On 11/21/18 4:01 AM, Mikhail Skorzhinskii wrote:

Sagi Grimberg [off-list ref] writes:
  > +static inline void nvme_tcp_queue_request(struct nvme_tcp_request *req)
  > +{
  > +	struct nvme_tcp_queue *queue = req->queue;
  > +
  > +	spin_lock_bh(&queue->lock);
  > +	list_add_tail(&req->entry, &queue->send_list);
  > +	spin_unlock_bh(&queue->lock);
  > +
  > +	queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work);
  > +}

May be I missing something, but why bother with bottom half version of
locking?

There are few places where this lock could be accessed:

  (1) From ->queue_rq() call;
  (2) From submitting new AEN request;
  (3) From receiving new R2T;

Which one if these originates from bottom half? Not 100% about queue_rq
data path, but (2) and (3) looks perfectly safe for me.

Actually, (3) in former versions was invoked in a soft-irq context which
was why this included the bh disable, but I guess it can be removed now.