HTTP/2 connections leak when reused_http_session() races with release_http_session() #1024

Describe the bug

reused_http_session() can put an h2 connection into InUsePool twice. When the connection goes idle, only one entry is removed. The other stays in InUsePool, which has no idle timer, so Pingora never closes the connection.

How it happens:

1. Task A calls reused_http_session(), which takes connection C out of InUsePool.
2. Stream B on C ends. release_http_session() can't find C, and C still has other streams, so it puts C back.
3. Task A puts C back too, without release_lock or a duplicate check. Now there are two entries.
4. When C goes idle, one entry moves to the idle pool and times out. The other keeps C open.

Relevant code:

• The insert in reused_http_session() doesn't take release_lock. release_http_session() does.
• PoolNode::insert() doesn't check for duplicates.

Pingora info

Pingora version: 0.9.0, still present on main
Rust version: 1.89
Operating system version: Debian 12

Steps to reproduce

Run test_reused_http_session_no_duplicate_in_use_entry from #1025 without the fix. It forces the race on every run.

I also reproduced it under load: 256 concurrent tasks in bursts against a local h2c upstream that never closes idle connections, with a 3 s idle_timeout.

Expected results

No connections left open after idle_timeout.

Observed results

The test finds one leftover entry in InUsePool. The load test left 4 to 10 sockets open in every run.

Additional context

#1025 fixes it by taking release_lock and removing the old entry before the insert.