Re: [RFC PATCH] ext4: increase the protection of drop nlink and ext4 inode destroy

[Andreas Dilger]

On Dec 26, 2016, at 5:34 AM, yi zhang <yi.zhang@xxxxxxxxxx> wrote:
> 
> Because of the disk and hardware issue, the ext4 filesystem have
> many errors, the inode->i_nlink of ext4 becomes zero abnormally
> but the dentry is still positive, it will cause memory corruption
> after the following process:
> 
> 1) Due to the inode->i_nlink is 0, this inode will be added into
> the orhpan list,
> 2) ext4_rename() cover this inode, and drop_nlink() will reverse
> the inode->i_nlink to 0xFFFFFFFF,
> 3) iput() add this inode to LRU,
> 4) evict() will call destroy_inode() to destroy this inode but
> skip removing it from the orphan list,
> 5) after this, the inode's memory address space will be used by
> other module, when the ext4 filesystem change the orphan list, it will
> trample other module's data and then may cause oops.
> 
> Although we cannot avoid hardware and disk errors, we can control the
> softwore error in the ext4 module, do not affect other modules and
> increase the difficulty of locating problems.
> 
> This patch avoid inode->i_nlink reverse and remove the inode form the

(typo) s/form/from/

> orphan list when destroy it if the list is not empty.
> Signed-off-by: yi zhang <yi.zhang@xxxxxxxxxx>
> ---
> fs/ext4/super.c | 1 +
> fs/inode.c      | 5 ++++-
> 2 files changed, 5 insertions(+), 1 deletion(-)
> 
> diff --git a/fs/ext4/super.c b/fs/ext4/super.c
> index 52b0530..617327e 100644
> --- a/fs/ext4/super.c
> +++ b/fs/ext4/super.c
> @@ -975,6 +975,7 @@ static void ext4_destroy_inode(struct inode *inode)
> 				EXT4_I(inode), sizeof(struct ext4_inode_info),
> 				true);
> 		dump_stack();
> +		ext4_orphan_del(NULL, inode);
> 	}
> 	call_rcu(&inode->i_rcu, ext4_i_callback);
> }
> diff --git a/fs/inode.c b/fs/inode.c
> index 88110fd..079d383 100644
> --- a/fs/inode.c
> +++ b/fs/inode.c
> @@ -279,7 +279,10 @@ static void destroy_inode(struct inode *inode)
> */
> void drop_nlink(struct inode *inode)
> {
> -	WARN_ON(inode->i_nlink == 0);
> +	if (WARN(inode->i_nlink == 0, "inode %lu nlink"
> +		" is already 0", inode->i_ino))

(style) the string should be kept on a single line instead of being
split, especially since it can fit easily.

(defect) this needs to have a newline.

	if (WARN(inode->i_nlink == 0,
		 "inode %lu nlink is already 0\n", inode->i_ino))

Cheers, Andreas

> +		return;
> +
> 	inode->__i_nlink--;
> 	if (!inode->i_nlink)
> 		atomic_long_inc(&inode->i_sb->s_remove_count);
> --
> 2.5.0
> 


Cheers, Andreas

Attachment: signature.asc
Description: Message signed with OpenPGP using GPGMail