Re: [PATCH 2/2] debugfs: return error values, not NULL

[Michal Hocko]

On Wed 23-01-19 13:26:26, Greg KH wrote:
> On Wed, Jan 23, 2019 at 01:13:50PM +0100, Michal Hocko wrote:
> > On Wed 23-01-19 12:55:35, Greg KH wrote:
> > > On Wed, Jan 23, 2019 at 12:06:28PM +0100, Michal Hocko wrote:
> > > > On Wed 23-01-19 11:28:14, Greg KH wrote:
> > > > > When an error happens, debugfs should return an error pointer value, not
> > > > > NULL.  This will prevent the totally theoretical error where a debugfs
> > > > > call fails due to lack of memory, returning NULL, and that dentry value
> > > > > is then passed to another debugfs call, which would end up succeeding,
> > > > > creating a file at the root of the debugfs tree, but would then be
> > > > > impossible to remove (because you can not remove the directory NULL).
> > > > > 
> > > > > So, to make everyone happy, always return errors, this makes the users
> > > > > of debugfs much simpler (they do not have to ever check the return
> > > > > value), and everyone can rest easy.
> > > > 
> > > > How come this is safe at all? Say you are creating a directory by
> > > > debugfs_create_dir and then feed the return value to debugfs_create_files
> > > > as a parent. In case of error you are giving it an invalid pointer and
> > > > likely blow up unless I miss something.
> > > 
> > > debugfs_create_files checks for invalid parents and will just refuse to
> > > create the file.  It's always done that.
> > 
> > I must be missing something because debugfs_create_files does
> > 	d_inode(parent)->i_private = data;
> > as the very first thing and that means that it dereferences an invalid
> > pointer right there.
> 
> debugfs_create_file() -> __debugfs_create_file() -> start_creating()
> and that function checks if parent is an error, which it aborts on, or
> if it is NULL, it sets parent to a valid value:
> 
>         /* If the parent is not specified, we create it in the root.
>          * We need the root dentry to do this, which is in the super
>          * block. A pointer to that is in the struct vfsmount that we
>          * have around.
>          */
>         if (!parent)
>                 parent = debugfs_mount->mnt_root;
> 
> I don't see any line that looks like:
> >       d_inode(parent)->i_private = data;
> in Linus's tree right now, what kernel version are you referring to?

Ohh, my bad. I have looked at debugfs_create_files which is a mq helper
around debugfs_create_file. But that is a good example why this patch is
dangerous anyway. blk_mq_debugfs_register simply checks for NULL and
debugfs_create_files doesn't expect ERR_PTR here. So you would have to
check each and every user to make sure you can do that.

> > > > I do agree that reporting errors is better than a simple catch all NULL
> > > > but this should have been done when introduced rather than now when most
> > > > callers simply check for NULL as a failure.
> > > 
> > > I'm fixing up all the "NULL is a failure" callsites in the kernel, see
> > > lkml for the first round of those patches.
> > 
> > You are merely removing them, which doesn't really help for this patch.
> 
> It doesn't hurt either, as if you really wanted to handle errors from debugfs
> properly, you have to check for IS_ERR() as well, because the filesystem can be
> compiled out (and then it returns an error pointer)

I would assume that this would be achieved by a direct config
dependency. E.g. BLK_DEBUG_FS. So the code doesn't even get compiled and
wouldn't ever encounter the ERR_PTR.

Yeah this whole debugfs API is broken and it would take a lot of time to
unclutter it but it definitely is not that simple as what this patch
does.
-- 
Michal Hocko
SUSE Labs