Re: [dpdk-dev] [PATCH] mk: enable next abi in static libs

[Neil Horman <nhorman@tuxdriver.com>]

On Mon, Jul 06, 2015 at 11:44:59PM +0200, Thomas Monjalon wrote:
> 2015-07-06 14:22, Neil Horman:
> > On Mon, Jul 06, 2015 at 03:49:50PM +0200, Thomas Monjalon wrote:
> > > 2015-07-06 09:35, Neil Horman:
> > > > On Mon, Jul 06, 2015 at 03:18:51PM +0200, Thomas Monjalon wrote:
> > > > > Any comment or ack?
> > > > > 
> > > > > 2015-07-03 00:05, Thomas Monjalon:
> > > > > > When a change makes really hard to keep ABI compatibility,
> > > > > > instead of waiting next release to break the ABI, it is smoother
> > > > > > to introduce the new code and enable it only for static libraries.
> > > > > > The flag RTE_NEXT_ABI may be used to "ifdef" the new code.
> > > > > > When the release is out, a dynamically linked application can use
> > > > > > the new shared libraries without rebuild while developpers can prepare
> > > > > > their application for the next ABI by reading the deprecation notice
> > > > > > and easily testing the new code.
> > > > > > When starting the next release cycle, the "ifdefs" will be removed
> > > > > > and the ABI break will be marked by incrementing LIBABIVER.
> > > > > > 
> > > > > > The new option CONFIG_RTE_NEXT_ABI is not defined in the configuration
> > > > > > templates because it is deduced from CONFIG_RTE_BUILD_SHARED_LIB.
> > > > > > It is automatically enabled for static libraries and disabled for
> > > > > > shared libraries.
> > > > > > It can be forced to another value by editing the generated .config file.
> > > > > > It shouldn't be enabled for shared libraries because it would break the
> > > > > > ABI without changing the version number LIBABIVER. That's why a warning
> > > > > > is printed in this case.
> > > > > > 
> > > > > > The guideline is also updated to integrate this new possibility.
> [...]
> > I'd be ok with it iff:
> > 
> > 1) It applies to static and shared ABI's together.  That is to say that setting
> > the NEXT_ABI config flag creates the same ABI changes regardless of other build
> > configuration.  It needs to be used in such a way that a consistent ABI is
> > presented when set, otherwise it won't be useful.
> 
> Yes the option trigger exactly the same ABI for static and shared libraries.
> But it's too complicated (at least for 2.1) to make LIBABIVER and version map
> dependant of this build-time option.

No, I think thats a bridge too far.  I'm not sure whats difficult about
overriding LIBABIVER in lib.rte.mk and bump all numbers 1 higher (or better
still just add a .1 to the end of it), by checking CONFIG_NEXT_ABI

As for maintaining the version map, I don't see any problem with duplicating the
map files, to a -next variant, and changing the CPU_LDFLAGS in rte.lib.mk based
on the NEXT_ABI config option again.

In fact, if this is a thing that people want, that might be beneficial, as
something else occurs to me.  I think you're going to want this to be a mandated
piece of the update process.  That is to say, if someone wants to deprecate an
aspect of the ABI, or change it, I think you'll want to mandate that they submit
the change at the same time they submit the deprecation notice, and simply
protect it with this NEXT_ABI config option.  That provides several advantages:

1) It ensures that the notice is submitted at the same time as the actual change
2) It ensures that the NEXT_ABI provides a complete view of what the next ABI
version looks like, not just a partial view of it

Adding a *-version-next.map file for each library makes adhering to the above
easier, and allows for an easy converstion, in that when its time to officially
update the ABI, fixing the version map is a matter of copying
<library>-version-next.map to <library>-version.map.

The use case that I'm thinking of here is as such:

Consider two ABI modifying updates, A and B.

The author of A writes his changes, submits them with appropriate ifdefs for
CONFIG_NEXT_ABI, along with a deprecation notice.

The author of B writes his changes, but doesn't submit them, instead submitting
only a deprecation notice, with plans to post the actuall patches after the
deprecation notice is shipped in a release

After the release CONFIG_NEXT_ABI exposes the ABI changes made by A, but not by
B (because they don't yet exist in the code).  I think to give users a complete
view of the NEXT_ABI, changes to the ABI, should be done by submitting the patch
(gated on the NEXT_ABI config option), along with the deprecation notice, at the
same time.  That way the ABI view is complete and consistent.  And if you do the
above bits with the cloned version map and LIBABIVER bump, its also consistent
between shared and static libraries.

> That's why, it should not be enabled to deploy shared libraries, though it can
> be used for tests and development.
> As static libraries are almost never packaged, they will be built and linked
> at the same time. That's why users of static libraries tend to prefer the
> newest ABI, which is the default in this case.
> 
> > 2) It only applies to the next ABI.  That is to say, it can't be a hodgepodge of
> > the next ABI and the one after that, and the one after that, or it won't provide
> > an appropriate preview for anyone.
> 
> If you mean the next ABI must be promoted as the standard ABI in the next release,
> yes: ifdefs will be cleaned when starting a new release.
> Thanks, I learnt the english word hodgepodge :)
> 
Je-mexcuse, une meli-melo? :)

I mean't what you indicate yes, and in addition to that, I just wanted to
clarify that this option could strictly _only_ apply to the very next ABI.  That
is to say, someone can't use this without also posting an ABI deprecation
notice, or we would find ourselves in a situation where something would only be
available in NEXT_ABI for more than one release, which would be unacceptable.
But I think we're saying the same thing.

>