[bgpd] Fix 0.99 shutdown regression, introduce Clearing and Deleted states
2006-09-14 Paul Jakma <paul.jakma@sun.com>
* (general) Fix some niggly issues around 'shutdown' and clearing
by adding a Clearing FSM wait-state and a hidden 'Deleted'
FSM state, to allow deleted peers to 'cool off' and hit 0
references. This introduces a slow memory leak of struct peer,
however that's more a testament to the fragility of the
reference counting than a bug in this patch, cleanup of
reference counting to fix this is to follow.
* bgpd.h: Add Clearing, Deleted states and Clearing_Completed
and event.
* bgp_debug.c: (bgp_status_msg[]) Add strings for Clearing and
Deleted.
* bgp_fsm.h: Don't allow timer/event threads to set anything
for Deleted peers.
* bgp_fsm.c: (bgp_timer_set) Add Clearing and Deleted. Deleted
needs to stop everything.
(bgp_stop) Remove explicit fsm_change_status call, the
general framework handles the transition.
(bgp_start) Log a warning if a start is attempted on a peer
that should stay down, trying to start a peer.
(struct .. FSM) Add Clearing_Completed
events, has little influence except when in state
Clearing to signal wait-state can end.
Add Clearing and Deleted states, former is a wait-state,
latter is a placeholder state to allow peers to disappear
quietly once refcounts settle.
(bgp_event) Try reduce verbosity of FSM state-change debug,
changes to same state are not interesting (Established->Established)
Allow NULL action functions in FSM.
* bgp_packet.c: (bgp_write) Use FSM events, rather than trying
to twiddle directly with FSM state behind the back of FSM.
(bgp_write_notify) ditto.
(bgp_read) Remove the vague ACCEPT_PEER peer_unlock, or else
this patch crashes, now it leaks instead.
* bgp_route.c: (bgp_clear_node_complete) Clearing_Completed
event, to end clearing.
(bgp_clear_route) See extensive comments.
* bgpd.c: (peer_free) should only be called while in Deleted,
peer refcounting controls when peer_free is called.
bgp_sync_delete should be here, not in peer_delete.
(peer_delete) Initiate delete.
Transition to Deleted state manually.
When removing peer from indices that provide visibility of it,
take great care to be idempotent wrt the reference counting
of struct peer through those indices.
Use bgp_timer_set, rather than replicating.
Call to bgp_sync_delete isn't appropriate here, sync can be
referenced while shutting down and finishing deletion.
(peer_group_bind) Take care to be idempotent wrt list references
indexing peers.
diff --git a/bgpd/bgp_packet.c b/bgpd/bgp_packet.c
index 8b024a1..da59d32 100644
--- a/bgpd/bgp_packet.c
+++ b/bgpd/bgp_packet.c
@@ -637,9 +637,7 @@
if (write_errno == EWOULDBLOCK || write_errno == EAGAIN)
break;
- BGP_EVENT_ADD (peer, BGP_Stop);
- peer->status = Idle;
- bgp_timer_set (peer);
+ BGP_EVENT_FLUSH_ADD (peer, TCP_fatal_error);
return 0;
}
if (num != writenum)
@@ -673,10 +671,8 @@
if (peer->v_start >= (60 * 2))
peer->v_start = (60 * 2);
- BGP_EVENT_ADD (peer, BGP_Stop);
- /*bgp_stop (peer);*/
- peer->status = Idle;
- bgp_timer_set (peer);
+ /* Flush any existing events */
+ BGP_EVENT_FLUSH_ADD (peer, BGP_Stop);
return 0;
case BGP_MSG_KEEPALIVE:
peer->keepalive_out++;
@@ -721,9 +717,7 @@
ret = writen (peer->fd, STREAM_DATA (s), stream_get_endp (s));
if (ret <= 0)
{
- BGP_EVENT_ADD (peer, BGP_Stop);
- peer->status = Idle;
- bgp_timer_set (peer);
+ BGP_EVENT_FLUSH_ADD (peer, TCP_fatal_error);
return 0;
}
@@ -743,10 +737,7 @@
if (peer->v_start >= (60 * 2))
peer->v_start = (60 * 2);
- /* We don't call event manager at here for avoiding other events. */
- bgp_stop (peer);
- peer->status = Idle;
- bgp_timer_set (peer);
+ BGP_EVENT_FLUSH_ADD (peer, BGP_Stop);
return 0;
}
@@ -2375,14 +2366,6 @@
if (BGP_DEBUG (events, EVENTS))
zlog_debug ("%s [Event] Accepting BGP peer delete", peer->host);
peer_delete (peer);
- /* we've lost track of a reference to ACCEPT_PEER somehow. It doesnt
- * _seem_ to be the 'update realpeer with accept peer' hack, yet it
- * *must* be.. Very very odd, but I give up trying to
- * root cause this - ACCEPT_PEER is a dirty hack, it should be fixed
- * instead, which would make root-causing this a moot point..
- * A hack because of a hack, appropriate.
- */
- peer_unlock (peer); /* god knows what reference... ACCEPT_PEER sucks */
}
return 0;
}