Segmentation fault when using class derived from Expression with multiple threads

[James Maddison]

The following example leads to a segmentation fault:

#!/usr/bin/env python
from dolfin import *
parameters["num_threads"] = 2
class TestExpression(Expression):
    def eval(self, value, x):
      value[:] = 1.0
      return
    def value_shape(self):
      return tuple()
mesh = Rectangle(0.0, 0.0, 1.0, 1.0, 10, 10)
space = FunctionSpace(mesh, "CG", 1)
test = TestFunction(space)
trial = TrialFunction(space)
mass = assemble(inner(test, trial) * dx)
rhs = assemble(inner(test, TestExpression()) * dx)
F = Function(space)
solve(mass, F.vector(), rhs)
print assemble(F * dx)

If TestExpression() is replaced with Constant(1.0) or Expression("1.0") then the test runs without error.

[Lawrence Mitchell (wence)]

Probably because eval() is being called from inside a threaded region, but doesn't grab the python GIL

[Johan Hake (johan-hake)]

Lawrence is right. We do not grab the GIL.

I have no experience with GIL, but a naive fix could be to add a function in the Python interface of DOLFIN to grab and release the GIL. These could then be called when threaded assemble is used.

[Johan Hake (johan-hake)]

As a workaround you can always use compiled expressions, which should be thread safe for most use cases.