Crash on start_vcd_trace

[kammoh]

On both linux and mac (python 3.7 and 3.8) all examples seg fault. This happens upon calling start_vcd_trace. I was able to trace down the crash to where the init callbacks in top->trace(tfp,99) is called in the C code. The crash happens in Top::traceInit: the userthis pointer seems to be corrupt! Unfortunately I was not able to get any further with pinpointing the cause.

Verilator version: master
PyVerilator: master

[acw1251]

Thanks for the report. Unfortunately I'm not able to reproduce the bug you've seen (I used python3.7 in ubuntu). Can you try running the tests in pyverilator/tests/test_pyverilator.py? You can do this by running either python3 -m pytest or python3 -m unittest in the pyverilator/tests folder. pytest should give easier to read output.

I tried running the examples using python3.7 (in the past I only used 3.5 and 3.6), and ran into issues because of the tcl communication being done with gtkwave (I think this was due to running this over SSH). I also ran into an issue with make failing. I think that was due to having a dirty build directory with some files that were created by an older version of pyverilator. If I removed the lines connecting to gtkwave and removed the build directory, the examples seemed to work (definitely no segfault).

Maybe you can also try taking simple_example.py and remove the three lines related to gtkwave (the call to sim.start_gtkwave and the two calls to sim.send_to_gtkwave)?

[kammoh]

Thank you for the quick reply! As I mentioned, this problem doesn't seem to have anything to do with either gtkwave or tcl wrappers. In addition to all the examples, the crash happens if I try to just generate a vcd trace (and this is what I actually need right now). I'm not sure if I'm doing the right thing here, but this is what I do:

sim = pyverilator.PyVerilator.build("top.v")
sim.start_vcd_trace('top.vcd') # <- seg-faults

Commenting out the lines with sim.*_gtkwave stops the crash, and I assume it's because there's no call to sim.start_gtkwave or eventually top->trace.
I also tried verilator stable as well as pyverilator versions as far back as 0.1.0-32-g3dc31d6.

Running python -m pytest in pyverilator/tests:

================================ test session starts =================================
platform linux -- Python 3.8.2, pytest-5.4.1, py-1.8.1, pluggy-0.13.1
rootdir: /home/kamyar/src/pyverilator
collected 7 items                                                                    


test_pyverilator.py Fatal Python error: Segmentation fault

Current thread 0x00007f97592a5740 (most recent call first):
  File "/home/kamyar/src/pyverilator/pyverilator/pyverilator.py", line 705 in start_vcd_trace
  File "/home/kamyar/src/pyverilator/pyverilator/tests/test_pyverilator.py", line 43 in test_pyverilator
  File "/home/kamyar/.local/lib/python3.8/site-packages/_pytest/python.py", line 184 in pytest_pyfunc_call
...
  File "/usr/lib/python3.8/runpy.py", line 86 in _run_code
  File "/usr/lib/python3.8/runpy.py", line 193 in _run_module_as_main
Segmentation fault (core dumped)

Python versions I tested inlucde : 3.8.2 and 3.7.7 both on Arch linux and macOS.

[kammoh]

Ok I was finally able to test on an Ubuntu 18.04 machine and it works without an issue.
Both python 3.6.9 and 3.8.2. gcc version is 7.5.0. Could this be a compiler option/compatibility issue?
using gcc 9.2.1 to compile both verilator and verilated binaries on the ubuntu machine also works fine. But I none of my development machines or servers are ubuntu :(

[acw1251]

I took a look in the code for things that could result in a segfault (specifically the python/c interface), and a few things stood out. I fixed all the issues I saw in 4d58d91. Can you try this out to see if you still get the segfault?

In the process of trying to reproduce the segfault, I updated verilator to master, and now it looks like some of the tests fail because of it. I'm going to look into what's going on with it.

[kammoh]

Thank you! This indeed solves the crash, and the vcd file is now correctly generated.
As you mentioned 5 of the tests (out of 14) fail on my machine, and the gtkwave example does not actually work.
One recent change I noticed in verilator is the way internal signal namespace is managed in the verilated cpp. I'm not sure if any of the failing tests depend on the previous behaviour.

[acw1251]

That's good news that the segfault doesn't happen any more. I've made issue #8 for the new failing tests. If you revert back to an older version of verilator, that should be fixed as well.

[acw1251]

Just a heads up, @kammoh , If you revert to verilator version v4.020 or before, you can get around issue #8.