raftstore: block in-memory pessimistic locks during the flashback (ti…

…kv#14859)

ref tikv#13303, close pingcap/tidb#44292

During the Flashback process, we should prevent any read or write operations on the in-memory pessimistic lock table
and clear it like rolling back other locks to ensure that Flashback can proceed smoothly.

Signed-off-by: JmPotato <ghzpotato@gmail.com>
Signed-off-by: tonyxuqqi <tonyxuqi@outlook.com>

components/raftstore-v2/src/operation/command/admin/flashback.rs

@@ -9,7 +9,10 @@ use kvproto::{
 use protobuf::Message;
 use raftstore::{
     coprocessor::RegionChangeReason,
-    store::metrics::{PEER_ADMIN_CMD_COUNTER, PEER_IN_FLASHBACK_STATE},
+    store::{
+        metrics::{PEER_ADMIN_CMD_COUNTER, PEER_IN_FLASHBACK_STATE},
+        LocksStatus,
+    },
     Result,
 };
 
@@ -85,7 +88,7 @@ impl<EK: KvEngine, ER: RaftEngine> Peer<EK, ER> {
     pub fn on_apply_res_flashback<T>(
         &mut self,
         store_ctx: &mut StoreContext<EK, ER, T>,
-        mut res: FlashbackResult,
+        #[allow(unused_mut)] mut res: FlashbackResult,
     ) {
         (|| {
             fail_point!("keep_peer_fsm_flashback_state_false", |_| {
@@ -114,6 +117,22 @@ impl<EK: KvEngine, ER: RaftEngine> Peer<EK, ER> {
             .unwrap();
         self.set_has_extra_write();
 
+        let mut pessimistic_locks = self.txn_context().ext().pessimistic_locks.write();
+        pessimistic_locks.status = if res.region_state.get_region().is_in_flashback {
+            // To prevent the insertion of any new pessimistic locks, set the lock status
+            // to `LocksStatus::IsInFlashback` and clear all the existing locks.
+            pessimistic_locks.clear();
+            LocksStatus::IsInFlashback
+        } else if self.is_leader() {
+            // If the region is not in flashback, the leader can continue to insert
+            // pessimistic locks.
+            LocksStatus::Normal
+        } else {
+            // If the region is not in flashback and the peer is not the leader, it
+            // cannot insert pessimistic locks.
+            LocksStatus::NotLeader
+        }
+
         // Compares to v1, v2 does not expire remote lease, because only
         // local reader can serve read requests.
     }

components/raftstore/src/store/fsm/peer.rs

@@ -6505,6 +6505,21 @@ where
         })());
         // Let the leader lease to None to ensure that local reads are not executed.
         self.fsm.peer.leader_lease_mut().expire_remote_lease();
+        let mut pessimistic_locks = self.fsm.peer.txn_ext.pessimistic_locks.write();
+        pessimistic_locks.status = if self.region().is_in_flashback {
+            // To prevent the insertion of any new pessimistic locks, set the lock status
+            // to `LocksStatus::IsInFlashback` and clear all the existing locks.
+            pessimistic_locks.clear();
+            LocksStatus::IsInFlashback
+        } else if self.fsm.peer.is_leader() {
+            // If the region is not in flashback, the leader can continue to insert
+            // pessimistic locks.
+            LocksStatus::Normal
+        } else {
+            // If the region is not in flashback and the peer is not the leader, it
+            // cannot insert pessimistic locks.
+            LocksStatus::NotLeader
+        }
     }
 
     fn on_ready_batch_switch_witness(&mut self, sw: SwitchWitness) {

components/raftstore/src/store/txn_ext.rs

@@ -125,6 +125,7 @@ pub enum LocksStatus {
     TransferringLeader,
     MergingRegion,
     NotLeader,
+    IsInFlashback,
 }
 
 impl fmt::Debug for PeerPessimisticLocks {

components/test_raftstore-v2/src/server.rs

@@ -885,6 +885,14 @@ impl<EK: KvEngine> Simulator<EK> for ServerCluster<EK> {
 
 impl<EK: KvEngine> Cluster<ServerCluster<EK>, EK> {
     pub fn must_get_snapshot_of_region(&mut self, region_id: u64) -> RegionSnapshot<EK::Snapshot> {
+        self.must_get_snapshot_of_region_with_ctx(region_id, SnapContext::default())
+    }
+
+    pub fn must_get_snapshot_of_region_with_ctx(
+        &mut self,
+        region_id: u64,
+        snap_ctx: SnapContext<'_>,
+    ) -> RegionSnapshot<EK::Snapshot> {
         let mut try_snapshot = || -> Option<RegionSnapshot<EK::Snapshot>> {
             let leader = self.leader_of_region(region_id)?;
             let store_id = leader.store_id;
@@ -897,7 +905,7 @@ impl<EK: KvEngine> Cluster<ServerCluster<EK>, EK> {
             let mut storage = self.sim.rl().storages.get(&store_id).unwrap().clone();
             let snap_ctx = SnapContext {
                 pb_ctx: &ctx,
-                ..Default::default()
+                ..snap_ctx.clone()
             };
             storage.snapshot(snap_ctx).ok()
         };

