Re: [dpdk-dev] [PATCH] add flow shared action API

[Ferruh Yigit <ferruh.yigit@intel.com>]

On 7/7/2020 6:24 PM, Ori Kam wrote:
> Hi Ferruh,
> 
>> -----Original Message-----
>> From: Ferruh Yigit <ferruh.yigit@intel.com>
>>
>> On 7/7/2020 7:21 AM, Ori Kam wrote:
>>> Hi Jerin,
>>>  Thanks you for your quick reply.
>>>
>>>> -----Original Message-----
>>>> From: Jerin Jacob <jerinjacobk@gmail.com>
>>>> Subject: Re: [dpdk-dev] [PATCH] add flow shared action API
>>>>
>>>> On Mon, Jul 6, 2020 at 7:02 PM Andrey Vesnovaty
>>>> <andrey.vesnovaty@gmail.com> wrote:
>>>>>
>>>>> Hi, Jerin.
>>>>
>>>> Hi Ori and Andrey,
>>>>
>>>>
>>>>>
>>>>> Please see below Ori's suggestion below to implement your
>>>> rte_flow_action_update() idea
>>>>> with some API changes of rte_flow_shared_action_xxx API changes.
>>>>>
>>>>> On Mon, Jul 6, 2020 at 3:28 PM Ori Kam <orika@mellanox.com> wrote:
>>>>>>
>>>>>> Hi Jerin,
>>>>>>
>>>>>>> -----Original Message-----
>>>>>>> From: Jerin Jacob <jerinjacobk@gmail.com>
>>>>>>> Sent: Monday, July 6, 2020 12:00 PM
>>>>>>> Subject: Re: [dpdk-dev] [PATCH] add flow shared action API
>>>>>>>
>>>>>>> On Sun, Jul 5, 2020 at 3:56 PM Ori Kam <orika@mellanox.com> wrote:
>>>>>>>>
>>>>>>>> Hi Jerin,
>>>>>>>> PSB,
>>>>>>>>
>>>>>>>> Thanks,
>>>>>>>> Ori
>>>>>>>>
>>>>>>>>> -----Original Message-----
>>>>>>>>> From: Jerin Jacob <jerinjacobk@gmail.com>
>>>>>>>>> Sent: Saturday, July 4, 2020 3:33 PM
>>>>>>>>> dpdk-dev <dev@dpdk.org>
>>>>>>>>> Subject: Re: [dpdk-dev] [PATCH] add flow shared action API
>>>>>>>>>
>>>>>>>>> On Sat, Jul 4, 2020 at 3:40 PM Andrey Vesnovaty
>>>>>>>>> <andrey.vesnovaty@gmail.com> wrote:
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>> Thanks,
>>>>>>>>>>
>>>>>>>>>> Andrey Vesnovaty
>>>>>>>>>> (+972)526775512 | Skype: andrey775512
>>>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>> [..Nip ..]
>>>>>>>>
>>>>>>>>>> I need to mention the locking issue once again.
>>>>>>>>>> If there is a need to maintain "shared session" in the generic
>>>> rte_flow
>>>>>>> layer
>>>>>>>>> all
>>>>>>>>>> calls to flow_create() with shared action & all delete need to take
>>>>>>>>> sharedsession
>>>>>>>>>> management locks at least for verification. Lock partitioning is also
>>>> bit
>>>>>>>>> problematic
>>>>>>>>>> since one flow may have more than one shared action.
>>>>>>>>>
>>>>>>>>> Then, I think better approach would be to introduce
>>>>>>>>> rte_flow_action_update() public
>>>>>>>>> API which can either take "const struct rte_flow_action []" OR shared
>>>>>>>>> context ID, to cater to
>>>>>>>>> both cases or something on similar lines. This would allow HW's
>>>>>>>>> without have  the shared context ID
>>>>>>>>> to use the action update.
>>>>>>>>
>>>>>>>> Can you please explain your idea?
>>>>>>>
>>>>>>> I see two types of HW schemes supporting action updates without going
>>>>>>> through call `rte_flow_destroy()` and call `rte_flow_create()`
>>>>>>> - The shared HW action context feature
>>>>>>> - The HW has "pattern" and "action" mapped to different HW objects
>> and
>>>>>>> action can be updated any time.
>>>>>>> Other than above-mentioned RSS use case, another use case would be to
>>>>>>> a) create rte_flow and set the action as DROP (Kind of reserving the HW
>>>> object)
>>>>>>> b) Update the action only when the rest of the requirements ready.
>>>>>>>
>>>>>>> Any API schematic that supports both notions of HW is fine with me.
>>>>>>>
>>>>>> I have an idea if the API will be changed to something like this,
>>>>>> Rte_flow_shared_action_update(uint16_port port, rte_shared_ctx *ctx,
>>>> rte_flow_action *action, error)
>>>>>> This will enable the application to send a different action than the original
>>>> one to be switched.
>>>>>> Assuming the PMD supports this.
>>>>>> Does it answer your concerns?
>>>>>
>>>>>
>>>>> This allows both:
>>>>> 1. Update action configuration
>>>>> 2. Replace action by some other action
>>>>> For 2 pure software implementation may carate shred action (that can be
>>>> shared
>>>>> with one flow only, depends on PMD) and later on
>>>> rte_flow_shared_action_update may replace this
