bug处理

.cursor/mcp.json

@@ -0,0 +1,16 @@
+{
+  "mcpServers": {
+    "ssh": {
+      "command": "cmd.exe",
+      "args": [
+        "/c",
+        "npx",
+        "-y",
+        "@aiondadotcom/mcp-ssh"
+      ],
+      "env": {
+        "ProgramData": "C:\\ProgramData"
+      }
+    }
+  }
+}

internal/service/transcoder.go

@@ -49,6 +49,9 @@ type TranscoderService struct {
 
 	mu          sync.Mutex
 	jobs        map[string]*hlsJob
+	// startGates serializes EnsureJobFrom / StopJob per media so concurrent
+	// playlist hits cannot spawn multiple ffmpeg writers into one HLS dir.
+	startGates  sync.Map // mediaID -> *sync.Mutex
 	strmResolve func(ctx context.Context, raw string) (*StrmPlayResult, error)
 	probe       *FFprobeService
 }
@@ -65,6 +68,8 @@ type hlsJob struct {
 	// startSec is the source seek offset fed to ffmpeg (-ss). The HLS
 	// playlist itself always starts at t=0 for that session.
 	startSec float64
+	// done is closed when the ffmpeg goroutine fully exits (after process death).
+	done chan struct{}
 }
 
 var (
@@ -122,6 +127,10 @@ func (t *TranscoderService) EnsureJobFrom(ctx context.Context, mediaID string, s
 		return "", ErrMediaNotFound
 	}
 
+	gate := t.mediaStartGate(mediaID)
+	gate.Lock()
+	defer gate.Unlock()
+
 	t.mu.Lock()
 	if existing, ok := t.jobs[mediaID]; ok {
 		if sameHLSStart(existing.startSec, startSec) {
@@ -129,10 +138,12 @@ func (t *TranscoderService) EnsureJobFrom(ctx context.Context, mediaID string, s
 			t.mu.Unlock()
 			return t.PlaylistPath(mediaID), nil
 		}
-		existing.cancel()
-		delete(t.jobs, mediaID)
+		prev := t.detachJobLocked(mediaID)
+		t.mu.Unlock()
+		waitJobExit(prev, 12*time.Second)
+	} else {
+		t.mu.Unlock()
 	}
-	t.mu.Unlock()
 
 	input, err := t.resolveTranscodeInput(ctx, m)
 	if err != nil {
@@ -146,19 +157,24 @@ func (t *TranscoderService) EnsureJobFrom(ctx context.Context, mediaID string, s
 
 	outDir := t.HLSDir(mediaID)
 	// Wipe prior segments so a mid-file restart cannot serve stale early chunks.
+	// Only safe after the previous ffmpeg has exited (waited above / via gate).
 	if err := resetHLSDir(outDir); err != nil {
 		return "", err
 	}
 
 	t.mu.Lock()
 	if existing, ok := t.jobs[mediaID]; ok {
+		// Another path (StopJob then a peer Ensure) should be rare under the gate;
+		// still reuse or replace safely.
 		if sameHLSStart(existing.startSec, startSec) {
 			t.touchJobLocked(mediaID)
 			t.mu.Unlock()
 			return t.PlaylistPath(mediaID), nil
 		}
-		existing.cancel()
-		delete(t.jobs, mediaID)
+		prev := t.detachJobLocked(mediaID)
+		t.mu.Unlock()
+		waitJobExit(prev, 12*time.Second)
+		t.mu.Lock()
 	}
 	if max := t.maxConcurrent(); max > 0 && len(t.jobs) >= max {
 		t.mu.Unlock()
@@ -174,15 +190,48 @@ func (t *TranscoderService) EnsureJobFrom(ctx context.Context, mediaID string, s
 		lastAccess: time.Now(),
 		encoder:    t.effectiveEncoder(),
 		startSec:   startSec,
+		done:       make(chan struct{}),
 	}
 	t.jobs[mediaID] = job
 	t.mu.Unlock()
 
 	helper.Go(t.log, "transcoder.monitorIdle", func() { t.monitorIdle(jobCtx, job) })
-	helper.Go(t.log, "transcoder.ffmpeg", func() { t.runFFmpeg(jobCtx, job, input) })
+	helper.Go(t.log, "transcoder.ffmpeg", func() {
+		defer close(job.done)
+		t.runFFmpeg(jobCtx, job, input)
+	})
 	return t.PlaylistPath(mediaID), nil
 }
 
