Re: ext4_mballoc_test: Internal error: Oops: map_id_range_down (kernel/user_namespace.c:318)

[Kemeng Shi]

on 2/29/2024 6:09 PM, Christian Brauner wrote:
> On Wed, Feb 28, 2024 at 11:33:36AM -0800, Guenter Roeck wrote:
>> On 2/28/24 11:26, Daniel Díaz wrote:
>>> Hello!
>>>
>>> On Wed, 28 Feb 2024 at 12:19, Naresh Kamboju <naresh.kamboju@xxxxxxxxxx> wrote:
>>>> Kunit ext4_mballoc_test tests found following kernel oops on Linux next.
>>>> All ways reproducible on all the architectures and steps to reproduce shared
>>>> in the bottom of this email.
>>>>
>>>> Reported-by: Linux Kernel Functional Testing <lkft@xxxxxxxxxx>
>>>>
>>
>> [ ... ]
>>
>>> +Guenter. Just the thing we were talking about, at about the same time.
>>>
>>
>> Good that others see the same problem. Thanks a lot for reporting!
> 
> Hm...
> 
> static struct super_block *mbt_ext4_alloc_super_block(void)
> {                                                                                                                                                                                                       struct ext4_super_block *es = kzalloc(sizeof(*es), GFP_KERNEL);
>         struct ext4_sb_info *sbi = kzalloc(sizeof(*sbi), GFP_KERNEL);
>         struct mbt_ext4_super_block *fsb = kzalloc(sizeof(*fsb), GFP_KERNEL);
> 
>         if (fsb == NULL || sbi == NULL || es == NULL)
>                 goto out;
> 
>         sbi->s_es = es;
>         fsb->sb.s_fs_info = sbi;
>         return &fsb->sb;
> 
> out:
>         kfree(fsb);
>         kfree(sbi);
>         kfree(es);
>         return NULL;
> }
> 
> That VFS level struct super_block that is returned from this function is
> never really initialized afaict? Therefore, sb->s_user_ns == NULL:
> 
> i_uid_write(sb, ...)
> -> NULL = i_user_ns(sb)
>    -> make_kuid(NULL)
>       -> map_id_range_down(NULL)
> 
> Outside of this test this can never be the case. See alloc_super() in
> fs/super.c. So to stop the bleeding this needs something like:
> 
> static struct super_block *mbt_ext4_alloc_super_block(void)
> {
>         struct ext4_super_block *es = kzalloc(sizeof(*es), GFP_KERNEL);
>         struct ext4_sb_info *sbi = kzalloc(sizeof(*sbi), GFP_KERNEL);
>         struct mbt_ext4_super_block *fsb = kzalloc(sizeof(*fsb), GFP_KERNEL);
> 
>         if (fsb == NULL || sbi == NULL || es == NULL)
>                 goto out;
> 
>         sbi->s_es = es;
>         fsb->sb.s_fs_info = sbi;
> +       fsb.sb.s_user_ns = &init_user_ns;
>         return &fsb->sb;
> 
> out:
>         kfree(fsb);
>         kfree(sbi);
>         kfree(es);
>         return NULL;
> }
> 
Hi Christian,
Thanks for the information. I'm looking at this too and I also found
root cause is sb.s_user_ns is NULL. I'm considering to get a
super_block with VFS level api sget_fc to fix this to avoid similar
problem when new unit tests are added or new member is added to
super_block.
Would like to hear more from you. Thanks!