[PATCH 1/1] Add 'Transport Interface' (triface) option. This can be used to specify the IP interface to use for the connection. The driver uses that to set SO_BINDTODEVICE on the socket before connecting.

[James Smart]

On 5/5/2021 7:31 AM, Belanger, Martin wrote:
...
>>>
>>> Given that this was the original intent for host_traddr, why not have
>>> host_traddr resolve the iface from the address and set sockopt
>>> SO_BINDTODEVICE on it?
>>>
>> That was my question, too.
>>
>> I would vastly prefer to not have another option to deal with (as it raises the
>> question whether to add it eg during 'nvme connect-all') And one could
>> argue that this was the intention of _having_ the host_traddr argument in
>> the first place ...
>>
>> Cheers,
>>
>> Hannes
> 
> Hi Sagi and Hannes,
> 
> Correct me if I'm wrong, but it sounds like host_traddr was primarily added for FC (at least it wasn't tested for TCP since it does not work in its current state). I'm not an expert on FC and maybe specifying an address is the right (and only) way to specify and interface for FC. For TCP, however, it's not advisable. Specifying an interface by its associated IP address is less intuitive than specifying the actual interface name and, in some cases, it simply won't work. That's because the association between interfaces and IP addresses is not predictable. IP addresses can be changed or can change by themselves over time (e.g. DHCP). 

HOST_TRADDR exists only in linux and has no strict definition other than 
what we define for it. I fully expect its value, just like TRADDR, will 
be transport specific.

For FC, we didn't want to use names based on driver instances, or on FC 
link addresses, so we chose WWNs that were specific to a FC port to 
identify it. The transport has the mechanisms to map the WWNs to a 
nvme-supporting FC host port.

Define whatever is necessary for rdma or TCP and what they put into 
HOST_TRADDR.

...
> 
> In conclusion, I believe that for TCP we need 2 options. One that can be used to specify an interface. And one that can be used to set the source address. And users should be allowed to use one or the other, or both, or none. Of course, the documentation for host_traddr will need some clarification. It should state that when used for TCP connection, this option only sets the source address. And the documentation for host_iface should say that this option only applies to TCP connections.

I don't like seeing 2 options. I'd rather you stuck with a single option 
and added formatting to the option so you could specify one or the other 
or both (or none by not specifying the option).

-- james