On 9/11/2018 4:25 PM, Jason Gunthorpe wrote: > On Tue, Sep 11, 2018 at 04:20:50PM -0500, Steve Wise wrote: >> Ignore the old date... >> >> Jason, is this good to go? > > I'm not sure, I think your old patch might have been the right > one.. But I haven't checked it all through again yet. > > But I'm still not sure if the kref should be allowed to work like > this at all.. > > Leaving a long term dangling filep in the object pointer is not how > this was really intended to work. > > Jason > Looking at the original commit (1e7710f3f656), ib_uverbs_async_event_close() sets is_closed, yet ib_uverbs_comp_event_close() does not. Seems like an oversight in that commit maybe? But as for the whole design of this part of the rdma core, I'm not to familiar with how it "should work" to comment further. I just saw the evidence in the dump and tracked down the issue. :) This is certainly a regression from cxgb4's perspective: These tests past prior to commit 1e7710f3f656, I believe. Steve
