Re: [RFC V2 PATCH 17/25] net/netpolicy: introduce netpolicy_pick_queue

[Daniel Borkmann]

On 08/04/2016 10:21 PM, John Fastabend wrote:
> On 16-08-04 12:36 PM, kan.liang@xxxxxxxxx wrote:
> > From: Kan Liang <kan.liang@xxxxxxxxx>
> >
> > To achieve better network performance, the key step is to distribute the
> > packets to dedicated queues according to policy and system run time
> > status.
> >
> > This patch provides an interface which can return the proper dedicated
> > queue for socket/task. Then the packets of the socket/task will be
> > redirect to the dedicated queue for better network performance.
> >
> > For selecting the proper queue, currently it uses round-robin algorithm
> > to find the available object from the given policy object list. The
> > algorithm is good enough for now. But it could be improved by some
> > adaptive algorithm later.
> >
> > The selected object will be stored in hashtable. So it does not need to
> > go through the whole object list every time.
> >
> > Signed-off-by: Kan Liang <kan.liang@xxxxxxxxx>
> > ---
> >   include/linux/netpolicy.h |   5 ++
> >   net/core/netpolicy.c      | 136 ++++++++++++++++++++++++++++++++++++++++++++++
> >   2 files changed, 141 insertions(+)
>
> There is a hook in the tx path now (recently added)
>
> # ifdef CONFIG_NET_EGRESS
>          if (static_key_false(&egress_needed)) {
>                  skb = sch_handle_egress(skb, &rc, dev);
>                  if (!skb)
>                          goto out;
>          }
> # endif
>
> that allows pushing any policy you like for picking tx queues. It would
> be better to use this mechanism. The hook runs 'tc' classifiers so
> either write a new ./net/sch/cls_*.c for this or just use ebpf to stick
> your policy in at runtime.
>
> I'm out of the office for a few days but when I get pack I can test that
> it actually picks the selected queue in all cases I know there was an
> issue with some of the drivers using select_queue awhile back.

+1, I tried to bring this up here [1] in the last spin. I think only very
few changes would be needed, f.e. on eBPF side to add a queue setting
helper function which is probably straight forward ~10loc patch; and with
regards to actually picking it up after clsact egress, we'd need to adapt
__netdev_pick_tx() slightly when CONFIG_XPS so it doesn't override it.

  [1] http://www.spinics.net/lists/netdev/msg386953.html