From aaca081cec39c6a120085c840acaf32c38885ab2 Mon Sep 17 00:00:00 2001
From: tangenta <tangenta@126.com>
Date: Wed, 31 Jul 2024 19:07:28 +0800
Subject: [PATCH] ddl: record get owner TS and compare it before runReorgJob
 quit (#55049)

close pingcap/tidb#54897
---
 pkg/ddl/ddl.go           | 15 ++++++++++++++-
 pkg/ddl/job_scheduler.go |  1 +
 pkg/ddl/reorg.go         | 24 +++++++++++++++++++++---
 3 files changed, 36 insertions(+), 4 deletions(-)

diff --git a/pkg/ddl/ddl.go b/pkg/ddl/ddl.go
index 48a021793cc73..236dc403a2ec6 100644
--- a/pkg/ddl/ddl.go
+++ b/pkg/ddl/ddl.go
@@ -519,6 +519,19 @@ type reorgContexts struct {
 	sync.RWMutex
 	// reorgCtxMap maps job ID to reorg context.
 	reorgCtxMap map[int64]*reorgCtx
+	beOwnerTS   int64
+}
+
+func (r *reorgContexts) getOwnerTS() int64 {
+	r.RLock()
+	defer r.RUnlock()
+	return r.beOwnerTS
+}
+
+func (r *reorgContexts) setOwnerTS(ts int64) {
+	r.Lock()
+	r.beOwnerTS = ts
+	r.Unlock()
 }
 
 func (dc *ddlCtx) getReorgCtx(jobID int64) *reorgCtx {
@@ -536,7 +549,7 @@ func (dc *ddlCtx) newReorgCtx(jobID int64, rowCount int64) *reorgCtx {
 		return existedRC
 	}
 	rc := &reorgCtx{}
-	rc.doneCh = make(chan error, 1)
+	rc.doneCh = make(chan reorgFnResult, 1)
 	// initial reorgCtx
 	rc.setRowCount(rowCount)
 	rc.mu.warnings = make(map[errors.ErrorID]*terror.Error)
diff --git a/pkg/ddl/job_scheduler.go b/pkg/ddl/job_scheduler.go
index 683af6bff20ef..b97c9de5cb3b8 100644
--- a/pkg/ddl/job_scheduler.go
+++ b/pkg/ddl/job_scheduler.go
@@ -112,6 +112,7 @@ func (l *ownerListener) OnBecomeOwner() {
 		sessPool:       l.ddl.sessPool,
 		delRangeMgr:    l.ddl.delRangeMgr,
 	}
+	l.ddl.reorgCtx.setOwnerTS(time.Now().Unix())
 	l.scheduler.start()
 }
 
diff --git a/pkg/ddl/reorg.go b/pkg/ddl/reorg.go
index c5295281ee686..2154974826a5a 100644
--- a/pkg/ddl/reorg.go
+++ b/pkg/ddl/reorg.go
@@ -65,7 +65,7 @@ type reorgCtx struct {
 	// If the reorganization job is done, we will use this channel to notify outer.
 	// TODO: Now we use goroutine to simulate reorganization jobs, later we may
 	// use a persistent job list.
-	doneCh chan error
+	doneCh chan reorgFnResult
 	// rowCount is used to simulate a job's row count.
 	rowCount int64
 	jobState model.JobState
@@ -80,6 +80,13 @@ type reorgCtx struct {
 	references atomicutil.Int32
 }
 
+// reorgFnResult records the DDL owner TS before executing reorg function, in order to help
+// receiver determine if the result is from reorg function of previous DDL owner in this instance.
+type reorgFnResult struct {
+	ownerTS int64
+	err     error
+}
+
 func newReorgExprCtx() exprctx.ExprContext {
 	evalCtx := contextstatic.NewStaticEvalContext(
 		contextstatic.WithSQLMode(mysql.ModeNone),
@@ -251,11 +258,13 @@ func (w *worker) runReorgJob(
 			return dbterror.ErrCancelledDDLJob
 		}
 
+		beOwnerTS := w.ddlCtx.reorgCtx.getOwnerTS()
 		rc = w.newReorgCtx(reorgInfo.Job.ID, reorgInfo.Job.GetRowCount())
 		w.wg.Add(1)
 		go func() {
 			defer w.wg.Done()
-			rc.doneCh <- reorgFn()
+			err := reorgFn()
+			rc.doneCh <- reorgFnResult{ownerTS: beOwnerTS, err: err}
 		}()
 	}
 
@@ -271,7 +280,16 @@ func (w *worker) runReorgJob(
 
 	// wait reorganization job done or timeout
 	select {
-	case err := <-rc.doneCh:
+	case res := <-rc.doneCh:
+		err := res.err
+		curTS := w.ddlCtx.reorgCtx.getOwnerTS()
+		if res.ownerTS != curTS {
+			d.removeReorgCtx(job.ID)
+			logutil.DDLLogger().Warn("owner ts mismatch, return timeout error and retry",
+				zap.Int64("prevTS", res.ownerTS),
+				zap.Int64("curTS", curTS))
+			return dbterror.ErrWaitReorgTimeout
+		}
 		// Since job is cancelled，we don't care about its partial counts.
 		if rc.isReorgCanceled() || terror.ErrorEqual(err, dbterror.ErrCancelledDDLJob) {
 			d.removeReorgCtx(job.ID)