On Thu, 2016-03-24 at 17:50 +0100, Willy Tarreau wrote:

On Thu, Mar 24, 2016 at 09:33:11AM -0700, Eric Dumazet wrote:

quoted

quoted

--- a/net/ipv4/inet_hashtables.c
+++ b/net/ipv4/inet_hashtables.c

@@ -189,6 +189,8 @@ static inline int compute_score(struct sock *sk, struct net *net,
                                return -1;
                        score += 4;
                }
+               if (sk->sk_reuseport)
+                       score++;

This wont work with BPF

quoted

                if (sk->sk_incoming_cpu == raw_smp_processor_id())
                        score++;

This one does not work either with BPF

But this *is* in 4.5. Does this mean that this part doesn't work anymore or
just that it's not usable in conjunction with BPF ? In this case I'm less
worried, because it would mean that we have a solution for non-BPF aware
applications and that BPF-aware applications can simply use BPF.

BPF can implement the CPU choice/pref itself. It has everything needed.

I don't try to reimplement something already available, but I'm confused
by a few points :
  - the code above already exists and you mention it cannot be used with BPF

_If_ you use BPF, then you can implement a CPU preference using BPF
instructions. It is a user choice.

  - for the vast majority of applications not using BPF, would the above *still*
    work (it worked in 4.4-rc at least)

  - it seems to me that for BPF to be usable on process shutting down, we'd
    need to have some form of central knowledge if the goal is to redefine
    how to distribute the load. In my case there are multiple independant
    processes forked on startup, so it's unclear to me how each of them could
    reconfigure BPF when shutting down without risking to break the other ones.
  - the doc makes me believe that BPF would require privileges to be unset, so
    that would not be compatible with a process shutting down which has already
    dropped its privileges after startup, but I could be wrong.

Thanks for your help on this,
Willy

The point is : BPF is the way to go, because it is expandable.

No more hard points coded forever in the kernel.

Really, when BPF can be the solution, we wont allow adding new stuff in
the kernel in the old way.