> On 11 Jan 2018, at 8:58 pm, Martin Pieuchot <m...@openbsd.org> wrote:
>
> On 11/01/18(Thu) 11:50, David Gwynne wrote:
>> while we were working on making the various pseudo interfaces you
>> stack on top of ethernet mpsafe, we split their input processing
>> off so they could be attacked one by one. they're all mpsafe now,
>> so this separation is not strictly necessary anymore.
>
> Well at that time we weren't sure how to split the work between
> CPUs. The current design allow to have multiple tasks doing some
> part of the work and, like we are pushing now, process all the
> incoming packets in the same context.
yes.
now though we assign interfaces to softnettqs, no matter what their
relationship to each other is. a physical interface could be on tq 0, and it's
child vlan on 1, which goes against the goal of processing all incoming packets
in the same context.
>
>> this moves carp and vlan input back into ether_input. [...]
>
> It does much more than that. It's also doing a conversion from mbuf to
> mbuf_list, it contains some IFF_RUNNING "fixes", it breaks carp(4)
> balancing, it breaks vlan/carp on top of trunk(4)...
it doesnt replace mbufs with mbuf_lists, it gets rid of the cookies for vlan
and carp input and gives them the list to requeue their packets on for input.
wrt to IFF_RUNNING, if we agree that we should be looking at it to know if an
interface is functional i can make a bunch of those fixes without oks.
when you say i break carp balancing, are you talking about the removal of the
PACKET_TAG_CARP_BAL_IP tagging? PACKET_TAG_CARP_BAL_IP is only used in
carp_lsdrop to clear the M_MCAST flag on the mbuf. M_MCAST wont be set on
packets destined for the carp interface because we check ac_enaddr before
checking if the packet is multicast or broadcast.
are you sure? carp and vlan on top of trunk should still function. however,
trunk or bridge/switch on vlan is broken though :(
>
>> this diff also gets rid of the use of the pseudo interfaces input
>> queues, it processes their packets off an mbuf list for each real
>> ethernet packet. if we can tie all the work done on behalf of a
>> physical ring to a single task it makes rx ring moderation for
>> physical interfaces a lot easier to implement.
>
> This needs to be done. But please don't mix that with a rewrite of
> the input handlers. I believe that's the easiest way to do that is
> to modify if_input() do process the packets directly if `ifp' is a
> pseudo-interface:
>
> if (ISSET(ifp->if_xflags, IFXF_CLONED) {
> struct ifih *ifih;
> struct srp_ref sr;
>
> NET_ASSERT_LOCKED();
>
> SRPL_FOREACH(ifih, &sr, &ifp->if_inputs, ifih_next) {
> if ((*ifih->ifih_input)(ifp, m, ifih->ifih_cookie))
> break;
> }
> SRPL_LEAVE(&sr)
> } else
> ifiq_input(&ifp->if_rcv, ml, 2048);
>
> If if_input() is called on a pseudo-interface we know we're already in
> a softnet process.
or a syscall.
the code above has pseudo interfaces recurse, where they'd loop either on ifih
or at the task level. probably not a huge concern though.
>
> We could also think of doing something similar for if_enqueue() and call
> if_start() directly for pseudo-interface.
ill think about that. ive had some other experiments in that area we could look
at too.
>
>> note that trunk and bridge/switch are still implemented using
>> interface input handlers at the moment.
>
> If you want to get rid of the input handlers, I'd suggest doing it in
> the beginning of a release cycle and for all pseudo drivers at once.
considering the trunk and bridge/switch issue, it probably is best to do them
all at once.
