| Message ID | 20210325093248.284678-1-gulam.mohamed@oracle.com (mailing list archive) |
|---|---|
| State | Accepted |
| Headers | show |
| Series | [V5] iscsi: Do Not set param when sock is NULL | expand |
On 3/25/21 4:32 AM, Gulam Mohamed wrote: > Description > =========== > 1. This Kernel panic could be due to a timing issue when there is a > race between the sync thread and the initiator was processing of > a login response from the target. The session re-open can be invoked > from two places: > a. Sessions sync thread when the iscsid restart > b. From iscsid through iscsi error handler > 2. The session reopen sequence is as follows in user-space > a. Disconnect the connection > b. Then send the stop connection request to the kernel > which releases the connection (releases the socket) > c. Queues the reopen for 2 seconds delay > d. Once the delay expires, create the TCP connection again by > calling the connect() call > e. Poll for the connection > f. When poll is successful i.e when the TCP connection is > established, it performs: > i. Creation of session and connection data structures > ii. Bind the connection to the session. This is the place > where we assign the sock to tcp_sw_conn->sock > iii. Sets parameters like target name, persistent address > iv. Creates the login pdu > v. Sends the login pdu to kernel > vi. Returns to the main loop to process further events. > The kernel then sends the login request over to the > target node > g. Once login response with success is received, it enters full > feature phase and sets the negotiable parameters like > max_recv_data_length,max_transmit_length, data_digest etc. > 3. While setting the negotiable parameters by calling > "iscsi_session_set_neg_params()", kernel panicked as sock was NULL > > What happened here is > --------------------- > 1. Before initiator received the login response mentioned in above > point 2.f.v, another reopen request was sent from the error > handler/sync session for the same session, as the initiator utils > was in main loop to process further events (as mentioned in point > 2.f.vi above). > 2. While processing this reopen, it stopped the connection which > released the socket and queued this connection and at this point > of time the login response was received for the earlier one > > Fix > --- > 1. Add new connection state ISCSI_CONN_BOUND in "enum iscsi_connection > _state" > 2. Set the connection state value to ISCSI_CONN_DOWN upon > iscsi_if_ep_disconnect() and iscsi_if_stop_conn() > 3. Set the connection state to the newly created value ISCSI_CONN_BOUND > after bind connection (transport->bind_conn()) > 4. In iscsi_set_param, return -ENOTCONN if the connection state is not > either ISCSI_CONN_BOUND or ISCSI_CONN_UP > > Signed-off-by: Gulam Mohamed <gulam.mohamed@oracle.com> > Reviewed-by: Mike Christie <michael.christie@oracle.com>
On Thu, 25 Mar 2021 09:32:48 +0000, Gulam Mohamed wrote: > Description > =========== > 1. This Kernel panic could be due to a timing issue when there is a > race between the sync thread and the initiator was processing of > a login response from the target. The session re-open can be invoked > from two places: > a. Sessions sync thread when the iscsid restart > b. From iscsid through iscsi error handler > 2. The session reopen sequence is as follows in user-space > a. Disconnect the connection > b. Then send the stop connection request to the kernel > which releases the connection (releases the socket) > c. Queues the reopen for 2 seconds delay > d. Once the delay expires, create the TCP connection again by > calling the connect() call > e. Poll for the connection > f. When poll is successful i.e when the TCP connection is > established, it performs: > i. Creation of session and connection data structures > ii. Bind the connection to the session. This is the place > where we assign the sock to tcp_sw_conn->sock > iii. Sets parameters like target name, persistent address > iv. Creates the login pdu > v. Sends the login pdu to kernel > vi. Returns to the main loop to process further events. > The kernel then sends the login request over to the > target node > g. Once login response with success is received, it enters full > feature phase and sets the negotiable parameters like > max_recv_data_length,max_transmit_length, data_digest etc. > 3. While setting the negotiable parameters by calling > "iscsi_session_set_neg_params()", kernel panicked as sock was NULL > > [...] Applied to 5.12/scsi-fixes, thanks! [1/1] iscsi: Do Not set param when sock is NULL https://git.kernel.org/mkp/scsi/c/9e67600ed6b8
