Re: [dpdk-dev] [PATCH v3] eventdev: update crypto adapter metadata structures

[Akhil Goyal <gakhil@marvell.com>]

Hi Abhinandan,
> > > >> >>
> > > >> >> >> In crypto adapter metadata, reserved bytes in request info
> > > >> >> >> structure is a space holder for response info. It enforces an
> > > >> >> >> order of operation if the structures are updated using memcpy
> > > >> >> >> to avoid overwriting response info. It is logical to move the
> > > >> >> >> reserved space out of request info. It also solves the
> > > >> >> >> ordering issue
> > > >> mentioned before.
> > > >> >> >I would like to understand what kind of ordering issue you have
> > > >> >> >faced with the current approach. Could you please give an
> > > >> >> >example/sequence
> > > >> >> and explain?
> > > >> >> >
> > > >> >>
> > > >> >> I have seen this issue with crypto adapter autotest (#n215).
> > > >> >>
> > > >> >> Example:
> > > >> >> rte_memcpy(&m_data.response_info, &response_info,
> > > >> >> sizeof(response_info)); rte_memcpy(&m_data.request_info,
> > > >> >> &request_info, sizeof(request_info));
> > > >> >>
> > > >> >> Here response info is getting overwritten by request info.
> > > >> >> Above lines can reordered to fix the issue, but can be ignored
> > > >> >> with this
> > > >> patch.
> > > >> >There is a reason for designing the metadata in this way.
> > > >> >Right now, sizeof (union rte_event_crypto_metadata) is 16 bytes.
> > > >> >So, the session based case needs just 16 bytes to store the data.
> > > >> >Whereas, for sessionless case each crypto_ops requires another 16
> > > bytes.
> > > >> >
> > > >> >By changing the struct in the following way you are doubling the
> > > >> >memory requirement.
> > > >> >With the changes, for sessionless case, each crypto op requires 32
> > > >> >bytes of space instead of 16 bytes and the mempool will be bigger.
> > > >> >This will have the perf impact too!
> > > >> >
> > > >> >You can just copy the individual members(cdev_id & queue_pair_id)
> > > >> >after the response_info.
> > > >> >OR You have a better way?
> > > >> >
> > > >>
> > > >> I missed the second word of event structure. It adds an extra 8
> > > >> bytes, which is not required.
> > > >I guess you meant not adding, it is overlapping with request
> information.
> > > >> Let me know, what you think of below change to metadata structure.
> > > >>
> > > >> struct rte_event_crypto_metadata {
> > > >> 	uint64_t response_info; // 8 bytes
> > > >With this, it lags clarity saying it is an event information.
> > > >> 	struct rte_event_crypto_request request_info; // 8 bytes };
> > > >>
> > > >> Total structure size is 16 bytes.
> > > >>
> > > >> Response info can be set like below in test app:
> > > >> 	m_data.response_info = response_info.event;
> > > >>
> > > >> The main aim of this patch is to decouple response info from request
> > info.
> > > >The decoupling was required as it was doing memcpy.
> > > >If you copy the individual members of request info(after response),
> > > >you don't require it.
> > >
> > > With this change, the application will be free of such constraints.
> >
> > [Anoob] Why don't we make it like,
> >
> > struct rte_event_crypto_metadata {
> >  	uint64_t ev_w0_template;
> >               /**< Template of the first word of ``struct rte_event``
> > (rte_event.event) to be set in the events generated by crypto adapter in
> > both RTE_EVENT_CRYPTO_ADAPTER_OP_NEW and
> > 	 * RTE_EVENT_CRYPTO_ADAPTER_OP_FORWARD modes.
> > 	*/
> >  	struct rte_event_crypto_request request_info; };
> >
> > This way, the usage would become more obvious and we will not have
> > additional space reserved as well.
> 
> Thanks for the suggestion. At this point, it is like an application without
> understanding the data structure/ internals
> has used memcpy and we are trying to fix it by changing the actual data
> structure instead of fixing the application!
> With this patch, the other applications(outside of dpdk) which are using the
> data structures in a right are forced to change!
> I don't think, this is the right way to handle this. I think, we should be
> updating the test case and documentation for this rather
> than changing the data structure.  I expect the next version of this patch
> should have those changes. Thanks!

The point here is about making the specification better and clearer to the user.
If the structure is not clear and is error prone if somebody does not follow
Proper documentation, we can modify it to reduce possible issues.

This is a cleanup which was added in deprecation notice as well in the last release.
This is a ABI break release and we should do this cleanup if it is legit.
Applications outside DPDK are notified as per the deprecation notice in the last release.
We have followed standard procedure to modify this structure,
hence, no need to worry about that.
We have provided, 2-3 suggestions to clean this up. Please suggest which
one is best for Intel use case.

Having first element as reserved in the structure doesn't look quite obvious.
This was also highlighted in the original patch, but was not taken up seriously
as we were not supporting it at that point.
See this
http://patches.dpdk.org/project/dpdk/patch/1524573807-168522-2-git-send-email-abhinandan.gujjar@intel.com/

-Akhil

Please note:
Avoid top posting for comments and remove unnecessary lines while commenting back.