On Fri, Jan 12, 2018 at 03:55:10PM +0000, Ferruh Yigit wrote:

On 1/12/2018 2:37 PM, Neil Horman wrote:

quoted

On Fri, Jan 12, 2018 at 11:50:12AM +0000, Ferruh Yigit wrote:

quoted

On 1/11/2018 9:29 PM, Neil Horman wrote:

quoted

On Thu, Jan 11, 2018 at 08:06:48PM +0000, Ferruh Yigit wrote:

quoted

On 12/13/2017 3:17 PM, Neil Horman wrote:

quoted

Document the need to add the __experimental tag to appropriate functions

Signed-off-by: Neil Horman <nhorman@tuxdriver.com>
CC: Thomas Monjalon <redacted>
CC: "Mcnamara, John" <redacted>
CC: Bruce Richardson <redacted>

<...>

quoted

quoted

quoted

 automatically marked as ``experimental`` to allow for a period of stabilization>>>  before they become part of a tracked ABI.

Full sentences for above statement:
"
Since changes to APIs are most likely immediately after their introduction, as
users begin to take advantage of those new APIs and start finding issues with
them, new DPDK APIs will be automatically marked as experimental to allow for a
period of stabilization before they become part of a tracked ABI.
"

This part is not related to this patchset, but it will be hard to maintain above
behavior, "automatically marked" is not automatic now and moving them to stable
after one release is also not automatic. Do you have any suggestion on how to
manage this, do you think can your script be expanded to cover these checks?

I would make the argument that this was never the case, but rather a statement
of principle.  I assert that because I can find no mechanism anywhere in our
build system that 'automatically' documented or marked a new API as experimental
(please correct me if I'm wrong here).  I think this was more meant to be a
directive to developers to do whatever coding was needed to preform such
marking/documentation in whatever style/format was current.  E.g. introducers of
a new API should document everything as EXPERIMENTAL using the appropriate
doxygen tag and version map tag.

In answer to your question, While we might be able to expand my script to check
for new API's and ensure they are marked as experimental, I don't think thats
the right place to do it, because that script is run at build time, where the
state of the tree is transient. A better place to do it would be with a git hook
at checkin time, or in the checkpatch script to flag new apis as experimental
right before those new API's are comitted.  Though I'm not a huge fan of that
either (at least not of the former suggestion).  I say that because I think we
need to allow developers the freedom to determine the supported status of any
new API that they add.  For example, it seems pretty clear that a new library
might want to have its entire API marked as experimental, but someone adding a
single new function to an existing API might be confident that the new function
is needed and should be immediately supported..

I think the better solution is to define the use of the EXPERIMENTAL tag in the
version map as the canonical location to define unstable API functions.  Doing
so immediately commits an author to ensuring that the corresponding function
definitions are marked with the __experimental tags, which in turn will cause
any downstream users to be alerted to the fact that they might be using those
API's in their code, so they can take appropriate action.  It still allows for
the Documentation to be out of sync, but alerting authors doing development I
think is the more important element here, as Documentation can be corrected more
easily than code in the field.

Thoughts?

After this point agree to using EXPERIMENTAL tag in the version map as standard,
but it will be hard to maintain "API is experimental for first release" without
help of any automated tool.

I completely agree, in fact I would say it is impossible to do without tooling,
with or without this change.  I think we need to do 1 of 2 things:

1) Add some code to checkpatch.pl to put up a warning if any new apis are added
without marking them as experimental

2) Change the documentation to be a suggestion rather than a requirement.

I'll look into doing (1), but I'm wondering if (2) is the more flexible way to
go. I'm hesitant to enforce the initial marking of new APIs as experimental.
Thoughts?

Neil