Switching between free-threaded and standard ABI does not invalidate build cache

[ngoldbaum]

Bug Description

See this bcrypt PR, which fixed a caching issue seen on github actions that led to a segfault. IMO this caching issue should have just led to a rebuild rather than a silently broken build.

Steps to Reproduce

I've only been able to trigger this on bcrypt's CI, but now that I understand what's happening I suspect that if you build PyO3 (or a project that depends on PyO3) with a gil-enabled interpreter and then try to rebuild the project with a free-threaded interpreter but keeping the build cache, you will get a successful build but the python extension will be broken.

Your operating system and version

MacOS

Your Python version (python --version)

3.13.1

Your Rust version (rustc --version)

1.84.0

Your PyO3 version

0.23.4

How did you install python? Did you use a virtualenv?

github actions setup-python action

Additional Info

No response

[ngoldbaum]

Managed to trigger this locally. Here are steps to reproduce the bug. This assumes you have a clone of the bcrypt repo and that you have 3.13.1 and 3.13.1t installed via pyenv and nox is installed into both interpreter's global environments.

pyenv global 3.13.1
cd bcrypt
nox -fb virtualenv -s tests
rm -r .nox
pyenv global 3.13.1t
nox -fb virtualenv -s tests

The second invocation uses a free-threaded interpreter, but it re-uses the target directory from the GIL-enabled build produced by the previous nox run

You should see a segmentation fault inside free-threaded Python trying to import bcrypt. This doesn't work because the bcrypt module was compiled with the GIL-enabled ABI.

collecting ... Fatal Python error: Segmentation fault

Current thread 0x00000001f6dfc240 (most recent call first):
  File "<frozen importlib._bootstrap>", line 488 in _call_with_frames_removed
  File "<frozen importlib._bootstrap_external>", line 1320 in create_module
  File "<frozen importlib._bootstrap>", line 813 in module_from_spec
  File "<frozen importlib._bootstrap>", line 921 in _load_unlocked
  File "<frozen importlib._bootstrap>", line 1331 in _find_and_load_unlocked
  File "<frozen importlib._bootstrap>", line 1360 in _find_and_load
  File "/Users/goldbaum/Documents/bcrypt/.nox/tests/lib/python3.13t/site-packages/bcrypt/__init__.py", line 13 in <module>
  File "<frozen importlib._bootstrap>", line 488 in _call_with_frames_removed
  File "<frozen importlib._bootstrap_external>", line 1026 in exec_module
  File "<frozen importlib._bootstrap>", line 935 in _load_unlocked
  File "<frozen importlib._bootstrap>", line 1331 in _find_and_load_unlocked
  File "<frozen importlib._bootstrap>", line 1360 in _find_and_load
  File "/Users/goldbaum/Documents/bcrypt/tests/test_bcrypt.py", line 6 in <module>
  File "/Users/goldbaum/Documents/bcrypt/.nox/tests/lib/python3.13t/site-packages/_pytest/assertion/rewrite.py", line 184 in exec_module
  File "<frozen importlib._bootstrap>", line 935 in _load_unlocked
  File "<frozen importlib._bootstrap>", line 1331 in _find_and_load_unlocked
  File "<frozen importlib._bootstrap>", line 1360 in _find_and_load
  File "<frozen importlib._bootstrap>", line 1387 in _gcd_import
  File "/Users/goldbaum/.pyenv/versions/3.13.1t/lib/python3.13t/importlib/__init__.py", line 88 in import_module

[ngoldbaum]

I tried and failed to fix this today following a suggestion from @davidhewitt: try adding cargo:rerun-if-changed for the path to the python executable. That way if python in a virtualenv gets replaced, the build gets triggered again.

For some reason I don't understand, that doesn't work. See the fix-python-build-cache branch on my fork. I tried adding it to the build.rs for both pyo3-build-config and pyo3-ffi, but for whatever reason cargo seems to think both crates are fresh when I try to compile after modifying the symlink in the virtualenv.

One odd thing is that if I use touch to modify the mtime of the python executable it does work!

Take a look at this copy/paste from a terminal session where I demonstrate this behavior: https://gist.github.com/ngoldbaum/10ec27e807e6b1668890e31afaa8da77

I had to hack the noxfile.py for bcrypt like this to be able to see the debug output from pip:

diff --git a/noxfile.py b/noxfile.py
index 4b1cd65..f83204a 100644
--- a/noxfile.py
+++ b/noxfile.py
@@ -7,7 +7,7 @@ nox.options.default_venv_backend = "uv|virtualenv"
 @nox.session
 def tests(session: nox.Session) -> None:
     session.install("coverage")
-    session.install(".[tests]")
+    session.install(".[tests]", "-v")

     session.run(
         "coverage", "run", "-m", "pytest", "--strict-markers", *session.posargs

Maybe this is a cargo caching bug? Deleting the file and recreating it the file with different content somehow doesn't trigger the rerun-if-changed check. It looks like checking for dirty builds based on hashes is in cargo but not yet stabilized, so we can't use that as an escape hatch.

[ngoldbaum]

I can trigger the same behavior in a test Rust project, so I think this is a cargo bug or a shortcoming of the rerun-if-changed feature. I filed an issue: rust-lang/cargo#15134

[ngoldbaum]

Until the upstream issue is fixed I think I have at least a way to generate a nicer error message.

If a host config file is found, instead of just blindly trusting it, we could call get_host_interpreter again and make sure the interpreter matches what's in the config file. If it doesn't, print an error suggesting setting e.g. PYO3_ENVIRONMENT_SIGNATURE and refuse to continue.

We could restrict the check for things we know definitely won't work, like switching from a GIL-disabled to GIL-enabled build or vice verse.

[ngoldbaum]

Argh, no, that doesn't work because the build script would need to run a second time, which of course it can't unless cargo tells it to.

Maybe this is just something we need to document for now as a possible issue people might run into?