Fix close_timeout for new started harvesters (elastic#3715)

If close_timeout is set, a file handler inside the harvester is closed after `close_timeout` is reached even if the output is blocked. For new which were found for harvesting during the output was blocked, it was possible that the Setup of the file happened but then the initial state could not be sent. So the harvester was never started, file handler was open but `close_timeout` this not apply. Setup and state update are now turned around. This has the affect, that in case of an error during the setup phase, the state must be revert again.

See elastic#3091 (comment)
(cherry picked from commit e13a7f2)

filebeat/prospector/prospector.go

@@ -299,16 +299,24 @@ func (p *Prospector) startHarvester(state file.State, offset int64) error {
 		return err
 	}
 
-	reader, err := h.Setup()
+	// State is directly updated and not through channel to make state update synchronous
+	err = p.updateState(input.NewEvent(state))
 	if err != nil {
-		return fmt.Errorf("Error setting up harvester: %s", err)
+		return err
 	}
 
-	// State is directly updated and not through channel to make state update immediate
-	// State is only updated after setup is completed successfully
-	err = p.updateState(input.NewEvent(state))
+	reader, err := h.Setup()
 	if err != nil {
-		return err
+		// Set state to finished True again in case of setup failure to make sure
+		// file can be picked up again by a future harvester
+		state.Finished = true
+
+		updateErr := p.updateState(input.NewEvent(state))
+		// This should only happen in the case that filebeat is stopped
+		if updateErr != nil {
+			logp.Err("Error updating state: %v", updateErr)
+		}
+		return fmt.Errorf("Error setting up harvester: %s", err)
 	}
 
 	p.registry.start(h, reader)