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);
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.
>
>>>> + if (unlikely(object)) {
>>>> + /* Someone else just created the object, bail out and retry. */
>>>> + spin_unlock(&inode->i_lock);
>>>> + kfree(new_object);
>>>> +
>>>> + rcu_read_lock();
>>>> + goto retry;
>>>> + }
>>>> +
>>>> + rcu_assign_pointer(inode_sec->object, new_object);
>>>> + /*
>>>> + * @inode will be released by hook_sb_delete() on its superblock
>>>> + * shutdown.
>>>> + */
>>>> + ihold(inode);
>>>> + spin_unlock(&inode->i_lock);
>>>> + return new_object;
>>>> +}
