`using PyCall` triggers segfault in unrelated threaded code

[marius311]

I have some code on Julia 1.3-rc4 that uses threads and does not use PyCall. I can run said code fine with JULIA_NUM_THREADS set to any number. However, if I do a using PyCall before running it, despite that this code ostensibly does not call into PyCall at all, it triggers a segfault whenever I have JULIA_NUM_THREADS set to 2 or more (although its fine with 1).

The error I get looks like:

julia> unrelated_code()
Fatal Python error: Segmentation fault

Thread 0x00002aaaaaae22c0 (most recent call first):

signal (11): Segmentation fault
in expression starting at REPL[5]:1
Segmentation fault

Here's the output I get from gdb.

Thread 3 "julia-debug" received signal SIGSEGV, Segmentation fault.
[Switching to Thread 0x2aaabda24700 (LWP 9749)]
0x00002aaaaadb02fb in jl_gc_safepoint () at /global/u1/m/marius/src/julia-1.3/src/jlapi.c:432
432         jl_gc_safepoint_(ptls);
(gdb) bt
#0  0x00002aaaaadb02fb in jl_gc_safepoint () at /global/u1/m/marius/src/julia-1.3/src/jlapi.c:432
#1  0x00002aaaaad99ddc in get_next_task (getsticky=0x2aaab87000d0) at /global/u1/m/marius/src/julia-1.3/src/partr.c:378
#2  0x00002aaaaad99ef5 in jl_task_get_next (getsticky=0x2aaab87000d0) at /global/u1/m/marius/src/julia-1.3/src/partr.c:403
#3  0x00002aaab03355fe in japi1_poptaskref_2219 (W=...) at task.jl:660
#4  0x00002aaab033603d in julia_wait_2217 () at task.jl:667
#5  0x00002aaabeae4de8 in japi1_task_done_hook_17285 (t=...) at task.jl:401
#6  0x00002aaaaad413d7 in jl_fptr_args (f=0x2aaab1aa9f70 <jl_system_image_data+13360560>, args=0x2aac48bffd88, nargs=1, m=0x2aaaba166ef0)
    at /global/u1/m/marius/src/julia-1.3/src/gf.c:1910
#7  0x00002aaaaad42070 in _jl_invoke (F=0x2aaab1aa9f70 <jl_system_image_data+13360560>, args=0x2aac48bffd88, nargs=1, mfunc=0x2aaab1aa9c30 <jl_system_image_data+13359728>, 
    world=26805) at /global/u1/m/marius/src/julia-1.3/src/gf.c:2130
#8  0x00002aaaaad42827 in jl_apply_generic (F=0x2aaab1aa9f70 <jl_system_image_data+13360560>, args=0x2aac48bffd88, nargs=1) at /global/u1/m/marius/src/julia-1.3/src/gf.c:2300
#9  0x00002aaaaad61333 in jl_apply (args=0x2aac48bffd80, nargs=2) at /global/u1/m/marius/src/julia-1.3/src/julia.h:1631
#10 0x00002aaaaad61b04 in jl_finish_task (t=0x2aaab87db0d0, resultval=0x2aaab0df0600 <jl_system_image_data+17472>) at /global/u1/m/marius/src/julia-1.3/src/task.c:181
#11 0x00002aaaaad62de6 in start_task () at /global/u1/m/marius/src/julia-1.3/src/task.c:669
#12 0x0000000000000000 in ?? ()
Backtrace stopped: Cannot access memory at address 0x2aac48c00000

Its 100% reproducible on this system, although not on others. The version of Python is 3.6.9 |Anaconda, Inc.| (default, Jul 30 2019, 19:10:26) \n[GCC 7.3.0] and PyCall is either 1.91.2 or master.

Any suggestions tracking this down further, or what I can do to fix it? Thanks.

Edit: Changed gdb output to use julia-debug.

[marius311]

I can't really reproduce this anymore.