drivers:tty:pty: Fix a race causing data loss on close
authorCorey Minyard <cminyard@mvista.com>
Tue, 24 Nov 2020 00:49:02 +0000 (18:49 -0600)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Thu, 7 Jan 2021 15:30:55 +0000 (16:30 +0100)
commit33d4ae98859873ddd49e22e4ca724387548b3d89
tree5e54534503026b25dacdd199e686d33e74a8f680
parentf446776ebffb9d3fb145b8d84f7f358f64cb54d6
drivers:tty:pty: Fix a race causing data loss on close

Remove the tty_vhangup() from the pty code and just release the
redirect.  The tty_vhangup() results in data loss and data out of order
issues.

If you write to a pty master an immediately close the pty master, the
receiver might get a chunk of data dropped, but then receive some later
data.  That's obviously something rather unexpected for a user.  It
certainly confused my test program.

It turns out that tty_vhangup() on the slave pty gets called from
pty_close(), and that causes the data on the slave side to be flushed,
but due to races more data can be copied into the slave side's buffer
after that.  Consider the following sequence:

thread1          thread2         thread3
-------          -------         -------
 |                |-write data into buffer,
 |                | n_tty buffer is filled
 |                | along with other buffers
 |                |-pty_close(master)
 |                |--tty_vhangup(slave)
 |                |---tty_ldisc_hangup()
 |                |----n_tty_flush_buffer()
 |                |-----reset_buffer_flags()
 |-n_tty_read()   |
 |--up_read(&tty->termios_rwsem);
 |                |------down_read(&tty->termios_rwsem)
 |                |------clear n_tty buffer contents
 |                |------up_read(&tty->termios_rwsem)
 |--tty_buffer_flush_work()       |
 |--schedules work calling        |
 |  flush_to_ldisc()              |
 |                                |-flush_to_ldisc()
 |                                |--receive_buf()
 |                                |---tty_port_default_receive_buf()
 |                                |----tty_ldisc_receive_buf()
 |                                |-----n_tty_receive_buf2()
 |                                |------n_tty_receive_buf_common()
 |                                |-------down_read(&tty->termios_rwsem)
 |                                |-------__receive_buf()
 |                                |       copies data into n_tty buffer
 |                                |-------up_read(&tty->termios_rwsem)
 |--down_read(&tty->termios_rwsem)
 |--copy buffer data to user

>From this sequence, you can see that thread2 writes to the buffer then
only clears the part of the buffer in n_tty.  The n_tty receive buffer
code then copies more data into the n_tty buffer.

But part of the vhangup, releasing the redirect, is still required to
avoid issues with consoles running on pty slaves.  So do that.
As far as I can tell, that is all that should be required.

Signed-off-by: Corey Minyard <cminyard@mvista.com>
Link: https://lore.kernel.org/r/20201124004902.1398477-3-minyard@acm.org
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
drivers/tty/pty.c
drivers/tty/tty_io.c