[ python-Bugs-1074011 ] write failure ignored in Py_Finalize()

[SourceForge.net]

Bugs item #1074011, was opened at 2004-11-26 18:02
Message generated for change (Comment added) made by rhettinger
You can respond by visiting: 
https://sourceforge.net/tracker/?func=detail&atid=105470&aid=1074011&group_id=5470

Category: Python Interpreter Core
Group: Python 2.3
Status: Open
Resolution: None
Priority: 6
Submitted By: Matthias Klose (doko)
>Assigned to: Tim Peters (tim_one)
Summary: write failure ignored in Py_Finalize()

Initial Comment:
[forwarded from http://bugs.debian.org/283108]

Write errors on stdout may be ignored, and hence may 
result in loss of valuable user data. 
 
Here's a quick demo: 
 
$ ./close-bug 
foo 
$ ./close-bug > /dev/full && echo unreported write failure 
unreported write failure 
$ cat close-bug 
#!/usr/bin/python 
import sys 
def main (): 
    try: 
        print 'foo' 
        sys.stdout.close () 
    except IOError, e: 
        sys.stderr.write ('write failed: %s\n' % e) 
        sys.exit (1) 
 
if __name__ == '__main__': 
    main () 


This particular failure comes from the following
unchecked fflush 
of stdout in pythonrun.c: 
 
  static void 
  call_ll_exitfuncs(void) 
  { 
          while (nexitfuncs > 0) 
                  (*exitfuncs[--nexitfuncs])(); 
 
          fflush(stdout); 
          fflush(stderr); 
  } 
 
Flushing the stream manually, python does raise an
exception.

Please note that simply adding a test for fflush
failure is 
not sufficient.  If you change the above to do this: 
 
  if (fflush(stdout) != 0) 
    { 
      ...handle error... 
    } 
 
It will appear to solve the problem. 
But here is a counterexample: 
 
import sys 
def main (): 
    try: 
        print "x" * 4095 
        print 
        sys.stdout.close () 
    except IOError, e: 
        sys.stderr.write ('write failed: %s\n' % e) 
        sys.exit (1) 

if __name__ == '__main__': 
    main () 
 
If you run the above with stdout redirected to /dev/full, 
it will silently succeed (exit 0) in spite of a write
failure. 
That's what happens on my debian unstable system. 
 
Instead of just checking the fflush return value, 
it should also check ferror: 
 
  if (fflush(stdout) != 0 || ferror(stdout)) 
    { 
      ...handle error... 
    } 



----------------------------------------------------------------------

>Comment By: Raymond Hettinger (rhettinger)
Date: 2004-12-19 16:47

Message:
Logged In: YES 
user_id=80475

Tim, what do you think?

----------------------------------------------------------------------

Comment By: Ben Hutchings (wom-work)
Date: 2004-12-06 19:33

Message:
Logged In: YES 
user_id=203860

OK, I can reproduce the remaining problem if I substitute
1023 for 4095. The culprit seems to be the unchecked fputs()
in PyFile_WriteString, which is used for the spaces and
newlines generated by the print statement but not for the
objects. I think that's a separate bug.

----------------------------------------------------------------------

Comment By: Jim Meyering (meyering)
Date: 2004-12-06 18:27

Message:
Logged In: YES 
user_id=41497

Even with python-2.4 (built fresh from CVS this morning),
I can still reproduce the problem on a Linux-2.6.9/ext3 system:

  /p/p/python-2.4/bin/python write-4096 > /dev/full && echo fail
  fail

The size that provokes the failure depends on the I/O block size
of your system, so you might need something as big as 131072
on some other type of system.

----------------------------------------------------------------------

Comment By: Ben Hutchings (wom-work)
Date: 2004-12-06 18:11

Message:
Logged In: YES 
user_id=203860

The second sample script works for me (in that it raises the
expected exception) with or without my patch. If the error
status of the file has been set by some other operation
without Python noticing it then surely that's a bug in the
code that calls that other operation.

----------------------------------------------------------------------

Comment By: Jim Meyering (meyering)
Date: 2004-12-02 04:20

Message:
Logged In: YES 
user_id=41497

Hi,
I submitted the original report (and text above).
The patch that just calls fflush is not enough, as the second sample script illustrates.
Here's a patch that does a little better

--- Python/sysmodule.c.~2.126.~ 2004-08-12 20:19:17.000000000 +0200
+++ Python/sysmodule.c  2004-12-02 09:59:09.058953816 +0100
@@ -927,6 +927,13 @@ settrace() -- set the global debug traci
 )
 /* end of sys_doc */ ;
 
+static int
+_check_and_flush (FILE *stream)
+{
+  int prev_fail = ferror (stream);
+  return fflush (stream) || prev_fail ? EOF : 0;
+}
+
 PyObject *
 _PySys_Init(void)
 {
@@ -941,8 +948,8 @@ _PySys_Init(void)
        sysdict = PyModule_GetDict(m);
 
        sysin = PyFile_FromFile(stdin, "<stdin>", "r", NULL);
-       sysout = PyFile_FromFile(stdout, "<stdout>", "w", NULL);
-       syserr = PyFile_FromFile(stderr, "<stderr>", "w", NULL);
+       sysout = PyFile_FromFile(stdout, "<stdout>", "w", _check_and_flush);
+       syserr = PyFile_FromFile(stderr, "<stderr>", "w", _check_and_flush);
        if (PyErr_Occurred())
                return NULL;
 #ifdef MS_WINDOWS

With that patch, the failing script now evokes a diagnostic
and nonzero exit.

  $ ./python write-4096 > /dev/full
  write failed: (0, 'Error')

But, as you can see, the diagnostic leaves a lot to be desired.
It should say ``write failes: [Errno 28] No space left on device''.
It'll take a more significant change to propagate errno from the
failing fputs/fwrite/etc. to where it can be used.

Jim at meyering.net

----------------------------------------------------------------------

Comment By: Ben Hutchings (wom-work)
Date: 2004-11-29 05:51

Message:
Logged In: YES 
user_id=203860

I can't see how to upload an attachment to this bug, so see
patch 1075147.

----------------------------------------------------------------------

Comment By: Ben Hutchings (wom-work)
Date: 2004-11-29 05:43

Message:
Logged In: YES 
user_id=203860

Each Python file object has a pointer to the function to be
called on the C stream when close is called on the file
object.  Normally this is fclose, but Python doesn't want
the standard streams closed so their file objects are
created with the function pointer set to NULL, making close
a no-op on the underlying files.  I'm going to attach a
patch that changes the function for stdout and stderr to be
fflush, so that the streams stay open but write errors are
detected at the time of an explicit close.  I don't see the
relevance of the exitfuncs.


----------------------------------------------------------------------

You can respond by visiting: 
https://sourceforge.net/tracker/?func=detail&atid=105470&aid=1074011&group_id=5470