On 11/23/2013 07:23 PM, One Thousand Gnomes wrote:

On Fri, 22 Nov 2013 10:59:24 -0500
Peter Hurley [off-list ref] wrote:

quoted

Only wakeup the _waiting_ reader, polls and/or writer(s).

Signed-off-by: Peter Hurley <redacted>
---
  drivers/tty/n_tty.c | 18 ++++++++++++------
  1 file changed, 12 insertions(+), 6 deletions(-)

diff --git a/drivers/tty/n_tty.c b/drivers/tty/n_tty.c
index 8f2356e..aae28a6 100644
--- a/drivers/tty/n_tty.c
+++ b/drivers/tty/n_tty.c

@@ -275,7 +275,8 @@ static void n_tty_check_unthrottle(struct tty_struct *tty)
  			return;
  		n_tty_set_room(tty);
  		n_tty_write_wakeup(tty->link);
-		wake_up_interruptible_poll(&tty->link->write_wait, POLLOUT);
+		if (waitqueue_active(&tty->link->write_wait))
+			wake_up_interruptible_poll(&tty->link->write_wait, POLLOUT);

Does this actually microbenchmark faster ?

Getting on and off the write_wait queue is actually pretty expensive for
the "other" pty (the writer), and the unnecessary wakeup from the reader
doesn't help.

The other chunks are gratuitous.

Regards,
Peter Hurley

PS - This came up because there is some worst-case behavior that I'm looking
into fixing.  When the userspace reader is very far behind (say because it's
reading char-by-char), it doesn't make sense to keep restarting the input
processing worker.