[RFC PATCH 07/10] pipe: Conditionalise wakeup in pipe_read() [ver #2]

[Konstantin Khlebnikov]

On 23/10/2019 23.18, David Howells wrote:
> Only do a wakeup in pipe_read() if we made space in a completely full
> buffer.  The producer shouldn't be waiting on pipe->wait otherwise.

We could go further and wakeup writer only when at least half of buffer is empty.
This gives better batching and reduces rate of context switches.

https://lore.kernel.org/lkml/157219118016.7078.16223055699799396042.stgit@buzz/T/#u

> 
> Signed-off-by: David Howells <dhowells at redhat.com>
> ---
> 
>   fs/pipe.c |   15 ++++++---------
>   1 file changed, 6 insertions(+), 9 deletions(-)
> 
> diff --git a/fs/pipe.c b/fs/pipe.c
> index 1274305772fb..e3a8f10750c9 100644
> --- a/fs/pipe.c
> +++ b/fs/pipe.c
> @@ -327,11 +327,13 @@ pipe_read(struct kiocb *iocb, struct iov_iter *to)
>   				spin_lock_irq(&pipe->wait.lock);
>   				tail++;
>   				pipe_commit_read(pipe, tail);
> -				do_wakeup = 0;
> -				wake_up_interruptible_sync_poll_locked(
> -					&pipe->wait, EPOLLOUT | EPOLLWRNORM);
> +				do_wakeup = 1;
> +				if (head - (tail - 1) == pipe->max_usage)
> +					wake_up_interruptible_sync_poll_locked(
> +						&pipe->wait, EPOLLOUT | EPOLLWRNORM);
>   				spin_unlock_irq(&pipe->wait.lock);
> -				kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
> +				if (head - (tail - 1) == pipe->max_usage)
> +					kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
>   			}
>   			total_len -= chars;
>   			if (!total_len)
> @@ -360,11 +362,6 @@ pipe_read(struct kiocb *iocb, struct iov_iter *to)
>   				ret = -ERESTARTSYS;
>   			break;
>   		}
> -		if (do_wakeup) {
> -			wake_up_interruptible_sync_poll(&pipe->wait, EPOLLOUT | EPOLLWRNORM);
> - 			kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
> -			do_wakeup = 0;
> -		}
>   		pipe_wait(pipe);
>   	}
>   	__pipe_unlock(pipe);
> 
>