2020-01-29 18:19:27

by Davidlohr Bueso

Subject: [PATCH] bcache: optimize barrier usage for Rmw atomic bitops

We can avoid the unnecessary barrier on non LL/SC architectures,
such as x86. Instead, use the smp_mb__after_atomic().

Signed-off-by: Davidlohr Bueso <[email protected]>
---
drivers/md/bcache/writeback.c | 6 +++---
1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/drivers/md/bcache/writeback.c b/drivers/md/bcache/writeback.c
index 4a40f9eadeaf..284ff10f480f 100644
--- a/drivers/md/bcache/writeback.c
+++ b/drivers/md/bcache/writeback.c
@@ -183,7 +183,7 @@ static void update_writeback_rate(struct work_struct *work)
*/
set_bit(BCACHE_DEV_RATE_DW_RUNNING, &dc->disk.flags);
/* paired with where BCACHE_DEV_RATE_DW_RUNNING is tested */
- smp_mb();
+ smp_mb__after_atomic();

/*
* CACHE_SET_IO_DISABLE might be set via sysfs interface,
@@ -193,7 +193,7 @@ static void update_writeback_rate(struct work_struct *work)
test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
clear_bit(BCACHE_DEV_RATE_DW_RUNNING, &dc->disk.flags);
/* paired with where BCACHE_DEV_RATE_DW_RUNNING is tested */
- smp_mb();
+ smp_mb__after_atomic();
return;
}

@@ -229,7 +229,7 @@ static void update_writeback_rate(struct work_struct *work)
*/
clear_bit(BCACHE_DEV_RATE_DW_RUNNING, &dc->disk.flags);
/* paired with where BCACHE_DEV_RATE_DW_RUNNING is tested */
- smp_mb();
+ smp_mb__after_atomic();
}

static unsigned int writeback_delay(struct cached_dev *dc,
--
2.16.4

2020-03-20 15:26:19

by Coly Li

Subject: Re: [PATCH] bcache: optimize barrier usage for Rmw atomic bitops

On 2020/1/30 2:08 上午, Davidlohr Bueso wrote:
> We can avoid the unnecessary barrier on non LL/SC architectures,
> such as x86. Instead, use the smp_mb__after_atomic().
>

Hi Davidlohr,

> Signed-off-by: Davidlohr Bueso <[email protected]>

Sorry for replying late. Indeed we don't need to worry about the extra
smp_mb() because the code patch gets called every 5 seconds by default,
which is very infrequent. But I like this more accurate patch, it is in
my testing already.

And by your inspiration, I also fix some other locations to use
smp_mb__before/after_atomic().

Thank you.

Coly Li

> ---
> drivers/md/bcache/writeback.c | 6 +++---
> 1 file changed, 3 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/md/bcache/writeback.c b/drivers/md/bcache/writeback.c
> index 4a40f9eadeaf..284ff10f480f 100644
> --- a/drivers/md/bcache/writeback.c
> +++ b/drivers/md/bcache/writeback.c
> @@ -183,7 +183,7 @@ static void update_writeback_rate(struct work_struct *work)
> */
> set_bit(BCACHE_DEV_RATE_DW_RUNNING, &dc->disk.flags);
> /* paired with where BCACHE_DEV_RATE_DW_RUNNING is tested */
> - smp_mb();
> + smp_mb__after_atomic();
>
> /*
> * CACHE_SET_IO_DISABLE might be set via sysfs interface,
> @@ -193,7 +193,7 @@ static void update_writeback_rate(struct work_struct *work)
> test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
> clear_bit(BCACHE_DEV_RATE_DW_RUNNING, &dc->disk.flags);
> /* paired with where BCACHE_DEV_RATE_DW_RUNNING is tested */
> - smp_mb();
> + smp_mb__after_atomic();
> return;
> }
>
> @@ -229,7 +229,7 @@ static void update_writeback_rate(struct work_struct *work)
> */
> clear_bit(BCACHE_DEV_RATE_DW_RUNNING, &dc->disk.flags);
> /* paired with where BCACHE_DEV_RATE_DW_RUNNING is tested */
> - smp_mb();
> + smp_mb__after_atomic();
> }
>
> static unsigned int writeback_delay(struct cached_dev *dc,
>


--

Coly Li