Sub threads execute in restricted mode

[yetanothermax]

BPO	1163563
Nosy	@mwhudson, @mhammond, @brettcannon, @belm0
Files	TestRestrictedThreadApplication.java: Java file (using jep) that generates error for Python versions > 2.3.4 threadmodule.c: Updated Module\threadmodule.c file for python 2.3.5 (same as threadmodule.c in 2.3.4) threadmodule.c: Updated threadmodule.c for 2.4.1 threadtest.c: Shows problem when running under 2.3.5 and above threadmodule.c: Correct threadmodule.c for 2.4.1 baseed on latest CVS maybe-better-1010677-fix.diff: mwh fix Support "bpo-" in Misc/NEWS #1 maybe-even-better-1010677-fix.diff: mwh fix Rename README to README.rst and enhance formatting #2

^{Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.}

Show more details

GitHub fields:

assignee = 'https://github.com/mwhudson'
closed_at = <Date 2005-12-08.12:49:49.000>
created_at = <Date 2005-03-15.08:56:26.000>
labels = ['interpreter-core', 'release-blocker']
title = 'Sub threads execute in restricted mode'
updated_at = <Date 2005-12-08.12:49:49.000>
user = 'https://bugs.python.org/yetanothermax'

bugs.python.org fields:

activity = <Date 2005-12-08.12:49:49.000>
actor = 'alephant'
assignee = 'mwh'
closed = True
closed_date = None
closer = None
components = ['Interpreter Core']
creation = <Date 2005-03-15.08:56:26.000>
creator = 'yetanothermax'
dependencies = []
files = ['1632', '1633', '1634', '1635', '1636', '1637', '1638']
hgrepos = []
issue_num = 1163563
keywords = []
message_count = 23.0
messages = ['24626', '24627', '24628', '24629', '24630', '24631', '24632', '24633', '24634', '24635', '24636', '24637', '24638', '24639', '24640', '24641', '24642', '24643', '24644', '24645', '24646', '24647', '24648']
nosy_count = 8.0
nosy_names = ['mwh', 'mhammond', 'brett.cannon', 'jbelmonte', 'yetanothermax', 'goatpunch', 'mrjohnson0', 'alephant']
pr_nums = []
priority = 'release blocker'
resolution = 'fixed'
stage = None
status = 'closed'
superseder = None
type = None
url = 'https://bugs.python.org/issue1163563'
versions = ['Python 2.4']

[yetanothermax]

I'm using the JEP product which allows integration
of Java with Python (see http://jepp.sourceforge.net) via
starting a Python interpreter in the same process as the
JVM.
This integrates with python via the C interface, allowing
the user to 'eval' python code (amongst other features).

It seems that since Python 2.3.5 any threads (using the
threading module) created via code that has been
evaluated through the jep.eval() interface (i.e.Python C
interface )are executed in restricted mode and cannot,
for example, create files. Code that is just evaled (i.e not
in a subthread) thread has no restrictions.

The error reported is:
IOError: file() constructor not accessible in restricted
mode

(see http://sourceforge.net/forum/forum.php?
thread_id=1247793&forum_id=376782) - I've given a JEP
specific example here.

There seems to be a similar problem with mod_python
(see
http://www.modpython.org/pipermail/mod_python/2005-
January/017129.html)

Reading through the release notes for Python 2.3.5
I see:
Bug bpo-754449: threading.Thread will no longer mask
exceptions raised during interpreter shutdown with
another exception caused by attempting to output the
initial exception. This fix also includes a backport of rev.
1.41 from HEAD.

This might be the problem as it seems to involve the
porting of 2.4 threading code back to the 2.3 tree.

