Hi, Guys:
It was found that the 3.12 kernel tty layer will lose or corrupt data
when have a full-duplex communication, especially in high baud rate, for
example 230k for my OMAP5 uart. Eventually I found there is lock missing
between copy data to ldisc layer buffer and copy data from the same
buffer to user space. I believe this issue existed since 3.8
kernel(since this kernel , it start to remove most of the spin-locks)
and I didn't find any fix even through 3.17 kernel. This patch was
tested to be works great with no any data loss again on 3.12 kernel.
This patch was built for the latest kernel, but I cannot test it.
Somebody may give a test.
I did try to use the existed lock atomic_read_lock, but it doesn’t work.
Signed-off-by: Hui Du <[email protected]>
---
drivers/tty/n_tty.c | 9 ++++++++-
1 file changed, 8 insertions(+), 1 deletion(-)
diff --git a/drivers/tty/n_tty.c b/drivers/tty/n_tty.c
index 2e900a9..6e5c6ae 100644
--- a/drivers/tty/n_tty.c
+++ b/drivers/tty/n_tty.c
@@ -125,6 +125,7 @@ struct n_tty_data {
struct mutex atomic_read_lock;
struct mutex output_lock;
+ struct mutex read_buf_lock;
};
static inline size_t read_cnt(struct n_tty_data *ldata)
@@ -1691,7 +1692,7 @@ n_tty_receive_buf_common(struct tty_struct *tty, const unsigned char *cp,
int room, n, rcvd = 0;
down_read(&tty->termios_rwsem);
-
+ mutex_lock(&ldata->read_buf_lock);
while (1) {
room = receive_room(tty);
n = min(count, room);
@@ -1710,6 +1711,7 @@ n_tty_receive_buf_common(struct tty_struct *tty, const unsigned char *cp,
tty->receive_room = room;
n_tty_check_throttle(tty);
+ mutex_unlock(&ldata->read_buf_lock);
up_read(&tty->termios_rwsem);
return rcvd;
@@ -1876,6 +1878,7 @@ static int n_tty_open(struct tty_struct *tty)
ldata->overrun_time = jiffies;
mutex_init(&ldata->atomic_read_lock);
mutex_init(&ldata->output_lock);
+ mutex_init(&ldata->read_buf_lock);
tty->disc_data = ldata;
reset_buffer_flags(tty->disc_data);
@@ -1939,6 +1942,7 @@ static int copy_from_read_buf(struct tty_struct *tty,
size_t tail = ldata->read_tail & (N_TTY_BUF_SIZE - 1);
retval = 0;
+ mutex_lock(&ldata->read_buf_lock);
n = min(read_cnt(ldata), N_TTY_BUF_SIZE - tail);
n = min(*nr, n);
if (n) {
@@ -1954,6 +1958,7 @@ static int copy_from_read_buf(struct tty_struct *tty,
*b += n;
*nr -= n;
}
+ mutex_unlock(&ldata->read_buf_lock);
return retval;
}
@@ -1992,6 +1997,7 @@ static int canon_copy_from_read_buf(struct tty_struct *tty,
bool eof_push = 0;
/* N.B. avoid overrun if nr == 0 */
+ mutex_lock(&ldata->read_buf_lock);
n = min(*nr, read_cnt(ldata));
if (!n)
return 0;
@@ -2052,6 +2058,7 @@ static int canon_copy_from_read_buf(struct tty_struct *tty,
ldata->push = 0;
tty_audit_push(tty);
}
+ mutex_unlock(&ldata->read_buf_lock);
return eof_push ? -EAGAIN : 0;
}
--
1.7.10.4
On 12/10/2014, 07:38 PM, Denis Du wrote:
>
> Hi, Guys:
Hi, are you sending this using some robot? I think I have seen like ten
copies of this patch already.
> It was found that the 3.12 kernel tty layer will lose or corrupt data
> when have a full-duplex communication, especially in high baud rate, for
> example 230k for my OMAP5 uart. Eventually I found there is lock missing
> between copy data to ldisc layer buffer and copy data from the same
> buffer to user space. I believe this issue existed since 3.8
> kernel(since this kernel , it start to remove most of the spin-locks)
> and I didn't find any fix even through 3.17 kernel. This patch was
> tested to be works great with no any data loss again on 3.12 kernel.
>
> This patch was built for the latest kernel, but I cannot test it.
> Somebody may give a test.
>
> I did try to use the existed lock atomic_read_lock, but it doesn’t work.
