When modifying netclassid, the command("echo 0x100001 > net_cls.classid") will
take more time on muti threads of one process, because the process create many
fds.
for example, one process exists 28000 fds and 60000 threads, echo command will
task 45 seconds.
Now, we only consider the main process when exec "iterate_fd", and the time is
52 milliseconds.

Signed-off-by: zhailiansen <zhailiansen@kuaishou.com>
---
 net/core/netclassid_cgroup.c | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/net/core/netclassid_cgroup.c b/net/core/netclassid_cgroup.c
index d6a70ae..78c903c 100644
--- a/net/core/netclassid_cgroup.c
+++ b/net/core/netclassid_cgroup.c
@@ -88,6 +88,12 @@ static void update_classid_task(struct task_struct *p, u32 classid)
 	};
 	unsigned int fd = 0;
 
+	/* Only update the leader task, when multi threads in this task,
+	 * so it can avoid the useless traversal.
+	 */
+	if (p && p != p->group_leader)
+		return;
+
 	do {
 		task_lock(p);
 		fd = iterate_fd(p->files, fd, update_classid_sock, &ctx);
-- 
1.8.3.1

On Sun, Oct 08, 2023 at 11:04:42AM +0800, zhailiansen wrote:

Hi,

thanks for your patch.
Some minor comments from my side.

> When modifying netclassid, the command("echo 0x100001 > net_cls.classid") will
> take more time on muti threads of one process, because the process create many

nit: muti -> multiple ?

> fds.
> for example, one process exists 28000 fds and 60000 threads, echo command will
> task 45 seconds.
> Now, we only consider the main process when exec "iterate_fd", and the time is
> 52 milliseconds.

Please consider line-wrapping the patch description at 75 bytes.

This patch seems best targeted at net-next, this information should be
included in the patch. So if post a v2 please consider using:

	Subject: [PATCH v2] ...

> Signed-off-by: zhailiansen <zhailiansen@kuaishou.com>

Please consider signing off using your real name.

e.g. Signed-off-by: Firstname Lastname <...>

> ---
>  net/core/netclassid_cgroup.c | 6 ++++++
>  1 file changed, 6 insertions(+)
> 
> diff --git a/net/core/netclassid_cgroup.c b/net/core/netclassid_cgroup.c
> index d6a70ae..78c903c 100644
> --- a/net/core/netclassid_cgroup.c
> +++ b/net/core/netclassid_cgroup.c
> @@ -88,6 +88,12 @@ static void update_classid_task(struct task_struct *p, u32 classid)
>  	};
>  	unsigned int fd = 0;
>  
> +	/* Only update the leader task, when multi threads in this task,
> +	 * so it can avoid the useless traversal.
> +	 */
> +	if (p && p != p->group_leader)
> +		return;

Can p ever be NULL?
It is dereferenced unconditionally by the existing call to task_lock() below.

> +
>  	do {
>  		task_lock(p);
>  		fd = iterate_fd(p->files, fd, update_classid_sock, &ctx);
> -- 
> 1.8.3.1
> 
>