Re: [PATCH net-next v2] route: Add a new fib_multipath_hash_policy base on cpu id for tunnel packet

[wenxu <wenxu@ucloud.cn>]

On 2019/2/22 下午6:26, Nikolay Aleksandrov wrote:
> On 22/02/2019 11:20, wenxu@ucloud.cn wrote:
>> From: wenxu <wenxu@ucloud.cn>
>>
>> Current fib_multipath_hash_policy can make hash based on the L3 or
>> L4. But it only work on the outer IP. So a specific tunnel always
>> has the same hash value. But a specific tunnel may contain so many
>> inner connections. However there is no good way for tunnel packet.
>> A specific tunnel route based on the percpu dst_cache. It will not
>> lookup route table for each packet.
>>
>> This patch provide a based cpu id hash policy. The different
>> connection run on different cpu and there will be different hash
>> value for percpu dst_cache.
>>
>> Signed-off-by: wenxu <wenxu@ucloud.cn>
>> ---
>>  net/ipv4/route.c           | 6 ++++++
>>  net/ipv4/sysctl_net_ipv4.c | 2 +-
>>  2 files changed, 7 insertions(+), 1 deletion(-)
>>
> Hi,
> When we had the same issue in the bonding, we added a new mode which used
> the flow dissector to get the inner headers and hash on them.
> I believe that is even easier nowadays, but I think people wanted to
> use different hash algorithms as well and last we discussed this we
> were talking about yet another bpf use to generate the hash.
> Now we got bpf_set_hash() which can be used to achieve this from various
> places, if you use that with hash_policy=1 (L4) then skb->hash will
> be used and you can achieve any hashing that you desire.
>
> Cheers,
>  Nik
>
>
Yes, we can set the skb->hash. But ip tunnel lookup the route table without skb.

The most important for performance issue most tunnel send work with dst_cache.

The dst_cache is percpu cache. So the number of dst_entry of a specific tunnel is the cpu cores .

So It's more better hash based on outer and cpu id.