Python's subprocess module has a race condition: Popen() constructor has a call to global "_cleanup()" function on whenever a Popen object gets created, and that call causes a check for all pending Popen objects whether their subprocess has exited - i.e. the poll() method is called for every active Popen object.

Now, if I create Popen object "foo" in one thread, and then a.wait(), and meanwhile I create another Popen object "bar" in another thread, then a.poll() might get called by _clean() right at the time when my first thread is running a.wait(). But those are not synchronized - each calls os.waitpid() if returncode is None, but the section which checks returncode and calls os.waitpid() is not protected as a critical section should be.

The following code illustrates the problem, and is known to break reasonably consistenly with Python2.4. Changes to subprocess in Python2.5 seems to address a somewhat related problem, but not this one.

import sys, os, threading, subprocess, time

class X(threading.Thread):
  def __init__(self, *args, **kwargs):
    super(X, self).__init__(*args, **kwargs)
    self.start()

def tt():
  s = subprocess.Popen(("/bin/ls", "-a", "/tmp"), stdout=subprocess.PIPE,
      universal_newlines=True)
  # time.sleep(1)
  s.communicate()[0]

for i in xrange(1000):
  X(target = tt)

This typically gives a few dozen errors like these:
Exception in thread Thread-795:
Traceback (most recent call last):
  File "/usr/lib/python2.4/threading.py", line 442, in __bootstrap
    self.run()
  File "/usr/lib/python2.4/threading.py", line 422, in run
    self.__target(*self.__args, **self.__kwargs)
  File "z.py", line 21, in tt
    s.communicate()[0]
  File "/usr/lib/python2.4/subprocess.py", line 1083, in communicate
    self.wait()
  File "/usr/lib/python2.4/subprocess.py", line 1007, in wait
    pid, sts = os.waitpid(self.pid, 0)
OSError: [Errno 10] No child processes

Note that uncommenting time.sleep(1) fixes the problem. So does wrapping subprocess.poll() and wait() with a lock. So does removing the call to global _cleanup() in Popen constructor.

Peter, could you take a look at this?  Thanks.

It appears that subprocess calls a module global "_cleanup()" whenever opening a new subprocess. This method is meant to reap any child processes that have terminated and have not explicitly cleaned up. These are processes you would expect to be cleaned up by GC, however, subprocess keeps a list of of all spawned subprocesses in _active until they are reaped explicitly so it can cleanup any that nolonger referenced anywhere else.

The problem is lots of methods, including poll() and wait(), check self.returncode and then modify it. Any non-atomic read/modify action is inherently non-threadsafe. And _cleanup() calls poll() on all un-reaped child processes. If two threads happen to try and spawn subprocesses at once, these _cleanup() calls collide..

The way to fix this depends on how thread-safe you want to make it. If you want to share popen objects between threads to wait()/poll() with impunity from any thread, you should add a recursive lock attribute to the Popen instance and have it lock/release it at the start/end of every method call.

If you only care about using popen objects in one thread at a time, then all you need to fix is the nasty "every popen created calls poll() on every other living popen object regardless of what thread started them,
and poll() is not threadsafe" behaviour.

Removing _cleanup() is one way, but it will then not reap child processes that you del'ed all references to (except the one in subprocess._active) before you checked they were done.

Probably another good idea is to not append and remove popen objects to _active directly, instead and add a popen.__del__() method that defers GC'ing of non-finished popen objects by adding them to _active. This
way, _active only contains un-reaped child processes that were due to be GC'ed. _cleanup() will then be responsible for polling and removing these popen objects from _active when they are done.

However, this alone will not fix things because you are still calling _cleanup() from different threads, it is still calling poll() on all these un-reaped processes, and poll() is not threadsafe. So... you either have to make poll() threadsafe (lock/unlock at the beginning/end of the method), or make _cleanup() threadsafe. The reason you can get away with making only _cleanup() threadsafe this way is _active will contain a list of processes that are not referenced anywhere else, so you know the only thing that will call poll() on them is the _cleanup() method.

Having just gone through that waffly description of the problems and various levels of fix, I think there are really only two fixes worth considering;

1) Make Popen instances fully threadsafe. Give them a recursive lock attribute and have every method acquire the lock at the start, and release it at the end.

2) Decide the "try to reap abandoned children at each Popen" idea was an ugly hack and abandon it. Remove _active and _cleanup(), and document that any child process not explicitly handled to completion will result in zombie child processes.

I like #2.  I don't see any use for threadsafe Popen instances, and I think that any self-respecting long-running server using Popen should properly manage its subprocesses anyway.  (And for short-running processes it doesn't really matter if you have a few zombies.)

