Meta-Issue: Mutli-threading leads to segfaults #1096

Activity

1. axsk commented on Jan 22, 2026

   

   axsk

   on Jan 22, 2026

   Author

   More actions

   Seeing that the last change to this repo is older then two years, is it correct to assume it has become unmaintained and no fix is to be expected?
2. topolarity commented on Mar 10, 2026

   

   topolarity

   on Mar 10, 2026

   Last edited by topolarity

   Contributor

   More actions

   Reproducer:

   using Base.Threads
   using PyCall
   o = PyObject(randn(3,3))
   t = Threads.@spawn begin
       println(Core.stderr, propertynames(o))
   end
   wait(t)

   $ julia +1.12 --project=. mwe.jl
   
   [119772] signal 11 (1): Segmentation fault
   in expression starting at /home/topolarity/repos/claude/PyCall.jl/mwe.jl:7
   _PyType_Lookup at /usr/lib/x86_64-linux-gnu/libpython3.12.so.1.0 (unknown line)
   _PyObject_LookupSpecial at /usr/lib/x86_64-linux-gnu/libpython3.12.so.1.0 (unknown line)
   PyObject_Dir at /usr/lib/x86_64-linux-gnu/libpython3.12.so.1.0 (unknown line)

   The issue seems to be that PyCall.jl does not manage thread states or the GIL: https://docs.python.org/3/c-api/threads.html

   Here's a "fixed" version of the example:

   using Base.Threads
   using PyCall
   ccall(:PyEval_InitThreads, Cvoid, ()) # enable threads / GIL
   o = PyObject(randn(3,3))
   tstate = ccall(:PyEval_SaveThread, Ptr{Cvoid}, ()) # Release GIL
   
   t = Threads.@spawn begin
       GIL = ccall(:PyGILState_Ensure, Ptr{Cvoid}, ())
       println(Core.stderr, propertynames(o))
       ccall(:PyGILState_Release, Cvoid, (Ptr{Cvoid},), GIL)
   end
   wait(t)
   
   ccall(:PyEval_RestoreThread, Cvoid, (Ptr{Cvoid},), tstate) # Re-acquire GIL
3. topolarity commented on Mar 10, 2026

   

   topolarity

   on Mar 10, 2026

   Contributor

   More actions

   Equivalent fix in PythonCall.jl: JuliaPy/PythonCall.jl#677

   Looks like they opt to leave the GIL-management to users for most of the API. For propertynames (which is called by the REPL tab-complete), they provide a mechanism to request execution on the main thread.
4. yuyichao commented on Oct 2, 2026

   

   yuyichao

   on Oct 2, 2026

   Collaborator

   More actions

   This is now causing significant problem whenever I want to use PyPlot on a new computer. There doesn't seem to be any alternative proposed here other than #1097. Given that it is relatively low impact on the user API (I've used the rebased version for a new months now) I'd like to just merge it soon. If there's a better approach we can switch the implementation later. I don't think this would cause any more of an API breakage since even if the new/final approach require changes in the user code, those code are already broken as of now and people have already been assuming that it works.
5. yuyichao

   closed this as completedin #1097on Oct 18, 2026