[BZ 1190] Make EOF sticky in stdio.zack/sticky-eof

C99 specifies that the EOF condition on a file is "sticky": once EOF
has been encountered, all subsequent reads should continue to return
EOF until the file is closed or something clears the "end-of-file
indicator" (e.g. fseek, clearerr).  This is arguably a change from
C89, where the wording was ambiguous; the BSDs always had sticky EOF,
but the System V lineage would attempt to read from the underlying fd
again.  GNU libc has followed System V for as long as we've been
using libio---the relevant chunk of code is

    int
    _IO_new_file_underflow (_IO_FILE *fp)
    {
      _IO_ssize_t count;
    #if 0
      /* SysV does not make this test; take it out for compatibility */
      if (fp->_flags & _IO_EOF_SEEN)
        return (EOF);
    #endif

That's been unchanged since before 1995.  This only matters if the
underlying file has changed in the meantime, of course; perhaps that's
why nobody got around to filing a bug report until 2005, six years
after C99 was published.  And nobody took that bug seriously until
2012, at which time there was a long but inconclusive discussion on
libc-alpha regarding whether it would break applications to change
anything, see <https://sourceware.org/ml/libc-alpha/2012-09/msg00343.html>.

It is my considered opinion that we should just go ahead and fix the
bug, and that a backward compatibility mode is not required, because
the BSDs have always had sticky EOF, so portable code has always had
to be prepared to deal with that behavior.  Nowadays, the lineages we
should be worrying most about compatibility with are all BSD-derived,
anyway.  Thus, this patch.

You might wonder if changing the _underflow impls is sufficient to
apply the C99 semantics to all of the many stdio functions that
perform input.  It should be enough to cover all paths to _IO_SYSREAD,
and the only other functions that call _IO_SYSREAD are the _seekoff
impls, which is OK because seeking clears EOF, and the _xsgetn impls,
which, as far as I can tell, are unused within glibc.

There is some question as to whether the test case in bug #19476
(one of the several duplicate bug reports) is valid, so instead I have
written a test case that uses a pseudoterminal to set up the necessary
conditions -- actually two test cases, one for narrow and one for wide
streams.  To facilitate this I added a new test-support function that
sets up a pair of pty file descriptors for you; it's almost the same
as BSD openpty, the only differences are that it allocates the
optionally-returned tty pathname with malloc, and that it crashes if
anything goes wrong.

zw

	[BZ 1190]
	* libio/fileops.c (_IO_new_file_underflow): Return EOF immediately
	if the _IO_EOF_SEEN bit is already set; update commentary.
	* libio/oldfileops.c (_IO_old_file_underflow): Likewise.
	* libio/wfileops.c (_IO_wfile_underflow): Likewise.

	* support/support_openpty.c, support/tty.h: New files.
	* support/Makefile (libsupport-routines): Add support_openpty.

	* libio/tst-fgetc-after-eof.c, wcsmbs/test-fgetwc-after-eof.c:
	New test cases.
	* libio/Makefile (tests): Add tst-fgetc-after-eof.
	* wcsmbs/Makefile (tests): Add tst-fgetwc-after-eof.

1 files changed, 3 insertions, 4 deletions

diff --git a/libio/fileops.c b/libio/fileops.c
index 79ad15351f..c9c5cbcc3c 100644
--- a/libio/fileops.c
+++ b/libio/fileops.c
@@ -468,11 +468,10 @@ int
 _IO_new_file_underflow (FILE *fp)
 {
   ssize_t count;
-#if 0
-  /* SysV does not make this test; take it out for compatibility */
+
+  /* C99 requires EOF to be "sticky".  */
   if (fp->_flags & _IO_EOF_SEEN)
-    return (EOF);
-#endif
+    return EOF;
 
   if (fp->_flags & _IO_NO_READS)
     {