Eric Dumazet <eric.duma...@gmail.com> wrote: > On 12/13/2018 03:03 AM, Florian Westphal wrote: > > Eric Dumazet <eric.duma...@gmail.com> wrote: > >> So, cloning would do an refcount_inc(), and deleting the extension would > >> do an refcount_dec_and_test() ? > >> > >> That is what I called an extra pair of atomic operations. > > > > If it replaces 1:1 current mptcp skb->private out-of-tree storage, then > > yes. > > > > One day I will write a book on the number of atomic operations done on a TCP > sendmsg() system call :/
Just to clarify: mptcp skb->private out-of-tree storage is maintened using atomic_ops too, so if its converted 1:1 nothing changes in the *mptcp tree*. If its possible to re-arrange the out-of-tree so that it doesn't need that (only used while skb is not cloned), then the same applies to the extension conversion. So this, as far as *this series* is concerened, a "future problem" that may or may not exist.