Anyway, adding Peter Hurley to CC who was working on eliminating locks
from this code lately. More precisely since 3.12 we have no locks there,
which would explain why are you seeing it starting 3.12.
> Signed-off-by: Hui Du <[email protected]>
>
> ---
> drivers/tty/n_tty.c | 9 ++++++++-
> 1 file changed, 8 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/tty/n_tty.c b/drivers/tty/n_tty.c
> index 2e900a9..6e5c6ae 100644
> --- a/drivers/tty/n_tty.c
> +++ b/drivers/tty/n_tty.c
> @@ -125,6 +125,7 @@ struct n_tty_data {
>
> struct mutex atomic_read_lock;
> struct mutex output_lock;
> + struct mutex read_buf_lock;
> };
>
> static inline size_t read_cnt(struct n_tty_data *ldata)
> @@ -1691,7 +1692,7 @@ n_tty_receive_buf_common(struct tty_struct *tty,
> const unsigned char *cp,
> int room, n, rcvd = 0;
>
> down_read(&tty->termios_rwsem);
> -
> + mutex_lock(&ldata->read_buf_lock);
> while (1) {
> room = receive_room(tty);
> n = min(count, room);
> @@ -1710,6 +1711,7 @@ n_tty_receive_buf_common(struct tty_struct *tty,
> const unsigned char *cp,
>
> tty->receive_room = room;
> n_tty_check_throttle(tty);
> + mutex_unlock(&ldata->read_buf_lock);
> up_read(&tty->termios_rwsem);
>
> return rcvd;
> @@ -1876,6 +1878,7 @@ static int n_tty_open(struct tty_struct *tty)
> ldata->overrun_time = jiffies;
> mutex_init(&ldata->atomic_read_lock);
> mutex_init(&ldata->output_lock);
> + mutex_init(&ldata->read_buf_lock);
>
> tty->disc_data = ldata;
> reset_buffer_flags(tty->disc_data);
> @@ -1939,6 +1942,7 @@ static int copy_from_read_buf(struct tty_struct *tty,
> size_t tail = ldata->read_tail & (N_TTY_BUF_SIZE - 1);
>
> retval = 0;
> + mutex_lock(&ldata->read_buf_lock);
> n = min(read_cnt(ldata), N_TTY_BUF_SIZE - tail);
> n = min(*nr, n);
> if (n) {
> @@ -1954,6 +1958,7 @@ static int copy_from_read_buf(struct tty_struct *tty,
> *b += n;
> *nr -= n;
> }
> + mutex_unlock(&ldata->read_buf_lock);
> return retval;
> }
>
> @@ -1992,6 +1997,7 @@ static int canon_copy_from_read_buf(struct
> tty_struct *tty,
> bool eof_push = 0;
>
> /* N.B. avoid overrun if nr == 0 */
> + mutex_lock(&ldata->read_buf_lock);
> n = min(*nr, read_cnt(ldata));
> if (!n)
> return 0;
> @@ -2052,6 +2058,7 @@ static int canon_copy_from_read_buf(struct
> tty_struct *tty,
> ldata->push = 0;
> tty_audit_push(tty);
> }
> + mutex_unlock(&ldata->read_buf_lock);
> return eof_push ? -EAGAIN : 0;
> }
>
--
js
suse labs
On 12/10/2014, 07:38 PM, Denis Du wrote:
>
> Hi, Guys:
Hi, are you sending this using some robot? I think I have seen like ten
copies of this patch already.
Sorry, I always have troubles to send the patch in correct format, even now I am not
100% sure this patch have the correct format. My company's firewall policy messed up
many things, I have to try my yahoo personal e-mail.
> It was found that the 3.12 kernel tty layer will lose or corrupt data
> when have a full-duplex communication, especially in high baud rate, for
> example 230k for my OMAP5 uart. Eventually I found there is lock missing
> between copy data to ldisc layer buffer and copy data from the same
> buffer to user space. I believe this issue existed since 3.8
> kernel(since this kernel , it start to remove most of the spin-locks)
> and I didn't find any fix even through 3.17 kernel. This patch was
> tested to be works great with no any data loss again on 3.12 kernel.
>
> This patch was built for the latest kernel, but I cannot test it.
> Somebody may give a test.
>
> I did try to use the existed lock atomic_read_lock, but it doesn’t work.
Anyway, adding Peter Hurley to CC who was working on eliminating locks
from this code lately. More precisely since 3.12 we have no locks there,
which would explain why are you seeing it starting 3.12.
