[ovs-dev] [dpdk-dev] Direct using of 'rte_eth_devices' in DPDK apps.

[Wiles, Keith]

On Nov 16, 2018, at 3:51 AM, Ananyev, Konstantin <konstantin.ananyev at intel.com<mailto:konstantin.ananyev at intel.com>> wrote:

Hi everyone,


Hi,

16/11/2018 09:42, Ilya Maximets:
Hi,
While discussing the ways to enable DPDK 18.11 new features in OVS
there was suggestions to use 'rte_eth_devices[]' array directly.
But this array is marked as '@internal' and also it located in
the internal header 'lib/librte_ethdev/rte_ethdev_core.h' with the
following disclaimer:

/**
* @file
*
* RTE Ethernet Device internal header.
*
* This header contains internal data types. But they are still part of the
* public API because they are used by inline functions in the published API.
*
* Applications should not use these directly.
*
*/

>From the other hand, test-pmd and some example apps in DPDK source
tree are using this array for various reasons.

So, is it OK to use this array directly or not?

Good question :)
Thanks for bringing this discussion.

As you said, it is public because of inline functions using it directly
for performance purpose. The DPDK API is bad for separating public and
internal stuff. And over time, there is not a lot of attention on trying
to not use internal symbols in applications.

In general we need to change the API, i.e. make 'rte_eth_devices' part
of a public API. Or change the test-pmd and example apps to stop
using it.

I agree we need to decide an option and make it clear.

We can try to make this variable private and add more public functions
to use it (I'm thinking at more iterators like sibling ones).
It would clarify the API.
It can be evaluated what is the real cost after compiler optimization
for Rx/Tx functions. It can also be evaluated to uninline functions.

On the other hand, we can wonder what is the real benefit of trying to
hide access to internal resources. Should we make all public?

In that case every change in any of such structures will be an ABI breakage.
Even now any change in rte_eth_dev is quite problematic because of that.
I think we better keep them private as much as possible and cleanup
our examples and testpmd code.
Konstantin

I Agree here, I have noticed a few places we allow direct access to internal data structures, which we need to restrict access by making them private with getter/setter functions or just getter/setter functions even if we can not make them private. At least with moving members and adding members we can state that it is not a ABI breakage as long as everyone uses the getter/setter functions. These functions could not be inline functions correct as that would still break API?


One more related question: Is it OK to access internal device
stuff using 'device' pointer obtained by 'rte_eth_dev_info'?
This looks really dangerous. It's unclear why pointers like this
exposed to user.

It's a lot easier to expose pointers than doing a good API for all uses.
We need to question what is really dangerous and what we want to avoid?

Regards,
Keith