On 10/09/2014 08:24 AM, Sebastian Hesselbarth wrote:

On 10/09/2014 04:47 PM, Thomas Petazzoni wrote:

quoted

Well, I initially remember that the original driver coming from Marvell
was using the HW PHY stuff, and that I changed that because it would
not integrate well with the kernel libphy.

A drawback of this is that because the hardware has built-in PHY
polling which triggers a MAC interrupt when the PHY status changes, they
typically don't wire up the PHY interrupt. Therefore, since we're not
able to use the MAC interrupt for PHY event notifications, we rely on
software PHY polling, which means that link up / link down events take
a few seconds to be noticed by the kernel. Unfortunately, I don't think
the hardware allows to use the hardware PHY polling to get link changes
interrupt, but not let the hardware configure the PHY itself.

Yeah, but that HW PHY stuff really only works properly with standard
compliant PHYs. In particular, the integrated Marvell PHY in Marvell
Berlin SoCs does not seem to reflect PHY status on BMCR properly /sigh/.
Anyway, I think we can live with PHY polling.

BTW, one thing I noticed here is that libphy calls adjust_link
over-and-over again although nothing has changed. I guess we can just
add some before/after comparison in the libphy state machine and only
call adjust_link when something has changed. I'll have to look closer
at the state machine first and maybe Florian can comment on this,
too.

There's basically nothing built in the generic libphy that would try to
limit the number of times the adjust_link() callback is invoked, some
changes went in the bcmgenet driver to avoid that, I have yet to see how
much of this logic is transferable to the libphy layer.
--
Florian