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

[Alexey Dobriyan <adobriyan@gmail.com>]

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.

Of course one bug could be fixed in multiple ways.

> If you look at integer overflows with security implications there is a
> 5 to 1 ratio of bugs that only affect 32-bit vs bugs that affect
> everything because it's just so much easier to overflow a 32-bit size.
> 
> aab98e2dbd64 ("ksmbd: fix integer overflows on 32 bit systems")
> 16ebb6f5b629 ("nfp: bpf: prevent integer overflow in nfp_bpf_event_output()")
> 09c4a6101532 ("rtc: tps6594: Fix integer overflow on 32bit systems")
> 55cf2f4b945f ("binfmt_flat: Fix integer overflow bug on 32 bit systems")
> fbbd84af6ba7 ("chelsio/chtls: prevent potential integer overflow on 32bit")
> bd96a3935e89 ("rdma/cxgb4: Prevent potential integer overflow on 32bit")
> d0257e089d1b ("RDMA/uverbs: Prevent integer overflow issue")

This one is good demonstration why BAO is better:
https://godbolt.org/z/14ofdfvhc

> 3c63d8946e57 ("svcrdma: Address an integer overflow")
> 7f33b92e5b18 ("NFSD: Prevent a potential integer overflow")
> 
> > 	int len;
> > 	if (__builtin_mul_overflow(replay_esn->bmp_len, 4, &len)) {
> > 		return true;
> > 	}
> > 	if (__builtin_add_overflow(len, sizeof(*replay_esn), &len)) {
> > 		return true;
> > 	}
> 
> This is so ugly...  :/  I'd prefer to just do open code the check at
> that point.
> 
> static inline int xfrm_replay_state_esn_len(struct xfrm_replay_state_esn *replay_esn)
> {
> 	if (replay_esn->bmp_len > (INT_MAX - sizeof(*replay_esn)) / sizeof(__u32))
> 		return -EINVAL;
> 	return sizeof(*replay_esn) + replay_esn->bmp_len * sizeof(__u32);
> }

You can't open code if you have something like this:

	X = a * b + c;

Second, the code is now effectively duplicated, once in overflow check,
second time in actual calculation.

BAO and BMO may look chatty but they're doing the right thing.