25/10/2017 12:05, Bruce Richardson: > On Wed, Oct 25, 2017 at 12:01:26PM +0200, Thomas Monjalon wrote: > > 25/10/2017 11:50, Richardson, Bruce: > > > From: Thomas Monjalon > > > > > On Tuesday 24 October 2017 01:59 AM, Thomas Monjalon wrote: > > > > > > 20/10/2017 14:31, Santosh Shukla: > > > > > >> Removed dpdk iova aware ABI deprecation notice, and updated ABI > > > > > >> change details in release_17.11.rst. > > > > > >> > > > > > >> Signed-off-by: Santosh Shukla <santosh.shu...@caviumnetworks.com> > > > > > >> Acked-by: John McNamara <john.mcnam...@intel.com> > > > > > >> --- > > > > > >> --- a/doc/guides/rel_notes/deprecation.rst > > > > > >> +++ b/doc/guides/rel_notes/deprecation.rst > > > > > >> -* eal: An ABI change is planned for 17.11 to make DPDK aware of > > > > > >> IOVA address > > > > > >> - translation scheme. > > > > > >> - Reference to phys address in EAL data-structure or functions may > > > > > >> change to > > > > > >> - IOVA address or more appropriate name. > > > > > >> - The change will be only for the name. > > > > > >> - Functional aspects of the API or data-structure will remain > > > > > >> same. > > > > > > Sorry, this series cannot be applied as is because it is breaking > > > > > > more than EAL API. The API of mbuf and mempool are also changed. > > > > > > We need to choose one of these three options: > > > > > > 1/ accept to break all API in 17.11 > > > > > > 2/ postpone the whole series to 18.02 > > > > > > > > > > Theme of series is to make dpdk iova aware so I would prefer option 1) > > > > or 2). > > > > > However I have no strong opinion on this topic. > > > > > Lets get more opinion from others about option 1/2/3. > > > > > > > > > > > 3/ rename only EAL API in 17.11 and postpone mbuf/mempool > > > > > > > > After discussing with Olivier it appeared there is a fourth solution. > > > > We should not break any API (EAL, mbuf, mempool). > > > > > > > > I would like to merge these changes in RC2, but keeping compatibility > > > > with > > > > old names: > > > > - When you rename a function or a type, you can define a macro for the > > > > old > > > > name, alias the new name. > > > > > > Note: using a macro doesn't prevent the ABI being broken if you rename a > > > public function. You'll need to use function versioning too. > > > > True > > We can use an inline function to avoid ABI breakage. > > Nope, inline function won't work either, since that ends up the same as > the macro and compiled into the end app, not the library ABI. You > need a public non-inline wrapper function to keep ABI, or else function > renaming via symbol versioning/mapping.
Ah ah ah, I'm writing before thinking :) Yes, the function must not be inlined. And generally speaking it is not an issue, even for performance critical functions. Adding one more function call in the path is not a bad thing for deprecated functions. I've seen another project (don't remember which one) adding a sleep() in deprecated functions and increasing the sleep time at each new release :)
