Re: [PATCH] riscv: remove DL_RO_DYN_SECTION

[Jim Wilson <jimw at sifive dot com>]

On Tue, Apr 9, 2019 at 4:42 AM Florian Weimer <fweimer@redhat.com> wrote:
> * David Abdurachmanov:

Do you or your employer have copyright assignments?  For a trivial
patch like this, a copyright assignment shouldn't be necessary, but
would be necessary if you want to contribute more patches.

> > While working on enabling D front-end (GDC) in GCC we noticed that druntime
> > was segfaulting if it is linked dynamically. This was tracked to
> > DL_RO_DYN_SECTION.

My analysis of this when it first came up...

I believe this was blindly copied from MIPS, and that no one on the
RISC-V side will know why it is there, because it was blindly copied
from MIPS.

Looking at the MIPS SVR4 ABI, I see
.dynamic This is the same as the generic ABI section of the same type, but
the MIPS-specific version does not include the SHF_WRITE at-
tribute.

I don't see any other reason in the spec for this to be read-only, but
making it read-only probably improves program security.  I don't see
an equivalent statement in the RISC-V ABI.

Looking at bfd/elfxx-mips.c _bfd_mips_elf_create_dynamic_sections(), we see code
  flags = (SEC_ALLOC | SEC_LOAD | SEC_HAS_CONTENTS | SEC_IN_MEMORY
           | SEC_LINKER_CREATED | SEC_READONLY);

  /* The psABI requires a read-only .dynamic section, but the VxWorks
     EABI doesn't.  */
  if (!htab->is_vxworks)
    {
      s = bfd_get_linker_section (abfd, ".dynamic");
      if (s != NULL)
        {
          if (! bfd_set_section_flags (abfd, s, flags))
            return FALSE;
        }
    }
that forces .dynamic to be read-only everywhere except vxworks.  I
don't see equivalent code in the RISC-V bfd port.

Looking at ld, in emulparams/, there are multiple mips scripts that
set TEXT_DYNAMIC, which forces .dynamic into the text segment instead
of the data segment.  This is then unset in the vxworks scripts.
Interestingly, hppa also sets TEXT_DYNAMIC, but the hppa support in
glibc is probably not well maintained.  Anyways, there is no
equivalent code for RISC-V.

So it appears unnecessary for the RISC-V glibc port.

If MIPS dynamic linking support is ever added to the druntime library,
then it will need the same change that you made for RISC-V, so in that
case it would not be a RISC-V specific change.  However, there aren't
a lot of people doing MIPS development work anymore, so this may never
come up.

Jim
PS Since I wrote my analysis, MIPS has been resurrected, so maybe
someone will care about that again.