bump patch version number

Summary:
GetAndRefSuperVersion() should not be called again in the same thread before ReturnAndCleanupSuperVersion() is called.

If we have a compaction filter that is using DB::Get, This will happen
```
CompactFiles() {
  GetAndRefSuperVersion() // -- first call
    ..
    CompactionFilter() {
      GetAndRefSuperVersion() // -- second call
      ReturnAndCleanupSuperVersion()
    }
    ..
  ReturnAndCleanupSuperVersion()
}
```

We solve this issue in the same way Iterator is solving it, but using GetReferencedSuperVersion()

This was discovered in https://github.com/facebook/mysql-5.6/issues/427 by alxyang
Closes https://github.com/facebook/rocksdb/pull/1803

Differential Revision: D4460155

Pulled By: IslamAbdelRahman

fbshipit-source-id: 5e54322

HISTORY.md

@@ -1,5 +1,5 @@
 # Rocksdb Change Log
-## Unreleased
+## 5.1.0 (01/13/2016)
 ### Public API Change
 * Support dynamically change `delete_obsolete_files_period_micros` option via SetDBOptions().
 * Added EventListener::OnExternalFileIngested which will be called when IngestExternalFile() add a file successfully.

db/column_family.cc

@@ -370,7 +370,8 @@ ColumnFamilyData::ColumnFamilyData(
       column_family_set_(column_family_set),
       pending_flush_(false),
       pending_compaction_(false),
-      prev_compaction_needed_bytes_(0) {
+      prev_compaction_needed_bytes_(0),
+      allow_2pc_(db_options.allow_2pc) {
   Ref();
 
   // Convert user defined table properties collector factories to internal ones.
@@ -492,6 +493,25 @@ ColumnFamilyOptions ColumnFamilyData::GetLatestCFOptions() const {
   return BuildColumnFamilyOptions(initial_cf_options_, mutable_cf_options_);
 }
 
+uint64_t ColumnFamilyData::OldestLogToKeep() {
+  auto current_log = GetLogNumber();
+
+  if (allow_2pc_) {
+    auto imm_prep_log = imm()->GetMinLogContainingPrepSection();
+    auto mem_prep_log = mem()->GetMinLogContainingPrepSection();
+
+    if (imm_prep_log > 0 && imm_prep_log < current_log) {
+      current_log = imm_prep_log;
+    }
+
+    if (mem_prep_log > 0 && mem_prep_log < current_log) {
+      current_log = mem_prep_log;
+    }
+  }
+
+  return current_log;
+}
+
 const double kIncSlowdownRatio = 0.8;
 const double kDecSlowdownRatio = 1 / kIncSlowdownRatio;
 const double kNearStopSlowdownRatio = 0.6;

db/column_family.h

@@ -239,6 +239,9 @@ class ColumnFamilyData {
   uint64_t GetTotalSstFilesSize() const;  // REQUIRE: DB mutex held
   void SetMemtable(MemTable* new_mem) { mem_ = new_mem; }
 
+  // calculate the oldest log needed for the durability of this column family
+  uint64_t OldestLogToKeep();
+
   // See Memtable constructor for explanation of earliest_seq param.
   MemTable* ConstructNewMemtable(const MutableCFOptions& mutable_cf_options,
                                  SequenceNumber earliest_seq);
@@ -404,6 +407,9 @@ class ColumnFamilyData {
   bool pending_compaction_;
 
   uint64_t prev_compaction_needed_bytes_;
+
+  // if the database was opened with 2pc enabled
+  bool allow_2pc_;
 };
 
 // ColumnFamilySet has interesting thread-safety requirements

db/compact_files_test.cc

@@ -253,6 +253,61 @@ TEST_F(CompactFilesTest, CapturingPendingFiles) {
   delete db;
 }
 
