Re: [BUG] Boot crash on v6.7-rc2

["Gustavo A. R. Silva" <gustavo@embeddedor.com>]

On 11/25/23 12:31, Kees Cook wrote:
> 
> 
> On November 25, 2023 9:54:28 AM PST, "Gustavo A. R. Silva" <gustavo@embeddedor.com> wrote:
>>
>>
>> On 11/24/23 09:28, Gustavo A. R. Silva wrote:
>>>
>>>
>>> On 11/24/23 04:24, Joey Gouly wrote:
>>>> Hi all,
>>>>
>>>> I just hit a boot crash on v6.7-rc2 (arm64, FVP model):
>>>
>>> [..]
>>>
>>>> Checking `struct neighbour`:
>>>>
>>>>      struct neighbour {
>>>>          struct neighbour __rcu    *next;
>>>>          struct neigh_table    *tbl;
>>>>      .. fields ..
>>>>          u8            primary_key[0];
>>>>      } __randomize_layout;
>>>>
>>>> Due to the `__randomize_layout`, `primary_key` field is being placed before `tbl` (actually it's the same address since it's a 0 length array). That means the memcpy() corrupts the tbl pointer.
>>>>
>>>> I think I just got unlucky with my CONFIG_RANDSTRUCT seed (I can provide it if needed), it doesn't look as if it's a new issue.
>>>
>>> It seems the issue is caused by this change that was recently added to -rc2:
>>>
>>> commit 1ee60356c2dc ("gcc-plugins: randstruct: Only warn about true flexible arrays")
>>>
>>> Previously, one-element and zero-length arrays were treated as true flexible arrays
>>> (however, they are "fake" flex arrays), and __randomize_layout would leave them
>>> untouched at the end of the struct; the same for proper C99 flex-array members. But
>>> after the commit above, that's no longer the case: Only C99 flex-array members will
>>> behave correctly (remaining untouched at end of the struct), and the other two types
>>> of arrays will be randomized.
>>
>> mmh... it seems that commit 1ee60356c2dc only prevents one-element arrays from being
>> treated as flex arrays, while the code should still keep zero-length arrays untouched:
>>
>>         if (typesize == NULL_TREE && TYPE_DOMAIN(fieldtype) != NULL_TREE &&
>>             TYPE_MAX_VALUE(TYPE_DOMAIN(fieldtype)) == NULL_TREE)
>>                 return true;
>>
>> -       if (typesize != NULL_TREE &&
>> -           (TREE_CONSTANT(typesize) && (!tree_to_uhwi(typesize) ||
>> -            tree_to_uhwi(typesize) == tree_to_uhwi(elemsize))))
>> -               return true;
>> -
> 
> This should be both the 0 and 1 checks. I think the original fix is correct: switch to a true flex array.

This code is new to me and I got a bit confused. Thanks for the clarification. :)

So, it'd be nice to apply this change:

https://lore.kernel.org/linux-hardening/b6c1c3ce-3ba0-4439-b0fb-2bb0c38586e0@embeddedor.com/

> 
>>
>> Sorry about the confusion.
>>
>>>
>>>>
>>>> I couldn't reproduce directly on v6.6 (the offsets for `tbl` and `primary_key` didn't overlap).
>>>> However I tried changing the zero-length-array to a flexible one:
>>>>
>>>>      +    DECLARE_FLEX_ARRAY(u8, primary_key);
>>>>      +    u8        primary_key[0];
> 
> Was this line supposed to be "-"?
> 
>>>>
>>>> Then the field offsets ended up overlapping, and I also got the same crash on v6.6.
>>>
>>> The right approach is to transform the zero-length array into a C99 flex-array member,
>>> like this:
>>>
>>> diff --git a/include/net/neighbour.h b/include/net/neighbour.h
>>> index 07022bb0d44d..0d28172193fa 100644
>>> --- a/include/net/neighbour.h
>>> +++ b/include/net/neighbour.h
>>> @@ -162,7 +162,7 @@ struct neighbour {
>>>           struct rcu_head         rcu;
>>>           struct net_device       *dev;
>>>           netdevice_tracker       dev_tracker;
>>> -       u8                      primary_key[0];
>>> +       u8                      primary_key[];
>>>    } __randomize_layout;
>>>
>>>    struct neigh_ops {
>>
>> In any case, I think we still should convert [0] to [ ].
> 
> I would expect the above to fix the problem. If it doesn't I'll need to take a closer look at the plugin...

I think this should fix the issue. Let me go create a proper patch for this.
I'll send it out, shortly.

--
Gustavo