Re: OpenAFS module libafs.ko uses GPL-only symbol '__put_devmap_managed_page'

[Dan Williams <dan.j.williams@intel.com>]

On Fri, Jun 15, 2018 at 9:43 PM, John Hubbard <jhubbard@nvidia.com> wrote:
> On 06/13/2018 12:51 PM, Dan Williams wrote:
>> [ adding Andrew, Christoph, and linux-mm ]
>>
>> On Wed, Jun 13, 2018 at 12:33 PM, Joe Gorse <jhgorse@gmail.com> wrote:
>>> Greetings,
>>>
>>> Please CC answers & comments to this email. Thanks! =)
>>>
>>> Our build is breaking as of
>>> commit e7638488434415aa478e78435cac8f0365737638
>>> Author: Dan Williams <dan.j.williams@intel.com>
>>> Date: Wed May 16 11:46:08 2018 -0700
>>>
>>> mm: introduce MEMORY_DEVICE_FS_DAX and CONFIG_DEV_PAGEMAP_OPS
>>> ... snip ...
>>> https://patchwork.kernel.org/patch/10412459/
>>>
>>> We do not directly use the GPL-only symbol '__put_devmap_managed_page'. It
>>> appears to be picked up from static-inlining in put_page(), which we need.
>>>
>>> How shall we proceed? Would it be reasonable to request the change of the
>>> GPL-only exports for this commit?
>>>
>>> Cheers,
>>> Joe Gorse
>>>
>>> P.S. The build failure, for the morbidly curious:
>>>> FATAL: modpost: GPL-incompatible module libafs.ko uses GPL-only symbol
>>>> '__put_devmap_managed_page'
>>>> scripts/Makefile.modpost:92: recipe for target '__modpost' failed
>>>> make[6]: *** [__modpost] Error 1
>>
>> I think the right answer here is to make __put_devmap_managed_page()
>> EXPORT_SYMBOL(), since features like devm_memremap_pages() want to
>> change the behavior of all users of put_page(). It again holds that
>> devm_memremap_pages() needs to become EXPORT_SYMBOL_GPL() because it,
>> not put_page(), is the interface that is leaking control of core
>> kernel state/infrastructure to its users.
>>
>
> Hi Dan and all,
>
> It looks like put_page() also picks up one more GPL symbol:
> devmap_managed_key.
>
> put_page
>     put_devmap_managed_page
>         devmap_managed_key
>
>     __put_devmap_managed_page
>
>
> So if the goal is to restore put_page() to be effectively EXPORT_SYMBOL
> again, then I think there would also need to be either a non-inlined
> wrapper for devmap_managed_key (awkward for a static key), or else make
> it EXPORT_SYMBOL, or maybe something else that's less obvious to me at the
> moment.

Right, certainly flipping the key is a kernel internal detail since it
is giving the dev_pagemap owner purview over all kernel page events,
but put_page() users are silent consumers. And you're right there's
currently no good way I see to export the 'producer' and 'consumer'
side of the key with different export types.