Re: [PATCH 4/10] Fix leaking of kernel heap addresses in net/

[Stephen Hemminger <shemminger@vyatta.com>]

On Fri, 12 Nov 2010 08:23:30 +0100
Eric Dumazet <eric.dumazet@gmail.com> wrote:

> Le jeudi 11 novembre 2010 à 21:34 -0500, Dan Rosenberg a écrit :
> > > I want whatever you replace it with to be equivalent for
> > > object tracking purposes.
> > 
> > In nearly all of the cases I fixed, the socket inode is already
> > provided, which serves as a perfectly good unique identifier.  Would you
> > prefer I include that information twice?
> 
> Oh well. Please read this answer carefuly.
> 
> Some facts to feed your next patch. I am very pleased you changed your
> mind and that we keep useful information in kernel log.
> 
> 1) Inode numbers are not guaranteed to be unique. Its a 32bit seq
> number, and we dont check another socket inode use the same inode number
> (after 2^32 allocations it can happens)
> 
> 2) /proc/net/ files can deliver same "line" of information several
> times, because of their implementation.
> 
> 3) Because of SLAB_DESTROY_BY_RCU, same 'kernel socket pointer' can be
> seen several times in /proc/net/tcp & /proc/net/udp, but really on
> different "sockets"
> 
> 4) Some good applications use both the socket pointer and inode number
> (tuple) to filter out the [2] problem. Dont break them, please ?
> Anything that might break an application must be at the very least
> tunable.
> 
> In my opinion, a good thing would be :
> 
> - Use a special printf() format , aka "secure pointer", as Thomas
> suggested.
> 
> - Make sure you print different opaque values for two different kernel
> pointers. This is mandatory.
> 
> - Make sure the NULL pointer stay as a NULL pointer to not let the
> hostile user know your secret, and to ease debugging stuff.
> 
> - Have security experts advice to chose a nice crypto function, maybe
> jenkin hash. Not too slow would be nice.
> 
> 
> static unsigned long securize_kpointers_rnd;
> 
> At boot time, stick a random value in this variable.
> (Maybe make sure the 5 low order bits are 0)
> 
> unsigned long opacify_kptr(unsigned long ptr)
> {
> 	if (ptr == 0)
> 		return ptr;
> 	if (capable(CAP_NET_ADMIN))
> 		return ptr;
> 
> 	return some_crypto_hash(ptr, &securize_kpointers_rnd);
> }
> 
> At least, use a central point, so that we can easily add/change the
> logic if needed.
> 
> Please provide this patch in kernel/printk.c for initial review, then
> once everybody is OK, you can send one patch for net tree.
> 
> No need to send 10 patches if we dont agree on the general principle.

Also, the whole idea needs to be under a config option, so only
the paranoid idiots turn it on.


--