+TEST_F(CompactFilesTest, CompactionFilterWithGetSv) {
+  class FilterWithGet : public CompactionFilter {
+   public:
+    virtual bool Filter(int level, const Slice& key, const Slice& value,
+                        std::string* new_value,
+                        bool* value_changed) const override {
+      if (db_ == nullptr) {
+        return true;
+      }
+      std::string res;
+      db_->Get(ReadOptions(), "", &res);
+      return true;
+    }
+
+    void SetDB(DB* db) {
+      db_ = db;
+    }
+
+    virtual const char* Name() const override { return "FilterWithGet"; }
+
+   private:
+    DB* db_;
+  };
+
+
+  std::shared_ptr<FilterWithGet> cf(new FilterWithGet());
+
+  Options options;
+  options.create_if_missing = true;
+  options.compaction_filter = cf.get();
+
+  DB* db = nullptr;
+  DestroyDB(db_name_, options);
+  Status s = DB::Open(options, db_name_, &db);
+  ASSERT_OK(s);
+
+  cf->SetDB(db);
+
+  // Write one L0 file
+  db->Put(WriteOptions(), "K1", "V1");
+  db->Flush(FlushOptions());
+
+  // Compact all L0 files using CompactFiles
+  rocksdb::ColumnFamilyMetaData meta;
+  db->GetColumnFamilyMetaData(&meta);
+  for (auto& file : meta.levels[0].files) {
+    std::string fname = file.db_path + "/" + file.name;
+    ASSERT_OK(
+        db->CompactFiles(rocksdb::CompactionOptions(), {fname}, 0));
+  }
+
+
+  delete db;
+}
+
 }  // namespace rocksdb
 
 int main(int argc, char** argv) {

db/compaction_job.cc

@@ -771,9 +771,9 @@ void CompactionJob::ProcessKeyValueCompaction(SubcompactionState* sub_compact) {
                    key, sub_compact->current_output_file_size) &&
                sub_compact->builder != nullptr) {
       CompactionIterationStats range_del_out_stats;
-      status =
-          FinishCompactionOutputFile(input->status(), sub_compact,
-                                     range_del_agg.get(), &range_del_out_stats);
+      status = FinishCompactionOutputFile(input->status(), sub_compact,
+                                          range_del_agg.get(),
+                                          &range_del_out_stats, &key);
       RecordDroppedKeys(range_del_out_stats,
                         &sub_compact->compaction_job_stats);
       if (!status.ok()) {

db/db_flush_test.cc

@@ -9,6 +9,7 @@
 
 #include "db/db_test_util.h"
 #include "port/stack_trace.h"
+#include "util/fault_injection_test_env.h"
 #include "util/sync_point.h"
 
 namespace rocksdb {
@@ -47,6 +48,38 @@ TEST_F(DBFlushTest, FlushWhileWritingManifest) {
 #endif  // ROCKSDB_LITE
 }
 
+TEST_F(DBFlushTest, SyncFail) {
+  std::unique_ptr<FaultInjectionTestEnv> fault_injection_env(
+      new FaultInjectionTestEnv(Env::Default()));
+  Options options;
+  options.disable_auto_compactions = true;
+  options.env = fault_injection_env.get();
+
+  SyncPoint::GetInstance()->LoadDependency(
+      {{"DBFlushTest::SyncFail:1", "DBImpl::SyncClosedLogs:Start"},
+       {"DBImpl::SyncClosedLogs:Failed", "DBFlushTest::SyncFail:2"}});
+  SyncPoint::GetInstance()->EnableProcessing();
+
+  Reopen(options);
+  Put("key", "value");
+  auto* cfd =
+      reinterpret_cast<ColumnFamilyHandleImpl*>(db_->DefaultColumnFamily())
+          ->cfd();
+  int refs_before = cfd->current()->TEST_refs();
+  FlushOptions flush_options;
+  flush_options.wait = false;
+  ASSERT_OK(dbfull()->Flush(flush_options));
+  fault_injection_env->SetFilesystemActive(false);
+  TEST_SYNC_POINT("DBFlushTest::SyncFail:1");
+  TEST_SYNC_POINT("DBFlushTest::SyncFail:2");
+  fault_injection_env->SetFilesystemActive(true);
+  dbfull()->TEST_WaitForFlushMemTable();
+  ASSERT_EQ("", FilesPerLevel());  // flush failed.
+  // Flush job should release ref count to current version.
+  ASSERT_EQ(refs_before, cfd->current()->TEST_refs());
+  Destroy(options);
+}
+
 }  // namespace rocksdb
 
 int main(int argc, char** argv) {

db/db_impl.cc

@@ -342,6 +342,7 @@ DBImpl::DBImpl(const DBOptions& options, const std::string& dbname)
       last_stats_dump_time_microsec_(0),
       next_job_id_(1),
       has_unpersisted_data_(false),
+      unable_to_flush_oldest_log_(false),
       env_options_(BuildDBOptions(immutable_db_options_, mutable_db_options_)),
       num_running_ingest_file_(0),
 #ifndef ROCKSDB_LITE
@@ -663,6 +664,10 @@ void DBImpl::MaybeDumpStats() {
 }
 
 uint64_t DBImpl::FindMinPrepLogReferencedByMemTable() {
+  if (!allow_2pc()) {
+    return 0;
+  }
+
   uint64_t min_log = 0;
 
