On Wed, Sep 13, 2023 at 08:09:57AM -0300, Christoph Hellwig wrote:
> Releasing an anon dev_t is a very common thing when freeing a
> super_block, as that's done for basically any not block based file
> system (modulo the odd mtd special case). So instead of requiring
> a special ->kill_sb helper and a lot of boilerplate in more complicated
> file systems, just release the anon dev_t in deactivate_locked_super if
> the super_block was using one.
>
> As the freeing is done after the main call to kill_super_notify, this
> removes the need for having two slightly different call sites for it.
Huh? At this stage in your series freeing is still in ->kill_sb()
instances, after the calls of kill_anon_super() you've turned into
the calls of generic_shutdown_super().
You do split it off into a separate method later in the series, but
at this point you are reopening the same UAF that had been dealt with
in dc3216b14160 "super: ensure valid info".
Either move the introduction of ->free_sb() before that one, or
split it into lifting put_anon_bdev() (left here) and getting rid
of kill_anon_super() (after ->free_sb() introduction).
On Thu, Sep 14, 2023 at 12:27:12AM +0100, Al Viro wrote:
> On Wed, Sep 13, 2023 at 08:09:57AM -0300, Christoph Hellwig wrote:
> > Releasing an anon dev_t is a very common thing when freeing a
> > super_block, as that's done for basically any not block based file
> > system (modulo the odd mtd special case). So instead of requiring
> > a special ->kill_sb helper and a lot of boilerplate in more complicated
> > file systems, just release the anon dev_t in deactivate_locked_super if
> > the super_block was using one.
> >
> > As the freeing is done after the main call to kill_super_notify, this
> > removes the need for having two slightly different call sites for it.
>
> Huh? At this stage in your series freeing is still in ->kill_sb()
> instances, after the calls of kill_anon_super() you've turned into
> the calls of generic_shutdown_super().
>
> You do split it off into a separate method later in the series, but
> at this point you are reopening the same UAF that had been dealt with
> in dc3216b14160 "super: ensure valid info".
>
> Either move the introduction of ->free_sb() before that one, or
> split it into lifting put_anon_bdev() (left here) and getting rid
> of kill_anon_super() (after ->free_sb() introduction).
Actually, looking at the final stage in the series, you still have
kill_super_notify() done *AFTER* ->free_sb() call. So the problem
persists until the very end...
On Thu, Sep 14, 2023 at 12:27:12AM +0100, Al Viro wrote:
> On Wed, Sep 13, 2023 at 08:09:57AM -0300, Christoph Hellwig wrote:
> > Releasing an anon dev_t is a very common thing when freeing a
> > super_block, as that's done for basically any not block based file
> > system (modulo the odd mtd special case). So instead of requiring
> > a special ->kill_sb helper and a lot of boilerplate in more complicated
> > file systems, just release the anon dev_t in deactivate_locked_super if
> > the super_block was using one.
> >
> > As the freeing is done after the main call to kill_super_notify, this
> > removes the need for having two slightly different call sites for it.
>
> Huh? At this stage in your series freeing is still in ->kill_sb()
> instances, after the calls of kill_anon_super() you've turned into
> the calls of generic_shutdown_super().
The above refers to freeing the anon dev_t, which at this stage is done
right after the kill_super_notify in generic_shutdown_super.
> You do split it off into a separate method later in the series, but
> at this point you are reopening the same UAF that had been dealt with
> in dc3216b14160 "super: ensure valid info".
How?
Old sequence before his patch:
deactivate_locked_super()
-> kill_anon_super()
-> generic_shutdown_super()
-> kill_super_notify()
-> free_anon_bdev()
-> kill_super_notify()
New sequence with this patch:
deactivate_locked_super()
-> generic_shutdown_super()
-> kill_super_notify()
-> free_anon_bdev()