Message-ID: <0272f671-802b-30b6-6ca2-2ffc1e205664@intel.com>
Date:   Wed, 27 Jun 2018 09:46:18 -0700
From:   "Samudrala, Sridhar" <sridhar.samudrala@...el.com>
To:     Jiri Pirko <jiri@...nulli.us>,
        Jakub Kicinski <jakub.kicinski@...ronome.com>
Cc:     Linux Netdev List <netdev@...r.kernel.org>,
        David Miller <davem@...emloft.net>,
        Jamal Hadi Salim <jhs@...atatu.com>,
        Cong Wang <xiyou.wangcong@...il.com>,
        Simon Horman <simon.horman@...ronome.com>,
        John Hurley <john.hurley@...ronome.com>,
        David Ahern <dsahern@...il.com>, mlxsw@...lanox.com
Subject: Re: [patch net-next 0/9] net: sched: introduce chain templates
 support with offloading to mlxsw

On 6/27/2018 12:50 AM, Jiri Pirko wrote:
> Tue, Jun 26, 2018 at 11:18:58PM CEST, jakub.kicinski@...ronome.com wrote:
>> On Tue, 26 Jun 2018 09:12:17 +0200, Jiri Pirko wrote:
>>> Tue, Jun 26, 2018 at 09:00:45AM CEST, jakub.kicinski@...ronome.com wrote:
>>>> On Mon, Jun 25, 2018 at 11:43 PM, Jiri Pirko <jiri@...nulli.us> wrote:
>>>>> Tue, Jun 26, 2018 at 06:58:50AM CEST, jakub.kicinski@...ronome.com wrote:
>>>>>> On Mon, 25 Jun 2018 23:01:39 +0200, Jiri Pirko wrote:
>>>>>>> From: Jiri Pirko <jiri@...lanox.com>
>>>>>>>
>>>>>>> For the TC clsact offload these days, some of HW drivers need
>>>>>>> to hold a magic ball. The reason is, with the first inserted rule inside
>>>>>>> HW they need to guess what fields will be used for the matching. If
>>>>>>> later on this guess proves to be wrong and user adds a filter with a
>>>>>>> different field to match, there's a problem. Mlxsw resolves it now with
>>>>>>> couple of patterns. Those try to cover as many match fields as possible.
>>>>>>> This aproach is far from optimal, both performance-wise and scale-wise.
>>>>>>> Also, there is a combination of filters that in certain order won't
>>>>>>> succeed.
>>>>>>>
>>>>>>> Most of the time, when user inserts filters in chain, he knows right away
>>>>>>> how the filters are going to look like - what type and option will they
>>>>>>> have. For example, he knows that he will only insert filters of type
>>>>>>> flower matching destination IP address. He can specify a template that
>>>>>>> would cover all the filters in the chain.
>>>>>> Perhaps it's lack of sleep, but this paragraph threw me a little off
>>>>>> the track.  IIUC the goal of this set is to provide a way to inform the
>>>>>> HW about expected matches before any rule is programmed into the HW.
>>>>>> Not before any rule is added to a particular chain.  One can just use
>>>>>> the first rule in the chain to make a guess about the chain, but thanks
>>>>>> to this set user can configure *all* chains before any rules are added.
>>>>> The template is per-chain. User can use template for chain x and
>>>>> not-use it for chain y. Up to him.
>>>> Makes sense.
>>>>
>>>> I can't help but wonder if it'd be better to associate the
>>>> constraints/rules with chains instead of creating a new "template"
>>>> object.  It seems more natural to create a chain with specific
>>>> constraints in place than add and delete template of which there can
>>>> be at most one to a chain...  Perhaps that's more about the user space
>>>> tc command line.  Anyway, not a strong objection, just a thought.
>>> Hmm. I don't think it is good idea. User should see the template in a
>>> "show" command per chain. We would have to have 2 show commands, one to
>>> list the template objects and one to list templates per chains. It makes
>>> things more complicated for no good reason. I think that this simple
>>> chain-lock is easier and serves the purpose.
>> Hm, I think the dump is fine, what I was thinking about was:
>>
>> # tc chain add dev dummy0 ingress chain_index 22 \
>>      ^^^^^
>> 	template proto ip \
>> 	^^^^^^^^
>> 	flower dst_mac 00:00:00:00:00:00/00:00:00:00:FF:FF
> Okay, I got it. I see 2 issues.
> 1) user might expect to add a chain without the template. But that does
>     not make sense really. Chains are created/deleted implicitly
>     according to refcount.
> 2) there is not chain object like this available to user. Adding it just
>     for template looks odd. Also, the "filter" and "template" are very
>     much alike. They both are added to a chain, they both implicitly
>     create chain if it does not exist, etc.
>
> if you don't like "tc filter template add dev dummy0 ingress", how
> about:
> "tc template add dev dummy0 ingress ..."
> "tc template add dev dummy0 ingress chain 22 ..."
> that makes more sense I think.

Isn't it possible to avoid introducing another keyword 'template',

Can't we just do
       tc chain add dev dummy0 ingress flower chain_index 0
to create a chain that takes any types of flower rules with index 0
and
      tc chain add dev dummy0 ingress flower chain_index 22
             dst_mac 00:00:00:00:00:00/00:00:00:00:FF:FF
      tc chain add dev dummy0 ingress flower chain_index 23
             dst_ip 192.168.0.0/16
to create 2 chains 22 and 23 that allow rules with specific fields.

To create a chain that takes decap rules like
   filter protocol ip pref 2 flower chain 25 handle 0x2
     dst_mac fe:24:9a:23:4c:5c
     src_mac ce:1d:58:eb:a0:35
     eth_type ipv4
     enc_dst_ip 192.168.100.20
     enc_src_ip 192.168.100.22
     enc_key_id 100
     enc_dst_port 4789

can we create a chain using this format

tc chain add dev dummy0 ingress flower chain_index 25
     dst_mac 00:00:00:00:00:00/FF:FF:FF:FF:FF:FF
     src_mac 00:00:00:00:00:00/FF:FF:FF:FF:FF:FF
     enc_dst_ip 192.168.100.0/24
     enc_src_ip 192.168.100.0/24
     eth_type ipv4/arp
     enc_key_id 0/ffff
     enc_dst_port 0/ffff

When adding a filter to a pre-defined chain, does the filter need
to specify all the fields that are included in the chain's template?

>
>
>> instead of:
>>
>> # tc filter template add dev dummy0 ingress \
>>      ^^^^^^^^^^^^^^^
>> 	proto ip chain_index 22 \
>> 	flower dst_mac 00:00:00:00:00:00/00:00:00:00:FF:FF
>>
>> And then delete becomes:
>>
>> # tc chain del dev dummy0 ingress chain_index 22
>> Error: The chain is not empty.
>>
>> The fact that template is very much like a filter is sort of an
>> implementation detail, from user perspective it may be more intuitive
>> to model template as an attribute of the chain, not a filter object
>> added to a chain.
>>
>> But I could well be the only person who feels that way :)

Powered by blists - more mailing lists