Re: linux-next: build failure after merge of the nvdimm tree

From: Dan Williams
Date: Mon Apr 09 2018 - 14:14:20 EST


On Mon, Apr 9, 2018 at 10:11 AM, Dan Williams <dan.j.williams@xxxxxxxxx> wrote:
> On Mon, Apr 9, 2018 at 12:34 AM, Oliver <oohall@xxxxxxxxx> wrote:
>> On Mon, Apr 9, 2018 at 1:38 PM, Oliver <oohall@xxxxxxxxx> wrote:
>>> On Mon, Apr 9, 2018 at 1:16 PM, Stephen Rothwell <sfr@xxxxxxxxxxxxxxxx> wrote:
>>>> Hi Dan,
>>>>
>>>> After merging the nvdimm tree, today's linux-next build (x86_64
>>>> allmodconfig) failed like this:
>>>>
>>>> ERROR: "of_node_to_nid" [drivers/nvdimm/of_pmem.ko] undefined!
>>>>
>>>> Caused by commit
>>>>
>>>> 717197608952 ("libnvdimm: Add device-tree based driver")
>>>>
>>>> X86 seems to not have a version of of_node_to_nid() even though CONFIG_OF
>>>> and CONFIG_NUMA are both 'y' in this build.
>>>
>>> It's a side effect of a driver selecting CONFIG_OF accidently. There's
>>> a patch to fix this in the drm-misc-next tree:
>>>
>>> https://lkml.org/lkml/2018/4/3/17
>>>
>>>> I have used the nvdimm tree from next-20180406 for today.
>>>
>>> That works too.
>>
>> Hmm, on closer inspection this is actually a bug in the of driver. The
>> patch above fixes the specific problem that we hit on ia64 due to
>> CONFIG_OF being selected by accident, but the underlying issue will
>> affect any platform that doesn't provide an implementation of
>> of_node_to_nid().
>>
>> The fundamental problem is that the various actual implementations of
>> of_node_to_nid (ppc, sparc and the generic one in of_numa.c) export
>> the symbol. While the fallback implementation does not because it is
>> defined as a weak symbol. As a result we'll get this build failure iff
>> there's a call to of_node_to_nid() in a module.
>>
>> The one-line fix is just to delete the call to of_node_to_nid() in
>> of_pmem.c. I have a patch that adds a Kconfig options and removes the
>> weak symbol games. That needs Acks from the ppc, sparc and DT
>> maintainers though so it'll take longer to organise.
>
> How about just marking OF_PMEM as bool for the time being until all of
> these other reworks land upstream?

I went ahead and just pushed this workaround for now:

https://git.kernel.org/pub/scm/linux/kernel/git/nvdimm/nvdimm.git/commit/?h=libnvdimm-for-next&id=291717b6fbdb175da88ae2144fc58d63a490128d