sync: kill git fetch process before SSH control master process

If the SSH control master process is killed while an active git
fetch is using its network socket, the underlying SSH client may
not realize the connection was broken.  This can lead to both the
client and the server waiting indefinitely for network messages
which will never be sent.

Work around the problem by keeping track of any processes that use
the tunnels we establish.  If we are about to kill any of the SSH
control masters that we started, ensure the clients using them are
successfully killed first.

Change-Id: Ida6c124dcb0c6a26bf7dd69cba2fbdc2ecd5b2fc
Signed-off-by: Shawn O. Pearce <sop@google.com>

git_config.py

@@ -23,7 +23,10 @@ from signal import SIGTERM
 from urllib2 import urlopen, HTTPError
 from error import GitError, UploadError
 from trace import Trace
-from git_command import GitCommand, ssh_sock
+
+from git_command import GitCommand
+from git_command import ssh_sock
+from git_command import terminate_ssh_clients
 
 R_HEADS = 'refs/heads/'
 R_TAGS  = 'refs/tags/'
@@ -391,6 +394,8 @@ def _open_ssh(host, port):
   return True
 
 def close_ssh():
+  terminate_ssh_clients()
+
   for key,p in _ssh_cache.iteritems():
     try:
       os.kill(p.pid, SIGTERM)