tomato-testing/toxav/audio_test.cc

#include "audio.h"

#include <gtest/gtest.h>

#include <algorithm>
#include <cmath>
#include <vector>

#include "../toxcore/logger.h"
#include "../toxcore/mono_time.h"
#include "../toxcore/network.h"
#include "../toxcore/os_memory.h"
#include "av_test_support.hh"
#include "rtp.h"

namespace {

using AudioTest = AvTest;

TEST_F(AudioTest, BasicNewKill)
{
    AudioTestData data;
    ACSession *ac = ac_new(mono_time, log, 123, AudioTestData::receive_frame, &data);
    ASSERT_NE(ac, nullptr);
    ac_kill(ac);
}

TEST_F(AudioTest, EncodeDecodeLoop)
{
    AudioTestData data;
    ACSession *ac = ac_new(mono_time, log, 123, AudioTestData::receive_frame, &data);
    ASSERT_NE(ac, nullptr);

    RtpMock rtp_mock;
    RTPSession *send_rtp = rtp_new(log, RTP_TYPE_AUDIO, mono_time, RtpMock::send_packet, &rtp_mock,
        nullptr, nullptr, nullptr, ac, RtpMock::audio_cb);
    RTPSession *recv_rtp = rtp_new(log, RTP_TYPE_AUDIO, mono_time, RtpMock::send_packet, &rtp_mock,
        nullptr, nullptr, nullptr, ac, RtpMock::audio_cb);
    rtp_mock.recv_session = recv_rtp;

    uint32_t sampling_rate = 48000;
    uint8_t channels = 1;
    size_t sample_count = 960;  // 20ms at 48kHz

    // Reconfigure to mono
    ASSERT_EQ(ac_reconfigure_encoder(ac, 48000, sampling_rate, channels), 0);

    std::vector<int16_t> pcm(sample_count * channels);
    for (size_t i = 0; i < pcm.size(); ++i) {
        pcm[i] = static_cast<int16_t>(i * 10);
    }

    std::vector<uint8_t> encoded(2000);
    int encoded_size = ac_encode(ac, pcm.data(), sample_count, encoded.data(), encoded.size());
    ASSERT_GT(encoded_size, 0);

    // Prepare payload: 4 bytes sampling rate + Opus data
    std::vector<uint8_t> payload(4 + static_cast<size_t>(encoded_size));
    uint32_t net_sr = net_htonl(sampling_rate);
    memcpy(payload.data(), &net_sr, 4);
    memcpy(payload.data() + 4, encoded.data(), static_cast<size_t>(encoded_size));

    // Send via RTP
    int rc = rtp_send_data(
        log, send_rtp, payload.data(), static_cast<uint32_t>(payload.size()), false);
    ASSERT_EQ(rc, 0);

    // Decode
    ac_iterate(ac);

    ASSERT_EQ(data.friend_number, 123u);
    ASSERT_EQ(data.sample_count, sample_count);
    ASSERT_EQ(data.channels, channels);
    ASSERT_EQ(data.sampling_rate, sampling_rate);
    ASSERT_EQ(data.last_pcm.size(), pcm.size());

    rtp_kill(log, send_rtp);
    rtp_kill(log, recv_rtp);
    ac_kill(ac);
}

TEST_F(AudioTest, EncodeDecodeRealistic)
{
    AudioTestData data;
    ACSession *ac = ac_new(mono_time, log, 123, AudioTestData::receive_frame, &data);
    ASSERT_NE(ac, nullptr);

    RtpMock rtp_mock;
    RTPSession *send_rtp = rtp_new(log, RTP_TYPE_AUDIO, mono_time, RtpMock::send_packet, &rtp_mock,
        nullptr, nullptr, nullptr, ac, RtpMock::audio_cb);
    RTPSession *recv_rtp = rtp_new(log, RTP_TYPE_AUDIO, mono_time, RtpMock::send_packet, &rtp_mock,
        nullptr, nullptr, nullptr, ac, RtpMock::audio_cb);
    rtp_mock.recv_session = recv_rtp;

    uint32_t sampling_rate = 48000;
    uint8_t channels = 1;
    size_t sample_count = 960;  // 20ms at 48kHz
    uint32_t bitrate = 48000;

