test_multiprocessing crashes on platforms that lack a working
sem_open(), despite it being turned off at compilation time by setting
HAVE_SEM_OPEN=0 in the Extension macros in setup.py

I think the multiprocessing module should disable the functionality
gracefully when it is missing from _multiprocessing.

Failure message:

test test_multiprocessing crashed -- <type 'exceptions.AttributeError'>:
'module' object has no attribute 'SemLock'
Traceback (most recent call last):
  File ".//Lib/test/regrtest.py", line 556, in runtest_inner
    indirect_test()
  File
"/usr/ports/lang/python/2.6/w-Python-2.6b3/Python-2.6b3/Lib/test/test_multiprocessing.py",
line 1758, in test_main
    ProcessesMixin.pool = multiprocessing.Pool(4)
  File
"/usr/ports/lang/python/2.6/w-Python-2.6b3/Python-2.6b3/Lib/multiprocessing/__init__.py",
line 226, in Pool
    return Pool(processes, initializer, initargs)
  File
"/usr/ports/lang/python/2.6/w-Python-2.6b3/Python-2.6b3/Lib/multiprocessing/pool.py",
line 84, in __init__
    self._setup_queues()
  File
"/usr/ports/lang/python/2.6/w-Python-2.6b3/Python-2.6b3/Lib/multiprocessing/pool.py",
line 130, in _setup_queues
    from .queues import SimpleQueue
  File
"/usr/ports/lang/python/2.6/w-Python-2.6b3/Python-2.6b3/Lib/multiprocessing/queues.py",
line 22, in <module>
    from multiprocessing.synchronize import Lock, BoundedSemaphore,
Semaphore, Condition
  File
"/usr/ports/lang/python/2.6/w-Python-2.6b3/Python-2.6b3/Lib/multiprocessing/synchronize.py",
line 29, in <module>
    SEM_VALUE_MAX = _multiprocessing.SemLock.SEM_VALUE_MAX
AttributeError: 'module' object has no attribute 'SemLock'
1 test failed:
    test_multiprocessing

Which platforms is this appearing on?

On Thu, 4 Sep 2008, Jesse Noller wrote:

> 
> Jesse Noller <jnoller@gmail.com> added the comment:
> 
> Which platforms is this appearing on?

OpenBSD, with this section added to setup.py:

@@ -1269,6 +1268,14 @@ class PyBuildExt(build_ext):
                 )
             libraries = []

+        elif platform.startswith('openbsd'):
+            macros = dict(                  # OpenBSD
+                HAVE_SEM_OPEN=0,            # Not implemented
+                HAVE_SEM_TIMEDWAIT=0,
+                HAVE_FD_TRANSFER=1,
+                )
+            libraries = []
+
         else:                                   # Linux and other unices
             macros = dict(
                 HAVE_SEM_OPEN=1,

2.6rc1 shows the same failure traceback on FreeBSD 6.3, which is covered
by a section of setup.py the same as Damien added for OpenBSD.

So the bug is actually in the multiprocessing module rather than the
unittest. If HAVE_SEM_OPEN is not defined then SemLock is never
built into _multiprocessing.so, but multiprocessing/syncronize.py
unconditionally depends on its presence.

I guess _multiprocessing could always define a dummy SemLock or
synchronize.py could check before it depends on it.

(it would be great to see this fixed for 2.6)

-d

Bumping up _ I'll need help with a patch

Looking at mp.synchronize, the whole module really does depend on a
working _multiprocessing.SemLock instance. If these platforms don't have
a native semaphore implementation, what is their basic inter-process
synchronisation primitive?

Once that is identified, then it should be possible to code either a C
or Python semaphore implementation based on that primitive.

For 2.6/3.0 it would probably be best to just disable the module
entirely on platforms that lack shareable semaphores (OpenBSD & FreeBSD
at least)

Jesse, how much (if any) of the rest of the package will work without
the synchronize module? If it isn't a lot, then it may be a matter of
just making this a cleaner ImportError and an expected test suite skip
on OpenBSD and FreeBSD.

(Unfortunately, our OpenBSD and FreeBSD buildbots are so unreliable that
they don't get much attention when they go red - it looks to me like the
OpenBSD buildbot isn't even managing to build _multiprocessing at the
moment, because HAVE_SEM_OPEN is incorrectly set to 1).

