iscsi_tcp: do not bind sockets that already have extra callbacks
PATCH v2 1/2 iscsi_tcp: do not bind sockets that already have extra callbacks
This attempts to avoid a situation where a misbehaving iscsi daemon passes a socket for a different iSCSI connection to BIND_CONN - which would result in infinite recursion and stack overflow. This will also prevent passing *other* sockets which had sk_user_data overridden, but that wouldn't have been safe anyways - since we throw away that pointer anyways. This does not cover all hypothetical scenarios where we pass bad sockets to BIND_CONN.
This also papers over a different bug - we allow a daemon to call BIND_CONN twice for the same connection - which would result in, at the least, failing to uninitialize/teardown the previous socket, which will be addressed separately.
--- drivers/scsi/iscsi_tcp.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/scsi/iscsi_tcp.c +1 -1
@@ -725,7 +725,7 @@ iscsi_sw_tcp_conn_bind(struct iscsi_cls_session *cls_session, } err = -EINVAL ; -if (!sk_is_tcp (sock->sk)) +if (!sk_is_tcp (sock->sk) || sock->sk->sk_user_data) goto free_socket; err =iscsi_conn_bind (cls_session, cls_conn, is_leading);
Signature · 1 line
2.45.1.288.g0e0cd299f1-goog
PATCH v2 2/2 libiscsi: disallow binding an already-bound connection
This fixes issue where misbehaving userspace initiator could bind the same connection multiple times, which would leak the old connection socket without cleaning it up.
For iscsi_tcp, it calls iscsi_suspend_tx directly in stop_conn. Update this to iscsi_conn_unbind, which matches the lifecycle of other drivers, and clears the CONN_FLAG_BOUND.
--- drivers/scsi/iscsi_tcp.c | 2 +- drivers/scsi/libiscsi.c | 6 ++++++ 2 files changed, 7 insertions(+), 1 deletion(-)
drivers/scsi/iscsi_tcp.c +1 -1
@@ -696,7 +696,7 @@ static void iscsi_sw_tcp_conn_stop(struct iscsi_cls_conn *cls_conn, int flag)wake_up_interruptible (sk_sleep (sock->sk));/* stop xmit side */ -iscsi_suspend_tx (conn); +iscsi_conn_unbind (cls_conn, true);/* stop recv side and release socket */ iscsi_sw_tcp_release_conn (conn);
drivers/scsi/libiscsi.c +6 -0
@@ -3453,6 +3453,12 @@ int iscsi_conn_bind(struct iscsi_cls_session *cls_session,struct iscsi_conn *conn = cls_conn->dd_data;spin_lock_bh (&session->frwd_lock); +if (test_bit (ISCSI_CONN_FLAG_BOUND , &conn->flags)) { +spin_unlock_bh (&session->frwd_lock); +return -EBUSY ; + } + +if (is_leading) session->leadconn = conn;
Signature · 1 line
2.45.1.288.g0e0cd299f1-goog
Mike Christie
maintainerRe: [PATCH v2 2/2] libiscsi: disallow binding an already-bound connection
Quoted 33 lines · Khazhismel Kumykov
This fixes issue where misbehaving userspace initiator could bind the same connection multiple times, which would leak the old connection socket without cleaning it up.
For
iscsi_tcp, it callsiscsi_suspend_txdirectly instop_conn. Update this toiscsi_conn_unbind, which matches the lifecycle of other drivers, and clears theCONN_FLAG_BOUND.Suggested-by Mike Christie <[email protected]>Signed-off-by Khazhismel Kumykov <[email protected]>--- drivers/scsi/iscsi_tcp.c | 2 +- drivers/scsi/libiscsi.c | 6 ++++++ 2 files changed, 7 insertions(+), 1 deletion(-)drivers/scsi/iscsi_tcp.c +1
-1@@ -696,7 +696,7 @@ static void iscsi_sw_tcp_conn_stop(struct iscsi_cls_conn *cls_conn, int flag)wake_up_interruptible (sk_sleep (sock->sk));/* stop xmit side */ -iscsi_suspend_tx (conn); +iscsi_conn_unbind (cls_conn, true);/* stop recv side and release socket */ iscsi_sw_tcp_release_conn (conn);drivers/scsi/libiscsi.c +0
-0
drivers/scsi/libiscsi.c int iscsi_conn_bind(struct iscsi_cls_session *cls_session, · line 3460
Quoted 8 lines
drivers/scsi/libiscsi.c +5
-0@@ -3453,6 +3453,12 @@ int iscsi_conn_bind(struct iscsi_cls_session *cls_session,struct iscsi_conn *conn = cls_conn->dd_data;spin_lock_bh (&session->frwd_lock); +if (test_bit (ISCSI_CONN_FLAG_BOUND , &conn->flags)) { +spin_unlock_bh (&session->frwd_lock); +return -EBUSY ; + } +
If some of the driver's bind_conn callout fails then it could leave the ISCSI_CONN_FLAG_BOUND bit set, and this will fail on a retry.
It's ok to call iscsi_conn_bind at the end of the bind_conn callout like how iscsi_tcp and cxgbi* do. So for iscsi_iser.c, be_iscsi.c, bnx2i_iscsi.c, and qedi_iscsi.c you need to move their iscsi_conn_bind to the end of their bind_conn callout.
You could also keep the the existing flow but add unwind/goto error handling to those drivers, but I think just moving the iscsi_conn_bind call to the end of the function is easier and it syncs all the drivers behavior which is nice.