One could add a __del__ method that registers zombies to be reaped later, but I don't think it's worth it, and __del__ has some serious issues of its own.  (If you really want to do this, use a weak reference callback instead of __del__ to do the zombie registration.)

Suddenly starting to leave Zombies is NOT an option for me. To prevent zombies, all applications using the subprocess module would need to be changed. This is a major breakage of backwards compatibility, IMHO. subprocess (as well as the popen2 module) has prevented zombies automatically from the beginning and I believe they should continue to do so (at least by default). 

A little bit of history: When I wrote the subprocess module, I stole the idea of calling _cleanup whenever a new instance is created from the popen2 module, since I thought it was nice, lightweight and avoided having a __del__ method (which I have some bad experience with, GC-related). This worked great for a long time. At 2006-04-10, however, martin.v.loewis committed patch 1467770 (revision r45234), to solve bug 1460493. It introduces a __del__ method and some more complexity. I must admit that I didn't review this patch in detail, but it looked like thread safeness was being taken care of. But apparently it isn't. 

Perhaps reverting to the original popen2 approach would solve the problem? Or does the popen2 module suffer from this bug as well? 

I think that we need to fix this once and for all, this time. We should probably also consider adding the option of not having subprocess auto-wait while we are at it, for those what would like to wait() themselves (there should be a tracker item for this, but I can't find it right now). 

Are we tight on time for fixing this bug? I'm out in the forest right now...

I just looked at popen2 and it has the same bug. I don't think __del__() related changes have anything to do with this. Popen.poll() and _cleanup() are non-threadsafe. When __init__() is called to create subprocesses simultaniously in two different threads, they both call _cleanup() and trample all over each other.

Removing _cleanup() will not leave zombies for popened processes that are correctly handled to completion. Using methods like communicate() and wait() will handle the process to completion and reap the child. 

Unfortunately, I think a fairly common use-case is people using popen to fire-and-forget subprocesses without using communicate() or wait(). These will not get reaped, and will end up as zombies.

I cannot think of an easy way to reap abandoned popen instances that is thread safe without using locks. At times like this I wish that the GIL was exposed as a recursive lock... we could have __cleanup__() acquire/release the GIL and make it atomic... no more race condition :-)

Actually, after thinking about it, there may be a way to make _cleanup() threadsafe without using explicit locks by leveraging off the GIL. If _active.pop() and _active.append() are atomic because of the GIL, then _cleanup() can be made threadsafe. To be truely threadsafe it also needs to make sure that _active contains only abandoned popen instances so that _cleanup() doesn't have it's inst.poll() calls collide with other threads that are still using those popen instances. This can be done by havin the popen instances added to _active only by Popen.__del__(), and only removed by _cleanup() when they are done.

If you want this fixed in 2.5.x, a new release is just around the corner, and time is very tight.  Otherwise, 2.6a1 is half a year off.

I just looked at subprocess in svs trunk and it looks like it might already be fixed in both subprocess.py and popen2.py. The part where __del__() adds abandoned Popen instances to _active and _cleanup() removes them is already there. _cleanup() has been made more thread resistant by catching and ignoring exceptions that arise when two _cleanups() try to remove the same instances. Popen.poll() has been made more robust by having it set self.returncode to an optional _deadstate argument value when poll gets an os.error from os.pidwait() on a child that gets cleaned up by another thread. I think there are still a few minor race conditions around Popen.poll(), but it will only affect what non-zero returncode you get... it's not so much "thread safe" as "thread robust".

I think the _deadstate argument approach to try and make Popen.poll() thread-robust is a bit hacky. I would rather see _cleanup() be properly threadsafe by having it remove the inst from _active before processing it and re-adding it back if it is still not finished. However, as it is now it looks like it is fixed...

I can create a patch to make current head a bit cleaner, if anyone is interested...

>I can create a patch to make current head a bit cleaner, if anyone is
>interested...

Sure, this would be great. 

I would also love a test case that reproduces this problem. 

There should be test cases aplenty - this bug was first reported in March 2006 and this is one of four incarnations of it. See also 1753891, 1754642, 1183780

I think all of these contain testcases (though two roughly coincide) and any fix would be wise to try them all out...

See http://bugs.python.org/issue1236 for a good repeatable testcase.

This or some variant also shows up with scons
(http://scons.tigris.org/issues/show_bug.cgi?id=1839) leading to some
nasty intermittent build failures.  Neal may remember how we addressed
this for a Process class in a past life.  Basically it's OK to collect
all the child exit codes if you record the results and return them when
requested. This would mean that waitpid and the like would have to check
a cached list of PIDs and exit statuses before actually waiting.  Don't
know if that's possible/any help in this case or not...

