Re: [PATCH net] vhost/vsock: annotate lockless accesses to guest_cid
From: Michael S. Tsirkin
Date: Wed Nov 18 2026 - 01:33:45 EST
On Wed, Nov 18, 2026 at 12:48:43AM +0800, Nguyen Dinh Phi wrote:
> vhost_vsock.guest_cid is written by vhost_vsock_set_cid() under
> vhost_vsock_mutex, but It is also read without that mutex in
> vhost_vsock_handle_tx_kick(), and vhost_vsock_get().
>
> KCSAN will report data-race if we issue VHOST_VSOCK_SET_GUEST_CID
> ioctl while the device has TX traffic. For example:
>
> write to 0xffff88810f1596ac of 4 bytes by task 7914 on cpu 1:
> vhost_vsock_dev_ioctl+0x25f/0x720
> __x64_sys_ioctl+0xc2/0x110
> x64_sys_call+0x299/0x20a0
> do_syscall_64+0xc9/0x510
> entry_SYSCALL_64_after_hwframe+0x76/0x7e
>
> read to 0xffff88810f1596ac of 4 bytes by task 7910 on cpu 3:
> vhost_vsock_handle_tx_kick+0x41b/0x740
> vhost_run_work_list+0x97/0xc0
> vhost_task_fn+0x98/0x1f0
> ret_from_fork+0x2e2/0x3c0
> ret_from_fork_asm+0x1a/0x30
>
> value changed: 0x0000004d -> 0x00000032
>
> Both readers load the CID once and works correctly with either the
> old or the new value.
>
> Mark the store with WRITE_ONCE() and the two lockless loads with
> READ_ONCE(), so KCSAN knows the race is intentional.
You make it sounds like it is just KCSAN false positive.
Is it?
if yes, this should be data_race. If no, document why.
>
> Fixes: 433fc58e6bf2 ("VSOCK: Introduce vhost_vsock.ko")
> Signed-off-by: Nguyen Dinh Phi <phind.uet@xxxxxxxxx>
> ---
> drivers/vhost/vsock.c | 6 +++---
> 1 file changed, 3 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/vhost/vsock.c b/drivers/vhost/vsock.c
> index 88a310390eb0..f26c59013075 100644
> --- a/drivers/vhost/vsock.c
> +++ b/drivers/vhost/vsock.c
> @@ -78,7 +78,7 @@ static struct vhost_vsock *vhost_vsock_get(u32 guest_cid, struct net *net)
>
> hash_for_each_possible_rcu(vhost_vsock_hash, vsock, hash, guest_cid,
> lockdep_is_held(&vhost_vsock_mutex)) {
> - u32 other_cid = vsock->guest_cid;
> + u32 other_cid = READ_ONCE(vsock->guest_cid);
>
> /* Skip instances that have no CID yet */
> if (other_cid == 0)
> @@ -572,7 +572,7 @@ static void vhost_vsock_handle_tx_kick(struct vhost_work *work)
> hdr = virtio_vsock_hdr(skb);
>
> /* Only accept correctly addressed packets */
> - if (le64_to_cpu(hdr->src_cid) == vsock->guest_cid &&
> + if (le64_to_cpu(hdr->src_cid) == READ_ONCE(vsock->guest_cid) &&
> le64_to_cpu(hdr->dst_cid) ==
> vhost_transport_get_local_cid())
> virtio_transport_recv_pkt(&vhost_transport, skb,
Will prevent compiler from hoisting the reads out of the loops.
Does it matter?
> @@ -859,7 +859,7 @@ static int vhost_vsock_set_cid(struct vhost_vsock *vsock, u64 guest_cid)
> if (vsock->guest_cid)
> hash_del_rcu(&vsock->hash);
>
> - vsock->guest_cid = guest_cid;
> + WRITE_ONCE(vsock->guest_cid, guest_cid);
> hash_add_rcu(vhost_vsock_hash, &vsock->hash, vsock->guest_cid);
> mutex_unlock(&vhost_vsock_mutex);
>
> --
> 2.53.0