fix: Oboe SIGSEGV on OpenSL ES + speaker toggle for native stream by torlando-tech · Pull Request #470 · torlando-tech/columba

torlando-tech · 2026-02-15T18:51:09Z

Summary

Bug 1 (COLUMBA-3A): Fix fatal SIGSEGV in Oboe audio callback on 32-bit OpenSL ES devices (Galaxy A13). The stop() + close() + reset() teardown sequence left a race window where the callback could fire after the stream object was freed. Fix: remove redundant stop() (Oboe's close() handles it), add destroyed_ atomic guard in onAudioReady().
Bug 2: Fix speaker toggle having no effect on Oboe playback path. Native Oboe streams bind to a hardware device at open time and don't follow setCommunicationDevice() routing changes. Fix: add restartStream() to close/reopen the Oboe stream, called from AudioDevice.setSpeakerphoneOn().

Files changed (LXST-kt submodule)

File	Changes
`oboe_playback_engine.h`	Add `destroyed_` atomic, `restartStream()` declaration
`oboe_playback_engine.cpp`	Fix `closeStream()`, `destroy()`, `stopStream()`; add `destroyed_` guard in callback; add `restartStream()`
`oboe_playback_jni.cpp`	Add `nativeRestartStream` JNI method
`NativePlaybackEngine.kt`	Add `restartStream()` + external declaration
`AudioDevice.kt`	Call `NativePlaybackEngine.restartStream()` in `setSpeakerphoneOn()`

Impact scope

These changes affect all ABIs (arm64-v8a, armeabi-v7a, x86_64), not just 32-bit. The SIGSEGV was only observed on OpenSL ES, but the teardown fix and destroyed_ guard apply to all Oboe streams. The speaker toggle restartStream() is also architecture-agnostic. Risk on 64-bit AAudio devices is low — close() already synchronizes callbacks reliably, and the atomic guard adds ~1ns per callback.

Test plan

Automated

compileDebugKotlin passes
TelephoneTest (65 tests) + ProfileTest (43 tests) pass

Manual — 32-bit device (primary target)

Deploy to Galaxy A13 or equivalent armeabi-v7a device
Make a voice call, verify audio plays through earpiece
Toggle speaker — audio should switch to speaker at normal volume
Toggle back to earpiece — audio should return to earpiece
End call — no crash on teardown
Repeat 3x to stress the closeStream() path
Check Sentry for COLUMBA-3A recurrence after release

Manual — 64-bit device (regression check)

Deploy to a 64-bit device (Pixel, Samsung S-series, etc.)
Make a voice call, verify audio plays through earpiece at normal volume
Toggle speaker mid-call — audio should switch with only a brief (~50ms) gap
Toggle back to earpiece — audio should return cleanly
End call normally — verify no crash, clean shutdown in logcat
Start a second call immediately after ending the first — verify no lingering state issues
During a call, check logcat for LXST:OboeEngine — verify:
- Stream opened: API=AAudio (confirms AAudio path on 64-bit)
- Restarting stream for audio routing change appears on speaker toggle
- Stream closed / Stream started pairs match (no orphaned streams)
If device supports headphones/BT: plug in during call, verify onErrorAfterClose recovery still works (existing behavior, not new)

🤖 Generated with Claude Code

sentry · 2026-02-15T18:54:46Z

Codecov Report

❌ Patch coverage is 84.31373% with 8 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
python/lxst_modules/call_manager.py	84.31%	8 Missing ⚠️

📢 Thoughts on this report? Let us know!

greptile-apps · 2026-02-15T18:54:47Z

Greptile Summary

This PR updates the LXST-kt submodule to fix two audio bugs in the native Oboe playback engine:

SIGSEGV fix (COLUMBA-3A): Removes the redundant stop() call before close() in stream teardown and adds a destroyed_ atomic guard in the onAudioReady callback to prevent use-after-free on 32-bit OpenSL ES devices (Galaxy A13). The close() call already stops the stream internally, and the atomic guard ensures any late-firing callback returns immediately.
Speaker toggle fix: Adds restartStream() which closes and reopens the Oboe stream when the speaker/earpiece toggle changes. Oboe streams bind to a hardware device at open time and don't follow setCommunicationDevice() routing changes, so a full close/reopen cycle is needed. The Kotlin side catches UnsatisfiedLinkError to gracefully handle cases where the native library isn't loaded.
Race fix (follow-up commit): Sets isPlaying_ = false in restartStream() before calling closeStream() to prevent onErrorAfterClose() from concurrently reopening the stream during a speaker toggle restart.

