2010-12-16 20:32:21

by Davide Libenzi

Subject: [patch] make pipes use event aware wakeups

I thought this was alreay merged, but this thread brought to my attention
it wasn't:

https://lkml.org/lkml/2010/12/15/234

This patch sends the events the wakeup refers to, so that epoll, and even
the new poll code in fs/select.c can avoid wakeups in events are not
matching the requested set.


Signed-off-by: Davide Libenzi <[email protected]>


- Davide


---
fs/pipe.c | 10 +++++-----
1 file changed, 5 insertions(+), 5 deletions(-)

Index: linux-2.6.mod/fs/pipe.c
===================================================================
--- linux-2.6.mod.orig/fs/pipe.c 2010-12-15 20:49:48.000000000 -0800
+++ linux-2.6.mod/fs/pipe.c 2010-12-15 20:51:42.000000000 -0800
@@ -441,7 +441,7 @@
break;
}
if (do_wakeup) {
- wake_up_interruptible_sync(&pipe->wait);
+ wake_up_interruptible_sync_poll(&pipe->wait, POLLOUT);
kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
}
pipe_wait(pipe);
@@ -450,7 +450,7 @@

/* Signal writers asynchronously that there is more room. */
if (do_wakeup) {
- wake_up_interruptible_sync(&pipe->wait);
+ wake_up_interruptible_sync_poll(&pipe->wait, POLLOUT);
kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
}
if (ret > 0)
@@ -612,7 +612,7 @@
break;
}
if (do_wakeup) {
- wake_up_interruptible_sync(&pipe->wait);
+ wake_up_interruptible_sync_poll(&pipe->wait, POLLIN);
kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
do_wakeup = 0;
}
@@ -623,7 +623,7 @@
out:
mutex_unlock(&inode->i_mutex);
if (do_wakeup) {
- wake_up_interruptible_sync(&pipe->wait);
+ wake_up_interruptible_sync_poll(&pipe->wait, POLLIN);
kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
}
if (ret > 0)
@@ -715,7 +715,7 @@
if (!pipe->readers && !pipe->writers) {
free_pipe_info(inode);
} else {
- wake_up_interruptible_sync(&pipe->wait);
+ wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLOUT);
kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
}

2010-12-16 20:35:30

by David Miller

Subject: Re: [patch] make pipes use event aware wakeups

From: Davide Libenzi <[email protected]>
Date: Thu, 16 Dec 2010 12:32:42 -0800 (PST)

> I thought this was alreay merged, but this thread brought to my attention
> it wasn't:
>
> https://lkml.org/lkml/2010/12/15/234
>
> This patch sends the events the wakeup refers to, so that epoll, and even
> the new poll code in fs/select.c can avoid wakeups in events are not
> matching the requested set.
>
>
> Signed-off-by: Davide Libenzi <[email protected]>

Acked-by: David S. Miller <[email protected]>

2010-12-16 20:50:20

by Eric Dumazet

Subject: Re: [patch] make pipes use event aware wakeups

Le jeudi 16 décembre 2010 à 12:32 -0800, Davide Libenzi a écrit :
> I thought this was alreay merged, but this thread brought to my attention
> it wasn't:
>
> https://lkml.org/lkml/2010/12/15/234
>
> This patch sends the events the wakeup refers to, so that epoll, and even
> the new poll code in fs/select.c can avoid wakeups in events are not
> matching the requested set.
>
>
> Signed-off-by: Davide Libenzi <[email protected]>
>
>
> - Davide
>
>
> ---
> fs/pipe.c | 10 +++++-----
> 1 file changed, 5 insertions(+), 5 deletions(-)
>
> Index: linux-2.6.mod/fs/pipe.c
> ===================================================================
> --- linux-2.6.mod.orig/fs/pipe.c 2010-12-15 20:49:48.000000000 -0800
> +++ linux-2.6.mod/fs/pipe.c 2010-12-15 20:51:42.000000000 -0800
> @@ -441,7 +441,7 @@
> break;
> }
> if (do_wakeup) {
> - wake_up_interruptible_sync(&pipe->wait);
> + wake_up_interruptible_sync_poll(&pipe->wait, POLLOUT);
> kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
> }
> pipe_wait(pipe);
> @@ -450,7 +450,7 @@
>
> /* Signal writers asynchronously that there is more room. */
> if (do_wakeup) {
> - wake_up_interruptible_sync(&pipe->wait);
> + wake_up_interruptible_sync_poll(&pipe->wait, POLLOUT);
> kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
> }
> if (ret > 0)
> @@ -612,7 +612,7 @@
> break;
> }
> if (do_wakeup) {
> - wake_up_interruptible_sync(&pipe->wait);
> + wake_up_interruptible_sync_poll(&pipe->wait, POLLIN);
> kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
> do_wakeup = 0;
> }
> @@ -623,7 +623,7 @@
> out:
> mutex_unlock(&inode->i_mutex);
> if (do_wakeup) {
> - wake_up_interruptible_sync(&pipe->wait);
> + wake_up_interruptible_sync_poll(&pipe->wait, POLLIN);
> kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
> }
> if (ret > 0)
> @@ -715,7 +715,7 @@
> if (!pipe->readers && !pipe->writers) {
> free_pipe_info(inode);
> } else {
> - wake_up_interruptible_sync(&pipe->wait);
> + wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLOUT);
> kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
> kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
> }

Indeed ;)

I am not sure the last bit is needed, but it should not matter.

Acked-by: Eric Dumazet <[email protected]>