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

[Baokun Li <libaokun1@huawei.com>]

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,

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.


Regards,
Baokun