Re: ethdev: hide internal structures

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

On 11/16/2021 5:54 PM, Tyler Retzlaff wrote:
> On Tue, Nov 16, 2021 at 09:32:15AM +0000, Ferruh Yigit wrote:
>>
>> Hi Tyler,
>>
>> It shouldn't be an API change, which API is changed?
> 
> exported declarations that were consumed by the application were removed
> from an installed header. anything making reference to rte_eth_devices[]
> will no longer compile.
> 
> any change that removes any identifier or macro visible to the application
> from an installed header is an api break.
> 
>> Existing binaries won't run and needs recompile, but shouldn't need to change
>> the code.
>> Unless application is accessing *internal* DPDK structs (which were exposed
>> to application because of some technical issues that above commit fixes).
> 
> the application was, but the access was to a symbol and identifier that
> had not been previously marked __rte_internal or __rte_experimental and thus
> assumed to be public.
> 
> just to be clear i agree with the change making these internal but there
> was virtually no warning.
> 
> https://doc.dpdk.org/guides-19.11/contributing/abi_policy.html
> 
> the exports and declarations need to be marked deprecated to give ample
> time before being removed in accordance with the abi policy.
> 
> i will ask that work be scheduled to identify the gap in the public api
> surface that access to these structures was providing rather than
> backing the change out. fortunately it is only schedule rather
> than service impacting since the application hadn't been deployed yet.
> 
> i thought someone was responsible for reviewing abi/api related changes
> on the board to understand the implications of changes like this?
> 

Sorry for the negative impact on your product, I can understand the
frustration.

The 'rte_eth_devices[]' was marked as '@internal' in the header file
since 2012 [1], so it is not new, but it was not marked programmatically,
only as comment in the header file.
Expectation was applications to not directly use it.


For long term ABI stability, this is a good step forward, although
the impact was known, best time for these kind of change is the 21.11
release, otherwise change needs to wait (at least) one more year.

This change has been discussed and accepted in the technical board [2],
and a deprecation notice has been sent to mail list [3] for notification.

Agree the announce was a little late than we normally do (although
only a month late than what defined in process), this is accepted by
the board to not miss the ABI break window (.11 release).
As you will recognize, not only ethdev, but a few more device abstraction
layer libraries had similar changes in this release.


[1]
f831c63cbe86 ("ethdev: minor changes")

[2]
https://mails.dpdk.org/archives/dev/2021-July/214662.html

[3]
https://patches.dpdk.org/project/dpdk/patch/20210826103500.2172550-1-ferruh.yigit@intel.com/