Hi,
On Tue, 2017-12-12 at 12:44 -0500, Craig Gallek wrote:
On Tue, Dec 12, 2017 at 8:09 AM, Paolo Abeni [off-list ref] wrote:
quoted
When a reuseport socket group is using a BPF filter to distribute
the packets among the sockets, we don't need to compute any hash
value, but the current reuseport_select_sock() requires the
caller to compute such hash in advance.
This patch reworks reuseport_select_sock() to compute the hash value
only if needed - missing or failing BPF filter. Since different
hash functions have different argument types - ipv4 addresses vs ipv6
ones - to avoid over-complicate the interface, reuseport_select_sock()
is now a macro.
Purely subjective, but I think a slightly more complicated function
signature for reuseport_select_sock (and reuseport_select_sock6?)
would look a little better than this macro. It would avoid needing to
expose the reuseport_info struct and would keep the rcu semantics
entirely within the function call (the fast-path memory access
semantics here are already non-trivial...)
Thanks for the feedback.
I was in doubt about the macro, too. The downside of using explicit
functions is the very long argument list and the need of 2 separate
functions for ipv4 and ipv6.
quoted
Additionally, the sk_reuseport test is move inside reuseport_select_sock,
to avoid some code duplication.
Overall this gives small but measurable performance improvement
under UDP flood while using SO_REUSEPORT + BPF.
Exciting, do you have some specific numbers here? I'd be interested
in knowing what kinds of loads you end up seeing improvements for.
this are the numbers I collected so far:
(ipv4)
socks nr vanilla(kpps) patched(kpps)
1 1747 1843
2 3109 3140
3 4480 4534
4 5796 5864
5 7063 7139
6 8168 8235
(ipv6)
socks nr vanilla(kpps) patched(kpps)
1 1433 1544
2 2537 2731
3 3622 3794
4 4689 4979
5 5738 6011
6 6671 6920
Cheers,
Paolo