homepage

I got a report that one of the tests for processing
(http://cheeseshop.python.org/pypi/processing) was failing with
 Fatal Python error: Invalid thread state for this thread
when run with a debug interpreter. This appears to be caused by the
interaction between os.fork() and threads. The following attached
program reliably reproduces the problem for me on Ubuntu edgy for x86.
All that happens is that the parent process starts a subthread then
forks, and then the child process starts a subthread.
With the normal interpreter I get
 started thread -1211028576 in process 18683
 I am thread -1211028576 in process 18683
 started thread -1211028576 in process 18685
 I am thread -1211028576 in process 18685
as expected, but with the debug interpreter I get
 started thread -1210782816 in process 18687
 I am thread -1210782816 in process 18687
 started thread -1210782816 in process 18689
 Fatal Python error: Invalid thread state for this thread
 [5817 refs]
Notice that the child process is reusing a thread id that was 
being used by the parent process at the time of the fork.
The code raising the error seems to be in pystate.c:
 PyThreadState *
 PyThreadState_Swap(PyThreadState *new)
 {
 PyThreadState *old = _PyThreadState_Current;
 _PyThreadState_Current = new;
 /* It should not be possible for more than one thread state
 to be used for a thread. Check this the best we can in
debug 
 builds.
 */
 #if defined(Py_DEBUG) && defined(WITH_THREAD)
 if (new) {
 PyThreadState *check = PyGILState_GetThisThreadState();
 if (check && check != new)
 Py_FatalError("Invalid thread state for this
thread");
 }
 #endif
 return old;
 }
It looks as though PyGILState_GetThisThreadState() is returning the
thread state of the thread from the parent process which has the same
id as the current thread. Therefore the check fails.
I think the thread local storage implementation in thread.c should
provide a function _PyThread_ReInitTLS() which PyOS_AfterFork() can
call. I think _PyThread_ReInitTLS() just needs to remove and free
each "struct key" in the linked list which does not match the id of
the calling thread.

The included patch against python2.51 fixes the problem for me.

Bug day task

we need this in before 2.6 is released.

Gregory, go ahead and apply and see if can stop the hell in the buildbots.

Updated version of roudkerk's patch. Adds the new function to
pythread.h and is based off of current trunk.
Note that Parser/intrcheck.c isn't used on my box, so it's completely
untested.
roudkerk's original analysis is correct. The TLS is never informed that
the old thread is gone, so when it seems the same id again it assumes it
is the old thread, which PyThreadState_Swap doesn't like.

Incidentally, it doesn't seem necessary to reinitialize the lock. Posix
duplicates the lock, so if you hold it when you fork your child will be
able to unlock it and use it as normal. Maybe there's some non-Posix
behaviour or something even more obscure when #401226 was done?
(reinitializing is essentially harmless though, so in no way should this
hold up release.)

I applied this in r64212.

The fix is required to run multiprocessing on Python 2.4 and 2.5, see
#4451. I suggest we fix the issue in 2.5.3. The fork-thread-patch-2
patch doesn't work on Python 2.5. I'm getting a segfault on my system:
test_connection (multiprocessing.tests.WithProcessesTestConnection) ...
Program received signal SIGSEGV, Segmentation fault.
[Switching to Thread 0x7fa2e999f6e0 (LWP 10594)]
0x000000000052065f in PyCFunction_Call (func=Cannot access memory at
address 0x7ffeffffffd8
) at Objects/methodobject.c:73
73 return (*meth)(self, arg);
Linux on AMD64 with Python 2.5 svn --with-pydebug.

I was wrong and the patch is right. Something is wrong in
multiprocessings connection_recvbytes_into() function for the old buffer
protocol. Somehow PyArg_ParseTuple(args, "w#|" ...) fucks up the
process' memory.
Martin, are you fine with the patch? fork-thread-patch-2 still applies
with some fuzzying.

Looks fine to me, somebody please backport it.
I'm concerned about the memory leak that this may cause, though (IIUC):
all the values are garbage, provided that they are pointers to dynamic
memory in the first place.
I think in the long run, we should allow to store a pointer to a release
function along with actual value to release.

Since this is a Python 2.5.3 issue, I'm lowering to deferred blocker
until after 3.0 and 2.6.1 are released.

Committed fork-thread-patch-2 as r67736 into 2.5 branch.