Re: [PATCH] elf: Support DT_RELR relative relocation format [BZ #27924]

["Fāng-ruì Sòng" <maskray@google.com>]

On Tue, Oct 12, 2021 at 7:10 AM H.J. Lu <hjl.tools@gmail.com> wrote:
>
> On Tue, Oct 12, 2021 at 1:18 AM Jan Beulich via Libc-alpha
> <libc-alpha@sourceware.org> wrote:
> >
> > On 11.10.2021 20:43, Fāng-ruì Sòng wrote:
> > > On Mon, Oct 11, 2021 at 12:48 AM Jan Beulich <jbeulich@suse.com> wrote:
> > >>
> > >> On 08.10.2021 08:57, Fangrui Song via Binutils wrote:
> > >>> --- a/elf/dynamic-link.h
> > >>> +++ b/elf/dynamic-link.h
> > >>> @@ -192,6 +192,33 @@ elf_machine_lazy_rel (struct link_map *map, struct r_scope_elem *scope[],
> > >>>  #  define ELF_DYNAMIC_DO_RELA(map, scope, lazy, skip_ifunc) /* Nothing to do.  */
> > >>>  # endif
> > >>>
> > >>> +# define ELF_DYNAMIC_DO_RELR(map)                                          \
> > >>> +  do {                                                                             \
> > >>> +    ElfW(Addr) l_addr = (map)->l_addr, base = 0, start;                            \
> > >>> +    const ElfW(Relr) *r = 0, *end = 0;                                             \
> > >>> +    if (!(map)->l_info[DT_RELR])                                           \
> > >>> +      break;                                                               \
> > >>> +    start = D_PTR((map), l_info[DT_RELR]);                                 \
> > >>> +    r = (const ElfW(Relr) *)start;                                         \
> > >>> +    end = (const ElfW(Relr) *)(start + (map)->l_info[DT_RELRSZ]->d_un.d_val); \
> > >>> +    for (; r < end; ++r) {                                                 \
> > >>> +      ElfW(Relr) entry = *r;                                               \
> > >>> +      if ((entry & 1) == 0) {                                                      \
> > >>> +     *((ElfW(Addr) *)(l_addr + entry)) += l_addr;                          \
> > >>> +     base = entry + sizeof(ElfW(Addr));                                    \
> > >>> +     continue;                                                             \
> > >>> +      }                                                                            \
> > >>> +      ElfW(Addr) offset = base;                                                    \
> > >>> +      do {                                                                 \
> > >>> +     entry >>= 1;                                                          \
> > >>> +     if ((entry & 1) != 0)                                                 \
> > >>> +       *((ElfW(Addr) *)(l_addr + offset)) += l_addr;                       \
> > >>> +     offset += sizeof(ElfW(Addr));                                         \
> > >>> +      } while (entry != 0);                                                \
> > >>> +      base += (8 * sizeof(ElfW(Relr)) - 1) * sizeof(ElfW(Addr));           \
> > >>
> > >> While in line with the proposed spec additions I'm afraid the uses of
> > >> ElfW(Addr) here aren't universally correct: You assume that ELF
> > >> container type (size) expresses an aspect of the ABI. While this is
> > >> indeed the case for several arch-es, I think this has been a mistake.
> > >> IA-64, while meanwhile mostly dead, is (was) an example where 64-bit
> > >> code can validly live in a 32-bit ELF container (at least as far as
> > >> the psABI is concerned; I have no idea whether glibc actually
> > >> followed the spec). There's a separate ELF header flag indicating the
> > >> ABI, and hence the size of a pointer.
> > >
> > > Thanks for chiming in.
> > >
> > > As of ia64 buildability, it works for me:
> > >
> > > scripts/build-many-glibcs.py /tmp/glibc-many compilers ia64-linux-gnu
> > > mkdir -p out/ia64; cd out/ia64
> > > ../../configure --prefix=/tmp/glibc/ia64 --host=ia64-linux-gnu
> > > CC=/tmp/glibc-many/install/compilers/ia64-linux-gnu/bin/ia64-glibc-linux-gnu-gcc
> > > CXX=/tmp/glibc-many/install/compilers/ia64-linux-gnu/bin/ia64-glibc-linux-gnu-g++
> > > make -j 50
> >
> > I didn't suggest the build would fail. What I said is that I don't
> > think the code is correct there.
> >
> > > As of the actual functionality, ugh, I cannot find ia64 in my Debian
> > > testing's qemu-user-static package:( So I cannot test.
> > >
> > > That said, gold and LLD don't support ia64.
> > > If we have a concern that ia64 may not work, the GNU ld maintainers
> > > can simply not add ia64 support:)
> >
> > But you realize that I took ia64 only as example, as that's where
> > I know ABI (pointer size) and ELF container size aren't connected.
> > As per my looking at merely EF_MIPS_* in context of reading
> > Joseph's reply, it might be that MIPS is another such example. But
> > I lack sufficient knowledge of MIPS ...
> >
>
> The new code should be tested and verified on all supported
> targets.  That is another reason to implement this in binutils
> ld first.

--pack-dyn-relocs=relr is well tested on arm, aarch64, and x86, and
works on popular arches like ppc64 as well.
For mips, it is no harm to keep the DT_RELR code path. Its
elf_machine_rel_relative is empty and it has no relative relocation
anyway.
I wish that our reasonable design decisions are not restricted by the
few retrocomputing architectures, especially when the concern is still
at the theoretical stage.
Well, if DT_RELR has issue, don't use it; nobody forces them to use DT_RELR.
After all, --pack-dyn-relocs=relr is opt in.

On the other side, DT_RELR is a really great addition to save code
size for non-exotic architectures. Distributions will benefit a lot.
LLD users can immediately benefit.
gold users can considerably leverage the feature relatively easily
since a patch is available.