fix: switch to SHUTTING_DOWN state unconditionally

During the shutdown sequence always switch to SHUTTING_DOWN.
Make sure that the rest of the code does not break if it can not switch to the desired
global state + some clean ups around state transitions.

Signed-off-by: Roman Gershman <[email protected]>

src/server/main_service.cc

@@ -884,7 +884,11 @@ void Service::Shutdown() {
   VLOG(1) << "Service::Shutdown";
 
   // We mark that we are shutting down. After this incoming requests will be
-  // rejected
+  // rejected.
+  mu_.lock();
+  global_state_ = GlobalState::SHUTTING_DOWN;
+  mu_.unlock();
+
   pp_.AwaitFiberOnAll([](ProactorBase* pb) {
     ServerState::tlocal()->EnterLameDuck();
     facade::Connection::ShutdownThreadLocal();
@@ -2503,27 +2507,20 @@ GlobalState Service::SwitchState(GlobalState from, GlobalState to) {
   return to;
 }
 
-void Service::RequestLoadingState() {
-  bool switch_state = false;
-  {
+bool Service::RequestLoadingState() {
+  if (SwitchState(GlobalState::ACTIVE, GlobalState::LOADING) == GlobalState::LOADING) {
     util::fb2::LockGuard lk(mu_);
-    ++loading_state_counter_;
-    if (global_state_ != GlobalState::LOADING) {
-      DCHECK_EQ(global_state_, GlobalState::ACTIVE);
-      switch_state = true;
-    }
-  }
-  if (switch_state) {
-    SwitchState(GlobalState::ACTIVE, GlobalState::LOADING);
+    loading_state_counter_++;
+    return true;
   }
+  return false;
 }
 
 void Service::RemoveLoadingState() {
   bool switch_state = false;
   {
     util::fb2::LockGuard lk(mu_);
-    DCHECK_EQ(global_state_, GlobalState::LOADING);
-    DCHECK_GT(loading_state_counter_, 0u);
+    CHECK_GT(loading_state_counter_, 0u);
     --loading_state_counter_;
     switch_state = loading_state_counter_ == 0;
   }
@@ -2532,11 +2529,6 @@ void Service::RemoveLoadingState() {
   }
 }
 
-GlobalState Service::GetGlobalState() const {
-  util::fb2::LockGuard lk(mu_);
-  return global_state_;
-}
-
 void Service::ConfigureHttpHandlers(util::HttpListenerBase* base, bool is_privileged) {
   // We skip authentication on privileged listener if the flag admin_nopass is set
   // We also skip authentication if requirepass is empty

src/server/main_service.h

@@ -78,11 +78,9 @@ class Service : public facade::ServiceInterface {
   // Upon switch, updates cached global state in threadlocal ServerState struct.
   GlobalState SwitchState(GlobalState from, GlobalState to) ABSL_LOCKS_EXCLUDED(mu_);
 
-  void RequestLoadingState() ABSL_LOCKS_EXCLUDED(mu_);
+  bool RequestLoadingState() ABSL_LOCKS_EXCLUDED(mu_);
   void RemoveLoadingState() ABSL_LOCKS_EXCLUDED(mu_);
 
-  GlobalState GetGlobalState() const ABSL_LOCKS_EXCLUDED(mu_);
-
   void ConfigureHttpHandlers(util::HttpListenerBase* base, bool is_privileged) final;
   void OnConnectionClose(facade::ConnectionContext* cntx) final;
 

src/server/replica.cc

@@ -417,7 +417,11 @@ error_code Replica::InitiatePSync() {
     io::PrefixSource ps{io_buf.InputBuffer(), Sock()};
 
     // Set LOADING state.
-    service_.RequestLoadingState();
+    if (!service_.RequestLoadingState()) {
+      return cntx_.ReportError(std::make_error_code(errc::state_not_recoverable),
+                               "Failed to enter LOADING state");
+    }
+
     absl::Cleanup cleanup = [this]() { service_.RemoveLoadingState(); };
 
     if (slot_range_.has_value()) {
@@ -502,10 +506,14 @@ error_code Replica::InitiateDflySync() {
     for (auto& flow : shard_flows_)
       flow->Cancel();
   };
+
   RETURN_ON_ERR(cntx_.SwitchErrorHandler(std::move(err_handler)));
 
   // Make sure we're in LOADING state.
-  service_.RequestLoadingState();
+  if (!service_.RequestLoadingState()) {
+    return cntx_.ReportError(std::make_error_code(errc::state_not_recoverable),
+                             "Failed to enter LOADING state");
+  }
 
   // Start full sync flows.
   state_mask_.fetch_or(R_SYNCING);

src/server/server_family.cc

@@ -1168,10 +1168,12 @@ void ServerFamily::SnapshotScheduling() {
     return;
   }
 
-  const auto loading_check_interval = std::chrono::seconds(10);
-  while (service_.GetGlobalState() == GlobalState::LOADING) {
-    schedule_done_.WaitFor(loading_check_interval);
-  }
+  ServerState* ss = ServerState::tlocal();
+  do {
+    if (schedule_done_.WaitFor(100ms)) {
+      return;
+    }
+  } while (ss->gstate() == GlobalState::LOADING);
 
   while (true) {
     const std::chrono::time_point now = std::chrono::system_clock::now();
@@ -1657,9 +1659,10 @@ GenericError ServerFamily::DoSave(bool ignore_state) {
 
