Lucid CVE-2012-3412

Ben Hutchings bhutchings at
Fri Sep 21 03:55:11 UTC 2012

On Fri, 2012-08-24 at 13:38 -0600, Tim Gardner wrote:
> On 08/24/2012 09:11 AM, Tim Gardner wrote:
> > On 08/24/2012 09:05 AM, Herton Ronaldo Krzesinski wrote:
> >> On Fri, Aug 24, 2012 at 07:58:34AM -0600, Tim Gardner wrote:
> >>>  static inline int netif_needs_gso(struct net_device *dev, struct sk_buff *skb)
> >>>  {
> >>> +	if (skb_is_gso(skb) &&
> >>> +		skb_shinfo(skb)->gso_segs > skb->dev->gso_max_segs)
> >>> +		return 0;
> >>
> >> Shouldn't be return 1 here? If the condition is true, we would clear the
> >> flags from features. If flags are cleared, when calling skb_gso_ok:
> >>
> >> net_gso_ok would always return 0
> >> skb_gso_ok would always return 0
> >> netif_needs_gso returns 1 because it does !skb_gso_ok
> >>
> >> Unless I'm missing something here. Anyway, hard to read these functions...
> >> I think just copying/clearing the flags and passing through skb_gso_ok
> >> would be better.
> >>
> > 
> > I guess I'm confused about when the flag is set. I was assuming if GSO
> > was set, then the driver handled 'generic segmentation offload'. Aren't
> > we checking for error conditions, e.g., if there are more segments then
> > the driver can handle, then _don't_ do GSO ?
> > 
> > rtg
> > 
> After some online discussion with Herton I think we've agreed to bag
> this mess. This is not a broad vulnerability, and the patch is proving
> to be more effort then its worth.

If you'd just asked, I could have given you the driver-only fix which
was applied to the out-of-tree version of sfc.  This would avoid the
need to backport any net core or TCP changes, which is comparatively
risky for older kernel versions.  (David Miller has covered 3.x now.)

The driver-only fix is to truncate TSO skbs to the maximum number of
segments (100).  This results in terrible throughput for any TCP stream
that exceeds that, but the limit is chosen to be high enough that
legitimate traffic should never hit it.  This is also the fix that RHEL
will use, due to its ABI stability requirements.

In the driver version in Debian squeeze and Ubuntu lucid (roughly
equivalent to Linux 2.6.33), the DMA ring sizes were constant, so the
fix can be simplified further:


Ben Hutchings, Staff Engineer, Solarflare
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.

More information about the kernel-team mailing list