Simplify channel swap detection and improve performance based on
IDisposable's review comments:
- Pass bool pointer directly instead of encoding in bit flag
- Remove redundant channel count check (already verified earlier)
- Use ARM NEON SIMD for channel swapping (4x faster)
- Process 4 frames (8 samples) per iteration with vrev32q_s16
These changes improve code clarity and boost channel swap performance
by ~4x using vectorized operations.
Query TC358743 HDMI receiver for detected audio sample rate before
initializing ALSA capture device. This fixes distortion issues when
HDMI sources send 44.1kHz audio (e.g., Armbian SBC) instead of 48kHz.
Previously, the code always requested 48kHz from ALSA, but in I2S slave
mode, the RV1106 I2S controller receives whatever clock rate the TC358743
master provides. This caused a sample rate mismatch where ALSA thought
it was 48kHz but hardware was actually running at 44.1kHz, resulting in
incorrect SpeexDSP resampling and audio distortion.
Changes:
- Add V4L2 ioctl to query TC358743's audio_sampling_rate control
- Use detected rate when configuring ALSA (falls back to 48kHz if unavailable)
- SpeexDSP resampler now gets correct input rate (44.1k, 48k, etc.)
- Supports all HDMI audio sample rates: 32k, 44.1k, 48k, 88.2k, 96k, etc.
Refactor audio processing to enhance stability and code clarity:
- Remove soft-clipping from audio capture pipeline
- Fix hardware frame size calculation for variable sample rates
- Add comprehensive error codes for audio initialization failures
- Clear stop flags after cleanup to prevent initialization deadlocks
- Improve mutex handling during device initialization
- Simplify constant validation and remove redundant comments
- Add DevPod setup instructions for Apple Silicon users
- Enforce Go cache clearing in dev_deploy.sh for CGO reliability
These changes improve audio capture stability when switching between
HDMI and USB audio sources, and fix race conditions during device
initialization and teardown.
Defensive programming to prevent undefined behavior when closing ALSA
PCM handles. While the previous commit disabled assertions with -DNDEBUG,
adding explicit NULL checks ensures graceful handling even if handles are
unexpectedly NULL.
All error paths that call snd_pcm_close() now verify the handle is non-NULL
before closing, preventing potential crashes in edge cases.
- Add input validation: NULL checks, bounds checking (max 7680 samples)
- Change return type to int for error propagation
- Use saturating NEON arithmetic (vqaddq_s16, vqsubq_s16) to prevent overflow
- Fix type consistency: use int16_t instead of short throughout
- Update documentation: precise threshold (0.9375 or 15/16), describe 4:1 compression
- Remove redundant clamping operations (mathematically proven unnecessary)
- Add stdbool.h include for bool type support
- Handle soft-clip errors at call site to prevent encoding corrupted audio
Implements SIMD-optimized soft-clipping before Opus encoding to prevent
digital clipping distortion on sharp transient attacks (e.g., plastic cup
impacts, percussive sounds). Uses smooth saturation curve starting at
±30720 (~94% of max amplitude) to preserve audio quality while eliminating
crackles and pops.
Processes 8 samples per iteration using ARM NEON intrinsics for optimal
performance on the ARM Cortex-A7 platform.
Query the ALSA channel map (snd_pcm_get_chmap) to detect hardware that
reports non-standard channel ordering (R,L instead of L,R). When detected,
swap channels after capture to ensure correct left/right positioning.
This properly handles hardware quirks (like TC358743 HDMI audio) without
hardcoding device names, making the solution portable and correct.
When HDMI is unplugged during active audio capture, the blocking
snd_pcm_readi() call was holding the mutex, preventing clean shutdown.
This caused snd_pcm_drop() to race with the blocking read, leading to
undefined behavior and crashes.