    ASSERT_EQ(ac_reconfigure_encoder(ac, bitrate, sampling_rate, channels), 0);

    double frequency = 440.0;
    double amplitude = 10000.0;
    const double pi = std::acos(-1.0);

    std::vector<int16_t> all_sent;
    std::vector<int16_t> all_recv;

    for (int frame = 0; frame < 50; ++frame) {
        std::vector<int16_t> pcm(sample_count * channels);
        for (size_t i = 0; i < sample_count; ++i) {
            double t = static_cast<double>(frame * sample_count + i) / sampling_rate;
            pcm[i] = static_cast<int16_t>(std::sin(2.0 * pi * frequency * t) * amplitude);
        }
        all_sent.insert(all_sent.end(), pcm.begin(), pcm.end());

        std::vector<uint8_t> encoded(2000);
        int encoded_size = ac_encode(ac, pcm.data(), sample_count, encoded.data(), encoded.size());
        ASSERT_GT(encoded_size, 0);

        std::vector<uint8_t> payload(4 + static_cast<size_t>(encoded_size));
        uint32_t net_sr = net_htonl(sampling_rate);
        memcpy(payload.data(), &net_sr, 4);
        memcpy(payload.data() + 4, encoded.data(), static_cast<size_t>(encoded_size));

        rtp_send_data(log, send_rtp, payload.data(), static_cast<uint32_t>(payload.size()), false);

        ac_iterate(ac);

        if (data.sample_count > 0) {
            all_recv.insert(all_recv.end(), data.last_pcm.begin(), data.last_pcm.end());
        }
    }

    ASSERT_FALSE(all_recv.empty());

    // Find the best match by trying different delays.
    // Jitter buffer delay (3 frames = 2880 samples) + Opus lookahead (~312 samples) = ~3192.
    double min_mse = 1e18;
    int best_delay = 0;

    // Search around the expected delay
    for (int delay = 3000; delay < 3500; ++delay) {
        double mse = 0;
        int count = 0;
        for (size_t i = 0; i < 2000; ++i) {  // Compare a decent chunk
            if (i + delay < all_sent.size() && i < all_recv.size()) {
                int diff = all_sent[i + delay] - all_recv[i];
                mse += static_cast<double>(diff) * diff;
                count++;
            }
        }
        if (count > 1000) {
            mse /= count;
            if (mse < min_mse) {
                min_mse = mse;
                best_delay = delay;
            }
        }
    }

    printf("Best audio delay found: %d samples, Min MSE: %f\n", best_delay, min_mse);

    // For 48kbps Opus, the MSE for a sine wave should be quite low once aligned.
    // 10M is about 20% of the signal power (50M), which is a safe threshold for verification.
    EXPECT_LT(min_mse, 10000000.0);

    rtp_kill(log, send_rtp);
    rtp_kill(log, recv_rtp);
    ac_kill(ac);
}

TEST_F(AudioTest, EncodeDecodeSiren)
{
    AudioTestData data;
    ACSession *ac = ac_new(mono_time, log, 123, AudioTestData::receive_frame, &data);
    ASSERT_NE(ac, nullptr);

    RtpMock rtp_mock;
    RTPSession *send_rtp = rtp_new(log, RTP_TYPE_AUDIO, mono_time, RtpMock::send_packet, &rtp_mock,
        nullptr, nullptr, nullptr, ac, RtpMock::audio_cb);
    RTPSession *recv_rtp = rtp_new(log, RTP_TYPE_AUDIO, mono_time, RtpMock::send_packet, &rtp_mock,
        nullptr, nullptr, nullptr, ac, RtpMock::audio_cb);
    rtp_mock.recv_session = recv_rtp;

    uint32_t sampling_rate = 48000;
    uint8_t channels = 1;
    size_t sample_count = 960;  // 20ms at 48kHz
    uint32_t bitrate = 64000;

    ASSERT_EQ(ac_reconfigure_encoder(ac, bitrate, sampling_rate, channels), 0);

    double amplitude = 10000.0;
    const double pi = std::acos(-1.0);

    std::vector<int16_t> all_sent;
    std::vector<int16_t> all_recv;

    // 1 second of audio (50 frames) is enough for a siren test
    for (int frame = 0; frame < 50; ++frame) {
        std::vector<int16_t> pcm(sample_count * channels);
        for (size_t i = 0; i < sample_count; ++i) {
            double t = static_cast<double>(frame * sample_count + i) / sampling_rate;
            // Linear frequency sweep from 50Hz to 440Hz over 1 second
            // f(t) = 50 + (440-50)/1 * t = 50 + 390t
            // phi(t) = 2*pi * integral(f(t)) = 2*pi * (50t + 195t^2)
            double phi = 2.0 * pi * (50.0 * t + 195.0 * t * t);
            pcm[i] = static_cast<int16_t>(std::sin(phi) * amplitude);
        }
        all_sent.insert(all_sent.end(), pcm.begin(), pcm.end());

        std::vector<uint8_t> encoded(2000);
        int encoded_size = ac_encode(ac, pcm.data(), sample_count, encoded.data(), encoded.size());
        ASSERT_GT(encoded_size, 0);

        std::vector<uint8_t> payload(4 + static_cast<size_t>(encoded_size));
        uint32_t net_sr = net_htonl(sampling_rate);
        memcpy(payload.data(), &net_sr, 4);
        memcpy(payload.data() + 4, encoded.data(), static_cast<size_t>(encoded_size));

        rtp_send_data(log, send_rtp, payload.data(), static_cast<uint32_t>(payload.size()), false);

        ac_iterate(ac);

        if (data.sample_count > 0) {
            all_recv.insert(all_recv.end(), data.last_pcm.begin(), data.last_pcm.end());
        }
    }

