[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Message-ID: <20160421132743.GL3408@twins.programming.kicks-ass.net>
Date: Thu, 21 Apr 2016 15:27:43 +0200
From: Peter Zijlstra <peterz@...radead.org>
To: Minfei Huang <mnghuan@...il.com>
Cc: mingo@...hat.com, linux-kernel@...r.kernel.org
Subject: Re: [PATCH] Remove useless test during exchanging values
On Thu, Apr 21, 2016 at 08:28:28PM +0800, Minfei Huang wrote:
> The value delta is correct as well, although calc_load_idle[idx] is
> equal to 0. Remove this useless test to improve performance, since this
> function is called more frequently.
>
> Signed-off-by: Minfei Huang <mnghuan@...il.com>
> ---
> kernel/sched/loadavg.c | 5 ++---
> 1 file changed, 2 insertions(+), 3 deletions(-)
>
> diff --git a/kernel/sched/loadavg.c b/kernel/sched/loadavg.c
> index ef71590..5a5d7ae 100644
> --- a/kernel/sched/loadavg.c
> +++ b/kernel/sched/loadavg.c
> @@ -216,10 +216,9 @@ void calc_load_exit_idle(void)
> static long calc_load_fold_idle(void)
> {
> int idx = calc_load_read_idx();
> - long delta = 0;
> + long delta;
>
> - if (atomic_long_read(&calc_load_idle[idx]))
> - delta = atomic_long_xchg(&calc_load_idle[idx], 0);
> + delta = atomic_long_xchg(&calc_load_idle[idx], 0);
This will actually degrade performance.
The read can have the cacheline in shared mode, while the xchg will
force it into exclusive mode. Also, the xchg atomic operation is a
really expensive instruction.
Powered by blists - more mailing lists