On Thu, Sep 16, 2021 at 9:01 PM Florian Weimer <fweimer@redhat.com> wrote:
>
> * Florian Weimer:
>
> > * Siddhesh Poyarekar:
> >
> >> On 9/16/21 11:33 PM, Florian Weimer wrote:
> >>> * H. J. Lu:
> >>>
> >>>> No.  Relocation of vDSO dynamic section is done by elf_get_dynamic_info.
> >>>> Here is a patch to remove the hack for vDSO.
> >>> I have concerns that this does not actually work on MIPS.  What
> >>> happens
> >>> if the object has a single RWX LOAD segment?  What kind of flags will
> >>> the link editor set on the DYNAMIC segment in this case?  RWX or R, RX?
> >>> I suspect it will be RWX, which means that we actually need the special
> >>> case.
> >>
> >> The DYNAMIC segment flags will be based on .dynamic, which is
> >> read-only for MIPS.
> >
> > Right, as far as I can tell BFD ld does not copy the W bit to the
> > DYNAMIC segment even if it is located inside an RW LOAD segment.  So
> > unless there are old binaries floating around which were linked
> > differently, the change should be correct for MIPS, too (and likewise
> > for RISC-V, although I did not check that).
>
> *sigh*
>
> I checked riscv64-linux-gnu-rv64imafdc-lp64d now, and binutils BFD ld is
> actually inconsistent with glibc there.  There, DYNAMIC is RW:
>
>   LOAD           0x0003f0 0x00000000000003f0 0x00000000000003f0 0x0001c0 0x0001c0 RW  0x1000
>   DYNAMIC        0x000440 0x0000000000000440 0x0000000000000440 0x000170 0x000170 RW  0x8
>
> So libphobos and others would have to change if we adopt H.J.'s patch, I
> think.  With it, RISC-V would be just like all the other non-MIPS
> targets.
>

Glibc is wrong here.  My patch will make it correct.

Here is the updated patch with the new commit log.

-- 
H.J.