lan966x: execute xdp_do_flush() before napi_complete_done()

Make sure that xdp_do_flush() is always executed before
napi_complete_done(). This is important for two reasons. First, a
redirect to an XSKMAP assumes that a call to xdp_do_redirect() from
napi context X on CPU Y will be followed by a xdp_do_flush() from the
same napi context and CPU. This is not guaranteed if the
napi_complete_done() is executed before xdp_do_flush(), as it tells
the napi logic that it is fine to schedule napi context X on another
CPU. Details from a production system triggering this bug using the
veth driver can be found following the first link below.

The second reason is that the XDP_REDIRECT logic in itself relies on
being inside a single NAPI instance through to the xdp_do_flush() call
for RCU protection of all in-kernel data structures. Details can be
found in the second link below.

Fixes: a825b61 ("net: lan966x: Add support for XDP_REDIRECT")
Signed-off-by: Magnus Karlsson <[email protected]>
Acked-by: Toke Høiland-Jørgensen <[email protected]>
Acked-by: Steen Hegelund <[email protected]>
Link: https://lore.kernel.org/r/[email protected]
Link: https://lore.kernel.org/all/[email protected]/
Signed-off-by: Jakub Kicinski <[email protected]>

drivers/net/ethernet/microchip/lan966x/lan966x_fdma.c

@@ -608,12 +608,12 @@ static int lan966x_fdma_napi_poll(struct napi_struct *napi, int weight)
 		lan966x_fdma_rx_reload(rx);
 	}
 
-	if (counter < weight && napi_complete_done(napi, counter))
-		lan_wr(0xff, lan966x, FDMA_INTR_DB_ENA);
-
 	if (redirect)
 		xdp_do_flush();
 
+	if (counter < weight && napi_complete_done(napi, counter))
+		lan_wr(0xff, lan966x, FDMA_INTR_DB_ENA);
+
 	return counter;
 }