[PATCH] IPIP tunnel performance improvement

[PATCH] IPIP tunnel performance improvement

[zhao ya]

From: Zhao Ya <marywangran0627@gmail.com>
Date: Sat, 27 Feb 2016 10:06:44 +0800
Subject: [PATCH] IPIP tunnel performance improvement

bypass the logic of each packet's own neighbour creation when using 
pointopint or loopback device.

Recently, in our tests, met a performance problem.
In a large number of packets with different target IP address through 
ipip tunnel, PPS will decrease sharply.

The output of perf top are as follows, __write_lock_failed is of the first:
  - 5.89% [kernel]		[k] __write_lock_failed
   -__write_lock_failed 					a
   -_raw_write_lock_bh						a
   -__neigh_create						a
   -ip_finish_output						a
   -ip_output							a
   -ip_local_out						a

The neighbour subsystem will create a neighbour object for each target 
when using pointopint device. When massive amounts of packets with diff-
erent target IP address to be xmit through a pointopint device, these 
packets will suffer the bottleneck at write_lock_bh(&tbl->lock) after 
creating the neighbour object and then inserting it into a hash-table 
at the same time. 

This patch correct it. Only one or little amounts of neighbour objects 
will be created when massive amounts of packets with different target IP 
address through ipip tunnel. 

As the result, performance will be improved.


Signed-off-by: Zhao Ya <marywangran0627@gmail.com>
Signed-off-by: Zhaoya <gaiuszhao@tencent.com>
---
 net/ipv4/ip_output.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/net/ipv4/ip_output.c b/net/ipv4/ip_output.c
index 64878ef..d7c0594 100644
--- a/net/ipv4/ip_output.c
+++ b/net/ipv4/ip_output.c
@@ -202,6 +202,8 @@ static int ip_finish_output2(struct net *net, struct sock *sk, struct sk_buff *s
 
 	rcu_read_lock_bh();
 	nexthop = (__force u32) rt_nexthop(rt, ip_hdr(skb)->daddr);
+	if (dev->flags & (IFF_LOOPBACK | IFF_POINTOPOINT))
+		nexthop = 0;
 	neigh = __ipv4_neigh_lookup_noref(dev, nexthop);
 	if (unlikely(!neigh))
 		neigh = __neigh_create(&arp_tbl, &nexthop, dev, false);

Re: [PATCH] IPIP tunnel performance improvement

[zhao ya]

BTW,before the version 3.5 kernel, the source code contains the logic.
2.6.32, for example, in arp_bind_neighbour function, there are the following logic:

__be32 nexthop = ((struct rtable *) DST) - > rt_gateway;
if (dev - > flags & (IFF_LOOPBACK | IFF_POINTOPOINT))
nexthop = 0;
n = __neigh_lookup_errno (
...

zhao ya said, at 2/27/2016 12:40 PM:
> From: Zhao Ya <marywangran0627@gmail.com>
> Date: Sat, 27 Feb 2016 10:06:44 +0800
> Subject: [PATCH] IPIP tunnel performance improvement
> 
> bypass the logic of each packet's own neighbour creation when using 
> pointopint or loopback device.
> 
> Recently, in our tests, met a performance problem.
> In a large number of packets with different target IP address through 
> ipip tunnel, PPS will decrease sharply.
> 
> The output of perf top are as follows, __write_lock_failed is of the first:
>   - 5.89% [kernel]		[k] __write_lock_failed
>    -__write_lock_failed 					a
>    -_raw_write_lock_bh						a
>    -__neigh_create						a
>    -ip_finish_output						a
>    -ip_output							a
>    -ip_local_out						a
> 
> The neighbour subsystem will create a neighbour object for each target 
> when using pointopint device. When massive amounts of packets with diff-
> erent target IP address to be xmit through a pointopint device, these 
> packets will suffer the bottleneck at write_lock_bh(&tbl->lock) after 
> creating the neighbour object and then inserting it into a hash-table 
> at the same time. 
> 
> This patch correct it. Only one or little amounts of neighbour objects 
> will be created when massive amounts of packets with different target IP 
> address through ipip tunnel. 
> 
> As the result, performance will be improved.
> 
> 
> Signed-off-by: Zhao Ya <marywangran0627@gmail.com>
> Signed-off-by: Zhaoya <gaiuszhao@tencent.com>
> ---
>  net/ipv4/ip_output.c | 2 ++
>  1 file changed, 2 insertions(+)
> 
> diff --git a/net/ipv4/ip_output.c b/net/ipv4/ip_output.c
> index 64878ef..d7c0594 100644
> --- a/net/ipv4/ip_output.c
> +++ b/net/ipv4/ip_output.c
> @@ -202,6 +202,8 @@ static int ip_finish_output2(struct net *net, struct sock *sk, struct sk_buff *s
>  
>  	rcu_read_lock_bh();
>  	nexthop = (__force u32) rt_nexthop(rt, ip_hdr(skb)->daddr);
> +	if (dev->flags & (IFF_LOOPBACK | IFF_POINTOPOINT))
> +		nexthop = 0;
>  	neigh = __ipv4_neigh_lookup_noref(dev, nexthop);
>  	if (unlikely(!neigh))
>  		neigh = __neigh_create(&arp_tbl, &nexthop, dev, false);
> 
> 

