Re: RFC: softint-based if_input

[Taylor R Campbell <campbell+netbsd-tech-kern%mumble.net@localhost>]

   Date: Thu, 4 Feb 2016 17:58:02 +0900
   From: Ryota Ozaki <ozaki-r%netbsd.org@localhost>

   So here is latest patches that apply the above fixes:
     http://www.netbsd.org/~ozaki-r/softint-if_input-percpuq3-diff.diff
     http://www.netbsd.org/~ozaki-r/softint-if_input-percpuq3.diff

Some little notes:

   --- a/sys/dev/pci/ixgbe/ixgbe.c
   +++ b/sys/dev/pci/ixgbe/ixgbe.c
   @@ -1650,7 +1650,12 @@ ixgbe_legacy_irq(void *arg)
   ...
   +#ifdef __NetBSD__
   +               /* Don't run ixgbe_rxeof in interrupt context */
   +               more = true;
   +#else
                   more = ixgbe_rxeof(que);
   +#endif

Is this necessary for the present patch?  It appears that ixgbe will
use the percpuq when ixgbe_rxeof calls if_input, so this shouldn't be
a problem for now.  Of course, maybe in the future we shall want to
make more extensive changes for polling or similar, but that's a
separate issue.

   --- a/sys/dev/usb/if_upl.c
   +++ b/sys/dev/usb/if_upl.c
   @@ -300,14 +300,14 @@ upl_attach(device_t parent, device_t self, void *aux)
   ...
           if_attach(ifp);
   -       if_alloc_sadl(ifp);
   +       if_register(ifp);

Should be either if_attach or if_initialize/if_register?

upl(4) seems to have a special if_input routine.  In this case, it
looks like you don't need if_percpuq or if_input at all --
ifp->_if_input will go straight to ip_pktq with no ethernet headers.

   --- a/sys/net/if.h
   +++ b/sys/net/if.h
   @@ -947,6 +950,21 @@ int        if_mcast_op(ifnet_t *, const unsigned long, const struct sockaddr *);
   ...
   +struct if_percpuq {
   +       struct ifnet    *ipq_ifp;
   +       void            *ipq_si;
   +       struct percpu   *ipq_ifqs;      /* struct ifqueue */
   +};

I would make this private to net/if.c -- no need to publish the guts
of if_percpuq at the moment.  Just need a forward declaration `struct
if_percpuq;' in net/if.h.

   +void   if_percpuq_enqueue(struct if_percpuq *, struct mbuf *);
   +struct mbuf *
   +       if_percpuq_dequeue(struct if_percpuq *);

I don't think these need to be published either for the moment -- they
can be static in net/if.c.

Another note on the API that occurred to me while reviewing: I've had
to determine whether a packet goes via a per-CPU queue and softint or
goes directly into, e.g., ether_input, based not on the call site
where that happens, but on whether the device did if_attach or
if_initialize/if_register.  This is relevant because the context of
the if_input call site determines whether it makes sense to defer to a
softint or go directly into the network stack.

It might be better to have two separate routines, say if_input and
if_schedule_input, with a KASSERT(ifp->ifp_percpuq != NULL) in
if_schedule_input.  On the other hand, that's hard to compile-test
unless we also push the percpuq into each driver's softc and require
the caller to pass it to if_schedule_input.  So maybe we should put
that off until then and just have callers use if_percpuq_enqueue.