Re: [dpdk-dev] [RFC v2 0/1] add flow action context API

[Jerin Jacob <jerinjacobk@gmail.com>]

On Wed, Jul 1, 2020 at 2:54 PM Andrey Vesnovaty
<andrey.vesnovaty@gmail.com> wrote:
>
>
> On Tue, Jun 30, 2020 at 12:52 PM Jerin Jacob <jerinjacobk@gmail.com> wrote:
>>
>> On Mon, Jun 29, 2020 at 3:52 PM Andrey Vesnovaty
>> <andrey.vesnovaty@gmail.com> wrote:
>> >
>> >
>> >
>> > On Sun, Jun 28, 2020 at 4:42 PM Jerin Jacob <jerinjacobk@gmail.com> wrote:
>> >>
>> >> On Sun, Jun 28, 2020 at 2:14 PM Andrey Vesnovaty
>> >> <andrey.vesnovaty@gmail.com> wrote:
>> >> >
>> >> > Hi
>> >> >
>> >> > On Fri, Jun 26, 2020 at 2:44 PM Jerin Jacob <jerinjacobk@gmail.com> wrote:
>> >> >>
>> >> >> On Sat, Jun 20, 2020 at 7:02 PM Andrey Vesnovaty
>> >> >> <andrey.vesnovaty@gmail.com> wrote:
>> >> >> >
>> >> >> > Hi, and thanks a lot for your RFC v1 comments.
>> >> >> >
>> >> >> > RFC v2 emphasize the intent for sharing the flow action:
>> >> >> > * The term 'action context' was unclear and replaced with
>> >> >> >    'shared action'.
>> >> >> > * RFC v2 subject became 'add flow shared action API'.
>> >> >> > * all proposed APIs renamed according the above.
>> >> >> >
>> >> >> > The new shared action is an independent entity decoupled from any flow
>> >> >> > while any flow can reuse such an action. Please go over the RFC
>> >> >> > description, it was almost entirely rewritten.
>> >> >> >
>> >> >> > @Jerin Jacob:
>> >> >> > Thanks again for your comments, it made me admit that v1 description was
>> >> >> > incomplete & unclear.  I hope v2 will be better at least in terms of
>> >> >> > clarity.
>> >> >>
>> >> >> The public API and its usage is very clear. Thanks for this RFC.
>> >> >
>> >> >
>> >> > My pleasure.
>> >> >>
>> >> >>
>> >> >> I think, RFC v2 still not addressing the concern raised in the
>> >> >> http://mails.dpdk.org/archives/dev/2020-June/169296.html.
>> >> >>
>> >> >> Since MLX hardware has an HW based shared object it is fine to have
>> >> >> public API based on that level of abstraction.
>> >> >> But at the PMD driver level we need to choose the correct abstraction
>> >> >> to support all PMD and support shared object scheme if possible.
>> >> >>
>> >> >> I purpose to introduce something below or similar
>> >> >>             int (*action_update)
>> >> >>                 (struct rte_eth_dev *,
>> >> >>                   struct rte_flow *flow,
>> >> >>                  const struct rte_flow_action [],
>> >> >>                  struct rte_flow_error *);
>> >> >
>> >> > Where this callback suppose to belong (struct rte_flow_ops)?
>> >>
>> >> Yes.
>> >>
>> >> > How should it be implemented by PMD?
>> >>
>> >> See below,
>> >>
>> >> > Is it about shared action and if "yes" why there is 'flow' argument?
>> >>
>> >> flow holds the "pattern" and "action" data as PMD specific handle.
>> >> So PMD, implementation can just change that action if it gets the PMD
>> >> specific handle.
>> >>
>> >>
>> >> >>
>> >> >>
>> >> >> in addition to: shared_action_create, shared_action_destroy,
>> >> >> shared_action_update, shared_action_query
>> >> >>
>> >> >> Have generic implementation of above, if action_update callback is not
>> >> >> NULL.
>> >> >
>> >> > "is not NULL" -> "is NULL"?
>> >>
>> >> Yes. When it is NULL.
>> >
>> >
>> > Jerin, few clarifications regarding generic implementation of shared action:
>> > Based on this conversation I'm assuming that generic implementation supposed to be something like:
>> > For each flow using some shared action:
>> > call ops-> action_update()
>> > If the assumption above correct:
>> > 1. taking into account that shared_action_update() is atomic, how can this deal with partial success: some flows may fail validation - should it:
>> >   1.1.lock all flows
>> >   1.2.validate all flows
>> >   1.3.update all flows
>> >   1.4. unlock
>>
>> Yes.
>
> This kind of locking in addition to shared session management requires locking of each flow_create/flow_destroy in addition to action_uodate callback implementation even if there are no shared actions at all. In other words it imposes an overhead on all PMDs that don't support shared action natively.

Yes. That's what my concern with implementing shared session if the
PMD only supports only action update for the given rte_flow *.
Another 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.




>>
>>
>> > 2. action_update callback is PMD specific & if it's unsupported there is no support for shared action any way
>>
>> Yes.
>>
>> > Please address the issues above
>>
>> >
>> >> >
>> >> >>
>> >> >> So that, it can work all PMDs and to
>> >> >> avoid the duplication of "complex" shared session management code.
>> >> >
>> >> > Do you mean shared action in use by multiple flows by "shared session"?
>> >>
>> >> Yes.
>> >
>> > Common 'shared session' management code:
>> > - can be reduced to atomic usage counter
>> > - maintaining list of flow using shared action expected to impact performance & not necessary for all PMD specific implementations
>> > Access to other shared resources hard to generalize because:
>> > - for some PMDs mutual exclusion is HW feature & no need to protect it in SW
>> > - for others there may be multiple resources & access to each one protected by different mechanism
>>
>> The general callback you can assume, it supports only action_update
>> based callback.
>> If PMD has mutual exclusion HW feature then it can override the
>> function pointers.
>>
>>
>>
>> > An observation related to action_update callback:
>> > If replaced (updated) action was shared then the flow won't be influenced any more by updates or removed shared action.