I've done some more digging into this for the FreeBSD case.

FreeBSD 6.3 and 7.0 both have sem_open, and the man pages suggest it 
should be fully functional. (see
http://www.freebsd.org/cgi/man.cgi?query=sem_open&apropos=0&sektion=0&manpath=FreeBSD+6.3-RELEASE&format=html)
There is a caveat on the length of the name, which I think could trigger
if the counter variable passed into SEM_CREATE() is >9999.  But as that 
variable seems like it can only ever be 0 (not sure this is intended!)
this shouldn't happen as it stands.

If I change HAVE_SEM_OPEN to 1 in setup.py, the _multiprocessing module
builds, but test_multiprocessing provokes a core dump in both cases.
The backtrace (on 6.3 i386) looks like:
#0  0x2820ef17 in ksem_open () from /lib/libc.so.6
#1  0x2820592c in sem_open () from /lib/libc.so.6
#2  0x284494a0 in semlock_new (type=0x2844b380, args=0x836443c, kwds=0x0)
    at
/home/andymac/build/python-svn/trunk-r66550/Modules/_multiprocessing/semaphore.c:439
#3  0x0809e710 in type_call (type=0x2844b380, args=0x836443c, kwds=0x0)
    at Objects/typeobject.c:731
#4  0x0806042a in PyObject_Call (func=0x2844b380, arg=0x836443c, kw=0x0)
    at Objects/abstract.c:2487
#5  0x080cc367 in PyEval_EvalFrameEx (f=0x84dd00c, throwflag=0)
    at Python/ceval.c:3890
#6  0x080cf9ac in PyEval_EvalCodeEx (co=0x8478800, globals=0x4e,
locals=0xa00,
    args=0x819b02c, argcount=4, kws=0x0, kwcount=0, defs=0x0, defcount=0,
    closure=0x0) at Python/ceval.c:2942
#7  0x08123f84 in function_call (func=0x8486ca4, arg=0x82e107c, kw=0x0)
    at Objects/funcobject.c:524
#8  0x0806042a in PyObject_Call (func=0x8486ca4, arg=0x82e107c, kw=0x0)
    at Objects/abstract.c:2487
#9  0x08069c15 in instancemethod_call (func=0x4e, arg=0x82e107c, kw=0x0)
    at Objects/classobject.c:2579
#10 0x0806042a in PyObject_Call (func=0x84857ac, arg=0x82e107c, kw=0x0)
    at Objects/abstract.c:2487
#11 0x080cc367 in PyEval_EvalFrameEx (f=0x826140c, throwflag=0)
    at Python/ceval.c:3890
---Type <return> to continue, or q <return> to quit---

On 7.0 amd64, the top of the backtrace scrolls off screen so I can't get 
the start of the trace (X not installed...).

To try and remove threads from the equation, due to FreeBSD 6.3 having 
an issue with fork() in a threaded build (see issue3864 for more info), 
I configured without threads (ie ./configure --without-threads) and 
the _multiprocessing module fails to build:

gcc -fPIC -fno-strict-aliasing -DNDEBUG -g -fwrapv -O3
-Wall-Wstrict-prototypes -DHAVE_SEM_OPEN=1 -DHAVE_FD_TRANSFER=1
-DHAVE_SEM_TIMEDWAIT=0 -IModules/_multiprocessing -I.
-I/home/andymac/build/python-svn/trunk-r66550/./Include -I. -IInclude
-I./Include -I/usr/local/include
-I/home/andymac/build/python-svn/trunk-r66550/Include
-I/home/andymac/build/python-svn/trunk-r66550 -c
/home/andymac/build/python-svn/trunk-r66550/Modules/_multiprocessing/semaphore.c
-o
build/temp.freebsd-6.3-RELEASE-i386-2.6/home/andymac/build/python-svn/trunk-r66550/Modules/_mult
iprocessing/semaphore.o
/home/andymac/build/python-svn/trunk-r66550/Modules/_multiprocessing/semaphore.c:
In function `semlock_acquire':
/home/andymac/build/python-svn/trunk-r66550/Modules/_multiprocessing/semaphore.c:314:
error: `_save' undeclared (first use in this function)
/home/andymac/build/python-svn/trunk-r66550/Modules/_multiprocessing/semaphore.c:314:
error: (Each undeclared identifier is reported only once
/home/andymac/build/python-svn/trunk-r66550/Modules/_multiprocessing/semaphore.c:314:
error: for each function it appears in.)

