Re: [PATCH 03/20] ext4: fix double brelse() the buffer of the extents path

[Baokun Li <libaokun@huaweicloud.com>]

On 2024/7/26 19:45, Ojaswin Mujoo wrote:
> On Wed, Jul 10, 2024 at 12:06:37PM +0800, libaokun@huaweicloud.com wrote:
>> From: Baokun Li <libaokun1@huawei.com>
>>
>> In ext4_ext_try_to_merge_up(), set path[1].p_bh to NULL after it has been
>> released, otherwise it may be released twice.
>>
>> An example of what triggers this is as follows:
>>
>>    split2    map    split1
>> |--------|-------|--------|
>>
>> ext4_ext_map_blocks
>>   ext4_ext_handle_unwritten_extents
>>    ext4_split_convert_extents
>>     // path->p_depth == 0
>>     ext4_split_extent
>>       // 1. do split1
>>       ext4_split_extent_at
>>         ext4_ext_insert_extent
>>           ext4_ext_create_new_leaf
>>             ext4_ext_grow_indepth
>>               le16_add_cpu(&neh->eh_depth, 1)
>>             ext4_find_extent
>>               path->p_depth = 1
>>           ext4_ext_try_to_merge
>>             ext4_ext_try_to_merge_up
>>               path->p_depth = 0
>>               brelse(path[1].p_bh)  ---> not set to NULL here
>>       // 2. update path
>>       ext4_find_extent
>>       // 3. do split2
>>       ext4_split_extent_at
>>         ext4_ext_insert_extent
>>           ext4_ext_create_new_leaf
>>             ext4_ext_grow_indepth
>>               le16_add_cpu(&neh->eh_depth, 1)
>>             ext4_find_extent
>>               path[0].p_bh = NULL;
>>               path->p_depth = 1
>>               read_extent_tree_block  ---> return err
>>               // path[1].p_bh is still the old value
>>   ext4_free_ext_path
>>    ext4_ext_drop_refs
>>     // path->p_depth == 1
>>     brelse(path[1].p_bh)  ---> brelse a buffer twice
> Hi Baokun,
>
> If i'm not wrong, in this trace, we'll enter ext4_ext_insert_extent() with
> gb_flags having EXT4_GET_BLOCKS_PRE_IO so we won't actually go for a
> merge_up.
>
> That being said, there seems to be a few places where we follow the
> split-insert pattern and it might be possible that one of those call
> sites might not be passing EXT4_GET_BLOCKS_PRE_IO and we'll the double
> free issue you mentioned. I'll check and update if I see anything.
Hi Ojaswin,

You're right. I am very sorry for the confusion.

The trace here is wrong, this patch should actually be placed after the two
UAF patches. Here ext4_ext_try_to_merge() is called when trying zeroout in
ext4_split_extent_at(). It is called when trying zeroout with or without
EXT4_GET_BLOCKS_PRE_IO.The correct trace is as follows:

   split2    map    split1
|--------|-------|--------|

ext4_ext_map_blocks
  ext4_ext_handle_unwritten_extents
   ext4_split_convert_extents
    // path->p_depth == 0
    ext4_split_extent
      // 1. do split1
      ext4_split_extent_at
        |ext4_ext_insert_extent
        |  ext4_ext_create_new_leaf
        |    ext4_ext_grow_indepth
        |      le16_add_cpu(&neh->eh_depth, 1)
        |    ext4_find_extent
        |      // return -ENOMEM
        |// get error and try zeroout
        |path = ext4_find_extent
        |  path->p_depth = 1
        |ext4_ext_try_to_merge
        |  ext4_ext_try_to_merge_up
        |    path->p_depth = 0
        |    brelse(path[1].p_bh)  ---> not set to NULL here
        |// zeroout success
      // 2. update path
      ext4_find_extent
      // 3. do split2
      ext4_split_extent_at
        ext4_ext_insert_extent
          ext4_ext_create_new_leaf
            ext4_ext_grow_indepth
              le16_add_cpu(&neh->eh_depth, 1)
            ext4_find_extent
              path[0].p_bh = NULL;
              path->p_depth = 1
              read_extent_tree_block  ---> return err
              // path[1].p_bh is still the old value
  ext4_free_ext_path
   ext4_ext_drop_refs
    // path->p_depth == 1
    brelse(path[1].p_bh)  ---> brelse a buffer twice

I'll adjust the order of the patches and correct the trace in the next 
version.
> On a separate note, isn't it a bit weird that we grow the tree indepth
> (which includes allocation, marking buffer dirty etc) only to later
> merge it up again and throwing all the changes we did while growing the
> tree. Surely we could optimize this particular case somehow right?
Sorry that my trace misled you. It seems reasonable to try to merge extent
in error handling.
>> Finally got the following WARRNING when removing the buffer from lru:
>>
>> ============================================
>> VFS: brelse: Trying to free free buffer
>> WARNING: CPU: 2 PID: 72 at fs/buffer.c:1241 __brelse+0x58/0x90
>> CPU: 2 PID: 72 Comm: kworker/u19:1 Not tainted 6.9.0-dirty #716
>> RIP: 0010:__brelse+0x58/0x90
>> Call Trace:
>>   <TASK>
>>   __find_get_block+0x6e7/0x810
>>   bdev_getblk+0x2b/0x480
>>   __ext4_get_inode_loc+0x48a/0x1240
>>   ext4_get_inode_loc+0xb2/0x150
>>   ext4_reserve_inode_write+0xb7/0x230
>>   __ext4_mark_inode_dirty+0x144/0x6a0
>>   ext4_ext_insert_extent+0x9c8/0x3230
>>   ext4_ext_map_blocks+0xf45/0x2dc0
>>   ext4_map_blocks+0x724/0x1700
>>   ext4_do_writepages+0x12d6/0x2a70
>> [...]
>> ============================================
>>
>> Fixes: ecb94f5fdf4b ("ext4: collapse a single extent tree block into the inode if possible")
>> Cc: stable@kernel.org
>> Signed-off-by: Baokun Li <libaokun1@huawei.com>
>> ---
>>   fs/ext4/extents.c | 1 +
>>   1 file changed, 1 insertion(+)
>>
>> diff --git a/fs/ext4/extents.c b/fs/ext4/extents.c
>> index 4d589d34b30e..657baf3991c1 100644
>> --- a/fs/ext4/extents.c
>> +++ b/fs/ext4/extents.c
>> @@ -1888,6 +1888,7 @@ static void ext4_ext_try_to_merge_up(handle_t *handle,
>>   	path[0].p_hdr->eh_max = cpu_to_le16(max_root);
>>   
>>   	brelse(path[1].p_bh);
>> +	path[1].p_bh = NULL;
> Anyways, I agree that adding this here is the right thing to do:
>
> Reviewed-by: Ojaswin Mujoo <ojaswin@linux.ibm.com>
>
> Thanks,
> Ojaswin
Thanks for the review!
>>   	ext4_free_blocks(handle, inode, NULL, blk, 1,
>>   			 EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET);
>>   }
>> -- 
>> 2.39.2
>>
-- 
With Best Regards,
Baokun Li