[ixp1200] Transmitting pkts out of Giga ports

[Kunze, Aaron]

Chompi,

I think I can answer some of these questions for you.  I believe your
intuition is correct and the transmit code is not filling the MAC
buffers fast enough.  The transmit code for this chapter was not tested
on the gigabit ports, so it's no surprise that the code is not fast
enough.  

However, the IXF1002 has a 2048 byte buffer per port.  So, you should be
able to send Ethernet packets of that size or less without underflowing.
The trick is to set the "tx send threshold" to something large.  This
threshold is the number of bytes that must be in the buffer before the
MAC starts transmitting.  If this threshold is set large enough, you
should never underflow.  This setting can be adjusted in the simulator,
and it can also be adjusted on the hardware using the IXF1002 driver.
If you are on Linux, this is done from user space via IOCTLs.  If you
are on VxWorks, this is done through function calls.

There are also two known bugs in the transmit code that we are trying to
put up on the errata web page.  This may be impacting your ability to
run the code on the hardware.  I have pasted the errata below.  It
should be up on the website soon.

For your last question, unfortunately, I couldn't find any microC code
for transmitting on gigabit ports, but that doesn't mean it doesn't
exist.  Maybe someone else on the list has something like that.

Aaron

Errata:

~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Page:   N/A
Location:  CD-ROM, Chapter09\transmit.c 
Description: 
    
The transmit code supplied in Chapter 9 does not properly handle packets
larger than 64 bytes.  There are two corrections:

First, lines 169 and 170 should read:

        port_state.current_data_ptr = 
            (void*)((unsigned int)(port_state.current_data_ptr) + (64 /
sizeof(long long)));

instead of:

        port_state.current_data_ptr = 
            (void*)((unsigned int)(port_state.current_data_ptr) + 64);

This correction accounts for the quad-word addressing of SDRAM on the
IXP1200.

The second correction is to add the following line after line 167:

        mpacket->eop = 0;

This ensures the the per-mpacket state maintained by the fill threads
gets properly initialized for each new packet.

------------------------------ 
Aaron Kunze 
Intel Research and Development 
aaron.kunze at intel.com 
----PGP Key ID: 0x81124B7C---- 
----NOT SPEAKING FOR INTEL---- 
---------GO BOILERS!!---------


> -----Original Message-----
> From: ixp1200-bounces at lists.cs.princeton.edu 
> [mailto:ixp1200-bounces at lists.cs.princeton.edu] On Behalf Of 
> chompi at softhome.net
> Sent: Tuesday, March 30, 2004 4:04 AM
> To: ixp1200 at lists.cs.princeton.edu
> Subject: [ixp1200] Transmitting pkts out of Giga ports
> 
> 
> 
> Hi all, 
> 
> I've recently tried sample code from chap9 Johnson and Kunze 
> book and in 
> DWB, the code works fine with Fast Eth ports. However, if I 
> try to use GE 
> ports, it only works when MTU<=192 (<=3 Mpkts per pkt). If 
> there are >3 
> Mpkts/pkt, the simulation halts with error "Transmit buffer 
> underflow". Does 
> this mean that the code is not filling up TFIFO fast enough?
> Anyways, even if I set MTU<=192, I still cannot send out any 
> traffic on real 
> hardware (we are using ENP-2506). L3Fwdr reference design 
> works fine on the 
> board so I do not expect any hardware or configuration problems.
> Did anybody encounter this issue before? Is there any sample 
> code (in micro 
> C) lying around that transmits data over GE ports? 
> 
> Any help is appreciated. 
> 
> Chompi. 
> _______________________________________________
> ixp1200 mailing list
> ixp1200 at lists.cs.princeton.edu
> https://lists.cs.princeton.edu/mailman/listinfo/ixp1200
>