Thread safe about xio_send_request and xio_release_response

[terrs]

Hi,

In nbdx_connect_work context, request will be sent in stack on_response->xio_release_response->xio_connection_xmit.

In nbdx_request context, request will be sent in stack
nbdx_transfer->xio_send_request->xio_connection_xmit.

Obviously, the two contexts are different, nbdx_connect_work
is xio connection context, nbdx_request is application context,
how to ensure xio send safely in this condition.

Thanks.

--haiwei.

[terrs]

Ok, maybe nobody met this bug, I paste more details and solution.

In tcp transport condition, there are xio context loop thread and application thread,
kernel_sendmsg in xio_tcp_xmit will be scheduled, so the two threads enter
xio_tcp_xmit repeatly, panic is coming cause tmp_work.msg_len is demaged.

get_cpu/put_cpu disable preempt, but not enought for kernel_sendmsg scheduled,
we need a flag to avoid calling xio_tcp_xmit repeatly by the two threads, though they
are running in same cpu.

[katyakats]

Nbdx is running over rdma, and we haven't tested it in tcp recently. Do you see this bug over rdma as well?