Don't recache req->vcl in wrk unless it will speed things up later.

bin/varnishd/cache/cache_req.c

@@ -217,12 +217,8 @@ Req_Cleanup(struct sess *sp, struct worker *wrk, struct req *req)
 	AZ(req->esi_level);
 	AZ(req->privs->magic);
 
-	if (req->vcl != NULL) {
-		if (wrk->vcl != NULL)
-			VCL_Rel(&wrk->vcl);
-		wrk->vcl = req->vcl;
-		req->vcl = NULL;
-	}
+	if (req->vcl != NULL)
+		VCL_Recache(wrk, &req->vcl);
 
 	/* Charge and log byte counters */
 	if (req->vsl->wid) {

bin/varnishd/cache/cache_varnishd.h

@@ -382,6 +382,7 @@ void VCL_Panic(struct vsb *, const struct vcl *);
 void VCL_Poll(void);
 void VCL_Ref(struct vcl *);
 void VCL_Refresh(struct vcl **);
+void VCL_Recache(struct worker *, struct vcl **);
 void VCL_Rel(struct vcl **);
 void VCL_TaskEnter(const struct vcl *, struct vrt_privs *);
 void VCL_TaskLeave(const struct vcl *, struct vrt_privs *);

bin/varnishd/cache/cache_vrt_vcl.c

@@ -76,17 +76,36 @@ VCL_Method_Name(unsigned m)
 void
 VCL_Refresh(struct vcl **vcc)
 {
+
+	while (vcl_active == NULL)
+		(void)usleep(100000);
+
 	if (*vcc == vcl_active)
 		return;
 	if (*vcc != NULL)
 		VCL_Rel(vcc);	/* XXX: optimize locking */
 
-	while (vcl_active == NULL)
-		(void)usleep(100000);
-
 	vcl_get(vcc, NULL);
 }
 
+void
+VCL_Recache(struct worker *wrk, struct vcl **vclp)
+{
+
+	AN(wrk);
+	AN(vclp);
+	CHECK_OBJ_NOTNULL(*vclp, VCL_MAGIC);
+
+	if (*vclp != vcl_active || wrk->vcl == vcl_active) {
+		VCL_Rel(vclp);
+		return;
+	}
+	if (wrk->vcl != NULL)
+		VCL_Rel(&wrk->vcl);
+	wrk->vcl = *vclp;
+	*vclp = NULL;
+}
+
 void
 VCL_Ref(struct vcl *vcl)
 {