fix: fix RegisterOnChange methods for journal and db_slice (dragonfly…

…db#3171)

* fix: fix RegisterOnChange methods for journal and db_slice. Call db_slice and journal callbacks atomically. Made a hack to avoid deadlock during SAVE

src/server/db_slice.cc

@@ -458,9 +458,7 @@ OpResult<DbSlice::PrimeItAndExp> DbSlice::FindInternal(const Context& cntx, std:
     if (!change_cb_.empty()) {
       auto bump_cb = [&](PrimeTable::bucket_iterator bit) {
         DVLOG(2) << "Running callbacks for key " << key << " in dbid " << cntx.db_index;
-        for (const auto& ccb : change_cb_) {
-          ccb.second(cntx.db_index, bit);
-        }
+        CallChangeCallbacks(cntx.db_index, bit);
       };
       db.prime.CVCUponBump(change_cb_.back().first, res.it, bump_cb);
     }
@@ -524,9 +522,7 @@ OpResult<DbSlice::AddOrFindResult> DbSlice::AddOrFindInternal(const Context& cnt
 
   // It's a new entry.
   DVLOG(2) << "Running callbacks for key " << key << " in dbid " << cntx.db_index;
-  for (const auto& ccb : change_cb_) {
-    ccb.second(cntx.db_index, key);
-  }
+  CallChangeCallbacks(cntx.db_index, key);
 
   // In case we are loading from rdb file or replicating we want to disable conservative memory
   // checks (inside PrimeEvictionPolicy::CanGrow) and reject insertions only after we pass max
@@ -975,9 +971,7 @@ void DbSlice::PreUpdate(DbIndex db_ind, Iterator it, std::string_view key) {
   FiberAtomicGuard fg;
 
   DVLOG(2) << "Running callbacks in dbid " << db_ind;
-  for (const auto& ccb : change_cb_) {
-    ccb.second(db_ind, ChangeReq{it.GetInnerIt()});
-  }
+  CallChangeCallbacks(db_ind, ChangeReq{it.GetInnerIt()});
 
   // If the value has a pending stash, cancel it before any modification are applied.
   // Note: we don't delete offloaded values before updates, because a read-modify operation (like
@@ -1089,6 +1083,13 @@ void DbSlice::ExpireAllIfNeeded() {
 
 uint64_t DbSlice::RegisterOnChange(ChangeCallback cb) {
   uint64_t ver = NextVersion();
+
+  // TODO rewrite this logic to be more clear
+  // this mutex lock is needed to check that this method is not called simultaneously with
+  // change_cb_ calls and journal_slice::change_cb_arr_ calls.
+  // It can be unlocked anytime because DbSlice::RegisterOnChange
+  // and journal_slice::RegisterOnChange calls without preemption
+  std::lock_guard lk(cb_mu_);
   change_cb_.emplace_back(ver, std::move(cb));
   return ver;
 }
@@ -1099,6 +1100,7 @@ void DbSlice::FlushChangeToEarlierCallbacks(DbIndex db_ind, Iterator it, uint64_
   // change_cb_ is ordered by version.
   DVLOG(2) << "Running callbacks in dbid " << db_ind << " with bucket_version=" << bucket_version
            << ", upper_bound=" << upper_bound;
+
   for (const auto& ccb : change_cb_) {
     uint64_t cb_version = ccb.first;
     DCHECK_LE(cb_version, upper_bound);
@@ -1113,6 +1115,7 @@ void DbSlice::FlushChangeToEarlierCallbacks(DbIndex db_ind, Iterator it, uint64_
 
 //! Unregisters the callback.
 void DbSlice::UnregisterOnChange(uint64_t id) {
+  lock_guard lk(cb_mu_);  // we need to wait until callback is finished before remove it
   for (auto it = change_cb_.begin(); it != change_cb_.end(); ++it) {
     if (it->first == id) {
       change_cb_.erase(it);
@@ -1506,4 +1509,10 @@ void DbSlice::OnCbFinish() {
   fetched_items_.clear();
 }
 
+void DbSlice::CallChangeCallbacks(DbIndex id, const ChangeReq& cr) const {
+  for (const auto& ccb : change_cb_) {
+    ccb.second(id, cr);
+  }
+}
+
 }  // namespace dfly

src/server/db_slice.h

@@ -469,6 +469,14 @@ class DbSlice {
   void PerformDeletion(Iterator del_it, DbTable* table);
   void PerformDeletion(PrimeIterator del_it, DbTable* table);
 
+  void LockChangeCb() const {
+    return cb_mu_.lock_shared();
+  }
+
+  void UnlockChangeCb() const {
+    return cb_mu_.unlock_shared();
+  }
+
  private:
   void PreUpdate(DbIndex db_ind, Iterator it, std::string_view key);
   void PostUpdate(DbIndex db_ind, Iterator it, std::string_view key, size_t orig_size);
@@ -523,6 +531,8 @@ class DbSlice {
     return version_++;
   }
 
