Merge branch 'patches/sh_bgp-write-cork' into dn42

Conflicts:
	lib/sockunion.c

10 files changed, 50 insertions, 82 deletions

diff --git a/bgpd/bgp_fsm.c b/bgpd/bgp_fsm.c
index 487ebddb..e18fc8cd 100644
--- a/bgpd/bgp_fsm.c
+++ b/bgpd/bgp_fsm.c
@@ -704,7 +704,7 @@ bgp_start (struct peer *peer)
 		    peer->fd);
 	  return -1;
 	}
-      BGP_READ_ON (peer->t_read, bgp_read, peer->fd);
+
       BGP_WRITE_ON (peer->t_write, bgp_write, peer->fd);
       break;
     }
diff --git a/bgpd/bgp_network.c b/bgpd/bgp_network.c
index be01bbd6..83b1142e 100644
--- a/bgpd/bgp_network.c
+++ b/bgpd/bgp_network.c
@@ -150,6 +150,7 @@ bgp_accept (struct thread *thread)
       zlog_err ("[Error] BGP socket accept failed (%s)", safe_strerror (errno));
       return -1;
     }
+  set_nonblocking (bgp_sock);
 
   if (BGP_DEBUG (events, EVENTS))
     zlog_debug ("[Event] BGP connection from host %s", inet_sutop (&su, buf));
diff --git a/bgpd/bgp_packet.c b/bgpd/bgp_packet.c
index b29bc1f0..1e2f95e3 100644
--- a/bgpd/bgp_packet.c
+++ b/bgpd/bgp_packet.c
@@ -107,8 +107,7 @@ bgp_connect_check (struct peer *peer)
   socklen_t slen;
   int ret;
 
-  /* Anyway I have to reset read and write thread. */
-  BGP_READ_OFF (peer->t_read);
+  /* Anyway I have to reset write thread. */
   BGP_WRITE_OFF (peer->t_write);
 
   /* Check file descriptor. */
@@ -597,7 +596,6 @@ bgp_write (struct thread *thread)
   struct stream *s; 
   int num;
   unsigned int count = 0;
-  int write_errno;
 
   /* Yes first of all get peer pointer. */
   peer = THREAD_ARG (thread);
@@ -610,46 +608,37 @@ bgp_write (struct thread *thread)
       return 0;
     }
 
-    /* Nonblocking write until TCP output buffer is full.  */
-  while (1)
+  s = bgp_write_packet (peer);
+  if (!s)
+    return 0;	/* nothing to send */
+
+  sockopt_cork (peer->fd, 1);
+
+  /* Nonblocking write until TCP output buffer is full.  */
+  do
     {
       int writenum;
-      int val;
-
-      s = bgp_write_packet (peer);
-      if (! s)
-	return 0;
-      
-      /* XXX: FIXME, the socket should be NONBLOCK from the start
-       * status shouldnt need to be toggled on each write
-       */
-      val = fcntl (peer->fd, F_GETFL, 0);
-      fcntl (peer->fd, F_SETFL, val|O_NONBLOCK);
 
       /* Number of bytes to be sent.  */
       writenum = stream_get_endp (s) - stream_get_getp (s);
 
       /* Call write() system call.  */
       num = write (peer->fd, STREAM_PNT (s), writenum);
-      write_errno = errno;
-      fcntl (peer->fd, F_SETFL, val);
-      if (num <= 0)
+      if (num < 0)
 	{
-	  /* Partial write. */
-	  if (write_errno == EWOULDBLOCK || write_errno == EAGAIN)
-	      break;
+	  /* write failed either retry needed or error */
+	  if (ERRNO_IO_RETRY(errno))
+	    break;
 
 	  BGP_EVENT_ADD (peer, TCP_fatal_error);
 	  return 0;
 	}
+
       if (num != writenum)
 	{
+	  /* Partial write */
 	  stream_forward_getp (s, num);
-
-	  if (write_errno == EAGAIN)
-	    break;
-
-	  continue;
+	  break;
 	}
 
       /* Retrieve BGP packet type. */
@@ -690,13 +679,14 @@ bgp_write (struct thread *thread)
 
