On Fri, Sep 15, 2023 at 03:28:14PM +0100, Al Viro wrote:
> On Fri, Sep 15, 2023 at 04:12:07PM +0200, Christian Brauner wrote:
> > + static void some_fs_kill_sb(struct super_block *sb)
> > + {
> > + struct some_fs_info *info = sb->s_fs_info;
> > +
> > + kill_*_super(sb);
> > + kfree(info);
> > + }
> > +
> > +It's best practice to never deviate from this pattern.
>
> The last part is flat-out incorrect. If e.g. fatfs or cifs ever switches
> to that pattern, you'll get UAF - they need freeing of ->s_fs_info
> of anything that ever had been mounted done with RCU delay; moreover,
> unload_nls() in fatfs needs to be behind the same.
>
> Lifetime rules for fs-private parts of superblock are really private to
> filesystem; their use by sget/sget_fc callbacks might impose restrictions
> on those, but that again is none of the VFS business.
PS: and no, we don't want to impose such RCU delay on every filesystem
out there; what's more, there's nothing to prohibit e.g. having ->s_fs_info
pointing to a refcounted fs-private object (possibly shared by various
superblocks), so freeing might very well be "drop the reference and destroy
if refcount has reached 0".
