Re: [PATCH v30 07/12] landlock: Support filesystem access-control

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Tue, Mar 23, 2021 at 8:22 PM Mickaël Salaün <mic@xxxxxxxxxxx> wrote:
> On 23/03/2021 18:49, Jann Horn wrote:
> > On Tue, Mar 23, 2021 at 4:54 PM Mickaël Salaün <mic@xxxxxxxxxxx> wrote:
> >> On 23/03/2021 01:13, Jann Horn wrote:
> >>>  On Tue, Mar 16, 2021 at 9:43 PM Mickaël Salaün <mic@xxxxxxxxxxx> wrote:
> >>>> Using Landlock objects and ruleset, it is possible to tag inodes
> >>>> according to a process's domain.
> >>> [...]
> >>>> +static void release_inode(struct landlock_object *const object)
> >>>> +       __releases(object->lock)
> >>>> +{
> >>>> +       struct inode *const inode = object->underobj;
> >>>> +       struct super_block *sb;
> >>>> +
> >>>> +       if (!inode) {
> >>>> +               spin_unlock(&object->lock);
> >>>> +               return;
> >>>> +       }
> >>>> +
> >>>> +       /*
> >>>> +        * Protects against concurrent use by hook_sb_delete() of the reference
> >>>> +        * to the underlying inode.
> >>>> +        */
> >>>> +       object->underobj = NULL;
> >>>> +       /*
> >>>> +        * Makes sure that if the filesystem is concurrently unmounted,
> >>>> +        * hook_sb_delete() will wait for us to finish iput().
> >>>> +        */
> >>>> +       sb = inode->i_sb;
> >>>> +       atomic_long_inc(&landlock_superblock(sb)->inode_refs);
> >>>> +       spin_unlock(&object->lock);
> >>>> +       /*
> >>>> +        * Because object->underobj was not NULL, hook_sb_delete() and
> >>>> +        * get_inode_object() guarantee that it is safe to reset
> >>>> +        * landlock_inode(inode)->object while it is not NULL.  It is therefore
> >>>> +        * not necessary to lock inode->i_lock.
> >>>> +        */
> >>>> +       rcu_assign_pointer(landlock_inode(inode)->object, NULL);
> >>>> +       /*
> >>>> +        * Now, new rules can safely be tied to @inode with get_inode_object().
> >>>> +        */
> >>>> +
> >>>> +       iput(inode);
> >>>> +       if (atomic_long_dec_and_test(&landlock_superblock(sb)->inode_refs))
> >>>> +               wake_up_var(&landlock_superblock(sb)->inode_refs);
> >>>> +}
> >>> [...]
> >>>> +static struct landlock_object *get_inode_object(struct inode *const inode)
> >>>> +{
> >>>> +       struct landlock_object *object, *new_object;
> >>>> +       struct landlock_inode_security *inode_sec = landlock_inode(inode);
> >>>> +
> >>>> +       rcu_read_lock();
> >>>> +retry:
> >>>> +       object = rcu_dereference(inode_sec->object);
> >>>> +       if (object) {
> >>>> +               if (likely(refcount_inc_not_zero(&object->usage))) {
> >>>> +                       rcu_read_unlock();
> >>>> +                       return object;
> >>>> +               }
> >>>> +               /*
> >>>> +                * We are racing with release_inode(), the object is going
> >>>> +                * away.  Wait for release_inode(), then retry.
> >>>> +                */
> >>>> +               spin_lock(&object->lock);
> >>>> +               spin_unlock(&object->lock);
> >>>> +               goto retry;
> >>>> +       }
> >>>> +       rcu_read_unlock();
> >>>> +
> >>>> +       /*
> >>>> +        * If there is no object tied to @inode, then create a new one (without
> >>>> +        * holding any locks).
> >>>> +        */
> >>>> +       new_object = landlock_create_object(&landlock_fs_underops, inode);
> >>>> +       if (IS_ERR(new_object))
> >>>> +               return new_object;
> >>>> +
> >>>> +       /* Protects against concurrent get_inode_object() calls. */
> >>>> +       spin_lock(&inode->i_lock);
> >>>> +       object = rcu_dereference_protected(inode_sec->object,
> >>>> +                       lockdep_is_held(&inode->i_lock));
> >>>
> >>> rcu_dereference_protected() requires that inode_sec->object is not
> >>> concurrently changed, but I think another thread could call
> >>> get_inode_object() while we're in landlock_create_object(), and then
> >>> we could race with the NULL write in release_inode() here? (It
> >>> wouldn't actually be a UAF though because we're not actually accessing
> >>> `object` here.) Or am I missing a lock that prevents this?
> >>>
> >>> In v28 this wasn't an issue because release_inode() was holding
> >>> inode->i_lock (and object->lock) during the NULL store; but in v29 and
> >>> this version the NULL store in release_inode() moved out of the locked
> >>> region. I think you could just move the NULL store in release_inode()
> >>> back up (and maybe add a comment explaining the locking rules for
