On 16-02-24 01:48 PM, David Miller wrote: > From: Saeed Mahameed <sae...@dev.mellanox.co.il> > Date: Wed, 24 Feb 2016 22:27:16 +0200 > >>>> Netdev features can be changed dynamically to off after vlan_vid_add >>>> was called, thus vlan_vid_del will skip ndo_vlan_rx_kill_vid and will >>>> leave the device driver with un-freed resources. >>> >>> Are you sure the fix isn't to make vlan_vid_add() check ->features instead >>> of ->hw_features. >> >> This is exactly what this fix suggests, "->features" is not consistent >> and can be turned ON/OFF between vlan_add/del which can leave the NIC >> driver in inconsistent state ! > > But the user changes the setting _exactly_ to control whether these > VLAN offloads occur or not. > >>> Should we really be trying to add VLAN filters when the user has >>> turned it off? >> >> Well, I think it is debatable, but the current implementation is not >> consistent, especially for adding vlan 0 by default and then the user >> disables the vlan filter, this will cause the stack to never call the >> nic ndo_vlan_rx_kill_vid for the pre added vlan 0 and vise versa call >> kill_vid without add_vid, BUG ? >> >> So i think we have two options, use this patch, and always trust to >> delegate vlan_vid_add/del to the NIC when it's HW supports it, and the >> nic will be smart enough to know what to do with it (in case vlan >> filter is enabled/disabled). Or, for each vlan we can remember if it >> was added to the NIC or not so the stack will know whether to clean it >> up or not. > > If automatically added VLANs are the issue, then we should specially > mark it such that it will get forcefully removed regardless of feature > settings. >
I suspect the hardware driver should flush the vid list in hardware when the feature flag is disabled and when it is enabled I guess we would need to do something like vlan_vids_add_by_dev() and do an add for all the upper dev vlans. I guess with some helper functions this could be done generally in the ethtool set_features code path by tracking down the vid list and calling all the del and add routines for each vlan. A bit more work but seems more correct then creating strange cases where hw_fatures and features don't actually work as expected. > The whole point of the separation of ->hw_features and ->features is > to separate what the card can do from what the user wants enabled or > not. > > Therefore offload operations should be triggered by ->features not > ->hw_features. > > Any test on ->hw_features that is not a validation of a ->features > change request is a BIG RED FLAG and almost always a bug. >