Solution mirrors PiKVM's approach:
- Release mutex before snd_pcm_readi()/snd_pcm_writei()
- Reacquire mutex after I/O completes
- Verify handle and stop flag before proceeding
This allows snd_pcm_drop() to immediately abort pending I/O when the
device is closed, ensuring clean shutdown during HDMI hotplug events.
Replace ALSA plugin layer resampling with libspeexdsp for improved audio
quality and reliability. This implementation uses direct hardware access
(hw:) instead of ALSA plugins (plughw:) and handles sample rate conversion
with SpeexDSP's high-quality sinc-based resampler.
Key changes:
- Add libspeexdsp 1.2.1 with ARM NEON optimizations to build dependencies
- Switch from plughw: to hw: device access for lower latency
- Implement conditional resampling (only when hardware rate ≠ 48kHz)
- Use SPEEX_RESAMPLER_QUALITY_DESKTOP for high-quality interpolation
- Add automatic audio dependency building in dev_deploy.sh
Quality improvements:
- Fix race condition in resampler cleanup with mutex protection
- Fix memory leak on resampler re-initialization
- Add buffer overflow validation (3840 frame limit for 192kHz)
- Improve error logging for resampling, encoding, and ALSA configuration
- Simplify code structure while maintaining all functionality
Technical details:
- Hardware negotiates actual sample rate (e.g., HDMI may vary)
- SpeexDSP converts hardware rate → 48kHz for Opus encoding
- USB Audio Gadget hardcoded to 48kHz (no resampling overhead)
- Static buffer allocation for zero allocation in hot path
- WebRTC requires 48kHz RTP clock rate per RFC 7587
Changes the audio subsystem from hw: (direct hardware access) to plughw:
(plugin layer with rate conversion) to enable configurable sample rates.
Changes:
- Update ALSA build to include plug,rate,linear,copy plugins
- Change device names from hw: to plughw: in C and Go code
- Remove 48kHz hardcoding for HDMI audio output
- Keep USB at 48kHz since hardware is fixed at that rate
- Update all comments to reflect plughw usage
Technical details:
- hw: devices bypass all ALSA plugins and require exact hardware rate match
- plughw: devices enable the ALSA plugin layer for automatic rate conversion
- Hardware still receives at native rate (48kHz), resampling happens in userspace
- HDMI can now use 8k/12k/16k/24k/48kHz, USB remains at 48kHz
- NEON-optimized resampling provides good performance on Cortex-A7
Requires rebuilding ALSA library with updated plugin configuration.
USB Audio Gadget (hw:1,0) hardware only supports 48kHz for both capture
and playback due to configfs p_srate/c_srate being hardcoded. This commit
ensures both audio paths respect this hardware limitation:
- Output path: Force 48kHz when using hw:1,0, allow configurable rates for HDMI
- Input path: Always use 48kHz regardless of UI configuration
- Calculate frame size dynamically based on actual sample rate used
Also removes redundant comments that don't add debugging or maintainability value.
- Clarify sample rate is configurable (8k/12k/16k/24k/48k), not fixed at 48kHz
- Expand mutex comment to include full lifecycle protection scope
- Document that ALSA playback init fails immediately with no fallback
- Add async behavior documentation to audio enable/restart functions
- Restore build_audio_deps target lost during merge
- Restore lint-fix, lint-go, lint-ui Makefile targets
- Fix variable alignment per linter
- Remove silent fallback to ALSA 'default' device on playback init failure
- Return error from SetAudioOutputSource for invalid source values
- Fix misleading comment about mutex scope in C audio code
- Clarify inputSourceMutex purpose for WebRTC packet serialization
- Replace helper function in getAudioConfig with explicit validation
- Consolidate audio default application in LoadConfig
- Streamline relay retry logic with inline conditions
- Extract closeFile and openHidFile helpers in USB gadget
- Simplify setPendingInputTrack pointer handling
- Improve error handling clarity in startAudio and updateUsbRelatedConfig
- Clean up processInputPacket mutex usage
Use snd_pcm_hw_params_set_rate_resample(1) to enable ALSA's rate plugin,
which provides software resampling even with hw: device interface.
