Re: [PATCH] net: check the length of the data before dereferencing it

[Eric Dumazet <eric.dumazet@gmail.com>]

On Sun, 2012-04-01 at 23:29 -0400, David Miller wrote:
> From: Eric Dumazet <eric.dumazet@gmail.com>
> Date: Mon, 02 Apr 2012 05:19:33 +0200
> 
> >> @@ -3845,6 +3845,8 @@ void tcp_parse_options(const struct sk_buff *skb, struct tcp_options_received *o
> >>  			length--;
> >>  			continue;
> >>  		default:
> >> +			if (length < 2)
> >> +				return;
> >>  			opsize = *ptr++;
> >>  			if (opsize < 2) /* "silly options" */
> >>  				return;
> > 
> > Acked-by: Eric Dumazet <eric.dumazet@gmail.com>
> 
> Tag Eric, you're it.
> 
> You ACK'd this patch, so you get to show how this is actually able
> to cause some kind of problem.
> 
> I assert that this is adding a useless test, that doesn't fix any kind
> of possible crash or misbehavior.  If length == 1 at the default:, the
> code will absolutely do the right thing.
> 
> Prove me wrong.

No problem.

You can have NOP,NOP,NOP,EVIL-OPTION

initial length=4  (multiple of 4)

We can read 5 bytes, and access 'out of bound' memory.

Usually not a problem since we have many bytes after our head.