Re: [dpdk-dev] [PATCH] devtools: skip the symbol check when map file under drivers

[Neil Horman <nhorman@tuxdriver.com>]

On Wed, May 22, 2019 at 11:54:13AM +0000, Jerin Jacob Kollanukkaran wrote:
> > -----Original Message-----
> > From: Bruce Richardson <bruce.richardson@intel.com>
> > Sent: Wednesday, May 22, 2019 4:21 PM
> > To: Jerin Jacob Kollanukkaran <jerinj@marvell.com>
> > Cc: Neil Horman <nhorman@tuxdriver.com>; dev@dpdk.org;
> > thomas@monjalon.net; stable@dpdk.org
> > Subject: Re: [dpdk-dev] [EXT] Re: [PATCH] devtools: skip the symbol check
> > when map file under drivers
> > 
> > > > Sorry, but I'm not ok with this, because many of our DPDK PMDs have
> > > > functions that get exported which are meant to be called by
> > > > applications directly.  The
> > >
> > > OK. Just to update my knowledge, Should those API needs to go through
> > > ABI/API depreciation process?
> > >
> > > Actually, I am concerned about the APIs, which is called between
> > > drviers not the application. For example,
> > > drivers/common/dpaax/rte_common_dpaax_version.map
> > >
> > > it is not interface to application rather it is for intra driver case.
> > > I think, I can change my logic to Skip the symbols which NOT starting with
> > rte_.
> > > Agree?
> > >
> > > Context:
> > > I am adding a new driver/common/octeontx2 directory and it has some
> > > API which Needs to shared between drivers not to the application. For
> > > me, it does not make sense to go through any ABI process in such case.
> > >
> > >
> > Maybe not, but other drivers will have APIs designed for apps to call directly -
> > some NIC drivers have them, and I suspect that rawdev drivers will need
> > them a lot. Therefore, it's best to have the drivers directory scanned by our
> > tooling.
> 
> Agreed. But all of those API  which called directly called from application
> is starts with rte_ symbol. How about skipping the symbols which is NOT start with rte_*
> example:
> drivers/common/octeontx/rte_common_octeontx_version.map
> drivers/common/dpaax/rte_common_dpaax_version.map
> 

No, that won't work.  If you export a function, it doesn't matter if its named
rte_* or not.  Its accessible from any library/application that cares to call
it, and so you have a responsibility to keep it stable for those users.

Currently the way we have around that is the use of the __rte_experimental tag.
Adding that tag to an exported function marks it as being unstable, and while
you can use it, it will generate a build time warning about its use, unless you
define ALLOW_EXPERIMENTAL_API.  You could use that, understanding that in-tree
drivers could use it safely, as you should always be keeping the API in sync
with its users, but thats not quite what you want I don't think.

Another solution (allbeit a slightly risky one), would be to bifurcate your
header files into a public and private version, with the private version
prototyping your driver-only functions properly, and the public version aliasing
them such that they generate a build time error indicating those functions
aren't available for public use (you can use the gcc static_assert macro I
believe).  Users could circumvent it by pulling the private header out of the
build, or just prototyping the functions themselves, but at that point a user is
asking for trouble anyway

Neil