This reverts commit 7036251180 ("tty: fix race in tty_fasync") and
commit b04da8bfdf ("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 7036251180 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>