Posix insists that the orientation of stderr not be changed by perror, even if stderr is not yet oriented.

Following is the quote from "XSH 2.5 Standard I/O Streams" [1]:

The perror(), psiginfo(), and psignal() functions shall behave as described above for the byte output functions if the stream is already byte-oriented, and shall behave as described above for the wide-character output functions if the stream is already wide-oriented. If the stream has no orientation, they shall behave as described for the byte output functions except that they shall not change the orientation of the stream.
And glibc attempts to implement Posix semantics. Unfortunately, it doesn't quite get it right.

Here is also quote from "perror()" reference page [2]:

The perror() function shall not change the orientation of the standard error stream.

And glibc attempts to implement Posix semantics. Unfortunately, it doesn't quite get it right.

Following are the tests when stderr is wide-oriented, multibyte-oriented and not oriented, prior to calling perror(). Tests 1) and 2) are OK. The issue is in test 3).

1) stderr is wide-oriented:

#include <stdio.h>
#include <wchar.h>
#include <errno.h>
int main(void)
{
  fwide(stderr, 1);
  errno = EINVAL;
  perror("");
  int x = fwide(stderr, 0);
  printf("fwide: %d\n",x);
  return 0;
}

$ ./a.out
Invalid argument
fwide: 1
$ ./a.out 2>/dev/null
fwide: 1

2) stderr is multibyte-oriented:

#include <stdio.h>
#include <wchar.h>
#include <errno.h>
int main(void)
{
  fwide(stderr, -1);
  errno = EINVAL;
  perror("");
  int x = fwide(stderr, 0);
  printf("fwide: %d\n",x);
  return 0;
}

$ ./a.out
Invalid argument
fwide: -1
$ ./a.out 2>/dev/null
fwide: -1

3) stderr is not oriented:

#include <stdio.h>
#include <wchar.h>
#include <errno.h>
int main(void)
{
  printf("initial fwide: %d\n", fwide(stderr, 0));
  errno = EINVAL;
  perror("");
  int x = fwide(stderr, 0);
  printf("fwide: %d\n", x);
  return 0;
}

$ ./a.out
initial fwide: 0
Invalid argument
fwide: 0
$ ./a.out 2>/dev/null
initial fwide: 0
fwide: -1

Of course, it is impossible to write to a stream without setting its orientation. So in an attempt to comply with this curious requirement, glibc attempts to make a new stream based on the same fd as stderr by using dup().

fileno() extracts the fd from a standard C library stream. dup() takes an fd, duplicates it, and returns the number of the copy. And fdopen() creates a standard C library stream from an fd. In short, that doesn't reopen stderr; rather, it creates (or attempts to create) a copy of stderr which can be written to without affecting the orientation of stderr.

Unfortunately, it doesn't work reliably because of the mode:

    fp = fdopen(fd, "w+");

That attempts to open a stream which allows both reading and writing. And it will work with the original stderr, which is just a copy of the console fd, originally opened for both reading and writing. But when you bind stderr to some other device with a redirect:

$ ./a.out 2>/dev/null

you are passing the executable an fd opened only for output. And  fdopen won't let you get away with that:

The application shall ensure that the mode of the stream as expressed by the mode argument is allowed by the file access mode of the open file description to which fildes refers.

The glibc implementation of fdopen actually checks, and returns NULL with errno set to EINVAL if you specify a mode which requires access rights not available to the fd.

I don't know why the glibc code uses "w+" as a mode argument, since it has no intention of reading from stderr.

Actually, you could just write directly to the new fd using write instead of fwrite, so I don't think fdopen is actually necessary at all.

[1] http://pubs.opengroup.org/onlinepubs/9699919799/functions/V2_chap02.html#tag_15_05

[2] http://pubs.opengroup.org/onlinepubs/9699919799/functions/perror.html