diff --git a/drivers/scsi/scsi_transport_iscsi.c b/drivers/scsi/scsi_transport_iscsi.c index 91074fd97f64..f4bf62b007a0 100644 --- a/drivers/scsi/scsi_transport_iscsi.c +++ b/drivers/scsi/scsi_transport_iscsi.c @@ -2475,6 +2475,7 @@ static void iscsi_if_stop_conn(struct iscsi_cls_conn *conn, int flag) */ mutex_lock(&conn_mutex); conn->transport->stop_conn(conn, flag); + conn->state = ISCSI_CONN_DOWN; mutex_unlock(&conn_mutex); } @@ -2901,6 +2902,13 @@ iscsi_set_param(struct iscsi_transport *transport, struct iscsi_uevent *ev) default: err = transport->set_param(conn, ev->u.set_param.param, data, ev->u.set_param.len); + if ((conn->state == ISCSI_CONN_BOUND) || + (conn->state == ISCSI_CONN_UP)) { + err = transport->set_param(conn, ev->u.set_param.param, + data, ev->u.set_param.len); + } else { + return -ENOTCONN; + } } return err; @@ -2960,6 +2968,7 @@ static int iscsi_if_ep_disconnect(struct iscsi_transport *transport, mutex_lock(&conn->ep_mutex); conn->ep = NULL; mutex_unlock(&conn->ep_mutex); + conn->state = ISCSI_CONN_DOWN; } transport->ep_disconnect(ep); @@ -3727,6 +3736,8 @@ iscsi_if_recv_msg(struct sk_buff *skb, struct nlmsghdr *nlh, uint32_t *group) ev->r.retcode = transport->bind_conn(session, conn, ev->u.b_conn.transport_eph, ev->u.b_conn.is_leading); + if (!ev->r.retcode) + conn->state = ISCSI_CONN_BOUND; mutex_unlock(&conn_mutex); if (ev->r.retcode || !transport->ep_connect) @@ -3966,7 +3977,8 @@ iscsi_conn_attr(local_ipaddr, ISCSI_PARAM_LOCAL_IPADDR); static const char *const connection_state_names[] = { [ISCSI_CONN_UP] = "up", [ISCSI_CONN_DOWN] = "down", - [ISCSI_CONN_FAILED] = "failed" + [ISCSI_CONN_FAILED] = "failed", + [ISCSI_CONN_BOUND] = "bound" }; static ssize_t show_conn_state(struct device *dev, diff --git a/include/scsi/scsi_transport_iscsi.h b/include/scsi/scsi_transport_iscsi.h index 8a26a2ffa952..fc5a39839b4b 100644 --- a/include/scsi/scsi_transport_iscsi.h +++ b/include/scsi/scsi_transport_iscsi.h @@ -193,6 +193,7 @@ enum iscsi_connection_state { ISCSI_CONN_UP = 0, ISCSI_CONN_DOWN, ISCSI_CONN_FAILED, + ISCSI_CONN_BOUND, }; struct iscsi_cls_conn {
Description =========== 1. This Kernel panic could be due to a timing issue when there is a race between the sync thread and the initiator was processing of a login response from the target. The session re-open can be invoked from two places: a. Sessions sync thread when the iscsid restart b. From iscsid through iscsi error handler 2. The session reopen sequence is as follows in user-space a. Disconnect the connection b. Then send the stop connection request to the kernel which releases the connection (releases the socket) c. Queues the reopen for 2 seconds delay d. Once the delay expires, create the TCP connection again by calling the connect() call e. Poll for the connection f. When poll is successful i.e when the TCP connection is established, it performs: i. Creation of session and connection data structures ii. Bind the connection to the session. This is the place where we assign the sock to tcp_sw_conn->sock iii. Sets parameters like target name, persistent address iv. Creates the login pdu v. Sends the login pdu to kernel vi. Returns to the main loop to process further events. The kernel then sends the login request over to the target node g. Once login response with success is received, it enters full feature phase and sets the negotiable parameters like max_recv_data_length,max_transmit_length, data_digest etc. 3. While setting the negotiable parameters by calling "iscsi_session_set_neg_params()", kernel panicked as sock was NULL What happened here is --------------------- 1. Before initiator received the login response mentioned in above point 2.f.v, another reopen request was sent from the error handler/sync session for the same session, as the initiator utils was in main loop to process further events (as mentioned in point 2.f.vi above). 2. While processing this reopen, it stopped the connection which released the socket and queued this connection and at this point of time the login response was received for the earlier one Fix --- 1. Add new connection state ISCSI_CONN_BOUND in "enum iscsi_connection _state" 2. Set the connection state value to ISCSI_CONN_DOWN upon iscsi_if_ep_disconnect() and iscsi_if_stop_conn() 3. Set the connection state to the newly created value ISCSI_CONN_BOUND after bind connection (transport->bind_conn()) 4. In iscsi_set_param, return -ENOTCONN if the connection state is not either ISCSI_CONN_BOUND or ISCSI_CONN_UP Signed-off-by: Gulam Mohamed <gulam.mohamed@oracle.com> --- Changes in V5: - Corrected the indentation and included the brackets for the else block in iscsi_set_param() function Changes in V4: - Included ISCSI_CONN_UP check in the if condition when the function iscsi_set_param() checks to see if iscsi parameter can be set or not to make it work for offload drivers also. Earlier version use to check only the ISCSI_CONN_BOUND Changes in V3: - Changed the approach from creating and checking a flag in iscsi_cls_conn structure, to a new connection state ISCSI_CONN_BOUND and checking it while setting the iscsi parameters drivers/scsi/scsi_transport_iscsi.c | 14 +++++++++++++- include/scsi/scsi_transport_iscsi.h | 1 + 2 files changed, 14 insertions(+), 1 deletion(-)