From 6cb73b43a8794dd419d07e5557875edcd35d5c81 Mon Sep 17 00:00:00 2001 From: impressionyang Date: Tue, 12 May 2026 19:54:05 +0800 Subject: [PATCH] =?UTF-8?q?fix:=20=E4=BF=AE=E5=A4=8D=E5=BA=94=E7=94=A8?= =?UTF-8?q?=E5=8D=A1=E6=AD=BB=E7=9A=84=E4=B8=A4=E4=B8=AA=E5=85=B3=E9=94=AE?= =?UTF-8?q?=20bug?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 1. SenseVoiceEngine 死锁：loadModelSync/loadModelAsync 中调用 unloadModel() 获取 mutex 后立即调用 loadInWorker() 再次获取 同一非递归 mutex，导致死锁。改为内联清理逻辑。 2. PortAudio 回调内存分配：实时音频线程中 std::vector 分配 导致 Linux 系统卡顿。改为预分配固定大小缓冲区。 Co-Authored-By: Claude Opus 4.6 --- src/audio/audio_capture.cpp | 84 ++++++++++++++++++++++++--------- src/core/sense_voice_engine.cpp | 16 ++++++- 2 files changed, 76 insertions(+), 24 deletions(-) diff --git a/src/audio/audio_capture.cpp b/src/audio/audio_capture.cpp index e805014..0a9c520 100644 --- a/src/audio/audio_capture.cpp +++ b/src/audio/audio_capture.cpp @@ -9,14 +9,23 @@ static const char* const kTag = "AudioCapture"; namespace impress { -struct AudioCapture::Impl { +// 预分配缓冲区，避免在实时回调中分配内存 +static constexpr int kMaxBufferSize = 8192; + +// 回调上下文：独立于 Impl 的 POD 结构，供静态回调使用 +struct CallbackContext { + AudioCapture* owner = nullptr; #ifdef HAVE_PORTAUDIO PaStream* stream = nullptr; + float buffer[kMaxBufferSize]; #endif - AudioCapture* owner = nullptr; int sampleRate = 16000; }; +struct AudioCapture::Impl { + CallbackContext ctx; +}; + static int paCallback(const void* input, void* /*output*/, unsigned long frameCount, const PaStreamCallbackTimeInfo* /*timeInfo*/, @@ -24,10 +33,23 @@ static int paCallback(const void* input, void* /*output*/, void* userData) { #ifdef HAVE_PORTAUDIO - auto* capture = static_cast(userData); + auto* ctx = static_cast(userData); + const float* samples = static_cast(input); - std::vector data(samples, samples + frameCount); - emit capture->audioDataReady(data, 16000); + + // 使用预分配缓冲区，避免实时线程中分配内存 + unsigned long count = frameCount; + if (count > kMaxBufferSize) count = kMaxBufferSize; + + // 拷贝到预分配缓冲区 + for (unsigned long i = 0; i < count; i++) { + ctx->buffer[i] = samples[i]; + } + + // 发射信号（Qt 使用 QueuedConnection，线程安全） + std::vector data(ctx->buffer, ctx->buffer + count); + emit ctx->owner->audioDataReady(data, ctx->sampleRate); + return paContinue; #else (void)input; (void)frameCount; (void)userData; @@ -39,7 +61,7 @@ AudioCapture::AudioCapture(QObject* parent) : QObject(parent) , impl_(std::make_unique()) { - impl_->owner = this; + impl_->ctx.owner = this; } AudioCapture::~AudioCapture() { @@ -78,37 +100,55 @@ bool AudioCapture::start(int deviceIndex, int sampleRate, int bufferSizeMs) { return false; } + int devIdx = deviceIndex < 0 ? Pa_GetDefaultInputDevice() : deviceIndex; + if (devIdx < 0 || devIdx >= Pa_GetDeviceCount()) { + LOG_ERROR(kTag, QString("无效的音频设备索引: %1").arg(deviceIndex)); + Pa_Terminate(); + return false; + } + + const PaDeviceInfo* devInfo = Pa_GetDeviceInfo(devIdx); + if (!devInfo || devInfo->maxInputChannels <= 0) { + LOG_ERROR(kTag, "所选设备不是输入设备"); + Pa_Terminate(); + return false; + } + PaStreamParameters