On Jun 24, 2009, at 1:27 PM, Anton Vorontsov wrote:

It appears that gianfar driver has the same problem[1] that I
just fixed for ucc_geth.

NFS boot using 10/half link takes about 10 minutes to complete:

The symptoms were observed on MPC8379E-RDB boards (eTSEC). Although
I didn't find where documentation forbids clearing Full Duplex bit
for non-MII/RMII modes, it's pretty distinct that the bit should be
set.

It's no wonder though, QE Ethernet and TSEC are pretty similar.

-			if (!(phydev->duplex))
-				tempval &= ~(MACCFG2_FULL_DUPLEX);
+			if (!phydev->duplex &&
+					(phyi == PHY_INTERFACE_MODE_MII ||
+					 phyi == PHY_INTERFACE_MODE_RMII))

Hmm....have you tested this on a GMII interface?  *Technically*, full  
duplex is required for GMII, as GMII is used only for gigabit.   
However, we've been treating the GMII interface type as an indicator  
that the PHY *has* a GMII connection to the NIC.  When gianfar detects  
the speed is 10/100 it switches to the compatible MII interface via  
this code, just below:

                         case 100:
                         case 10:
                                 tempval =
                                     ((tempval & ~(MACCFG2_IF)) |  
MACCFG2_MII);


My concern is that you will be detecting the GMII interface, and  
disallowing half-duplex, despite the fact that the interface is  
actually running at 10 or 100 Mbit.

Andy