Re: less size_t please (was Re: [PATCH net] xfrm: fix integer overflow in xfrm_replay_state_esn_len())

[Dan Carpenter <dan.carpenter@linaro.org>]

On Thu, Feb 06, 2025 at 08:06:55PM +0300, Alexey Dobriyan wrote:
> On Thu, Jan 30, 2025 at 07:15:15PM +0300, Dan Carpenter wrote:
> > On Thu, Jan 30, 2025 at 04:44:42PM +0300, Alexey Dobriyan wrote:
> > > > -static inline unsigned int xfrm_replay_state_esn_len(struct xfrm_replay_state_esn *replay_esn)
> > > > +static inline size_t xfrm_replay_state_esn_len(struct xfrm_replay_state_esn *replay_esn)
> > > >  {
> > > > -	return sizeof(*replay_esn) + replay_esn->bmp_len * sizeof(__u32);
> > > > +	return size_add(sizeof(*replay_esn), size_mul(replay_esn->bmp_len, sizeof(__u32)));
> > > 
> > > Please don't do this.
> > > 
> > > You can (and should!) make calculations and check for overflow at the
> > > same time. It's very efficient.
> > > 
> > > > 1) Use size_add() and size_mul().  This change is necessary for 32bit systems.
> > > 
> > > This bloats code on 32-bit.
> > > 
> > 
> > I'm not sure I understand.  On 32-bit systems a size_t and an unsigned
> > int are the same size.  Did you mean to say 64-bit?
> 
> It looks like yes.
> 
> > Declaring sizes as u32 leads to integer overflows like this one.
> 
> No, the problem is unchecked C addition and mixing types which confuses
> people (in the opposite direction too -- there were fake CVEs because
> someone thought "size_t len" in write hooks could be big enough).
> 
> The answer is to use single type as much as possible and using checked
> additions on-the-go at every binary operator if possible.

In the write_hooks examples, we fixed those by moving to size_t.

64bit types are safer because 2**64 is a superset of 2**32.  Anything
which can overflow 64bits can overflow 32bits.  So obviously 64bits is
safer.

But it's surprising the extent of it.  We avoid using ulong types in
UAPI because it's a headache for 32bit support.  So normally we get
an u32 number_items from the user.  That's 4 billion.  It's a small
number and it's actually pretty hard for it to lead to an integer
overflow on 64bit systems.  The struct_size() function is basically
not needed if you're on 64bit and you declare your length variables as
size_t.

The rest of the kernel has an assumption that sizes are saved in size_t.
The size_add() and struct_size() macros rely on it.  In networking there
are a number of functions like sock_kmalloc() which truncate the size
parameter to int and they just make me itch to look at.

regards,
dan carpenter