Re: [PATCH 1/1] ext4: fix crash on BUG_ON in ext4_alloc_group_tables

[Baokun Li <libaokun1@huawei.com>]

On 2024/9/26 17:23, Aleksandr Mikhalitsyn wrote:
> On Thu, Sep 26, 2024 at 10:50 AM Baokun Li <libaokun1@huawei.com> wrote:
>> On 2024/9/26 0:17, Aleksandr Mikhalitsyn wrote:
>>> On Wed, Sep 25, 2024 at 5:57 PM Jan Kara <jack@suse.cz> wrote:
>>>> On Wed 25-09-24 16:33:24, Alexander Mikhalitsyn wrote:
>>>>> [   33.882936] EXT4-fs (dm-5): mounted filesystem 8aaf41b2-6ac0-4fa8-b92b-77d10e1d16ca r/w with ordered data mode. Quota mode: none.
>>>>> [   33.888365] EXT4-fs (dm-5): resizing filesystem from 7168 to 786432 blocks
>>>>> [   33.888740] ------------[ cut here ]------------
>>>>> [   33.888742] kernel BUG at fs/ext4/resize.c:324!
>>>> Ah, I was staring at this for a while before I understood what's going on
>>>> (it would be great to explain this in the changelog BTW).  As far as I
>>>> understand commit 665d3e0af4d3 ("ext4: reduce unnecessary memory allocation
>>>> in alloc_flex_gd()") can actually make flex_gd->resize_bg larger than
>>>> flexbg_size (for example when ogroup = flexbg_size, ngroup = 2*flexbg_size
>>>> - 1) which then confuses things. I think that was not really intended and
>>> Hi Jan,
>>>
>>> First of all, thanks for your reaction/review on this one ;-)
>>>
>>> You are absolutely right, have just checked with our reproducer and
>>> this modification:
>>>
>>> diff --git a/fs/ext4/resize.c b/fs/ext4/resize.c
>>> index e04eb08b9060..530a918f0cab 100644
>>> --- a/fs/ext4/resize.c
>>> +++ b/fs/ext4/resize.c
>>> @@ -258,6 +258,8 @@ static struct ext4_new_flex_group_data
>>> *alloc_flex_gd(unsigned int flexbg_size,
>>>                   flex_gd->resize_bg = 1 << max(fls(last_group - o_group + 1),
>>>                                                 fls(n_group - last_group));
>>>
>>> +       BUG_ON(flex_gd->resize_bg > flexbg_size);
>>> +
>>>           flex_gd->groups = kmalloc_array(flex_gd->resize_bg,
>>>                                           sizeof(struct ext4_new_group_data),
>>>                                           GFP_NOFS);
>>>
>>> and yes, it crashes on this BUG_ON. So it looks like instead of making
>>> flex_gd->resize_bg to be smaller
>>> than flexbg_size in most cases we can actually have an opposite effect
>>> here. I guess we really need to fix alloc_flex_gd() too.
>>>
>>>> instead of fixing up ext4_alloc_group_tables() we should really change
>>>> the logic in alloc_flex_gd() to make sure flex_gd->resize_bg never exceeds
>>>> flexbg size. Baokun?
>>> At the same time, if I understand the code right, as we can have
>>> flex_gd->resize_bg != flexbg_size after
>>> 5d1935ac02ca5a ("ext4: avoid online resizing failures due to oversized
>>> flex bg") and
>>> 665d3e0af4d3 ("ext4: reduce unnecessary memory allocation in alloc_flex_gd()")
>>> we should always refer to flex_gd->resize_bg value which means that
>>> ext4_alloc_group_tables() fix is needed too.
>>> Am I correct in my understanding?
>> Hi Alex,
> Hi Baokun,
>
>> These two are not exactly equivalent.
>>
>> The flex_gd->resize_bg is only used to determine how many block groups we
>> allocate memory to, i.e., the maximum number of block groups per resize.
>> And the flexbg_size is used to make some judgement on flexible block
>> groups, for example, the BUG_ON triggered in the issue is to make sure
>> src_group and last_group must be in the same flexible block group.
> Huge thanks for explaining this!
>
> Then I guess it's better if you send a patch with your fix.
> Feel free to add my Tested-by tag.
Okay, I'll send a patch later.
>
> Question to you and Jan. Do you guys think that it makes sense to try
> to create a minimal reproducer for this problem without Incus/LXD involved?
> (only e2fsprogs, lvm tools, etc)
>
> I guess this test can be put in the xfstests test suite, right?
>
> Kind regards,
> Alex
I think it makes sense, and it's good to have more use cases to look
around some corners. If you have an idea, let it go.

Regards, Baokun