help('modules spam') causes segfault.

When pydoc tries goes though the files it does the following in the ModuleScanner class.

(minimal example)

>>> for importer, modname, ispkg in pkgutil.walk_packages():
...     if modname == 'test.badsyntax_pep3120':
...         loader = importer.find_module(modname)
... 
Segmentation fault

Adding:

   if modname=='test.badsyntax_pep3120':
      continue

At the top of the for loop will suppress the segfault in pydoc by skipping that file.


A bit further probing narrowed it down to this...

>>> loader = pkgutil.get_loader('test.badsyntax_pep3120')
Segmentation fault

I'm not familiar with the pkgutil module so I didn't go further.

Looks like tok->filename isn't set in PyTokenizer_FromFile:


Index: Parser/tokenizer.c
===================================================================
--- Parser/tokenizer.c  (revision 82984)
+++ Parser/tokenizer.c  (working copy)
@@ -818,6 +818,7 @@
     tok->cur = tok->inp = tok->buf;
     tok->end = tok->buf + BUFSIZ;
     tok->fp = fp;
+    tok->filename = "XXX";
     tok->prompt = ps1;
     tok->nextprompt = ps2;
     if (enc != NULL) {

To be a little clearer: Since tok->filename is used as a flag in other
places, I'm not sure where to set it without breaking other things.
This is the location of the segfault and the following would "fix" it
(using a placeholder for the name):


Index: Parser/tokenizer.c
===================================================================
--- Parser/tokenizer.c  (revision 83019)
+++ Parser/tokenizer.c  (working copy)
@@ -582,6 +582,8 @@
     if (badchar) {
         /* Need to add 1 to the line number, since this line
            has not been counted, yet.  */
+        if (tok->filename == NULL)
+            tok->filename = "<file>";
         PyErr_Format(PyExc_SyntaxError,
             "Non-UTF-8 code starting with '\\x%.2x' "
             "in file %.200s on line %i, "

Trying to look at this a bit further...

The PyErr_Format function then calls PyUnicode_FromFormatV in Objects/unicodeobject.c to do the actual formating.

It looks like there have been a number of issues of this type. Search the tracker for unicodeobject.c and PyUnicode_FromFormatV.

This may be related to http://bugs.python.org/issue7330.

The error happens when Null is passed to strlen in (unicodeobject.c, line 860)

Passing NULL to a string format function is probably in the category of don't do that.

Stefans solution of checking for NULL before calling PyErr_Format looks to me to be correct.  As he notes, the same technique is used in another place.

It might be good to add a note somewhere not to pass NULL to C string format functions.

This is by far the simplest fix for this.  See patch file.

This patch is what Stefan Krah suggested and I agree unless someone a lot more familiar with the import process can take a look at this and re-factor things so the filename is passed along with the file descriptor so the error can include it.  I tried and each change required other changes in order to get the correct failure and error messages.  I gave up when the tests for non existent file started failing with ']' on the first line.

I can't commit this, so could someone who can take a look at the patch and move this along.

The diff which I posted was not really meant as the ultimate fix. It's a
workaround. I think the real filename should be passed down to
PyTokenizer_FromFile or
PyTokenizer_FindEncoding.

I added you to this Victor because it looks like what your doing to rewrite the imports to work with Unicode (issue:9425) overlaps this.

See the test in the patch.

Your rewrite may fix this as the segfault has to do with getting the file encoding.  My apologies if this isn't the case.

Ron

The test in the patch isn't quite right. The following still fails.

Python 3.2a3+ (py3k:85719, Oct 18 2010, 22:32:47) 
[GCC 4.4.3] on linux2
Type "help", "copyright", "credits" or "license" for more information.
>>> import imp
>>> imp.find_module('test/badsyntax_pep3120')
Segmentation fault

help no longer segfaults, but the find_module call still does; updating title.  The patch does cure the segfault, but as Stefan says it isn't the best fix since having '<file>' in the error message instead of the real file name isn't very useful.

Pydoc skips the badsysntax_pep3120 file for now.  When this gets fixed that workaround should be removed.  The work around is commented and refers to this issue #.

p3k_i9313.diff is just a workaround, not the correct fix. The problem is that PyTokenizer_FindEncoding() doesn't get the filename.

I wrote tokenizer_encoding_filename.patch which add PyTokenizer_FindEncodingFilename() and patch import.c and traceback.c to pass the filename.

Hum, I'm not sure that my patch works if the locale encoding is not UTF-8: import.c manipulates path in the filesystem encoding, whereas PyTokenizer_FindEncodingFilename() expects UTF-8 filename.

See this patch as a draft, I will try to fix the encoding issue.

See also #9738 (Document the encoding of functions bytes arguments of the C API) to check which encoding is expected :-p

I'm working on #3080 which changes the import machinery to use Unicode instead of byte strings, and so it will be possible to fix this issue correctly.

New changeset 7b8d625eb6e4 by Victor Stinner in branch 'default':
Issue #9319: Include the filename in "Non-UTF8 code ..." syntax error.
http://hg.python.org/cpython/rev/7b8d625eb6e4

> Hum, I'm not sure that my patch works if the locale encoding is not
> UTF-8: import.c manipulates path in the filesystem encoding, whereas
> PyTokenizer_FindEncodingFilename() expects UTF-8 filename.

Thanks to my work on #3080, the import machinery (and other functions using Python modules and filenames) manipulates filenames as Unicode, and so I don't have to care about the filename encoding anymore.

6e9dc970ac0e (of issue #10785) fixed the crash, 7b8d625eb6e4 added the filename into the SyntaxError.

How about applying the workaround patch to Python 3.2? An unprecise error message is way better than a segfault.

New changeset fa5e348889c2 by Victor Stinner in branch '3.2':
Issue #9319: Fix a crash on parsing a Python source code without encoding
http://hg.python.org/cpython/rev/fa5e348889c2

Fixed in 3.2 (fa5e348889c2) and 3.3 (7b8d625eb6e4). The bug is a regression introduced in Python 3.2, so Python 3.1 doesn't need to be fixed.

New changeset 701069f9888c by Victor Stinner in branch '3.2':
Issue #9319: Fix the unit test
http://hg.python.org/cpython/rev/701069f9888c

New changeset bb62908896fe by Jesus Cea in branch 'default':
Correctly merging #9319 into 3.3?
http://hg.python.org/cpython/rev/bb62908896fe

> Correctly merging #9319 into 3.3?

No, this issue was fixed differently in Python 3.3. I see that you reverted (bb62908896fe) this merge (except the test, which is a good idea).

Can the workaround be removed from Lib/pydoc.py:2001 ?
I tried to remove it from 3.2 and help('modules spam') seems to work fine.

New changeset ce0687a8383b by Nick Coghlan in branch 'default':
Issue #9319: Remove the workaround for this since fixed problem from pydoc
http://hg.python.org/cpython/rev/ce0687a8383b