Fix busy Libevent loops (infinite loops in Shadow)

There is a bug causing busy loops in Libevent and infinite loops in
the Shadow simulator. A connection that is marked for close, wants
to flush, is held open to flush, but is rate limited (the token
bucket is empty) triggers the bug.

This commit fixes the bug. Details are below.

This currently happens on read and write callbacks when the active
socket is marked for close. In this case, Tor doesn't actually try
to complete the read or write (it returns from those methods when
marked), but instead tries to clear the connection with
conn_close_if_marked(). Tor will not close a marked connection that
contains data: it must be flushed first. The bug occurs when this
flush operation on the marked connection can not occur because the
connection is rate-limited (its write token bucket is empty).

The fix is to detect when rate limiting is preventing a marked
connection from properly flushing. In this case, it should be
flagged as read/write_blocked_on_bandwidth and the read/write events
de-registered from Libevent. When the token bucket gets refilled, it
will check the associated read/write_blocked_on_bandwidth flag, and
add the read/write event back to Libevent, which will cause it to
fire. This time, it will be properly flushed and closed.

The reason that both read and write events are both de-registered
when the marked connection can not flush is because both result in
the same behavior. Both read/write events on marked connections will
never again do any actual reads/writes, and are only useful to
trigger the flush and close the connection. By setting the
associated read/write_blocked_on_bandwidth flag, we ensure that the
event will get added back to Libevent, properly flushed, and closed.

Why is this important? Every Shadow event occurs at a discrete time
instant. If Tor does not properly deregister Libevent events that
fire but result in Tor essentially doing nothing, Libevent will
repeatedly fire the event. In Shadow this means infinite loop,
outside of Shadow this means wasted CPU cycles.

1 files changed, 14 insertions, 0 deletions

diff --git a/src/or/main.c b/src/or/main.c
index 163dd5d0fb..a5ad1d9648 100644
--- a/src/or/main.c
+++ b/src/or/main.c
@@ -848,6 +848,20 @@ conn_close_if_marked(int i)
                            "Holding conn (fd %d) open for more flushing.",
                            (int)conn->s));
         conn->timestamp_lastwritten = now; /* reset so we can flush more */
+      } else if (sz == 0) { /* retval is also 0 */
+    	/* Connection must flush before closing, but it's being rate-limited.
+    	   Let's remove from Libevent, and mark it as blocked on bandwidth so it
+    	   will be re-added on next token bucket refill. Prevents busy Libevent
+    	   loops where we keep ending up here and returning 0 until we are no
+    	   longer blocked on bandwidth. */
+        if (connection_is_reading(conn)) {
+          conn->read_blocked_on_bw = 1;
+          connection_stop_reading(conn);
+        }
+        if (connection_is_writing(conn)) {
+          conn->write_blocked_on_bw = 1;
+          connection_stop_writing(conn);
+        }
       }
       return 0;
     }