       /* OK we send packet so delete it. */
       bgp_packet_delete (peer);
-
-      if (++count >= BGP_WRITE_PACKET_MAX)
-	break;
     }
+  while (++count < BGP_WRITE_PACKET_MAX &&
+	 (s = bgp_write_packet (peer)) != NULL);
   
   if (bgp_write_proceed (peer))
     BGP_WRITE_ON (peer->t_write, bgp_write, peer->fd);
+  else
+    sockopt_cork (peer->fd, 0);
   
   return 0;
 }
@@ -705,7 +695,7 @@ bgp_write (struct thread *thread)
 static int
 bgp_write_notify (struct peer *peer)
 {
-  int ret;
+  int ret, val;
   u_char type;
   struct stream *s; 
 
@@ -715,7 +705,10 @@ bgp_write_notify (struct peer *peer)
     return 0;
   assert (stream_get_endp (s) >= BGP_HEADER_SIZE);
 
-  /* I'm not sure fd is writable. */
+  /* Put socket in blocking mode. */
+  val = fcntl (peer->fd, F_GETFL, 0);
+  fcntl (peer->fd, F_SETFL, val & ~O_NONBLOCK);
+
   ret = writen (peer->fd, STREAM_DATA (s), stream_get_endp (s));
   if (ret <= 0)
     {
@@ -2237,12 +2230,13 @@ bgp_read_packet (struct peer *peer)
     return 0;
 
   /* Read packet from fd. */
-  nbytes = stream_read_unblock (peer->ibuf, peer->fd, readsize);
+  nbytes = stream_read_try (peer->ibuf, peer->fd, readsize);
 
   /* If read byte is smaller than zero then error occured. */
   if (nbytes < 0) 
     {
-      if (errno == EAGAIN)
+      /* Transient error should retry */
+      if (nbytes == -2)
 	return -1;
 
       plog_err (peer->log, "%s [Error] bgp_read_packet error: %s",
@@ -2319,22 +2313,14 @@ bgp_read (struct thread *thread)
   peer = THREAD_ARG (thread);
   peer->t_read = NULL;
 
-  /* For non-blocking IO check. */
-  if (peer->status == Connect)
-    {
-      bgp_connect_check (peer);
-      goto done;
-    }
-  else
+  if (peer->fd < 0)
     {
-      if (peer->fd < 0)
-	{
-	  zlog_err ("bgp_read peer's fd is negative value %d", peer->fd);
-	  return -1;
-	}
-      BGP_READ_ON (peer->t_read, bgp_read, peer->fd);
+      zlog_err ("bgp_read peer's fd is negative value %d", peer->fd);
+      return -1;
     }
 
+  BGP_READ_ON (peer->t_read, bgp_read, peer->fd);
+
   /* Read packet header to determine type of the packet */
   if (peer->packet_size == 0)
     peer->packet_size = BGP_HEADER_SIZE;
diff --git a/bgpd/bgpd.c b/bgpd/bgpd.c
index 882fe37c..a03025f6 100644
--- a/bgpd/bgpd.c
+++ b/bgpd/bgpd.c
@@ -2257,6 +2257,9 @@ peer_change_action (struct peer *peer, afi_t afi, safi_t safi,
   if (CHECK_FLAG (peer->sflags, PEER_STATUS_GROUP))
     return;
 
+  if (peer->status != Established)
+    return;
+
   if (type == peer_change_reset)
     bgp_notify_send (peer, BGP_NOTIFY_CEASE,
 		     BGP_NOTIFY_CEASE_CONFIG_CHANGE);
diff --git a/lib/sockunion.c b/lib/sockunion.c
index d1f6d80b..5de3bcfc 100644
--- a/lib/sockunion.c
+++ b/lib/sockunion.c
@@ -551,6 +551,16 @@ sockopt_v6only (int family, int sock)
   return 0;
 }
 