   // we must look through the memtables for two phase transactions
@@ -707,6 +712,11 @@ void DBImpl::MarkLogAsContainingPrepSection(uint64_t log) {
 }
 
 uint64_t DBImpl::FindMinLogContainingOutstandingPrep() {
+
+  if (!allow_2pc()) {
+    return 0;
+  }
+
   std::lock_guard<std::mutex> lock(prep_heap_mutex_);
   uint64_t min_log = 0;
 
@@ -1832,6 +1842,7 @@ Status DBImpl::WriteLevel0TableForRecovery(int job_id, ColumnFamilyData* cfd,
 }
 
 Status DBImpl::SyncClosedLogs(JobContext* job_context) {
+  TEST_SYNC_POINT("DBImpl::SyncClosedLogs:Start");
   mutex_.AssertHeld();
   autovector<log::Writer*, 1> logs_to_sync;
   uint64_t current_log_number = logfile_number_;
@@ -1868,6 +1879,7 @@ Status DBImpl::SyncClosedLogs(JobContext* job_context) {
     MarkLogsSynced(current_log_number - 1, true, s);
     if (!s.ok()) {
       bg_error_ = s;
+      TEST_SYNC_POINT("DBImpl::SyncClosedLogs:Failed");
       return s;
     }
   }
@@ -1918,6 +1930,8 @@ Status DBImpl::FlushMemTableToOutputFile(
   // is unlocked by the current thread.
   if (s.ok()) {
     s = flush_job.Run(&file_meta);
+  } else {
+    flush_job.Cancel();
   }
 
   if (s.ok()) {
@@ -2135,7 +2149,7 @@ Status DBImpl::CompactFiles(
                        immutable_db_options_.info_log.get());
 
   // Perform CompactFiles
-  SuperVersion* sv = GetAndRefSuperVersion(cfd);
+  SuperVersion* sv = cfd->GetReferencedSuperVersion(&mutex_);
   {
     InstrumentedMutexLock l(&mutex_);
 
@@ -2147,7 +2161,12 @@ Status DBImpl::CompactFiles(
                          input_file_names, output_level,
                          output_path_id, &job_context, &log_buffer);
   }
-  ReturnAndCleanupSuperVersion(cfd, sv);
+  if (sv->Unref()) {
+    mutex_.Lock();
+    sv->Cleanup();
+    mutex_.Unlock();
+    delete sv;
+  }
 
   // Find and delete obsolete files
   {
@@ -2505,7 +2524,7 @@ Status DBImpl::SetDBOptions(
       mutable_db_options_ = new_options;
 
       if (total_log_size_ > GetMaxTotalWalSize()) {
-        FlushColumnFamilies();
+        MaybeFlushColumnFamilies();
       }
 
       persist_options_status = PersistOptions();
@@ -2752,7 +2771,7 @@ void DBImpl::MarkLogsSynced(
       ++it;
     }
   }
-  assert(logs_.empty() || logs_[0].number > up_to ||
+  assert(!status.ok() || logs_.empty() || logs_[0].number > up_to ||
          (logs_.size() == 1 && !logs_[0].getting_synced));
   log_sync_cv_.SignalAll();
 }
@@ -4698,9 +4717,8 @@ Status DBImpl::WriteImpl(const WriteOptions& write_options,
          versions_->GetColumnFamilySet()->NumberOfColumnFamilies() == 1);
 
   if (UNLIKELY(!single_column_family_mode_ &&
-               !alive_log_files_.begin()->getting_flushed &&
                total_log_size_ > GetMaxTotalWalSize())) {
-    FlushColumnFamilies();
+    MaybeFlushColumnFamilies();
   } else if (UNLIKELY(write_buffer_manager_->ShouldFlush())) {
     // Before a new memtable is added in SwitchMemtable(),
     // write_buffer_manager_->ShouldFlush() will keep returning true. If another
@@ -5018,28 +5036,40 @@ Status DBImpl::WriteImpl(const WriteOptions& write_options,
   return status;
 }
 
-void DBImpl::FlushColumnFamilies() {
+void DBImpl::MaybeFlushColumnFamilies() {
   mutex_.AssertHeld();
 
