Re: [PATCH 1/2] ext2: fix race between setxattr and write back

[Jan Kara]

On Fri 11-08-23 11:38:56, Ye Bin wrote:
> There's a issue as follows:
> Block Allocation Reservation Windows Map (ext2_try_to_allocate_with_rsv):
> reservation window 0x000000006f105382 start: 0, end: 0
> reservation window 0x000000008fd1a555 start: 1044, end: 1059
> Window map complete.
> kernel BUG at fs/ext2/balloc.c:1158!
> invalid opcode: 0000 [#1] PREEMPT SMP KASAN
> RIP: 0010:ext2_try_to_allocate_with_rsv.isra.0+0x15c4/0x1800
> Call Trace:
>  <TASK>
>  ext2_new_blocks+0x935/0x1690
>  ext2_new_block+0x73/0xa0
>  ext2_xattr_set2+0x74f/0x1730
>  ext2_xattr_set+0x12b6/0x2260
>  ext2_xattr_user_set+0x9c/0x110
>  __vfs_setxattr+0x139/0x1d0
>  __vfs_setxattr_noperm+0xfc/0x370
>  __vfs_setxattr_locked+0x205/0x2c0
>  vfs_setxattr+0x19d/0x3b0
>  do_setxattr+0xff/0x220
>  setxattr+0x123/0x150
>  path_setxattr+0x193/0x1e0
>  __x64_sys_setxattr+0xc8/0x170
>  do_syscall_64+0x35/0x80
>  entry_SYSCALL_64_after_hwframe+0x63/0xcd
> 
> Above issue may happens as follows:
>         setxattr                             write back
> ext2_xattr_set
>   ext2_xattr_set2
>     ext2_new_block
>       ext2_new_blocks
>         ext2_try_to_allocate_with_rsv
> 	  alloc_new_reservation
> 	  --> group=0 [0, 1023] rsv [1016, 1023]
> 	                                do_writepages
> 					  mpage_writepages
> 					    write_cache_pages
> 					      __mpage_writepage
> 					        ext2_get_block
> 						  ext2_get_blocks
> 						   ext2_alloc_branch
> 						    ext2_new_blocks
> 						     ext2_try_to_allocate_with_rsv
> 						       alloc_new_reservation
> 		                     -->group=1 [1024, 2047] rsv [1044, 1059]
> 	  if ((my_rsv->rsv_start > group_last_block) ||
> 	      (my_rsv->rsv_end < group_first_block)
> 	      rsv_window_dump
> 	      BUG();
> Now ext2 mkwrite delay allocate new blocks. So there maybe allocate blocks when
> do write back. However, there is no concurrent protection between
> ext2_xattr_set() and do_writepages().
> To solve about issue hold '&ei->truncate_mutex' lock when new block for xattr.
> 
> Signed-off-by: Ye Bin <yebin10@xxxxxxxxxx>

Thanks for the patch! I agree the handling of reservation window and its
use for block allocation needs to be protected by ei->i_truncate_mutex.
However in this particular case of xattr allocation where we want to
allocate just one block which is completely independent of file data, I'd
rather choose to make ext2_new_blocks() ignore the reservation window (set
my_rsv to NULL). There's already a logic at the beginning of
ext2_new_blocks() deciding whether to use the reservation window or not and
we could extend it - probably by adding flags argument to it a introducing
a NORESERVE flag.

Also as a preparatory patch, I'd just remove ext2_new_block() and opencode
it in the xattr code since it has only that one user anyway.

								Honza



> ---
>  fs/ext2/balloc.c | 8 +++++++-
>  1 file changed, 7 insertions(+), 1 deletion(-)
> 
> diff --git a/fs/ext2/balloc.c b/fs/ext2/balloc.c
> index c8049c90323d..039f655febfd 100644
> --- a/fs/ext2/balloc.c
> +++ b/fs/ext2/balloc.c
> @@ -1432,8 +1432,14 @@ ext2_fsblk_t ext2_new_blocks(struct inode *inode, ext2_fsblk_t goal,
>  ext2_fsblk_t ext2_new_block(struct inode *inode, unsigned long goal, int *errp)
>  {
>  	unsigned long count = 1;
> +	struct ext2_inode_info *ei = EXT2_I(inode);
> +	ext2_fsblk_t ret;
> +
> +	mutex_lock(&ei->truncate_mutex);
> +	ret = ext2_new_blocks(inode, goal, &count, errp);
> +	mutex_unlock(&ei->truncate_mutex);
>  
> -	return ext2_new_blocks(inode, goal, &count, errp);
> +	return ret;
>  }
>  
>  #ifdef EXT2FS_DEBUG
> -- 
> 2.31.1
> 
-- 
Jan Kara <jack@xxxxxxxx>
SUSE Labs, CR