Le Mon, 16 Jan 2023 10:53:49 +0000,
"Russell King (Oracle)" [off-list ref] a écrit :

On Mon, Jan 16, 2023 at 11:39:23AM +0100, Clément Léger wrote:

quoted

diff --git a/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c b/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c
index f2247b8cf0a3..88c941003855 100644
--- a/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c
+++ b/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c

@@ -3818,6 +3818,12 @@ static int __stmmac_open(struct net_device *dev,
 		}
 	}
 
+	/* We need to setup the phy & PCS before accessing the stmmac registers
+	 * because in some cases (RZ/N1), if the stmmac IP is not clocked by the
+	 * PCS, hardware init will fail because it lacks a RGMII RX clock.
+	 */
+	phylink_start(priv->phylink);  

So what happens if you end up with the mac_link_up method being called
at this point in the driver, before the hardware has been setup ?

If you use a fixed-link, that's a real possibility.

I actually have this setup. On the board, one GMAC is connected to a
DSA switch using a fixed-link and the other using the PCS such as added
by this series.

From what I see, indeed, the mac_link_up() function is called before
stmmac_hw_setup(). This does not seems to have any effect on my setup
(except making it working of course) but I agree this is clearly not
ideal.

What I could do is adding a function in the miic pcs driver that could
be called from my rzn1 stmmac probe function to actually configure the
PCS at probe time based on the detected "phy-mode". Does that seems
better to you ?

Thanks,

-- 
Clément Léger,
Embedded Linux and Kernel engineer at Bootlin
https://bootlin.com