tty: Fix race condition if flushing tty flip buffers
authorPeter Hurley <peter@hurleysoftware.com>
Wed, 20 Mar 2013 17:20:43 +0000 (13:20 -0400)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Wed, 10 Apr 2013 00:05:02 +0000 (17:05 -0700)
As Ilya Zykov identified in his patch 'PROBLEM: Race condition in
tty buffer's function flush_to_ldisc()', a race condition exists
which allows a parallel flush_to_ldisc() to flush and free the tty
flip buffers while those buffers are in-use. For example,

  CPU 0                         |  CPU 1                                  |  CPU 2
                                | flush_to_ldisc()                        |
                                |  grab spin lock                         |
tty_buffer_flush()              |                                         | flush_to_ldisc()
 wait for spin lock             |                                         |  wait for spin lock
                                |   if (!test_and_set_bit(TTYP_FLUSHING)) |
                                |    while (next flip buffer)             |
                                |     ...                                 |
                                |     drop spin lock                      |
  grab spin lock                |                                         |
   if (test_bit(TTYP_FLUSHING)) |                                         |
    set_bit(TTYP_FLUSHPENDING)  |      receive_buf()                      |
    drop spin lock              |                                         |
                                |                                         |   grab spin lock
                                |                                         |    if (!test_and_set_bit(TTYP_FLUSHING))
                                |                                         |    if (test_bit(TTYP_FLUSHPENDING))
                                |                                         |    __tty_buffer_flush()

CPU 2 has just flushed and freed all tty flip buffers while CPU 1 is
transferring data from the head flip buffer.

The original patch was rejected under the assumption that parallel
flush_to_ldisc() was not possible. Because of necessary changes to
the workqueue api, work items can execute in parallel on SMP.

This patch differs slightly from the original patch by testing for
a pending flush _after_ each receive_buf(), since TTYP_FLUSHPENDING
can only be set while the lock is dropped around receive_buf().

Reported-by: Ilya Zykov <linux@izyk.ru>
Signed-off-by: Peter Hurley <peter@hurleysoftware.com>
Acked-by: Ilya Zykov <linux@izyk.ru>
Cc: Jiri Slaby <jslaby@suse.cz>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
drivers/tty/tty_buffer.c

index 578aa75..9121c1f 100644 (file)
@@ -449,11 +449,6 @@ static void flush_to_ldisc(struct work_struct *work)
                                tty_buffer_free(port, head);
                                continue;
                        }
-                       /* Ldisc or user is trying to flush the buffers
-                          we are feeding to the ldisc, stop feeding the
-                          line discipline as we want to empty the queue */
-                       if (test_bit(TTYP_FLUSHPENDING, &port->iflags))
-                               break;
                        if (!tty->receive_room)
                                break;
                        if (count > tty->receive_room)
@@ -465,17 +460,20 @@ static void flush_to_ldisc(struct work_struct *work)
                        disc->ops->receive_buf(tty, char_buf,
                                                        flag_buf, count);
                        spin_lock_irqsave(&buf->lock, flags);
+                       /* Ldisc or user is trying to flush the buffers.
+                          We may have a deferred request to flush the
+                          input buffer, if so pull the chain under the lock
+                          and empty the queue */
+                       if (test_bit(TTYP_FLUSHPENDING, &port->iflags)) {
+                               __tty_buffer_flush(port);
+                               clear_bit(TTYP_FLUSHPENDING, &port->iflags);
+                               wake_up(&tty->read_wait);
+                               break;
+                       }
                }
                clear_bit(TTYP_FLUSHING, &port->iflags);
        }
 
-       /* We may have a deferred request to flush the input buffer,
-          if so pull the chain under the lock and empty the queue */
-       if (test_bit(TTYP_FLUSHPENDING, &port->iflags)) {
-               __tty_buffer_flush(port);
-               clear_bit(TTYP_FLUSHPENDING, &port->iflags);
-               wake_up(&tty->read_wait);
-       }
        spin_unlock_irqrestore(&buf->lock, flags);
 
        tty_ldisc_deref(disc);