Re: [PATCH net v2] wireguard: wait for per-peer crypto during removal

[Chris Arges <carges@cloudflare.com>]

On 2026-09-16 12:06:08, Jason A. Donenfeld wrote:
> On Sun, Sep 13, 2026 at 07:19:08AM -0500, Chris J Arges wrote:
> > Calling peer_remove_after_dead() currently flushes device-wide packet
> > crypto and handshake workqueues while holding RTNL. This is problematic as
> > unrelated peers can continue adding work to those queues, blocking other
> > tasks that want to take the RTNL lock.
> > 
> > Instead, this patch tracks pending crypto handoffs for each peer using a
> > counter. After marking the peer dead, synchronize_net() prevents new
> > submissions. Next, wait for pending crypto workers to schedule TX work and
> > for RX NAPI to drain the peer's RX queue. Then flush only the peer's TX
> > packet and handshake work.
> > 
> > This scopes teardown synchronization to the removed peer and prevents
> > unrelated peers from extending the RTNL hold time.
> >  struct wg_device;
> > @@ -161,6 +162,7 @@ static inline int wg_queue_enqueue_per_device_and_peer(
> >  	 */
> >  	if (unlikely(!wg_prev_queue_enqueue(peer_queue, skb)))
> >  		return -ENOSPC;
> > +	atomic_inc(&PACKET_PEER(skb)->packet_crypt_pending);
> >  
> >  	/* Then we queue it up in the device queue, which consumes the
> >  	 * packet as soon as it can.
> > @@ -182,6 +184,8 @@ static inline void wg_queue_enqueue_per_peer_tx(struct sk_buff *skb, enum packet
> >  	atomic_set_release(&PACKET_CB(skb)->state, state);
> >  	queue_work_on(wg_cpumask_choose_online(&peer->serial_work_cpu, peer->internal_id),
> >  		      peer->device->packet_crypt_wq, &peer->transmit_packet_work);
> > +	if (atomic_dec_and_test(&peer->packet_crypt_pending))
> > +		wake_up_var(&peer->packet_crypt_pending);
> >  	wg_peer_put(peer);
> >  }
> >  
> > diff --git a/drivers/net/wireguard/receive.c b/drivers/net/wireguard/receive.c
> > index 824bbefce61c..bb35e3205491 100644
> > --- a/drivers/net/wireguard/receive.c
> > +++ b/drivers/net/wireguard/receive.c
> > @@ -476,9 +476,11 @@ int wg_packet_rx_poll(struct napi_struct *napi, int budget)
> >  
> >  next:
> >  		wg_noise_keypair_put(keypair, false);
> > -		wg_peer_put(peer);
> >  		if (unlikely(free))
> >  			dev_kfree_skb(skb);
> > +		if (atomic_dec_and_test(&peer->packet_crypt_pending))
> > +			wake_up_var(&peer->packet_crypt_pending);
> > +		wg_peer_put(peer);
> 
> This adds two atomic updates to a per-peer counter for every RX packet
> and TX batch. That could introduce contention across crypto workers. I
> suppose it'd be good to see some measurements in if this changes
> anything. Certainly it should change _something_. Question is by how
> much.
> 
> Jason

Jason,

I benchmarked a wg peer on a 2 vCPU and 8 vCPU setup with 1200 byte
UDP payloads. Goal was stressing the path where these atomics are getting
incremented.
- patched 2 vCPU test showed ~1-2% reduction in throughput
- patched 8 vCPU test showed ~6% reduction in throughput (more contention)

My main goal is reducing the amount of time holding rtnl_lock when we remove a
peer. In our systems we frequently get hangs due to workloads bringing up and
tearing down wg peers. So perhaps I'll need to look into another approach
where we're not introducing something like a counter into the hotpath.

Thanks,
--chris