On Mon, Sep 28, 2020 at 1:34 AM Mark Mielke <mark.mielke@xxxxxxxxx> wrote:
>
> Kernel may fail to boot or devices may fail to come up when
> initializing iscsi_tcp devices starting with Linux 5.8.
>
> Marc Dionne identified the cause in RHBZ#1877345.
>
> Commit a79af8a64d39 ("[SCSI] iscsi_tcp: use iscsi_conn_get_addr_param
> libiscsi function") introduced getpeername() within the session spinlock.
>
> Commit 1b66d253610c ("bpf: Add get{peer, sock}name attach types for
> sock_addr") introduced BPF_CGROUP_RUN_SA_PROG_LOCK() within getpeername,
> which acquires a mutex and when used from iscsi_tcp devices can now lead
> to "BUG: scheduling while atomic:" and subsequent damage.
>
> This commit ensures that the spinlock is released before calling
> getpeername() or getsockname(). sock_hold() and sock_put() are
> used to ensure that the socket reference is preserved until after
> the getpeername() or getsockname() complete.
>
> Reported-by: Marc Dionne <marc.c.dionne@xxxxxxxxx>
> Link: https://bugzilla.redhat.com/show_bug.cgi?id=1877345
> Link: https://lkml.org/lkml/2020/7/28/1085
> Link: https://lkml.org/lkml/2020/8/31/459
> Fixes: a79af8a64d39 ("[SCSI] iscsi_tcp: use iscsi_conn_get_addr_param libiscsi function")
> Fixes: 1b66d253610c ("bpf: Add get{peer, sock}name attach types for sock_addr")
> Signed-off-by: Mark Mielke <mark.mielke@xxxxxxxxx>
> Cc: stable@xxxxxxxxxxxxxxx
> ---
> drivers/scsi/iscsi_tcp.c | 22 +++++++++++++++-------
> 1 file changed, 15 insertions(+), 7 deletions(-)
>
> diff --git a/drivers/scsi/iscsi_tcp.c b/drivers/scsi/iscsi_tcp.c
> index b5dd1caae5e9..d10efb66cf19 100644
> --- a/drivers/scsi/iscsi_tcp.c
> +++ b/drivers/scsi/iscsi_tcp.c
> @@ -736,6 +736,7 @@ static int iscsi_sw_tcp_conn_get_param(struct iscsi_cls_conn *cls_conn,
> struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
> struct iscsi_sw_tcp_conn *tcp_sw_conn = tcp_conn->dd_data;
> struct sockaddr_in6 addr;
> + struct socket *sock;
> int rc;
>
> switch(param) {
> @@ -747,13 +748,17 @@ static int iscsi_sw_tcp_conn_get_param(struct iscsi_cls_conn *cls_conn,
> spin_unlock_bh(&conn->session->frwd_lock);
> return -ENOTCONN;
> }
> + sock = tcp_sw_conn->sock;
> + sock_hold(sock->sk);
> + spin_unlock_bh(&conn->session->frwd_lock);
> +
> if (param == ISCSI_PARAM_LOCAL_PORT)
> - rc = kernel_getsockname(tcp_sw_conn->sock,
> + rc = kernel_getsockname(sock,
> (struct sockaddr *)&addr);
> else
> - rc = kernel_getpeername(tcp_sw_conn->sock,
> + rc = kernel_getpeername(sock,
> (struct sockaddr *)&addr);
> - spin_unlock_bh(&conn->session->frwd_lock);
> + sock_put(sock->sk);
> if (rc < 0)
> return rc;
>
> @@ -775,6 +780,7 @@ static int iscsi_sw_tcp_host_get_param(struct Scsi_Host *shost,
> struct iscsi_tcp_conn *tcp_conn;
> struct iscsi_sw_tcp_conn *tcp_sw_conn;
> struct sockaddr_in6 addr;
> + struct socket *sock;
> int rc;
>
> switch (param) {
> @@ -789,16 +795,18 @@ static int iscsi_sw_tcp_host_get_param(struct Scsi_Host *shost,
> return -ENOTCONN;
> }
> tcp_conn = conn->dd_data;
> -
> tcp_sw_conn = tcp_conn->dd_data;
> - if (!tcp_sw_conn->sock) {
> + sock = tcp_sw_conn->sock;
> + if (!sock) {
> spin_unlock_bh(&session->frwd_lock);
> return -ENOTCONN;
> }
> + sock_hold(sock->sk);
> + spin_unlock_bh(&session->frwd_lock);
>
> - rc = kernel_getsockname(tcp_sw_conn->sock,
> + rc = kernel_getsockname(sock,
> (struct sockaddr *)&addr);
> - spin_unlock_bh(&session->frwd_lock);
> + sock_put(sock->sk);
> if (rc < 0)
> return rc;
>
> --
> 2.28.0
>
Works for me and prevents the iscsid crash.
Tested-by: Marc Dionne <marc.c.dionne@xxxxxxxxx>
