Re: TG3 fix for slow switches (Was: TG3 driver failure on HP 16-way)

[Peter Chubb <peterc@xxxxxxxxxxxxxxxxxx>]

>>>>> "David" == David S Miller <davem@xxxxxxxxxxxxx> writes:

David> On Tue, 21 Dec 2004 11:12:23 +1100 Peter Chubb
David> <peterc@xxxxxxxxxxxxxxxxxx> wrote:

>> The problem is that the driver gives up before the switch that the
>> NIC is connected to has finished the negotiation phase.
>> 
>> Here's a simple patch.  I changed the way the loop works too,
>> because tg3_readphys() sets *val to 0xffffffff if it fails.

David> This patch shouldn't be needed.  This waiting loop is just an
David> optimization in case we can negotiation quickly.

David> You need to figure out why that isn't working correctly.

It doesn't work because tg3_readphy sets bmsr to 0xffffffff even if it
can't read the value.  This breaks that loop early; and because
BBMSR_LSTATUS is set, all sorts of things happen before the card is ready.


Why is tg3_readphys returning 0xffffffff if it can't read a value anyway?

Here's another possible fix.

===== drivers/net/tg3.c 1.222 vs edited =====
Index: linux-2.6.10-rc3/drivers/net/tg3.c
===================================================================
--- linux-2.6.10-rc3.orig/drivers/net/tg3.c     2004-12-21 00:25:34.000000000 
+0000
+++ linux-2.6.10-rc3/drivers/net/tg3.c  2004-12-21 01:00:56.586108274 +0000
@@ -1554,10 +1554,11 @@
                }
        }
 
-       bmsr = 0;
+
        for (i = 0; i < 100; i++) {
                tg3_readphy(tp, MII_BMSR, &bmsr);
-               tg3_readphy(tp, MII_BMSR, &bmsr);
+               if (tg3_readphy(tp, MII_BMSR, &bmsr))
+                       bmsr = 0;
                if (bmsr & BMSR_LSTATUS)
                        break;
                udelay(40);