Re: [PATCH 1/2] blk-mq: introduce blk_mq_complete_request_sync()

[Ming Lei <ming.lei@redhat.com>]

On Mon, Mar 18, 2019 at 08:49:24AM -0700, Bart Van Assche wrote:
> On Mon, 2019-03-18 at 23:16 +0800, Ming Lei wrote:
> > I am not familiar with SRP, could you explain what SRP initiator driver
> > will do when the controller is in bad state? Especially about dealing with
> > in-flight IO requests under this situation.
> 
> Hi Ming,
> 
> Just like the NVMeOF initiator driver, the SRP initiator driver uses an
> RDMA RC connection for all of its communication over the network. If
> communication between initiator and target fails the target driver will
> close the connection or one of the work requests that was posted by the
> initiator driver will complete with an error status (wc->status !=
> IB_WC_SUCCESS). In the latter case the function srp_handle_qp_err() will
> try to reestablish the connection between initiator and target after a
> certain delay:
> 
> 	if (delay > 0)
> 		queue_delayed_work(system_long_wq, &rport->reconnect_work,
> 				   1UL * delay * HZ);
> 
> SCSI timeouts may kick the SCSI error handler. That results in calls of
> the srp_reset_device() and/or srp_reset_host() functions. srp_reset_host()
> terminates all outstanding requests after having disconnected the RDMA RC
> connection.

Looks the approach of NVMe's error handler is basically similar with
above.

And NVMe just uses 'blk_mq_tagset_busy_iter(nvme_cancel_request)' to abort
in-flight requests, and I guess SCSI FC may use driver's approach to do that.

> Disconnecting the RC connection first guarantees that there
> are no concurrent request completion calls from the regular completion
> path and from the error handler.

Looks no concurrent request completion guarantee requires driver's
specific implementation.

However, this patch provides one simple approach for NVMe, then no
driver specific sync mechanism is needed.

Thanks,
Ming