    ASSERT_FALSE(all_recv.empty());

    auto calculate_mse_at = [&](int delay, size_t window) {
        double mse = 0;
        int count = 0;
        for (size_t i = 0; i < window; ++i) {
            int sent_idx = static_cast<int>(i) + delay;
            if (sent_idx >= 0 && static_cast<size_t>(sent_idx) < all_sent.size()
                && i < all_recv.size()) {
                int diff = all_sent[static_cast<size_t>(sent_idx)] - all_recv[i];
                mse += static_cast<double>(diff) * diff;
                count++;
            }
        }
        return count > 0 ? mse / count : 1e18;
    };

    // Two-stage search for speed
    double min_mse = 1e18;
    int coarse_best = 0;

    // 1. Coarse search
    for (int delay = -5000; delay < 5000; delay += 100) {
        double mse = calculate_mse_at(delay, 5000);
        if (mse < min_mse) {
            min_mse = mse;
            coarse_best = delay;
        }
    }

    // 2. Fine search around coarse best
    int best_delay = coarse_best;
    for (int delay = coarse_best - 100; delay <= coarse_best + 100; ++delay) {
        double mse = calculate_mse_at(delay, 10000);
        if (mse < min_mse) {
            min_mse = mse;
            best_delay = delay;
        }
    }

    printf("Best siren audio delay found: %d samples, Min MSE: %f\n", best_delay, min_mse);

    // For 64kbps Opus, the MSE for a siren wave should be reasonably low once aligned.
    EXPECT_LT(min_mse, 20000000.0);

    rtp_kill(log, send_rtp);
    rtp_kill(log, recv_rtp);
    ac_kill(ac);
}
TEST_F(AudioTest, ReconfigureEncoder)
{
    AudioTestData data;
    ACSession *ac = ac_new(mono_time, log, 123, AudioTestData::receive_frame, &data);
    ASSERT_NE(ac, nullptr);

    // Initial state: 48kHz, mono, 48kbps
    // Change to 24kHz, stereo, 32kbps
    int rc = ac_reconfigure_encoder(ac, 32000, 24000, 2);
    ASSERT_EQ(rc, 0);

    size_t sample_count = 480;  // 20ms at 24kHz
    uint8_t channels = 2;
    std::vector<int16_t> pcm(sample_count * channels, 0);

    std::vector<uint8_t> encoded(1000);
    int encoded_size = ac_encode(ac, pcm.data(), sample_count, encoded.data(), encoded.size());
    ASSERT_GT(encoded_size, 0);

    ac_kill(ac);
}

TEST_F(AudioTest, GetFrameDuration)
{
    AudioTestData data;
    ACSession *ac = ac_new(mono_time, log, 123, AudioTestData::receive_frame, &data);
    ASSERT_NE(ac, nullptr);

    // Default duration in audio.c is 120ms (AUDIO_MAX_FRAME_DURATION_MS)
    EXPECT_EQ(ac_get_lp_frame_duration(ac), 120u);

    ac_kill(ac);
}

TEST_F(AudioTest, QueueInvalidMessage)
{
    AudioTestData data;
    ACSession *ac = ac_new(mono_time, log, 123, AudioTestData::receive_frame, &data);
    ASSERT_NE(ac, nullptr);

    RtpMock rtp_mock;
    // Create a video RTP session but try to queue to audio session
    RTPSession *video_rtp = rtp_new(log, RTP_TYPE_VIDEO, mono_time, RtpMock::send_packet, &rtp_mock,
        nullptr, nullptr, nullptr, ac, RtpMock::audio_cb);
    RTPSession *audio_recv_rtp = rtp_new(log, RTP_TYPE_AUDIO, mono_time, RtpMock::send_packet,
        &rtp_mock, nullptr, nullptr, nullptr, ac, RtpMock::audio_cb);
    rtp_mock.recv_session = audio_recv_rtp;

    std::vector<uint8_t> dummy_video(100, 0);
    int rc = rtp_send_data(
        log, video_rtp, dummy_video.data(), static_cast<uint32_t>(dummy_video.size()), true);
    ASSERT_EQ(rc, 0);

