Re: [PATCH] BaseTools/GenFw AARCH64: fix up GOT based relative relocations

["Ard Biesheuvel" <ard.biesheuvel@linaro.org>]

On Wed, 4 Sep 2019 at 08:37, Leif Lindholm <leif.lindholm@linaro.org> wrote:
>
> On Wed, Sep 04, 2019 at 05:01:58AM -0700, Ard Biesheuvel wrote:
> > On Wed, 4 Sep 2019 at 04:49, Leif Lindholm <leif.lindholm@linaro.org> wrote:
> > >
> > > On Tue, Sep 03, 2019 at 09:17:33PM -0700, Ard Biesheuvel wrote:
> > > > We take great care to avoid GOT based relocations in EDK2 executables,
> > > > primarily because they are pointless - we don't care about things like
> > > > the CoW footprint or relocations that target read-only sections, and so
> > > > GOT entries only bloat the binary.
> > > >
> > > > However, in some cases (e.g., when building the relocatable PrePi SEC
> > > > module in ArmVirtPkg with the CLANG38 toolchain), we may end up with
> > > > some GOT based relocations nonetheless, which break the build since
> > > > GenFw does not know how to deal with them.
> > > >
> > > > The relocations emitted in this case are ADRP/LDR instruction pairs
> > > > that are annotated as GOT based, which means that it is the linker's
> > > > job to emit the GOT entry and tag it with an appropriate dynamic
> > > > relocation that ensures that the correct absolute value is stored into
> > > > the GOT entry when the executable is loaded. This dynamic relocation
> > > > not visible to GenFw, and so populating the PE/COFF relocation section
> > > > for these entries is non-trivial.
> > > >
> > > > Since each ADRP/LDR pair refers to a single symbol that is local to the
> > > > binary (given that shared libraries are not supported), we can actually
> > > > convert the ADRP/LDR pair into an ADRP/ADD pair that produces the symbol
> > > > address directly rather than loading it from memory. This leaves the
> > > > GOT entry in the binary, but since it is now unused, it is no longer
> > > > necessary to emit a PE/COFF relocation entry for it.
> > > >
> > > > Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
> > >
> > > This is a very neat fix. My only concern is that I am not able to
> > > reproduce the issue on my local Buster with clang7 (default). Is it
> > > reproducible with clang8?
> > >
> >
> > I managed to reproduce it on Ubuntu Bionic with clang 6. It may also
> > be related to the version of ld.gold or the LLVM gold plugin.
> >
> > You should be able to test this patch for correctness by stripping the
> > no-pie/no-pic options from the GCC5 command line, and checking any
> > produced .dll with readelf -r to see whether any GOT based relocations
> > were emitted, and whether the resulting binary still runs. I will do
> > the same locally.
>
> By removing the -no-pic/-no-pie flags, I get the GCC5 profile to
> display the issue. And by adding this patch, the build issue goes
> away.
>
> The image remains bootable with -kernel.
>
> I won't claim to have double checked the arithmetic/encodings, but
> apart from that:
> Reviewed-by: Leif Lindholm <leif.lindholm@linaro.org>
>

Thanks all

Pushed as adb59b633c12..d2687f23c909


> > > > ---
> > > >  BaseTools/Source/C/GenFw/Elf64Convert.c | 28 +++++++++++++++++++-
> > > >  1 file changed, 27 insertions(+), 1 deletion(-)
> > > >
> > > > diff --git a/BaseTools/Source/C/GenFw/Elf64Convert.c b/BaseTools/Source/C/GenFw/Elf64Convert.c
> > > > index 3d6319c821e9..d574300ac4fe 100644
> > > > --- a/BaseTools/Source/C/GenFw/Elf64Convert.c
> > > > +++ b/BaseTools/Source/C/GenFw/Elf64Convert.c
> > > > @@ -1017,6 +1017,31 @@ WriteSections64 (
> > > >          } else if (mEhdr->e_machine == EM_AARCH64) {
> > > >
> > > >            switch (ELF_R_TYPE(Rel->r_info)) {
> > > > +            INT64 Offset;
> > > > +
> > > > +          case R_AARCH64_LD64_GOT_LO12_NC:
> > > > +            //
> > > > +            // Convert into an ADD instruction - see R_AARCH64_ADR_GOT_PAGE below.
> > > > +            //
> > > > +            *(UINT32 *)Targ &= 0x3ff;
> > > > +            *(UINT32 *)Targ |= 0x91000000 | ((Sym->st_value & 0xfff) << 10);
> > > > +            break;
> > > > +
> > > > +          case R_AARCH64_ADR_GOT_PAGE:
> > > > +            //
> > > > +            // This relocation points to the GOT entry that contains the absolute
> > > > +            // address of the symbol we are referring to. Since EDK2 only uses
> > > > +            // fully linked binaries, we can avoid the indirection, and simply
> > > > +            // refer to the symbol directly. This implies having to patch the
> > > > +            // subsequent LDR instruction (covered by a R_AARCH64_LD64_GOT_LO12_NC
> > > > +            // relocation) into an ADD instruction - this is handled above.
> > > > +            //
> > > > +            Offset = (Sym->st_value - (Rel->r_offset & ~0xfff)) >> 12;
> > > > +
> > > > +            *(UINT32 *)Targ &= 0x9000001f;
> > > > +            *(UINT32 *)Targ |= ((Offset & 0x1ffffc) << (5 - 2)) | ((Offset & 0x3) << 29);
> > > > +
> > > > +            /* fall through */
> > > >
> > > >            case R_AARCH64_ADR_PREL_PG_HI21:
> > > >              //
> > > > @@ -1037,7 +1062,6 @@ WriteSections64 (
> > > >                // Attempt to convert the ADRP into an ADR instruction.
> > > >                // This is only possible if the symbol is within +/- 1 MB.
> > > >                //
> > > > -              INT64 Offset;
> > > >
> > > >                // Decode the ADRP instruction
> > > >                Offset = (INT32)((*(UINT32 *)Targ & 0xffffe0) << 8);
> > > > @@ -1212,6 +1236,8 @@ WriteRelocations64 (
> > > >              case R_AARCH64_LDST32_ABS_LO12_NC:
> > > >              case R_AARCH64_LDST64_ABS_LO12_NC:
> > > >              case R_AARCH64_LDST128_ABS_LO12_NC:
> > > > +            case R_AARCH64_ADR_GOT_PAGE:
> > > > +            case R_AARCH64_LD64_GOT_LO12_NC:
> > > >                //
> > > >                // No fixups are required for relative relocations, provided that
> > > >                // the relative offsets between sections have been preserved in
> > > > --
> > > > 2.17.1
> > > >