Re: [PATCH 1/3] linux: Use INTERNAL_SYSCALL on fstatat{64}

[Adhemerval Zanella via Libc-alpha <libc-alpha@sourceware.org>]

On 15/10/2020 12:45, Florian Weimer wrote:
> * Adhemerval Zanella via Libc-alpha:
> 
>> +  return INTERNAL_SYSCALL_ERROR_P (r)
>> +	 ? INLINE_SYSCALL_ERROR_RETURN_VALUE (-r)
>> +	 : 0;
> 
> I'm not really happy how the expression -r breaks the INTERNAL_*
> abstraction.  Do you really have to change the calling convention for
> __cp_kstat_stat?  I expect that if you don't do that (and make calls to
> __cp_kstat_stat tail calls), you can preserve the abstraction.

The INTERNAL_* abstractions is something I would like to get rid of
since currently INTERNAL_SYSCALL follows the same convention for all
architectures (meaning that there is no extra auxiliary variable that
hold whether the syscall has failed, as previously for powerpc and
sparc) and Linux assumes that errno values are all in the range of
[-4096UL, -1UL].

We can now simplify the INTERNAL_SYSCALL tests to check for specific
errno value without the need to resorting the extra macros.  I really
see that we can move to something simpler and more readable internal
APIs as:

  #define INTERNAL_SYSCALL_CALL (syscall, ...) ...

  static inline _Bool internal_syscall_failed (unsigned long int val)
  {
    return val > -4096UL;
  }

  static inline long int __syscall_ret (unsigned long int val)
  {
    if (internal_syscall_failed (val))
      {
        errno = -val;
        return -1;
      }
    return 0;
  }

  #define INLINE_SYSCALL_CALL (...) \
   __syscall_ret (NTERNAL_SYSCALL_CALL (__VA_ARGS__))


And so we can write more specific syscall wrappers as:

  int r = INTERNAL_SYSCALL_CALL (syscall, ...);
  if (r == 0 || -r == ENOSYS)
    return __syscall_ret (r);
  r = INTERNAL_SYSCALL_CALL (syscall_fallback, ...);
  return __syscall_ret (r);

Or similar if there is no need to handle errno:

  return -INTERNAL_SYSCALL_CALL (syscall, ...);