> >>> landlock_inode(...)->object)?
> >>>
> >>> (Or alternatively you could use rcu_dereference_raw() with a comment
> >>> explaining that the read pointer is only used to check for NULL-ness,
> >>> and that it is guaranteed that the pointer can't change if it is NULL
> >>> and we're holding the lock. But that'd be needlessly complicated, I
> >>> think.)
> >>
> >> To reach rcu_assign_pointer(landlock_inode(inode)->object, NULL) in
> >> release_inode() or in hook_sb_delete(), the
> >> landlock_inode(inode)->object need to be non-NULL,
> >
> > Yes.
> >
> >> which implies that a
> >> call to get_inode_object(inode) either "retry" (because release_inode is
> >> only called by landlock_put_object, which set object->usage to 0) until
> >> it creates a new object, or reuses the existing referenced object (and
> >> increments object->usage).
> >
> > But it can be that landlock_inode(inode)->object only becomes non-NULL
> > after get_inode_object() has checked
> > rcu_dereference(inode_sec->object).
> >
> >> The worse case would be if
> >> get_inode_object(inode) is called just before the
> >> rcu_assign_pointer(landlock_inode(inode)->object, NULL) from
> >> hook_sb_delete(), which would result in an object with a NULL underobj,
> >> which is the expected behavior (and checked by release_inode).
> >
> > The scenario I'm talking about doesn't involve hook_sb_delete().
> >
> >> The line rcu_assign_pointer(inode_sec->object, new_object) from
> >> get_inode_object() can only be reached if the underlying inode doesn't
> >> reference an object,
> >
> > Yes.
> >
> >> in which case hook_sb_delete() will not reach the
> >> rcu_assign_pointer(landlock_inode(inode)->object, NULL) line for this
> >> same inode.
> >>
> >> This works because get_inode_object(inode) is mutually exclusive to
> >> itself with the same inode (i.e. an inode can only point to an object
> >> that references this same inode).
> >
> > To clarify: You can concurrently call get_inode_object() multiple
> > times on the same inode, right? There are no locks held on entry to
> > that function.
> >
> >> I tried to explain this with the comment "Protects against concurrent
> >> get_inode_object() calls" in get_inode_object(), and the comments just
> >> before both rcu_assign_pointer(landlock_inode(inode)->object, NULL).
> >
> > The scenario I'm talking about is:
> >
> > Initially the inode does not have an associated landlock_object. There
> > are two threads A and B. Thread A is going to execute
> > get_inode_object(). Thread B is going to execute get_inode_object()
> > followed immediately by landlock_put_object().
> >
> > thread A: enters get_inode_object()
> > thread A: rcu_dereference(inode_sec->object) returns NULL
> > thread A: enters landlock_create_object()
> > thread B: enters get_inode_object()
> > thread B: rcu_dereference(inode_sec->object) returns NULL
> > thread B: calls landlock_create_object()
> > thread B: sets inode_sec->object while holding inode->i_lock
> > thread B: leaves get_inode_object()
> > thread B: enters landlock_put_object()
> > thread B: object->usage drops to 0, object->lock is taken
> > thread B: calls release_inode()
> > thread B: drops object->lock
> > thread A: returns from landlock_create_object()
> > thread A: takes inode->i_lock
> >
> > At this point, thread B will run:
> >
> >     rcu_assign_pointer(landlock_inode(inode)->object, NULL);
> >
> > while thread A runs:
> >
> >     rcu_dereference_protected(inode_sec->object,
> >         lockdep_is_held(&inode->i_lock));
> >
> > meaning there is a (theoretical) data race, since
> > rcu_dereference_protected() doesn't use READ_ONCE().
>
> Hum, I see, that is what I was missing. And that explain why there is
> (in practice) no impact on winning the race.
>
> I would prefer to use rcu_access_pointer() instead of
> rcu_dereference_protected() to avoid pitfall, and it reflects what I was
> expecting:
>
> --- a/security/landlock/fs.c
> +++ b/security/landlock/fs.c
> @@ -117,9 +117,7 @@ static struct landlock_object
> *get_inode_object(struct inode *const inode)
>
>         /* Protects against concurrent get_inode_object() calls. */
>         spin_lock(&inode->i_lock);
> -       object = rcu_dereference_protected(inode_sec->object,
> -                       lockdep_is_held(&inode->i_lock));
> -       if (unlikely(object)) {
> +       if (unlikely(rcu_access_pointer(inode_sec->object))) {
>                 /* Someone else just created the object, bail out and
> retry. */
>                 spin_unlock(&inode->i_lock);
>                 kfree(new_object);

