[RFC PATCH 1/4] nptl: use futex_time64 for lll_futex_syscall() if possible

Adhemerval Zanella Netto adhemerval.zanella@linaro.org
Wed Nov 11 13:54:34 GMT 2026



On 11/11/26 09:56, Arnd Bergmann wrote:
> On Wed, Nov 11, 2026, at 13:08, Adhemerval Zanella Netto wrote:
> 
>>
>> This diffstat seems wrong, futex-internal.h is listed as
>> 1 insertion and 5 deletions, but only a single-line change is posted.
>>
>> I guess it is the
>>
>>   #ifndef __NR_futex_time64
>>   # define __NR_futex_time64 __NR_futex
>>   #endif
>>
>> block at line 72.  Removing it on its own would break every 64-bit ABI,
>> because nptl/futex-internal.c uses futex_time64 unconditionally.
> 
> Yes, that's what I get for editing the patches manually. I removed
> that change from my git tree when I noticed it would break 64-bit
> systems but did not recreate the text when I did the same change
> in the local patch file.
> 
>> Moving it to lowlevellock-futex.h works, and I think it is what you
>> want anyway (see below). 
> 
> I had thought about that but not actually done it because redirecting
> syscall numbers with incompatible ABIs does seem fragile, even though
> I understand this one is probably safe and there are other users in
> glibc doing the same thing.
> 
>>> diff --git a/sysdeps/nptl/futex-internal.h b/sysdeps/nptl/futex-internal.h
>>> index 502f89476a24..8934a5ee7932 100644
>>> --- a/sysdeps/nptl/futex-internal.h
>>> +++ b/sysdeps/nptl/futex-internal.h
>>> @@ -123,7 +119,7 @@ futex_fatal_error (void)
>>>  static __always_inline int
>>>  futex_wait (unsigned int *futex_word, unsigned int expected, int private)
>>>  {
>>> -  int err = lll_futex_timed_wait (futex_word, expected, NULL, private);
>>> +  int err = lll_futex_wait (futex_word, expected, private);
>>>    switch (err)
>>>      {
>>>      case 0:
>>
>> This misses the two other places in nptl that call the time32 futex
>> directly with no timeout:
>>
>>   - lll_futex_wait_cancel, a few lines below in this same header, used by
>>     the GLIBC_2.0 __old_sem_wait compat symbol (i386 and the other ABIs
>>     that have it). With the time32 futex gone, it returns -1/ENOSYS
>>     instead of blocking, and it is silent regression (and unfortunately we
>>     don't have a testcase for it).
> 
> This one should be covered by my change to lll_futex_syscall()
> 
>>   - nptl/pthread_mutex_trylock.c, the FUTEX_TRYLOCK_PI call on the robust
>>     PI owner-died path. On ENOSYS the code continues as if the kernel had
>>     finished the owner death, it clears FUTEX_OWNER_DIED and returns
>>     EOWNERDEAD. I am almost sure it is wrong.
> 
> This one I did miss.
> 
>> Instead of two copies of lll_futex_syscall, I would pick the syscall
>> number once, use the existing fallback idiom, and use it everywhere.
> 
> Fair enough.
> 
> On a related note, if both glibc and musl end up defining the
> __NR_...time64 macros this way for 64-bit architectures, I wonder
> if we should just do the same thing in the kernel and use the time64
> names as the canonical names for the macros. There are several
> ways to do this, and each of them risks breaking something, but
> it does seem appealing to have the definitions in the kernel do
> something along the lines of:
> 
> // in the generated asm/unistd.h file from the kernel, drop __NR_futex
> // in favor of explicit __NR_futex_time32/__NR_futex_time64 to match
> // the internal implementation
> 
> #if __WORDSIZE==32
> #define __NR_futex_time32 98
> #define __NR_futex_time64 422
> #define __NR_gettimeofday_time32 169
> ...
> #else
> #define __NR_futex_time64 98
> #define __NR_gettimeofday_time64 169
> ...
> #endif
> 
> // in a new architecture-independent header included from there
> #if __WORDSIZE==64 || (defined(__x86_64__) && defined(__ILP32__))
> # define __NR_futex __NR_futex_time64
> # define __NR_gettimeofday __NR_gettimeofday_time64
> ...
> #elif !defined(__USE_TIME_BITS64)
> # define __NR_futex __NR_futex_time32
> # define __NR_gettimeofday __NR_gettimeofday_time32
> ...
> #else
> // leave the old macros undefined when building with __USE_TIME_BITS64
> #endif
> 
> This would forcibly break any source code that attempts to
> call the time32 syscalls by their traditional name in a source
> file built with -D_TIME_BITS=64, which would otherwise be a bug
> that may go unnoticed for a long time.

The _TIME_BITS is only meant to be used by programs built against glibc, we
only use it on tests (and configuring glibc with this should trigger an
error). And I am not sure if it would simplify things, internal calls
usually only uses 64 bit time_t variants now (either through wrappers like
lowlevellock.h or by routing explicitly to internal symbols). The 32 bit
time_t variants are only provided for ABI compatibility.

Or do you mean by external users, through syscall()? In this case I am
not sure if routines the SYS_ macro depending of _TIME_BITS value is a 
good idea.



More information about the Libc-alpha mailing list