On Wed, Jun 29, 2016 at 2:44 PM, Eric Dumazet <eric.duma...@gmail.com> wrote: > On Wed, 2016-06-29 at 09:41 -0700, Andy Lutomirski wrote: > >> Overall, it looks like there's overhead of something like 50ns for >> each ahash invocation vs the shash equivalent. It's not huge, but >> it's there. (This is cache-hot. I bet it's considerably worse if >> cache-cold, because ahash will require a lot more code cache lines as >> well as the extra cache lines involved in the scatterlist and whatever >> arch stuff is needed to map back and forth between virtual and >> physical addresses. > > I am kind of mystified seeing someone caring about TCP MD5, other than > just making sure it wont crash the host when it needs to be used ;) > > The real useful work would be to use a jump label so that we can avoid > spending cycles for non TCP MD5 sessions, when a host never had to use > any MD5 negotiation. > > >
I don't care about TCP MD5 performance at all. Ease of maintenance is nice, though, and maybe there are other places in the kernel where performance does matter. --Andy