Ah, yeah, that should work. I had forgotten about rcu_access_pointer().

> But I'm not sure about your proposition to move the NULL store in
> release_inode() back up. Do you mean to add back the inode lock in
> release_inode() like this?
>
> --- a/security/landlock/fs.c
> +++ b/security/landlock/fs.c
> @@ -59,16 +59,12 @@ static void release_inode(struct landlock_object
> *const object)
>          * Makes sure that if the filesystem is concurrently unmounted,
>          * hook_sb_delete() will wait for us to finish iput().
>          */
> +       spin_lock(&inode->i_lock);
>         sb = inode->i_sb;
>         atomic_long_inc(&landlock_superblock(sb)->inode_refs);
>         spin_unlock(&object->lock);
> -       /*
> -        * Because object->underobj was not NULL, hook_sb_delete() and
> -        * get_inode_object() guarantee that it is safe to reset
> -        * landlock_inode(inode)->object while it is not NULL.  It is therefore
> -        * not necessary to lock inode->i_lock.
> -        */
>         rcu_assign_pointer(landlock_inode(inode)->object, NULL);
> +       spin_unlock(&inode->i_lock);
>         /*
>          * Now, new rules can safely be tied to @inode with get_inode_object().
>          */
>
>
> I would prefer to avoid nested locks if it is not necessary though.

Hm, yeah, you have a point there.

Doing it locklessly does make the locking rules a little complicated
though, and you'll have to update the comment inside struct
landlock_inode_security. At the moment, it says:

* @object: Weak pointer to an allocated object.  All writes (i.e.
* creating a new object or removing one) are protected by the
* underlying inode->i_lock.  Disassociating @object from the inode is
* additionally protected by @object->lock, from the time @object's
* usage refcount drops to zero to the time this pointer is nulled out.

which isn't true anymore.




[Index of Archives]     [Kernel Newbies]     [Security]     [Netfilter]     [Bugtraq]     [Linux FS]     [Yosemite Forum]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Linux RAID]     [Samba]     [Video 4 Linux]     [Device Mapper]     [Linux Resources]

  Powered by Linux