On Thu, Aug 13, 2026 at 09:12:36PM +0900, Daehyeon Ko wrote:
virtio_transport_recv_pkt() first looks up a socket using the full source
and destination tuple. If that misses, it falls back to a bound-socket
lookup using only the destination address. The fallback is needed for
listening and connecting sockets, but it can also select an established
socket that remains in the bound table.
As a result, a packet from an unrelated source can be dispatched to a
non-listening socket. In TCP_SYN_SENT, a source-blind RESPONSE marks the
selected socket established while retaining its original remote address.
Subsequent RW packets can likewise be delivered through the
destination-only fallback.
This was reproduced with two capless processes under different UIDs. The
attacker discovered the victim tuple through unprivileged AF_VSOCK
sock_diag and injected a chosen 16-byte payload into the victim established
loopback socket. The legitimate peer received none of those bytes.
I don't understand this, what it means? If the receiver doesn't receive
those injected bytes, should be fine, no?
After taking the socket lock, verify that packets for non-listening sockets
come from the peer stored in remote_addr. Listening sockets continue to
accept packets from any source.
IMO this description should be improved, it's quite hard to follow.
This is an idea IIUC the issue:
virtio_transport_recv_pkt() looks up sockets in two steps: first by
the full source and destination tuple in the connected table, then by
destination only in the bound table. The fallback is needed for
listening and connecting sockets, but it never validates the source
address in the packet header. Since sockets remain in the bound table
after connect(), a packet from any source can be dispatched to a
non-listening socket.
An unprivileged process can discover the target tuple through AF_VSOCK
sock_diag (no capabilities required), send a forged RESPONSE to a
TCP_SYN_SENT socket to establish the connection, and then inject
arbitrary data through subsequent RW packets, all delivered via the
destination-only fallback.
After lock_sock(), verify that the packet source matches the
remote_addr stored during connect().
Fixes: 06a8fc78367d ("VSOCK: Introduce virtio_vsock_common.ko")
Cc: [email protected]
Signed-off-by: Daehyeon Ko <[email protected]>
---
net/vmw_vsock/virtio_transport_common.c | 10 +++++++---
1 file changed, 7 insertions(+), 3 deletions(-)
diff --git a/net/vmw_vsock/virtio_transport_common.c
b/net/vmw_vsock/virtio_transport_common.c
index 8becad812..f73e0406a 100644
--- a/net/vmw_vsock/virtio_transport_common.c
+++ b/net/vmw_vsock/virtio_transport_common.c
@@ -1822,11 +1822,15 @@ void virtio_transport_recv_pkt(struct virtio_transport
*t,
lock_sock(sk);
- /* Check if sk has been closed or assigned to another transport before
- * lock_sock (note: listener sockets are not assigned to any transport)
+ /* Check if sk has been closed, assigned to another transport, or if the
+ * packet is from a different peer than the one connected to sk. These
+ * properties could have changed before lock_sock. Listener sockets are
+ * not assigned to any transport and accept packets from any peer.
Please keeps changes minimal, why adding "These properties could have
changed before lock_sock." and changing the phrase related to listener
socket?
Also the new part is not clear IMO, we should explain why we are adding
this new check, not what we are doing which is clear looking at the
code.
IMO we should just add something like this (feel free to change it):
Also verify the packet source: the bound-table fallback lookup
matches by destination only, so a non-listening socket can be
reached by a packet from an unrelated peer. Drop it unless the
source matches the peer the socket is connecting or connected to.
*/
if (sock_flag(sk, SOCK_DONE) ||
- (sk->sk_state != TCP_LISTEN && vsk->transport != &t->transport)) {
+ (sk->sk_state != TCP_LISTEN &&
+ (vsk->transport != &t->transport ||
+ !vsock_addr_equals_addr(&src, &vsk->remote_addr)))) {
Do we need to add the same check also in the VMCI transport?
Hyper-V seems different.
Thanks,
Stefano
(void)virtio_transport_reset_no_sock(t, skb, net);
release_sock(sk);
sock_put(sk);
base-commit: 9006c116dd111d457bf5d074990210f70a4ad2c8
--
2.54.0
