Re: [PATCH net-next] net: sock_rps_record_flow() is for connected sockets

[Tom Herbert <tom@herbertland.com>]

On Thu, Dec 8, 2016 at 12:05 PM, Hannes Frederic Sowa
<hannes@stressinduktion.org> wrote:
> Hello,
>
> On Thu, Dec 8, 2016, at 20:15, Tom Herbert wrote:
>> On Thu, Dec 8, 2016 at 10:02 AM, Eric Dumazet <eric.dumazet@gmail.com>
>> wrote:
>> > On Thu, 2016-12-08 at 09:49 -0800, Tom Herbert wrote:
>> >
>> >> Of course that would only help on systems where no one enable encaps,
>> >> ie. looks good in the the simple benchmarks but in real life if just
>> >> one socket enables encap everyone else takes the hit. Alternatively,
>> >> maybe we could do early demux when we do the lookup in GRO to
>> >> eliminate the extra lookup?
>> >
>> > Well, if you do the lookup in GRO, wont it be done for every incoming
>> > MSS, instead of once per GRO packet ?
>>
>> We should be able to avoid that. We already do the lookup for every
>> UDP packet going into GRO, would only need to take the refcnt once for
>> the whole GRO packet.
>>
>> >
>> > Anyway, the flooded UDP sockets out there are not normally connected
>>
>> We still should be able to use early demux in that case, just can't
>> avoid the route lookup. I wonder if we might be able to cache a soft
>> route maybe for the last local destination received to help the
>> unconnected sockets case...
>>
>> In any case, I can take a look at of doing early demux from with UDP GRO.
>
> Early demux already breaks ip rules: we might set up a rule so an
> incoming packet might depending on the rule not find an input route at
> all and would be forwarded. Same problem might occur with VRF, when you
> have multiple ip addresses in different "realms".
>
> That said, I don't see why we can't be more aggressive for GRO in the
> unconnected case: we simply must make sure that the current namespace
> holds the ip address, which is simply a hash lookup. After that we can
> even accept packets for a wildcard bounded socket.
>
> Probably we should disable this logic as soon as soon as vrf and/or
> rules are active to have correct semantics.
>
All this gets dicey in the presence of encapsulation. One problem is
that we can't tell when or if a packet crosses network namespace just
by parsing the packet. It's a subset of the general problem of
correctly identifying packets outside of the terminal protocol
processing (we've already talked about the incorrectness of devices
that identify UDP encapsulation based on port numbers). But even
without encapsulation there is still the problem as you point out with
vrf, IPvlan, etc. I think the answer thus far has been to hand wave
and rely on probability, for instance identifying UDP encapsulation by
port number in device probably works almost all of the time. Matching
an unconnected UDP socket in GRO and then accepting a route associated
with that probably would also work nearly all the time. Maybe if we
can quantify the dependencies that early parsing has in this area,
other mechanisms (vrf) might be able to do something intelligent to
ensure correctness-- does seem like a hard problem though!

Tom

> Bye,
> Hannes