+func (t *TranscoderService) mediaStartGate(mediaID string) *sync.Mutex {
+	v, _ := t.startGates.LoadOrStore(mediaID, &sync.Mutex{})
+	return v.(*sync.Mutex)
+}
+
+// detachJobLocked cancels and removes a job from the map without waiting.
+// Caller must hold t.mu and must waitJobExit afterwards before wiping the HLS dir.
+func (t *TranscoderService) detachJobLocked(mediaID string) *hlsJob {
+	j, ok := t.jobs[mediaID]
+	if !ok {
+		return nil
+	}
+	j.cancel()
+	delete(t.jobs, mediaID)
+	return j
+}
+
+func waitJobExit(job *hlsJob, timeout time.Duration) {
+	if job == nil || job.done == nil {
+		return
+	}
+	timer := time.NewTimer(timeout)
+	defer timer.Stop()
+	select {
+	case <-job.done:
+	case <-timer.C:
+	}
+}
+
 func sameHLSStart(a, b float64) bool {
 	const tol = 0.75
 	if a < b {

internal/service/transcoder_cmd_unix.go

@@ -0,0 +1,14 @@
+//go:build unix
+
+package service
+
+import (
+	"os/exec"
+	"syscall"
+)
+
+// setFFmpegSysProcAttr puts ffmpeg in its own process group so cancel can
+// tear down the whole group (and any helpers) reliably on Linux/macOS.
+func setFFmpegSysProcAttr(cmd *exec.Cmd) {
+	cmd.SysProcAttr = &syscall.SysProcAttr{Setpgid: true}
+}

internal/service/transcoder_cmd_windows.go

@@ -0,0 +1,9 @@
+//go:build windows
+
+package service
+
+import "os/exec"
+
+func setFFmpegSysProcAttr(cmd *exec.Cmd) {
+	// Windows: CommandContext cancel is enough for the single ffmpeg process.
+}

internal/service/transcoder_ffmpeg_probe.go

@@ -34,6 +34,7 @@ func (t *TranscoderService) runFFmpeg(ctx context.Context, job *hlsJob, input tr
 	args := buildFFmpegArgsForInput(t.cfg, input, playlist, segments)
 
 	cmd := exec.CommandContext(ctx, bin, args...) // #nosec G204 -- bin is resolved by resolveFFmpegPath and args are passed without a shell.
+	setFFmpegSysProcAttr(cmd)
 	cmd.Stderr = os.Stderr
 
 	t.log.Info("transcode started",
@@ -43,20 +44,25 @@ func (t *TranscoderService) runFFmpeg(ctx context.Context, job *hlsJob, input tr
 		zap.Float64("start_sec", input.StartSec),
 	)
 	t.hub.Publish("transcode", map[string]any{
-		"media_id": job.mediaID,
-		"encoder":  job.encoder,
-		"status":   "started",
+		"media_id":  job.mediaID,
+		"encoder":   job.encoder,
+		"status":    "started",
+		"start_sec": input.StartSec,
 	})
 
 	if err := cmd.Run(); err != nil && !errors.Is(ctx.Err(), context.Canceled) {
 		t.log.Warn("ffmpeg exited",
 			zap.String("media_id", job.mediaID),
+			zap.Float64("start_sec", input.StartSec),
 			zap.Error(err),
 		)
 	}
 
 	t.mu.Lock()
-	delete(t.jobs, job.mediaID)
+	// Only drop the map entry if we are still the registered generation.
+	if cur, ok := t.jobs[job.mediaID]; ok && cur == job {
+		delete(t.jobs, job.mediaID)
+	}
 	t.mu.Unlock()
 
 	t.hub.Publish("transcode", map[string]any{

internal/service/transcoder_jobs.go

@@ -45,14 +45,18 @@ func (t *TranscoderService) WaitReady(ctx context.Context, mediaID string, timeo
 	}
 }
 
