Fix an assertion failure in error handler (#13251)

Summary:
we saw this [assertion](https://github.com/facebook/rocksdb/blob/02b4197544f758bdf84d80fe9319238611848c48/db/error_handler.cc#L576) failing in crash test. The LOG shows that there's a call to SetOptions() concurrent to ResumeImpl(). It's possible that while waiting for error recovery flush (with mutex released), SetOptions() failed to write to MANIFEST and added a file to be quarantined. This triggered the assertion failure when ResumeImpl() calls ClearBGError().

This PR fixes the issue by setting background error when SetOptions() fails to write to MANIFEST.

Pull Request resolved: #13251

Test Plan: monitor future crash test failures.

Reviewed By: hx235

Differential Revision: D67660106

Pulled By: cbi42

fbshipit-source-id: 1b52bb23005c4b544f8f9bceefd3b9dcbaf0edfa

db/db_impl/db_impl.cc

@@ -1299,11 +1299,15 @@ Status DBImpl::SetOptions(
       VersionEdit dummy_edit;
       s = versions_->LogAndApply(cfd, new_options, read_options, write_options,
                                  &dummy_edit, &mutex_, directories_.GetDbDir());
+      if (!versions_->io_status().ok()) {
+        assert(!s.ok());
+        error_handler_.SetBGError(versions_->io_status(),
+                                  BackgroundErrorReason::kManifestWrite);
+      }
       // Trigger possible flush/compactions. This has to be before we persist
       // options to file, otherwise there will be a deadlock with writer
       // thread.
       InstallSuperVersionAndScheduleWork(cfd, &sv_context, new_options);
-
       persist_options_status =
           WriteOptionsFile(write_options, true /*db_mutex_already_held*/);
       bg_cv_.SignalAll();

db/error_handler.cc

@@ -573,6 +573,8 @@ Status ErrorHandler::ClearBGError() {
 
   // Signal that recovery succeeded
   if (recovery_error_.ok()) {
+    // If this assertion fails, it means likely bg error is not set after a
+    // file is quarantined during MANIFEST write.
     assert(files_to_quarantine_.empty());
     Status old_bg_error = bg_error_;
     // old_bg_error is only for notifying listeners, so may not be checked

db/version_set.cc

@@ -5939,6 +5939,8 @@ Status VersionSet::LogAndApply(
     }
     TEST_SYNC_POINT_CALLBACK("VersionSet::LogAndApply:WakeUpAndDone", mu);
 #endif /* !NDEBUG */
+    // FIXME: One MANIFEST write failure can cause all writes to SetBGError,
+    // should only SetBGError once.
     return first_writer.status;
   }