On Wed, Jul 19, 2017 at 10:14:54AM -0600, David Ahern wrote:
> On 7/19/17 1:02 AM, Jiri Pirko wrote:
> > @@ -2094,6 +2106,40 @@ mlxsw_sp_fib_entry_should_offload(const struct
> > mlxsw_sp_fib_entry *fib_entry)
> > }
> > }
> >
> > +static void
> > +mlxsw_sp_fib6_entry_offload_set(struct mlxsw_sp_fib_entry *fib_entry)
> > +{
> > + struct mlxsw_sp_fib6_entry *fib6_entry;
> > + struct mlxsw_sp_rt6 *mlxsw_sp_rt6;
> > +
> > + fib6_entry = container_of(fib_entry, struct mlxsw_sp_fib6_entry,
> > + common);
> > + list_for_each_entry(mlxsw_sp_rt6, &fib6_entry->rt6_list, list) {
> > + struct rt6_info *rt = mlxsw_sp_rt6->rt;
> > +
> > + write_lock_bh(&rt->rt6i_table->tb6_lock);
> > + rt->rt6i_flags |= RTF_OFFLOAD;
> > + write_unlock_bh(&rt->rt6i_table->tb6_lock);
>
> Seems wrong. A device driver should not be taking FIB table locks.
Will remove this in v2.
[...]
> > +static int mlxsw_sp_nexthop6_init(struct mlxsw_sp *mlxsw_sp,
> > + struct mlxsw_sp_nexthop_group *nh_grp,
> > + struct mlxsw_sp_nexthop *nh,
> > + const struct rt6_info *rt)
> > +{
> > + struct net_device *dev = rt->dst.dev;
> > + struct mlxsw_sp_rif *rif;
> > + int err;
> > +
> > + nh->nh_grp = nh_grp;
> > + memcpy(&nh->gw_addr, &rt->rt6i_gateway, sizeof(nh->gw_addr));
> > +
> > + if (!dev)
> > + return 0;
> > +
> > + rif = mlxsw_sp_rif_find_by_dev(mlxsw_sp, dev);
> > + if (!rif)
> > + return 0;
>
> rif == 0 means the dst device is not related to a port owned by this
> driver?
Yes.
>
>
> A lot to process so I am sure I missed the answer to these:
>
> 1. How do you handle host routes for local addresses? IPv6 inserts the
> host and anycast routes with the device set to 'lo' (or VRF device)
> instead of the device with the address. I have a patch to change this,
> but needs more testing
In mlxsw_sp_fib6_entry_type_set() we check for RTF_LOCAL and set the
FIB entry type to MLXSW_SP_FIB_ENTRY_TYPE_TRAP. Packets hitting these
routes will be trapped with IP2ME trap ID towards the CPU.
> 2. How are routes with devices unrelated to ports owned by this driver
> handled?
They are handled just like any other route, but they don't have a valid
RIF (for directly connected routes) or an adjacency group (for
gatewayed routes), so the check in mlxsw_sp_fib_entry_should_offload()
will return false and they will be programmed to the device with trap
action, but using a trap ID (RTR_INGRESS0) with a lower traffic class
than IP2ME, so packets that actually need to be locally received by the
CPU have a better QoS.