+  void CallChangeCallbacks(DbIndex id, const ChangeReq& cr) const;
+
  private:
   ShardId shard_id_;
   uint8_t caching_mode_ : 1;
@@ -544,6 +554,12 @@ class DbSlice {
   // Used in temporary computations in Acquire/Release.
   mutable absl::flat_hash_set<uint64_t> uniq_fps_;
 
+  // To ensure correct data replication, we must serialize the buckets that each running command
+  // will modify, followed by serializing the command to the journal. We use a mutex to prevent
+  // interleaving between bucket and journal registrations, and the command execution with its
+  // journaling. LockChangeCb is called before the callback, and UnlockChangeCb is called after
+  // journaling is completed. Register to bucket and journal changes is also does without preemption
+  mutable util::fb2::SharedMutex cb_mu_;
   // ordered from the smallest to largest version.
   std::vector<std::pair<uint64_t, ChangeCallback>> change_cb_;
 

src/server/detail/save_stages_controller.cc

@@ -254,7 +254,10 @@ void SaveStagesController::SaveDfs() {
 
   // Save shard files.
   auto cb = [this](Transaction* t, EngineShard* shard) {
+    // a hack to avoid deadlock in Transaction::RunCallback(...)
+    shard->db_slice().UnlockChangeCb();
     SaveDfsSingle(shard);
+    shard->db_slice().LockChangeCb();
     return OpStatus::OK;
   };
   trans_->ScheduleSingleHop(std::move(cb));
@@ -294,7 +297,10 @@ void SaveStagesController::SaveRdb() {
   }
 
   auto cb = [snapshot = snapshot.get()](Transaction* t, EngineShard* shard) {
+    // a hack to avoid deadlock in Transaction::RunCallback(...)
+    shard->db_slice().UnlockChangeCb();
     snapshot->StartInShard(shard);
+    shard->db_slice().LockChangeCb();
     return OpStatus::OK;
   };
   trans_->ScheduleSingleHop(std::move(cb));

src/server/journal/journal_slice.cc

@@ -199,13 +199,14 @@ void JournalSlice::AddLogRecord(const Entry& entry, bool await) {
 }
 
 uint32_t JournalSlice::RegisterOnChange(ChangeCallback cb) {
-  lock_guard lk(cb_mu_);
+  // mutex lock isn't needed due to iterators are not invalidated
   uint32_t id = next_cb_id_++;
   change_cb_arr_.emplace_back(id, std::move(cb));
   return id;
 }
 
 void JournalSlice::UnregisterOnChange(uint32_t id) {
+  // we need to wait until callback is finished before remove it
   lock_guard lk(cb_mu_);
   auto it = find_if(change_cb_arr_.begin(), change_cb_arr_.end(),
                     [id](const auto& e) { return e.first == id; });

src/server/journal/journal_slice.h

@@ -47,7 +47,6 @@ class JournalSlice {
   void UnregisterOnChange(uint32_t);
 
   bool HasRegisteredCallbacks() const {
-    std::shared_lock lk(cb_mu_);
     return !change_cb_arr_.empty();
   }
 
@@ -62,8 +61,8 @@ class JournalSlice {
   std::optional<base::RingBuffer<JournalItem>> ring_buffer_;
   base::IoBuf ring_serialize_buf_;
 
-  mutable util::fb2::SharedMutex cb_mu_;
-  std::vector<std::pair<uint32_t, ChangeCallback>> change_cb_arr_ ABSL_GUARDED_BY(cb_mu_);
+  mutable util::fb2::SharedMutex cb_mu_;  // to prevent removing callback during call
+  std::list<std::pair<uint32_t, ChangeCallback>> change_cb_arr_;
 
   LSN lsn_ = 1;
 

src/server/transaction.cc

@@ -627,6 +627,7 @@ void Transaction::RunCallback(EngineShard* shard) {
   DCHECK_EQ(shard, EngineShard::tlocal());
 
   RunnableResult result;
+  shard->db_slice().LockChangeCb();
   try {
     result = (*cb_ptr_)(this, shard);
 
@@ -664,7 +665,10 @@ void Transaction::RunCallback(EngineShard* shard) {
   // Log to journal only once the command finished running
   if ((coordinator_state_ & COORD_CONCLUDING) || (multi_ && multi_->concluding)) {
     LogAutoJournalOnShard(shard, result);
+    shard->db_slice().UnlockChangeCb();
     MaybeInvokeTrackingCb();
+  } else {
+    shard->db_slice().UnlockChangeCb();
   }
 }
 
@@ -1247,9 +1251,11 @@ OpStatus Transaction::RunSquashedMultiCb(RunnableType cb) {
   DCHECK_EQ(unique_shard_cnt_, 1u);
 
   auto* shard = EngineShard::tlocal();
+  shard->db_slice().LockChangeCb();
   auto result = cb(this, shard);
   shard->db_slice().OnCbFinish();
   LogAutoJournalOnShard(shard, result);
+  shard->db_slice().UnlockChangeCb();
   MaybeInvokeTrackingCb();
 
   DCHECK_EQ(result.flags, 0);  // if it's sophisticated, we shouldn't squash it