I did more testing yesterday and today and got at another crash related
to the
same issue addressed by this patch, except this time it doesn't happen
during initialization
time. Changing the link type any time after initial setup triggers the
same bug of wrong indexing.
This is not limited to nbr_self, all neighbors will be indexed
incorrectly after we change the link type.
Of course this is why we have interface up/down events to reset the
state and delete all neighbors.
Since the code changes the interface config (link type in this case)
before going to ISM_Down state,
indexing rules on link is flipped causing all subsequent lookups to
fail. In my patch below I added an assertion to
catch such cases:
+ rn = route_node_lookup (oi->nbrs, &p);
+ if (rn){
+ /* We found the neighbor!
+ * now make sure it is not the exact same neighbor
+ * structure that we are about to free
+ */
+ assert (nbr != rn->info);
+ }
+ route_unlock_node (rn);
My rational was that we shouldn't get into this case, it should be
handled earlier. Now I think I should
just drop the assertion and set rn->info to NULL. Any reason why I
shouldn't do that? Also. Do I need to
unlock twice, so the code would like like this:
+ rn = route_node_lookup (oi->nbrs, &p);
+ if (rn){
+ /* We found the neighbor!
+ * now make sure it is not the exact same neighbor
+ * structure that we are about to free
+ */
+ if (nbr == rn->info);{
+ rn->info = NULL;
+ route_unlock_node (rn);
+ }
+ }
+ route_unlock_node (rn);
Thanks,
Jafar
On 4/18/2016 11:15 AM, Jafar Al-Gharaibeh wrote:
ospfd keeps a list of neighbor routers for each configured interface. This
list is indexed using the neighbor router id in case of point-to-point and
virtual link types, otherwise the list is indexed using the neighbor's
source IP (RFC 2328, page 96). The router adds itself as a "pseudo" neighbor
on each link, and also keeps a pointer called (nbr_self) to the neighbor
structure. This takes place when the interface is first configured. Currently
ospfd adds this pseudo neighbor before the link parameters are fully
configure,
including whether the link type is point-to-point or virtual link. This causes
the pseudo neighbor to be always indexed using the source IP address
regardless
of th link type. For point-to-point and virtual links, this causes the lookup
for the pseudo neighbor to always fail because the lookup is done using the
router id whereas the neighbor was added using its source IP address.
This becomes really problematic if there is a state change that requires a
rebuild of nbr_self, changing the router id for example. When resetting
nbr_self, the router first tries to remove the pseudo neighbor form its
neighbor list on each link by looking it up and resetting any references to it
before freeing the neighbor structure. since the lookup fails to retrieve any
references in the case of point-to-point and virtual links the neighbor
structure is freed leaving dangling references to it. Any access to the
neighbor list after that is bound to stumble over this dangling pointer
causing ospfd to crash.
Signed-off-by: Jafar Al-Gharaibeh <[email protected]>
---
ospfd/ospf_interface.c | 8 +++++---
ospfd/ospf_neighbor.c | 29 ++++++++++++++++++++++++++++-
ospfd/ospfd.c | 6 +++---
3 files changed, 36 insertions(+), 7 deletions(-)
diff --git a/ospfd/ospf_interface.c b/ospfd/ospf_interface.c
index f41c4f1..a221851 100644
--- a/ospfd/ospf_interface.c
+++ b/ospfd/ospf_interface.c
@@ -232,8 +232,8 @@ ospf_if_new (struct ospf *ospf, struct interface *ifp,
struct prefix *p)
/* Set default values. */
ospf_if_reset_variables (oi);
- /* Add pseudo neighbor. */
- oi->nbr_self = ospf_nbr_new (oi);
+ /* Set pseudo neighbor to Null */
+ oi->nbr_self = NULL;
oi->ls_upd_queue = route_table_init ();
oi->t_ls_upd_event = NULL;
@@ -910,7 +910,9 @@ ospf_vl_new (struct ospf *ospf, struct ospf_vl_data
*vl_data)
if (IS_DEBUG_OSPF_EVENT)
zlog_debug ("ospf_vl_new(): set associated area to the backbone");
- ospf_nbr_add_self (voi);
+ /* Add pseudo neighbor. */
+ ospf_nbr_self_reset (voi);
+
ospf_area_add_if (voi->area, voi);
ospf_if_stream_set (voi);
diff --git a/ospfd/ospf_neighbor.c b/ospfd/ospf_neighbor.c
index 23a6417..a4246f3 100644
--- a/ospfd/ospf_neighbor.c
+++ b/ospfd/ospf_neighbor.c
@@ -181,6 +181,31 @@ ospf_nbr_delete (struct ospf_neighbor *nbr)
route_unlock_node (rn);
}
+ else
+ {
+ /*
+ * This neighbor was not found, but bfore we move on and
+ * free the neighbor structre, make sure that it was not
+ * indexed incorrectly and ended up in the "worng" place
+ */
+
+ /* reverse the lookup rules */
+ if (oi->type == OSPF_IFTYPE_VIRTUALLINK ||
+ oi->type == OSPF_IFTYPE_POINTOPOINT)
+ p.u.prefix4 = nbr->src;
+ else
+ p.u.prefix4 = nbr->router_id;
+
+ rn = route_node_lookup (oi->nbrs, &p);
+ if (rn){
+ /* We found the neighbor!
+ * now make sure it is not the exact same neighbor
+ * structure that we are about to free
+ */
+ assert (nbr != rn->info);
+ }
+ route_unlock_node (rn);
+ }
/* Free ospf_neighbor structure. */
ospf_nbr_free (nbr);
@@ -207,7 +232,9 @@ ospf_nbr_bidirectional (struct in_addr *router_id,
void
ospf_nbr_self_reset (struct ospf_interface *oi)
{
- ospf_nbr_delete (oi->nbr_self);
+ if (oi->nbr_self)
+ ospf_nbr_delete (oi->nbr_self);
+
oi->nbr_self = ospf_nbr_new (oi);
ospf_nbr_add_self (oi);
}
diff --git a/ospfd/ospfd.c b/ospfd/ospfd.c
index c317ed8..b656c88 100644
--- a/ospfd/ospfd.c
+++ b/ospfd/ospfd.c
@@ -768,9 +768,6 @@ add_ospf_interface (struct connected *co, struct ospf_area
*area)
oi->params = ospf_lookup_if_params (co->ifp, oi->address->u.prefix4);
oi->output_cost = ospf_if_get_output_cost (oi);
- /* Add pseudo neighbor. */
- ospf_nbr_add_self (oi);
-
/* Relate ospf interface to ospf instance. */
oi->ospf = area->ospf;
@@ -779,6 +776,9 @@ add_ospf_interface (struct connected *co, struct ospf_area *area)
skip network type setting. */
oi->type = IF_DEF_PARAMS (co->ifp)->type;
+ /* Add pseudo neighbor. */
+ ospf_nbr_self_reset (oi);
+
ospf_area_add_if (oi->area, oi);
/* if router_id is not configured, dont bring up
_______________________________________________
Quagga-dev mailing list
[email protected]
https://lists.quagga.net/mailman/listinfo/quagga-dev