components/test_raftstore/src/server.rs

@@ -788,6 +788,14 @@ impl Simulator for ServerCluster {
 
 impl Cluster<ServerCluster> {
     pub fn must_get_snapshot_of_region(&mut self, region_id: u64) -> RegionSnapshot<RocksSnapshot> {
+        self.must_get_snapshot_of_region_with_ctx(region_id, Default::default())
+    }
+
+    pub fn must_get_snapshot_of_region_with_ctx(
+        &mut self,
+        region_id: u64,
+        snap_ctx: SnapContext<'_>,
+    ) -> RegionSnapshot<RocksSnapshot> {
         let mut try_snapshot = || -> Option<RegionSnapshot<RocksSnapshot>> {
             let leader = self.leader_of_region(region_id)?;
             let store_id = leader.store_id;
@@ -800,7 +808,7 @@ impl Cluster<ServerCluster> {
             let mut storage = self.sim.rl().storages.get(&store_id).unwrap().clone();
             let snap_ctx = SnapContext {
                 pb_ctx: &ctx,
-                ..Default::default()
+                ..snap_ctx.clone()
             };
             storage.snapshot(snap_ctx).ok()
         };

src/storage/mvcc/reader/reader.rs

@@ -5,9 +5,10 @@ use std::ops::Bound;
 
 use engine_traits::{CF_DEFAULT, CF_LOCK, CF_WRITE};
 use kvproto::{
-    errorpb::{self, EpochNotMatch, StaleCommand},
+    errorpb::{self, EpochNotMatch, FlashbackInProgress, StaleCommand},
     kvrpcpb::Context,
 };
+use raftstore::store::LocksStatus;
 use tikv_kv::{SnapshotExt, SEEK_BOUND};
 use txn_types::{Key, Lock, OldValue, TimeStamp, Value, Write, WriteRef, WriteType};
 
@@ -146,6 +147,8 @@ pub struct MvccReader<S: EngineSnapshot> {
     term: u64,
     #[allow(dead_code)]
     version: u64,
+
+    allow_in_flashback: bool,
 }
 
 impl<S: EngineSnapshot> MvccReader<S> {
@@ -164,6 +167,7 @@ impl<S: EngineSnapshot> MvccReader<S> {
             fill_cache,
             term: 0,
             version: 0,
+            allow_in_flashback: false,
         }
     }
 
@@ -182,6 +186,7 @@ impl<S: EngineSnapshot> MvccReader<S> {
             fill_cache: !ctx.get_not_fill_cache(),
             term: ctx.get_term(),
             version: ctx.get_region_epoch().get_version(),
+            allow_in_flashback: false,
         }
     }
 
@@ -267,6 +272,13 @@ impl<S: EngineSnapshot> MvccReader<S> {
                     err.set_epoch_not_match(EpochNotMatch::default());
                     return Some(Err(KvError::from(err).into()));
                 }
+                // If the region is in the flashback state, it should not be allowed to read the
+                // locks.
+                if locks.status == LocksStatus::IsInFlashback && !self.allow_in_flashback {
+                    let mut err = errorpb::Error::default();
+                    err.set_flashback_in_progress(FlashbackInProgress::default());
+                    return Some(Err(KvError::from(err).into()));
+                }
 
                 locks.get(key).map(|(lock, _)| {
                     // For write commands that are executed in serial, it should be impossible
@@ -768,6 +780,10 @@ impl<S: EngineSnapshot> MvccReader<S> {
     pub fn snapshot(&self) -> &S {
         &self.snapshot
     }
+
+    pub fn set_allow_in_flashback(&mut self, set_allow_in_flashback: bool) {
+        self.allow_in_flashback = set_allow_in_flashback;
+    }
 }
 
