⚠️ PyO3 0.23.0, 0.23.1, 0.23.2 have soundness bug in compile configuration

[davidhewitt]

I've seen recent segmentation faults from builds shipped in watchfiles and jiter, as well as unexpected errors in pydantic-core's test suite.

My current hypothesis is that the root cause is #4497, which mistakenly changed reading of PYO3_CONFIG_FILE env var such that we don't emit cargo:rerun-if-changed block for it.

As such, wheels built in bulk for many interpreters with maturin are not rebuilt properly and all contain code cfg-d for the first interpreter built for.

I will continue to verify that this is the cause, and work on a patch ASAP.

I think we will need to yank 0.23.0 through 0.23.2 and file a security advisory, as wheels built with these versions are likely highly unstable and will fail in unpredictable ways.

[davidhewitt]

fat finger submit, details incoming...

[alexdotgov2]

This "only" impacts the situation where you're doing multiple builds in the same dir (or caching it), is that right?

[ngoldbaum]

I don't think this is the root cause of issues in 0.23, but anyone who tries to set up free-threaded wheels may run into Quansight-Labs/setup-python#5 on Windows, which also affects CI environments setting up multiple python versions simultaneously. Just sharing here since others might be experimenting with building free-threaded wheels using Maturin.

[davidhewitt]

This "only" impacts the situation where you're doing multiple builds in the same dir (or caching it), is that right?

Yes, typically this would be due to using maturin build -i 3.8 3.9 3.10 3.11 etc. Projects like cryptography are likely unaffected because they're only building the one abi3 wheel.