On Wed, Apr 23, 2025 at 05:53:12PM +0200, Luigi Leonardi wrote:
Hi Michal,
On Mon, Apr 21, 2025 at 11:50:41PM +0200, Michal Luczaj wrote:
Currently vsock's lingering effectively boils down to waiting (or timing
out) until packets are consumed or dropped by the peer; be it by receiving
the data, closing or shutting down the connection.
To align with the semantics described in the SO_LINGER section of man
socket(7) and to mimic AF_INET's behaviour more closely, change the logic
of a lingering close(): instead of waiting for all data to be handled,
block until data is considered sent from the vsock's transport point of
view. That is until worker picks the packets for processing and decrements
virtio_vsock_sock::bytes_unsent down to 0.
Note that such lingering is limited to transports that actually implement
vsock_transport::unsent_bytes() callback. This excludes Hyper-V and VMCI,
under which no lingering would be observed.
The implementation does not adhere strictly to man page's interpretation of
SO_LINGER: shutdown() will not trigger the lingering. This follows AF_INET.
Signed-off-by: Michal Luczaj <m...@rbox.co>
---
net/vmw_vsock/virtio_transport_common.c | 13 +++++++++++--
1 file changed, 11 insertions(+), 2 deletions(-)
diff --git a/net/vmw_vsock/virtio_transport_common.c
b/net/vmw_vsock/virtio_transport_common.c
index
7f7de6d8809655fe522749fbbc9025df71f071bd..aeb7f3794f7cfc251dde878cb44fdcc54814c89c
100644
--- a/net/vmw_vsock/virtio_transport_common.c
+++ b/net/vmw_vsock/virtio_transport_common.c
@@ -1196,12 +1196,21 @@ static void virtio_transport_wait_close(struct sock
*sk, long timeout)
{
if (timeout) {
DEFINE_WAIT_FUNC(wait, woken_wake_function);
+ ssize_t (*unsent)(struct vsock_sock *vsk);
+ struct vsock_sock *vsk = vsock_sk(sk);
+
+ /* Some transports (Hyper-V, VMCI) do not implement
+ * unsent_bytes. For those, no lingering on close().
+ */
+ unsent = vsk->transport->unsent_bytes;
+ if (!unsent)
+ return;
IIUC if `unsent_bytes` is not implemented, virtio_transport_wait_close
basically does nothing. My concern is that we are breaking the
userspace due to a change in the behavior: Before this patch, with a
vmci/hyper-v transport, this function would wait for SOCK_DONE to be
set, but not anymore.
Wait, we are in virtio_transport_common.c, why we are talking about
Hyper-V and VMCI?
I asked to check `vsk->transport->unsent_bytes` in the v1, because this
code was part of af_vsock.c, but now we are back to virtio code, so I'm
confused...
Stefano
add_wait_queue(sk_sleep(sk), &wait);
do {
- if (sk_wait_event(sk, &timeout,
- sock_flag(sk, SOCK_DONE), &wait))
+ if (sk_wait_event(sk, &timeout, unsent(vsk) == 0,
+ &wait))
break;
} while (!signal_pending(current) && timeout);
-- 2.49.0
Thanks,
Luigi