    // Iterate should NOT trigger callback because payload type was wrong
    ac_iterate(ac);
    EXPECT_EQ(data.sample_count, 0u);

    rtp_kill(log, video_rtp);
    rtp_kill(log, audio_recv_rtp);
    ac_kill(ac);
}

TEST_F(AudioTest, JitterBufferDuplicate)
{
    AudioTestData data;
    ACSession *ac = ac_new(mono_time, log, 123, AudioTestData::receive_frame, &data);
    ASSERT_NE(ac, nullptr);

    RtpMock rtp_mock;
    rtp_mock.auto_forward = false;
    RTPSession *send_rtp = rtp_new(log, RTP_TYPE_AUDIO, mono_time, RtpMock::send_packet, &rtp_mock,
        nullptr, nullptr, nullptr, ac, RtpMock::audio_cb);
    RTPSession *recv_rtp = rtp_new(log, RTP_TYPE_AUDIO, mono_time, RtpMock::send_packet, &rtp_mock,
        nullptr, nullptr, nullptr, ac, RtpMock::audio_cb);
    rtp_mock.recv_session = recv_rtp;

    uint8_t dummy_data[100] = {0};
    uint32_t net_sr = net_htonl(48000);
    memcpy(dummy_data, &net_sr, 4);

    rtp_send_data(log, send_rtp, dummy_data, sizeof(dummy_data), false);
    ASSERT_EQ(rtp_mock.captured_packets.size(), 1u);

    // Feed the same packet twice
    rtp_receive_packet(
        recv_rtp, rtp_mock.captured_packets[0].data(), rtp_mock.captured_packets[0].size());
    rtp_receive_packet(
        recv_rtp, rtp_mock.captured_packets[0].data(), rtp_mock.captured_packets[0].size());

    // First iterate should process the packet
    ac_iterate(ac);
    EXPECT_GT(data.sample_count, 0u);
    data.sample_count = 0;

    // Second iterate should NOT process anything (duplicate was dropped in queue)
    ac_iterate(ac);
    EXPECT_EQ(data.sample_count, 0u);

    rtp_kill(log, send_rtp);
    rtp_kill(log, recv_rtp);
    ac_kill(ac);
}

TEST_F(AudioTest, JitterBufferOutOfOrder)
{
    AudioTestData data;
    ACSession *ac = ac_new(mono_time, log, 123, AudioTestData::receive_frame, &data);
    ASSERT_NE(ac, nullptr);

    RtpMock rtp_mock;
    rtp_mock.auto_forward = false;
    RTPSession *send_rtp = rtp_new(log, RTP_TYPE_AUDIO, mono_time, RtpMock::send_packet, &rtp_mock,
        nullptr, nullptr, nullptr, ac, RtpMock::audio_cb);
    RTPSession *recv_rtp = rtp_new(log, RTP_TYPE_AUDIO, mono_time, RtpMock::send_packet, &rtp_mock,
        nullptr, nullptr, nullptr, ac, RtpMock::audio_cb);
    rtp_mock.recv_session = recv_rtp;

    uint8_t dummy_data[100] = {0};
    uint32_t net_sr = net_htonl(48000);
    memcpy(dummy_data, &net_sr, 4);

    // Capture 3 packets
    rtp_send_data(log, send_rtp, dummy_data, sizeof(dummy_data), false);
    rtp_send_data(log, send_rtp, dummy_data, sizeof(dummy_data), false);
    rtp_send_data(log, send_rtp, dummy_data, sizeof(dummy_data), false);
    ASSERT_EQ(rtp_mock.captured_packets.size(), 3u);

    // Receive in order 0, 2, 1
    rtp_receive_packet(
        recv_rtp, rtp_mock.captured_packets[0].data(), rtp_mock.captured_packets[0].size());
    rtp_receive_packet(
        recv_rtp, rtp_mock.captured_packets[2].data(), rtp_mock.captured_packets[2].size());
    rtp_receive_packet(
        recv_rtp, rtp_mock.captured_packets[1].data(), rtp_mock.captured_packets[1].size());

    // Iterate once, should process all 3 packets (because ac_iterate now loops)
    data.sample_count = 0;
    ac_iterate(ac);
    EXPECT_GT(data.sample_count, 0u);

    // Subsequent iterate should find nothing
    data.sample_count = 0;
    ac_iterate(ac);
    EXPECT_EQ(data.sample_count, 0u);

    rtp_kill(log, send_rtp);
    rtp_kill(log, recv_rtp);
    ac_kill(ac);
}

TEST_F(AudioTest, PacketLossConcealment)
{
    AudioTestData data;
    ACSession *ac = ac_new(mono_time, log, 123, AudioTestData::receive_frame, &data);
    ASSERT_NE(ac, nullptr);

