Re: [PATCH 3/6] y2038: Introduce struct __timeval64 - new internal glibc type

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

On 21/01/2020 13:45, Lukasz Majewski wrote:
> Hi Arnd, Adhemerval,
> 
>> On Mon, Jan 20, 2020 at 5:46 PM Adhemerval Zanella
>> <adhemerval.zanella@linaro.org> wrote:
>>> On 20/01/2020 10:00, Arnd Bergmann wrote:  
>>>> On Mon, Jan 20, 2020 at 1:34 PM Lukasz Majewski <lukma@denx.de>
>>>> wrote:  
>>>>>>> On Sun, Jan 19, 2020 at 4:22 PM Lukasz Majewski <lukma@denx.de>
>>>>>>> wrote:  
>>>>>>>>> On Sat, Jan 18, 2020 at 11:48 PM Alistair Francis
>>>>>>>>> <alistair23@gmail.com> wrote:  
>>>>>>>
>>>>>>> It would certainly help if this could be consistent across
>>>>>>> architectures and libraries.  
>>>>>>
>>>>>> Ok. Thanks for the input.  
>>>>>
>>>>> Would it be OK to have:
>>>>>
>>>>> struct __timeval64
>>>>> {
>>>>>   __time64_t tv_sec;         /* Seconds */
>>>>>   __int64_t tv_usec;       /* Microseconds */
>>>>> };  
>>>>
>>>> This would not work if you pass it into a kernel interface that
>>>> expects a 32-bit suseconds_t on sparc64, but it should work as an
>>>> internal type on all 32-bit architectures.
>>>>  
>>>>> I would prefer to avoid changing typedef of suseconds_t as it may
>>>>> affect struct timeval related operations.
>>>>>
>>>>> Using explicitly __int64_t for tv_usec seems better option,
>>>>> isn't it?  
>>>>
>>>> Maybe a __suseconds64_t that happens to be 32-bit wide on
>>>> sparc64?  
>>>
>>> If timeval64 kernel ABI expects tv_usec being a different type
>>> depending of the architecture, it is an indication we should
>>> parametrize with a __suseconds64_t.
>>>
>>> As a side note, why tv_usec for timeval64 is a 64-bit value? Is it
>>> related to some alignment constraint?  
>>
>> I think it's mostly for simplicity: we do need the microseconds to be
>> in the low bytes for compatibility with the 64-bit syscalls in the
>> kernel, so it's either a 64-bit microseconds value or the same
>> conditional padding that exists in timespec, with an extra special
>> case for sparc64.
>>
>> The 64-bit case is slightly simpler for the few kernel interfaces that
>> take a timeval as input and then can avoid conditionally zeroing out
>> the upper 32 bits before checking the lower 32 bits again 1000000.
>> I think that is only needed for clock_adjtimex, VIDIOC_QBUF and
>> PPSETTIME though, everything else only passes a timeval
>> to user space.
> 
> I would opt for simpler approach :-)
> 
> Is the below code correct?
> ./include/time.h
> 
> struct __timeval64
> {
>    __time64_t tv_sec;         /* Seconds */
>    __suseconds64_t tv_usec;       /* Microseconds */
> };
> 
> However, I'm a bit confused on the correct place to define
> __suseconds64_t.
> 
> Shall it be added to posix/bits/types.h?
> __STD_TYPE __SUSECONDS_T_TYPE __suseconds_t; /* Signed count of
> microseconds.  */


Yes, the idea is to add on posix/bits/types.h:

__STD_TYPE __SUSECONDS64_T_TYPE __suseconds64_t;

Then on generic bits bits/typesizes.h:

#define __SUSECONDS_T_TYPE      __SQUAD_TYPE;

And then also define it on each Linux bits/typesizes.h.

> 
> If yes, then wouldn't we break the compatibility by adding new type to
> this file?

What do you mean by break compatibility? It is a new type on an
installed header used by new types.