[RFC PATCH 4/8] mm: Separate fault info out of 'struct vm_fault'

[Nick Desaulniers]

On Thu, Jan 14, 2021 at 11:41 AM Will Deacon <will at kernel.org> wrote:
>
> On Thu, Jan 14, 2021 at 11:09:01AM -0800, Linus Torvalds wrote:
> > On Thu, Jan 14, 2021 at 11:00 AM Will Deacon <will at kernel.org> wrote:
> > >
> > > I tried that initially, but I found that I had to make all of the
> > > members const to get it to work, at which point the anonymous struct
> > > wasn't really adding anything. Did I just botch the syntax?
> >
> > I'm not sure what you tried. But this stupid test-case sure works for me:
> >
> >     struct hello {
> >         const struct {
> >                 unsigned long address;
> >         };
> >         unsigned int flags;
> >     };
> >
> >     extern int fn(struct hello *);
> >
> >     int test(void)
> >     {
> >         struct hello a = {
> >                 .address = 1,
> >         };
> >         a.flags = 0;
> >         return fn(&a);
> >     }
> >
> > and because "address" is in that unnamed constant struct, you can only
> > set it within that initializer, and cannot do
> >
> >         a.address = 0;
> >
> > without an error (the way you _can_ do "a.flags = 0").
> >
> > I don't see naming the struct making a difference - apart from forcing
> > that big rename patch, of course.
> >
> > But maybe we're talking about different issues?
>
> Urgh...
>
> We _are_ both on the same page, and your reply above had me thinking I've
> lost the plot, so I went back to the start. Check out v5.11-rc3 and apply
> this patch:
>
>
> diff --git a/include/linux/mm.h b/include/linux/mm.h
> index ecdf8a8cd6ae..1eb950865450 100644
> --- a/include/linux/mm.h
> +++ b/include/linux/mm.h
> @@ -514,11 +514,14 @@ static inline bool fault_flag_allow_retry_first(unsigned int flags)
>   * pgoff should be used in favour of virtual_address, if possible.
>   */
>  struct vm_fault {
> -       struct vm_area_struct *vma;     /* Target VMA */
> +       const struct {
> +               struct vm_area_struct *vma;     /* Target VMA */
> +               gfp_t gfp_mask;                 /* gfp mask to be used for allocations */
> +               pgoff_t pgoff;                  /* Logical page offset based on vma */
> +               unsigned long address;          /* Faulting virtual address */
> +       };
> +
>         unsigned int flags;             /* FAULT_FLAG_xxx flags */
> -       gfp_t gfp_mask;                 /* gfp mask to be used for allocations */
> -       pgoff_t pgoff;                  /* Logical page offset based on vma */
> -       unsigned long address;          /* Faulting virtual address */
>         pmd_t *pmd;                     /* Pointer to pmd entry matching
>                                          * the 'address' */
>         pud_t *pud;                     /* Pointer to pud entry matching
>
>
> Sure enough, an arm64 defconfig builds perfectly alright with that change,
> but it really shouldn't. I'm using clang 11.0.5, so I had another go with
> GCC 9.2.1 and bang:
>
> mm/filemap.c: In function ‘filemap_map_pages’:
> mm/filemap.c:2963:16: error: assignment of member ‘address’ in read-only object
>  2963 |   vmf->address += (xas.xa_index - last_pgoff) << PAGE_SHIFT;
>       |                ^~
> make[1]: *** [scripts/Makefile.build:279: mm/filemap.o] Error 1
> make[1]: *** Waiting for unfinished jobs....
> make: *** [Makefile:1805: mm] Error 2
> make: *** Waiting for unfinished jobs....
>
> Nick -- any clue what's happening here? We would like that const anonymous
> struct to behave like a const struct member, as the alternative (naming the
> thing) results in a lot of refactoring churn.

Weird, looks like a bug to me in Clang, filed
https://bugs.llvm.org/show_bug.cgi?id=48755.
-- 
Thanks,
~Nick Desaulniers