On Tue, Feb 02, 2021 at 07:42:11PM +0000, Saleem, Shiraz wrote: > > > Only loosely following the arguments here, but one of the requirements > > > of the driver-op scheme is that the notifying agent needs to know the > > > target device. With the notifier-chain approach the target device > > > becomes anonymous to the notifier agent. > > > > Yes, and you need to have an aux device in the first place. The netdev side has > > neither of this things. > > But we do. The ice PCI driver is thing spawning the aux device. And > we are trying to do something directed here specifically between the > ice PCI driver and the irdma aux driver. Seems the notifier chain > approach, from the comment above, is less directed and when you want > to broadcast events from core driver to multiple registered > subscribers. Yes, generally for good design the net and rdma drivers should be peers, using similar interfaces, otherwise there will be trouble answering the question what each peice of code is for, and if a net change breaks rdma land. > > I think it would be a bit odd to have extensive callbacks that > > are for RDMA only, that suggests something in the core API is not general enough. > > Yes there are some domain specific ops. But it is within the > boundary of how the aux bus should be used no? Sure is, but I'm not sure it is a great design of a driver. In the end I don't care alot about which thing you pick, so long as the peer layer is fused with aux bus and there isn't a 2nd registration layer for devices. Jason
