Re: [PATCH v3 16/16] objtool,x86: Rewrite retpoline thunk calls

[Fangrui Song]

On 2021-06-04, 'Nick Desaulniers' via Clang Built Linux wrote:
> On Fri, Jun 4, 2021 at 1:50 PM Nick Desaulniers <ndesaulniers@xxxxxxxxxx> wrote:
> > (Manually replying to https://lore.kernel.org/lkml/CAFJ_xbq06nfaEWtVNLtg7XCJrQeQ9wCs4Zsoi5Y_HP3Dx0iTRA@xxxxxxxxxxxxxx/)
> >
> > Hi Peter,
> > We're also tracking 2 recent regressions that look like they've come from this
> > patch.
> >
> > https://github.com/ClangBuiltLinux/linux/issues/1384
> > https://github.com/ClangBuiltLinux/linux/issues/1388
> >
> > (Both in linux-next at the moment).
> >
> > The first, it looks like a boot failure. The second is a warning from the
> > linker on a kernel module; even readelf seems unhappy with the results of the
> > output from objtool.
> >
> > I can more easily reproduce the latter, so I'm working on getting a smaller
> > reproducer. I'll let you know when I have it, but wanted to report it ASAP.
>
> Sent a pretty big attachment, privately.  I was able to capture the
> before/after with:
>
> $ $ echo 'CONFIG_GCOV_KERNEL=n
> CONFIG_KASAN=n
> CONFIG_LTO_CLANG_THIN=y' >allmod.config
> $ OBJTOOL_ARGS="--backup" make -kj"$(nproc)" KCONFIG_ALLCONFIG=1
> LLVM=1 LLVM_IAS=1 all
>
> It looks like
>
> $ ./tools/objtool/objtool orc generate  --module  --no-fp
> --no-unreachable  --retpoline  --uaccess  --mcount
> drivers/gpu/drm/amd/amdgpu/amdgpu.lto.o; ld.lld -r -m elf_x86_64
> -plugin-opt=-code-model=kernel -plugin-opt=-stack-alignment=8
> --thinlto-cache-dir=.thinlto-cache -mllvm -import-instr-limit=5
> -plugin-opt=-warn-stack-size=2048 --build-id=sha1  -T
> scripts/module.lds -o drivers/gpu/drm/amd/amdgpu/amdgpu.ko
> drivers/gpu/drm/amd/amdgpu/amdgpu.lto.o
> drivers/gpu/drm/amd/amdgpu/amdgpu.mod.o
>
> is producing the linker error:
>
> ld.lld: error: drivers/gpu/drm/amd/amdgpu/amdgpu.lto.o:
> SHT_SYMTAB_SHNDX has 79581 entries, but the symbol table associated
> has 79582
>
> Readelf having issues with the output:
> $ readelf -s amdgpu.lto.o.orig
> <works fine>
> $ readelf -s amdgpu.lto.o
> readelf: Error: Reading 73014451695 bytes extends past end of file for
> string table
> $ llvm-readelf -s amdgpu.lto.o
> llvm-readelf: error: 'amdgpu.lto.o': unable to continue dumping, the
> file is corrupt: section table goes past the end of file
>
> `file` having issues:
> $ file drivers/gpu/drm/amd/amdgpu/amdgpu.lto.o
> drivers/gpu/drm/amd/amdgpu/amdgpu.lto.o: ELF 64-bit LSB relocatable,
> x86-64, version 1 (SYSV), no section header
>
> for comparison:
> $ file ./drivers/spi/spi-ath79.lto.o
> ./drivers/spi/spi-ath79.lto.o: ELF 64-bit LSB relocatable, x86-64,
> version 1 (SYSV), not stripped

tools/objtool/elf.c:elf_add_symbol may not update .symtab_shndx .
Speaking of llvm-objcopy, it finalizes the content of .symtab_shndx when .symtab
is finalized. objtool may want to adopt a similar approach.

read_symbols searches for the section ".symtab_shndx". It'd be better to
use the section type SHT_SYMTAB_SHNDX.