Comment 85 for bug 1569925

Long story short -> https://goo.gl/vPyh8C

Basically the block device enqueues the last request (a SYNC scsi command coming from sd_shutdown) for every scsi device there is on the system. Unfortunately, since the OS is shutting down, in between the block request and its execution, we have userland (systemd) killing iscsid, without proper logout, and/or removing the network.

What happens next is that the mid layer (SCSI) tries to deliver the request through the transport layer (iscsi_tcp_sw) but it fails since the transport layer checks the session status and finds out that the session is not in LOGIN state.

The default behaviour of the transport layer (iscsi_tcp_sw) in such situation is to tell the mid-layer to keep resetting the request timeout timer while it tries to recover (something that will never happen because the network is gone).

Changing that default behavior to state that the scsi command was NOT handled by the transport layer (iscsi_tcp_sw) implies in making the scsi timeout function to try to "abort" the scsi command, which also creates other commands that will timeout because of the transport layer.

Best scenario so far was to change BLK_EH_NOT_HANDLED for BLK_EH_HANDLED in the scsi_times_out function and make the kernel to be able to shutdown. By doing that, I'm confirming to block device something that DID NOT happen, meaning that the command never left the transport layer.

This might be ONE of possible ways to fix this: I can mark in the transport layer that I have timed-out DURING the shutdown procedure, cancelling all the block device requests without having to invoke the scsi error handling mechanism, generating more traffic in transport layer (what would also cause more timeouts, causing a loop in the problematic sequence).

Anyway, I'll get back to this next week and hopefully identify best course of action.