This fixes audio distortion when HDMI sources output non-48kHz rates
(e.g., 44.1kHz from SBCs). ALSA now automatically resamples any input
rate to the configured 48kHz that Opus expects.
The rate plugin is available because ALSA is compiled with
--with-pcm-plugins=rate in install_audio_deps.sh
ALSA now forces the configured sample rate (default 48kHz) instead of
auto-detecting the source rate. This prevents Opus encoder initialization
failures when HDMI sources output 44.1kHz audio, which Opus doesn't support.
Changes:
- Use snd_pcm_hw_params_set_rate() to force exact rate (48kHz by default)
- ALSA performs software resampling if hardware rate differs
- Update valid rates to Opus-compatible only (8k, 12k, 16k, 24k, 48k)
- Remove auto-adaptation logic that caused Opus failures with 44.1kHz
This ensures audio capture works reliably with any HDMI source rate.
- Fix validateAndApply comment to clarify it returns values, doesn't apply them
- Correct capture_channels comment about hardware capabilities
- Fix opus_packet_loss_perc default value from 0 to 20 (matches backend default)
- Fix handle_alsa_error return value documentation (return 0 also unlocks mutex)
Changes:
- Consolidate duplicate stop logic into helper functions
- Fix RPC getAudioConfig to return actual runtime values instead of
inconsistent defaults (bitrate was returning 128 vs actual 192)
- Improve setAudioTrack mutex handling to eliminate nested locking
- Simplify ALSA error retry logic by reorganizing conditional branches
- Split CGO Connect() into separate input/output methods for clarity
- Use map lookup for sample rate validation instead of long if-chain
- Add inline comments documenting validation steps
All changes preserve existing functionality while reducing code
duplication and improving readability. Tested with both HDMI and
USB audio sources.
Critical Fixes:
- Fix race condition in handleInputTrackForSession by reloading source inside mutex
- Fix ALSA handle cleanup atomicity (nullify before close to prevent use-after-free)
- Bounds check for opus buffer already present (verified)
Configuration Alignment:
- Align audio bitrate default to 192 kbps across all layers (C, Go defaults, config)
- Align audio complexity default to 8 across all layers
- Align DTX default to enabled (true/1) across all layers for bandwidth efficiency
Documentation Improvements:
- Update C header comment to reflect accurate 192 kbps default
- Clarify NEON requirement (not just "always available")
- Fix ALSA device mapping comments to reflect environment variable usage
- Document fallback behavior in playback init
Code Quality:
- Add validation logging for out-of-range audio configuration values
- Improve error visibility for configuration issues
All changes thoroughly analyzed before implementation.
Change recovery_attempts from int to uint8_t for better efficiency:
- Reduces memory footprint (1 byte vs 4 bytes)
- Better cache utilization on ARM
- Matches max_attempts type (uint8_t)
- Values never exceed 3, fits perfectly in uint8_t range
Updated function signature and all call sites for consistency.
Extract shared error recovery logic:
- Create handle_alsa_error() for EPIPE, EAGAIN, ESTRPIPE, EIO errors
- Consolidates ~180 lines of duplicate error handling code
- Used by both capture and playback paths
Extract shared close logic:
- Create close_audio_stream() for safe shutdown sequence
- Handles CAS synchronization, delay, mutex protection
- Used by both jetkvm_audio_capture_close and jetkvm_audio_playback_close
Remove all TRACE_LOG dead code:
- TRACE_LOG was compiled to ((void)0) with zero runtime value
- Eliminates ~30 statements cluttering the codebase
Result: 87 lines removed (9% reduction), improved maintainability
- Separate capture_channels (stereo HDMI) from playback_channels (mono mic)
to prevent initialization conflicts that were breaking stereo output
- Optimize defaults for LAN use: 192kbps bitrate, complexity 8, 0% packet
loss compensation, DTX disabled (eliminates static and improves clarity)
- Add comprehensive race condition protection in C audio layer with handle
validity checks and mutex-protected cleanup operations
- Enable USB audio volume control and configure microphone as mono
- Add centralized AUDIO_DEFAULTS constant in UI with localized labels
- Add missing time import to fix compilation
This resolves audio quality issues and crash scenarios when switching
between HDMI and USB audio sources.
