Merge pull request #11817 from tangcong/fix-mvcc-deadlock-bug

mvcc: fix a deadlock bug in mvcc

CHANGELOG-3.5.md

@@ -105,6 +105,7 @@ Note that any `etcd_debugging_*` metrics are experimental and subject to change.
 - Improve [count-only range performance](https://github.com/etcd-io/etcd/pull/11771).
 - Remove [redundant storage restore operation to shorten the startup time](https://github.com/etcd-io/etcd/pull/11779).
   - With 40 million key test data,it can shorten the startup time from 5 min to 2.5 min.
+- [Fix deadlock bug in mvcc](https://github.com/etcd-io/etcd/pull/11817).
 
 ### Package `embed`
 

mvcc/kvstore.go

@@ -145,14 +145,18 @@ func NewStore(lg *zap.Logger, b backend.Backend, le lease.Lessor, ci cindex.Cons
 
 func (s *store) compactBarrier(ctx context.Context, ch chan struct{}) {
 	if ctx == nil || ctx.Err() != nil {
-		s.mu.Lock()
 		select {
 		case <-s.stopc:
 		default:
+			// fix deadlock in mvcc,for more information, please refer to pr 11817.
+			// s.stopc is only updated in restore operation, which is called by apply
+			// snapshot call, compaction and apply snapshot requests are serialized by
+			// raft, and do not happen at the same time.
+			s.mu.Lock()
 			f := func(ctx context.Context) { s.compactBarrier(ctx, ch) }
 			s.fifoSched.Schedule(f)
-		}
 			s.mu.Unlock()
+		}
 		return
 	}
 	close(ch)