 GenericError ServerFamily::DoSaveCheckAndStart(bool new_version, string_view basename,
                                                Transaction* trans, bool ignore_state) {
-  auto state = service_.GetGlobalState();
+  auto state = ServerState::tlocal()->gstate();
+
   // In some cases we want to create a snapshot even if server is not active, f.e in takeover
-  if (!ignore_state && (state != GlobalState::ACTIVE)) {
+  if (!ignore_state && (state != GlobalState::ACTIVE && state != GlobalState::SHUTTING_DOWN)) {
     return GenericError{make_error_code(errc::operation_in_progress),
                         StrCat(GlobalStateName(state), " - can not save database")};
   }
@@ -2242,6 +2245,8 @@ void ServerFamily::Info(CmdArgList args, const CommandContext& cmd_cntx) {
     absl::StrAppend(&info, a1, ":", a2, "\r\n");
   };
 
+  ServerState* ss = ServerState::tlocal();
+
   if (should_enter("SERVER")) {
     auto kind = ProactorBase::me()->GetKind();
     const char* multiplex_api = (kind == ProactorBase::IOURING) ? "iouring" : "epoll";
@@ -2467,7 +2472,7 @@ void ServerFamily::Info(CmdArgList args, const CommandContext& cmd_cntx) {
     append("last_saved_file", save_info.file_name);
     append("last_success_save_duration_sec", save_info.success_duration_sec);
 
-    size_t is_loading = service_.GetGlobalState() == GlobalState::LOADING;
+    unsigned is_loading = (ss->gstate() == GlobalState::LOADING);
     append("loading", is_loading);
     append("saving", is_saving);
     append("current_save_duration_sec", curent_durration_sec);
@@ -2752,7 +2757,8 @@ void ServerFamily::ReplicaOfInternal(CmdArgList args, Transaction* tx, SinkReply
     util::fb2::LockGuard lk(replicaof_mu_);  // Only one REPLICAOF command can run at a time
 
     // We should not execute replica of command while loading from snapshot.
-    if (ServerState::tlocal()->is_master && service_.GetGlobalState() == GlobalState::LOADING) {
+    ServerState* ss = ServerState::tlocal();
+    if (ss->is_master && ss->gstate() == GlobalState::LOADING) {
       builder->SendError(kLoadingErr);
       return;
     }
@@ -2766,7 +2772,7 @@ void ServerFamily::ReplicaOfInternal(CmdArgList args, Transaction* tx, SinkReply
 
     // If NO ONE was supplied, just stop the current replica (if it exists)
     if (replicaof_args->IsReplicaOfNoOne()) {
-      if (!ServerState::tlocal()->is_master) {
+      if (!ss->is_master) {
         CHECK(replica_);
 
         SetMasterFlagOnAllThreads(true);  // Flip flag before clearing replica
@@ -2776,8 +2782,9 @@ void ServerFamily::ReplicaOfInternal(CmdArgList args, Transaction* tx, SinkReply
         StopAllClusterReplicas();
       }
 
-      CHECK_EQ(service_.SwitchState(GlobalState::LOADING, GlobalState::ACTIVE), GlobalState::ACTIVE)
-          << "Server is set to replica no one, yet state is not active!";
+      // May not switch if we are at shutdown state, and the process is shutting down
+      // at the same time.
+      service_.SwitchState(GlobalState::LOADING, GlobalState::ACTIVE);
 
       return builder->SendOk();
     }

tests/dragonfly/replication_test.py

@@ -1976,18 +1976,22 @@ async def test_replicaof_reject_on_load(df_factory, df_seeder_factory):
     replica.stop()
     replica.start()
     c_replica = replica.client()
+
+    @assert_eventually
+    async def check_replica_isloading():
+        persistence = await c_replica.info("PERSISTENCE")
+        assert persistence["loading"] == 1
+
+    # If this fails adjust `keys` and the `assert dbsize >= 30000` above.
+    # Keep in mind that if the assert False is triggered below, it doesn't mean
+    # that there is a bug because it could be the case that while executing
+    # INFO PERSISTENCE df is in loading state but when we call REPLICAOF df
+    # is no longer in loading state and the assertion false is triggered.
+    await check_replica_isloading()
+
     # Check replica of not alowed while loading snapshot
-    try:
-        # If this fails adjust `keys` and the `assert dbsize >= 30000` above.
-        # Keep in mind that if the assert False is triggered below, it doesn't mean
-        # that there is a bug because it could be the case that while executing
-        # INFO PERSISTENCE df is in loading state but when we call REPLICAOF df
-        # is no longer in loading state and the assertion false is triggered.
-        assert "loading:1" in (await c_replica.execute_command("INFO PERSISTENCE"))
+    with pytest.raises(aioredis.BusyLoadingError):
         await c_replica.execute_command(f"REPLICAOF localhost {master.port}")
-        assert False
-    except aioredis.BusyLoadingError as e:
-        assert "Dragonfly is loading the dataset in memory" in str(e)
 
     # Check one we finish loading snapshot replicaof success
     await wait_available_async(c_replica, timeout=180)