[ovs-discuss] [Solved][Bug] flow deletion before transmission end - transmission stalled

Justin Pettit jpettit at nicira.com
Tue Nov 30 20:12:16 UTC 2010


On Nov 30, 2010, at 11:50 AM, Christian Fischer wrote:

> On Tuesday 30 November 2010 20:38:54 Christian Fischer wrote:
>> On Tuesday 30 November 2010 19:28:07 Ben Pfaff wrote:
>>> On Tue, Nov 30, 2010 at 09:18:43AM +0100, Christian Fischer wrote:
>>>> On Saturday 27 November 2010 23:52:20 Ben Pfaff wrote:
>>>>> Was the problem and solution anything worth documenting so that we
>>>>> could help others to avoid this problem in the future?
>>>> 
>>>> the problem was self made.
>>>> Having the same mtu for physical interfaces like the virtual interface
>>>> mtu (ovs, both 1500) isn't a good idea if one likes to use VLANs. 1518
>>>> is a reasonable value.
>>>> 
>>>> Well, both, mtu_setting() from InterfaceReconfigure.py and ovs-vsctl
>>>> add port, have configured 1500 bytes, but I haven't verified the
>>>> settings, and therefore it was my fault.
>>> 
>>> I'm surprised that this was necessary.  Most NIC drivers allow an extra
>>> 4 bytes when a VLAN tag is present.  What driver does your NIC use?
>> 
>> e1000e: Intel(R) PRO/1000 Network Driver - 1.1.2-NAPI
> 
> and igb Intel(R) Gigabit Ethernet Network Driver - version 2.1.9

This is an issue in the Intel driver that we've seen internally.  The driver calculates the size of the receive FIFO based on whether a VLAN group is defined.  OVS doesn't use VLAN groups, so the driver sets the maximum size to 1514 bytes, but frames with VLANs are 1518.  Jesse reworked the way VLANs are handled in Linux to make this class of problems go away, which should make it into 2.6.37.  We may look at patching the driver to work with existing kernels.  If we do that, we'll forward the patch to you.

--Justin






More information about the discuss mailing list