GeezerVison

Reject exported close calls before taking the data lock

When one thread passes this mmap to _SSLSocket.write, crates/stdlib/src/openssl.rs:3073-3104 holds the mmap guard via data.borrow_buf() across a potentially unbounded allow_threads call. A concurrent mapped.close() now blocks here instead of observing the already-nonzero export count and raising BufferError; because this mutex wait remains attached to the interpreter, it can also participate in a stop-the-world deadlock. Keep an unlocked export-count fast path before this acquisition and recheck under the lock to preserve the race fix; check_resizeable needs the same ordering.

Useful? React with 👍 / 👎.

Thanks, you're right. Fixed in aa7ac19. close() and check_resizeable() now reject existing exports before acquiring the mapping mutex, while retaining the check under the lock to prevent invalidating a concurrent export.

Both new regressions hold an actual borrowed mmap buffer across the operation. They failed before this change and now pass on Linux and Windows.

Remove new assertions disallowed by the test policy

This block adds new test setup, assertions, and data to a test file, but the repository's critical test policy permits test-file edits only for adding or removing expected-failure markers (and their associated TODO comments). The added snippet therefore needs to be removed or the repository policy changed before landing.

AGENTS.md reference: AGENTS.md:L273-L279

Useful? React with 👍 / 👎.