-  WriteContext context;
-
   if (alive_log_files_.begin()->getting_flushed) {
     return;
   }
 
-  uint64_t flush_column_family_if_log_file = alive_log_files_.begin()->number;
-  alive_log_files_.begin()->getting_flushed = true;
+  auto oldest_alive_log = alive_log_files_.begin()->number;
+  auto oldest_log_with_uncommited_prep = FindMinLogContainingOutstandingPrep();
+
+  if (allow_2pc() &&
+      unable_to_flush_oldest_log_ &&
+      oldest_log_with_uncommited_prep > 0 &&
+      oldest_log_with_uncommited_prep <= oldest_alive_log) {
+    // we already attempted to flush all column families dependent on
+    // the oldest alive log but the log still contained uncommited transactions.
+    // the oldest alive log STILL contains uncommited transaction so there
+    // is still nothing that we can do.
+    return;
+  }
+
+  WriteContext context;
+
   Log(InfoLogLevel::INFO_LEVEL, immutable_db_options_.info_log,
       "Flushing all column families with data in WAL number %" PRIu64
       ". Total log size is %" PRIu64 " while max_total_wal_size is %" PRIu64,
-      flush_column_family_if_log_file, total_log_size_, GetMaxTotalWalSize());
+      oldest_alive_log, total_log_size_, GetMaxTotalWalSize());
   // no need to refcount because drop is happening in write thread, so can't
   // happen while we're in the write thread
   for (auto cfd : *versions_->GetColumnFamilySet()) {
     if (cfd->IsDropped()) {
       continue;
     }
-    if (cfd->GetLogNumber() <= flush_column_family_if_log_file) {
+    if (cfd->OldestLogToKeep() <= oldest_alive_log) {
       auto status = SwitchMemtable(cfd, &context);
       if (!status.ok()) {
         break;
@@ -5049,6 +5079,26 @@ void DBImpl::FlushColumnFamilies() {
     }
   }
   MaybeScheduleFlushOrCompaction();
+
+  // we only mark this log as getting flushed if we have successfully
+  // flushed all data in this log. If this log contains outstanding prepred
+  // transactions then we cannot flush this log until those transactions are commited.
+
+  unable_to_flush_oldest_log_ = false;
+
+  if (allow_2pc()) {
+    if (oldest_log_with_uncommited_prep == 0 ||
+        oldest_log_with_uncommited_prep > oldest_alive_log) {
+      // this log contains no outstanding prepared transactions
+      alive_log_files_.begin()->getting_flushed = true;
+    } else {
+      Log(InfoLogLevel::WARN_LEVEL, immutable_db_options_.info_log,
+          "Unable to release oldest log due to uncommited transaction");
+      unable_to_flush_oldest_log_ = true;
+    }
+  } else {
+    alive_log_files_.begin()->getting_flushed = true;
+  }
 }
 
 uint64_t DBImpl::GetMaxTotalWalSize() const {

db/db_impl.h

@@ -308,6 +308,16 @@ class DBImpl : public DB {
                            ColumnFamilyHandle* column_family = nullptr,
                            bool disallow_trivial_move = false);
 
+  void TEST_MaybeFlushColumnFamilies();
+
+  bool TEST_UnableToFlushOldestLog() {
+    return unable_to_flush_oldest_log_;
+  }
+
+  bool TEST_IsLogGettingFlushed() {
+    return alive_log_files_.begin()->getting_flushed;
+  }
+
   // Force current memtable contents to be flushed.
   Status TEST_FlushMemTable(bool wait = true,
                             ColumnFamilyHandle* cfh = nullptr);
@@ -734,7 +744,7 @@ class DBImpl : public DB {
   // REQUIRES: mutex locked
   Status PersistOptions();
 
-  void FlushColumnFamilies();
+  void MaybeFlushColumnFamilies();
 
   uint64_t GetMaxTotalWalSize() const;
 
@@ -994,6 +1004,15 @@ class DBImpl : public DB {
   // Used when disableWAL is true.
   bool has_unpersisted_data_;
 