    RtpMock rtp_mock;
    rtp_mock.auto_forward = false;
    RTPSession *send_rtp = rtp_new(log, RTP_TYPE_AUDIO, mono_time, RtpMock::send_packet, &rtp_mock,
        nullptr, nullptr, nullptr, ac, RtpMock::audio_cb);
    RTPSession *recv_rtp = rtp_new(log, RTP_TYPE_AUDIO, mono_time, RtpMock::send_packet, &rtp_mock,
        nullptr, nullptr, nullptr, ac, RtpMock::audio_cb);
    rtp_mock.recv_session = recv_rtp;

    uint8_t dummy_data[100] = {0};
    uint32_t net_sr = net_htonl(48000);
    memcpy(dummy_data, &net_sr, 4);

    // Send packet 0 and deliver it immediately.
    rtp_send_data(log, send_rtp, dummy_data, sizeof(dummy_data), false);
    rtp_receive_packet(
        recv_rtp, rtp_mock.captured_packets[0].data(), rtp_mock.captured_packets[0].size());
    ac_iterate(ac);
    EXPECT_GT(data.sample_count, 0u);
    data.sample_count = 0;

    // Send packets 1 through 5 but do not deliver them, creating a gap in the sequence.
    for (int i = 0; i < 5; ++i) {
        rtp_send_data(log, send_rtp, dummy_data, sizeof(dummy_data), false);
    }

    // Send and deliver packet 6. The gap (1-5) exceeds the jitter buffer capacity (3).
    rtp_send_data(log, send_rtp, dummy_data, sizeof(dummy_data), false);
    rtp_receive_packet(
        recv_rtp, rtp_mock.captured_packets[6].data(), rtp_mock.captured_packets[6].size());

    // The next iteration should trigger Packet Loss Concealment (PLC) for the missing packets.
    // In audio.c, a return code of 2 from jbuf_read indicates that PLC should be performed.
    ac_iterate(ac);
    EXPECT_GT(data.sample_count, 0u);

    rtp_kill(log, send_rtp);
    rtp_kill(log, recv_rtp);
    ac_kill(ac);
}

TEST_F(AudioTest, JitterBufferReset)
{
    AudioTestData data;
    ACSession *ac = ac_new(mono_time, log, 123, AudioTestData::receive_frame, &data);
    ASSERT_NE(ac, nullptr);

    RtpMock rtp_mock;
    rtp_mock.auto_forward = false;
    RTPSession *send_rtp = rtp_new(log, RTP_TYPE_AUDIO, mono_time, RtpMock::send_packet, &rtp_mock,
        nullptr, nullptr, nullptr, ac, RtpMock::audio_cb);
    RTPSession *recv_rtp = rtp_new(log, RTP_TYPE_AUDIO, mono_time, RtpMock::send_packet, &rtp_mock,
        nullptr, nullptr, nullptr, ac, RtpMock::audio_cb);
    rtp_mock.recv_session = recv_rtp;

    uint8_t dummy_data[100] = {0};
    uint32_t net_sr = net_htonl(48000);
    memcpy(dummy_data, &net_sr, 4);

    rtp_send_data(log, send_rtp, dummy_data, sizeof(dummy_data), false);
    rtp_receive_packet(
        recv_rtp, rtp_mock.captured_packets[0].data(), rtp_mock.captured_packets[0].size());
    ac_iterate(ac);

    // The jitter buffer size is (capacity * 4) rounded up to the next power of 2.
    // With AUDIO_JITTERBUFFER_COUNT = 3, the size is 16.
    // A jump in sequence number greater than the buffer size triggers a full reset of the jitter
    // buffer.
    for (int i = 0; i < 20; ++i) {
        rtp_send_data(log, send_rtp, dummy_data, sizeof(dummy_data), false);
    }

    // Deliver the latest packet, which is well beyond the current window.
    rtp_receive_packet(
        recv_rtp, rtp_mock.captured_packets.back().data(), rtp_mock.captured_packets.back().size());

    // The session should recover after the reset and process the new packet normally.
    data.sample_count = 0;
    ac_iterate(ac);
    EXPECT_GT(data.sample_count, 0u);

    rtp_kill(log, send_rtp);
    rtp_kill(log, recv_rtp);
    ac_kill(ac);
}

TEST_F(AudioTest, DecoderReconfigureCooldown)
{
    AudioTestData data;
    ACSession *ac = ac_new(mono_time, log, 123, AudioTestData::receive_frame, &data);
    ASSERT_NE(ac, nullptr);

