Re: linux-next: build warning after merge of the akpm tree

[Dmitry Vyukov]

On Mon, Dec 7, 2020 at 1:08 PM Dmitry Vyukov <dvyukov@xxxxxxxxxx> wrote:
> > > Hi all,
> > >
> > > After merging the akpm tree, today's linux-next build (powerpc
> > > allyesconfig) produced warnings like this:
> > >
> > > kernel/kcov.c:296:14: warning: conflicting types for built-in function '__sanitizer_cov_trace_switch'; expected 'void(long unsigned int,  void *)' [-Wbuiltin-declaration-mismatch]
> > >   296 | void notrace __sanitizer_cov_trace_switch(u64 val, u64 *cases)
> > >       |              ^~~~~~~~~~~~~~~~~~~~~~~~~~~~
> >
> > Odd.  clang wants that signature, according to
> > https://clang.llvm.org/docs/SanitizerCoverage.html.  But gcc seems to
> > want a different signature.  Beats me - best I can do is to cc various
> > likely culprits ;)
> >
> > Which gcc version?  Did you recently update gcc?
> >
> > > ld: warning: orphan section `.data..Lubsan_data177' from `arch/powerpc/oprofile/op_model_pa6t.o' being placed in section `.data..Lubsan_data177'
> > >
> > > (lots of these latter ones)
> > >
> > > I don't know what produced these, but it is in the akpm-current or
> > > akpm trees.
>
> I can reproduce this in x86_64 build as well but only if I enable
> UBSAN as well. There were some recent UBSAN changes by Kees, so maybe
> that's what affected the warning.
> Though, the warning itself looks legit and unrelated to UBSAN. In
> fact, if the compiler expects long and we accept u64, it may be broken
> on 32-bit arches...

No, I think it works, the argument should be uint64.

I think both gcc and clang signatures are correct and both want
uint64_t. The question is just how uint64_t is defined :) The old
printf joke that one can't write portable format specifier for
uint64_t.

What I know so far:
clang 11 does not produce this warning even with obviously wrong
signatures (e.g. short).
I wasn't able to trigger it with gcc on 32-bits at all. KCOV is not
supported on i386 and on arm I got no warnings even with obviously
wrong signatures (e.g. short).
Using "(unsigned long val, void *cases)" fixes the warning on x86_64.

I am still puzzled why gcc considers this as a builtin because we
don't enable -fsanitizer-coverage on this file. I am also puzzled how
UBSAN affects things.

We could change the signature to long, but it feels wrong/dangerous
because the variable should really be 64-bits (long is broken on
32-bits).
Or we could introduce a typedef that is long on 64-bits and 'long
long' on 32-bits.