OS throws away large packets

[Sam Leffler]

Yehonatan Yossef wrote:
>  
>
>   
>> -----Original Message-----
>> From: Tom Judge [mailto:tom at tomjudge.com] 
>> Sent: Monday, April 28, 2008 8:21 PM
>> To: Mr Y
>> Cc: freebsd-questions at freebsd.org; freebsd-net at freebsd.org
>> Subject: Re: OS throws away large packets
>>
>> Mr Y wrote:
>>     
>>> Hi all,
>>>
>>> I'm trying to implement Large Recieve Offload for an 
>>>       
>> Ethernet driver 
>>     
>>> on FreeBSD 6.3, but all my >MTU packets are being thrown by the OS.
>>> I'm using mbuf chains in this imlpementation, each mbuf is 
>>>       
>> a cluster 
>>     
>>> of MCLBYTES bytes. They are linked by the m_next pointer.
>>> The first packet being thrown away is 2945 bytes long. 
>>>       
>> Wireshark shows 
>>     
>>> the packet that is being passed to the OS is correct.
>>>
>>> Do I need to set some OS parameter to make it recieve mbuf chains?
>>>
>>> Please help.
>>>
>>>       
>> Hi Yony,
>>
>> I seem to remember some discussion about this list last year 
>> see the following threads:
>>
>>
>>     
> http://lists.freebsd.org/pipermail/freebsd-net/2007-September/015250.htm
> l
>   
> http://lists.freebsd.org/pipermail/freebsd-net/2007-September/015350.htm
> l
>   
> >From my limited reading of these threads just now and possibly bad
> memory.  It would seem that the MRU to MTU relationship is defined in
> the nic driver rather than 
>   
>> enforced further up the stack or at least that seamed to be the case
>>     
> with the bce driver.
>   
>> Hope this is helpful,
>>
>> Tom
>>     
>
> Hi Tom,
>
> >From what I understand these threads are referring to the bce hardware
> configuration (bus configuration) and driver mbuf allocation size. Am I
> correct?
> In my case I'm not trying to receive packets >MTU from the HW, but to
> chain mbuf clusters, each is MCLBYTES long, and pass the mbuf chain to
> the OS.
> Since tcpdump (analyzed by wireshark) catches the packets above the
> driver and reports a good packet (and 2945 bytes long), I assume my
> driver functionality is ok. From what I know tcpdump is supposed to
> immitate the way the stack sees the packet, yet it is discarded.
> My logic says there is an OS parameter handled by the driver (at net
> device init time for example) that will set the OS to receive large mbuf
> chains, or a kernel tcp parameter. Is the tcp stack submitted to the mtu
> somehow?
>
>   
I don't see where you've identified what version of the os you're 
working with.  There's a check in the 802.3 input path on earlier 
systems to discard frames >mtu.  This was removed not too long ago with 
LRO in mind; check the history of sys/net/if_ethersubr.c.

    Sam