fix unexpected slow query during GC running after stop 1 tikv-server (tikv#899)

Signed-off-by: crazycs520 <crazycs520@gmail.com>

Author: crazycs520

internal/locate/region_request.go

@@ -402,7 +402,7 @@ func (state *tryFollower) next(bo *retry.Backoffer, selector *replicaSelector) (
 		}
 		targetReplica = selector.replicas[idx]
 		// Each follower is only tried once
-		if !targetReplica.isExhausted(1) {
+		if !targetReplica.isExhausted(1) && targetReplica.store.getLivenessState() != unreachable {
 			state.lastIdx = idx
 			selector.targetIdx = idx
 			break
@@ -628,8 +628,8 @@ func (state *accessFollower) onSendFailure(bo *retry.Backoffer, selector *replic
 }
 
 func (state *accessFollower) isCandidate(idx AccessIndex, replica *replica) bool {
-	// the epoch is staled or retry exhausted.
-	if replica.isEpochStale() || replica.isExhausted(1) {
+	// the epoch is staled or retry exhausted, or the store is unreachable.
+	if replica.isEpochStale() || replica.isExhausted(1) || replica.store.getLivenessState() == unreachable {
 		return false
 	}
 	// The request can only be sent to the leader.

internal/locate/region_request3_test.go

@@ -278,6 +278,12 @@ func refreshEpochs(regionStore *regionStore) {
 	}
 }
 
+func refreshLivenessStates(regionStore *regionStore) {
+	for _, store := range regionStore.stores {
+		atomic.StoreUint32(&store.livenessState, uint32(reachable))
+	}
+}
+
 func AssertRPCCtxEqual(s *testRegionRequestToThreeStoresSuite, rpcCtx *RPCContext, target *replica, proxy *replica) {
 	s.Equal(rpcCtx.Store, target.store)
 	s.Equal(rpcCtx.Peer, target.peer)
@@ -567,6 +573,7 @@ func (s *testRegionRequestToThreeStoresSuite) TestReplicaSelector() {
 	// Test accessFollower state with kv.ReplicaReadFollower request type.
 	req = tikvrpc.NewReplicaReadRequest(tikvrpc.CmdGet, &kvrpcpb.GetRequest{}, kv.ReplicaReadFollower, nil)
 	refreshEpochs(regionStore)
+	refreshLivenessStates(regionStore)
 	replicaSelector, err = newReplicaSelector(cache, regionLoc.Region, req)
 	s.Nil(err)
 	s.NotNil(replicaSelector)
@@ -680,10 +687,13 @@ func (s *testRegionRequestToThreeStoresSuite) TestSendReqWithReplicaSelector() {
 	region, err := s.cache.LocateRegionByID(s.bo, s.regionID)
 	s.Nil(err)
 	s.NotNil(region)
+	regionStore := s.cache.GetCachedRegionWithRLock(region.Region).getStore()
+	s.NotNil(regionStore)
 
 	reloadRegion := func() {
 		s.regionRequestSender.replicaSelector.region.invalidate(Other)
 		region, _ = s.cache.LocateRegionByID(s.bo, s.regionID)
+		regionStore = s.cache.GetCachedRegionWithRLock(region.Region).getStore()
 	}
 
 	hasFakeRegionError := func(resp *tikvrpc.Response) bool {
@@ -715,6 +725,7 @@ func (s *testRegionRequestToThreeStoresSuite) TestSendReqWithReplicaSelector() {
 	s.Equal(sender.replicaSelector.targetIdx, AccessIndex(1))
 	s.True(bo.GetTotalBackoffTimes() == 1)
 	s.cluster.StartStore(s.storeIDs[0])
+	atomic.StoreUint32(&regionStore.stores[0].livenessState, uint32(reachable))
 
 	// Leader is updated because of send success, so no backoff.
 	bo = retry.NewBackoffer(context.Background(), -1)
@@ -734,6 +745,7 @@ func (s *testRegionRequestToThreeStoresSuite) TestSendReqWithReplicaSelector() {
 	s.True(hasFakeRegionError(resp))
 	s.Equal(bo.GetTotalBackoffTimes(), 1)
 	s.cluster.StartStore(s.storeIDs[1])
+	atomic.StoreUint32(&regionStore.stores[1].livenessState, uint32(reachable))
 
