frequent case: this is an established connection that needs to forward some
// data to the host stream. Also works for a connection that has begun shutting // data to the host stream. Also works for a connection that has begun shutting
// down, but the peer still has some data to send. // down, but the peer still has some data to send.
ConnState::Established | ConnState::PeerClosed(_, false) ConnState::Established
| ConnState::PeerClosed(_, false)
| ConnState::LocalClosed(false)
if pkt.op() == uapi::VSOCK_OP_RW => if pkt.op() == uapi::VSOCK_OP_RW =>
{ {
if pkt.buf().is_none() { if pkt.buf().is_none() {
@ -359,6 +358,26 @@ where
} }
} }
// The peer wants to shut down while the local stream's read side is already closed.
// We only care about the send indication: if the peer will no longer send data, both
// directions are now closed and we can schedule termination.
ConnState::LocalClosed(ref mut recv_off)
if pkt.op() == uapi::VSOCK_OP_SHUTDOWN =>
{
let send_off = pkt.flags() & uapi::VSOCK_FLAGS_SHUTDOWN_SEND != 0;
if send_off && !*recv_off {
*recv_off = true;
if self.tx_buf.is_empty() {
self.pending_rx.insert(PendingRx::Rst);
} else {
self.expiry = Some(
Instant::now()
+ Duration::from_millis(defs::CONN_SHUTDOWN_TIMEOUT_MS),
);
}
}
}
// A credit update from our peer is valid only in a state which allows data // A credit update from our peer is valid only in a state which allows data
// transfer towards the peer. // transfer towards the peer.
ConnState::Established | ConnState::PeerInit | ConnState::PeerClosed(false, _) ConnState::Established | ConnState::PeerInit | ConnState::PeerClosed(false, _)
@ -369,7 +388,10 @@ where
// A credit request from our peer is valid only in a state which allows data // A credit request from our peer is valid only in a state which allows data
// transfer from the peer. We'll respond with a credit update packet. // transfer from the peer. We'll respond with a credit update packet.
ConnState::Established | ConnState::PeerInit | ConnState::PeerClosed(_, false) ConnState::Established
| ConnState::PeerInit
| ConnState::PeerClosed(_, false)
| ConnState::LocalClosed(false)
if pkt.op() == uapi::VSOCK_OP_CREDIT_REQUEST => if pkt.op() == uapi::VSOCK_OP_CREDIT_REQUEST =>
{ {
self.pending_rx.insert(PendingRx::CreditUpdate); self.pending_rx.insert(PendingRx::CreditUpdate);
@ -424,7 +446,7 @@ where
// We're generally interested in being notified when data can be read from the host // We're generally interested in being notified when data can be read from the host
// stream, unless we're in a state which doesn't allow moving data from host to guest. // stream, unless we're in a state which doesn't allow moving data from host to guest.
match self.state { match self.state {
ConnState::Killed | ConnState::LocalClosed | ConnState::PeerClosed(true, _) => (), ConnState::Killed | ConnState::LocalClosed(_) | ConnState::PeerClosed(true, _) => (),
_ if self.need_credit_update_from_peer() => (), _ if self.need_credit_update_from_peer() => (),
_ => evset.insert(epoll::Events::EPOLLIN), _ => evset.insert(epoll::Events::EPOLLIN),
} }
@ -468,7 +490,10 @@ where
// If this connection was shutting down, but is waiting to drain the TX buffer // If this connection was shutting down, but is waiting to drain the TX buffer
// before forceful termination, the wait might be over. // before forceful termination, the wait might be over.
if self.state == ConnState::PeerClosed(true, true) && self.tx_buf.is_empty() { if (self.state == ConnState::PeerClosed(true, true)
|| self.state == ConnState::LocalClosed(true))
&& self.tx_buf.is_empty()
{
self.pending_rx.insert(PendingRx::Rst); self.pending_rx.insert(PendingRx::Rst);
} else if self.peer_needs_credit_update() { } else if self.peer_needs_credit_update() {
// If we've freed up some more buffer space, we may need to let the peer know it // If we've freed up some more buffer space, we may need to let the peer know it
@ -973,7 +998,7 @@ mod unit_tests {
} }
// A recv attempt in an invalid state should yield an instant reset packet. // A recv attempt in an invalid state should yield an instant reset packet.
