[graingert]

Summary

_CachedRequest.wrapped_receive (used by BaseHTTPMiddleware) creates a new Request.stream() async generator on every call, takes one chunk with __anext__(), and drops it while it is still suspended. Each one is then finalized by the GC through the event loop's asyncgen finalizer hook. Request.body() and the form parsers in Request._get_form() can leak the same way when parsing raises (e.g. MultiPartException for too many files) or is cancelled mid-stream.

This is the leak behind:

• Exception when calling request.json() with three or more stacked BaseHTTPMiddleware #3066 – Exception ignored in: <async_generator object Request.stream ...> from uvloop.loop.Loop._asyncgen_finalizer_hook with stacked BaseHTTPMiddlewares.
• Fix debug-mode segfault when the asyncgen finalizer hook runs during frame teardown MagicStack/uvloop#756 – the same finalizer path segfaults the interpreter under uvloop debug mode; its reproducer is two stacked BaseHTTPMiddlewares and a streaming upload.

It also provides the reproducer that #3093 was missing for the unused _wrapped_rc_stream.

Changes:

• Wrap every internal use of self.stream() (wrapped_receive, body(), both form parsers) in contextlib.aclosing. Closing a stream generator after one chunk is safe: its position lives on the request (_receive / _stream_consumed), not in the generator, and nothing runs on GeneratorExit.
• Drop the unused _wrapped_rc_stream.
• Remove the filterwarnings ignore for trio's Async generator 'starlette.requests.Request.stream' was garbage collected... ResourceWarning. With the ignore removed, 21 tests fail on main; with this change the full suite passes, so the trio test runs now guard against regressions.
• Fix tests that leaked their own async generators (unclosed receive() generators, and breaking out of request.stream()), which the stricter warning config surfaced.

User code that breaks out of async for chunk in request.stream() can still leak a generator; it may be worth a docs note recommending aclosing there as a follow-up.

Checklist

• I understand that this PR may be closed in case there was no previous discussion. (This doesn't apply to typos!)
• I've added a test for each change that was introduced, and I tried as much as possible to make a single atomic change.
• I've updated the documentation accordingly.

🤖 Generated with Claude Code

[cubic-dev-ai]

No issues found across 4 files

_{Re-trigger cubic}

[codspeed]

Merging this PR will improve performance by 63.8%

⚡ 6 improved benchmarks
✅ 236 untouched benchmarks

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
⚡	Memory	test_multipart[max-fields]	4,295 B	864 B	×5
⚡	Memory	test_multipart[max-part-size]	4.2 KB	2 KB	×2.1
⚡	Simulation	test_multipart[file-512KiB]	3.6 ms	2.8 ms	+28.41%
⚡	Simulation	test_urlencoded_form[max-part-size]	558 µs	468.5 µs	+19.1%
⚡	Simulation	test_multipart[boundary-like-file]	6.8 ms	6.1 ms	+11.68%
⚡	Simulation	test_multipart[max-part-size]	848.9 µs	770.2 µs	+10.22%

Tip

Curious why performance improved? Comment @codspeedbot explain why performance improved on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing graingert:aclose-request-stream (1a7a23f) with main (7cb2f09)}

[Kludex]

Thanks!