nvme-tcp: i/o errors and stalled host from failure to send command pdus

[Sagi Grimberg]

> Question: queue->more_requests is updated under a lock. However, not
> all readers of queue->more_requests acquire the lock. The smp_id/smp_id2
> values seem to show that the same queue was being processed on
> different CPUs. Is this a race between the kworker and inline
> submission?

The entire send path in io_work is wrapped with the send_mutex. Which
part is not protected with the lock?