ctx.conn.state = ConnState::LocalClosed; ctx.conn.state = ConnState::LocalClosed(true);
ctx.notify_epollin(); ctx.notify_epollin();
ctx.recv(); ctx.recv();
assert_eq!(ctx.pkt.op(), uapi::VSOCK_OP_RST); assert_eq!(ctx.pkt.op(), uapi::VSOCK_OP_RST);
@ -981,25 +1006,39 @@ mod unit_tests {
#[test] #[test]
fn test_local_close() { fn test_local_close() {
// When the host stream's read side is closed (e.g. peer did shutdown(SHUT_WR)),
// we should only indicate SHUTDOWN_SEND (we won't send anymore), not SHUTDOWN_RCV,
// since the guest can still send data to us.
let mut ctx = CsmTestContext::new_established(); let mut ctx = CsmTestContext::new_established();
let mut stream = TestStream::new(); let mut stream = TestStream::new();
stream.read_state = StreamState::Closed; stream.read_state = StreamState::Closed;
ctx.set_stream(stream); ctx.set_stream(stream);
ctx.notify_epollin(); ctx.notify_epollin();
ctx.recv(); ctx.recv();
// When the host-side stream is closed, we can neither send not receive any more data.
// Therefore, the vsock shutdown packet that we'll deliver to the guest must contain both
// the no-more-send and the no-more-recv indications.
assert_eq!(ctx.pkt.op(), uapi::VSOCK_OP_SHUTDOWN); assert_eq!(ctx.pkt.op(), uapi::VSOCK_OP_SHUTDOWN);
assert_ne!(ctx.pkt.flags() & uapi::VSOCK_FLAGS_SHUTDOWN_SEND, 0); assert_ne!(ctx.pkt.flags() & uapi::VSOCK_FLAGS_SHUTDOWN_SEND, 0);
assert_ne!(ctx.pkt.flags() & uapi::VSOCK_FLAGS_SHUTDOWN_RCV, 0); assert_eq!(ctx.pkt.flags() & uapi::VSOCK_FLAGS_SHUTDOWN_RCV, 0);
// The kill timer should now be armed. // The connection should be in half-closed state, not fully closed.
assert!(ctx.conn.will_expire()); assert_eq!(ctx.conn.state, ConnState::LocalClosed(false));
assert!(
ctx.conn.expiry().unwrap() // No kill timer should be set for half-close.
< Instant::now() + Duration::from_millis(defs::CONN_SHUTDOWN_TIMEOUT_MS) assert!(!ctx.conn.will_expire());
);
// Guest should still be able to send data to us.
let data = &[1, 2, 3, 4];
ctx.init_data_pkt(data);
ctx.send();
assert_eq!(ctx.conn.stream.write_buf, data.to_vec());
// When the guest also shuts down sending, the connection should be fully closed.
ctx.init_pkt(uapi::VSOCK_OP_SHUTDOWN, 0)
.set_flags(uapi::VSOCK_FLAGS_SHUTDOWN_SEND);
ctx.send();
assert_eq!(ctx.conn.state, ConnState::LocalClosed(true));
assert!(ctx.conn.has_pending_rx());
ctx.recv();
assert_eq!(ctx.pkt.op(), uapi::VSOCK_OP_RST);
} }
#[test] #[test]

6
virtio-devices/src/vsock/csm/mod.rs
View file

@ -52,8 +52,10 @@ pub enum ConnState {
PeerInit, PeerInit,
/// The connection handshake has been performed successfully, and data can now be exchanged. /// The connection handshake has been performed successfully, and data can now be exchanged.
Established, Established,
/// The host (AF_UNIX) socket was closed. /// The host (AF_UNIX) socket's read side was closed (e.g. the host peer did
LocalClosed, /// shutdown(SHUT_WR)). The bool indicates whether receiving from the guest is also shut down
/// (i.e. the connection is fully closed from the local perspective).
LocalClosed(bool),
/// A VSOCK_OP_SHUTDOWN packet was received from the guest. The tuple represents the guest R/W /// A VSOCK_OP_SHUTDOWN packet was received from the guest. The tuple represents the guest R/W
/// indication: (will_not_recv_anymore_data, will_not_send_anymore_data). /// indication: (will_not_recv_anymore_data, will_not_send_anymore_data).
PeerClosed(bool, bool), PeerClosed(bool, bool),