+int
+sockopt_cork (int sock, int onoff)
+{
+#ifdef TCP_CORK
+  return setsockopt (sock, IPPROTO_TCP, TCP_CORK, &onoff, sizeof(onoff));
+#else
+  return 0;
+#endif
+}
+
 /* If same family and same prefix return 1. */
 int
 sockunion_same (union sockunion *su1, union sockunion *su2)
diff --git a/lib/sockunion.h b/lib/sockunion.h
index db145cf2..2e8ce2d0 100644
--- a/lib/sockunion.h
+++ b/lib/sockunion.h
@@ -103,6 +103,7 @@ extern int sockopt_v6only (int family, int sock);
 extern int sockunion_bind (int sock, union sockunion *, 
                            unsigned short, union sockunion *);
 extern int sockopt_ttl (int family, int sock, int ttl);
+extern int sockopt_cork (int sock, int onoff);
 extern int sockunion_socket (union sockunion *su);
 extern const char *inet_sutop (union sockunion *su, char *str);
 extern enum connect_result sockunion_connect (int fd, union sockunion *su, 
diff --git a/lib/stream.c b/lib/stream.c
index 983330ff..bff74d7b 100644
--- a/lib/stream.c
+++ b/lib/stream.c
@@ -711,32 +711,6 @@ stream_read (struct stream *s, int fd, size_t size)
   return nbytes;
 }
 
-/* Read size from fd. */
-int
-stream_read_unblock (struct stream *s, int fd, size_t size)
-{
-  int nbytes;
-  int val;
-  
-  STREAM_VERIFY_SANE(s);
-  
-  if (STREAM_WRITEABLE (s) < size)
-    {
-      STREAM_BOUND_WARN (s, "put");
-      return 0;
-    }
-  
-  val = fcntl (fd, F_GETFL, 0);
-  fcntl (fd, F_SETFL, val|O_NONBLOCK);
-  nbytes = read (fd, s->data + s->endp, size);
-  fcntl (fd, F_SETFL, val);
-
-  if (nbytes > 0)
-    s->endp += nbytes;
-  
-  return nbytes;
-}
-
 ssize_t
 stream_read_try(struct stream *s, int fd, size_t size)
 {
diff --git a/lib/stream.h b/lib/stream.h
index 3e4ba7b4..486a3f93 100644
--- a/lib/stream.h
+++ b/lib/stream.h
@@ -181,10 +181,6 @@ extern u_int32_t stream_get_ipv4 (struct stream *);
    Use stream_read_try instead.  */
 extern int stream_read (struct stream *, int, size_t);
 
-/* Deprecated: all file descriptors should already be non-blocking.
-   Will be removed.  Use stream_read_try instead. */
-extern int stream_read_unblock (struct stream *, int, size_t);
-
 /* Read up to size bytes into the stream.
    Return code:
      >0: number of bytes read
diff --git a/lib/thread.c b/lib/thread.c
index 70f8defa..1e714417 100644
--- a/lib/thread.c
+++ b/lib/thread.c
@@ -958,7 +958,6 @@ thread_fetch (struct thread_master *m, struct thread *fetch)
   struct thread *thread;
   fd_set readfd;
   fd_set writefd;
-  fd_set exceptfd;
   struct timeval timer_val;
   struct timeval timer_val_bg;
   struct timeval *timer_wait;
@@ -984,7 +983,6 @@ thread_fetch (struct thread_master *m, struct thread *fetch)
       /* Structure copy.  */
       readfd = m->readfd;
       writefd = m->writefd;
-      exceptfd = m->exceptfd;
       
       /* Calculate select wait timer if nothing else to do */
       quagga_get_relative (NULL);
@@ -995,7 +993,7 @@ thread_fetch (struct thread_master *m, struct thread *fetch)
 	  (!timer_wait || (timeval_cmp (*timer_wait, *timer_wait_bg) > 0)))
 	timer_wait = timer_wait_bg;
       
-      num = select (FD_SETSIZE, &readfd, &writefd, &exceptfd, timer_wait);
+      num = select (FD_SETSIZE, &readfd, &writefd, NULL, timer_wait);
       
       /* Signals should get quick treatment */
       if (num < 0)
diff --git a/lib/thread.h b/lib/thread.h
index e5b1cb35..a6aed173 100644
--- a/lib/thread.h
+++ b/lib/thread.h
@@ -53,7 +53,6 @@ struct thread_master
   struct thread_list background;
   fd_set readfd;
   fd_set writefd;
-  fd_set exceptfd;
   unsigned long alloc;
 };