    RtpMock rtp_mock;
    rtp_mock.auto_forward = false;
    RTPSession *send_rtp = rtp_new(log, RTP_TYPE_AUDIO, mono_time, RtpMock::send_packet, &rtp_mock,
        nullptr, nullptr, nullptr, ac, RtpMock::audio_cb);
    RTPSession *recv_rtp = rtp_new(log, RTP_TYPE_AUDIO, mono_time, RtpMock::send_packet, &rtp_mock,
        nullptr, nullptr, nullptr, ac, RtpMock::audio_cb);
    rtp_mock.recv_session = recv_rtp;

    uint8_t dummy_data[100] = {0};
    uint32_t net_sr_48 = net_htonl(48000);
    uint32_t net_sr_24 = net_htonl(24000);

    // 1. Reconfigure to 24kHz. The initial sampling rate is 48kHz.
    memcpy(dummy_data, &net_sr_24, 4);
    rtp_send_data(log, send_rtp, dummy_data, sizeof(dummy_data), false);
    rtp_receive_packet(
        recv_rtp, rtp_mock.captured_packets.back().data(), rtp_mock.captured_packets.back().size());
    ac_iterate(ac);
    EXPECT_EQ(data.sampling_rate, 24000u);
    data.sampling_rate = 0;

    // 2. Advance time by only 100ms. This is less than the 500ms cooldown required for decoder
    // reconfiguration.
    tm.t += 100;
    mono_time_update(mono_time);

    // 3. Attempt to reconfigure back to 48kHz.
    memcpy(dummy_data, &net_sr_48, 4);
    rtp_send_data(log, send_rtp, dummy_data, sizeof(dummy_data), false);
    rtp_receive_packet(
        recv_rtp, rtp_mock.captured_packets.back().data(), rtp_mock.captured_packets.back().size());

    // Reconfiguration should be rejected due to the cooldown, so the callback should not be
    // invoked.
    ac_iterate(ac);
    EXPECT_EQ(data.sampling_rate, 0u);

    // 4. Advance time beyond the 500ms cooldown period (measured from the first reconfiguration).
    tm.t += 500;
    mono_time_update(mono_time);

    // 5. Attempt reconfiguration to 48kHz again.
    rtp_send_data(log, send_rtp, dummy_data, sizeof(dummy_data), false);
    rtp_receive_packet(
        recv_rtp, rtp_mock.captured_packets.back().data(), rtp_mock.captured_packets.back().size());

    // Reconfiguration should now succeed.
    ac_iterate(ac);
    EXPECT_EQ(data.sampling_rate, 48000u);

    rtp_kill(log, send_rtp);
    rtp_kill(log, recv_rtp);
    ac_kill(ac);
}

TEST_F(AudioTest, QueueDummyMessage)
{
    AudioTestData data;
    ACSession *ac = ac_new(mono_time, log, 123, AudioTestData::receive_frame, &data);
    ASSERT_NE(ac, nullptr);

    RtpMock rtp_mock;
    // RTP_TYPE_AUDIO + 2 is the dummy type
    RTPSession *dummy_rtp = rtp_new(log, RTP_TYPE_AUDIO + 2, mono_time, RtpMock::send_packet,
        &rtp_mock, nullptr, nullptr, nullptr, ac, RtpMock::audio_cb);
    RTPSession *audio_recv_rtp = rtp_new(log, RTP_TYPE_AUDIO, mono_time, RtpMock::send_packet,
        &rtp_mock, nullptr, nullptr, nullptr, ac, RtpMock::audio_cb);
    rtp_mock.recv_session = audio_recv_rtp;

    std::vector<uint8_t> dummy_payload(100, 0);
    int rc = rtp_send_data(
        log, dummy_rtp, dummy_payload.data(), static_cast<uint32_t>(dummy_payload.size()), false);
    ASSERT_EQ(rc, 0);

    // Iterate should NOT trigger callback because it was a dummy packet
    ac_iterate(ac);
    EXPECT_EQ(data.sample_count, 0u);

    rtp_kill(log, dummy_rtp);
    rtp_kill(log, audio_recv_rtp);
    ac_kill(ac);
}

TEST_F(AudioTest, LatePacketReset)
{
    AudioTestData data;
    ACSession *ac = ac_new(mono_time, log, 123, AudioTestData::receive_frame, &data);
    ASSERT_NE(ac, nullptr);

