On Fri, 2009-06-12 at 07:05 -0700, Jean-Louis Dupond wrote:

Hello,

I'm experiencing a problem with my "Broadcom Corporation NetXtreme
BCM5722 Gigabit Ethernet PCI Express" network card in my Dell R300
servers. When booting the server, the network card sometimes doesn't
advertise gigabit speeds, and so it auto-negotiates @ 10mbit. The speed
can then be set to gigabit with mii-tool, but its not a good solution!
On the other hand, sometimes when it boots, it just works perfectly, and
advertises gigabit speeds & auto-negotiates on gigabit!

I tried to debug the driver a bit, and found some strange behavior
:) Only 2 values were different in the writephy function when it booted
@ 10mbit & gigabit.

I found something obvious in the tg3_readphy & tg3_writephy methods.

quoted

807         while (loops != 0) {
808                 udelay(10);
809                 frame_val = tr32(MAC_MI_COM);
810                 if ((frame_val & MI_COM_BUSY) == 0) {
811                         udelay(5);
812                         frame_val = tr32(MAC_MI_COM);
813                         break;
814                 }
815                 loops -= 1;
816         }

This code was written like this to make sure we get the correct MDIO
data.  The data is supposed to be valid when the MI_COM_BUSY bit is
cleared.  But on some chips, the data may not be ready until some
microseconds after the BUSY bit is cleared.

When you see the "wrong" speed being established, please provide the mii
register dump using mii-tool -vvv eth0.  We'll then be able to see what
we advertised and what the other side advertised.

Thanks.

On line 809 frame_val is filled with a read (tr32) from the device.
On line 810 we check if its not 'MI_COM_BUSY'
If not then we have a delay of 5, and then we read it again ?!
I don't know why the value is read twice! I checked with some other
drivers (also broadcom) and there the read command was given some other
argument when reading the BUSY state, and then in the if, it was really
fetching the data. But in this case, we only have 2 times the same
argument!

With the original code the server booted into 10mbit the half of the
boots! When I removed line 811 & 812, then it booted into 10mbit only
1/20 times ! Which is way better ! But its still not fully fixed!

---------------------------------------------------------------------------

Today I found the programmers documentation on the Broadcom website and it
mentions the following:

quoted

// Write value to MI communication register
Mi_Communication_Register = Value32
// Now read back MI Communication register until the start bit
// has been cleared or we have timed out (>5000 reads)
Loopcount = 5000
While (LoopCount > 0)
Begin
Value32 = Mi_Communication_Register
If (!(Value32 | 0x20000000)) then BREAK loop
Else Loopcount--
End

Here u can see its NOT needed to read the value twice.
Also there is no delay mentioned, so I removed it, and rebooted 20 times
without 1 time on 10mbit !


Feel free to give any other solutions / comments !

Sincerely,
Jean-Louis Dupond
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html