Re: [PATCH net-next 0/4] tcp: better smp listener behavior

[Daniel Borkmann <daniel@iogearbox.net>]

On 10/13/2015 11:22 AM, Tobias Klauser wrote:
> On 2015-10-09 at 20:02:47 +0200, Daniel Borkmann <daniel@iogearbox.net> wrote:
>> On 10/09/2015 12:50 PM, Eric Dumazet wrote:
>>> On Thu, 2015-10-08 at 20:42 -0700, Tom Herbert wrote:
>>>> On Thu, Oct 8, 2015 at 8:37 AM, Eric Dumazet <edumazet@google.com> wrote:
>>>>> As promised in last patch series, we implement a better SO_REUSEPORT
>>>>> strategy, based on cpu affinities if selected by the application.
>>>>>
>>>>> We also moved sk_refcnt out of the cache line containing the lookup
>>>>> keys, as it was considerably slowing down smp operations because
>>>>> of false sharing. This was simpler than converting listen sockets
>>>>> to conventional RCU (to avoid sk_refcnt dirtying)
>>>>>
>>>>> Could process 6.0 Mpps SYN instead of 4.2 Mpps on my test server.
>>>>>
>>>> Is this IPv4, IPv6, or some combination of the two ? :-)
>>>
>>> IPv4 only (mostly because I was using trafgen and its csumtcp() only
>>> deals with IPv4 and I am lazy)
>>
>> Agreed, will fix that in trafgen. ;) Thanks!
>
> Daniel, I'd have som preliminary patches for trafgen ready to introduce
> the csumudp6 and csumtcp6 trafgen helper functions (UDP/TCP checksum
> using IPv6 pseudo headers). If you want I can push them after some
> further testing later today.

Was off yesterday, so looks like you beat me to it. So, yes, perfect,
please do! :)

> Eric, would these to helpers be sufficient for your use case or do you
> need any additional trafgen functionality?