> Signed-off-by: Hui Du <[email protected]>
>
> ---
> drivers/tty/n_tty.c | 9 ++++++++-
> 1 file changed, 8 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/tty/n_tty.c b/drivers/tty/n_tty.c
> index 2e900a9..6e5c6ae 100644
> --- a/drivers/tty/n_tty.c
> +++ b/drivers/tty/n_tty.c
> @@ -125,6 +125,7 @@ struct n_tty_data {
>
> struct mutex atomic_read_lock;
> struct mutex output_lock;
> + struct mutex read_buf_lock;
> };
>
> static inline size_t read_cnt(struct n_tty_data *ldata)
> @@ -1691,7 +1692,7 @@ n_tty_receive_buf_common(struct tty_struct *tty,
> const unsigned char *cp,
> int room, n, rcvd = 0;
>
> down_read(&tty->termios_rwsem);
> -
> + mutex_lock(&ldata->read_buf_lock);
> while (1) {
> room = receive_room(tty);
> n = min(count, room);
> @@ -1710,6 +1711,7 @@ n_tty_receive_buf_common(struct tty_struct *tty,
> const unsigned char *cp,
>
> tty->receive_room = room;
> n_tty_check_throttle(tty);
> + mutex_unlock(&ldata->read_buf_lock);
> up_read(&tty->termios_rwsem);
>
> return rcvd;
> @@ -1876,6 +1878,7 @@ static int n_tty_open(struct tty_struct *tty)
> ldata->overrun_time = jiffies;
> mutex_init(&ldata->atomic_read_lock);
> mutex_init(&ldata->output_lock);
> + mutex_init(&ldata->read_buf_lock);
>
> tty->disc_data = ldata;
> reset_buffer_flags(tty->disc_data);
> @@ -1939,6 +1942,7 @@ static int copy_from_read_buf(struct tty_struct *tty,
> size_t tail = ldata->read_tail & (N_TTY_BUF_SIZE - 1);
>
> retval = 0;
> + mutex_lock(&ldata->read_buf_lock);
> n = min(read_cnt(ldata), N_TTY_BUF_SIZE - tail);
> n = min(*nr, n);
> if (n) {
> @@ -1954,6 +1958,7 @@ static int copy_from_read_buf(struct tty_struct *tty,
> *b += n;
> *nr -= n;
> }
> + mutex_unlock(&ldata->read_buf_lock);
> return retval;
> }
>
> @@ -1992,6 +1997,7 @@ static int canon_copy_from_read_buf(struct
> tty_struct *tty,
> bool eof_push = 0;
>
> /* N.B. avoid overrun if nr == 0 */
> + mutex_lock(&ldata->read_buf_lock);
> n = min(*nr, read_cnt(ldata));
> if (!n)
> return 0;
> @@ -2052,6 +2058,7 @@ static int canon_copy_from_read_buf(struct
> tty_struct *tty,
> ldata->push = 0;
> tty_audit_push(tty);
> }
> + mutex_unlock(&ldata->read_buf_lock);
> return eof_push ? -EAGAIN : 0;
> }
>
--
js
suse labs
On 12/10/2014 01:50 PM, Jiri Slaby wrote:
> On 12/10/2014, 07:38 PM, Denis Du wrote:
>>
>> Hi, Guys:
>
> Hi, are you sending this using some robot? I think I have seen like ten
> copies of this patch already.
>
>> It was found that the 3.12 kernel tty layer will lose or corrupt data
>> when have a full-duplex communication, especially in high baud rate, for
>> example 230k for my OMAP5 uart. Eventually I found there is lock missing
>> between copy data to ldisc layer buffer and copy data from the same
>> buffer to user space. I believe this issue existed since 3.8
>> kernel(since this kernel , it start to remove most of the spin-locks)
>> and I didn't find any fix even through 3.17 kernel. This patch was
>> tested to be works great with no any data loss again on 3.12 kernel.
>>
>> This patch was built for the latest kernel, but I cannot test it.
>> Somebody may give a test.
>>
>> I did try to use the existed lock atomic_read_lock, but it doesn’t work.
>
> Anyway, adding Peter Hurley to CC who was working on eliminating locks
> from this code lately. More precisely since 3.12 we have no locks there,
> which would explain why are you seeing it starting 3.12.
Yeah, sorry. A new version of gcc for ARM exposed the lack of barriers for
the head pointer in raw mode. I got tied up on something else, but this
is my #1 priority now.
I hope to have something testable in the next day or two.
Sorry for the delay.
Regards,
Peter Hurley