nvmet-tcp timeout issue

[Engel, Amit]

Hello Sagi et al.

We see a potential bug that leads to nvmet-tcp timeout in some cases

It appears that there are cases where kernel_sendpage() fails and nvmet_tcp_write_space() is never called
This leads to a timeout due to a 'stuck' IO

Based on the code, as long as SOCK_NOSPACE is set, upon kernel_sendpage() failure, nvmet_tcp_write_space callback is being called
And the io_work will be scheduled (queue_work_on) - NO issue
But we found a case where kernel_sendpage() fails and SOCK_NOSPACE bit is not being set.
In this case, we believe that nvmet_tcp_write_space() is never called. This will lead to a IO timeout.

In addition, it appears that some scenarios will lead to unnecessary cpu stress:
In case that nvmet_try_send_response() returns with 0
nvmet_tcp_try_send_one() will return 1 as well as nvmet_tcp_try_send()
it means that 'pending' flag in nvmet_tcp_io_work is set to 'true' and will cause unnecessary cpu stress

Will be glad to hear your feedback for both of the above issues

Thanks
Amit Engel