 	// Leader is changed. No backoff.
 	reloadRegion()
@@ -750,7 +762,7 @@ func (s *testRegionRequestToThreeStoresSuite) TestSendReqWithReplicaSelector() {
 	resp, _, err = sender.SendReq(bo, req, region.Region, time.Second)
 	s.Nil(err)
 	s.True(hasFakeRegionError(resp))
-	s.Equal(bo.GetTotalBackoffTimes(), 2) // The unreachable leader is skipped
+	s.Equal(bo.GetTotalBackoffTimes(), 3)
 	s.False(sender.replicaSelector.region.isValid())
 	s.cluster.ChangeLeader(s.regionID, s.peerIDs[0])
 
@@ -1098,3 +1110,81 @@ func (s *testRegionRequestToThreeStoresSuite) TestReplicaReadFallbackToLeaderReg
 	// after region error returned, the region should be invalidated.
 	s.False(region.isValid())
 }
+
+func (s *testRegionRequestToThreeStoresSuite) TestAccessFollowerAfter1TiKVDown() {
+	var leaderAddr string
+	s.regionRequestSender.client = &fnClient{fn: func(ctx context.Context, addr string, req *tikvrpc.Request, timeout time.Duration) (response *tikvrpc.Response, err error) {
+		// Returns error when accesses non-leader.
+		if leaderAddr != addr {
+			return nil, context.DeadlineExceeded
+		}
+		return &tikvrpc.Response{Resp: &kvrpcpb.GetResponse{
+			Value: []byte("value"),
+		}}, nil
+	}}
+
+	req := tikvrpc.NewRequest(tikvrpc.CmdGet, &kvrpcpb.GetRequest{
+		Key: []byte("key"),
+	})
+	req.ReplicaReadType = kv.ReplicaReadMixed
+
+	loc, err := s.cache.LocateKey(s.bo, []byte("key"))
+	s.Nil(err)
+	region := s.cache.GetCachedRegionWithRLock(loc.Region)
+	s.NotNil(region)
+	regionStore := region.getStore()
+	leaderAddr = regionStore.stores[regionStore.workTiKVIdx].addr
+	s.NotEqual(leaderAddr, "")
+	for i := 0; i < 10; i++ {
+		bo := retry.NewBackofferWithVars(context.Background(), 100, nil)
+		resp, _, _, err := s.regionRequestSender.SendReqCtx(bo, req, loc.Region, time.Second, tikvrpc.TiKV)
+		s.Nil(err)
+		s.NotNil(resp)
+
+		// Since send req to follower will receive error, then all follower will be marked as unreachable and epoch stale.
+		allFollowerStoreEpochStale := true
+		for i, store := range regionStore.stores {
+			if i == int(regionStore.workTiKVIdx) {
+				continue
+			}
+			if store.epoch == regionStore.storeEpochs[i] {
+				allFollowerStoreEpochStale = false
+				break
+			} else {
+				s.Equal(store.getLivenessState(), unreachable)
+			}
+		}
+		if allFollowerStoreEpochStale {
+			break
+		}
+	}
+
+	// mock for GC leader reload all regions.
+	bo := retry.NewBackofferWithVars(context.Background(), 10, nil)
+	_, err = s.cache.BatchLoadRegionsWithKeyRange(bo, []byte(""), nil, 1)
+	s.Nil(err)
+
+	loc, err = s.cache.LocateKey(s.bo, []byte("key"))
+	s.Nil(err)
+	region = s.cache.GetCachedRegionWithRLock(loc.Region)
+	s.NotNil(region)
+	regionStore = region.getStore()
+	for i, store := range regionStore.stores {
+		if i == int(regionStore.workTiKVIdx) {
+			continue
+		}
+		// After reload region, the region epoch will be updated, but the store liveness state is still unreachable.
+		s.Equal(store.epoch, regionStore.storeEpochs[i])
+		s.Equal(store.getLivenessState(), unreachable)
+	}
+
+	for i := 0; i < 100; i++ {
+		bo := retry.NewBackofferWithVars(context.Background(), 1, nil)
+		resp, _, retryTimes, err := s.regionRequestSender.SendReqCtx(bo, req, loc.Region, time.Second, tikvrpc.TiKV)
+		s.Nil(err)
+		s.NotNil(resp)
+		// since all follower'store is unreachable, the request will be sent to leader, the backoff times should be 0.
+		s.Equal(0, bo.GetTotalBackoffTimes())
+		s.Equal(0, retryTimes)
+	}
+}