Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.34 as permitted sender) client-ip=23.128.96.34;
Message-ID: <d7294586-04a4-49f7-8f5f-2dd66c8b4cde@embeddedor.com>
Date:   Sat, 25 Nov 2023 11:54:28 -0600
MIME-Version: 1.0
User-Agent: Mozilla Thunderbird
Subject: Re: [BUG] Boot crash on v6.7-rc2
From:   "Gustavo A. R. Silva" <gustavo@embeddedor.com>
To:     Joey Gouly <joey.gouly@arm.com>, Kees Cook <keescook@chromium.org>,
        linux-hardening@vger.kernel.org, netdev@vger.kernel.org
Cc:     "David S. Miller" <davem@davemloft.net>,
        Eric Dumazet <edumazet@google.com>,
        Jakub Kicinski <kuba@kernel.org>,
        Paolo Abeni <pabeni@redhat.com>, linux-kernel@vger.kernel.org,
        "Gustavo A. R. Silva" <gustavoars@kernel.org>
References: <20231124102458.GB1503258@e124191.cambridge.arm.com>
 <7086f60f-9f74-4118-a10c-d98b9c6cc8eb@embeddedor.com>
Content-Language: en-US
In-Reply-To: <7086f60f-9f74-4118-a10c-d98b9c6cc8eb@embeddedor.com>
Content-Type: text/plain; charset=UTF-8; format=flowed
Content-Transfer-Encoding: 8bit
Precedence: bulk


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;
-

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];
>>
>> 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 [ ].

--
Gustavo