I've attached a Java file which shows the problem when 
using JEP.
The error output is:
Exception in thread Thread-1:
Traceback (most recent call last):
  File "C:\Python24\Lib\threading.py", line 442, in 
__bootstrap
  File "<string>", line 5, in run
IOError: file() constructor not accessible in restricted 
mode

2.4.1c1 (#63, Mar 10 2005, 10:36:41) [MSC v.1310 32
bit (Intel)]
Creating file from main thread...
Done
Creating file from sub thread...
Done

[goatpunch]

Logged In: YES
user_id=967140

The same problem definately does occur with mod_python,
I've found it to occur with Python 2.3.5 and 2.4 on Win32:

http://www.modpython.org/pipermail/mod_python/2005-
March/017802.html

[brettcannon]

Logged In: YES
user_id=357491

To answer the comment on bug bpo-754449 and the email I got
from someone, there is no way my fix for that bug would
cause this. I only touched the 'threading' module, nothing
else. That is in pure Python and thus has no direct way of
modifying what does and does not execute in restricted mode.
If I understand how restricted execution works that is all
set at the C level.

Plus the fix only deals with the teardown code; it in no way
interfaces with how threads are executed. All changed code
come after the thread and completed execution. The only
thing it does ahead of time is store a reference to stdout.

In other words it ain't my fault to the best of my
knowledge. =)

[belm0]

Logged In: YES
user_id=282299

This problem also occurs on Linux. Note however that
I've traced the source of the symptom to a different bug
fix, namely [1010677] thread Module Breaks
PyGILState_Ensure()
(http://sourceforge.net/tracker/index.php?func=detail&aid=1010677&group_id=5470&atid=305470).
Specifically, the change from v2.56 to v2.56.8.1 of
threadmodule.c.

[yetanothermax]

Logged In: YES
user_id=1218811

I've finally got access to a C compiler and it looks like
the 'pure' Python changes for bpo-754449 are in the clear
(apologies to bcannon).

The problem actually seems to relate to changes to
threadmodule.c of the Python Modules directory.
In versions greater than 2.3.4 it uses the PyGILState_XXX
constructs t_bootstrap() function.
These do not seem to be compatible with running sub
interpreters as JEP (and probably mod_python) do.
The python documentation
(http://docs.python.org/api/threads.html) says "Note that the
PyGILState_() functions assume there is only one global
interpreter (created automatically by Py_Initialize()). Python
still supports the creation of additional interpreters (using
Py_NewInterpreter()), but mixing multiple interpreters and the
PyGILState_() API is unsupported. "
I've compiled a new versions of the python2X.dll for both 2.3.5
and 2.4.1 changing the use of these PyGIL_StateXXX
constructs back to the 2.3.4 form and this seems to resolve
the problem (I've attached the updated threadmodule.c files
for both 2.3.5 and 2.4.1).

[yetanothermax]

Logged In: YES
user_id=1218811

Here's the updated threadmodule.c for 2.4.1

[yetanothermax]

Logged In: YES
user_id=1218811

I didn't see jbelmote's comment, perhaps my updated
threadmodule.c will reopen the bpo-1010677 bug.

[mhammond]

Logged In: YES
user_id=14198

Can anyone tell me specifically what Python C API function
is involved here? ie, the "entry point" that these external
threads use to call into Python.

[mrjohnson0]

Logged In: YES
user_id=1044885

In interactive mode, Jep uses Py_NewInterpreter,
Py_CompileString, and PyRun_String for its eval().

This file handles java -> python:
http://204.156.146.230/cgi-bin/viewcvs.cgi/jep/pyembed.c?rev=1.29&content-type=text/vnd.viewcvs-markup

You'd be looking for pyembed_eval().

[yetanothermax]

Logged In: YES
user_id=1218811

I've attached a file (threadtest.c) that is adapted from the
the way JEP uses Py_NewInterpreter.
This runs correctly
under 2.3.4 but fails under 2.3.5 and later Pythons.

[yetanothermax]

Logged In: YES
user_id=1218811

This problem prevents us from upgrading to Python 2.4.
I've attached a another 'fixed' threadmodule.c which is based
upon the lastest CVS release.(Only change from the last
threadmodule.c attachment is the removal of a '\n' from an
error message).
The threadtest.c program demonstrates the problem (seems
to be the use of Py_NewInterpreter() with the
PyGIL_StateXXX functions in the current threadmodule.c).
How do we get this solved for the next Python release?