Confidence Score: 4/5

This PR is safe to merge — it fixes a real SIGSEGV crash and adds necessary speaker toggle support with sound synchronization patterns.
The SIGSEGV fix is well-reasoned: removing the redundant stop() call and adding a destroyed_ atomic guard is the correct approach for OpenSL ES's less deterministic callback lifecycle. The restartStream race fix (setting isPlaying_=false before closeStream) properly prevents onErrorAfterClose from concurrent re-opening. The Kotlin-side UnsatisfiedLinkError catch is a defensive guard for when native code isn't loaded. The one deduction is for the theoretical concurrent access to the stream_ shared_ptr between restartStream and onErrorAfterClose without a mutex, though this is low-probability in practice.
oboe_playback_engine.cpp — the stream lifecycle methods (restartStream, openStream, closeStream, onErrorAfterClose) share stream_ without a mutex, relying only on atomic flag ordering. Manual testing on both 32-bit and 64-bit devices is critical.

Important Files Changed

Filename	Overview
LXST-kt	Submodule pointer update (4af529c → 8859783) bringing in two fixes: SIGSEGV guard for Oboe audio callback on OpenSL ES teardown, and speaker toggle restart for native Oboe streams. One minor thread-safety concern in `restartStream` but low practical risk.

Sequence Diagram

sequenceDiagram
    participant KT as AudioDevice.kt (Kotlin)
    participant NPE as NativePlaybackEngine.kt
    participant JNI as oboe_playback_jni.cpp
    participant OPE as OboePlaybackEngine
    participant Oboe as Oboe AudioStream
    participant CB as SCHED_FIFO Callback

    Note over KT: User toggles speaker
    KT->>NPE: NativePlaybackEngine.restartStream()
    NPE->>JNI: nativeRestartStream()
    JNI->>OPE: restartStream()
    OPE->>OPE: isPlaying_ = false
    OPE->>Oboe: stream_->close()
    Note over CB: Late callback sees isPlaying_=false → Stop
    Oboe-->>OPE: close() returns
    OPE->>OPE: stream_.reset()
    OPE->>Oboe: builder.openStream(stream_)
    OPE->>OPE: isPlaying_ = true
    OPE->>Oboe: stream_->requestStart()
    Oboe->>CB: onAudioReady() resumes

    Note over OPE: destroy() path (teardown)
    OPE->>OPE: destroyed_ = true
    OPE->>Oboe: stream_->close()
    Note over CB: Late callback sees destroyed_=true → memset(0) + Stop
    OPE->>OPE: stream_.reset()

_{Last reviewed commit: 77a4629}

torlando-tech · 2026-02-15T20:17:39Z

@greptileai

greptile-apps

_{1 file reviewed, 1 comment}

_{Edit Code Review Agent Settings | Greptile}

LXST-kt

Update LXST-kt submodule with fixes for: - SIGSEGV in Oboe callback on 32-bit OpenSL ES devices (COLUMBA-3A) - Speaker toggle having no effect on native Oboe playback stream Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

LXST-kt

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

__link_closed() was calling Kotlin callbacks (signal + onCallEnded) while holding _call_handler_lock. The signal triggered Kotlin hangup() synchronously, running NativePlaybackEngine.destroy() which blocks on Oboe stream close — keeping the Python lock held for the entire duration and preventing subsequent call() from acquiring it. Fix: move Kotlin callbacks outside the lock (same pattern as hangup()) and dispatch hangup() async in Kotlin's STATUS_AVAILABLE handler. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Voice calls degraded progressively from 60ms to 130ms per-frame arrival rate over 20 seconds, even on fast 1-hop local WiFi links. Root cause: each RNS.Packet.send() holds the Python GIL for encryption (AES-256-CBC + HMAC-SHA256) and transport dispatch, creating a feedback loop where both devices' TX/RX paths compete for GIL time. Batch 3 audio frames per RNS.Packet.send() call, reducing crypto overhead from ~16.7 calls/sec to ~5.6 calls/sec (67% reduction). The LXST wire format already supports frame lists ({0x01: [f1, f2, f3]}) and the receiver already handles both single frames and lists. Results: packet arrival rate stable at ~60ms/frame for 40+ seconds, zero silence callbacks, zero PLC, buffer steady at 6-9 frames. Also updates LXST-kt submodule with adaptive playout drain for ring buffer latency bounding during packet bursts and speaker toggles. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

sentry · 2026-02-16T19:49:44Z

python/lxst_modules/call_manager.py

+        if link is not None and self._tx_batch:
+            try:
+                self._flush_tx_batch(link)


