Re: [PATCH] sparc64: Define p4d_page() instead of stubbing it to NULL

[Rosen Penev]

On Tue, Oct 6, 2026 at 1:31 AM Nathan Chancellor <nathan@xxxxxxxxxx> wrote:
>
> On Tue, Oct 06, 2026 at 09:59:19AM +0200, Kjetil Oftedal wrote:
> > On Mon, 5 Oct 2026 at 22:41, Rosen Penev <rosenp@xxxxxxxxx> wrote:
> > >
> > > A sparc64 allmodconfig build with LLVM fails in vmalloc_to_page():
> > >
> > >   mm/vmalloc.c: error: performing pointer arithmetic on a null pointer
> > >   has undefined behavior [-Werror,-Wnull-pointer-arithmetic]
> > >
> > > sparc64 defines p4d_page() as NULL, on the assumption that it is only
> > > used by stubbed out hugetlb gup code. Generic code now uses it in
> > > vmalloc_to_page(), ptdump and KASAN. In vmalloc_to_page() the NULL is
> > > used in pointer arithmetic, which clang diagnoses even though the
> > > branch is unreachable on sparc64.
> > >
> >
> > Has this been reported as a bug to the clang project?
> > As it throwing an error on a rather obviously unreachable branch.
>
> Technically the error part is self-inflicted on the part of the kernel
> since allmodconfig enables CONFIG_WERROR. That said, while it might be
> nice for the compiler not to warn in any case of dead code, that's just
> not how it works for either GCC or clang (none of these should be
> visible if that were the case):
LLM suggests as an alternative:

/* sparc64 has no p4d leaf mappings, so this is never called */
static inline struct page *p4d_page(p4d_t p4d)
{
      return NULL;
}

Not sure how to proceed.
>   https://godbolt.org/z/K3nfW4n5f
>
> --
> Cheers,
> Nathan