On Thu, Apr 23, 2020 at 05:15:22PM +0000, Saleem, Shiraz wrote: > > Subject: Re: [RFC PATCH v5 01/16] RDMA/irdma: Add driver framework > > definitions > > > > On Thu, Apr 23, 2020 at 12:32:48AM +0000, Saleem, Shiraz wrote: > > > > > we have a split initialization design for gen2 and future products. > > > phase1 is control path resource initialization in irdma_probe_dev and > > > phase-2 is the rest of the resources with the ib registration at the > > > end of irdma_open. irdma_close must de-register the ib device which > > > will take care of ibdev free too. So it makes sense to keep allocation > > > of the ib device in irdma_open. > > > > The best driver pattern is to allocate the ib_device at the very start of probe() and > > use this to anchor all the device resources and memories. > > > > The whole close/open thing is really weird, you should get rid of it. > maybe I missing something. But why is it weird? Because the RDMA driver should exist as its own entity. It does not shutdown unless the remove() method on is struct device_driver is closed. So what exactly are open/cose supposed to be doing? I think it is a left over of trying to re-implement the driver model. > underlying configuration changes and reset management for the physical > function need a light-weight mechanism which is realized with the close/open > from netdev PCI drv --> rdma drv. > Without a teardown and re-add of virtual device off the bus. Yes, that is exactly right. If you have done something so disruptive that the ib_device needs to be destroyed then you should unplug/replug the entire virtual bus device, that is the correct and sane thing to do. There is no 'light weight' here, destroying the ib_device is incredibly expensive and disruptive. Jason
