Re: [PATCH v3 0/3] Add a futex interface to glibc (sys/futex.h)

On 25/08/26 14:46, Paul Eggert wrote:
> On 2026-08-25 09:50, Adhemerval Zanella wrote:
>> The
>> futex_timedwait supports both time_t sizes with the usual TIME_BITS=64
>> redirection
> 
> There's no backward-compatibility need for this, since it's a new glibc API. How about if this function supports only 64-bit time_t instead?
> 
> This issue is independent of what glibc does internally. Internally glibc may use helper functions that act like futex_timedwait but support 32-bit time_t, but I don't see why the 32-bit time_t functionality needs to be exposed to user apps.
> 
> At least, if there is a real need for futex_timedwait to support 32-bit time_t, that need should be explained in the patch and/or the documentation. It's confusing that some of the new API supports 32-bit time_t but other parts of it (e.g., futex_waitv) do not.

I was not fully sure about this and on my first draft I made futex_timedwait
64-bit time only. I only changed after I checked on libstdc++ code that uses
futex (libstdc++-v3/src/c++11/futex.cc and libstdc++-v3/libsupc++/guard.cc) 
and noticed that futex.cc redefine the struct timespec because the TU is not 
built with _TIME_BITS=64. I did not dig further, but I think it might be quite
complex to enable it on libstdc++.

Another reason, which I also do not have a strong opinion, is that futex does
have a 32-bit syscall (different than futex_waitv which is 64-bit only).

But it should be reasonable to only support futex_timedwait with 64-bit
time_t.