Ensure writes to WAL tail during FlushWAL(true /* sync */) will be synced (facebook#10560)

ajkr · Connor1996 · commit f815de7641e3 · 2024-03-25T11:14:04.000+08:00
Summary: WAL append and switch can both happen between `FlushWAL(true /* sync */)`'s sync operations and its call to `MarkLogsSynced()`. We permit this since locks need to be released for the sync operations. Such an appended/switched WAL is both inactive and incompletely synced at the time `MarkLogsSynced()` processes it. Prior to this PR, `MarkLogsSynced()` assumed all inactive WALs were fully synced and removed them from consideration for future syncs. That was wrong in the scenario described above and led to the latest append(s) never being synced. This PR changes `MarkLogsSynced()` to only remove inactive WALs from consideration for which all flushed data has been synced. Pull Request resolved: facebook#10560 Test Plan: repro unit test for the scenario described above. Without this PR, it fails on "key2" not found Reviewed By: riversand963 Differential Revision: D38957391 Pulled By: ajkr fbshipit-source-id: da77175eba97ff251a4219b227b3bb2d4843ed26
diff --git a/HISTORY.md b/HISTORY.md
@@ -1,6 +1,8 @@
 # Rocksdb Change Log
 
 ## Unreleased
+### Bug Fixes
+* Fixed bug where `FlushWAL(true /* sync */)` (used by `GetLiveFilesStorageInfo()`, which is used by checkpoint and backup) could cause parallel writes at the tail of a WAL file to never be synced.
 
 ### Bug Fixes
 
diff --git a/db/db_impl/db_impl.cc b/db/db_impl/db_impl.cc
@@ -1488,20 +1488,28 @@ void DBImpl::MarkLogsSynced(uint64_t up_to, bool synced_dir,
     auto& wal = *it;
     assert(wal.IsSyncing());
 
-    if (logs_.size() > 1) {
+    if (wal.number < logs_.back().number) {
+      // Inactive WAL
       if (immutable_db_options_.track_and_verify_wals_in_manifest &&
           wal.GetPreSyncSize() > 0) {
         synced_wals->AddWal(wal.number, WalMetadata(wal.GetPreSyncSize()));
       }
-      logs_to_free_.push_back(wal.ReleaseWriter());
-      it = logs_.erase(it);
+      if (wal.GetPreSyncSize() == wal.writer->file()->GetFlushedSize()) {
+        // Fully synced
+        logs_to_free_.push_back(wal.ReleaseWriter());
+        it = logs_.erase(it);
+      } else {
+        assert(wal.GetPreSyncSize() < wal.writer->file()->GetFlushedSize());
+        wal.FinishSync();
+        ++it;
+      }
     } else {
+      assert(wal.number == logs_.back().number);
+      // Active WAL
       wal.FinishSync();
       ++it;
     }
   }
-  assert(logs_.empty() || logs_[0].number > up_to ||
-         (logs_.size() == 1 && !logs_[0].IsSyncing()));
   log_sync_cv_.SignalAll();
 }
 
diff --git a/db/db_write_test.cc b/db/db_write_test.cc
@@ -386,6 +386,51 @@ TEST_P(DBWriteTest, UnflushedPutRaceWithTrackedWalSync) {
   Close();
 }
 
+TEST_P(DBWriteTest, InactiveWalFullySyncedBeforeUntracked) {
+  // Repro bug where a WAL is appended and switched after
+  // `FlushWAL(true /* sync */)`'s sync finishes and before it untracks fully
+  // synced inactive logs. Previously such a WAL would be wrongly untracked
+  // so the final append would never be synced.
+  Options options = GetOptions();
+  std::unique_ptr<FaultInjectionTestEnv> fault_env(
+      new FaultInjectionTestEnv(env_));
+  options.env = fault_env.get();
+  Reopen(options);
+
+  ASSERT_OK(Put("key1", "val1"));
+
+  SyncPoint::GetInstance()->SetCallBack(
+      "DBImpl::SyncWAL:BeforeMarkLogsSynced:1", [this](void* /* arg */) {
+        ASSERT_OK(Put("key2", "val2"));
+        ASSERT_OK(dbfull()->TEST_SwitchMemtable());
+      });
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->EnableProcessing();
+
+  ASSERT_OK(db_->FlushWAL(true /* sync */));
+
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
+  ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearAllCallBacks();
+
+  ASSERT_OK(Put("key3", "val3"));
+
+  ASSERT_OK(db_->FlushWAL(true /* sync */));
+
+  Close();
+
+  // Simulate full loss of unsynced data. This should drop nothing since we did
+  // `FlushWAL(true /* sync */)` before `Close()`.
+  fault_env->DropUnsyncedFileData();
+
+  Reopen(options);
+
+  ASSERT_EQ("val1", Get("key1"));
+  ASSERT_EQ("val2", Get("key2"));
+  ASSERT_EQ("val3", Get("key3"));
+
+  // Need to close before `fault_env` goes out of scope.
+  Close();
+}
+
 TEST_P(DBWriteTest, IOErrorOnWALWriteTriggersReadOnlyMode) {
   std::unique_ptr<FaultInjectionTestEnv> mock_env(
       new FaultInjectionTestEnv(env_));
