Re: [PATCH net] macsec: Fix use-after-free while sending the offloading packet

[Sabrina Dubroca <sd@queasysnail.net>]

2024-10-15, 21:46:26 +0800, Jianbo Liu wrote:
> 
> 
> On 10/15/2024 6:28 PM, Sabrina Dubroca wrote:
> > 2024-10-15, 17:57:59 +0800, Jianbo Liu wrote:
> > > 
> > > 
> > > On 10/15/2024 4:56 PM, Sabrina Dubroca wrote:
> > > > 2024-10-14, 12:07:20 +0300, Tariq Toukan wrote:
> > > > > From: Jianbo Liu <jianbol@nvidia.com>
> > > > > 
> > > > > KASAN reports the following UAF. The metadata_dst, which is used to
> > > > > store the SCI value for macsec offload, is already freed by
> > > > > metadata_dst_free() in macsec_free_netdev(), while driver still use it
> > > > > for sending the packet.
> > > > > 
> > > > > To fix this issue, dst_release() is used instead to release
> > > > > metadata_dst. So it is not freed instantly in macsec_free_netdev() if
> > > > > still referenced by skb.
> > > > 
> > > > Ok. Then that packet is going to get dropped when it reaches the
> > > > driver, right? At this point the TXSA we need shouldn't be configured
> > > 
> > > I think so because dst's output should be updated.
> > 
> > What updates the dst when we're deleting the macsec device? And this
> > is just a metadata_dst, it's only useful to hold the SCI.
> > 
> 
> You are right. It's not updated.
> 
> > But I guess we would have the same issue when the macsec device still
> > exists but the TXSA is gone, so hopefully this is handled well by all
> > drivers.
> > 
> 
> And for now, I'd rather focus on fixing the kernel crash caused by UAF.

Ok, but please take a look at it very soon. Sending packets
unencrypted when they should be encrypted can be just as bad as
crashing the system.

-- 
Sabrina