Re: [PATCH 05/22] mm: export alloc_pages_vma

[Dan Williams]

On Tue, Jun 25, 2019 at 8:01 AM Michal Hocko <mhocko@xxxxxxxxxx> wrote:
>
> On Tue 25-06-19 09:23:17, Christoph Hellwig wrote:
> > On Mon, Jun 24, 2019 at 11:24:48AM -0700, Dan Williams wrote:
> > > I asked for this simply because it was not exported historically. In
> > > general I want to establish explicit export-type criteria so the
> > > community can spend less time debating when to use EXPORT_SYMBOL_GPL
> > > [1].
> > >
> > > The thought in this instance is that it is not historically exported
> > > to modules and it is safer from a maintenance perspective to start
> > > with GPL-only for new symbols in case we don't want to maintain that
> > > interface long-term for out-of-tree modules.
> > >
> > > Yes, we always reserve the right to remove / change interfaces
> > > regardless of the export type, but history has shown that external
> > > pressure to keep an interface stable (contrary to
> > > Documentation/process/stable-api-nonsense.rst) tends to be less for
> > > GPL-only exports.
> >
> > Fully agreed.  In the end the decision is with the MM maintainers,
> > though, although I'd prefer to keep it as in this series.
>
> I am sorry but I am not really convinced by the above reasoning wrt. to
> the allocator API and it has been a subject of many changes over time. I
> do not remember a single case where we would be bending the allocator
> API because of external modules and I am pretty sure we will push back
> heavily if that was the case in the future.

This seems to say that you have no direct experience of dealing with
changing symbols that that a prominent out-of-tree module needs? GPU
drivers and the core-mm are on a path to increase their cooperation on
memory management mechanisms over time, and symbol export changes for
out-of-tree GPU drivers have been a significant source of friction in
the past.

> So in this particular case I would go with consistency and export the
> same way we do with other functions. Also we do not want people to
> reinvent this API and screw that like we have seen in other cases when
> external modules try reimplement core functionality themselves.

Consistency is a weak argument when the cost to the upstream community
is negligible. If the same functionality was available via another /
already exported interface *that* would be an argument to maintain the
existing export policy. "Consistency" in and of itself is not a
precedent we can use more widely in default export-type decisions.

Effectively I'm arguing EXPORT_SYMBOL_GPL by default with a later
decision to drop the _GPL. Similar to how we are careful to mark sysfs
interfaces in Documentation/ABI/ that we are not fully committed to
maintaining over time, or are otherwise so new that there is not yet a
good read on whether they can be made permanent.