On Mon, Dec 10, 2018 at 02:15:18PM -0600, Steve Wise wrote: > > On 12/10/2018 2:07 PM, Jason Gunthorpe wrote: > > On Mon, Dec 10, 2018 at 01:47:39PM -0600, Steve Wise wrote: > >> On 12/7/2018 1:25 PM, Jason Gunthorpe wrote: > >>> I don't like messing with module refcount to solve a locking problem.. > >>> Here something like this compile-tested thing.. > >>> > >>> Instead of implementing dellink as a function call implement it as a > >>> flag 'ALLOW_USER_UNREGISTER'. The core code simply calls > >>> ib_unregister_device_and_put() upon dellink. > >>> > >> Why don't we just mandate that if a driver uses rdma_link_ops, then it > >> must support the core doing the ib_unregister_device_and_put() on a > >> DELLINK nlmsg? > > Yes, exactly. > > > > But conversely we must prevent dellink from running if the device was > > not created by a compatible driver > > That'll be handled by virtue that only devices created via NEWLINK will > have the driver's link_ops pointer cached in the ib_device. this is fine, but do we need link_ops to be stored if we are just calling ib_unregister_device_and_put() There is not really anyneed for a dellink callback, right? Jason
