Date:   Wed, 15 Feb 2023 11:46:52 +0000
From:   Tudor Ambarus <tudor.ambarus@...aro.org>
To:     Theodore Ts'o <tytso@....edu>, Jun Nie <jun.nie@...aro.org>
Cc:     "Darrick J. Wong" <djwong@...nel.org>, adilger.kernel@...ger.ca,
        linux-ext4@...r.kernel.org, linux-kernel@...r.kernel.org,
        Lee Jones <joneslee@...gle.com>
Subject: Re: [PATCH] ext4: reject 1k block fs on the first block of disk

Hi, Ted!

On 2/15/23 04:32, Theodore Ts'o wrote:
> On Wed, Jan 04, 2023 at 09:58:03AM +0800, Jun Nie wrote:
>> Darrick J. Wong <djwong@...nel.org> 于2023年1月4日周三 03:17写道：
>>>
>>> On Thu, Dec 29, 2022 at 09:45:02AM +0800, Jun Nie wrote:
>>>> For 1k-block filesystems, the filesystem starts at block 1, not block 0.
>>>> If start_fsb is 0, it will be bump up to s_first_data_block. Then
>>>> ext4_get_group_no_and_offset don't know what to do and return garbage
>>>> results (blockgroup 2^32-1). The underflow make index
>>>> exceed es->s_groups_count in ext4_get_group_info() and trigger the BUG_ON.
>>>>
>>>> Fixes: 4a4956249dac0 ("ext4: fix off-by-one fsmap error on 1k block filesystems")
>>>> Link: https://syzkaller.appspot.com/bug?id=79d5768e9bfe362911ac1a5057a36fc6b5c30002
>>>> Reported-by: syzbot+6be2b977c89f79b6b153@...kaller.appspotmail.com
>>>> Signed-off-by: Jun Nie <jun.nie@...aro.org>
>>>> ---
>>>>   fs/ext4/fsmap.c | 6 ++++++
>>>>   1 file changed, 6 insertions(+)
>>>>
>>>> diff --git a/fs/ext4/fsmap.c b/fs/ext4/fsmap.c
>>>> index 4493ef0c715e..1aef127b0634 100644
>>>> --- a/fs/ext4/fsmap.c
>>>> +++ b/fs/ext4/fsmap.c
>>>> @@ -702,6 +702,12 @@ int ext4_getfsmap(struct super_block *sb, struct ext4_fsmap_head *head,
>>>>                if (handlers[i].gfd_dev > head->fmh_keys[0].fmr_device)
>>>>                        memset(&dkeys[0], 0, sizeof(struct ext4_fsmap));
>>>>
>>>> +             /*
>>>> +              * Re-check the range after above limit operation and reject
>>>> +              * 1K fs on block 0 as fs should start block 1. */
>>>> +             if (dkeys[0].fmr_physical ==0 && dkeys[1].fmr_physical == 0)
>>>> +                     continue;
>>>
>>> ...and if this filesystem has 4k blocks, and therefore *does* define a
>>> block 0?
>>
>> Yes, this is a real corner case test :-)
> 
> So I'm really nervous about this change.  I don't understand the code;
> and I don't understand how the reproducer works.  I can certainly
> reproduce it using the reproducer found here[1], but it seems to
> require running multiple processes all creating loop devices and then
> running FS_IOC_GETMAP.
> 
> [1] https://syzkaller.appspot.com/bug?id=79d5768e9bfe362911ac1a5057a36fc6b5c30002
> 
> If I change the reproducer to just run the execute_one() once, it
> doesn't trigger the bug.  It seems to only trigger when you have
> multiple processes all racing to create a loop device, mount the file
> system, try running FS_IOC_GETMAP --- and then delete the loop device
> without actually unmounting the file system.  Which is **weird***.
> 
> I've tried taking the image, and just running "xfs_io -c fsmap /mnt",
> and that doesn't trigger it either.
> 
> And I don't understand the reply to Darrick's question about why it's
> safe to add the check since for 4k block file systems, block 0 *is*
> valid.
> 
> So if someone can explain to me what is going on here with this code
> (there are too many abstractions and what's going on with keys is just
> making my head hurt), *and* what the change actually does, and how to
> reproduce the problem with a ***simple*** reproducer -- the syzbot
> mess doesn't count, that would be great.  But applying a change that I
> don't understand to code I don't understand, to fix a reproducer which
> I also doesn't understand, just doesn't make me feel comfortable.
> 

Let me share what I understood until now. The low key is zeroed. The
high key is defined and uses a fmr_physical of value zero, which is
smaller than the first data block for the 1k-block ext4 fs (which starts
at offset 1024).

-> ext4_getfsmap_datadev()
   keys[0].fmr_physical = 0, keys[1].fmr_physical = 0
   bofs = le32_to_cpu(sbi->s_es->s_first_data_block) = 1, eofs = 256
   start_fsb = keys[0].fmr_physical = 1, end_fsb = keys[1].fmr_physical = 0
   -> ext4_get_group_no_and_offset()
     blocknr = 1, le32_to_cpu(es->s_first_data_block) =1
   start_ag = 0, first_cluster = 0
   ->
     blocknr = 0, le32_to_cpu(es->s_first_data_block) =1
   end_ag = 4294967295, last_cluster = 8191

   Then there's a loop that stops when info->gfi_agno <= end_ag; that 
will trigger the BUG_ON in ext4_get_group_info() as the group nr exceeds 
EXT4_SB(sb)->s_groups_count)
   -> ext4_mballoc_query_range()
     -> ext4_mb_load_buddy()
       -> ext4_mb_load_buddy_gfp()
         -> ext4_get_group_info()

It's an out of bounds request and Darrick suggested to not return any
mapping for the byte range 0-1023 for the 1k-block filesystem. The
alternative would be to return -EINVAL when the high key starts at
fmr_phisical of value zero for the 1k-block fs.

In order to reproduce this one would have to create an 1k-block ext4 fs
and to pass a high key with fmr_physical of value zero, thus I would
expect to reproduce it with something like this:
xfs_io -c 'fsmap -d 0 0' /mnt/scratch

However when doing this I notice that in
xfsprogs-dev/io/fsmap.c l->fmr_device and h->fmr_device will have value
zero, FS_IOC_GETFSMAP is called and then we receive no entries
(head->fmh_entries = 0). Now I'm trying to see what I do wrong, and how
to reproduce the bug.

Cheers,
ta

Powered by blists - more mailing lists