bgpd: Ensure community data is freed in some cases.

Customer has this valgrind trace:

Direct leak of 2829120 byte(s) in 70728 object(s) allocated from:
  0 in community_new ../bgpd/bgp_community.c:39
  1 in community_uniq_sort ../bgpd/bgp_community.c:170
  2 in route_set_community ../bgpd/bgp_routemap.c:2342
  3 in route_map_apply_ext ../lib/routemap.c:2673
  4 in subgroup_announce_check ../bgpd/bgp_route.c:2367
  5 in subgroup_process_announce_selected ../bgpd/bgp_route.c:2914
  6 in group_announce_route_walkcb ../bgpd/bgp_updgrp_adv.c:199
  7 in hash_walk ../lib/hash.c:285
  8 in update_group_af_walk ../bgpd/bgp_updgrp.c:2061
  9 in group_announce_route ../bgpd/bgp_updgrp_adv.c:1059
 10 in bgp_process_main_one ../bgpd/bgp_route.c:3221
 11 in bgp_process_wq ../bgpd/bgp_route.c:3221
 12 in work_queue_run ../lib/workqueue.c:282

The above leak detected by valgrind was from a screenshot so I copied it
by hand.  Any mistakes in line numbers are purely from my transcription.
Additionally this is against a slightly modified 8.5.1 version of FRR.
Code inspection of 8.5.1 -vs- latest master shows the same problem
exists.  Code should be able to be followed from there to here.

What is happening:

There is a route-map being applied that modifes the outgoing community
to a peer.  This is saved in the attr copy created in
subgroup_process_announce_selected.  This community pointer is not
interned.  So the community->refcount is still 0.  Normally when
a prefix is announced, the attr and the prefix are placed on a
adjency out structure where the attribute is interned.  This will
cause the community to be saved in the community hash list as well.
In a non-normal operation when the decision to send is aborted after
the route-map application, the attribute is just dropped and the
pointer to the community is just dropped too, leading to situations
where the memory is leaked.  The usage of bgp suppress-fib would
would be a case where the community is caused to be leaked.
Additionally the previous commit where an unsuppress-map is used
to modify the outgoing attribute but since unsuppress-map was
not considered part of outgoing policy the attribute would be dropped as
well.  This pointer drop also extends to any dynamically allocated
memory saved by the attribute pointer that was not interned yet as well.

So let's modify the return case where the decision is made to
not send the prefix to the peer to always just flush the attribute
to ensure memory is not leaked.

Fixes: #15459
Signed-off-by: Donald Sharp <sharpd@nvidia.com>
(cherry picked from commit addff17a555240a4ccb5d0c5733a780256837651)

1 files changed, 1 insertions, 1 deletions

diff --git a/bgpd/bgp_updgrp.h b/bgpd/bgp_updgrp.h
index d4c6ecfdbb..5e8b5374d6 100644
--- a/bgpd/bgp_updgrp.h
+++ b/bgpd/bgp_updgrp.h
@@ -441,7 +441,7 @@ extern struct bgp_adj_out *bgp_adj_out_alloc(struct update_subgroup *subgrp,
 extern void bgp_adj_out_remove_subgroup(struct bgp_dest *dest,
 					struct bgp_adj_out *adj,
 					struct update_subgroup *subgrp);
-extern void bgp_adj_out_set_subgroup(struct bgp_dest *dest,
+extern bool bgp_adj_out_set_subgroup(struct bgp_dest *dest,
 				     struct update_subgroup *subgrp,
 				     struct attr *attr,
 				     struct bgp_path_info *path);