Fix FUSE_RELEASE protocol reply processing

This commit fixes the potential unexpected errors
of original handling of FUSE_RELEASE responses while
keep the same behavior (ignoring any reply).

pkg/sentry/fsimpl/fuse/dev.go

@@ -204,8 +204,11 @@ func (fd *DeviceFD) readLocked(ctx context.Context, dst usermem.IOSequence, opts
 
 	// Fully done with this req, remove it from the queue.
 	fd.queue.Remove(req)
-	if req.hdr.Opcode == linux.FUSE_RELEASE {
+
+	// Remove noReply ones from map of requests expecting a reply.
+	if req.noReply {
 		fd.numActiveRequests -= 1
+		delete(fd.completions, req.hdr.Unique)
 	}
 
 	return int64(n), nil
@@ -296,6 +299,10 @@ func (fd *DeviceFD) writeLocked(ctx context.Context, src usermem.IOSequence, opt
 
 			fut, ok := fd.completions[hdr.Unique]
 			if !ok {
+				if fut.hdr.Unique == linux.FUSE_RELEASE {
+					// Currently we simply discard the reply for FUSE_RELEASE.
+					return n + src.NumBytes(), nil
+				}
 				// Server sent us a response for a request we never sent?
 				return 0, syserror.EINVAL
 			}

pkg/sentry/fsimpl/fuse/file.go

@@ -84,7 +84,12 @@ func (fd *fileDescription) Release(ctx context.Context) {
 	}
 	kernelTask := kernel.TaskFromContext(ctx)
 	// ignoring errors and FUSE server reply is analogous to Linux's behavior.
-	req, _ := conn.NewRequest(auth.CredentialsFromContext(ctx), uint32(kernelTask.ThreadID()), fd.inode().NodeID, opcode, &in)
+	req, err := conn.NewRequest(auth.CredentialsFromContext(ctx), uint32(kernelTask.ThreadID()), fd.inode().NodeID, opcode, &in)
+	if err != nil {
+		// No way to invoke Call() with an errored request.
+		return
+	}
+	req.noReply = true
 	conn.CallAsync(kernelTask, req)
 }