Re: [PATCH net] af_packet: move notifier's packet_dev_mc out of rcu critical section

[Willem de Bruijn <willemdebruijn.kernel@gmail.com>]

Jakub Kicinski wrote:
> On Tue, 20 May 2025 22:41:30 -0400 Willem de Bruijn wrote:
> > > @@ -4277,6 +4280,13 @@ static int packet_notifier(struct notifier_block *this,
> > >  		}
> > >  	}
> > >  	rcu_read_unlock();
> > > +
> > > +	/* packet_dev_mc might grab instance locks so can't run under rcu */
> > > +	list_for_each_entry_safe(ml, tmp, &mclist, remove_list) {
> > > +		packet_dev_mc(dev, ml, -1);
> > > +		kfree(ml);
> > > +	}
> > > +  
> > 
> > Just verifying my understanding of the not entirely obvious locking:
> > 
> > po->mclist modifications (add, del, flush, unregister) are all
> > protected by the RTNL, not the RCU. The RCU only protects the sklist
> > and by extension the sks on it. So moving the mclist operations out of
> > the RCU is fine.
> > 
> > The delayed operation on the mclist entry is still within the RTNL
> > from unregister_netdevice_notifier. Which matter as it protects not
> > only the list, but also the actual operations in packet_dev_mc, such
> > as inc/dec on dev->promiscuity and associated dev_change_rx_flags.
> > And new packet_mclist.remove_list too.
> 
> Matches my understanding FWIW, but this will be a great addition 
> to the commit message. Let's add it in v2..
> 
> > >  	return NOTIFY_DONE;
> > >  }
> > >  
> > > diff --git a/net/packet/internal.h b/net/packet/internal.h
> > > index d5d70712007a..1e743d0316fd 100644
> > > --- a/net/packet/internal.h
> > > +++ b/net/packet/internal.h
> > > @@ -11,6 +11,7 @@ struct packet_mclist {
> > >  	unsigned short		type;
> > >  	unsigned short		alen;
> > >  	unsigned char		addr[MAX_ADDR_LEN];
> > > +	struct list_head	remove_list;  
> > 
> > INIT_LIST_HEAD on alloc in packet_mc_add?
> 
> Just to be clear this is an "entry node" not a "head node",
> is it common to init "entry nodes"? 

I wasn't sure. A small sample from net/core showed that many do, e.g.,
napi->poll_list. But not all, e.g., failover->list just calls
list_add_tail immediately.

I suspect, and from that it seems, that it is safe to not explicitly
initalize entry nodes if you know what you're doing / how they're
used.

But whether that is actually intended to work, especially with more
involved debugging (such as LIST_POISON) and invariant checking
(__list_add_valid), I don't know.

I did not find any authoritative documentation that says you have too,
so I guess it's fine. But not ideal.