On 5/11/21 7:13 PM, Sagi Grimberg wrote:

quoted

quoted

I've been thinking about what you said regarding the need to repeat
the -w option when two interfaces share the same IP address. I think
we're looking at the problem from a different point of view. The
current implementation uses an IP address to identify an interface.
I, on the other hand, believe that the best way to identify an
interface is by its "interface name or index". In previous emails, I
provided examples of the problems that may occur when using an IP
address to identify an interface. For example, one can assign the
same IP address to different interfaces making it impossible to
distinguish interfaces by their IP address alone. Another example is
that the low level APIs (e.g. setsockopt(SO_BINDTODEVICE) don’t even
require the source IP address. They only need the interface
name/index. So, why go through the trouble of performing a reverse
address lookup to retrieve the interface name/index when the address is

not used at all?

quoted

quoted

By the way, if nvme-cli/linux-nvme allowed specifying interfaces by
name/index, then we would not really need to repeat the -w option
unless we also wanted to set the source address at the same time.
Setting the source address is a completely different thing from
setting the interface. One should be allowed to set one independently
from the other, or both, or none.

If you look at how ping is implemented, they do not infer the
interface from the IP address. If one wants to force ping to go over
an interface, then one must provide the interface by name/index using
the -I option. If one wants to change the source IP address (without
forcing a specific interface), then one provides the IP address to
the -I option. It's simple and intuitive. And ping also supports
appending the interface to the Destination IP using the '%' delimiter
for IPv6-only as per RFC4007.

I think that nvme-cli/linux-nvme should follow the ping approach.
Interfaces should never be inferred from source IP addresses, but
instead be clearly identified by their name or index. And setting the
source address should be independent from setting the interface.

I'm starting to think that we are going in circles, I'm getting to the
point that having host_iface is the right way to go.

We can have nvme-cli convert <addr>%iface notation to
"..,host_traddr=<addr>,host_iface=<iface>,.." when creating the
controller string...

That was my thinking, too; the only _other_ way would be to have a
host_traddr _imply_ a binding to the underlying interface.
We could define that, but having a single option doing _two_ different
things is not good software design.
So let's do the 'host_iface' thingie.

Cheers,

Hannes
--
Dr. Hannes Reinecke                Kernel Storage Architect
hare@suse.de                              +49 911 74053 688
SUSE Software Solutions GmbH, Maxfeldstr. 5, 90409 Nürnberg HRB 36809
(AG Nürnberg), Geschäftsführer: Felix Imendörffer

Thanks Hannes and Sagi. I will resubmit the 'host_iface' patch after running more IPv6 tests. Sagi made me realize that IPv6 has a few subtle differences and I need to double check that the current host_iface implementation works as well for IPv6 as it does for IPv4.
Martin
_______________________________________________
Linux-nvme mailing list
Linux-nvme@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-nvme