+
+  // if an attempt was made to flush all column families that
+  // the oldest log depends on but uncommited data in the oldest
+  // log prevents the log from being released.
+  // We must attempt to free the dependent memtables again
+  // at a later time after the transaction in the oldest
+  // log is fully commited.
+  bool unable_to_flush_oldest_log_;
+
   static const int KEEP_LOG_FILE_NUM = 1000;
   // MSVC version 1800 still does not have constexpr for ::max()
   static const uint64_t kNoTimeOut = port::kMaxUint64;

db/db_impl_debug.cc

@@ -19,6 +19,11 @@ uint64_t DBImpl::TEST_GetLevel0TotalSize() {
   return default_cf_handle_->cfd()->current()->storage_info()->NumLevelBytes(0);
 }
 
+void DBImpl::TEST_MaybeFlushColumnFamilies() {
+  InstrumentedMutexLock l(&mutex_);
+  MaybeFlushColumnFamilies();
+}
+
 int64_t DBImpl::TEST_MaxNextLevelOverlappingBytes(
     ColumnFamilyHandle* column_family) {
   ColumnFamilyData* cfd;

db/db_range_del_test.cc

@@ -112,6 +112,81 @@ TEST_F(DBRangeDelTest, CompactionOutputFilesExactlyFilled) {
   db_->ReleaseSnapshot(snapshot);
 }
 
+TEST_F(DBRangeDelTest, MaxCompactionBytesCutsOutputFiles) {
+  // Ensures range deletion spanning multiple compaction output files that are
+  // cut by max_compaction_bytes will have non-overlapping key-ranges.
+  // https://github.com/facebook/rocksdb/issues/1778
+  const int kNumFiles = 2, kNumPerFile = 1 << 8, kBytesPerVal = 1 << 12;
+  Options opts = CurrentOptions();
+  opts.comparator = test::Uint64Comparator();
+  opts.disable_auto_compactions = true;
+  opts.level0_file_num_compaction_trigger = kNumFiles;
+  opts.max_compaction_bytes = kNumPerFile * kBytesPerVal;
+  opts.memtable_factory.reset(new SpecialSkipListFactory(kNumPerFile));
+  // Want max_compaction_bytes to trigger the end of compaction output file, not
+  // target_file_size_base, so make the latter much bigger
+  opts.target_file_size_base = 100 * opts.max_compaction_bytes;
+  Reopen(opts);
+
+  // snapshot protects range tombstone from dropping due to becoming obsolete.
+  const Snapshot* snapshot = db_->GetSnapshot();
+
+  // It spans the whole key-range, thus will be included in all output files
+  ASSERT_OK(db_->DeleteRange(WriteOptions(), db_->DefaultColumnFamily(),
+                             GetNumericStr(0),
+                             GetNumericStr(kNumFiles * kNumPerFile - 1)));
+  Random rnd(301);
+  for (int i = 0; i < kNumFiles; ++i) {
+    std::vector<std::string> values;
+    // Write 1MB (256 values, each 4K)
+    for (int j = 0; j < kNumPerFile; j++) {
+      values.push_back(RandomString(&rnd, kBytesPerVal));
+      ASSERT_OK(Put(GetNumericStr(kNumPerFile * i + j), values[j]));
+    }
+    // extra entry to trigger SpecialSkipListFactory's flush
+    ASSERT_OK(Put(GetNumericStr(kNumPerFile), ""));
+    dbfull()->TEST_WaitForFlushMemTable();
+    ASSERT_EQ(i + 1, NumTableFilesAtLevel(0));
+  }
+
+  dbfull()->TEST_CompactRange(0, nullptr, nullptr, nullptr,
+                              true /* disallow_trivial_move */);
+  ASSERT_EQ(0, NumTableFilesAtLevel(0));
+  ASSERT_GE(NumTableFilesAtLevel(1), 2);
+
+  std::vector<std::vector<FileMetaData>> files;
+  dbfull()->TEST_GetFilesMetaData(db_->DefaultColumnFamily(), &files);
+
+  for (size_t i = 0; i < files[1].size() - 1; ++i) {
+    ASSERT_TRUE(InternalKeyComparator(opts.comparator)
+                    .Compare(files[1][i].largest, files[1][i + 1].smallest) <
+                0);
+  }
+  db_->ReleaseSnapshot(snapshot);
+}
+
+TEST_F(DBRangeDelTest, SentinelsOmittedFromOutputFile) {
+  // Regression test for bug where sentinel range deletions (i.e., ones with
+  // sequence number of zero) were included in output files.
+  // snapshot protects range tombstone from dropping due to becoming obsolete.
+  const Snapshot* snapshot = db_->GetSnapshot();
+
+  // gaps between ranges creates sentinels in our internal representation
+  std::vector<std::pair<std::string, std::string>> range_dels = {{"a", "b"}, {"c", "d"}, {"e", "f"}};
+  for (const auto& range_del : range_dels) {
+    ASSERT_OK(db_->DeleteRange(WriteOptions(), db_->DefaultColumnFamily(),
+                               range_del.first, range_del.second));
+  }
+  ASSERT_OK(db_->Flush(FlushOptions()));
+  ASSERT_EQ(1, NumTableFilesAtLevel(0));
+
+  std::vector<std::vector<FileMetaData>> files;
+  dbfull()->TEST_GetFilesMetaData(db_->DefaultColumnFamily(), &files);
+  ASSERT_GT(files[0][0].smallest_seqno, 0);
+
+  db_->ReleaseSnapshot(snapshot);
+}
+
 TEST_F(DBRangeDelTest, FlushRangeDelsSameStartKey) {
   db_->Put(WriteOptions(), "b1", "val");
   ASSERT_OK(
@@ -475,6 +550,36 @@ TEST_F(DBRangeDelTest, GetCoveredKeyFromSst) {
   db_->ReleaseSnapshot(snapshot);
 }
 
