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

[Hannes Frederic Sowa <hannes@stressinduktion.org>]

On Tue, Sep 03, 2013 at 11:19:14AM -0600, Stephen Warren wrote:
> On 09/03/2013 05:37 AM, Thomas Graf wrote:
> > Allocating skbs when sending out neighbour discovery messages
> > currently uses sock_alloc_send_skb() based on a per net namespace
> > socket and thus share a socket wmem buffer space.
> > 
> > If a netdevice is temporarily unable to transmit due to carrier
> > loss or for other reasons, the queued up ndisc messages will cosnume
> > all of the wmem space and will thus prevent from any more skbs to
> > be allocated even for netdevices that are able to transmit packets.
> > 
> > The number of neighbour discovery messages sent is very limited,
> > use of alloc_skb() bypasses the socket wmem buffer size enforcement
> > while the manual call to skb_set_owner_w() maintains the socket
> > reference needed for the IPv6 output path.
> > 
> > This patch has orginally been posted by Eric Dumazet in a modified
> > form.
> 
> Tested-by: Stephen Warren <swarren@nvidia.com>
> 
> Although I do note something slightly odd:
> 
> next-20130830 had an issue, and reverting V1 of this patch solved it.
> 
> However, in next-20130903, if I revert the revert of V1 of this patch, I
> don't see any issue; it appears that the problem was some interaction
> between V1 of this patch and something else in next-20130830.
> 
> Either way, this patch doesn't seem to introduce any issue when applied
> on top of either next-20130830 with V1 reverted, or on top of
> next-20130903, so it's fine.

Could either of you run the v1 version of the patch with CONFIG_PROVE_LOCKING
enabled? I also do think there is some side-effect we don't understand yet.

Thanks,

  Hannes