gro_cells: mark napi struct as not busy poll candidates

Rolf Neugebauer reported very long delays at netns dismantle.

Eric W. Biederman was kind enough to look at this problem
and noticed synchronize_net() occurring from netif_napi_del() that was
added in linux-4.5

Busy polling makes no sense for tunnels NAPI.
If busy poll is used for sessions over tunnels, the poller will need to
poll the physical device queue anyway.

netif_tx_napi_add() could be used here, but function name is misleading,
and renaming it is not stable material, so set NAPI_STATE_NO_BUSY_POLL
bit directly.

This will avoid inserting gro_cells napi structures in napi_hash[]
and avoid the problematic synchronize_net() (per possible cpu) that
Rolf reported.

Fixes: 93d05d4 ("net: provide generic busy polling to all NAPI drivers")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: Rolf Neugebauer <rolf.neugebauer@docker.com>
Reported-by: Eric W. Biederman <ebiederm@xmission.com>
Acked-by: Cong Wang <xiyou.wangcong@gmail.com>
Tested-by: Rolf Neugebauer <rolf.neugebauer@docker.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

include/net/gro_cells.h

@@ -68,6 +68,9 @@ static inline int gro_cells_init(struct gro_cells *gcells, struct net_device *de
 		struct gro_cell *cell = per_cpu_ptr(gcells->cells, i);
 
 		__skb_queue_head_init(&cell->napi_skbs);
+
+		set_bit(NAPI_STATE_NO_BUSY_POLL, &cell->napi.state);
+
 		netif_napi_add(dev, &cell->napi, gro_cell_poll, 64);
 		napi_enable(&cell->napi);
 	}