+TEST_F(DBRangeDelTest, GetCoveredMergeOperandFromMemtable) {
+  const int kNumMergeOps = 10;
+  Options opts = CurrentOptions();
+  opts.merge_operator = MergeOperators::CreateUInt64AddOperator();
+  Reopen(opts);
+
+  for (int i = 0; i < kNumMergeOps; ++i) {
+    std::string val;
+    PutFixed64(&val, i);
+    db_->Merge(WriteOptions(), "key", val);
+    if (i == kNumMergeOps / 2) {
+      // deletes [0, 5]
+      db_->DeleteRange(WriteOptions(), db_->DefaultColumnFamily(), "key",
+                       "key_");
+    }
+  }
+
+  ReadOptions read_opts;
+  std::string expected, actual;
+  ASSERT_OK(db_->Get(read_opts, "key", &actual));
+  PutFixed64(&expected, 30);  // 6+7+8+9
+  ASSERT_EQ(expected, actual);
+
+  expected.clear();
+  read_opts.ignore_range_deletions = true;
+  ASSERT_OK(db_->Get(read_opts, "key", &actual));
+  PutFixed64(&expected, 45);  // 0+1+2+...+9
+  ASSERT_EQ(expected, actual);
+}
+
 TEST_F(DBRangeDelTest, GetIgnoresRangeDeletions) {
   Options opts = CurrentOptions();
   opts.max_write_buffer_number = 4;

db/db_test2.cc

@@ -2216,6 +2216,18 @@ TEST_F(DBTest2, ManualCompactionOverlapManualCompaction) {
 
   rocksdb::SyncPoint::GetInstance()->DisableProcessing();
 }
+
+TEST_F(DBTest2, OptimizeForPointLookup) {
+  Options options = CurrentOptions();
+  Close();
+  options.OptimizeForPointLookup(2);
+  ASSERT_OK(DB::Open(options, dbname_, &db_));
+
+  ASSERT_OK(Put("foo", "v1"));
+  ASSERT_EQ("v1", Get("foo"));
+  Flush();
+  ASSERT_EQ("v1", Get("foo"));
+}
 #endif  // ROCKSDB_LITE
 
 }  // namespace rocksdb

db/flush_job.cc

@@ -230,6 +230,12 @@ Status FlushJob::Run(FileMetaData* file_meta) {
   return s;
 }
 
+void FlushJob::Cancel() {
+  db_mutex_->AssertHeld();
+  assert(base_ != nullptr);
+  base_->Unref();
+}
+
 Status FlushJob::WriteLevel0Table() {
   AutoThreadOperationStageUpdater stage_updater(
       ThreadStatus::STAGE_FLUSH_WRITE_L0);

db/flush_job.h

@@ -67,9 +67,11 @@ class FlushJob {
 
   ~FlushJob();
 
-  // Require db_mutex held
+  // Require db_mutex held.
+  // Once PickMemTable() is called, either Run() or Cancel() has to be call.
   void PickMemTable();
   Status Run(FileMetaData* file_meta = nullptr);
+  void Cancel();
   TableProperties GetTableProperties() const { return table_properties_; }
 
  private:

db/memtable.cc

@@ -560,7 +560,7 @@ static bool SaveValue(void* arg, const char* entry) {
     ValueType type;
     UnPackSequenceAndType(tag, &s->seq, &type);
 
