Fix race in tty_fasync() properly
authorLinus Torvalds <torvalds@linux-foundation.org>
Sun, 7 Feb 2010 18:11:23 +0000 (10:11 -0800)
committerLinus Torvalds <torvalds@linux-foundation.org>
Sun, 7 Feb 2010 18:26:01 +0000 (10:26 -0800)
This reverts commit 703625118069 ("tty: fix race in tty_fasync") and
commit b04da8bfdfbb ("fnctl: f_modown should call write_lock_irqsave/
restore") that tried to fix up some of the fallout but was incomplete.

It turns out that we really cannot hold 'tty->ctrl_lock' over calling
__f_setown, because not only did that cause problems with interrupt
disables (which the second commit fixed), it also causes a potential
ABBA deadlock due to lock ordering.

Thanks to Tetsuo Handa for following up on the issue, and running
lockdep to show the problem.  It goes roughly like this:

 - f_getown gets filp->f_owner.lock for reading without interrupts
   disabled, so an interrupt that happens while that lock is held can
   cause a lockdep chain from f_owner.lock -> sighand->siglock.

 - at the same time, the tty->ctrl_lock -> f_owner.lock chain that
   commit 703625118069 introduced, together with the pre-existing
   sighand->siglock -> tty->ctrl_lock chain means that we have a lock
   dependency the other way too.

So instead of extending tty->ctrl_lock over the whole __f_setown() call,
we now just take a reference to the 'pid' structure while holding the
lock, and then release it after having done the __f_setown.  That still
guarantees that 'struct pid' won't go away from under us, which is all
we really ever needed.

Reported-and-tested-by: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Acked-by: Greg Kroah-Hartman <gregkh@suse.de>
Acked-by: Américo Wang <xiyou.wangcong@gmail.com>
Cc: stable@kernel.org
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
drivers/char/tty_io.c
fs/fcntl.c

index c6f3b48be9dd68cc37c4847eb8cab4b268ec4c90..dcb9083ecde0982dd061720be8e2e324d3bde5ed 100644 (file)
@@ -1951,8 +1951,10 @@ static int tty_fasync(int fd, struct file *filp, int on)
                        pid = task_pid(current);
                        type = PIDTYPE_PID;
                }
-               retval = __f_setown(filp, pid, type, 0);
+               get_pid(pid);
                spin_unlock_irqrestore(&tty->ctrl_lock, flags);
+               retval = __f_setown(filp, pid, type, 0);
+               put_pid(pid);
                if (retval)
                        goto out;
        } else {
index 5ef953e6f908ea2d6f68061310a106fd0179d15a..97e01dc0d95fc4fe8464d729ce94d32b888b567d 100644 (file)
@@ -199,9 +199,7 @@ static int setfl(int fd, struct file * filp, unsigned long arg)
 static void f_modown(struct file *filp, struct pid *pid, enum pid_type type,
                      int force)
 {
-       unsigned long flags;
-
-       write_lock_irqsave(&filp->f_owner.lock, flags);
+       write_lock_irq(&filp->f_owner.lock);
        if (force || !filp->f_owner.pid) {
                put_pid(filp->f_owner.pid);
                filp->f_owner.pid = get_pid(pid);
@@ -213,7 +211,7 @@ static void f_modown(struct file *filp, struct pid *pid, enum pid_type type,
                        filp->f_owner.euid = cred->euid;
                }
        }
-       write_unlock_irqrestore(&filp->f_owner.lock, flags);
+       write_unlock_irq(&filp->f_owner.lock);
 }
 
 int __f_setown(struct file *filp, struct pid *pid, enum pid_type type,