svn commit: r271504 - in head/sys: dev/oce dev/vmware/vmxnet3 dev/xen/netfront net netinet ofed/drivers/net/mlx4

[Hans Petter Selasky]

On 09/13/14 22:54, Rick Macklem wrote:
> Hans Petter Selasky wrote:
>> On 09/13/14 22:35, Hans Petter Selasky wrote:
>>>
>>> I see the default set to (65536 - 4). I don't know why you
>>> subtracted 4
>>> but I would have expected the max ethernet header length to be
>>> subtracted
>>> here?
>>
>> You mean to say that the default should be
>>
>> 65535 - ethernet header - vlan header ?
>>
> Almost. It is actually:
> 65536 - ethernet header - vlan header
> (or the min of IP_MAXPACKET vs 32*MCLBYTES - ethernet header - vlan header
>   if you want to cover your butt for the case where the value of
>   MCLBYTES is changed) IP_MAXPACKET (65535) comes from the fact that
> some devices use the iplen field of the ip header in the TSO segment
> for its length, I think? (Some do not and can support TSO segments
> greater than IP_MAXPACKET in length, but again, the default shouldn't
> assume this nor should it assume the device does the vlan header in
> hardware --> "- vlan header" to be safe for default.)
>
> This is because there are lots of broken drivers (basically any one
> that has a limit of 32 transmit segments) and this at least makes
> them work correctly. Unfortunately they still do a lot of m_defrag()
> calls for this case, but with a patch like yours, the drivers may
> eventually get patched to use the max_frags setting and then avoid
> the need to do m_defrag() calls.
>
> rick

Hi Rick,

Thanks for shedding light into existing practices.

--HPS