[PATCH] Fix section type of .eh_frame on Linux x86_64

[Adhemerval Zanella]

On 13/03/2020 13:07, Fangrui Song wrote:
> On 2020-03-13, H.J. Lu wrote:
>> On Fri, Mar 13, 2020 at 7:04 AM Adhemerval Zanella
>> <adhemerval.zanella@linaro.org> wrote:
>>>
>>>
>>>
>>> On 13/03/2020 10:12, H.J. Lu wrote:
>>> > On Fri, Mar 13, 2020 at 5:28 AM Adhemerval Zanella via Libc-alpha
>>> > <libc-alpha@sourceware.org> wrote:
>>> >>
>>> >>
>>> >>
>>> >> On 11/03/2020 18:27, Fangrui Song via Libc-alpha wrote:
>>> >>> Clang since https://reviews.llvm.org/D73999 will error for the wrong
>>> >>> sh_type.
>>> >>
>>> >> It will make eh_frame section of sigaction object to have the
>>> >> SHT_X86_64_UNWIND, but it seems that 'ld.bfd' at least ignore and
>>> >> set the resulting eh_frame sh_type for libc.so to SHT_PROGBITS
>>> >> anyway. The 'as' also still generate SHT_PROGBITS for code that
>>> >> requires a eh_frame (C++ with exception handling that emits a
>>> >> gcc_except_table section, for instance).
>>> >
>>> > SHT_X86_64_UNWIND was added so that linker can group all
>>> > SHT_X86_64_UNWIND sections together without checking
>>> > .eh_frame section name.   Other than that, linker treats .eh_frame like
>>> > other SHT_PROGBITS sections.   Since .eh_frame is also listed as a
>>> > special section, similar to .got and .plt, SHT_X86_64_UNWIND isn't
>>> > really required.
>>>
>>> Thanks for the explanation.
>>>
>>> >
>>> >> Setting the eh_frame in assembly routines is not a common practice,
>>> >> the only other code that I could find that actually does it is
>>> >> Linux.  For i686 vDSO is also uses 'progbits':
>>> >>
>>> >> arch/x86/entry/vdso/vdso32/sigreturn.S
>>> >>
>>> >>  36         .section .eh_frame,"a",@progbits
>>> >>  37 .LSTARTFRAMEDLSI1:
>>> >>
>>> >> The change should be ok, but I would like to understand better why
>>> >> exactly SHT_X86_64_UNWIND should be used for eh_frame, why binutils
>>> >> does not seems to use it to eh_frame, and why clang is now not
>>> >> accepting the eh_frame with SHT_PROGBITS type.
>>> >>
>>> >
>>> > Linker should check .eh_frame section name, and optionally check
>>> > SHT_X86_64_UNWIND.   Because of this, i386 psABI only specifies
>>> > special .eh_frame section without SHT_386_UNWIND.
>>> >
>>> > So I think there is no need to change glibc and lld should be changed.
>>> >
> 
> clang integrated assembler is more rigid.
> https://github.com/llvm/llvm-project/blob/master/llvm/lib/MC/MCObjectFileInfo.cpp#L488
> .eh_frame is a pre-allocated recognized section.
> "Redeclaring" (via .section directive) it with different sh_flags or sh_type will error.

The question now is whether this specific test make sense.  From 
H.J.Lu explanation, it seems clang strictness in not allowing
eh_frame to be defined as SHT_PROGBITS does not make much sense
because the section will be set a SHT_PROGBITS in any case
(even if a module defines it as SHT_X86_64_UNWIND).

> 
>>> I was confused why binutils was not throwing a similar issue.  So
>>> the idea is SHT_X86_64_UNWIND is optional on eh_frame definitions,
>>> not mandatory, and being optional I agree that compiler/linker
>>> should not emit an error in such case.
>>
>> It is unfortunate that x86-64 psABI specifies both SHT_X86_64_UNWIND
>> and .eh_frame.  Only one is needed, not both.
> 
> I know little about GNU as internals
> (https://sourceware.org/legacy-ml/binutils/2020-02/msg00129.html),
> but it seems .cfi* generated .eh_frame contents don't conflict with
> .section declared .eh_frame
> 
> Though, .section .eh_frame,"a",@unwind; .section .eh_frame,"a",@progbits
> will error as well.
> 
> https://sourceware.org/git/?p=binutils-gdb.git;a=commit;h=d2b2c203e1ff2daf8cfeab0906084f9389e66246 (2004) added SHT_X86_64_UNWIND to binutils.
> There was no further information why it was added, though.