[zebra] Bug #268, Fix race between add/delete of routes, sanitise rib queueing
2006-07-27 Paul Jakma <paul.jakma@sun.com>
* rib.h: (struct rib) Add a route_node rn_status flag field,
this has to be copied every time head RIB of a route_node
changes.
Remove the rib lock field, not needed - see below.
Add a status field for RIB-private flags.
* zebra_rib.c: Add a global for the workqueue hold time, useful
for testing.
(general) Fix for bug #268. Problem originally
detailed by Simon Bryden in [quagga-dev 4001].
Essentially, add/delete of a RIB must happen /before/ the
queue. Best-path selection (ie rib_process) and reaping of
freed RIBs can then be done after queueing. Only the route_node
is queued - no important RIB state (i.e. whether a RIB is to be
deleted) is queued.
(struct zebra_queue_node_t) Disappears, no longer need to
track multiple things on the queue, only the route_node.
(rib_{lock,unlock}) removed, RIBs no longer need to be
refcounted, no longer queued.
(rib_queue_qnode_del) Removed, deleted RIBs no longer deleted
via the queue.
(rib_queue_add_qnode) deleted
(rib_queue_add) Only the route_node is queued for best-path
selection, we can check whether it is already queued or
not and avoid queueing same node twice - struct rib * argument
is not needed.
(rib_link/unlink) (un)link RIB from route_node.
(rib_{add,del}node) Front-end to updates of a RIB.
(rib_process) Reap any deleted RIBs via rib_unlink.
Unset the route_node 'QUEUED' flag.
(General) Remove calls to rib_queue_add where add/del node was
called - not needed, update calls where not.
Ignore RIB_ENTRY_REMOVEd ribs in loops through route_nodes
diff --git a/zebra/ChangeLog b/zebra/ChangeLog
index a9bb0d0..fde3f51 100644
--- a/zebra/ChangeLog
+++ b/zebra/ChangeLog
@@ -7,9 +7,39 @@
* test_main.c: Test harness for zebra, currently just to test the
RIB.
* Makefile.am: Build testzebra using above.
+ * debug.{c,h}: Add 'debug zebra rib' and 'debug zebra rib queue'.
+ * rib.h: (struct rib) Add a route_node rn_status flag field,
+ this has to be copied every time head RIB of a route_node
+ changes.
+ Remove the rib lock field, not needed - see below.
+ Add a status field for RIB-private flags.
* zebra_rib.c: Add a global for the workqueue hold time, useful
for testing.
- * debug.{c,h}: Add 'debug zebra rib' and 'debug zebra rib queue'.
+ (general) Fix for bug #268. Problem originally detailed by
+ Simon Bryden in [quagga-dev 4001].
+ Essentially, add/delete of a RIB must happen /before/ the
+ queue. Best-path selection (ie rib_process) and reaping of
+ freed RIBs can then be done after queueing. Only the route_node
+ is queued - no important RIB state (i.e. whether a RIB is to be
+ deleted) is queued.
+ (struct zebra_queue_node_t) Disappears, no longer need to
+ track multiple things on the queue, only the route_node.
+ (rib_{lock,unlock}) removed, RIBs no longer need to be
+ refcounted, no longer queued.
+ (rib_queue_qnode_del) Removed, deleted RIBs no longer deleted
+ via the queue.
+ (rib_queue_add_qnode) deleted
+ (rib_queue_add) Only the route_node is queued for best-path
+ selection, we can check whether it is already queued or
+ not and avoid queueing same node twice - struct rib * argument
+ is not needed.
+ (rib_link/unlink) (un)link RIB from route_node.
+ (rib_{add,del}node) Front-end to updates of a RIB.
+ (rib_process) Reap any deleted RIBs via rib_unlink.
+ Unset the route_node 'QUEUED' flag.
+ (General) Remove calls to rib_queue_add where add/del node was
+ called - not needed, update calls where not.
+ Ignore RIB_ENTRY_REMOVEd ribs in loops through route_nodes
2006-07-27 Rumen Svobodnikov <rumen@telecoms.bg>