[PATCH] elf: dl-minimal malloc needs to respect fundamental alignment

[H.J. Lu]

On Thu, Jun 23, 2016 at 8:01 AM, Florian Weimer <fweimer@redhat.com> wrote:
> On 06/21/2016 03:20 PM, H.J. Lu wrote:
>>
>> On Tue, Jun 21, 2016 at 6:06 AM, Florian Weimer <fweimer@redhat.com>
>> wrote:
>>>
>>> On 06/21/2016 03:00 PM, H.J. Lu wrote:
>>>
>>>>> MALLOC_ALIGNMENT is potentially larger.  malloc/tst-malloc-thread-fail
>>>>> tests
>>>>> for alignment.  To my knowledge, it passes on all regularly tested
>>>>> architectures after commit dea39b13e2958a7f0e75b5594a06d97d61cc439f.
>>>>
>>>>
>>>>
>>>> MALLOC_ALIGNMENT is kind of mapped to the malloc alignment of
>>>> a psABI.  Shouldn't ld.so malloc have the same alignment of libc malloc?
>>>
>>>
>>>
>>> I don't see why.  MALLOC_ALIGNMENT has to match both the ABI constraint
>>> and
>>> the malloc/malloc.c implementation constraint (which requires a minimum
>>> alignment of 2 * sizeof (size_t)).
>>
>>
>> My understanding is since the minimum constraint of malloc alignment
>> <= ABI alignment, MALLOC_ALIGNMENT == ABI alignment.  Do you
>> have a glibc platform where it isn't true?
>
>
> If I'm not mistaken, m68k has fundamental alignment 2 (the alignment of
> long, double, and long double), but our malloc still aligns to 8 (2 * sizeof
> (size_t)).
>

Malloc alignment has been an isssie.  GCC defines
MALLOC_ABI_ALIGNMENT as

#ifndef MALLOC_ABI_ALIGNMENT
#define MALLOC_ABI_ALIGNMENT BITS_PER_WORD
#endif

which is incorrect for Linux and unusable:

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=36159

I'd like to see a reasonable solution to address malloc alignment
on Linux, not just in ld.so.


-- 
H.J.