> > > You could make a good guess at matching to two together, but it is
> > > error prone. Phys are low level things which the user is not really
> > > involved in. They interact with interface names. ethtool, ip, etc, all
> > > use interface names. In fact, i don't know of any tool which uses
> > > phydev names.
> >
> > So... proposal:
> >
> > Users should not be dealing with sysfs interface directly, anyway. We
> > should have a tool for that. It can live in kernel/tools somewhere, I
> > guess.
>
> We already have one, ethtool(1).
>
> >
> > Would we name leds phy0:... (with simple incrementing number), and
> > expose either interface name or phydev name as a attribute?
> >
> > So user could do
> >
> > cat /sys/class/leds/phy14:green:foobar/netdev
> > lan5@eth1:
I forgot about network name spaces. There can be multiple interfaces
with the name eth0, each in its own network namespace. For your
proposal to work, /sys/class/leds/phy14:green:foobar needs to be in
the network namespace, so it is only visible to other processes in the
same name space, and lan5@eth1 is then unique to that namespace.
> Which is the wrong way around. ethtool will be passed the interface
> name and an PHY descriptor of some sort, and it has to go search
> through all the LEDs to find the one with this attribute. I would be
> much more likely to add a sysfs link from
> /sys/class/net/lan5/phy:left:green to
> /sys/class/leds/phy14:left:green.
I need to test a bit, but i think this works. Everything under
/sys/class/net is network namespace aware. You only see
/sys/class/net/lan5 if there is a lan5 is in the current name space,
and you see the current namespaces version of lan5.. A sysfs symlink
out of namespace to /sys/class/led should work, assuming
/sys/class/led is namespace unaware, and phy14 is unique across all
network name spaces. But you cannot have a link in the opposite
direction from /sys/class/led/phy14 to /sys/class/net/lan5, since it
has no idea which lan5 to symlink to.
Andrew