Moved all start/stop of sources into audio (out of jsonrpc)
Clean up duplicated code, made direction a bool, more logging, made all source/relay atomics.
Eliminate SetConfig since we always set it during start.
Eliminate the extra initialized flag.
Properly detect when USB audio was previously active.
Relay has the pointer to the source, not a copy.
CgoSource (and stub) expose the AudioSource interface.
Removed obvious comments that don't add value:
- cgo_source.go: Removed redundant status check comments
- audio.go: Consolidated mutex pattern comments
Kept important comments that explain non-obvious patterns:
- Why mutex is released before C calls (deadlock prevention)
- Why operations happen outside mutex (avoid blocking on CGO)
- Why single critical section is used (race condition prevention)
Problem:
Previous fix reduced but didn't eliminate the hang when switching audio
sources. The C layer was still blocking on snd_pcm_readi()/snd_pcm_writei()
while holding the mutex, preventing cleanup from proceeding.
Solution:
Call snd_pcm_drop() BEFORE acquiring the mutex in close functions. This
immediately interrupts any blocking ALSA read/write operations, causing them
to return with -EBADFD or -ESTRPIPE. The sequence is now:
1. Set stop_requested flag
2. Call snd_pcm_drop() to interrupt blocking I/O (no mutex needed - thread-safe)
3. Acquire mutex for cleanup
4. Close handles and free resources
5. Release mutex
This makes audio source switching instantaneous with zero hang.
Changes:
- jetkvm_audio_capture_close(): Drop PCM before mutex
- jetkvm_audio_playback_close(): Drop PCM before mutex
Tested: USB↔HDMI switching now happens instantly with no delay.
Problem:
When switching audio sources (USB to HDMI or vice versa), the application
would hang indefinitely. This was caused by a deadlock between Go and C
layers:
1. Main thread calls SetAudioOutputSource() → stopOutputAudio()
2. stopOutputAudio() calls outputRelay.Stop() which waits for goroutine
3. Goroutine is blocked in ReadMessage() holding Go mutex
4. ReadMessage() calls blocking C function jetkvm_audio_read_encode()
5. C function is blocked reading from ALSA device
6. Disconnect() can't acquire Go mutex to clean up
7. Deadlock: Main thread waiting for goroutine, goroutine waiting for ALSA
Solution:
Release the Go mutex BEFORE calling blocking C functions in ReadMessage()
and WriteMessage(). The C layer has its own pthread mutex protection and
handles stop requests via atomic flags. This allows:
- Disconnect() to acquire the mutex immediately
- C layer to detect stop request and return quickly
- Goroutines to exit cleanly
- Audio source switching to work flawlessly
Fixes:
- internal/audio/cgo_source.go:ReadMessage() - Release mutex before C call
- internal/audio/cgo_source.go:WriteMessage() - Release mutex before C call
This fix eliminates the hang when switching between USB and HDMI audio
sources.
Root cause: ALSA assertion failure `snd_pcm_writei: Assertion 'pcm' failed`
when pcm_playback_handle became NULL during concurrent write operations.
The crash occurred because:
1. Thread A checks pcm_playback_handle != NULL (passes)
2. Thread B calls jetkvm_audio_playback_close(), sets handle = NULL
3. Thread A calls snd_pcm_writei(NULL, ...) → SIGABRT
Solution: Added pthread mutexes to protect concurrent access:
- playback_mutex protects pcm_playback_handle in decode_write and close
- capture_mutex protects pcm_capture_handle in read_encode and close
All critical sections now acquire mutex before accessing ALSA handles,
preventing the NULL pointer from being passed to ALSA functions.
