Re: Update on the glibc segfault issue on Alpha

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

On 04/01/23 06:25, John Paul Adrian Glaubitz wrote:
> Hi Adhemerval!
> 
> On 1/3/23 13:09, Adhemerval Zanella Netto wrote:
>> Thanks, this commits helps narrow down the issue.  The 73fc4e28b9464f0e refactor did not
>> add the GL(dl_phdr) and GL(dl_phnum) for static case, relying on the __ehdr_start symbol
>> to get the correct values.
>>
>> The issue is for some archs, alpha for instance, the hidden weak reference is not making
>> the static linker to define the __ehdr_start address correctly: it is being set to 0 and
>> thus GL(dl_phdr) and GL(dl_phnum) are set to invalid values.
>>
>> And I am not sure if the hidden weak __ehdr_start does work on all architectures, so I
>> think it would be safer to just restore the previous behavior to setup GL(dl_phdr) and
>> GL(dl_phnum) for static and we can simplify __ehdr_start fallback case to not use
>> a weak ref (as for PIE). I am checking if the following patch trigger any regression,
>> at least for alpha it fixes the static failures:
>>
>> diff --git a/csu/libc-start.c b/csu/libc-start.c
>> index 543560f36c..63a3eceaea 100644
>> --- a/csu/libc-start.c
>> +++ b/csu/libc-start.c
>> @@ -271,18 +271,10 @@ LIBC_START_MAIN (int (*main) (int, char **, char ** MAIN_AUXVEC_DECL),
>>            So we can set up _dl_phdr and _dl_phnum even without any
>>            information from auxv.  */
>>
>> -      extern const ElfW(Ehdr) __ehdr_start
>> -# if BUILD_PIE_DEFAULT
>> -       __attribute__ ((visibility ("hidden")));
>> -# else
>> -       __attribute__ ((weak, visibility ("hidden")));
>> -      if (&__ehdr_start != NULL)
>> -# endif
>> -        {
>> -          assert (__ehdr_start.e_phentsize == sizeof *GL(dl_phdr));
>> -          GL(dl_phdr) = (const void *) &__ehdr_start + __ehdr_start.e_phoff;
>> -          GL(dl_phnum) = __ehdr_start.e_phnum;
>> -        }
>> +      extern const ElfW(Ehdr) __ehdr_start attribute_hidden;
>> +      assert (__ehdr_start.e_phentsize == sizeof *GL(dl_phdr));
>> +      GL(dl_phdr) = (const void *) &__ehdr_start + __ehdr_start.e_phoff;
>> +      GL(dl_phnum) = __ehdr_start.e_phnum;
>>       }
>>
>>     __tunables_init (__environ);
>> diff --git a/sysdeps/unix/sysv/linux/dl-parse_auxv.h b/sysdeps/unix/sysv/linux/dl-parse_auxv.h
>> index bf9374371e..5913c9d6e5 100644
>> --- a/sysdeps/unix/sysv/linux/dl-parse_auxv.h
>> +++ b/sysdeps/unix/sysv/linux/dl-parse_auxv.h
>> @@ -56,6 +56,10 @@ void _dl_parse_auxv (ElfW(auxv_t) *av, dl_parse_auxv_t auxv_values)
>>     if (GLRO(dl_sysinfo_dso) != NULL)
>>       GLRO(dl_sysinfo) = auxv_values[AT_SYSINFO];
>>   #endif
>> +#ifndef SHARED
>> +  GL(dl_phdr) = (void*) auxv_values[AT_PHDR];
>> +  GL(dl_phnum) = auxv_values[AT_PHENT];
>> +#endif
>>
>>     DL_PLATFORM_AUXV
>>   }
> 
> I can confirm that this patch fixes the problem for me. I have uploaded a manually built glibc
> package with the patch applied to Debian »unreleased« so that the buildds can resume building
> packages again.
> 
> Would it be possible to backport this patch to 2.36 once it has been merged with the current
> development tree? The reason is that I don't think that Debian is going to switch to anything
> beyond 2.36 anytime soon due to the upcoming feature freeze in January.

I have sent a more detailed patch [1] and I will probably backport to all affected releases
once it is accepted.