[syslinux] SYSLINUX 2.06 released

James Courtier-Dutton James at superbug.demon.co.uk
Wed Aug 27 12:37:05 PDT 2003


Josef Siemes wrote:
> Hi,
> 
> "H. Peter Anvin" <hpa at zytor.com> schrieb am 26.08.03 16:38:01:
> 
>>James Courtier-Dutton wrote:
>>
>>>Do you have a network trace of this problem happening. Did you check for 
>>>the DF bit being set of not?
>>>
>>>Maybe some PXE stacks are doing MTU Path discovery which then means that 
>>> they should not see any fragmented packets.
> 
> 
>>It would be the TFTP server not doing so, but it should then get back 
>>the ICMP NEED FRAGMENT and allow the packet to be fragmented.  This is 
>>UDP, not TCP, so it cannot force smaller IP datagrams to be sent.
>>
>>The big problem is that the PXE stack will generally do the wrong thing 
>>with the fragmented incoming packet.  Most PXE stacks are buggier than 
>>hell, and it doesn't take much tweaking.
> 
> 
> Since I have reported this problem some more explanation of what was happening:
> 
> The server is set to MTU of 1492. This is a token-ring network with some ethernet
> bridges, and these bridges demand 1492, they don't work with 1500. With this
> every packet sent from the tftpd with packet size 1500 (this is blksize 1480
> in tftp) gets fragmented into two packets, the last one containing the 8 bytes
> that didn't fit into the first one. The effect was that any download (even of
> pxelinux config file) by pxelinux failed in mysterious ways: Sometimes pxelinux
> reported errors in the config file, and the downloaded kernel didn't start at all.
> 
> The client is a IBM PCI token ring card, and this doesn't reassemble the packets
> correct. I don't know if it set the DF bit. 
> 
> The effect of this is that the download once again generated more packets as
> were needed, even if the PXE rom was able to reassemble the packets as it
> should have done. So the best solution for this would be to make this configurable
> and let the user (or the config file) decide which blksize to use.
> 
> BTW: Still need to test 2.06 ...
> 
> Regards,
> 
> Josef
> __________________________________________________________________________
> Die sicherste Form der Kommunikation: E-Mails verschluesseln, Spam-Filter,
> Adressverifizierung, digitale Unterschrift: http://freemail.web.de
> 
> 
> 
Ah. Token ring...I think it is 10 years since I needed to play with a 
Token ring network. Token ring networks can handle much larger packets 
(4096+) than ethernet (<=1492), so I am not surprised that it is not 
very good at joining fragments.
Token ring is so old. Is there a very good reason for keeping it, and 
not moving to a switched ethernet network?
The clients containing the Token ring cards must also be very old.

Anyway, thanks for the explanation.

Cheers
James





More information about the Syslinux mailing list