[PATCH 3/3] miidev: fix 1G wrong detection

[Eric Bénard]

Hi,

Le Fri, 7 Sep 2012 11:28:33 +0200,
Johannes Stezenbach <js at sig21.net> a écrit :

> Hi,
> 
> On Fri, Sep 07, 2012 at 11:02:18AM +0200, Eric Bénard wrote:
> > Le Fri, 7 Sep 2012 10:52:16 +0200,
> > Johannes Stezenbach <js at sig21.net> a écrit :
> > > On Thu, Sep 06, 2012 at 09:39:31PM +0200, Eric Bénard wrote:
> > > > since 99e72c8bbdbdc690025a5868d831f1fe79ad56fc on an i.MX51 based board,
> > > > I get : "phy0: Link is up - 1000/Full". It seems miidev tries to probe
> > > > the PHY to early and gets 0x3ffff which leads to the wrong capabilities
> > > > setting.
> > > 
> > > Hm, MII registers are only 16bit, why does your mii_read()
> > > implementation return 0x3ffff?
> > > 
> > in fec_imx it returns the 32 bit register. I though we could mask it to
> > only return the data but that wouldn't solve the problem as the tests
> > in miidev would fail because the data is 0xFFFF.
> 
> Well, the check for the PHY ID registers was added for a purpose:
> It allows barebox to print a useful error message if it can't talk
> to the PHY, which is *much* better than letting you guess why
> you ethernet doesn't work.  You could change the check to ">= 0xffff"
> but it looks strange.  IMHO it would be better to mask the invalid
> bits in your mii_read().
> 
that's the cleanest solution, but in the present case if we do that
miidev will fail to probe on this board.

> > > Also, what exactly do you mean by "too early"?  Your code
> > > shouldn't call mii_register() before the MDIO clock is stable.
> > > 
> > fec_imx.c does that.
> 
> But why is it too early?  What do you need to wait for?
> Maybe something in eth_device.open() enables MDIO?
> 
> Maybe it is actually better to defer PHY probing until
> eth_device.open() is called, to save a few milliseconds during
> boot from flash when ethernet isn't used.  But I have a board which can
> have different PHY (e.g. 100Mbit or 1Gbit), and where one
> of them doesn't answer to address 0, so I need to probe the address.
> Then I would need to defer the mii_register() also
> until eth_device.open()?
> 
sure fec_imx may need some rework to fit with mii_dev in trhe actual
state.

Eric