Re: [PATCH] ubsan: Implement __ubsan_handle_alignment_assumption

[Nathan Chancellor <natechancellor@gmail.com>]

On Tue, Jan 12, 2021 at 01:15:42PM -0800, Nick Desaulniers wrote:
> On Tue, Jan 12, 2021 at 12:55 PM Nathan Chancellor
> <natechancellor@gmail.com> wrote:
> >
> > When building ARCH=mips 32r2el_defconfig with CONFIG_UBSAN_ALIGNMENT:
> >
> > ld.lld: error: undefined symbol: __ubsan_handle_alignment_assumption
> > >>> referenced by slab.h:557 (include/linux/slab.h:557)
> > >>>               main.o:(do_initcalls) in archive init/built-in.a
> > >>> referenced by slab.h:448 (include/linux/slab.h:448)
> > >>>               do_mounts_rd.o:(rd_load_image) in archive init/built-in.a
> > >>> referenced by slab.h:448 (include/linux/slab.h:448)
> > >>>               do_mounts_rd.o:(identify_ramdisk_image) in archive init/built-in.a
> > >>> referenced 1579 more times
> >
> > Implement this for the kernel based on LLVM's
> > handleAlignmentAssumptionImpl because the kernel is not linked against
> > the compiler runtime.
> >
> > Link: https://github.com/ClangBuiltLinux/linux/issues/1245
> > Link: https://github.com/llvm/llvm-project/blob/llvmorg-11.0.1/compiler-rt/lib/ubsan/ubsan_handlers.cpp#L151-L190
> > Signed-off-by: Nathan Chancellor <natechancellor@gmail.com>
> > ---
> >  lib/ubsan.c | 28 ++++++++++++++++++++++++++++
> >  lib/ubsan.h |  6 ++++++
> >  2 files changed, 34 insertions(+)
> >
> > diff --git a/lib/ubsan.c b/lib/ubsan.c
> > index 3e3352f3d0da..a1e6cc9993f8 100644
> > --- a/lib/ubsan.c
> > +++ b/lib/ubsan.c
> > @@ -427,3 +427,31 @@ void __ubsan_handle_load_invalid_value(void *_data, void *val)
> >         ubsan_epilogue();
> >  }
> >  EXPORT_SYMBOL(__ubsan_handle_load_invalid_value);
> > +
> > +void __ubsan_handle_alignment_assumption(void *_data, unsigned long ptr,
> > +                                        unsigned long align,
> > +                                        unsigned long offset)
> > +{
> > +       struct alignment_assumption_data *data = _data;
> > +       unsigned long real_ptr;
> > +
> > +       if (suppress_report(&data->location))
> > +               return;
> > +
> > +       ubsan_prologue(&data->location, "alignment-assumption");
> > +
> > +       if (offset)
> > +               pr_err("assumption of %lu byte alignment (with offset of %lu byte) for pointer of type %s failed",
> > +                      align, offset, data->type->type_name);
> > +       else
> > +               pr_err("assumption of %lu byte alignment for pointer of type %s failed",
> > +                      align, data->type->type_name);
> > +
> > +       real_ptr = ptr - offset;
> > +       pr_err("%saddress is %lu aligned, misalignment offset is %lu bytes",
> > +              offset ? "offset " : "", BIT(ffs(real_ptr)),
> 
> if real_ptr is an unsigned long, do we want to use `__ffs(real_ptr) +
> 1` here rather than ffs which takes an int?  It seems the kernel is
> missing a definition of ffsl. :(

Why the + 1? I think if we use __ffs (which it seems like we should), I
think that needs to become

BIT(real_ptr ? __ffs(real_ptr) : 0)

I have made that change locally and will send it for v2 in a day or so
to give Kees some time to check it out.

Thanks for the review!
Nathan