[PATCH net-next] gen_estimator: change the lock order for better perfermance

[PATCH net-next] gen_estimator: change the lock order for better perfermance

[Hong Zhiguo]

From: Hong Zhiguo <zhiguohong@tencent.com>

e->stats_lock is usually taken by fast path to update stats.
In the old order, fast path should wait for write_lock(&est_lock).
Even though it's only one line inside the write_lock(&est_lock),
but if there's interrupt or page fault, a lot of spin on
e->stats_lock will be wasted in fast path.

Signed-off-by: Hong Zhiguo <zhiguohong@tencent.com>
---
 net/core/gen_estimator.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/net/core/gen_estimator.c b/net/core/gen_estimator.c
index 6b5b6e7..bad7f5f 100644
--- a/net/core/gen_estimator.c
+++ b/net/core/gen_estimator.c
@@ -120,11 +120,11 @@ static void est_timer(unsigned long arg)
 		u32 npackets;
 		u32 rate;
 
-		spin_lock(e->stats_lock);
 		read_lock(&est_lock);
 		if (e->bstats == NULL)
 			goto skip;
 
+		spin_lock(e->stats_lock);
 		nbytes = e->bstats->bytes;
 		npackets = e->bstats->packets;
 		brate = (nbytes - e->last_bytes)<<(7 - idx);
@@ -136,9 +136,9 @@ static void est_timer(unsigned long arg)
 		e->last_packets = npackets;
 		e->avpps += (rate >> e->ewma_log) - (e->avpps >> e->ewma_log);
 		e->rate_est->pps = (e->avpps+0x1FF)>>10;
+		spin_unlock(e->stats_lock);
 skip:
 		read_unlock(&est_lock);
-		spin_unlock(e->stats_lock);
 	}
 
 	if (!list_empty(&elist[idx].list))
-- 
1.8.1.2

Re: [PATCH net-next] gen_estimator: change the lock order for better perfermance

[Eric Dumazet]

On Tue, 2013-09-17 at 16:38 +0800, Hong Zhiguo wrote:
> From: Hong Zhiguo <zhiguohong@tencent.com>
> 
> e->stats_lock is usually taken by fast path to update stats.
> In the old order, fast path should wait for write_lock(&est_lock).
> Even though it's only one line inside the write_lock(&est_lock),
> but if there's interrupt or page fault, a lot of spin on
> e->stats_lock will be wasted in fast path.

1) net-next is not open

2) There is no way a page fault can happen in this path.

3) This patch is wrong.

Current order is there for good reasons.

Have you really tried LOCKDEP, before sending a patch changing lock
order ?

答复: [PATCH net-next] gen_estimator: change the lock order for better perfermance(Internet mail)

[zhiguohong(洪志国)]

Actually I did try LOCKDEP and didn't get any warning.

Could you pointed out why it's wrong?

-----邮件原件-----
发件人: Eric Dumazet [mailto:eric.dumazet@gmail.com] 
发送时间: 2013年9月17日 19:12
收件人: Hong Zhiguo
抄送: netdev@vger.kernel.org; davem@davemloft.net; stephen@networkplumber.org; zhiguohong(洪志国)
主题: Re: [PATCH net-next] gen_estimator: change the lock order for better perfermance(Internet mail)

On Tue, 2013-09-17 at 16:38 +0800, Hong Zhiguo wrote:
> From: Hong Zhiguo <zhiguohong@tencent.com>
> 
> e->stats_lock is usually taken by fast path to update stats.
> In the old order, fast path should wait for write_lock(&est_lock).
> Even though it's only one line inside the write_lock(&est_lock), but 
> if there's interrupt or page fault, a lot of spin on
> e->stats_lock will be wasted in fast path.

1) net-next is not open

2) There is no way a page fault can happen in this path.

3) This patch is wrong.

Current order is there for good reasons.

Have you really tried LOCKDEP, before sending a patch changing lock order ?