It appears that some support is there for a single threaded build, but
is incomplete (there's a similar problem in socket_connections.c, but
the module build bails before then).

If its not to be supported on single threaded builds (which would be a
big shame in my opinion!) then the code should make this explicit,
otherwise the single threaded build case needs to be fixed.

I'm still trying to understand the core dump in the multithreaded build
- unfortunately I'm not terribly familiar with gdb or with debugging
from cores (and the actual failure appears to be triggering in the C
library for which I currently don't appear to have symbols...)

Any suggests on how/where to dig further on this?

Oops - meant to add that the actual reported cause of the core dump is
"Bad system call".

Also, the OpenBSD man pages make clear that shared semaphores aren't
supported and sem_open() doesn't exist:
http://www.openbsd.org/cgi-bin/man.cgi?query=sem_open&apropos=0&sektion=0&manpath=OpenBSD+Current&arch=i386&format=html

I've been thinking about this - Right now, having a working mp.synchronize 
module, and thread support is key to package currently. For 2.6 - it's 
really too late to try to mock up a working mp.synchronize module, or 
significantly change the package.

My recommendation (I'm going to work on a patch to do this) is to not 
support fbsd/openbsd in this release cycle, which is unfortunate. Any 
other thoughts?

Well, even if 2.6 slipped (which it is looking probably won't happen),
how much time would you need to deal with this? Sounds like you just
won't be able to get to it even with an extra week. So that means
multiprocessing is just not supported on FreeBSD and OpenBSD at this
moment. Sucks, but hopefully it can get fixed in the future. And if
people complain, bug them to get a reliable buildbot going.

Agreed - Jesse, can you work up a patch that generates a clean import
error when _multiprocessing.SemLock can't be defined (due to
HAVE_SEM_OPEN=0 or a single-threaded build), adds test_multiprocessing
to the expected skips for FreeBSD and OpenBSD, and updates the
multiprocessing docs to note the restriction to systems with a working
sem_open() implementation?

Improving the *BSD and single-threaded build compatibility of the
multiprocessing package will just have to be high on the to-do list for
2.6.1.

This may also be worth mentioning in the release notes - Barry's call on
that one.

Hows this error look:

>>> import multiprocessing
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/Users/jesse/open_source/subversion/python-
trunk/Lib/multiprocessing/__init__.py", line 178, in <module>
    " function, see issue 3770.")
multiprocessing.SemaphoreImportError: This platform lacks a functioning 
sem_open implementation and thereforce, the required synchronization 
primitives needed will not function, see issue 3770.
>>>

looks good to me

On Mon, Sep 29, 2008 at 11:20 AM, Jesse Noller <report@bugs.python.org> wrote:
>
> Jesse Noller <jnoller@gmail.com> added the comment:
>
> Hows this error look:
>
>>>> import multiprocessing
> Traceback (most recent call last):
>  File "<stdin>", line 1, in <module>
>  File "/Users/jesse/open_source/subversion/python-
> trunk/Lib/multiprocessing/__init__.py", line 178, in <module>
>    " function, see issue 3770.")
> multiprocessing.SemaphoreImportError: This platform lacks a functioning
> sem_open implementation and thereforce, the required synchronization
> primitives needed will not function, see issue 3770.

Is "thereforce" an actual word? Otherwise it looks fine to me.

> Is "thereforce" an actual word? Otherwise it looks fine to me.
>

Yeah, I caught that. Rather than disable the entire package, which
would be frustrating to many - I've changed it to only disable
mp.synchronize for now, patch is pending my final build and test/doc
check locally

Here's a patch, works on my machine. Please review it and make sure it 
satisfies what we've spoken about.

Could use confirmation from Damien and Andrew that they now see the
expected skips with the patch applied, but otherwise looks good to me.

I can confirm that the patch works on OpenBSD -current. Only one nit:

Does this line in Lib/test/test_multiprocessing.py need to be there?