-// StopJob cancels a running ffmpeg process for mediaID, if any.
+// StopJob cancels a running ffmpeg process for mediaID, if any, and waits
+// briefly for it to exit so a subsequent EnsureJobFrom cannot race-write the
+// same HLS directory.
 func (t *TranscoderService) StopJob(mediaID string) {
+	gate := t.mediaStartGate(mediaID)
+	gate.Lock()
+	defer gate.Unlock()
+
 	t.mu.Lock()
-	defer t.mu.Unlock()
-	if j, ok := t.jobs[mediaID]; ok {
-		j.cancel()
-		delete(t.jobs, mediaID)
-	}
+	prev := t.detachJobLocked(mediaID)
+	t.mu.Unlock()
+	waitJobExit(prev, 12*time.Second)
 }
 
 // TouchJob records client activity for the HLS playlist or segment. The idle
@@ -73,10 +77,13 @@ func (t *TranscoderService) touchJobLocked(mediaID string) {
 // StopAll terminates every running transcode (called on graceful shutdown).
 func (t *TranscoderService) StopAll() {
 	t.mu.Lock()
-	defer t.mu.Unlock()
-	for id, j := range t.jobs {
-		j.cancel()
-		delete(t.jobs, id)
+	pending := make([]*hlsJob, 0, len(t.jobs))
+	for id := range t.jobs {
+		pending = append(pending, t.detachJobLocked(id))
+	}
+	t.mu.Unlock()
+	for _, j := range pending {
+		waitJobExit(j, 5*time.Second)
 	}
 }
 

web/src/pages/PlayerPage.tsx

@@ -242,13 +242,21 @@ export function PlayerPage() {
   }, [id, modeParam, directOnly])
 
   // Wire up the actual <video> element when we know the mode.
+  // Depend on media.id (not the media object): refreshing duration after
+  // MANIFEST_PARSED must not remount HLS or it storms EnsureJob / DELETE.
+  const mediaId = media?.id
+  const mediaRef = useRef(media)
+  mediaRef.current = media
   useEffect(() => {
-    if (!media || !ref.current) return
+    if (!mediaId || !ref.current) return
+    const currentMedia = mediaRef.current
+    if (!currentMedia) return
     teardownHls()
 
     const video = ref.current
+    const durationSec = currentMedia.duration_sec || 0
     if (mode === 'hls') {
-      const url = hlsURL(media.id, hlsStartSec)
+      const url = hlsURL(mediaId, hlsStartSec)
       void import('hls.js').then(({ default: HlsCtor }) => {
         if (HlsCtor.isSupported()) {
           const hls = new HlsCtor({ enableWorker: true, lowLatencyMode: false })
@@ -256,9 +264,9 @@ export function PlayerPage() {
           hls.attachMedia(video)
           hls.on(HlsCtor.Events.MANIFEST_PARSED, () => {
             // .strm 入库时常缺 duration；转码启动时会补探测，这里刷新一次给进度条总时长。
-            if ((media.duration_sec || 0) > 0) return
+            if (durationSec > 0) return
             mediaAPI
-              .get(media.id)
+              .get(mediaId)
               .then((fresh) => {
                 if ((fresh.duration_sec || 0) > 0) setMedia(fresh)
               })
@@ -290,23 +298,22 @@ export function PlayerPage() {
         setMode('direct')
       })
     } else {
-      video.src = streamURL(media.id)
-      if (hlsUnavailable && needsTranscodeForBrowser(media)) {
+      video.src = streamURL(mediaId)
+      if (hlsUnavailable && needsTranscodeForBrowser(currentMedia)) {
         setPlayerError('当前正在直连播放原始文件；此封装或音轨浏览器兼容性有限，可能只有画面没有声音。请配置本机 ffmpeg 后切回 HLS 转码播放。')
       }
       void video.play().catch(() => undefined)
     }
     return () => teardownHls()
-  }, [hlsUnavailable, hlsStartSec, media, mode, params, setParams, teardownHls])
+  }, [hlsUnavailable, hlsStartSec, mediaId, mode, params, setParams, teardownHls])
 
   // Stop the host ffmpeg job only when leaving HLS for this media (not on mid-file seek restarts).
   useEffect(() => {
-    if (!media || mode !== 'hls') return
-    const mediaId = media.id
+    if (!mediaId || mode !== 'hls') return
     return () => {
       api.delete(`/hls/${encodeURIComponent(mediaId)}`).catch(() => undefined)
     }
-  }, [media, mode])
+  }, [mediaId, mode])
 
   // 自动拉取已有的播放进度并恢复播放位置
   useEffect(() => {