numpy-wsgi-deadlock
Build
docker build -t numpy-wsgi-deadlock .
Run
docker run -p 8000:80 --name hello numpy-wsgi-deadlock
Reproduce
curl http://localhost:8000/
Debug
docker top hello
sudo gdb attach <pid-of-wsgi-process>
You will see the deadlock on thread 4:
Thread 4 (Thread 0x7fae53194700 (LWP 28580)):
#0 0x00007fae58dcc536 in futex_abstimed_wait_cancelable (private=0, abstime=0x0, expected=0, futex_word=0x55a7437040e0) at ../sysdeps/unix/sysv/linux/futex-internal.h:205
#1 do_futex_wait (sem=sem@entry=0x55a7437040e0, abstime=0x0) at sem_waitcommon.c:111
#2 0x00007fae58dcc5e4 in __new_sem_wait_slow (sem=0x55a7437040e0, abstime=0x0) at sem_waitcommon.c:181
#3 0x00007fae55064768 in PyThread_acquire_lock () from target:/usr/lib/x86_64-linux-gnu/libpython2.7.so.1.0
#4 0x00007fae54fd6556 in PyEval_RestoreThread () from target:/usr/lib/x86_64-linux-gnu/libpython2.7.so.1.0
#5 0x00007fae5508eb96 in PyGILState_Ensure () from target:/usr/lib/x86_64-linux-gnu/libpython2.7.so.1.0
#6 0x00007fae365764a6 in _error_handler (method=method@entry=0, errobj=errobj@entry=0x7fae49104758, errtype=errtype@entry=0x7fae365da9f1 "underflow", retstatus=retstatus@entry=4, first=first@entry=0x7fae53193890) at numpy/core/src/umath/ufunc_object.c:124
#7 0x00007fae3657bc6f in PyUFunc_handlefperr (errmask=521, errobj=0x7fae49104758, retstatus=retstatus@entry=4, first=first@entry=0x7fae53193890) at numpy/core/src/umath/ufunc_object.c:214
#8 0x00007fae36588de9 in double_power (a=0x7fae59753240, b=<optimized out>, modulo=0x7fae5547dd50 <_Py_NoneStruct>) at numpy/core/src/umath/scalarmath.c.src:1168
#9 0x00007fae55031a17 in ?? () from target:/usr/lib/x86_64-linux-gnu/libpython2.7.so.1.0
#10 0x00007fae54fd7e0a in PyEval_EvalFrameEx () from target:/usr/lib/x86_64-linux-gnu/libpython2.7.so.1.0
#11 0x00007fae5514615c in PyEval_EvalCodeEx () from target:/usr/lib/x86_64-linux-gnu/libpython2.7.so.1.0
#12 0x00007fae5509a5b0 in ?? () from target:/usr/lib/x86_64-linux-gnu/libpython2.7.so.1.0
#13 0x00007fae55032543 in PyObject_Call () from target:/usr/lib/x86_64-linux-gnu/libpython2.7.so.1.0
#14 0x00007fae55145587 in PyEval_CallObjectWithKeywords () from target:/usr/lib/x86_64-linux-gnu/libpython2.7.so.1.0
#15 0x00007fae5550c2d6 in ?? () from target:/usr/lib/apache2/modules/mod_wsgi.so
#16 0x00007fae5550fa7c in ?? () from target:/usr/lib/apache2/modules/mod_wsgi.so
#17 0x00007fae58dc4494 in start_thread (arg=0x7fae53194700) at pthread_create.c:333
#18 0x00007fae58b06aff in clone () at ../sysdeps/unix/sysv/linux/x86_64/clone.S:97
Cleanup
docker kill hello
docker rm hello
Workaround
Default wsgi config will create python sub interpreters which numpy does not support. Forcing wsgi to use the global interpreter will work around this issue.
WSGIApplicationGroup %{GLOBAL}
See this issue for details.
Also see WSGIApplicationGroup description on this page for details.