tomas16
commented
9 months ago I figured out the problem: when scene goes out of scope in the python function hit_test, the underlying C++ objects are cleaned up. However, execution of the kernel is triggered by the np.asarray() call in main, so by the time the kernel runs the memory for scene has been freed.
There are 2 workarounds that I've tested successfully:
- Put the
np.asarraycall inside thehit_testfunction, so the last line becomesreturn np.asarray(was_hit) - Create the
sceneobject inmainand pass it in as an argument tohit_test
I'm not familiar enough with the code to quickly fix this myself. However if kernels keep a reference to their dependencies, it may be a simple matter of incrementing the refcount of the scene object there and decrementing it in the kernel's destructor.
rtabbara
This issue replaces #204.
Save the following code as
segfault.py:Run it with AddressSanitizer:
It looks like mitsuba is calling embree for intersection testing while the whole embree acceleration structure has already been deallocated by a different thread. Note the heap-use-after-free doesn't always occur on the exact same line, which doesn't matter as the real issue seems to be related to erroneously decrementing the refcount of some (all?) variables.
For smaller wavefront sizes, the heap-use-after-free doesn't always occur.
If you run this without AddressSanitizer, it's likely you will get the correct result. However this is all derived from issue #204, where the behavior is one of (segfault, correct result, incorrect result). In all 3 cases though, AddressSanitizer catches this heap-use-after-free. So while this issue doesn't cause this toy example to fail, it does cause trouble in real code.
Versions