Re: [PATCH for-4.18] x86/time: Fix UBSAN failure in __update_vcpu_system_time()

[Andrew Cooper <andrew.cooper3@xxxxxxxxxx>]

On 02/11/2023 8:33 am, Jan Beulich wrote:
> On 01.11.2023 21:37, Andrew Cooper wrote:
>> As reported:
>>
>>   (XEN) 
>> ================================================================================
>>   (XEN) UBSAN: Undefined behaviour in arch/x86/time.c:1542:32
>>   (XEN) member access within null pointer of type 'union vcpu_info_t'
>>   (XEN) ----[ Xen-4.19-unstable  x86_64  debug=y ubsan=y  Not tainted ]----
>>   ...
>>   (XEN) Xen call trace:
>>   (XEN)    [<ffff82d040345036>] R 
>> common/ubsan/ubsan.c#ubsan_epilogue+0xa/0xd2
>>   (XEN)    [<ffff82d0403456e8>] F __ubsan_handle_type_mismatch+0x133/0x49b
>>   (XEN)    [<ffff82d040345b4a>] F __ubsan_handle_type_mismatch_v1+0xfa/0xfc
>>   (XEN)    [<ffff82d040623356>] F 
>> arch/x86/time.c#__update_vcpu_system_time+0x212/0x30f
>>   (XEN)    [<ffff82d040623461>] F update_vcpu_system_time+0xe/0x10
>>   (XEN)    [<ffff82d04062389d>] F 
>> arch/x86/time.c#local_time_calibration+0x1f7/0x523
>>   (XEN)    [<ffff82d0402a64b5>] F common/softirq.c#__do_softirq+0x1f4/0x31a
>>   (XEN)    [<ffff82d0402a67ad>] F do_softirq+0x13/0x15
>>   (XEN)    [<ffff82d0405a95dc>] F arch/x86/domain.c#idle_loop+0x2e0/0x367
>>   (XEN)
>>   (XEN) 
>> ================================================================================
>>
>> It is not valid to derive a pointer from vcpu_info() prior to checking that
>> the underlying map pointer is good.
>>
>> Reorder actions so the NULL pointer check is first.
>>
>> Fixes: 20279afd7323 ("x86: split populating of struct vcpu_time_info into a 
>> separate function")
>> Signed-off-by: Andrew Cooper <andrew.cooper3@xxxxxxxxxx>
> In the interest of silencing the checker
> Acked-by: Jan Beulich <jbeulich@xxxxxxxx>

Thanks.

> However, ...
>
>> 4.18 blocker, or we'll need to issue an XSA/CVE.
> ... I dare to disagree (or to at least be uncertain) here, and I further
> dare to question correctness of the checker (which suggests that it might
> be helpful to point out in the description which version of the compiler
> it was). Quoting part of a footnote of the respective part of the C99
> spec, describing the unary * operator:
>
> "Thus, &*E is equivalent to E (even if E is a null pointer), ..."
>
> This imo leaves it at best ambiguous whether there's any actual UB here.

I'd argue that this isn't relevant.

What matters how the compiler interprets the spec, and in this case GCC
is telling us explicitly that it does consider this to be UB.

Whether there is a practical consequence is a different question, but ...

> Furthermore, if you deem this XSA-worthy despite the generated code not
> resulting in any real misbehavior

... we've issued XSAs for this class of issue before.  XSA-328 is the
one I specifically remember, but I'm sure we've done others too.

In this case, an unprivileged guest can control the NULL-ness here, so
if there's a practical consequence from the compiler then the guest can
definitely tickle that consequence.

Alternatively, the security team could decide to change it's stance on
this class of issues.

> , code patterns like
> (found in free_heap_pages())
>
>             struct page_info *predecessor = pg - mask;
>
>             /* Merge with predecessor block? */
>             if ( !mfn_valid(page_to_mfn(predecessor)) ||
>
> or (found in get_page_from_l1e())
>
>     struct page_info *page = mfn_to_page(_mfn(mfn));
>     ...
>     valid = mfn_valid(_mfn(mfn));
>
>     if ( !valid ||
>
> would be in the same class (access outside of array bounds), just that the
> checker cannot spot those without producing false positives (simply because
> it doesn't know frame_table[]'s bounds). We're fully aware of the existence
> of such code, and we've decided to - if at all - only eliminate such cases
> (slowly) as respective code is touched anyway.

I don't agree with describing these as the same class.  NULL deference
is different to OoB, even if they overlap from an underlying mechanics
point of view.

Nevertheless, I've raised that "valid" pattern with the security team
before, and I would certainly prefer to express it differently.

But neither of them trigger UBSAN.  GCC can't see any wiggle room to
potentially optimise, and I expect it's because __mfn_valid() is in an
external call.

If we had working LTO, I'd be interested to see that alters the UBSAN
determination or not.

>
> Jan
>
> PS: argo.c:find_ring_mfn() has
>
>     *mfn = page_to_mfn(page);
>     if ( !mfn_valid(*mfn) )
>         ret = -EINVAL;
>
> which, while not at risk of yielding any UB, is an entirely pointless
> check (as the underlying assumption needs to be that the struct
> page_info * must already be a valid pointer, or else _its_ obtaining was
> possibly UB). Having such checks in the code is at best misleading.
>

Yeah, that does looks wonky.  Luckily ARGO is still experimental.

Although given the topics on the call today, I wonder whether this is
really appropriate, seeing as ARGO underpins the security model on OpenXT...

I did have some plans (for copious free time, of course) to investigate
coccinelle and use it like Linux does, in order to spot and correct
known anti-patterns.

~Andrew