Red Hat Bugzilla – Bug 616841
RFE: Support setting of rx, tx, sg, tso, ufo, gso, gro, lro against physical network interfaces
Last modified: 2013-01-23 14:57:04 EST
Description of problem: Large receive offload is a feature that isn't supported by Linux bridges, nor virtio backend drivers. If a physical NIC has LRO enabled by default, you'll get dropped packets on the host eth3: received packets cannot be forwarded while LRO is enabled eth3: received packets cannot be forwarded while LRO is enabled eth3: received packets cannot be forwarded while LRO is enabled Thus we need to be able to disable LRO on physical NICs. The ifcfg-xxx files appears to support these settings, albeit with a deprecated option Deprecated, but supported: ETHTOOL_OPTS=... Any device-specific options supported by ethtool. For example, if you wanted to force 100Mb full duplex: ETHTOOL_OPTS="speed 100 duplex full autoneg off" Note that changing speed or duplex settings almost always requires disabling autonegotiation with 'autoneg off'. Long term, this should be done by sysadmin-written udev rules. So we have two options - Expose this tunable via the interface XML (and by extension netcf) and use the deprecated setting - Expose this tunable via the node device APIs and create udev rules We likely need to expose whether each physical NIC is capable of gro, lro, etc via the node device APIs in any case, otherwise you won't know whether the tunable is actually supported on the NIC in question. This could push the balance in favour of a pure nodedevice API instead of nodedev + interface Version-Release number of selected component (if applicable): libvirt-0.8.2 How reproducible: Hardware dependant Steps to Reproduce: 1. Get a machine witha NIC supporting LRO 2. Enable LRO 3. Put the NIC into a bridge 4. Boot a guest with the bridge 5. Observe dropped packets Actual results: Expected results: Additional info:
some notes: an additional idea is to make libvirt verify that NIC is compatible with virt, and give the admin hints if it is not.
Bug 743176 points out another problem with having LRO turned on on a physical ethernet - when ip_forward is enabled, the device will be reset.
Michal, if we take the udev route, what would the design look like?
Investigating down a rat hole of BZs, I've found that this has been remedied in the RHEL6 kernel as of 2.6.32-230 (See Bug 772317). The fix is to notice that LRO is turned on when attaching a netdev to a bridge or bond (neither of which support LRO), and simply disable it at that time. This makes perfect sense, since there is no possible reason that having it enabled would be desired. Upstream uses a different mechanism, but the result is the same - if LRO should be disabled for a particular device due to it being attached to some other driver that doesn't support LRO, then it will be automatically disabled. This all means that there is no longer any need for libvirt to provide a mechanism for enabling/disabling LRO, so I'm closing this BZ.