-    if ((type == kTypeValue || type == kTypeDeletion) &&
+    if ((type == kTypeValue || type == kTypeMerge) &&
         range_del_agg->ShouldDelete(Slice(key_ptr, key_length))) {
       type = kTypeRangeDeletion;
     }

db/memtable_list.h

@@ -221,6 +221,8 @@ class MemTableList {
   // PickMemtablesToFlush() is called.
   void FlushRequested() { flush_requested_ = true; }
 
+  bool HasFlushRequested() { return flush_requested_; }
+
   // Copying allowed
   // MemTableList(const MemTableList&);
   // void operator=(const MemTableList&);

db/range_del_aggregator.cc

@@ -420,9 +420,10 @@ void RangeDelAggregator::AddToBuilder(
       RangeTombstone tombstone;
       if (collapse_deletions_) {
         auto next_tombstone_map_iter = std::next(tombstone_map_iter);
-        if (next_tombstone_map_iter == stripe_map_iter->second.raw_map.end()) {
-          // it's the sentinel tombstone
-          break;
+        if (next_tombstone_map_iter == stripe_map_iter->second.raw_map.end() ||
+            tombstone_map_iter->second.seq_ == 0) {
+          // it's a sentinel tombstone
+          continue;
         }
         tombstone.start_key_ = tombstone_map_iter->first;
         tombstone.end_key_ = next_tombstone_map_iter->first;

db/version_set.h

@@ -520,6 +520,8 @@ class Version {
     return next_;
   }
 
+  int TEST_refs() const { return refs_; }
+
   VersionStorageInfo* storage_info() { return &storage_info_; }
 
   VersionSet* version_set() { return vset_; }

db/write_thread.cc

@@ -317,7 +317,7 @@ void WriteThread::LaunchParallelFollowers(ParallelGroup* pg,
 
   while (w != pg->last_writer) {
     // Writers that won't write don't get sequence allotment
-    if (!w->CallbackFailed()) {
+    if (!w->CallbackFailed() && w->ShouldWriteToMemtable()) {
       sequence += WriteBatchInternal::Count(w->batch);
     }
     w = w->link_newer;

include/rocksdb/version.h

@@ -5,8 +5,8 @@
 #pragma once
 
 #define ROCKSDB_MAJOR 5
-#define ROCKSDB_MINOR 0
-#define ROCKSDB_PATCH 0
+#define ROCKSDB_MINOR 1
+#define ROCKSDB_PATCH 2
 
 // Do not use these. We made the mistake of declaring macros starting with
 // double underscore. Now we have to live with our choice. We'll deprecate these

util/fault_injection_test_env.cc

@@ -149,7 +149,7 @@ Status TestWritableFile::Flush() {
 
 Status TestWritableFile::Sync() {
   if (!env_->IsFilesystemActive()) {
-    return Status::OK();
+    return Status::IOError("FaultInjectionTestEnv: not active");
   }
   // No need to actual sync.
   state_.pos_at_last_sync_ = state_.pos_;

util/io_posix.cc

@@ -265,11 +265,6 @@ Status PosixRandomAccessFile::Read(uint64_t offset, size_t n, Slice* result,
     // An error: return a non-ok status
     s = IOError(filename_, errno);
   }
-  if (!use_direct_io()) {
-    // we need to fadvise away the entire range of pages because
-    // we do not want readahead pages to be cached.
-    Fadvise(fd_, 0, 0, POSIX_FADV_DONTNEED);  // free OS pages
-  }
   *result = Slice(scratch, (r < 0) ? 0 : n - left);
   return s;
 }

util/options.cc

@@ -700,7 +700,7 @@ ColumnFamilyOptions* ColumnFamilyOptions::OptimizeForPointLookup(
   block_based_options.block_cache =
       NewLRUCache(static_cast<size_t>(block_cache_size_mb * 1024 * 1024));
   table_factory.reset(new BlockBasedTableFactory(block_based_options));
-  memtable_factory.reset(NewHashLinkListRepFactory());
+  memtable_prefix_bloom_size_ratio = 0.02;
   return this;
 }
 

utilities/transactions/transaction_test.cc

@@ -1355,6 +1355,108 @@ TEST_P(TransactionTest, TwoPhaseLogRollingTest) {
   delete cfb;
 }
 
