From: David Ahern <dsah...@kernel.org>
Date: Wed, 5 Dec 2018 15:34:09 -0800
> @@ -270,37 +270,25 @@ static inline bool neigh_key_eq128(const struct
> neighbour *n, const void *pkey)
> (n32[2] ^ p32[2]) | (n32[3] ^ p32[3])) == 0;
> }
>
> -static inline struct neighbour *___neigh_lookup_noref(
> - struct neigh_table *tbl,
> - bool (*key_eq)(const struct neighbour *n, const void *pkey),
> - __u32 (*hash)(const void *pkey,
> - const struct net_device *dev,
> - __u32 *hash_rnd),
> - const void *pkey,
> - struct net_device *dev)
> +static inline struct neighbour *__neigh_lookup_noref(struct neigh_table *tbl,
> + const void *pkey,
> + struct net_device *dev)
> {
Sorry, we can't do this.
The whole point of how this is laid out is so that the entire hash traversal,
including the hash function, is expanded inline.
This demux is extremely critical on the output side, it must be the
smallest number of cycles possible. It was the only way I could justify
not caching neigh entries in the routes any more when I wrote this code.
Even before retpoline, putting an indirect call here is painful. With
retpoline it is deadly.
Please avoid removing the full inline expansion of the neigh lookup in the ipv6
and ipv4 data paths.
Thank you.
