Re: segfaults in _dl_fixup since 4af6982

[Adhemerval Zanella <adhemerval.zanella@linaro.org>]

On 13/10/2021 11:51, Manuel Lauss wrote:
> This patch works on x64_64 now, the 32bit x86 build breaks.

Thanks for checking it, I know what I got wrong in first place.
The 32-bit break is expected, I added only x86_64 support.  I will
prepare a patch to work on all architectures and run make check
on the architecture I have available.

> 
> Manuel
> 
> On Wed, Oct 13, 2021 at 4:41 PM Adhemerval Zanella
> <adhemerval.zanella@linaro.org> wrote:
>>
>>
>>
>> On 13/10/2021 10:55, Manuel Lauss wrote:
>>> Applied this patch on top of git head, still same error.
>>>
>>> Manuel
>>
>> Sigh, it seems I got the elf_get_dynamic_info() sideways.  Could you check
>> this slight modified patch? It is x86_64 only for now.
>>
>> And I am puzzled why I am not seeing any issues in the environments I am
>> checking...
>>
>>>
>>> On Wed, Oct 13, 2021 at 3:47 PM Adhemerval Zanella
>>> <adhemerval.zanella@linaro.org> wrote:
>>>>
>>>>
>>>>
>>>> On 13/10/2021 08:56, Adhemerval Zanella wrote:
>>>>>
>>>>>
>>>>> On 13/10/2021 08:44, Adhemerval Zanella wrote:
>>>>>>
>>>>>>
>>>>>> On 13/10/2021 07:58, Adhemerval Zanella wrote:
>>>>>>>
>>>>>>>
>>>>>>>> On 13 Oct 2021, at 06:19, Szabolcs Nagy <szabolcs.nagy@arm.com> wrote:
>>>>>>>>
>>>>>>>> The 10/13/2021 08:42, Manuel Lauss via Libc-alpha wrote:
>>>>>>>>> [resend with libc-alpha]
>>>>>>>>> Hi Adhemerval,
>>>>>>>>>
>>>>>>>>> Your recent commit 4af6982e4c9fc465ffb7a54b794aaaa134241f05 ("elf: Fix
>>>>>>>>> elf_get_dynamic_info definition")  causes a lot (but not all) of my
>>>>>>>>> applications
>>>>>>>>> to segfault:
>>>>>>>>>
>>>>>>>>> Program received signal SIGSEGV, Segmentation fault.
>>>>>>>>> _dl_fixup (l=<optimized out>, reloc_arg=<optimized out>) at dl-runtime.c:146
>>>>>>>>> 146       return elf_machine_fixup_plt (l, result, refsym, sym, reloc,
>>>>>>>>> rel_addr, value);
>>>>>>>>> (gdb) bt
>>>>>>>>> #0  _dl_fixup (l=<optimized out>, reloc_arg=<optimized out>) at dl-runtime.c:146
>>>>>>>>> #1  0x00007ffff7fe251a in _dl_runtime_resolve_xsavec () at
>>>>>>>>> ../sysdeps/x86_64/dl-trampoline.h:126
>>>>>>>>> #2  0x000055555555634e in main (argc=1, argv=0x7fffffffd568) at
>>>>>>>>> iconvconfig.c:288
>>>>>>>>>
>>>>>>>>> Reverting the commit fixes the problem. Gcc, python, etc seem to be
>>>>>>>>> not affected, while for example iconvconfig and openssh are.
>>>>>>>>
>>>>>>>> on aarch64 buildbot i see
>>>>>>>>
>>>>>>>> FAIL: nptl/test-cond-printers
>>>>>>>> FAIL: nptl/test-condattr-printers
>>>>>>>> FAIL: nptl/test-mutex-printers
>>>>>>>> FAIL: nptl/test-mutexattr-printers
>>>>>>>> FAIL: nptl/test-rwlock-printers
>>>>>>>> FAIL: nptl/test-rwlockattr-printers
>>>>>>>>
>>>>>>>> and
>>>>>>>>
>>>>>>>> $ cat nptl/test-cond-printers.out
>>>>>>>> Error: Response does not match the expected pattern.
>>>>>>>> Command: start
>>>>>>>> Expected pattern: main
>>>>>>>> Response:  Temporary breakpoint 1 at 0xac4: file test-cond-printers.c, line 36.
>>>>>>>> Starting program: /work/glibc-aarch64-linux/build/build/nptl/test-cond-printers
>>>>>>>>
>>>>>>>> Program received signal SIGSEGV, Segmentation fault.
>>>>>>>> _dl_fixup (l=<optimized out>, reloc_arg=<optimized out>) at dl-runtime.c:146
>>>>>>>> 146       return elf_machine_fixup_plt (l, result, refsym, sym, reloc, rel_addr, value);
>>>>>>>> (gdb)
>>>>>>>>
>>>>>>>>
>>>>>>>> it seems to be a bind-now pie where PLT relocs are not yet
>>>>>>>> processed when __libc_start_main is called but PLTGOT is
>>>>>>>> already marked readonly so _dl_fixup crashes.
>>>>>>>
>>>>>>> I will check this out.
>>>>>>>
>>>>>>
>>>>>> I am trying to reproduce it on aarch64-linux-gnu without much success (using
>>>>>> gcc 11.1.1 20210702 and ld 2.36.1.20210702). Neither ssh or iconvconfig fails
>>>>>> when running through the built glibc (using ubuntu 21.04 one).
>>>>>>
>>>>>> Any idea on how to reproduce it?
>>>>>>
>>>>>
>>>>> Sigh... I think this is another fallback of RTLD_BOOTSTRAP.  I fixed the
>>>>> elf_get_dynamic_info(), but there are lot of more usage that changes
>>>>> the semantic of the nested functions (specifically on elf/do-rel.h
>>>>> and dl-machine.h).
>>>>>
>>>>> I will try to fix, I think we should define RTLD_BOOTSTRAP before the
>>>>> fist dynamic-link.h in rtld.
>>>>>
>>>>
>>>> Could you check if the attached patch fixes it?  The rtld.c now should get
>>>> all the required function with the expected semantic (RTLD_BOOTSTRAP defined).