Traceback (most recent call last):
  File "/usr/lib/scons-0.97.0d20070918/SCons/Taskmaster.py", line 194,
in execute
    self.targets[0].build()
  File "/usr/lib/scons-0.97.0d20070918/SCons/Node/__init__.py", line
365, in build
    stat = apply(executor, (self,), kw)
  File "/usr/lib/scons-0.97.0d20070918/SCons/Executor.py", line 139, in
__call__
    return self.do_execute(target, kw)
  File "/usr/lib/scons-0.97.0d20070918/SCons/Executor.py", line 129, in
do_execute
    status = apply(act, (self.targets, self.sources, env), kw)
  File "/usr/lib/scons-0.97.0d20070918/SCons/Action.py", line 332, in
__call__
    stat = self.execute(target, source, env)
  File "/usr/lib/scons-0.97.0d20070918/SCons/Action.py", line 479, in
execute
    result = spawn(shell, escape, cmd_line[0], cmd_line, ENV)
  File "/usr/lib/scons-0.97.0d20070918/SCons/Platform/posix.py", line
104, in spawnvpe_spawn
    return exec_spawnvpe([sh, '-c', string.join(args)], env)
  File "/usr/lib/scons-0.97.0d20070918/SCons/Platform/posix.py", line
68, in exec_spawnvpe
    stat = os.spawnvpe(os.P_WAIT, l[0], l, env)
  File "/sw/pkgs/python-2.3.5_00/lib/python2.3/os.py", line 553, in spawnvpe
    return _spawnvef(mode, file, args, env, execvpe)
  File "/sw/pkgs/python-2.3.5_00/lib/python2.3/os.py", line 504, in
_spawnvef
    wpid, sts = waitpid(pid, 0)
OSError: [Errno 10] No child processes
scons: building terminated because of errors.

Looking at the subprocess.py code it occurred to me that it never checks
if the value of self.pid returned by os.fork is -1, this would mean that
later it runs waitpid with -1 as the first argument, putting it into
promiscuous (wait for any process in the group) mode.  This seems like a
bad idea...

> Looking at the subprocess.py code it occurred to me that it never
> checks if the value of self.pid returned by os.fork is -1

What makes you think os.fork(0 can return -1? It's not C you know...

Good question.  The documentation I was reading was mute on the subject 
so I made a "reasonable" guess.  Does it throw an exception instead?

Guido van Rossum wrote:
> Guido van Rossum added the comment:
>
>   
>> Looking at the subprocess.py code it occurred to me that it never
>> checks if the value of self.pid returned by os.fork is -1
>>     
>
> What makes you think os.fork(0 can return -1? It's not C you know...
>
> _____________________________________
> Tracker <report@bugs.python.org>
> <http://bugs.python.org/issue1731717>
> _____________________________________
>

Yes, like all system calls in the os module.

"""Basically it's OK to collect
all the child exit codes if you record the results and return them when
requested. This would mean that waitpid and the like would have to check
a cached list of PIDs and exit statuses before actually waiting."""

note that this would still have problems.  PIDs are recycled by the OS
so as soon as you've waited on one and the process dies, the OS is free
to launch a new process using it.  If the new process happens to be
another one of ours launched by subprocess that would be a problem. 
Make the cached map of pids -> exit codes be a map of pids -> [list of
exit codes] instead?

I am unable to reproduce this problem in today's trunk (2.6a) on OSX 10.4.

I'm not sure what the POSIX standards say on this (and MS-Windows may go 
it's own contrary way), but for most real systems the PID is a running 
count (generally 32 bit or larger today) which would have to cycle all 
the way around to repeat.  It's actually done this way on purpose to 
provide the longest possible time between IDs getting reused.  I suspect 
that having it cycle and reuse an ID isn't an issue in practice, and 
keeping a list of results leaves you with the problem of figuring out 
which PID 55367 they're talking about...  Not to mention that if you're 
leaving child process results unchecked for long enough for the PID 
counter to cycle, there are other problems with the application. ;-)

Gregory P. Smith wrote:
> Gregory P. Smith <greg@krypto.org> added the comment:
>
> """Basically it's OK to collect
> all the child exit codes if you record the results and return them when
> requested. This would mean that waitpid and the like would have to check
> a cached list of PIDs and exit statuses before actually waiting."""
>
> note that this would still have problems.  PIDs are recycled by the OS
> so as soon as you've waited on one and the process dies, the OS is free
> to launch a new process using it.  If the new process happens to be
> another one of ours launched by subprocess that would be a problem. 
> Make the cached map of pids -> exit codes be a map of pids -> [list of
> exit codes] instead?
>
> _____________________________________
> Tracker <report@bugs.python.org>
> <http://bugs.python.org/issue1731717>
> _____________________________________
>

Which system uses a 32 bit PID?  Not Linux, or FreeBSD, or OS X - they
all use 16 bit PIDs.