[mwhudson]

Logged In: YES
user_id=6656

I saw the subject and thought "Hmm, should raise the priority of this so it
gets at least thought about hard by the next release", then saw that you
did that...

How do we get this solved for the next Python release?

I don't know. It's possible it should be raised on python-dev (it seems
fairly quiet just now, so it might be a good time :).

Basically, it seems that the whole multiple interpreters facility is screwed,
correct?

[yetanothermax]

Logged In: YES
user_id=1218811

It looks like with Python versions > 2.3.4 threads are off limits
for multiple interpreters.
If that's your definition of screwed then I guess the answer's
yes.
I'm surprised more mod_python users haven't been bitten by
this.
Shall I raise theproblem on the list (I've already done so on
comp.lang.python a while ago)

[yetanothermax]

Logged In: YES
user_id=1218811

I've created a patch for threadmodule.c
http://sourceforge.net/tracker/index.php?
func=detail&aid=1203393&group_id=5470&atid=305470

[mwhudson]

Logged In: YES
user_id=6656

Hmm. threadtest.c just crashed for me with a debug build:

Fatal Python error: Invalid thread state for this thread
Abort trap

This is in the call to Py_NewInterpreter() ... (eek).

[yetanothermax]

Logged In: YES
user_id=1218811

Have you tried the threadtest.c with a debug build after
applying the patch to threadmodule.c?
The patch is here
http://sourceforge.net/tracker/index.php?
func=detail&aid=1203393&group_id=5470&atid=305470
Cheers,
Max

[mwhudson]

Logged In: YES
user_id=6656

(Grr, SF hates me today)

Why, are you under the impression that that could make a difference here? (hint: it can't).

The problem with Py_NewInterpreter was an over-zealous check in PythreadState_Swap(), which I've tweaked. A red herring.

On to the real problem... I see the problem, and can now reproduce it with your test code, but isn't your patch just a reversion of the patch that was attached to bug bpo-1010677? Why won't applying it bring that patch back?

[mwhudson]

Logged In: YES
user_id=6656

So here's my take on a fix to 1010677, relative to CVS HEAD.

It's something of a hack, and I certainly wouldn't condone applying it to the
trunk as is but I would be interested in hearing if it fixes your problems.

(after applying your patch the test suite failed to complete, btw -- please
try this yourself next time :)

[yetanothermax]

Logged In: YES
user_id=1218811

Good news - I've made a couple of tests and the patch seems
to fix the problem.

With regards to the earlier patch:

• apologies for not running the test suite - I feel suitably
  chastised :) & I'll be sure to run it in the future.
• I did note that it would probably re-raise bpo-1010677
  (see 2005-04-20 13:51 comment) but guessed it was
  more important to contribute the patch so that someone
  who has better thread implementation knowledge could
  look into it...

[mwhudson]

Logged In: YES
user_id=6656

That's good to hear.

Tim: could you glance over the attached? If you can't be
assed to read all the various reports, I've attempted to
summarise the problem(s) here:

http://mail.python.org/pipermail/python-dev/2005-June/054258.html

If you don't have time, please at least say so...

yetanothermax: thanks for the confirmation. Could you also
try this cleaner version of the patch?

Wrt your patch, it took me too long to work out that it was
simply unapplying the earlier changes -- and you don't need
to supply a patch for that, even creaky old CVS makes *that*
easy :)

[yetanothermax]

Logged In: YES
user_id=1218811

The maybe-even-better-1010677-fix also seems to work

[mwhudson]

Logged In: YES
user_id=6656

Fixed, in

Misc/NEWS revision 1.1301
Modules/threadmodule.c revision 2.64
Python/pystate.c revision 2.42

(I made a tiny tweak in addition to the latest patch
attached to this report).

yetanothermax: Thanks for the report, and the testing of
patches!

[alephant]

Logged In: YES
user_id=1400247

This is apparantly fixed in 2.4.2 despite the absence of an
entry in the NEWS for that release.