Description of problem:
A VLAN'd vnet interface on a VM; host has bridge->bond->eth0/1.
TCP comms fail when the "eth1" interface in the VM has tx csum enabled (the
default), works with it disabled. Reported under SIP on IPv6; replicated with
ssh on IPv4 and v6. Comms target system has similar configuration.
First data-bearing transmitted packet from ssh client as captured on
bridge of host has bad cksum. This might be expected if the offload
request is passed to the host with each packet, however disabling offload
on the bridge does not help (and commands on the client to enable offload
complete with no error).
Version-Release number of selected component (if applicable):
kernel-431.1.1 and -431.26.1 (self-built)
Steps to Reproduce:
1. Networking config as above
2. ssh <target>
No password prompt
Prompt for password
It should actually be kernel. Component updated.
The proposed solution isn't complete. Found additional issues related to this
functionality. Trying to address.
Patch has been NACKed for rhel6.6.
This problem is being resolved in a differnet way and for a different BZ.
This one can now be marked/closed as duplicate.
This problem was introduced by enabbling TSO/GSO offloading support for vlan devices on virtio-net. The problem is really in the host kernel hw drivers where a lot of these drivers have VLAN hw acceleration enabled. Thus, these host drivers make an assumption that any packet that needs TSO/GSO or checksum acceleration will not have a vlan header present. This is not the case for packets now generated by virtual machines.
We are currently tracking different HW that has this issue in Bug 1132972.
As a workaround, you can disable checksum and TSO offloading on the vlan device.
*** This bug has been marked as a duplicate of bug 1132588 ***