]> git.karo-electronics.de Git - karo-tx-linux.git/commit
tty: fix port buffer locking
authorVegard Nossum <vegard.nossum@oracle.com>
Thu, 11 May 2017 10:18:52 +0000 (12:18 +0200)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Thu, 18 May 2017 14:43:55 +0000 (16:43 +0200)
commit925bb1ce47f429f69aad35876df7ecd8c53deb7e
treed9dfa3cf733954e03995ce75dfdfa8a0a4ba4827
parent11d4d32158eeaf36fe1073a4a260193d7a19ccf1
tty: fix port buffer locking

tty_insert_flip_string_fixed_flag() is racy against itself when called
from the ioctl(TCXONC, TCION/TCIOFF) path [1] and the flush_to_ldisc()
workqueue path [2].

The problem is that port->buf.tail->used is modified without consistent
locking; the ioctl path takes tty->atomic_write_lock, whereas the workqueue
path takes ldata->output_lock.

We cannot simply take ldata->output_lock, since that is specific to the
N_TTY line discipline.

It might seem natural to try to take port->buf.lock inside
tty_insert_flip_string_fixed_flag() and friends (where port->buf is
actually used/modified), but this creates problems for flush_to_ldisc()
which takes it before grabbing tty->ldisc_sem, o_tty->termios_rwsem,
and ldata->output_lock.

Therefore, the simplest solution for now seems to be to take
tty->atomic_write_lock inside tty_port_default_receive_buf(). This lock
is also used in the write path [3] with a consistent ordering.

[1]: Call Trace:
 tty_insert_flip_string_fixed_flag
 pty_write
 tty_send_xchar                     // down_read(&o_tty->termios_rwsem)
                                    // mutex_lock(&tty->atomic_write_lock)
 n_tty_ioctl_helper
 n_tty_ioctl
 tty_ioctl                          // down_read(&tty->ldisc_sem)
 do_vfs_ioctl
 SyS_ioctl

[2]: Workqueue: events_unbound flush_to_ldisc
Call Trace:
 tty_insert_flip_string_fixed_flag
 pty_write
 tty_put_char
 __process_echoes
 commit_echoes                      // mutex_lock(&ldata->output_lock)
 n_tty_receive_buf_common
 n_tty_receive_buf2
 tty_ldisc_receive_buf              // down_read(&o_tty->termios_rwsem)
 tty_port_default_receive_buf       // down_read(&tty->ldisc_sem)
 flush_to_ldisc                     // mutex_lock(&port->buf.lock)
 process_one_work

[3]: Call Trace:
 tty_insert_flip_string_fixed_flag
 pty_write
 n_tty_write                        // mutex_lock(&ldata->output_lock)
                                    // down_read(&tty->termios_rwsem)
 do_tty_write (inline)              // mutex_lock(&tty->atomic_write_lock)
 tty_write                          // down_read(&tty->ldisc_sem)
 __vfs_write
 vfs_write
 SyS_write

The bug can result in about a dozen different crashes depending on what
exactly gets corrupted when port->buf.tail->used points outside the
buffer.

The patch passes my LOCKDEP/PROVE_LOCKING testing but more testing is
always welcome.

Found using syzkaller.

Cc: <stable@vger.kernel.org>
Signed-off-by: Vegard Nossum <vegard.nossum@oracle.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
drivers/tty/tty_port.c