优化，排查项目问题

2026-09-30 11:46:36 +08:00 · 2026-09-05 12:34:17 +08:00
parent 7fa05391e1
commit 1407b9b5c4
85 changed files with 1940 additions and 638 deletions
@@ -6,6 +6,7 @@ import (
 	"errors"
 	"fmt"
 	"strings"
+	"time"

 	"github.com/glebarez/sqlite"
 	"go.uber.org/zap"
@@ -73,6 +74,14 @@ func configureConnectionPool(db *gorm.DB, cfg *config.Config) error {
 	if cfg.Database.MaxIdleConns > 0 {
 		sqlDB.SetMaxIdleConns(cfg.Database.MaxIdleConns)
 	}
+	// 连接生命周期：默认 0 意味着 Postgres 重启/故障切换后的陈旧连接
+	// 永不过期，首次复用才报错，运行期断连恢复慢且可能批量报错。
+	if isPostgres(db) {
+		sqlDB.SetConnMaxLifetime(time.Hour)
+		sqlDB.SetConnMaxIdleTime(10 * time.Minute)
+	} else if isSQLite(db) {
+		sqlDB.SetConnMaxLifetime(24 * time.Hour)
+	}
 	return nil
 }

@@ -9,7 +9,7 @@ const mediaSearchIndexSchemaVersion = 2

 func ensureMediaSearchIndex(db *gorm.DB) error {
 	if err := ensureMediaSearchMetaTable(db); err != nil {
-		return nil
+		return err // meta 表创建失败必须上抛，不能静默掩盖
 	}
 	version := currentMediaSearchIndexVersion(db)
 	if version != mediaSearchIndexSchemaVersion {
@@ -132,13 +132,19 @@ func ensureEmbyMountsCompatibility(db *gorm.DB) error {
 			return err
 		}
 	}
-	// 针对已有数据：如果存在多个 sort_order=0/NULL 的记录，按创建时间顺序赋予稳定递增的序号
+	// 针对已有数据：只给 sort_order=0/NULL 的行按创建时间补号（从现有
+	// 最大值之后递增），不能整表重排——此前无条件按 created_at 从 0 重新
+	// 编号，会把用户自定义的顺序覆盖掉。
 	var zeroCount int64
-	if err := db.Model(&model.EmbyMount{}).Where("sort_order = 0 OR sort_order IS NULL").Count(&zeroCount).Error; err == nil && zeroCount > 1 {
+	if err := db.Model(&model.EmbyMount{}).Where("sort_order = 0 OR sort_order IS NULL").Count(&zeroCount).Error; err == nil && zeroCount > 0 {
+		// max 只统计非 0 行：sort_order=0 与 NULL 同样视为“未分配”，
+		// 全部为 0 时从 0 开始编号（与迁移前的初始化语义一致）。
+		var maxOrder int
+		_ = db.Raw("SELECT COALESCE(MAX(sort_order), -1) FROM emby_mounts WHERE sort_order > 0").Scan(&maxOrder).Error
 		var mounts []model.EmbyMount
-		if err := db.Order("created_at asc, id asc").Find(&mounts).Error; err == nil {
+		if err := db.Where("sort_order = 0 OR sort_order IS NULL").Order("created_at asc, id asc").Find(&mounts).Error; err == nil {
 			for i, m := range mounts {
-				_ = db.Exec("UPDATE emby_mounts SET sort_order = ? WHERE id = ?", i, m.ID).Error
+				_ = db.Exec("UPDATE emby_mounts SET sort_order = ? WHERE id = ?", maxOrder+1+i, m.ID).Error
 			}
 		}
 	}
@@ -50,28 +50,43 @@ func copyModelTables(src, target *gorm.DB, batchSize int) (map[string]int64, int
 		if modelType.Kind() != reflect.Ptr {
 			return tableCounts, totalCopied, fmt.Errorf("model %T is not a pointer", m)
 		}
-		sliceType := reflect.SliceOf(modelType.Elem())
-		slicePtr := reflect.New(sliceType)
-		if err := src.Unscoped().Find(slicePtr.Interface()).Error; err != nil {
-			return tableCounts, totalCopied, fmt.Errorf("read sqlite table %s: %w", table, err)
-		}
-		filtered := slicePtr.Elem()
+		var primaryKeySet map[string]struct{}
 		if targetCount > 0 {
-			primaryKeySet, err := targetPrimaryKeySet(target, table, primaryColumns)