    RtpMock rtp_mock;
    rtp_mock.auto_forward = false;
    RTPSession *send_rtp = rtp_new(log, RTP_TYPE_AUDIO, mono_time, RtpMock::send_packet, &rtp_mock,
        nullptr, nullptr, nullptr, ac, RtpMock::audio_cb);
    RTPSession *recv_rtp = rtp_new(log, RTP_TYPE_AUDIO, mono_time, RtpMock::send_packet, &rtp_mock,
        nullptr, nullptr, nullptr, ac, RtpMock::audio_cb);
    rtp_mock.recv_session = recv_rtp;

    uint8_t dummy_data[100] = {0};
    uint32_t net_sr = net_htonl(48000);
    memcpy(dummy_data, &net_sr, 4);

    // 1. Send and process the first packet.
    rtp_send_data(log, send_rtp, dummy_data, sizeof(dummy_data), false);  // seq 0
    rtp_receive_packet(
        recv_rtp, rtp_mock.captured_packets[0].data(), rtp_mock.captured_packets[0].size());
    ac_iterate(ac);
    ASSERT_GT(data.sample_count, 0u);
    data.sample_count = 0;

    // 2. Buffer another packet with a different sampling rate (24kHz) but don't process it yet.
    uint32_t net_sr_24 = net_htonl(24000);
    memcpy(dummy_data, &net_sr_24, 4);
    rtp_send_data(log, send_rtp, dummy_data, sizeof(dummy_data), false);  // seq 1
    rtp_receive_packet(
        recv_rtp, rtp_mock.captured_packets[1].data(), rtp_mock.captured_packets[1].size());

    // 3. Receive the late packet (seq 0) again.
    // This triggers the bug: (uint32_t)(0 - 1) > 16, causing a full jitter buffer reset.
    rtp_receive_packet(
        recv_rtp, rtp_mock.captured_packets[0].data(), rtp_mock.captured_packets[0].size());

    // 4. Try to process the next packet.
    // Due to the bug, packet 1 was cleared. We will likely get PLC (48kHz) instead of packet 1
    // (24kHz).
    ac_iterate(ac);

    // If the bug is present, sampling_rate will be 48000 (from PLC) instead of 24000.
    EXPECT_EQ(data.sampling_rate, 24000u);

    rtp_kill(log, send_rtp);
    rtp_kill(log, recv_rtp);
    ac_kill(ac);
}

TEST_F(AudioTest, InvalidSamplingRate)
{
    AudioTestData data;
    ACSession *ac = ac_new(mono_time, log, 123, AudioTestData::receive_frame, &data);
    ASSERT_NE(ac, nullptr);

    RtpMock rtp_mock;
    rtp_mock.auto_forward = false;
    RTPSession *send_rtp = rtp_new(log, RTP_TYPE_AUDIO, mono_time, RtpMock::send_packet, &rtp_mock,
        nullptr, nullptr, nullptr, ac, RtpMock::audio_cb);
    RTPSession *recv_rtp = rtp_new(log, RTP_TYPE_AUDIO, mono_time, RtpMock::send_packet, &rtp_mock,
        nullptr, nullptr, nullptr, ac, RtpMock::audio_cb);
    rtp_mock.recv_session = recv_rtp;

    // 1. Send a packet with an absurdly large sampling rate.
    uint8_t malicious_data[100] = {0};
    uint32_t net_sr = net_htonl(1000000000);  // 1 GHz
    memcpy(malicious_data, &net_sr, 4);
    // Add some dummy Opus data so it's not too short
    malicious_data[4] = 0x08;

    rtp_send_data(log, send_rtp, malicious_data, sizeof(malicious_data), false);
    rtp_receive_packet(
        recv_rtp, rtp_mock.captured_packets.back().data(), rtp_mock.captured_packets.back().size());

    // This packet should fail reconfiguration and be discarded.
    ac_iterate(ac);
    EXPECT_EQ(data.sample_count, 0u);

    // 2. Trigger PLC. It should NOT use the malicious sampling rate.
    // Send 5 packets to create a gap.
    for (int i = 0; i < 5; ++i) {
        rtp_send_data(log, send_rtp, malicious_data, sizeof(malicious_data), false);
    }
    // Deliver the next one.
    rtp_send_data(log, send_rtp, malicious_data, sizeof(malicious_data), false);
    rtp_receive_packet(
        recv_rtp, rtp_mock.captured_packets.back().data(), rtp_mock.captured_packets.back().size());

    // Next iterate triggers PLC. If it uses 1GHz, it might overflow/crash.
    ac_iterate(ac);

    rtp_kill(log, send_rtp);
    rtp_kill(log, recv_rtp);
    ac_kill(ac);
}

TEST_F(AudioTest, ShortPacket)
{
    AudioTestData data;
    ACSession *ac = ac_new(mono_time, log, 123, AudioTestData::receive_frame, &data);
    ASSERT_NE(ac, nullptr);

