Re: [PATCH net v2] vhost/net: Fix receive queue test in busy poll
From: Eugenio Perez Martin
Date: Fri Nov 20 2026 - 07:04:41 EST
On Fri, Nov 20, 2026 at 4:39 AM raoxu <raoxu@xxxxxxxxxxxxx> wrote:
>
> From: Xu Rao <raoxu@xxxxxxxxxxxxx>
>
> sock_has_rx_data() returns a positive value when receive data is
> available. For sockets without a peek_len() callback, however, its
> fallback returns skb_queue_empty(), which has the opposite meaning.
>
> An AF_PACKET SOCK_RAW socket is accepted by get_raw_socket() as a
> vhost-net backend, and its socket operations do not provide peek_len().
> With busy polling enabled for this backend, an empty receive queue is
> treated as ready and a nonempty queue as not ready. This can make
> vhost_net_busy_poll() stop polling prematurely or fail to stop for
> queued receive data, and invert its receive-side scheduling decision.
> TUN and TAP backends provide peek_len() and are unaffected.
>
> Negate skb_queue_empty() so the fallback matches the callback and the
> readiness checks in vhost_net_busy_poll().
>
> The issue was found during AI-assisted code review. The helper's
> incorrect result was confirmed on x86_64 running 6.6.0-amd64-desktop
> using a temporary diagnostic module containing the original and fixed
> helpers. An AF_PACKET SOCK_RAW socket, without PACKET_RX_RING, was
> bound to loopback in a private network namespace. Across three cycles
> of sending and receiving a 60-byte frame, the original helper returned
> 1 for an empty queue and 0 for a queued frame. The fixed helper returned
> 0 and 1 respectively, matching socket readability in all nine checks.
> This tests helper copies on real sockets, not the running vhost-net
> busy-poll path; the vhost-net module was not replaced.
>
> The patched drivers/vhost/net.o was also build-tested on the net tree
> with x86_64 GCC 12.3 and W=1 under allmodconfig and allyesconfig.
> No end-to-end vhost tests or kernel selftests were run.
>
> Fixes: 1576d9860599 ("tun: switch to use skb array for tx")
> Cc: stable@xxxxxxxxxxxxxxx
Acked-by: Eugenio Pérez <eperezma@xxxxxxxxxx>
> Signed-off-by: Xu Rao <raoxu@xxxxxxxxxxxxx>
> ---
> Changes in v2:
> - Target net and add the introducing commit's Fixes tag.
> - Explain the raw packet socket backend and busy-poll decisions.
> - State the AI-assisted discovery method and document the diagnostic
> module test, its results and its helper-level coverage.
> - Document the targeted build tests.
> - Keep the code change unchanged from v1.
>
> v1: https://lore.kernel.org/virtualization/20261110051642.249640-1-raoxu@xxxxxxxxxxxxx/
>
> AI assistance: OpenCode was used to inspect the history, prepare and run
> the diagnostic module test, run build checks, and revise this message.
> The code change is unchanged from v1. The discovery method was confirmed
> by the author.
>
> drivers/vhost/net.c | 2 +-
> 1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/drivers/vhost/net.c b/drivers/vhost/net.c
> index 4570a4ffbab2..10ed2d55ff07 100644
> --- a/drivers/vhost/net.c
> +++ b/drivers/vhost/net.c
> @@ -538,7 +538,7 @@ static int sock_has_rx_data(struct socket *sock)
> if (sock->ops->peek_len)
> return sock->ops->peek_len(sock);
>
> - return skb_queue_empty(&sock->sk->sk_receive_queue);
> + return !skb_queue_empty(&sock->sk->sk_receive_queue);
> }
>
> static void vhost_net_busy_poll_try_queue(struct vhost_net *net,
>
> base-commit: 435cbdaf9b03687498dc16fe4af8d16a977e21d0
> --
> 2.50.1
>