+#import multiprocessing.SemaphoreImportError

On Sep 29, 2008, at 6:36 PM, Damien Miller <report@bugs.python.org>  
wrote:

>
> Damien Miller <djmdjm@users.sourceforge.net> added the comment:
>
> I can confirm that the patch works on OpenBSD -current. Only one nit:
>
> Does this line in Lib/test/test_multiprocessing.py need to be there?
>
> +#import multiprocessing.SemaphoreImportError
>

Oops - my bad. I'll remove it and check it tonight

checked in r66688, lowering from rb to crit to address post 2.6 final

The checked in change has the planned effect on FreeBSD 6.3 i386. I
can't check on my 7.0 amd64 box as I can't quickly put a current source
tree on it.

POSIX semaphores should be fixed in 8-CURRENT, pending MFC.

There are rtld + malloc issues in FreeBSD. Python multiprocessing's
use of POSIX threads is not strictly POSIX compliant, as it tries
to do a lot more than just call exec() or async-signal-safe POSIX
APIs after fork()-ing. There is a degree of reluctance in the camp
to fix for these reasons...

In the meantime, you may wish to try building Python 2.6 on FreeBSD
using GNU Pth, here is a patch:

http://people.freebsd.org/~bms/dump/python26-fbsd-pth.patch

thanks!
BMS

Sorry, I hit the wrong bug

Closing; we've removed hard-coded platform variables for a better autoconf 
approach. We currently skip the test suite on platforms which don't 
support or have what we need.

I'm trying to follow the history of this issue, as it does not seem fully resolved to me. While trying to package something for Debian, one of the buildd [0] logs for hurd-i386 points to this issue as the cause of a failure [1]. This occurs with Python 3.7, though Debian's Python 3.7 may involve some distro patches on top of upstream source. If the build log is inaccessible or incomprehensible, the relevant sections are:

    ...
    The following additional packages will be installed:
    ... libpython3-stdlib libpython3.7-minimal libpython3.7-stdlib ... python3 python3-minimal python3.7 python3.7-minimal
    ...
    /usr/bin/ctest --force-new-ctest-process -j1
    Test project /<<PKGBUILDDIR>>/obj-i686-gnu
        Start 1: tests
    1/1 Test #1: tests ............................***Failed    0.10 sec
    Traceback (most recent call last):
      File "/usr/lib/python3.7/multiprocessing/synchronize.py", line 28, in <module>
        from _multiprocessing import SemLock, sem_unlink
    ImportError: cannot import name 'SemLock' from '_multiprocessing' (/usr/lib/python3.7/lib-dynload/_multiprocessing.cpython-37m-i386-gnu.so)

    During handling of the above exception, another exception occurred:

    Traceback (most recent call last):
      File "/<<PKGBUILDDIR>>/tests/run-tests.py", line 46, in <module>
        print_lock = multiprocessing.Lock()
      File "/usr/lib/python3.7/multiprocessing/context.py", line 66, in Lock
        from .synchronize import Lock
      File "/usr/lib/python3.7/multiprocessing/synchronize.py", line 32, in <module>
        " synchronization primitives needed will not" +
    ImportError: This platform lacks a functioning sem_open implementation, therefore, the required synchronization primitives needed will not function, see issue 3770.

Do you know what is happening here? Maybe the compounded ImportError I'm seeing *was* the resolution referred to in prior comments, but if so it seems strange to reference this issue in the error message.

The hurd-i386 platform is not a priority for me, so if the answer is "multiprocessing doesn't work there" I am perfectly OK with this. Also I realise I am asking you to recall >10 year old memories so I apologise in advance for any re-opened wounds.

If you need to see the source for what buildd is actually testing here, it's the debian/v2020.11.01-1 tag of [2].

  [0]: https://wiki.debian.org/buildd
  [1]: https://buildd.debian.org/status/fetch.php?pkg=rumur&arch=hurd-i386&ver=2020.01.11-1&stamp=1579290012&raw=0
  [2]: https://github.com/Smattr/rumur

Matthew, this tracker issue has been closed for over 10 years; posting new questions or comments to such is unlikely to get a response (I just happened to see your comment in passing).  If you believe you have identified a new issue with current versions of Python, you should search for newer issues and, if appropriate, open a new one with a reference to this one.  However, before doing that, I would do a web search on "cannot import name 'SemLock'"; you should find a number of references to situations that seem similar to yours with the cause being the lack of a working sem_open implementation as noted in the ImportError message, a condition that might just be a system configuration issue on the build system.  Good luck!