Bug: The _tx_batch list is accessed by hangup() and receive_audio_packet() without a lock, creating a race condition that can cause lost audio frames during call teardown.
_{Severity: HIGH}

Suggested Fix

Wrap all accesses and modifications to _tx_batch and its related state variables within both the hangup() and receive_audio_packet() methods using the self._call_handler_lock. This will ensure that operations on the transmit buffer are atomic and prevent concurrent modification from different threads.

Prompt for AI Agent

Review the code at the location below. A potential bug has been identified by an AI agent. Verify if this is a real issue. If it is, propose a fix; if not, explain why it's not valid. Location: python/lxst_modules/call_manager.py#L436-L438 Potential issue: A race condition exists due to unsynchronized access to the `_tx_batch` list from multiple threads. The `hangup()` method can read, modify, and clear `_tx_batch` without a lock, while the `receive_audio_packet()` method concurrently appends data to it, also without a lock. This can lead to lost audio frames, as one thread might replace the list object (`_tx_batch = []`) while another thread is still operating on a stale reference. This can also cause state corruption for diagnostic counters like `_tx_batch_count`.

…oggle fix: Oboe SIGSEGV on OpenSL ES + speaker toggle for native stream

greptile-apps bot reviewed Feb 15, 2026

View reviewed changes

LXST-kt Outdated Show resolved Hide resolved

torlando-tech and others added 3 commits February 15, 2026 22:40

fix: update LXST-kt with restartStream race fix

372a182

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

fix: update LXST-kt with getXRunCount TOCTOU fix

a4a4348

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

torlando-tech force-pushed the fix/oboe-sigsegv-speaker-toggle branch from a58e4c2 to a4a4348 Compare February 16, 2026 03:41

feat: update LXST-kt with Opus PLC and increased prebuffer

ac17fb0

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

sentry bot reviewed Feb 16, 2026

View reviewed changes

LXST-kt Outdated Show resolved Hide resolved

torlando-tech and others added 4 commits February 15, 2026 23:46

fix: update LXST-kt with PLC decoder race guard

7c397b1

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

fix: update LXST-kt with stream lifecycle serialization

1e9636e

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

sentry bot reviewed Feb 16, 2026

View reviewed changes

torlando-tech merged commit 64b5767 into main Feb 16, 2026
10 checks passed

torlando-tech deleted the fix/oboe-sigsegv-speaker-toggle branch February 16, 2026 20:21

torlando-tech added a commit that referenced this pull request Feb 17, 2026

Merge pull request #470 from torlando-tech/fix/oboe-sigsegv-speaker-t…

818cfce

…oggle fix: Oboe SIGSEGV on OpenSL ES + speaker toggle for native stream

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

fix: Oboe SIGSEGV on OpenSL ES + speaker toggle for native stream#470

fix: Oboe SIGSEGV on OpenSL ES + speaker toggle for native stream#470
torlando-tech merged 8 commits intomainfrom
fix/oboe-sigsegv-speaker-toggle

torlando-tech commented Feb 15, 2026

Uh oh!

sentry bot commented Feb 15, 2026 •

edited

Loading

Uh oh!

greptile-apps bot commented Feb 15, 2026 •

edited

Loading

Uh oh!

torlando-tech commented Feb 15, 2026

Uh oh!

greptile-apps bot left a comment

Uh oh!

Uh oh!

Uh oh!

sentry bot Feb 16, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Uh oh!

Conversation

torlando-tech commented Feb 15, 2026

Summary

Files changed (LXST-kt submodule)

Impact scope

Test plan

Automated

Manual — 32-bit device (primary target)

Manual — 64-bit device (regression check)

Uh oh!

sentry bot commented Feb 15, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

greptile-apps bot commented Feb 15, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Greptile Summary

Confidence Score: 4/5

Important Files Changed

Sequence Diagram

Uh oh!

torlando-tech commented Feb 15, 2026

Uh oh!

greptile-apps bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

sentry bot Feb 16, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

sentry bot commented Feb 15, 2026 •

edited

Loading

greptile-apps bot commented Feb 15, 2026 •

edited

Loading