On Thu, Dec 06, 2018 at 02:53:36PM -0500, Doug Ledford wrote: > Then we end up having to duplicate the three or so bugfix patches where > we moved roce sgid resolution out of the critical path because it can > sleep. We end up making the code problems we have worse. Better to > redesign this to remove that atomic assumption. Hence why I had them > drop the stuff they had. The resolution is a slightly different issue, that was to do with creating the rdma_ah_attr. Once the attr is created then the create_ah driver callback is non-sleepable.. > If we get to the point of having two link layers that can't comply > with the atomic requirement, then I think it's time to look at > changing the requirement. Sure, and the people who want it changed are responsible :) Don't just sweep the issue under the rung by ignoring the API contract 'because it works for me today' Jason
