Audio Thread Safety & Lock-Free Patterns

The audio thread in JUCE/Tracktion Engine runs under strict real-time constraints. Any blocking or unbounded operation causes audible glitches (clicks, dropouts, silence). This skill covers what is forbidden, how to communicate safely between threads, and the patterns used in this codebase.

What Is Forbidden on the Audio Thread

All of the following can block, allocate, or take unbounded time. Never do any of these inside Plugin::applyToBuffer(), AudioProcessor::processBlock(), or any code called from the audio callback.

1. Memory Allocation

cpp

// FORBIDDEN - all of these may call malloc/new
auto* p = new MyObject();                  // heap allocation
std::vector<float> temp(numSamples);       // heap allocation on resize
juce::String s = "level: " + juce::String(value);  // heap allocation
juce::Array<int> arr;
arr.add(42);                               // may reallocate
myStdVector.push_back(x);                  // may reallocate

Instead: Pre-allocate in initialise() or use stack-allocated fixed-size buffers.

cpp

// OK - stack allocation with known size
float temp[2048];

// OK - pre-allocated in initialise(), reused in applyToBuffer()
std::vector<float> scratchBuffer;  // member variable, resized in initialise()

2. Locks and Mutexes

cpp

// FORBIDDEN - may block waiting for another thread
juce::ScopedLock sl(criticalSection);
std::lock_guard<std::mutex> lock(mutex);
std::unique_lock<std::mutex> ul(mutex);

Instead: Use atomics, lock-free queues, or juce::AbstractFifo.

3. MessageManager Calls

cpp

// FORBIDDEN - posts to message thread, may allocate, may block
juce::MessageManager::callAsync([this] { /* ... */ });
sendChangeMessage();
triggerAsyncUpdate();  // OK only from non-audio thread

4. File I/O

cpp

// FORBIDDEN - unbounded latency
juce::File("x.wav").loadFileAsData();
fopen(), fread(), fwrite();
DBG("value: " + juce::String(x));  // writes to stderr

5. Objective-C Message Sends (macOS)

cpp

// FORBIDDEN - ObjC runtime may take locks
// Any call that crosses into ObjC (NSLog, CoreFoundation bridged calls, etc.)

6. Unbounded Operations

cpp

// FORBIDDEN - unknown iteration count
while (!ready.load()) { /* spin */ }  // unbounded spin
for (auto& item : dynamicContainer) { /* ... */ }  // if container can grow

Lock-Free Communication Patterns

Pattern 1: std::atomic for Simple Values

Use for single values shared between audio and UI threads. memory_order_relaxed is sufficient when there is no ordering dependency between multiple variables.

cpp

class MyPlugin : public te::Plugin {
    std::atomic<float> gainLevel{1.0f};

    // Message thread (UI/parameter change):
    void setGain(float g) {
        gainLevel.store(g, std::memory_order_relaxed);
    }

    // Audio thread:
    void applyToBuffer(const PluginRenderContext& rc) override {
        float g = gainLevel.load(std::memory_order_relaxed);
        rc.destBuffer->applyGain(0, rc.bufferNumSamples, g);
    }
};

Pattern 2: Atomic Exchange for Peak Metering

The audio thread writes a running maximum. The UI thread reads and resets in one atomic operation, ensuring no peaks are lost and no lock is needed.

cpp

class MyPlugin : public te::Plugin {
    std::atomic<float> peakLeft{0.0f};
    std::atomic<float> peakRight{0.0f};

    // Audio thread: update running peak
    void applyToBuffer(const PluginRenderContext& rc) override {
        auto& buf = *rc.destBuffer;
        float newPeakL = buf.getMagnitude(0, rc.bufferStartSample, rc.bufferNumSamples);
        float newPeakR = buf.getMagnitude(1, rc.bufferStartSample, rc.bufferNumSamples);

        // Only store if new peak is greater than current
        auto prevL = peakLeft.load(std::memory_order_relaxed);
        if (newPeakL > prevL)
            peakLeft.store(newPeakL, std::memory_order_relaxed);

        auto prevR = peakRight.load(std::memory_order_relaxed);
        if (newPeakR > prevR)
            peakRight.store(newPeakR, std::memory_order_relaxed);
    }

    // UI thread (timer callback): read and reset
    float consumePeakLeft() {
        return peakLeft.exchange(0.0f, std::memory_order_relaxed);
    }
    float consumePeakRight() {
        return peakRight.exchange(0.0f, std::memory_order_relaxed);
    }
};

Pattern 3: juce::CachedValue for ValueTree-Backed Properties

CachedValue<T> caches a ValueTree property in a local variable that is safe to read on the audio thread (the cached copy is updated on the message thread via listener, but the read is just a plain member access).

cpp

class MyPlugin : public te::Plugin {
    juce::CachedValue<float> levelParam;
    juce::CachedValue<float> panParam;

    void initialise(const PluginInitialisationInfo&) override {
        // Bind to ValueTree properties (message thread)
        levelParam.referTo(state, IDs::level, nullptr, 0.8f);
        panParam.referTo(state, IDs::pan, nullptr, 0.0f);
    }

    void applyToBuffer(const PluginRenderContext& rc) override {
        // Safe to read cached value on audio thread
        float level = levelParam.get();
        float pan = panParam.get();
        // ... apply to buffer ...
    }
};