 #[cfg(test)]

src/storage/txn/commands/flashback_to_version.rs

@@ -95,6 +95,7 @@ impl<S: Snapshot, L: LockManager> WriteCommand<S, L> for FlashbackToVersion {
     fn process_write(mut self, snapshot: S, context: WriteContext<'_, L>) -> Result<WriteResult> {
         let mut reader =
             MvccReader::new_with_ctx(snapshot.clone(), Some(ScanMode::Forward), &self.ctx);
+        reader.set_allow_in_flashback(true);
         let mut txn = MvccTxn::new(TimeStamp::zero(), context.concurrency_manager);
         match self.state {
             FlashbackToVersionState::RollbackLock {

src/storage/txn/commands/flashback_to_version_read_phase.rs

@@ -153,6 +153,7 @@ impl<S: Snapshot> ReadCommand<S> for FlashbackToVersionReadPhase {
     fn process_read(self, snapshot: S, statistics: &mut Statistics) -> Result<ProcessResult> {
         let tag = self.tag().get_str();
         let mut reader = MvccReader::new_with_ctx(snapshot, Some(ScanMode::Forward), &self.ctx);
+        reader.set_allow_in_flashback(true);
         // Filter out the SST that does not have a newer version than `self.version` in
         // `CF_WRITE`, i.e, whose latest `commit_ts` <= `self.version` in the later
         // scan. By doing this, we can only flashback those keys that have version

tests/integrations/raftstore/test_flashback.rs

@@ -13,14 +13,76 @@ use kvproto::{
     raft_cmdpb::{AdminCmdType, CmdType, RaftCmdRequest, RaftCmdResponse, Request},
     raft_serverpb::RegionLocalState,
 };
-use raftstore::store::Callback;
+use raftstore::store::{Callback, LocksStatus};
 use test_raftstore::*;
 use test_raftstore_macro::test_case;
-use txn_types::WriteBatchFlags;
+use tikv::storage::kv::SnapContext;
+use txn_types::{Key, PessimisticLock, WriteBatchFlags};
 
 const TEST_KEY: &[u8] = b"k1";
 const TEST_VALUE: &[u8] = b"v1";
 
+#[test_case(test_raftstore::new_server_cluster)]
+#[test_case(test_raftstore_v2::new_server_cluster)]
+fn test_flashback_with_in_memory_pessimistic_locks() {
+    let mut cluster = new_cluster(0, 3);
+    cluster.cfg.raft_store.raft_heartbeat_ticks = 20;
+    cluster.run();
+    cluster.must_transfer_leader(1, new_peer(1, 1));
+
+    let region = cluster.get_region(TEST_KEY);
+    // Write a pessimistic lock to the in-memory pessimistic lock table.
+    {
+        let snapshot = cluster.must_get_snapshot_of_region(region.get_id());
+        let txn_ext = snapshot.txn_ext.unwrap();
+        let mut pessimistic_locks = txn_ext.pessimistic_locks.write();
+        assert!(pessimistic_locks.is_writable());
+        pessimistic_locks
+            .insert(vec![(
+                Key::from_raw(TEST_KEY),
+                PessimisticLock {
+                    primary: TEST_KEY.to_vec().into_boxed_slice(),
+                    start_ts: 10.into(),
+                    ttl: 3000,
+                    for_update_ts: 20.into(),
+                    min_commit_ts: 30.into(),
+                    last_change_ts: 5.into(),
+                    versions_to_last_change: 3,
+                    is_locked_with_conflict: false,
+                },
+            )])
+            .unwrap();
+        assert_eq!(pessimistic_locks.len(), 1);
+    }
+    // Prepare flashback.
+    cluster.must_send_wait_flashback_msg(region.get_id(), AdminCmdType::PrepareFlashback);
+    // Check the in-memory pessimistic lock table.
+    {
+        let snapshot = cluster.must_get_snapshot_of_region_with_ctx(
+            region.get_id(),
+            SnapContext {
+                allowed_in_flashback: true,
+                ..Default::default()
+            },
+        );
+        let txn_ext = snapshot.txn_ext.unwrap();
+        let pessimistic_locks = txn_ext.pessimistic_locks.read();
+        assert!(!pessimistic_locks.is_writable());
+        assert_eq!(pessimistic_locks.status, LocksStatus::IsInFlashback);
+        assert_eq!(pessimistic_locks.len(), 0);
+    }
+    // Finish flashback.
+    cluster.must_send_wait_flashback_msg(region.get_id(), AdminCmdType::FinishFlashback);
+    // Check the in-memory pessimistic lock table.
+    {
+        let snapshot = cluster.must_get_snapshot_of_region(region.get_id());
+        let txn_ext = snapshot.txn_ext.unwrap();
+        let pessimistic_locks = txn_ext.pessimistic_locks.read();
+        assert!(pessimistic_locks.is_writable());
+        assert_eq!(pessimistic_locks.len(), 0);
+    }
+}
+
 #[test_case(test_raftstore::new_node_cluster)]
 #[test_case(test_raftstore_v2::new_node_cluster)]
 fn test_allow_read_only_request() {