Re: [PATCH] dump_stack: convert generic dump_stack into a weak symbol

[Arnd Bergmann]

On Wed, Mar 7, 2018 at 3:21 AM, Sergey Senozhatsky
<sergey.senozhatsky.work@xxxxxxxxx> wrote:
> Hello Arnd,
>
> On (03/06/18 14:27), Arnd Bergmann wrote:
> [..]
>> As we are now removing blackfin, based on the latest discussion, this
>> part should no longer be necessary.
>
> When is this going to happen? 4.17?

Originally I planned to wait a few more releases, but the last maintainer
has commented that he will now send a patch for immediate removal,
so 4.17 is almost certain at this point.

> [..]
>> nds32 currently only exists in linux-next, not in the mainline kernel.
>> If it's the only architecture that does something different from everyone
>> else, I think we should change nds32.
>>
>> I looked at the nds32 show_stack() implementation now and it
>> seems to me that it is completely unnecessary, as the implementation
>> from lib/dump_stack.c does basically the same thing (by calling
>> show_stack(NULL, NULL)).
>
> Interesting point. I'll leave it to nds32 maintainers.
> OTOH blackfin is still in linux-next, so I assume we need
> that __weak dump_stack() for the time being.

I did the review of all the nds32 patches, and would have commented
on this if I had noticed it earlier. I see no reason not to change it,
and would suggest that you continue under the assumption that
nds32 is going to be fixed, leaving it up to Greentime to add a fix
to his tree before he sends the pull request.

>> > +asmlinkage __weak __visible void dump_stack(void)
>> >  {
>> >         __dump_stack();
>> >  }
>>
>> Weak symbols are generally discouraged in the kernel. We have
>> them in a couple of places, but I find them rather confusing as they
>> make it harder to figure out what is actually going on.
>
> Honestly, I kind of find __weak less confusing than EXPORT_SYMBOL(dump_stack)
> in 3 different places. __weak hints that the symbol likely will be overridden
> somewhere, while EXPORT_SYMBOL() does not (at least not to me). Dunno.

It's not either/or, they are both wrong ;-)

The EXPORT_SYMBOL() is not the thing that makes it work. The duplicate
declaration today only works because lib/dump_stack.o is listed as lib-y
in the Makefile instead of obj-y. On blackfin and nds32, this causes the entire
file to just not be included in the final vmlinux, because there are no
references to it.

With your patch, I would actually expect the lib/dump_stack.o file
to still not be picked up, so now you have a missing EXPORT_SYMBOL()
on the two unusual architectures until the point when you add another
(referenced) symbol to it.

       Arnd