On Tue, Mar 10, 2020 at 04:07:14PM +0100, Oleg Nesterov wrote:
>
> Annoyingly, this triggers a lot of sparse_error's in pre-process.c:collect_arg().
> And just in case, of course this is not specific to dissect/sindex, ./sparse or
> anything else will equally complain.
>
> For example,
>
> 1011 static inline bool page_expected_state(struct page *page,
> 1012 unsigned long check_flags)
> 1013 {
> 1014 if (unlikely(atomic_read(&page->_mapcount) != -1))
> 1015 return false;
> 1016
> 1017 if (unlikely((unsigned long)page->mapping |
> 1018 page_ref_count(page) |
> 1019 #ifdef CONFIG_MEMCG
> 1020 (unsigned long)page->mem_cgroup |
> 1021 #endif
> 1022 (page->flags & check_flags)))
> 1023 return false;
> 1024
> 1025 return true;
> 1026 }
>
> leads to
>
> mm/page_alloc.c:1019:1: error: directive in macro's argument list
> mm/page_alloc.c:1021:1: error: directive in macro's argument list
>
> and it is not immediately clear why. Yes, because "unlikely" is a macro.
>
> Can't we simply remove this sparse_error() ? "#if" inside the macro's args
> is widely used in kernel, gcc doesn't complain, afaics pre-process.c handles
> this case correctly.
I'm quite reluctant to simply suppress it.
My (contradictory) point of view is that it is because it's not
immediately clear there is a problem that the warning is needed
but, OTOH, people and the Standard, want to use macros transparently
so a macro wrapping a function call should behave just like directly
calling the function. And yes both Sparse and GCC seem to be able
to handle this, so it's maybe only a restriction for more primtive
preprocessors. I dunno.
Some arguments/justifications for the arning can be found at:
https://www.spinics.net/lists/kernel/msg1636994.html
Anyway, only a warning should be issued (I'll send a patch for this).
I also wouldn't mind to add a new warning flag to suppress it,
something like -Wno-directive-within-macro.
-- Luc
