I'm attaching the patch fixing one more problem noticed with non-zero nop_interval: There is a couple of problems when a connection is closed by iscsi_tcp_nop_work_handler, and there are uncompleted scsi commands for that connection: 1) usr/iscsi/iscsid.c:iscsi_scsi_cmd_done assumes that a closed connection has .state == STATE_CLOSE. The assumption fails if the connection is closed not as a reaction to STATE_CLOSE, but by a raw conn_close call outside of rx/tx handlers (like it happens for unanswered NOP-INs). This commit takes care to set .state = STATE_CLOSE in ep_close callback, so incomplete scsi commands will be freed on completion. 2) the connection doesn't go away from iscsi_tcp_conn_list until its refcount reaches zero (see iscsi_tcp_release). It makes iscsi_tcp_nop_work_handler call conn_close for the same connection repeatedly, as long as it has incomplete scsi commands (hence non-zero refcount). This commit zeroes out nop_interval for closed connections, so next iscsi_tcp_nop_work_handler calls won't try to close the connection any more.
From ac1b732b612fe2c188dcc2441e5dd247d3452836 Mon Sep 17 00:00:00 2001 From: Anton Kovalenko <anton.kovalenko@xxxxxxxxxxx> Date: Tue, 30 Aug 2016 19:02:51 +0300 Subject: [PATCH] Disable NOP-IN and adjust state for closed conn There is a couple of problems when a connection is closed by iscsi_tcp_nop_work_handler, and there are uncompleted scsi commands for that connection: 1) usr/iscsi/iscsid.c:iscsi_scsi_cmd_done assumes that a closed connection has .state == STATE_CLOSE. The assumption fails if the connection is closed not as a reaction to STATE_CLOSE, but by a raw conn_close call outside of rx/tx handlers (like it happens for unanswered NOP-INs). This commit takes care to set .state = STATE_CLOSE in ep_close callback, so incomplete scsi commands will be freed on completion. 2) the connection doesn't go away from iscsi_tcp_conn_list until its refcount reaches zero (see iscsi_tcp_release). It makes iscsi_tcp_nop_work_handler call conn_close for the same connection repeatedly, as long as it has incomplete scsi commands (hence non-zero refcount). This commit zeroes out nop_interval for closed connections, so next iscsi_tcp_nop_work_handler calls won't try to close the connection any more. --- usr/iscsi/iscsi_tcp.c | 2 ++ 1 file changed, 2 insertions(+) diff --git a/usr/iscsi/iscsi_tcp.c b/usr/iscsi/iscsi_tcp.c index 4cc2cf4..536f22e 100644 --- a/usr/iscsi/iscsi_tcp.c +++ b/usr/iscsi/iscsi_tcp.c @@ -508,6 +508,8 @@ static size_t iscsi_tcp_close(struct iscsi_connection *conn) struct iscsi_tcp_connection *tcp_conn = TCP_CONN(conn); tgt_event_del(tcp_conn->fd); + conn->state = STATE_CLOSE; + tcp_conn->nop_interval = 0; return 0; } -- 2.9.3
