Re: [PATCH] scsi: iscsi_tcp: Don't read sk_user_data from the recv work

From: Mike Christie

Date: Mon Nov 16 2026 - 13:12:36 EST


On 11/10/26 4:24 PM, Praneeth Balasubramanian wrote:
> With recv_from_iscsi_q=1, iscsi_sw_tcp_data_ready() queues conn->recvwork, and iscsi_sw_tcp_recv_data_work() runs the receive path under lock_sock() only. At the end of iscsi_sw_tcp_recv_data(), iscsi_sw_sk_state_check() reads the connection
>
>
> With recv_from_iscsi_q=1, iscsi_sw_tcp_data_ready() queues
> conn->recvwork, and iscsi_sw_tcp_recv_data_work() runs the receive path
> under lock_sock() only. At the end of iscsi_sw_tcp_recv_data(),
> iscsi_sw_sk_state_check() reads the connection from sk->sk_user_data.
> Its comment says it must be called with sk_callback_lock, which the work
> does not hold.
>
> On teardown, iscsi_sw_tcp_release_conn() first calls
> iscsi_sw_tcp_conn_restore_callbacks(), which sets sk->sk_user_data to
> NULL, and only then flushes recvwork in iscsi_suspend_rx(). If queued
> work runs in between and the socket is in TCP_CLOSE or TCP_CLOSE_WAIT,
> iscsi_sw_sk_state_check() dereferences NULL->session. With a delay
> added at the start of the work to widen this window, a target-side
> connection close gives:
>
> Oops: general protection fault, probably for non-canonical address ...
> KASAN: null-ptr-deref in range [0x0000000000000010-0x0000000000000017]
> Workqueue: iscsi_q_3 iscsi_sw_tcp_recv_data_work
> RIP: 0010:iscsi_sw_tcp_recv_data+0x219/0x420
>
> Both callers already have the connection. iscsi_sw_tcp_recv_data() gets
> it as an argument, and iscsi_sw_tcp_state_change() reads it under
> sk_callback_lock. Pass it in instead of reading sk_user_data again. The
> connection stays valid until iscsi_suspend_rx() has flushed the work.
>
> The unlocked read was found with an LLM-based tool that checks callers
> against the locking rules stated in kernel comments.
>
> Fixes: f1d269765ee2 ("scsi: iscsi: Run recv path from workqueue")
> Cc: stable@xxxxxxxxxxxxxxx
> Signed-off-by: Praneeth Balasubramanian <praneeth@xxxxxxxx>
> ---
>
> Notes:
> This was found by an LLM-based static analysis (GPT-5.6) that checks
> kernel code against the rules stated in its comments, as part of research
> at UCSB SecLab. The patch and changelog were drafted with Claude Opus 5.5.
> I reviewed the patch and take responsibility for it.
>
> recv_from_iscsi_q is off by default, and setting up the session needs
> root on the initiator, so I am sending this as a regular fix rather than
> a security report.
>
> Testing: two QEMU VMs on next-20261023, an LIO target and an open-iscsi
> initiator booted with iscsi_tcp.recv_from_iscsi_q=1, using CHAP and
> continuous reads. The initiator kernel had a 20 ms msleep() added at the
> start of iscsi_sw_tcp_recv_data_work(). Closing the TCP connection from
> the target side produced the oops above. I have only seen it with the
> added delay.
>
> The patch applies to mainline 9ba46df34488 and to next-20261109.
>
> drivers/scsi/iscsi_tcp.c | 14 ++++++++------
> 1 file changed, 8 insertions(+), 6 deletions(-)
>
> diff --git a/drivers/scsi/iscsi_tcp.c b/drivers/scsi/iscsi_tcp.c
> index 9260b1c9b0e0..88b1f255ae9b 100644
> --- a/drivers/scsi/iscsi_tcp.c
> +++ b/drivers/scsi/iscsi_tcp.c
> @@ -104,18 +104,20 @@ static int iscsi_sw_tcp_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
>
> /**
> * iscsi_sw_sk_state_check - check socket state
> + * @conn: iscsi connection the socket belongs to
> * @sk: socket
> *
> * If the socket is in CLOSE or CLOSE_WAIT we should
> * not close the connection if there is still some
> * data pending.
> *
> - * Must be called with sk_callback_lock.
> + * The caller passes @conn rather than having it read from
> + * sk->sk_user_data. The recv work runs without sk_callback_lock,
> + * and connection teardown may already have cleared sk_user_data.
> */
> -static inline int iscsi_sw_sk_state_check(struct sock *sk)
> +static inline int iscsi_sw_sk_state_check(struct iscsi_conn *conn,
> + struct sock *sk)
> {
> - struct iscsi_conn *conn = sk->sk_user_data;
> -
> if ((sk->sk_state == TCP_CLOSE_WAIT || sk->sk_state == TCP_CLOSE) &&
> (conn->session->state != ISCSI_STATE_LOGGING_OUT) &&
> !atomic_read(&sk->sk_rmem_alloc)) {
> @@ -148,7 +150,7 @@ static void iscsi_sw_tcp_recv_data(struct iscsi_conn *conn)
> * unmap it now. */
> iscsi_tcp_segment_unmap(&tcp_conn->in.segment);
>
> - iscsi_sw_sk_state_check(sk);
> + iscsi_sw_sk_state_check(conn, sk);
> }
>
> static void iscsi_sw_tcp_recv_data_work(struct work_struct *work)
> @@ -202,7 +204,7 @@ static void iscsi_sw_tcp_state_change(struct sock *sk)
> return;
> }
>
> - iscsi_sw_sk_state_check(sk);
> + iscsi_sw_sk_state_check(conn, sk);

I think after we have done iscsi_sw_tcp_conn_restore_callbacks it's
a bug to be accessing sk_user_data from inside iscsi_tcp and expect
that it's non-NULL.

We don't want to allow that to happen so we should take the
sk_callback_lock when callling iscsi_sw_tcp_recv_data from
iscsi_sw_tcp_recv_data_work like is done from iscsi_sw_tcp_data_ready.
Or, call iscsi_suspend_rx before iscsi_sw_tcp_conn_restore_callbacks
and then add some code so after iscsi_suspend_rx has returned we
know that racing iscsi_conn_queue_recv calls are handled.