    RtpMock rtp_mock;
    rtp_mock.auto_forward = false;
    RTPSession *send_rtp = rtp_new(log, RTP_TYPE_AUDIO, mono_time, RtpMock::send_packet, &rtp_mock,
        nullptr, nullptr, nullptr, ac, RtpMock::audio_cb);
    RTPSession *recv_rtp = rtp_new(log, RTP_TYPE_AUDIO, mono_time, RtpMock::send_packet, &rtp_mock,
        nullptr, nullptr, nullptr, ac, RtpMock::audio_cb);
    rtp_mock.recv_session = recv_rtp;

    // 1. Send a packet that is too short (only sampling rate, no Opus data).
    // The protocol requires 4 bytes SR + at least 1 byte Opus data.
    uint8_t short_data[4] = {0, 0, 0xBB, 0x80};  // 48000

    // rtp_send_data might not like 4 bytes if it expects more, but let's see.
    rtp_send_data(log, send_rtp, short_data, sizeof(short_data), false);
    rtp_receive_packet(
        recv_rtp, rtp_mock.captured_packets.back().data(), rtp_mock.captured_packets.back().size());

    // This should not crash. In debug it might hit an assert.
    // In production it might do an OOB read.
    ac_iterate(ac);

    rtp_kill(log, send_rtp);
    rtp_kill(log, recv_rtp);
    ac_kill(ac);
}

TEST_F(AudioTest, JitterBufferWrapAround)
{
    AudioTestData data;
    ACSession *ac = ac_new(mono_time, log, 123, AudioTestData::receive_frame, &data);
    ASSERT_NE(ac, nullptr);

    RtpMock rtp_mock;
    rtp_mock.auto_forward = false;
    RTPSession *send_rtp = rtp_new(log, RTP_TYPE_AUDIO, mono_time, RtpMock::send_packet, &rtp_mock,
        nullptr, nullptr, nullptr, ac, RtpMock::audio_cb);
    RTPSession *recv_rtp = rtp_new(log, RTP_TYPE_AUDIO, mono_time, RtpMock::send_packet, &rtp_mock,
        nullptr, nullptr, nullptr, ac, RtpMock::audio_cb);
    rtp_mock.recv_session = recv_rtp;

    uint8_t dummy_data[100] = {0};
    uint32_t net_sr = net_htonl(48000);
    memcpy(dummy_data, &net_sr, 4);

    // Send enough packets to reach the sequence number wrap-around point (0xFFFF -> 0x0000).
    // We detect the current sequence number to minimize the number of iterations.
    uint16_t seq = 0;
    {
        rtp_send_data(log, send_rtp, dummy_data, sizeof(dummy_data), false);
        const uint8_t *pkt = rtp_mock.captured_packets.back().data();
        seq = (pkt[3] << 8) | pkt[4];
        rtp_receive_packet(recv_rtp, pkt, rtp_mock.captured_packets.back().size());
        rtp_mock.captured_packets.clear();
        ac_iterate(ac);
    }

    // Aim for sequence number 65532 to be the last processed packet before the gap.
    int to_send = (65532 - seq + 65536) % 65536;
    for (int i = 0; i < to_send; ++i) {
        rtp_send_data(log, send_rtp, dummy_data, sizeof(dummy_data), false);
        const uint8_t *pkt = rtp_mock.captured_packets.back().data();
        rtp_receive_packet(recv_rtp, pkt, rtp_mock.captured_packets.back().size());
        rtp_mock.captured_packets.clear();
        ac_iterate(ac);
    }

    // Now 'bottom' should be at 65533 (next expected).
    data.sample_count = 0;

    // Create a gap of 2 missing packets: 65533, 65534.
    // Packet 65535 is delivered. Gap is 2. Capacity is 3. Should NOT trigger PLC.
    rtp_send_data(log, send_rtp, dummy_data, sizeof(dummy_data), false);  // 65533 (dropped)
    rtp_send_data(log, send_rtp, dummy_data, sizeof(dummy_data), false);  // 65534 (dropped)
    rtp_send_data(log, send_rtp, dummy_data, sizeof(dummy_data), false);  // 65535 (delivered)
    rtp_receive_packet(
        recv_rtp, rtp_mock.captured_packets.back().data(), rtp_mock.captured_packets.back().size());

    // Iteration should result in no frames processed because the gap is within capacity.
    // If there is a bug in wrap-around distance calculation, it will trigger PLC here.
    ac_iterate(ac);
    EXPECT_EQ(data.sample_count, 0u);

    rtp_kill(log, send_rtp);
    rtp_kill(log, recv_rtp);
    ac_kill(ac);
}

}  // namespace