2021-08-23 00:07:39

by Nguyen Dinh Phi

Subject: [PATCH v2] tty: Fix data race between tiocsti() and flush_to_ldisc()

The ops->receive_buf() may be accessed concurrently from these two
functions. If the driver flushes data to the line discipline
receive_buf() method while tiocsti() is waiting for the
ops->receive_buf() to finish its work, the data race will happen.

For example:
tty_ioctl |tty_ldisc_receive_buf
->tioctsi | ->tty_port_default_receive_buf
| ->tty_ldisc_receive_buf
->hci_uart_tty_receive | ->hci_uart_tty_receive
->h4_recv | ->h4_recv

In this case, the h4 receive buffer will be overwritten by the
latecomer, and we will lost the data.

Hence, change tioctsi() function to use the exclusive lock interface
from tty_buffer to avoid the data race.

Signed-off-by: Nguyen Dinh Phi <[email protected]>
Reported-by: [email protected]
---
V2:
- Remove FIXME comment.

drivers/tty/tty_io.c | 4 ++--
1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/drivers/tty/tty_io.c b/drivers/tty/tty_io.c
index e8532006e960..6616d4a0d41d 100644
--- a/drivers/tty/tty_io.c
+++ b/drivers/tty/tty_io.c
@@ -2290,8 +2290,6 @@ static int tty_fasync(int fd, struct file *filp, int on)
* Locking:
* Called functions take tty_ldiscs_lock
* current->signal->tty check is safe without locks
- *
- * FIXME: may race normal receive processing
*/

static int tiocsti(struct tty_struct *tty, char __user *p)
@@ -2307,8 +2305,10 @@ static int tiocsti(struct tty_struct *tty, char __user *p)
ld = tty_ldisc_ref_wait(tty);
if (!ld)
return -EIO;
+ tty_buffer_lock_exclusive(tty->port);
if (ld->ops->receive_buf)
ld->ops->receive_buf(tty, &ch, &mbz, 1);
+ tty_buffer_unlock_exclusive(tty->port);
tty_ldisc_deref(ld);
return 0;
}
--
2.25.1

2021-08-23 08:53:53

by Jiri Slaby

Subject: Re: [PATCH v2] tty: Fix data race between tiocsti() and flush_to_ldisc()

On 23. 08. 21, 2:06, Nguyen Dinh Phi wrote:
> The ops->receive_buf() may be accessed concurrently from these two
> functions. If the driver flushes data to the line discipline
> receive_buf() method while tiocsti() is waiting for the
> ops->receive_buf() to finish its work, the data race will happen.
>
> For example:
> tty_ioctl |tty_ldisc_receive_buf
> ->tioctsi | ->tty_port_default_receive_buf
> | ->tty_ldisc_receive_buf
> ->hci_uart_tty_receive | ->hci_uart_tty_receive
> ->h4_recv | ->h4_recv
>
> In this case, the h4 receive buffer will be overwritten by the
> latecomer, and we will lost the data.
>
> Hence, change tioctsi() function to use the exclusive lock interface
> from tty_buffer to avoid the data race.

This makes sense. It mimics paste_selection() -- the lock order is
preexisting. The normal path processing has the locks inverted, not sure
why it doesn't matter :):
flush_to_ldisc
mutex_lock(&buf->lock); <--------- A (mutex)
receive_buf()
tty_port_default_receive_buf()
disc = tty_ldisc_ref(tty); <----- B (ldsem for READ)

Reviewed-by: Jiri Slaby <[email protected]>

> Signed-off-by: Nguyen Dinh Phi <[email protected]>
> Reported-by: [email protected]
> ---
> V2:
> - Remove FIXME comment.
>
> drivers/tty/tty_io.c | 4 ++--
> 1 file changed, 2 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/tty/tty_io.c b/drivers/tty/tty_io.c
> index e8532006e960..6616d4a0d41d 100644
> --- a/drivers/tty/tty_io.c
> +++ b/drivers/tty/tty_io.c
> @@ -2290,8 +2290,6 @@ static int tty_fasync(int fd, struct file *filp, int on)
> * Locking:
> * Called functions take tty_ldiscs_lock
> * current->signal->tty check is safe without locks
> - *
> - * FIXME: may race normal receive processing
> */
>
> static int tiocsti(struct tty_struct *tty, char __user *p)
> @@ -2307,8 +2305,10 @@ static int tiocsti(struct tty_struct *tty, char __user *p)
> ld = tty_ldisc_ref_wait(tty);

Here (and in paste_selection()), it is lock "B (ldsem) for READ".


> if (!ld)
> return -EIO;
> + tty_buffer_lock_exclusive(tty->port);

and "A (mutex)".

> if (ld->ops->receive_buf)
> ld->ops->receive_buf(tty, &ch, &mbz, 1);
> + tty_buffer_unlock_exclusive(tty->port);
> tty_ldisc_deref(ld);
> return 0;
> }

thanks,
--
js
suse labs