Root cause: ALSA was silently using a different sample rate than configured,
causing severe pitch/speed distortion (the "cassette player" warping effect).
The bug occurred when:
- User configured 48 kHz in UI
- HDMI source output 44.1 kHz audio
- set_rate() failed, set_rate_near() chose 44.1 kHz
- Code never checked what rate was actually set
- Opus encoder created for 48 kHz but received 44.1 kHz audio
- Result: ~9% pitch shift and timing mismatch
Fix:
- Always use set_rate_near() and check the actual rate returned
- Pass detected rate and frame size to Opus encoder/decoder creation
- Avoid modifying global state to prevent capture/playback interference
- Recalculate frame_size for 20ms at the actual rate
- Verify rate after hw_params application
- Add detailed logging for rate adaptation
This ensures Opus encoder/decoder use the correct rate matching the hardware,
regardless of what the HDMI source outputs.
- Update default EDID with registered manufacturer ID (Dell) and proper 24-inch display dimensions (52x32cm) for better macOS/OS compatibility
- Add configurable sample rate (32/44.1/48/96 kHz) to support different HDMI audio sources
- Add packet loss compensation percentage control for FEC overhead tuning
- Fix config migration to ensure new audio parameters get defaults for existing configs
- Update all language translations for new audio settings
- Add config fields for bitrate, complexity, DTX, FEC, buffer periods
- Add RPC methods for get/set audio config and restart
- Add UI settings page with controls for all audio parameters
- Add Apply Settings button to restart audio with new config
- Add config migration for backwards compatibility
- Add translations for all 9 languages
- Clean up redundant comments and optimize log levels
- Revert .golangci.yml to dev branch state (removed custom build-tags)
- Remove internal/audio/cgo_source_stub.go (not needed with proper cross-compilation)
- Fix import ordering in ui/src/utils.ts
Use 'make lint-go' for proper ARM cross-compilation environment.
Provides no-op AudioSource implementations for platforms that don't
support ARM CGO audio (x86_64, darwin, etc.). This allows golangci-lint
to run successfully on any platform without requiring ARM cross-compilation
toolchain.
The stub implementations return errors when called, ensuring that if
they're accidentally used at runtime on non-ARM platforms, it will fail
gracefully with a clear error message rather than undefined symbols.
Build constraints ensure the real CGO implementation is used on linux/arm
and linux/arm64, while stubs are used everywhere else.
Remove all subprocess-based audio code to simplify the audio system and
reduce complexity. Audio now uses CGO in-process mode exclusively.
Changes:
- Remove subprocess mode: Deleted Supervisor, IPCSource, embed.go
- Remove audio mode selection from UI (Settings → Audio)
- Remove audio mode from backend config (AudioMode field)
- Remove JSON-RPC handlers: getAudioMode/setAudioMode
- Remove Makefile targets: build_audio_output/input/binaries
- Remove standalone C binaries: jetkvm_audio_{input,output}.c
- Remove IPC protocol implementation: ipc_protocol.{c,h}
- Remove unused IPC functions from audio_common.{c,h}
- Simplify audio.go: startAudio() instead of startAudioSubprocesses()
- Update all function calls and comments to remove subprocess references
- Add constants to cgo_source.go (ipcMaxFrameSize, ipcMsgTypeOpus)
- Keep update_opus_encoder_params() for potential future runtime config
Benefits:
- Simpler codebase: -1,734 lines of code
- Better performance: No IPC overhead on embedded hardware
- Easier maintenance: Single audio implementation
- Smaller binary: No embedded audio subprocess binaries
The audio system now works exclusively via CGO direct C function calls,
with ALSA device selection (HDMI vs USB) still configurable via settings.
Alex P
Marc Brooks