bgpd: write NOTIFY non-blockingly

switching the socket to blocking may well block the entire bgpd process
for some time if our peer is overloaded (which may well be the original
reason for the NOTIFY)

The error handling is slightly different from the previous ML discussion
on this;  buffer exhaustion isn't technically a fatal TCP error, and we
should probably proceed with FSM actions according to a sent NOTIFY
(adjusting timers) even if we didn't manage to get the NOTIFY onto the
wire.

Acked-by: Leonid Rosenboim <lrosenbo@wrs.com>
Signed-off-by: David Lamparter <equinox@opensourcerouting.org>

1 files changed, 5 insertions, 4 deletions

diff --git a/bgpd/bgp_packet.c b/bgpd/bgp_packet.c
index d115353f..d71df082 100644
--- a/bgpd/bgp_packet.c
+++ b/bgpd/bgp_packet.c
@@ -719,14 +719,15 @@ bgp_write_notify (struct peer *peer)
     return 0;
   assert (stream_get_endp (s) >= BGP_HEADER_SIZE);
 
-  /* Put socket in blocking mode. */
-  val = fcntl (peer->fd, F_GETFL, 0);
-  fcntl (peer->fd, F_SETFL, val & ~O_NONBLOCK);
-
   /* Stop collecting data within the socket */
   sockopt_cork (peer->fd, 0);
 
+  /* socket is in nonblocking mode, if we can't deliver the NOTIFY, well,
+   * we only care about getting a clean shutdown at this point. */
   ret = write (peer->fd, STREAM_DATA (s), stream_get_endp (s));
+
+  /* only connection reset/close gets counted as TCP_fatal_error, failure
+   * to write the entire NOTIFY doesn't get different FSM treatment */
   if (ret <= 0)
     {
       BGP_EVENT_ADD (peer, TCP_fatal_error);