This idea is from Alexey Kuznetsov.

It is common for policies to be non-prefixed.  And for
that case we can optimize lookups, insert, etc. quite
a bit.

For each direction, we have a dynamically sized policy
hash table for non-prefixed policies.  We also have a
hash table on policy->index.

For prefixed policies, we have a list per-direction which
we will consult on lookups when a non-prefix hashtable
lookup fails.

This still isn't as efficient as I would like it.  There
are four immediate problems:

1) Lots of excessive refcounting, which can be fixed just
   like xfrm_state was
2) We do 2 hash probes on insert, one to look for dups and
   one to allocate a unique policy->index.  Althought I wonder
   how much this matters since xfrm_state inserts do up to
   3 hash probes and that seems to perform fine.
3) xfrm_policy_insert() is very complex because of the priority
   ordering and entry replacement logic.
4) Lots of counter bumping, in addition to policy refcounts,
   in the form of xfrm_policy_count[].  This is merely used
   to let code path(s) know that some IPSEC rules exist.  So
   this count is indexed per-direction, maybe that is overkill.
Signed-off-by: David S. Miller <davem@davemloft.net>