queueing: remove per-peer queuesjd/remove-per-peer-queues

Previously, having many peers would result in many napi structs, which
could make lookups in the napi_hash in net/core/dev.c slow. So, we move
to using a single napi struct per device.

The best solution would be to replace napi_hash with an idr or just get
rid of it all together and use straight pointers. However, that isn't the
case currently, so we work with what is and begrudgingly remove per-peer
queues. On the upside, it means we reduce the per-peer memory usage by
about 8k/16k, but on the downside it means that napi_gro_receive is
called on a unified list, which might result in less GRO speedups on
systems with many peers active at once.

However, if napi_hash does ever go away, we should consider reverting
this commit.

Since this means moving to unified packet queues, flushing at peer
removal is something of a problem. So we make the slightly dubious
modification of just not flushing, and letting our reference counters do
the work. This in turn required some small changes to ensure that the
reference counter will, at some point in the future, still reach zero,
and not be kept alive by non-stop packet ingress.

Co-developed-by: Jonathan Neuschäfer <j.neuschaefer@gmx.net>

1 files changed, 2 insertions, 2 deletions

diff --git a/src/queueing.h b/src/queueing.h
index c17b0d8..2869675 100644
--- a/src/queueing.h
+++ b/src/queueing.h
@@ -139,12 +139,12 @@ static inline void queue_enqueue_per_peer(struct crypt_queue *queue, struct sk_b
 	peer_put(peer);
 }
 
-static inline void queue_enqueue_per_peer_napi(struct crypt_queue *queue, struct sk_buff *skb, enum packet_state state)
+static inline void queue_enqueue_per_device_napi(struct crypt_queue *queue, struct sk_buff *skb, enum packet_state state)
 {
 	struct wireguard_peer *peer = peer_rcu_get(PACKET_PEER(skb));
 
 	atomic_set(&PACKET_CB(skb)->state, state);
-	napi_schedule(&peer->napi);
+	napi_schedule(&peer->device->napi);
 	peer_put(peer);
 }