AIX, HP-UX, Solaris, 64 bit Linux, ...  Even in the Linux x86 header 
files there's a mix of int and short.  The last time I had to do the 
math on how long it would take the PID to cycle was probably on an AIX 
box and it was a very long time.

Jean-Paul Calderone wrote:
> Jean-Paul Calderone <exarkun@divmod.com> added the comment:
>
> Which system uses a 32 bit PID?  Not Linux, or FreeBSD, or OS X - they
> all use 16 bit PIDs.
>
> ----------
> nosy: +exarkun
>
> _____________________________________
> Tracker <report@bugs.python.org>
> <http://bugs.python.org/issue1731717>
> _____________________________________
>

PIDs cycle quite frequently.  You must be thinking of something else. 
Even 64 bit Linux has a maximum PID of 2 ** 15 - 1.

fwiw, I see pids cycle in a reasonable amount of time on modern
production linux systems today.  its a fact of life, we should deal with
it. :)

Bad design stays bad design: the hope that pids don't get reused soon
breaks two assumptions:
1) I don't have to wait() for a child process soon. It's the programs
business.
2) Pids cycle: there are security patches to make pids of future
processes hard to predict by assigning random free pids.

I actually was about to report a bug on subprocess when I bumped into
this one. My problem is pretty thread-less but related:
How do I kill() a process if it's pid got recycled behind my back??

In my opinion the module should work the "Python" way of doing things,
otherwise programmers will be surprised, as I was:
a) don't do my work (no wait() for things I care of/have reference to)
b) do your work (clean up things I don't care of/have no reference to)

If that's not possible, how does subprocess actually "intends to replace
os.spawn*"? [Library Reference]

It is still possible to extend the Popen() call to reflect if the caller
wants a P_NOWAITO or P_NOWAIT behavior, but the closer we get to the os,
the less sense does it make to replace some other modules...

Update for my comment: python2.5.2 does not show that problem.
python2.4.5 did. I repeat, 2.5.2 does not clean up processes which are
still referenced, but it does clean unreferenced ones.

This popped up on the mailing list today.

Please give this bug a proper review. I *believe* it's fixed but I'm not
sure.

Any more information on this?

I had this happen to me today, I used Popen and os.waitpid(p.pid, 0) and
I got an "OSError: [Errno 10] No child processes". I haven't tried
adding a sleep, since it's unreliable. I'm using python 2.5.2, Windows
XP. I haven't tried this on linux yet as the problem arose while testing
my app for windows.

I always get a subprocess error when using embedded python 2.6.2:

File "/usr/lib/python2.6/subprocess.py", line 1123, in wait:     pid,
sts = os.waitpid(self.pid, 0): OSError: [Errno 10] No child processes

Example library and main program are attached.

I have confirmed that none of the original test cases referenced here or
in the referenced issues fails on python 2.5.4, 2.6.2, trunk, or py3k on
linux.

The test case attached to this ticket I cannot test as I don't
understand how one would run it (and exim appears to be involved).

Given the discussion (specifically msg32219) I think this ticket should
be closed as fixed; unless, Yonas, you can provide a reproducible test
case that does not involve third party software, or someone else can
reproduce your exim case.

If someone wants to propose a patch to refactor subprocess, IMO that
should be opened as a new feature request ticket.

To test with exim4 is easy. To reproduce on Ubuntu Jaunty:

1. apt-get install exim4-daemon-heavy

2. echo "local_scan = /usr/lib/exim4/local_scan/libpyexim.so" >
/etc/exim4/conf.d/main/15_py-exim_plugin_path

3. cd /usr/lib/exim4/local_scan

4. Compile mylib, output will be "libpyexim.so":

gcc `python2.6-config --cflags` -c -fPIC mylib.c -o mylib.o 
gcc -Xlinker -export-dynamic -lpython2.6 -lnsl -lpthread -ldl -lutil -lm
-lz -shared -Wl,-soname,libpyexim.so -o libpyexim.so  mylib.o

5. Restart server:
/etc/init.d/exim4 restart

6. Send some mail:
cat mail.txt | sendmail -t

(contents of mail.txt):
Content-type: text/plain
To: your_username@localhost
From: foobar@example.com
Subject: test

Hello world.

Also, copy exim_local_scan2.py to /usr/lib/python2.6/

I'm assuming that exim4 is reading config files from /etc/exim4/conf.d/*.

To make sure, you can enforce "split file mode" by running `sudo
dpkg-reconfigure exim4-config`. It should be one of the last questions
present to you.

I'm afraid I'm not going to be installing exim in order to test this. 
Perhaps someone else will be interested enough to do so, perhaps not :)

Copying files into /usr/lib/pythonx.x is a very odd thing to do, by the
way (though it should have no impact on the bug).