On Wed, Mar 04, 2020 at 08:54:25PM +0800, wenxu wrote: > > 在 2020/3/4 5:53, Pablo Neira Ayuso 写道: [...] > The indirect block infrastructure is designed by the driver guys. The callbacks > is used for building and finishing relationship between the tunnel device and > the hardware devices. Such as the tunnel device come in and go away and the hardware > device come in and go away. The relationship between the tunnel device and the > hardware devices is so subtle. I understand that this mechanism provides a way for the driver to subscribe to tunnel devices that might be offloaded. > > Probably not a requirement in your case, but the same net_device might > > be used in several flowtables. Your patch is flawed there and I don't > > see an easy way to fix this. > > The same tunnel device can only be added to one offloaded flowtables. This is a limitation that needs to be removed. There are requirements to allow to make the same tunnel device be part of another flowtable. > The tunnel device can build the relationship with the hardware > devices one time in the dirver. This is protected by > flow_block_cb_is_busy and xxx_indr_block_cb_priv in driver. > > > I know there is no way to use ->ndo_setup_tc for tunnel devices, but > > you could have just make it work making it look consistent to the > > ->ndo_setup_tc logic. > > I think the difficulty is how to find the hardware device for tunnel > device to set the rule to the hardware. Right, this is the problem that the infrastructure is solving, however, it's a bit of a twisty way to address the problem. > > I'm inclined to apply this patch though, in the hope that this all can > > be revisited later to get it in line with the ->ndo_setup_tc approach. > > However, probably I'm hoping for too much. I have applied this patchset to nf-next. Probably, there might be a chance to revisit this indirect block infrastructure. Thank you.
