[PATCH net-next] tcp: add a scheduling point in established_get_first()

[PATCH net-next] tcp: add a scheduling point in established_get_first()

[Jian Wen]

Kubernetes[1] is going to stick with /proc/net/tcp for a while.

This commit reduces the scheduling latency introduced by established_get_first(),
similar to commit acffb584cda7 ("net: diag: add a scheduling point in inet_diag_dump_icsk()").

In our environment, the scheduling latency affects:
1. the performance of latency-sensitive services like Redis
2. the delay of synchronize_net() that is called with RTNL is locked
   12 times when Dockerd is deleting a container

[1] https://github.com/google/cadvisor/blob/v0.47.2/container/libcontainer/handler.go#L130

Signed-off-by: Jian Wen <wenjian1@xiaomi.com>
---
 net/ipv4/tcp_ipv4.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/net/ipv4/tcp_ipv4.c b/net/ipv4/tcp_ipv4.c
index fd365de4d5ff..3271848e9c9a 100644
--- a/net/ipv4/tcp_ipv4.c
+++ b/net/ipv4/tcp_ipv4.c
@@ -57,6 +57,7 @@
 #include <linux/init.h>
 #include <linux/times.h>
 #include <linux/slab.h>
+#include <linux/sched.h>
 
 #include <net/net_namespace.h>
 #include <net/icmp.h>
@@ -2456,6 +2457,7 @@ static void *established_get_first(struct seq_file *seq)
 				return sk;
 		}
 		spin_unlock_bh(lock);
+		cond_resched();
 	}
 
 	return NULL;
-- 
2.25.1

Re: [PATCH net-next] tcp: add a scheduling point in established_get_first()

[Eric Dumazet]

On Fri, Jun 30, 2023 at 9:18 AM Jian Wen <wenjianhn@gmail.com> wrote:
>
> Kubernetes[1] is going to stick with /proc/net/tcp for a while.
>
> This commit reduces the scheduling latency introduced by established_get_first(),
> similar to commit acffb584cda7 ("net: diag: add a scheduling point in inet_diag_dump_icsk()").
>
> In our environment, the scheduling latency affects:
> 1. the performance of latency-sensitive services like Redis
> 2. the delay of synchronize_net() that is called with RTNL is locked
>    12 times when Dockerd is deleting a container
>
> [1] https://github.com/google/cadvisor/blob/v0.47.2/container/libcontainer/handler.go#L130
>
> Signed-off-by: Jian Wen <wenjian1@xiaomi.com>
> ---
>  net/ipv4/tcp_ipv4.c | 2 ++
>  1 file changed, 2 insertions(+)
>
> diff --git a/net/ipv4/tcp_ipv4.c b/net/ipv4/tcp_ipv4.c
> index fd365de4d5ff..3271848e9c9a 100644
> --- a/net/ipv4/tcp_ipv4.c
> +++ b/net/ipv4/tcp_ipv4.c
> @@ -57,6 +57,7 @@
>  #include <linux/init.h>
>  #include <linux/times.h>
>  #include <linux/slab.h>
> +#include <linux/sched.h>
>
>  #include <net/net_namespace.h>
>  #include <net/icmp.h>
> @@ -2456,6 +2457,7 @@ static void *established_get_first(struct seq_file *seq)
>                                 return sk;
>                 }
>                 spin_unlock_bh(lock);
> +               cond_resched();
>         }
>
>         return NULL;
> --
> 2.25.1
>
Hi Jian, thanks for your patch.

Few points:

- Note that net-next is currently closed (merge window)

- Also, /proc interface does not hold RTNL, not sure why you mention
RTNL in the changelog,
and not other mutexes in the kernel that also would be impacted by the
long duration of established_get_first() ?

- The cond_resched() should be done even if all buckets are empty ?

- Using inet_diag, Kubernetes could list both IPv4/IPv6 sockets in one dump,
and benefit from more modern interface (with cond_resched() already there)

[PATCH v2 net-next] tcp: add a scheduling point in established_get_first()

[Jian Wen]

Kubernetes[1] is going to stick with /proc/net/tcp for a while.

This commit reduces the scheduling latency introduced by
established_get_first(), similar to commit acffb584cda7 ("net: diag:
add a scheduling point in inet_diag_dump_icsk()").

In our environment, the scheduling latency affects the performance of
latency-sensitive services like Redis.

Changes in V2 :
 - call cond_resched() before checking if a bucket is empty as
   suggested by Eric Dumazet
 - removed the delay of synchronize_net() from the commit message

[1] https://github.com/google/cadvisor/blob/v0.47.2/container/libcontainer/handler.go#L130

Signed-off-by: Jian Wen <wenjian1@xiaomi.com>
---
 net/ipv4/tcp_ipv4.c | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/net/ipv4/tcp_ipv4.c b/net/ipv4/tcp_ipv4.c
index fd365de4d5ff..cecd5a135e64 100644
--- a/net/ipv4/tcp_ipv4.c
+++ b/net/ipv4/tcp_ipv4.c
@@ -57,6 +57,7 @@
 #include <linux/init.h>
 #include <linux/times.h>
 #include <linux/slab.h>
+#include <linux/sched.h>
 
 #include <net/net_namespace.h>
 #include <net/icmp.h>
@@ -2446,6 +2447,8 @@ static void *established_get_first(struct seq_file *seq)
 		struct hlist_nulls_node *node;
 		spinlock_t *lock = inet_ehash_lockp(hinfo, st->bucket);
 
+		cond_resched();
+
 		/* Lockless fast path for the common case of empty buckets */
 		if (empty_bucket(hinfo, st))
 			continue;
-- 
2.25.1

Re: [PATCH v2 net-next] tcp: add a scheduling point in established_get_first()

[Simon Horman]

On Tue, Jul 11, 2023 at 11:24:05AM +0800, Jian Wen wrote:
> Kubernetes[1] is going to stick with /proc/net/tcp for a while.
> 
> This commit reduces the scheduling latency introduced by
> established_get_first(), similar to commit acffb584cda7 ("net: diag:
> add a scheduling point in inet_diag_dump_icsk()").
> 
> In our environment, the scheduling latency affects the performance of
> latency-sensitive services like Redis.
> 
> Changes in V2 :
>  - call cond_resched() before checking if a bucket is empty as
>    suggested by Eric Dumazet
>  - removed the delay of synchronize_net() from the commit message
> 
> [1] https://github.com/google/cadvisor/blob/v0.47.2/container/libcontainer/handler.go#L130
> 
> Signed-off-by: Jian Wen <wenjian1@xiaomi.com>

Reviewed-by: Simon Horman <simon.horman@corigine.com>

Re: [PATCH v2 net-next] tcp: add a scheduling point in established_get_first()

[patchwork-bot+netdevbpf]

Hello:

This patch was applied to netdev/net-next.git (main)
by Jakub Kicinski <kuba@kernel.org>:

On Tue, 11 Jul 2023 11:24:05 +0800 you wrote:
> Kubernetes[1] is going to stick with /proc/net/tcp for a while.
> 
> This commit reduces the scheduling latency introduced by
> established_get_first(), similar to commit acffb584cda7 ("net: diag:
> add a scheduling point in inet_diag_dump_icsk()").
> 
> In our environment, the scheduling latency affects the performance of
> latency-sensitive services like Redis.
> 
> [...]

Here is the summary with links:
  - [v2,net-next] tcp: add a scheduling point in established_get_first()
    https://git.kernel.org/netdev/net-next/c/9f4a7c930284

You are awesome, thank you!
-- 
Deet-doot-dot, I am a bot.
https://korg.docs.kernel.org/patchwork/pwbot.html