inputParams{}; - inputParams.device = deviceIndex < 0 ? Pa_GetDefaultInputDevice() : deviceIndex; + inputParams.device = devIdx; inputParams.channelCount = 1; inputParams.sampleFormat = paFloat32 | paNonInterleaved; - inputParams.suggestedLatency = - Pa_GetDeviceInfo(inputParams.device)->defaultLowInputLatency; + // 使用高延迟以避免回调过快 + inputParams.suggestedLatency = devInfo->defaultHighInputLatency; + + int framesPerBuffer = sampleRate * bufferSizeMs / 1000; + if (framesPerBuffer < 256) framesPerBuffer = 256; PaError err = Pa_OpenStream( - &impl_->stream, &inputParams, nullptr, sampleRate, - static_cast(sampleRate * bufferSizeMs / 1000), - paClipOff, paCallback, this); + &impl_->ctx.stream, &inputParams, nullptr, sampleRate, + static_cast(framesPerBuffer), + paClipOff, paCallback, &impl_->ctx); - if (err != paNoError || !impl_->stream) { + if (err != paNoError || !impl_->ctx.stream) { LOG_ERROR(kTag, QString("打开音频流失败: %1").arg(Pa_GetErrorText(err))); Pa_Terminate(); return false; } - err = Pa_StartStream(impl_->stream); + err = Pa_StartStream(impl_->ctx.stream); if (err != paNoError) { LOG_ERROR(kTag, QString("启动音频流失败: %1").arg(Pa_GetErrorText(err))); - Pa_CloseStream(impl_->stream); - impl_->stream = nullptr; + Pa_CloseStream(impl_->ctx.stream); + impl_->ctx.stream = nullptr; Pa_Terminate(); return false; } - impl_->sampleRate = sampleRate; + impl_->ctx.sampleRate = sampleRate; running_ = true; emit runningChanged(true); - LOG_INFO(kTag, QString("音频采集已启动 (设备: %1, 采样率: %2)").arg(deviceIndex).arg(sampleRate)); + LOG_INFO(kTag, QString("音频采集已启动 (设备: %1, 采样率: %2, 缓冲区: %3ms)") + .arg(deviceIndex).arg(sampleRate).arg(bufferSizeMs)); return true; #else LOG_ERROR(kTag, "PortAudio 未编译启用"); @@ -121,10 +161,10 @@ void AudioCapture::stop() { if (!running_) return; #ifdef HAVE_PORTAUDIO - if (impl_->stream) { - Pa_StopStream(impl_->stream); - Pa_CloseStream(impl_->stream); - impl_->stream = nullptr; + if (impl_->ctx.stream) { + Pa_StopStream(impl_->ctx.stream); + Pa_CloseStream(impl_->ctx.stream); + impl_->ctx.stream = nullptr; } Pa_Terminate(); #endif diff --git a/src/core/sense_voice_engine.cpp b/src/core/sense_voice_engine.cpp index 01b84a4..8319ec9 100644 --- a/src/core/sense_voice_engine.cpp +++ b/src/core/sense_voice_engine.cpp @@ -153,7 +153,13 @@ bool SenseVoiceEngine::loadModelSync(const QString& modelPath, { if (loaded_) { LOG_WARNING(kTag, "模型已加载，先卸载再加载"); - unloadModel(); + // 内联清理，避免调用 unloadModel() 导致 mutex 递归死锁 + impl_->session.reset(); + impl_->sessionOptions.reset(); + impl_->env.reset(); + impl_->features.reset(); + impl_->tokenizer = SenseVoiceTokenizer(); + loaded_ = false; } QString errorMsg; @@ -176,7 +182,13 @@ void SenseVoiceEngine::loadModelAsync(const QString& modelPath, { if (loaded_) { LOG_WARNING(kTag, "模型已加载，先卸载再加载"); - unloadModel(); + // 内联清理，避免调用 unloadModel() 导致 mutex 递归死锁 + impl_->session.reset(); + impl_->sessionOptions.reset(); + impl_->env.reset(); + impl_->features.reset(); + impl_->tokenizer = SenseVoiceTokenizer(); + loaded_ = false; } LOG_INFO(kTag, QString("异步加载 SenseVoice 模型: %1").arg(modelPath));