audio: fix OpenAL buffer leak (~26 MB/min) — alDeleteBuffers never succeeded

[marcoThede]

The problem

Long sessions kept growing until the machine started swapping. Measured on
macOS/arm64: +25.7 MB/min, perfectly linear over 97 samples, no plateau.
After roughly two hours the process sat at ~5 GB and the system was thrashing
(2.2% CPU, 5.1M pageouts on an 8 GB machine).

97% of the heap growth came from a single allocator:

GameEngine::update()
└─ OpenALAudioManager::update()
   └─ processPlayingList()
      └─ notifyOfAudioCompletion()
         └─ startNextLoop()
            └─ playSample3D()
               └─ OpenALAudioFileCache::getBufferForFile()
                  └─ decodeFFmpeg()
                     └─ alBufferData → alBufferStorageDirectSOFT

Root cause

releaseOpenAudioFile() only detached the buffer from playing sources when
m_openCount > 0. But the cache reference is dropped (closeBuffer) before
the still-playing source is handed a new buffer — so a buffer with
m_openCount == 0 can very much still be bound to a live OpenAL source.
alDeleteBuffers() then fails with AL_INVALID_OPERATION, and nobody checks
the return value.

Instrumenting alGetError() right after the call made it unambiguous:

before:  0 of 1158 alDeleteBuffers calls succeeded
after:   880 of 889 succeeded

Not a single audio buffer was ever freed for the lifetime of the process.

Result

	before	after
successful buffer deletions	0 / 1158	880 / 889
memory growth	25.7 MB/min	1.45 MB/min
per hour	~1.6 GB	~87 MB

Measured over 11 minutes, then re-confirmed on the committed tree. The
remainder is ordinary cache warm-up — the footprint now also drops again as
the cache evicts, which it never did before.

The commits

1. account for decoded PCM size — decodeFFmpeg() accumulates the decoded
   size into m_fileSize, but getBufferForFile() immediately overwrote it
   with file->size() (the compressed size), so the cache under-reported its
   own footprint. Explicitly not the cause of the leak — verified by
   applying it alone, which changed nothing (+1392 buffers/3 min vs +1451).
   Wrong accounting regardless.
2. fix OpenAL buffer leak — the actual fix described above.
3. skip cache lookup for audio events without a filename — TurretMoveLoop
   and RadarEvent ship with an empty Sounds = in the retail INIZH.big.
   TurretMoveLoop is loop all random, so it regenerated the empty filename and
   walked down to TheFileSystem->openFile("") every single frame. Nothing
   leaked, but up to 12813 futile calls in five minutes against 2948 real ones.

Commit 1 is the weakest of the three and can be dropped without affecting the
fix, if you prefer to keep this tight.

Reproducing

No gameplay required — the main menu is enough, since the menu music loops
and every loop re-decodes a sample. Watch GeneralsXZH in Activity Monitor, or:

footprint -p <pid>      # not `ps -o rss` — heap/leaks suspend the process and
                        # RSS then drops by hundreds of MB without any real free

Note that leaks reports nothing useful here (only ~18 KB of system XPC
cycles): the buffers stay reachable, so this is a cache that never drains
rather than a classic leak.

Scope

Core/GameEngineDevice/Source/OpenALAudioDevice/ is shared, so this affects
Linux as well as macOS. Windows uses Miles Audio and is unaffected.
Verified on macOS 15 / arm64 (MoltenVK + DXVK + SDL3); I have not been able to
test the Linux build.

[Copilot]

🟡 Changes recommended

alDeleteBuffers() failures are still not checked/logged, which risks silent regressions of the “buffers never freed” behavior if deletion fails again for any reason.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Pull request overview

This PR addresses a long-session memory growth issue in the shared OpenAL audio cache (Core/GameEngineDevice/...) by ensuring OpenAL buffers can actually be deleted (preventing unbounded buffer accumulation), while also improving cache sizing accuracy and avoiding wasteful cache lookups for empty sound filenames.

Changes:

• Always detach OpenAL buffers from any still-referencing sources before attempting deletion, avoiding AL_INVALID_OPERATION on alDeleteBuffers.
• Fix cache size accounting by keeping decoded PCM size (as accumulated during FFmpeg decode) instead of overwriting with compressed file size.
• Skip cache lookup/work when the resolved filename is empty (avoids repeated openFile("") calls for stock events with empty Sounds = lists).

File summaries

File	Description
Core/GameEngineDevice/Source/OpenALAudioDevice/OpenALAudioCache.cpp	Fixes OpenAL buffer lifetime handling, corrects cache footprint accounting, and avoids futile empty-filename lookups.

Review details

• Files reviewed: 1/1 changed files
• Comments generated: 1
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.