Re: [PATCH v0 1/6] nvme-fabrics: introduce connect_sync option

[Hannes Reinecke]

On 2/16/24 10:49, Christoph Hellwig wrote:
> On Fri, Feb 16, 2024 at 09:45:21AM +0100, Daniel Wagner wrote:
> > The TCP and RDMA transport are doing a synchronous connect, meaning the
> > syscal returns with the final result, that is. it either failed or
> > succeeded.
> >
> > This isn't the case for FC. This transport just setups and triggers
> > the connect and returns without waiting on the result.
>
> That's really weird and unexpected.  James, can you explain the reason
> behind this?
Reason is that the initial connect attempt might fail with an temporary 
failure, and will need to be retried. And rather than implementing two 
methods for handling this (one for the initial connect, and another one
for reconnect where one _has_ to use a workqueue) as eg TCP and RDMA
has implemented it FC is using a single code path for handling both.

Temporary failure on initial connect is far more likely on FC than on
other transports due to the way how FC-NVMe is modelled; essentially
one has to log into the remote port for each protocol. So if you run
in a dual fabric (with both FCP and NVMe) you'll need to log into the
same remote port twice. Depending on the implementation the target might 
only be capable of handling one port login at the same time, so the
other one will be failed with a temporary error.
That's why it's a common issue with FC. It _might_ happen with TCP, too,
but apparently not regularly otherwise we would have seen quite some
failures here; TCP can't really handle temporary failures for the
initial connect.

Cheers,

Hannes
--
Dr. Hannes Reinecke                  Kernel Storage Architect
hare@xxxxxxx                                +49 911 74053 688
SUSE Software Solutions GmbH, Frankenstr. 146, 90461 Nürnberg
HRB 36809 (AG Nürnberg), GF: I. Totev, A. McDonald, W. Knoblich