+			primaryKeySet, err = targetPrimaryKeySet(target, table, primaryColumns)
 			if err != nil {
 				return tableCounts, totalCopied, err
 			}
-			filtered = filterRowsMissingInTarget(target, table, primaryColumns, filtered, primaryKeySet)
 		}
-		if filtered.Len() == 0 {
-			continue
+		// 分页流式读取：此前整表一次性 Find 进内存，media 表几十万行、
+		// 每行含 overview/genres 等长文本时可达数百 MB，迁移过程有 OOM
+		// 风险。源库在迁移期间是静态的，offset 分页安全。
+		const readBatch = 1000
+		copiedForTable := int64(0)
+		for offset := 0; ; offset += readBatch {
+			batchPtr := reflect.New(reflect.SliceOf(modelType.Elem()))
+			if err := src.Unscoped().Limit(readBatch).Offset(offset).Find(batchPtr.Interface()).Error; err != nil {
+				return tableCounts, totalCopied, fmt.Errorf("read sqlite table %s: %w", table, err)
+			}
+			batch := batchPtr.Elem()
+			if batch.Len() == 0 {
+				break
+			}
+			filtered := batch
+			if primaryKeySet != nil {
+				filtered = filterRowsMissingInTarget(target, table, primaryColumns, batch, primaryKeySet)
+			}
+			if filtered.Len() > 0 {
+				filteredPtr := reflect.New(filtered.Type())
+				filteredPtr.Elem().Set(filtered)
+				if err := target.Clauses(clause.OnConflict{DoNothing: true}).CreateInBatches(filteredPtr.Interface(), batchSize).Error; err != nil {
+					return tableCounts, totalCopied, fmt.Errorf("copy sqlite table %s: %w", table, err)
+				}
+				copiedForTable += int64(filtered.Len())
+			}
+			if batch.Len() < readBatch {
+				break
+			}
 		}
-		filteredPtr := reflect.New(filtered.Type())
-		filteredPtr.Elem().Set(filtered)
-		if err := target.Clauses(clause.OnConflict{DoNothing: true}).CreateInBatches(filteredPtr.Interface(), batchSize).Error; err != nil {
-			return tableCounts, totalCopied, fmt.Errorf("copy sqlite table %s: %w", table, err)
-		}
-		copiedForTable := int64(filtered.Len())
 		tableCounts[table] = copiedForTable
 		totalCopied += copiedForTable
 	}
@@ -5,12 +5,21 @@ import (
 	"fmt"
 	"path/filepath"
 	"strings"
+	"sync"
+	"sync/atomic"
+	"time"

 	"gorm.io/gorm"

 	"github.com/truewhile/MeBox/internal/config"
 )

+// sqliteGateHoldLimit 是写闸持有者的最长合法持有时长。语句级写闸在 SQL
+// 执行 panic 时 After 回调不会运行，令牌会泄漏并让后续所有写入永久等锁；
+// 超过该时长的持有者按泄漏强制回收（60s 内单条写语句远未到，正常写路径
+// 不受影响）。
+const sqliteGateHoldLimit = 60 * time.Second
+
 func installSQLiteWriteGate(db *gorm.DB) {
 	if db == nil {
 		return
@@ -22,15 +31,18 @@ func installSQLiteWriteGate(db *gorm.DB) {
 		if tx.Statement != nil && tx.Statement.Context != nil {
 			ctx = tx.Statement.Context
 		}
-		if err := gate.Lock(ctx); err != nil {
+		holder, err := gate.Lock(ctx)
+		if err != nil {
 			_ = tx.AddError(err)
 			return
 		}
-		tx.InstanceSet(lockedKey, struct{}{})
+		tx.InstanceSet(lockedKey, holder)
 	}
 	unlock := func(tx *gorm.DB) {
-		if _, ok := tx.InstanceGet(lockedKey); ok {
-			gate.Unlock()
+		if holder, ok := tx.InstanceGet(lockedKey); ok {
+			if h, ok := holder.(*sqliteGateHolder); ok {
+				gate.Unlock(h)
+			}
 		}
 	}
 	rawLock := func(tx *gorm.DB) {
@@ -64,38 +76,76 @@ func isReadOnlySQL(sql string) bool {
 	return false
 }

