Re: [PATCH] xen-netfront: drop skb when skb->len > 65535

[Wei Liu <wei.liu2@citrix.com>]

On Sat, Mar 02, 2013 at 02:54:17AM +0000, Ian Campbell wrote:
> On Fri, 2013-03-01 at 17:00 +0000, Wei Liu wrote:
> > On Fri, 2013-03-01 at 16:48 +0000, Jan Beulich wrote:
> > > >>> On 01.03.13 at 17:31, Wei Liu <wei.liu2@citrix.com> wrote:
> > > > The `size' field of Xen network wired format is uint16_t, anything bigger 
> > > > than
> > > > 65535 will cause overflow.
> > > > 
> > > > The punishment introduced by XSA-39 is quite harsh - DomU is disconnected when
> > > > it's discovered to be sending corrupted skbs. However, it looks like Linux
> > > > kernel will generate some bad skbs sometimes, so drop those skbs before
> > > > sending to over netback to avoid being disconnected.
> > > 
> > > While fixing the frontend is certainly desirable, we can't expect
> > > everyone to deploy fixed netfronts in all their VMs - some OS
> > > versions used in there may even be out of service. So we
> > > ought to find a way to also more gracefully deal with the
> > > situation in netback, without re-opening the security issue
> > > that prompted those changes.
> > > 
> > 
> > Regarding the punishment bit, I think its worth discussing it a bit.
> 
> Yes, the trick is figuring out what to do without reintroducing the
> softlockup which XSA-39 fixed.
> 
> Perhaps we should allow silently consume (and drop) oversize skbs and
> only shutdown the rings if they also consume too many (FSVO too many)
> slots?
> 
> > But the bug is always there, it drew no attention until revealed by
> > XSA-39. It ought to be fixed anyway. :-)
> 
> I would have sworn that skb->len was also limited to 64k, but looking at
> the header I see it is actually an int and the only limit of that sort
> is related to MAX_SKB_FRAGS (which doesn't actually limit the total
> size).

I had the impression that skb->len was limited to 64k, too. But it
turned out I was wrong.

> 
> OOI how big were the skbs you were seeing?

As Nick (npegg@linode.com) pointed out in his email, he saw size 65538.
I can reproduce this as well by setting vif's mtu to 100 then run iperf.
100 was just a random number I came up with when I played with
fragmentation.

> 
> Not that it really matters but do we have a handle on why the prexisting
> bug didn't already cause connectivity issues? Does the retransmit (which
> I suppose must be happening) somehow end up using a smaller skb size?
> 

Not sure. I didn't have enough time to look into this yesterday. :-(

> BTW you mean "wire protocol" not "wired protocol" in the comments etc.
> 

Yes.


Wei.

> Ian.
>