@@ -170,9 +170,18 @@ std::optional<Replica::LastMasterSyncData> Replica::Stop() {
170170 sync_fb_.JoinIfNeeded ();
171171 DVLOG (1 ) << " MainReplicationFb stopped " << this ;
172172 acks_fb_.JoinIfNeeded ();
173- for (auto & flow : shard_flows_) {
174- flow.reset ();
175- }
173+
174+ proactor_->Await ([this ]() {
175+ // Destructor is blocking, so other fibers can observe partial state
176+ // of flows during clean up. To avoid this, we move them and clear the
177+ // member before the preemption point
178+ auto shard_flows = std::move (shard_flows_);
179+ shard_flows_.clear ();
180+ for (auto & flow : shard_flows) {
181+ flow.reset ();
182+ }
183+ shard_flows_.clear ();
184+ });
176185
177186 if (last_journal_LSNs_.has_value ()) {
178187 return LastMasterSyncData{master_context_.master_repl_id , last_journal_LSNs_.value ()};
@@ -501,29 +510,45 @@ error_code Replica::InitiatePSync() {
501510 return error_code{};
502511}
503512
504- // Initialize and start sub-replica for each flow.
505- error_code Replica::InitiateDflySync (std::optional<LastMasterSyncData> last_master_sync_data) {
506- auto start_time = absl::Now ( );
507-
508- // Initialize MultiShardExecution.
509- multi_shard_exe_. reset ( new MultiShardExecution () );
513+ void Replica::InitializeShardFlows () {
514+ decltype (shard_flows_) shard_flows_copy;
515+ shard_flows_copy. resize (master_context_. num_flows );
516+ DCHECK (!shard_flows_copy. empty ());
517+ thread_flow_map_ = Partition (shard_flows_copy. size ());
518+ const size_t pool_sz = shard_set-> pool ()-> size ( );
510519
511- // Initialize shard flows.
512- shard_flows_.resize (master_context_.num_flows );
513- DCHECK (!shard_flows_.empty ());
514- for (unsigned i = 0 ; i < shard_flows_.size (); ++i) {
515- // Transfer LSN state for partial sync
520+ for (size_t i = 0 ; i < shard_flows_copy.size (); ++i) {
516521 uint64_t partial_sync_lsn = 0 ;
517- if (shard_flows_[i]) {
522+ if (!shard_flows_. empty () && shard_flows_[i]) {
518523 partial_sync_lsn = shard_flows_[i]->JournalExecutedCount ();
519524 }
520- shard_flows_ [i].reset (
525+ shard_flows_copy [i].reset (
521526 new DflyShardReplica (server (), master_context_, i, &service_, multi_shard_exe_));
522527 if (partial_sync_lsn > 0 ) {
523- shard_flows_ [i]->SetRecordsExecuted (partial_sync_lsn);
528+ shard_flows_copy [i]->SetRecordsExecuted (partial_sync_lsn);
524529 }
525530 }
526- thread_flow_map_ = Partition (shard_flows_.size ());
531+
532+ shard_set->pool ()->AwaitFiberOnAll ([pool_sz, this , &shard_flows_copy](auto index, auto * ctx) {
533+ for (unsigned i = index; i < shard_flows_copy.size (); i += pool_sz) {
534+ shard_flows_copy[i]->SetSocketThread (ProactorBase::me ());
535+ }
536+ });
537+ // now update shard_flows on proactor thread
538+ shard_flows_ = std::move (shard_flows_copy);
539+ }
540+
541+ // Initialize and start sub-replica for each flow.
542+ error_code Replica::InitiateDflySync (std::optional<LastMasterSyncData> last_master_sync_data) {
543+ auto start_time = absl::Now ();
544+
545+ // Initialize MultiShardExecution.
546+ multi_shard_exe_.reset (new MultiShardExecution ());
547+
548+ // Initialize shard flows. The update to the shard_flows_ should be done by this thread.
549+ // Otherwise, there is a race condition between GetSummary() and the shard_flows_[i].reset()
550+ // below.
551+ InitializeShardFlows ();
527552
528553 // Blocked on until all flows got full sync cut.
529554 BlockingCounter sync_block{unsigned (shard_flows_.size ())};
@@ -1210,11 +1235,12 @@ error_code Replica::ParseReplicationHeader(base::IoBuf* io_buf, PSyncResponse* d
12101235
12111236auto Replica::GetSummary () const -> Summary {
12121237 auto f = [this ]() {
1238+ DCHECK (this );
12131239 auto last_io_time = LastIoTime ();
12141240
1215- // Note: we access LastIoTime from foreigh thread in unsafe manner. However, specifically here
1216- // it's unlikely to cause a real bug.
1217- for ( const auto & flow : shard_flows_) { // Get last io time from all sub flows.
1241+ for ( const auto & flow : shard_flows_) {
1242+ DCHECK ( Proactor () == ProactorBase::me ());
1243+ DCHECK ( flow);
12181244 last_io_time = std::max (last_io_time, flow->LastIoTime ());
12191245 }
12201246
@@ -1246,25 +1272,14 @@ auto Replica::GetSummary() const -> Summary {
12461272 return res;
12471273 };
12481274
1249- if (Sock ())
1250- return Proactor ()->AwaitBrief (f);
1251-
1252- /* *
1253- * when this branch happens: there is a very short grace period
1254- * where Sock() is not initialized, yet the server can
1255- * receive ROLE/INFO commands. That period happens when launching
1256- * an instance with '--replicaof' and then immediately
1257- * sending a command.
1258- *
1259- * In that instance, we have to run f() on the current fiber.
1260- */
1261- return f ();
1275+ return Proactor ()->AwaitBrief (f);
12621276}
12631277
12641278std::vector<uint64_t > Replica::GetReplicaOffset () const {
12651279 std::vector<uint64_t > flow_rec_count;
12661280 flow_rec_count.resize (shard_flows_.size ());
12671281 for (const auto & flow : shard_flows_) {
1282+ DCHECK (flow.get ());
12681283 uint32_t flow_id = flow->FlowId ();
12691284 uint64_t rec_count = flow->JournalExecutedCount ();
12701285 DCHECK_LT (flow_id, shard_flows_.size ());
0 commit comments