Re: [PATCH v2] ipv6: Don't depend on per socket memory for neighbour discovery messages

[Thomas Graf <tgraf@suug.ch>]

On 09/03/13 at 01:56pm, Hannes Frederic Sowa wrote:
> Hi!
> 
> On Tue, Sep 03, 2013 at 01:37:01PM +0200, Thomas Graf wrote:
> > @@ -389,6 +386,11 @@ static struct sk_buff *ndisc_alloc_skb(struct net_device *dev,
> >  	skb_reserve(skb, hlen + sizeof(struct ipv6hdr));
> >  	skb_reset_transport_header(skb);
> >  
> > +	/* Manually assign socket ownership as we avoid calling
> > +	 * sock_alloc_send_pskb() to bypass wmem buffer limits
> > +	 */
> > +	skb_set_owner_w(skb, sk);
> > +
> >  	return skb;
> >  }
> 
> Do you know why this is needed? From the report it seemed to me that we might
> have a deadlock on idev->lock and I couldn't find the culprit.
> 
> When I tested your change on x86_64 I did not experience this.

I also didn't see any problems running v1 of the patch which
is confusing as the IPv6 output path assumes a socket reference
in various places as Dave pointed out correctly.

I don't see why the problem would be limited to ARM.

> Maybe someone with arm could try this patch with CONFIG_PROVE_LOCKING?

I would certainly welcome that.