nvme-tcp: fix regression that causes sporadic requests to time out

[ Upstream commit 3770a42b ]

When we queue requests, we strive to batch as much as possible and also
signal the network stack that more data is about to be sent over a socket
with MSG_SENDPAGE_NOTLAST. This flag looks at the pending requests queued
as well as queue->more_requests that is derived from the block layer
last-in-batch indication.

We set more_request=true when we flush the request directly from
.queue_rq submission context (in nvme_tcp_send_all), however this is
wrongly assuming that no other requests may be queued during the
execution of nvme_tcp_send_all.

Due to this, a race condition may happen where:

 1. request X is queued as !last-in-batch
 2. request X submission context calls nvme_tcp_send_all directly
 3. nvme_tcp_send_all is preempted and schedules to a different cpu
 4. request Y is queued as last-in-batch
 5. nvme_tcp_send_all context sends request X+Y, however signals for
    both MSG_SENDPAGE_NOTLAST because queue->more_requests=true.

==> none of the requests is pushed down to the wire as the network
stack is waiting for more data, both requests timeout.

To fix this, we eliminate queue->more_requests and only rely on
the queue req_list and send_list to be not-empty.

Fixes: 122e5b9f ("nvme-tcp: optimize network stack with setting msg flags according to batch size")
Reported-by: Jonathan Nicklin <jnicklin@blockbridge.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Tested-by: Jonathan Nicklin <jnicklin@blockbridge.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Sasha Levin <sashal@kernel.org>