There is currently (and has always been it seems) a problem
with flush_to_ldisc().
The below diff is against 2.4.20 and contains some stuff made in 2.5,
as well, but the important thing is to set TTY_DONT_FLIP
while processing the flip buffer.
Otherwise, on slow systems or systems with high load and
fast flipping (high baudrates) a new flip might start before
the next flip has been processed, resulting in that characters
could be lost or perhaps arrive out of order.
A driver can write data to a flip buffer that really isn't processed yet.
At least that is what I think happens...
Does it make sense?
/Johan
--- linux/drivers/char/tty_io.c 2 Dec 2002 08:13:09 -0000 1.16
+++ linux/drivers/char/tty_io.c 30 Jul 2003 10:06:04 -0000
@@ -1910,24 +1910,23 @@ static void flush_to_ldisc(void *private
int count;
unsigned long flags;
- if (test_bit(TTY_DONT_FLIP, &tty->flags)) {
+ /* Have TTY_DONT_FLIP set while processing the flip buffer */
+ if (test_and_set_bit(TTY_DONT_FLIP, &tty->flags)) {
+ /* Already set, process later. */
queue_task(&tty->flip.tqueue, &tq_timer);
return;
}
+ save_flags(flags); cli();
if (tty->flip.buf_num) {
cp = tty->flip.char_buf + TTY_FLIPBUF_SIZE;
fp = tty->flip.flag_buf + TTY_FLIPBUF_SIZE;
- tty->flip.buf_num = 0;
-
- save_flags(flags); cli();
+ tty->flip.buf_num = 0;
tty->flip.char_buf_ptr = tty->flip.char_buf;
tty->flip.flag_buf_ptr = tty->flip.flag_buf;
} else {
cp = tty->flip.char_buf;
fp = tty->flip.flag_buf;
- tty->flip.buf_num = 1;
-
- save_flags(flags); cli();
+ tty->flip.buf_num = 1;
tty->flip.char_buf_ptr = tty->flip.char_buf +
TTY_FLIPBUF_SIZE;
tty->flip.flag_buf_ptr = tty->flip.flag_buf +
TTY_FLIPBUF_SIZE;
}
@@ -1936,6 +1935,7 @@ static void flush_to_ldisc(void *private
restore_flags(flags);
tty->ldisc.receive_buf(tty, cp, fp, count);
+ clear_bit(TTY_DONT_FLIP, &tty->flags);
}
Could the silence on this matter be interpreted as
1. No one knows how the tty flipping works (including me I guess)?
2. No one cares about slow machines with high interrupt load?
3. Bad timing, try again...
/Johan
----- Original Message -----
From: <[email protected]>
To: <[email protected]>
Sent: Wednesday, July 30, 2003 12:16 PM
Subject: Safer flush_to_ldisc()
> There is currently (and has always been it seems) a problem
> with flush_to_ldisc().
> The below diff is against 2.4.20 and contains some stuff made in 2.5,
> as well, but the important thing is to set TTY_DONT_FLIP
> while processing the flip buffer.
> Otherwise, on slow systems or systems with high load and
> fast flipping (high baudrates) a new flip might start before
> the next flip has been processed, resulting in that characters
> could be lost or perhaps arrive out of order.
> A driver can write data to a flip buffer that really isn't processed yet.
>
> At least that is what I think happens...
> Does it make sense?
>
> /Johan
>
>
>
> --- linux/drivers/char/tty_io.c 2 Dec 2002 08:13:09 -0000 1.16
> +++ linux/drivers/char/tty_io.c 30 Jul 2003 10:06:04 -0000
> @@ -1910,24 +1910,23 @@ static void flush_to_ldisc(void *private
> int count;
> unsigned long flags;
>
> - if (test_bit(TTY_DONT_FLIP, &tty->flags)) {
> + /* Have TTY_DONT_FLIP set while processing the flip buffer */
> + if (test_and_set_bit(TTY_DONT_FLIP, &tty->flags)) {
> + /* Already set, process later. */
> queue_task(&tty->flip.tqueue, &tq_timer);
> return;
> }
> + save_flags(flags); cli();
> if (tty->flip.buf_num) {
> cp = tty->flip.char_buf + TTY_FLIPBUF_SIZE;
> fp = tty->flip.flag_buf + TTY_FLIPBUF_SIZE;
> - tty->flip.buf_num = 0;
> -
> - save_flags(flags); cli();
> + tty->flip.buf_num = 0;
> tty->flip.char_buf_ptr = tty->flip.char_buf;
> tty->flip.flag_buf_ptr = tty->flip.flag_buf;
> } else {
> cp = tty->flip.char_buf;
> fp = tty->flip.flag_buf;
> - tty->flip.buf_num = 1;
> -
> - save_flags(flags); cli();
> + tty->flip.buf_num = 1;
> tty->flip.char_buf_ptr = tty->flip.char_buf +
> TTY_FLIPBUF_SIZE;
> tty->flip.flag_buf_ptr = tty->flip.flag_buf +
> TTY_FLIPBUF_SIZE;
> }
> @@ -1936,6 +1935,7 @@ static void flush_to_ldisc(void *private
> restore_flags(flags);
>
> tty->ldisc.receive_buf(tty, cp, fp, count);
> + clear_bit(TTY_DONT_FLIP, &tty->flags);
> }
>
> -
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to [email protected]
> More majordomo info at http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at http://www.tux.org/lkml/
>