+TEST_P(TransactionTest, TwoPhaseLogRollingTest2) {
+  DBImpl* db_impl = reinterpret_cast<DBImpl*>(db->GetRootDB());
+
+  Status s;
+  ColumnFamilyHandle *cfa, *cfb;
+
+  ColumnFamilyOptions cf_options;
+  s = db->CreateColumnFamily(cf_options, "CFA", &cfa);
+  ASSERT_OK(s);
+  s = db->CreateColumnFamily(cf_options, "CFB", &cfb);
+  ASSERT_OK(s);
+
+  WriteOptions wopts;
+  wopts.disableWAL = false;
+  wopts.sync = true;
+
+  auto cfh_a = reinterpret_cast<ColumnFamilyHandleImpl*>(cfa);
+  auto cfh_b = reinterpret_cast<ColumnFamilyHandleImpl*>(cfb);
+
+  TransactionOptions topts1;
+  Transaction* txn1 = db->BeginTransaction(wopts, topts1);
+  s = txn1->SetName("xid1");
+  ASSERT_OK(s);
+  s = txn1->Put(cfa, "boys", "girls1");
+  ASSERT_OK(s);
+
+  Transaction* txn2 = db->BeginTransaction(wopts, topts1);
+  s = txn2->SetName("xid2");
+  ASSERT_OK(s);
+  s = txn2->Put(cfb, "up", "down1");
+  ASSERT_OK(s);
+
+  // prepre transaction in LOG A
+  s = txn1->Prepare();
+  ASSERT_OK(s);
+
+  // prepre transaction in LOG A
+  s = txn2->Prepare();
+  ASSERT_OK(s);
+
+  // regular put so that mem table can actually be flushed for log rolling
+  s = db->Put(wopts, "cats", "dogs1");
+  ASSERT_OK(s);
+
+  auto prepare_log_no = txn1->GetLogNumber();
+
+  // roll to LOG B
+  s = db_impl->TEST_FlushMemTable(true);
+  ASSERT_OK(s);
+
+  // now we pause background work so that
+  // imm()s are not flushed before we can check their status
+  s = db_impl->PauseBackgroundWork();
+  ASSERT_OK(s);
+
+  ASSERT_GT(db_impl->TEST_LogfileNumber(), prepare_log_no);
+  ASSERT_GT(cfh_a->cfd()->GetLogNumber(), prepare_log_no);
+  ASSERT_EQ(cfh_a->cfd()->GetLogNumber(), db_impl->TEST_LogfileNumber());
+  ASSERT_EQ(db_impl->TEST_FindMinLogContainingOutstandingPrep(),
+            prepare_log_no);
+  ASSERT_EQ(db_impl->TEST_FindMinPrepLogReferencedByMemTable(), 0);
+
+  // commit in LOG B
+  s = txn1->Commit();
+  ASSERT_OK(s);
+
+  ASSERT_EQ(db_impl->TEST_FindMinPrepLogReferencedByMemTable(), prepare_log_no);
+
+  ASSERT_TRUE(!db_impl->TEST_UnableToFlushOldestLog());
+
+  // request a flush for all column families such that the earliest
+  // alive log file can be killed
+  db_impl->TEST_MaybeFlushColumnFamilies();
+  // log cannot be flushed because txn2 has not been commited
+  ASSERT_TRUE(!db_impl->TEST_IsLogGettingFlushed());
+  ASSERT_TRUE(db_impl->TEST_UnableToFlushOldestLog());
+
+  // assert that cfa has a flush requested
+  ASSERT_TRUE(cfh_a->cfd()->imm()->HasFlushRequested());
+
+  // cfb should not be flushed becuse it has no data from LOG A
+  ASSERT_TRUE(!cfh_b->cfd()->imm()->HasFlushRequested());
+
+  // cfb now has data from LOG A
+  s = txn2->Commit();
+  ASSERT_OK(s);
+
+  db_impl->TEST_MaybeFlushColumnFamilies();
+  ASSERT_TRUE(!db_impl->TEST_UnableToFlushOldestLog());
+
+  // we should see that cfb now has a flush requested
+  ASSERT_TRUE(cfh_b->cfd()->imm()->HasFlushRequested());
+
+  // all data in LOG A resides in a memtable that has been
+  // requested for a flush
+  ASSERT_TRUE(db_impl->TEST_IsLogGettingFlushed());
+
+  delete txn1;
+  delete txn2;
+  delete cfa;
+  delete cfb;
+}
 /*
  * 1) use prepare to keep first log around to determine starting sequence
  * during recovery.