On Thu, 20 Apr 2017, Peter Zijlstra wrote:
For opt spinning we need to specifically know who would be next in order, again, doesn't matter how many, just who's next.
I've sent a v3 with a more precise description of this, which I hope is to your satisfaction. Given a clear tree iteration/order defined by interval trees (indexed by lowpoint and treats duplicates as inorder traversal), it is not something I would wish to alter. Over the weekend I've been experimenting more with still taking the tree->lock, but spinning while blocking ranges is 1 and 'owner' (in this case the first overlapping node, remembered when we did the initial lookup adding to the tree, _with_ the tree->lock held) is on_cpu. This would maintain the order and prevent blocking for threads that are about (?) to receive the lock. While I have somewhat of a patch, I'm tired and have not had the chance to even test the thing, so I went ahead and sent v3 anyway to not delay further. Thanks, Davidlohr
