Date:   Wed, 19 Aug 2020 05:48:03 -0400
From:   Jamal Hadi Salim <jhs@...atatu.com>
To:     Cong Wang <xiyou.wangcong@...il.com>
Cc:     David Miller <davem@...emloft.net>,
        Linux Kernel Network Developers <netdev@...r.kernel.org>,
        Jiri Pirko <jiri@...nulli.us>,
        Ariel Levkovich <lariel@...lanox.com>
Subject: Re: [PATCH net-next 1/1] net/sched: Introduce skb hash classifier

On 2020-08-17 3:47 p.m., Cong Wang wrote:
> On Mon, Aug 17, 2020 at 4:19 AM Jamal Hadi Salim <jhs@...atatu.com> wrote:
>>

[..]
>> There is no ambiguity of intent in the fw case, there is only one field.
>> In the case of having multiple fields it is ambigious if you
>> unconditionally look.
>>
>> Example: policy says to match skb mark of 5 and hash of 3.
>> If packet arrives with skb->mark is 5 and skb->hash is 3
>> very clearly matched the intent of the policy.
>> If packet arrives withj skb->mark 7 and hash 3 it clearly
>> did not match the intent. etc.
> 
> This example clearly shows no ambiguous, right? ;)
> 

Ambigious only from the perspective of relational AND vs OR
(your original pseudo code had it in OR relation).

> 
>>
>>> But if filters were put in a global hashtable, the above would be
>>> much harder to implement.
>>>
>>
>> Ok, yes. My assumption has been you will have some global shared
>> structure where all filters will be installed on.
> 
> Sure, if not hashtable, we could simply put them in a list:
> 
> list_for_each_filter {
>    if (filter_parameter_has_hash) {
>      match skb->hash with cls->param_hash
>    }
>    if (filter_parameter_has_mark) {
>      match skb->mark with cls->param_mark
>    }
> }
> 

Yes, that would work - but iteration is linear.

> 
>>
>> I think i may have misunderstood all along what you were saying
>> which is:
>>
>> a) add the rules so they are each _independent with different
>>      priorities_ in a chain.
> 
> Yes, because this gives users freedom to pick a different prio
> from its value (hash or mark).
>

ok.

> 
>>
>> b)  when i do lookup for packet arrival, i will only see a filter
>>    that matches "match mark 5 and hash 3" (meaning there is no
>>    ambiguity on intent). If packet data doesnt match policy then
>>    i will iterate to another filter on the chain list with lower
>>    priority.
> 
> Right. Multiple values mean AND, not OR, so if you specify
> mark 5 and hash 3, it will match skb->mark==5 && skb->hash==3.
> If not matched, it will continue the iteration until the end.
>

That would remove the ambiguity (assuming iteration with "continue"
to create the AND effect).

>>
>> Am i correct in my understanding?
>>
>> If i am - then we still have a problem with lookup scale in presence
>> of a large number of filters since essentially this approach
>> is linear lookup (similar problem iptables has). I am afraid
>> a hash table or something with similar principle goals is needed.
> 
> Yeah, this is why I asked you whether we have to put them in a
> hashtable in previous emails, as hashtable organizes them with
> a key, it is hard to combine multiple fields in one key and allow
> to extend easily in the future. But other people smarter than me
> may have better ideas here.

To achieve reasonable performance (with many filters) I dont think
there is escape from having something that is centralized
(per priority) - sort of what fw, u32 or flower do. A hash table is
the most common approach; i was hoping that IDR maybe useful since the
skb->hash maps nicely to "32 bit key" but Vlad was saying at the
tc workshop that he was seeing bottlenecks with IDR.

I will test a few hash algorithms (including common one like jenkins)
for this use case.
Problem right now is we have that rcu-inspired upper bucket limit
(which exists) in fw as well: If i have 1M entries which can only be
spread to 256 buckets perfectly then i have ~4K entries per bucket
which are a linked list. So we are still not doing so well. Do you have
time to make that better for fw (since you are an rcu maestro).

cheers,
jamal

Powered by blists - more mailing lists