-// sqliteWriteGate serializes in-process SQLite writes while respecting the
-// statement context, so request cancellation can break out of a queued write.
+// sqliteWriteGate serializes in-process SQLite writes. 所有权令牌（而非裸
+// 信号量）保证只有持有者本人能释放；持有超时按泄漏自动回收，避免一次
+// panic 让进程的 SQLite 写入半永久性瘫痪。
 type sqliteWriteGate struct {
-	ch chan struct{}
+	mu    sync.Mutex
+	cond  *sync.Cond
+	owner *sqliteGateHolder
 }

+type sqliteGateHolder struct {
+	id       uint64
+	acquired time.Time
+}
+
+var sqliteGateHolderSeq atomic.Uint64
+
 func newSQLiteWriteGate() *sqliteWriteGate {
-	return &sqliteWriteGate{ch: make(chan struct{}, 1)}
+	g := &sqliteWriteGate{}
+	g.cond = sync.NewCond(&g.mu)
+	return g
 }

-func (g *sqliteWriteGate) Lock(ctx context.Context) error {
-	select {
-	case g.ch <- struct{}{}:
-		return nil
-	default:
-	}
+func (g *sqliteWriteGate) Lock(ctx context.Context) (*sqliteGateHolder, error) {
+	g.mu.Lock()
+	defer g.mu.Unlock()
 	if ctx == nil {
 		ctx = context.Background()
 	}
-	select {
-	case g.ch <- struct{}{}:
-		return nil
-	case <-ctx.Done():
-		return ctx.Err()
+	// ctx 取消时唤醒等待者（cond 无法感知 ctx，用旁路 goroutine 广播）。
+	if done := ctx.Done(); done != nil {
+		stop := make(chan struct{})
+		defer close(stop)
+		go func() {
+			select {
+			case <-done:
+				g.cond.Broadcast()
+			case <-stop:
+			}
+		}()
+	}
+	for {
+		if g.owner == nil {
+			holder := &sqliteGateHolder{
+				id:       sqliteGateHolderSeq.Add(1),
+				acquired: time.Now(),
+			}
+			g.owner = holder
+			return holder, nil
+		}
+		if ctx.Err() != nil {
+			return nil, ctx.Err()
+		}
+		if time.Since(g.owner.acquired) > sqliteGateHoldLimit {
+			// 持有者疑似 panic 泄漏（After 回调未执行）：强制回收。
+			g.owner = nil
+			g.cond.Broadcast()
+			continue
+		}
+		g.cond.Wait()
 	}
 }

-func (g *sqliteWriteGate) Unlock() {
-	select {
-	case <-g.ch:
-	default:
+func (g *sqliteWriteGate) Unlock(h *sqliteGateHolder) {
+	g.mu.Lock()
+	defer g.mu.Unlock()
+	if h == nil || g.owner != h {
+		return
 	}
+	g.owner = nil
+	g.cond.Broadcast()
 }

 func buildSQLiteDSN(cfg *config.Config) string {
@@ -104,7 +154,9 @@ func buildSQLiteDSN(cfg *config.Config) string {
 		// keep as-is to respect user-provided relative paths.
 		dbPath = filepath.Clean(dbPath)
 	}
-	dsn := dbPath + "?_pragma=foreign_keys(1)"
+	// _txlock=immediate：事务以写锁开始。此前 deferred BEGIN 在并发事务
+	// 升级写锁时会绕过 busy_timeout 直接报 SQLITE_BUSY。
+	dsn := dbPath + "?_txlock=immediate&_pragma=foreign_keys(1)"
 	if cfg.Database.WALMode {
 		dsn += "&_pragma=journal_mode(WAL)&_pragma=synchronous(NORMAL)"
 	}