>>>>> action with some other action by handle returned from
>>>> rte_flow_shared_action_create
>>>>> Doesign between 1 and 2 is per PMD.
>>>>
>>>> struct rte_flow * object holds the driver representation of the
>>>> pattern + action.
>>>> So in order to update the action, we would need struct rte_flow * in API.
>>>>
>>> Why is that? The idea is to change the action, the action itself is connected to
>> flows.
>>> The PMD can save in the shared_ctx all flows that are connected to this
>> action.
>>>
>>>> I think, simple API change would be to accommodate "rte_shared_ctx
>>>> *ctx, rte_flow_action *action" modes
>>>> without introducing the emulation for one or other mode, will be.
>>>>
>>>> enum rte_flow_action_update_type {
>>>>               RTE_FLOW_ACTION_UPDATE_TYPE_SHARED_ACTION,
>>>>               RTE_FLOW_ACTION_UPDATE_TYPE_ACTION,
>>>> };
>>>>
>>>> struct rte_flow_action_update_type_param {
>>>>          enum rte_flow_action_update_type type;
>>>>          union {
>>>>                      struct rte_flow_action_update_type_shared_action_param {
>>>>                                 rte_shared_ctx *ctx;
>>>>                       } shared_action;
>>>>                       struct rte_flow_action_update_type_shared_action_param {
>>>>                                 rte_flow *flow,
>>>>                                  rte_flow_action *action;
>>>>                       } action;
>>>>          }
>>>> }
>>>>
>>> Thank you for the idea but I fall to see how your suggested API is simpler than
>> the one suggested by me?
>>> In my suggestion the PMD simply needs to check if the new action and
>> change the
>>> context and to that action, or just change parameters in the action, if it is the
>> same action.
>>>
>>> Let's go with the original patch API modified to support like you requested
>> also changing the action,
>>> based on my comments.
>>>
>>>> rte_flow_action_update(uint16_port port, struct
>>>> rte_flow_action_update_type_param  *param, error)
>>>>
>>>>>
>>>>>>
>>>>>>>
>>>>>>>> As I can see if we use the flow_action array it may result in bugs.
>>>>>>>> For example, the application created two flows with the same RSS (not
>>>> using
>>>>>>> the context)
>>>>>>>> Then he wants to change one flow to use different RSS, but the result
>> will
>>>> that
>>>>>>> both flows
>>>>>>>> will be changed.
>>>>>>>
>>>>>>> Sorry. I don't quite follow this.
>>>>>>>
>>>>>> I was trying to show that there must be some context. But I don’t think this
>> is
>>>> relevant to
>>>>>> your current ideas.
>>>>>>
>>>>>>>> Also this will enforce the PMD to keep track on all flows which will have
>>>>>>> memory penalty for
>>>>>>>> some PMDs.
>>
>> Hi Ori, Andrey,
>>
>> This is a set of new APIs and we are very close to the -rc1, so we have only a
>> few days to close the feature to merge them for this release.
>>
>> Also accompanying PMD and testpmd implementation with the proposed API
>> changes
>> looks missing.
>>
>> We can either postpone the patchset to next release to give time for more
>> PMD
>> owners to participate, which can give better API for long term.
>> Or try to to squeeze into this release taking into account that the APIs will be
>> experimental.
>>
>> What do you think, what is you schedule for the feature, do you have room to
>> postpone it?
> Not so much it is an important API for Mellanox.

Got it.

> 
>> If not, first existing discussions needs to resolved, and it is good to have the
>> PMD and testpmd implementations, do you think can this be done for next few
>> days?
>>
> I think that this is the correct API to implement, I fully agree that this API is experimental
> just like any other new API, and might change based on comments and use cases.
> I know that Mellanox is committed to this feature and that Andrey is working around the clock 
> to complete the missing parts, and should have a version by tomorrow (July 8th ) evening.

OK

> (with update to flow filtering sample app, testpmd will not be ready by RC1, but it will be for RC2)
> We would like very much to push it in this version.

OK, please conclude the existing discussion before finalizing.