Pattern 4: std::atomic<bool> for Flags and Triggers

Use for one-shot triggers (e.g., pad hit, reset signal) or boolean state flags.

cpp

class DrumPad {
    std::atomic<bool> triggered{false};

    // UI/MIDI thread: fire trigger
    void hit() {
        triggered.store(true, std::memory_order_relaxed);
    }

    // Audio thread: consume trigger
    void processBlock(juce::AudioBuffer<float>& buffer) {
        if (triggered.exchange(false, std::memory_order_relaxed)) {
            // Start sample playback from beginning
            playbackPosition = 0;
        }
        // ... render audio ...
    }
};

Pattern 5: juce::AbstractFifo / Lock-Free Queues

Use when you need to pass variable-sized data or multiple messages between threads.

cpp

class MeterBridge {
    juce::AbstractFifo fifo{512};
    std::array<float, 512> buffer{};

    // Audio thread: write meter values
    void pushMeterValue(float value) {
        const auto scope = fifo.write(1);
        if (scope.blockSize1 > 0)
            buffer[(size_t)scope.startIndex1] = value;
    }

    // UI thread: read meter values
    float popMeterValue() {
        float result = 0.0f;
        const auto scope = fifo.read(1);
        if (scope.blockSize1 > 0)
            result = buffer[(size_t)scope.startIndex1];
        return result;
    }
};

Tracktion Engine Specifics

Plugin Lifecycle & Threading

code

Message Thread                    Audio Thread
──────────────                    ────────────
Plugin::initialise()              Plugin::applyToBuffer()
Plugin::deinitialise()              (called every audio block)
Plugin::restorePluginStateFromValueTree()
ValueTree listeners fire

•Plugin::applyToBuffer(const PluginRenderContext& rc) runs on the audio thread.
•Plugin::initialise() and Plugin::deinitialise() run on the message thread.
•Never access Edit&, ValueTree, or UndoManager from applyToBuffer(). Use CachedValue instead.

PluginRenderContext Quick Reference

cpp

void applyToBuffer(const PluginRenderContext& rc) override {
    auto& audio = *rc.destBuffer;          // juce::AudioBuffer<float>&
    auto& midi  = *rc.bufferForMidiMessages; // MidiMessageArray&
    int startSample = rc.bufferStartSample;
    int numSamples  = rc.bufferNumSamples;
    double sampleRate = sampleRateValue;   // from initialise()
}

Rack Wrapping

When a plugin is rack-wrapped (inserted into a RackType), Tracktion Engine manages the audio routing. The plugin's applyToBuffer() is still called on the audio thread, but the buffer routing is handled by the rack. Initialise/deinitialise lifecycle is managed by the rack's node graph.

Common Patterns in This Codebase

Per-Chain Peak Metering

cpp

// In a multi-chain plugin (e.g., drum grid with multiple output chains):
struct Chain {
    std::atomic<float> peak{0.0f};
    // ... other chain state ...
};

std::array<Chain, 16> chains;

// Audio thread: update peak for each chain
void applyToBuffer(const PluginRenderContext& rc) override {
    for (int i = 0; i < numActiveChains; ++i) {
        float mag = getChainMagnitude(i, rc);
        auto prev = chains[i].peak.load(std::memory_order_relaxed);
        if (mag > prev)
            chains[i].peak.store(mag, std::memory_order_relaxed);
    }
}

// UI thread: consume peaks for meter display
float consumePeak(int chainIndex) {
    return chains[chainIndex].peak.exchange(0.0f, std::memory_order_relaxed);
}

Pad Trigger Flags

cpp

// Array of atomic trigger flags, one per pad
std::array<std::atomic<bool>, 16> padTriggers{};

// MIDI/UI thread
void triggerPad(int padIndex) {
    padTriggers[padIndex].store(true, std::memory_order_relaxed);
}

// Audio thread
void applyToBuffer(const PluginRenderContext& rc) override {
    for (int i = 0; i < 16; ++i) {
        if (padTriggers[i].exchange(false, std::memory_order_relaxed)) {
            startPadPlayback(i);
        }
    }
}

CachedValue for Level/Pan

cpp

// Backed by ValueTree so values persist and can be automated
juce::CachedValue<float> level, pan;

void initialise(const PluginInitialisationInfo& info) override {
    level.referTo(state, IDs::level, nullptr, 0.8f);
    pan.referTo(state, IDs::pan, nullptr, 0.0f);
}

void applyToBuffer(const PluginRenderContext& rc) override {
    float l = level.get();
    float p = pan.get();
    // Apply gain and panning to buffer...
}

Debugging Checklist

If you hear clicks, dropouts, or glitches:

•Search for allocations in applyToBuffer() - look for new, String, Array, vector operations
•Search for locks - grep for ScopedLock, lock_guard, CriticalSection in audio path
•Check for DBG() calls in audio code - these do file I/O
•Verify pre-allocation - all buffers sized in initialise(), not in applyToBuffer()
•Check CachedValue usage - ensure referTo() is called in initialise(), not in applyToBuffer()