nfp: flower: prevent offload if rhashtable insert fails

For flow offload adds, if the rhash insert code fails, the flow will still
have been offloaded but the reference to it in the driver freed.

Re-order the offload setup calls to ensure that a flow will only be written
to FW if a kernel reference is held and stored in the rhashtable. Remove
this hashtable entry if the offload fails.

Fixes: c01d0efa ("nfp: flower: use rhashtable for flow caching")
Signed-off-by: NJohn Hurley <john.hurley@netronome.com>
Reviewed-by: NPieter Jansen van Vuuren <pieter.jansenvanvuuren@netronome.com>
Reviewed-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

drivers/net/ethernet/netronome/nfp/flower/offload.c

@@ -476,17 +476,17 @@ nfp_flower_add_offload(struct nfp_app *app, struct net_device *netdev,
 	if (err)
 		goto err_destroy_flow;
 
-	err = nfp_flower_xmit_flow(netdev, flow_pay,
-				   NFP_FLOWER_CMSG_TYPE_FLOW_ADD);
-	if (err)
-		goto err_release_metadata;
-
 	flow_pay->tc_flower_cookie = flow->cookie;
 	err = rhashtable_insert_fast(&priv->flow_table, &flow_pay->fl_node,
 				     nfp_flower_table_params);
 	if (err)
 		goto err_release_metadata;
 
+	err = nfp_flower_xmit_flow(netdev, flow_pay,
+				   NFP_FLOWER_CMSG_TYPE_FLOW_ADD);
+	if (err)
+		goto err_remove_rhash;
+
 	port->tc_offload_cnt++;
 
 	/* Deallocate flow payload when flower rule has been destroyed. */
@@ -494,6 +494,10 @@ nfp_flower_add_offload(struct nfp_app *app, struct net_device *netdev,
 
 	return 0;
 
+err_remove_rhash:
+	WARN_ON_ONCE(rhashtable_remove_fast(&priv->flow_table,
+					    &flow_pay->fl_node,
+					    nfp_flower_table_params));
 err_release_metadata:
 	nfp_modify_flow_metadata(app, flow_pay);
 err_destroy_flow: