Re: Patch review status for 64-bit time_t patches.

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

On 03/06/2021 00:56, Carlos O'Donell wrote:
> On 5/31/21 5:09 PM, Carlos O'Donell wrote:
>> On 5/27/21 5:46 PM, Carlos O'Donell wrote:
>>> Adhemerval, Lukasz,
>>>
>>> I'm at ~6/25 patches reviewed. I started not by reviewing but by building additional
>>> glibc packages for Fedora Rawhide and then testing in Rawhide.
>>>
>>> Testing looks good in Fedora Rawhide on top of Florian's recent merge in of the
>>> most recent development changes.
>>>
>>> I'll continue reviewing these as discussed in order to ensure we can merge them
>>> in a timely fashion.
>>
>> Just a quick status update.
>>
>> I'm at ~18/25 patches reviewed.
>>
>> I had a somewhat long discussion with Adhemerval on IRC about the new dual-time
>> ABI and the dropping of the __glibc_reserved* entries form the public ABI. This
>> is OK for the new ABI, but we must be careful we don't ever define __USE_TIME_BITS64
>> for any existing 64-bit time_t ABIs since it would be an ABI regression.
>>
>> I'm out of time today, but I'll come back tomorrow to try finish the reviews.
>  
> I'm at ~21/25 patches reviewed.
> 
> I went back over the sysvipc ABI again looking to see if I could make it match
> up by putting back the reserved structure members to match the kernel ABI and
> I could. 
> 
> I'm going to recommend we change to match the kernel ABI from a legacy perspective.

Hi Carlos,

I tried to follow your suggestion here and although feasible, it would required
to provide a special struct_msqid64_ds_helper.h (and other similar sysvipc 
headers) for some specific architectures: arm, mips (all abis), and s390.

This is due the special alignment requirement of changing the fields with
__time_t to __time64. For instance, with even the extra two reserved fields
for 'msqid_ds', the size increases from 88 to 96 on arm.

And these extra arch-specific definitions is one thing I would like to
avoid. As I commented with you on IRC, the sysvipc is really an legacy API
and I don't expect the kernel to provide new control mechanism (that
would require us to extend the control structures.