[PATCH v3 23/29] configure: Add --with-ld-relro-load-gaps configure option and test

[Adhemerval Zanella Netto]

On 26/12/24 10:16, Florian Weimer wrote:
> * Adhemerval Zanella Netto:
> 
>> On 29/11/24 13:17, Florian Weimer wrote:
>>> * Adhemerval Zanella Netto:
>>>
>>>> On 29/09/24 13:36, Florian Weimer wrote:
>>>>> Related to binutils bug 28743.
>>>>>
>>>>> The new test elf/tst-gaps-ldso verifies that ld.so does not have gaps
>>>>> if the configure check indicates that the linker does not produce such
>>>>> gaps, and the architecture supports a single page size only.
>>>>>
>>>>> The companion test elf/tst-load-alignment checks that the LOAD
>>>>> segments produced during the build permit loading glibc on the
>>>>> largest supported page size.
>>>>
>>>> I failing to see the usefulness of this new configure check, if I understand
>>>> correctly on following patches it uses as an optimization on rtld.c and
>>>> dl-find_object.c to avoid some extra work for l_contiguous. It also seems
>>>> that it used solely on cold-path (initialization).
>>>>
>>>> Do we really need this configure switch?
>>>
>>> The test is required to disable the gap test if binutils is known to be
>>> buggy and produces gaps even on targets where just one page size is
>>> supported.  This bug is fixed in later binutils.  We still do not have a
>>> way to produce gapless binaries that support multiple page sizes.
>>
>> Can't we just add this test as default and mark the tests as XFAIL in
>> case of on a faulty binutils?
> 
> Do you mean XFAILing it unconditionally, without checking the binutils
> version?  Or maybe XFAIL it only for --without-ld-relro-load-gaps,
> and default to that?

XFAILing or just let if fail if caller builds glibc with a binutils
with this issue.  It seems like a critical fix that someone packing
glibc should be aware of.