Re: [PATCH] IPIP tunnel performance improvement

[Cong Wang]

On Fri, Feb 26, 2016 at 8:40 PM, zhao ya <marywangran0627@gmail.com> wrote:
> From: Zhao Ya <marywangran0627@gmail.com>
> Date: Sat, 27 Feb 2016 10:06:44 +0800
> Subject: [PATCH] IPIP tunnel performance improvement
>
> bypass the logic of each packet's own neighbour creation when using
> pointopint or loopback device.
>
> Recently, in our tests, met a performance problem.
> In a large number of packets with different target IP address through
> ipip tunnel, PPS will decrease sharply.
>
> The output of perf top are as follows, __write_lock_failed is of the first:
>   - 5.89% [kernel]              [k] __write_lock_failed
>    -__write_lock_failed                                         a
>    -_raw_write_lock_bh                                          a
>    -__neigh_create                                              a
>    -ip_finish_output                                            a
>    -ip_output                                                   a
>    -ip_local_out                                                a
>
> The neighbour subsystem will create a neighbour object for each target
> when using pointopint device. When massive amounts of packets with diff-
> erent target IP address to be xmit through a pointopint device, these
> packets will suffer the bottleneck at write_lock_bh(&tbl->lock) after
> creating the neighbour object and then inserting it into a hash-table
> at the same time.
>
> This patch correct it. Only one or little amounts of neighbour objects
> will be created when massive amounts of packets with different target IP
> address through ipip tunnel.
>
> As the result, performance will be improved.

Well, you just basically revert another bug fix:

commit 0bb4087cbec0ef74fd416789d6aad67957063057
Author: David S. Miller <davem@davemloft.net>
Date:   Fri Jul 20 16:00:53 2012 -0700

    ipv4: Fix neigh lookup keying over loopback/point-to-point devices.

    We were using a special key "0" for all loopback and point-to-point
    device neigh lookups under ipv4, but we wouldn't use that special
    key for the neigh creation.

    So basically we'd make a new neigh at each and every lookup :-)

    This special case to use only one neigh for these device types
    is of dubious value, so just remove it entirely.

    Reported-by: Eric Dumazet <eric.dumazet@gmail.com>
    Signed-off-by: David S. Miller <davem@davemloft.net>

which would bring the neigh entries counting problem back...

Did you try to tune the neigh gc parameters for your case?

Thanks.

Re: [PATCH] IPIP tunnel performance improvement

[zhao ya]

Yes, I did, but have no effect.

I want to ask is, why David's patch not used.

Thanks.



Cong Wang said, at 2/27/2016 2:29 PM:
> On Fri, Feb 26, 2016 at 8:40 PM, zhao ya <marywangran0627@gmail.com> wrote:
>> From: Zhao Ya <marywangran0627@gmail.com>
>> Date: Sat, 27 Feb 2016 10:06:44 +0800
>> Subject: [PATCH] IPIP tunnel performance improvement
>>
>> bypass the logic of each packet's own neighbour creation when using
>> pointopint or loopback device.
>>
>> Recently, in our tests, met a performance problem.
>> In a large number of packets with different target IP address through
>> ipip tunnel, PPS will decrease sharply.
>>
>> The output of perf top are as follows, __write_lock_failed is of the first:
>>   - 5.89% [kernel]              [k] __write_lock_failed
>>    -__write_lock_failed                                         a
>>    -_raw_write_lock_bh                                          a
>>    -__neigh_create                                              a
>>    -ip_finish_output                                            a
>>    -ip_output                                                   a
>>    -ip_local_out                                                a
>>
>> The neighbour subsystem will create a neighbour object for each target
>> when using pointopint device. When massive amounts of packets with diff-
>> erent target IP address to be xmit through a pointopint device, these
>> packets will suffer the bottleneck at write_lock_bh(&tbl->lock) after
>> creating the neighbour object and then inserting it into a hash-table
>> at the same time.
>>
>> This patch correct it. Only one or little amounts of neighbour objects
>> will be created when massive amounts of packets with different target IP
>> address through ipip tunnel.
>>
>> As the result, performance will be improved.
> 
> Well, you just basically revert another bug fix:
> 
> commit 0bb4087cbec0ef74fd416789d6aad67957063057
> Author: David S. Miller <davem@davemloft.net>
> Date:   Fri Jul 20 16:00:53 2012 -0700
> 
>     ipv4: Fix neigh lookup keying over loopback/point-to-point devices.
> 
>     We were using a special key "0" for all loopback and point-to-point
>     device neigh lookups under ipv4, but we wouldn't use that special
>     key for the neigh creation.
> 
>     So basically we'd make a new neigh at each and every lookup :-)
> 
>     This special case to use only one neigh for these device types
>     is of dubious value, so just remove it entirely.
> 
>     Reported-by: Eric Dumazet <eric.dumazet@gmail.com>
>     Signed-off-by: David S. Miller <davem@davemloft.net>
> 
> which would bring the neigh entries counting problem back...
> 
> Did you try to tune the neigh gc parameters for your case?
> 
> Thanks.
>