Re: CVE-2021-3600 aka bpf: Fix 32 bit src register truncation on div/mod

[Thadeu Lima de Souza Cascardo]

On Thu, Aug 26, 2021 at 12:23:37PM +0200, Pavel Machek wrote:
> Hi!
> 
> As far as I can tell, CVE-2021-3600 is still a problem for 4.14 and
> 4.19.
> 
> Unfortunately, those kernels lack BPF_JMP32 support, and that support
> is too big and intrusive to backport.
> 
> So I tried to come up with solution without JMP32 support... only to
> end up with not seeing the bug in the affected code.
> 
> Changelog says:
> 
>     bpf: Fix 32 bit src register truncation on div/mod
>     
>     While reviewing a different fix, John and I noticed an oddity in one of the
>     BPF program dumps that stood out, for example:
>     
>       # bpftool p d x i 13
>        0: (b7) r0 = 808464450
>        1: (b4) w4 = 808464432
>        2: (bc) w0 = w0
>        3: (15) if r0 == 0x0 goto pc+1
>        4: (9c) w4 %= w0
>       [...]
>     
>     In line 2 we noticed that the mov32 would 32 bit truncate the original src
>     register for the div/mod operation. While for the two operations the dst
>     register is typically marked unknown e.g. from adjust_scalar_min_max_vals()
>     the src register is not, and thus verifier keeps tracking original bounds,
>     simplified:
> 
> So this explains "mov32 w0, w0" is problematic, and fixes the bug by
> replacing it with jmp32. Unfortunately, I can't do that in 4.19; plus
> I don't really see how the bug is solved -- we avoided adding mov32
> sequence that triggers the problem, but the problematic sequence could
> still be produced by the userspace, no?
> 
> Does adjust_scalar_min_max_vals still need fixing?
> 
> Do you have any hints how to solve this in 4.19?
> 
> Best regards,
> 								Pavel
> -- 
> DENX Software Engineering GmbH,      Managing Director: Wolfgang Denk
> HRB 165235 Munich, Office: Kirchenstr.5, D-82194 Groebenzell, Germany

Hi, Pavel.

I have just sent the fixes for 4.14. I sent fixes for 4.19 last Friday.

The problem here is that the verifier assumes the source register has a given
value, but the fixups change that value to something else when it is truncated.

The fixups run after the verifier, so a similar sequence produced by userspace
will be correctly verified, so no fixes are necessary on adjust_scalar_min_max
for this specific issue. The fixed-up code is not verified again.

The challenge in providing those fixes to 4.14 and 4.19 is the absence of JMP32
in those kernels. So, AX was taken as a temporary, so it would still work on
JITs.

Cascardo.