From: Hillf Danton <hdanton@sina.com>
    
Once a key's reference count has been reduced to 0, the garbage collector
thread may destroy it at any time and so key_put() is not allowed to touch
the key after that point.  The most key_put() is normally allowed to do is
to touch key_gc_work as that's a static global variable.

However, in an effort to speed up the reclamation of quota, this is now
done in key_put() once the key's usage is reduced to 0 - but now the code
is looking at the key after the deadline, which is forbidden.

Fix this on an expedited basis[*] by taking a ref on the key->user struct
and caching the key length before dropping the refcount so that we can
reduce the quota afterwards if we dropped the last ref.

[*] This is going to hurt key_put() performance, so a better way is
probably necessary, such as sticking the dead key onto a queue for the
garbage collector to pick up rather than having it scan the serial number
registry.

Fixes: 9578e327b2b4 ("keys: update key quotas in key_put()")
Reported-by: syzbot+6105ffc1ded71d194d6d@syzkaller.appspotmail.com
Tested-by: syzbot+6105ffc1ded71d194d6d@syzkaller.appspotmail.com
Suggested-by: Hillf Danton <hdanton@sina.com>
Signed-off-by: David Howells <dhowells@redhat.com>
cc: Jarkko Sakkinen <jarkko@kernel.org>
cc: Kees Cook <kees@kernel.org>
cc: keyrings@vger.kernel.org
cc: stable@vger.kernel.org # v6.10+
---
 security/keys/key.c |   19 ++++++++++++++-----
 1 file changed, 14 insertions(+), 5 deletions(-)

diff --git a/security/keys/key.c b/security/keys/key.c
index 3d7d185019d3..1e6028492355 100644
--- a/security/keys/key.c
+++ b/security/keys/key.c
@@ -645,21 +645,30 @@ EXPORT_SYMBOL(key_reject_and_link);
  */
 void key_put(struct key *key)
 {
+	int quota_flag;
+	unsigned short len;
+	struct key_user *user;
+
 	if (key) {
 		key_check(key);
 
+		quota_flag = test_bit(KEY_FLAG_IN_QUOTA, &key->flags);
+		len = key->quotalen;
+		user = key->user;
+		refcount_inc(&user->usage);
 		if (refcount_dec_and_test(&key->usage)) {
 			unsigned long flags;
 
 			/* deal with the user's key tracking and quota */
-			if (test_bit(KEY_FLAG_IN_QUOTA, &key->flags)) {
-				spin_lock_irqsave(&key->user->lock, flags);
-				key->user->qnkeys--;
-				key->user->qnbytes -= key->quotalen;
-				spin_unlock_irqrestore(&key->user->lock, flags);
+			if (quota_flag) {
+				spin_lock_irqsave(&user->lock, flags);
+				user->qnkeys--;
+				user->qnbytes -= len;
+				spin_unlock_irqrestore(&user->lock, flags);
 			}
 			schedule_work(&key_gc_work);
 		}
+		key_user_put(user);
 	}
 }
 EXPORT_SYMBOL(key_put);

On 03/18, David Howells wrote:
>
> --- a/security/keys/key.c
> +++ b/security/keys/key.c
> @@ -645,21 +645,30 @@ EXPORT_SYMBOL(key_reject_and_link);
>   */
>  void key_put(struct key *key)
>  {
> +	int quota_flag;
> +	unsigned short len;
> +	struct key_user *user;
> +
>  	if (key) {
>  		key_check(key);
>  
> +		quota_flag = test_bit(KEY_FLAG_IN_QUOTA, &key->flags);
> +		len = key->quotalen;
> +		user = key->user;
> +		refcount_inc(&user->usage);
>  		if (refcount_dec_and_test(&key->usage)) {
>  			unsigned long flags;
>  
>  			/* deal with the user's key tracking and quota */
> -			if (test_bit(KEY_FLAG_IN_QUOTA, &key->flags)) {
> -				spin_lock_irqsave(&key->user->lock, flags);
> -				key->user->qnkeys--;
> -				key->user->qnbytes -= key->quotalen;
> -				spin_unlock_irqrestore(&key->user->lock, flags);
> +			if (quota_flag) {
> +				spin_lock_irqsave(&user->lock, flags);
> +				user->qnkeys--;
> +				user->qnbytes -= len;
> +				spin_unlock_irqrestore(&user->lock, flags);
>  			}
>  			schedule_work(&key_gc_work);
>  		}
> +		key_user_put(user);

Do we really need the unconditional refcount_inc / key_user_put ?

	void key_put(struct key *key)
	{
		if (key) {
			struct key_user *user = NULL;
			unsigned short len;

			key_check(key);

			if (test_bit(KEY_FLAG_IN_QUOTA, &key->flags)) {
				len = key->quotalen;
				user = key->user;
				refcount_inc(&user->usage);
			}

			if (refcount_dec_and_test(&key->usage)) {
				unsigned long flags;

				/* deal with the user's key tracking and quota */
				if (user) {
					spin_lock_irqsave(&user->lock, flags);
					user->qnkeys--;
					user->qnbytes -= len;
					spin_unlock_irqrestore(&user->lock, flags);
				}
				schedule_work(&key_gc_work);
			}

			if (user)
				key_user_put(user);
		}
	}

looks a bit more clear/simple to me...

Oleg.

Either way... I know nothing about security/key, but grep -w key_put
finds

 *  When it is no longer required, the key should be released using::

	void key_put(struct key *key);

    Or::

	void key_ref_put(key_ref_t key_ref);

    These can be called from interrupt context.
                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

in Documentation/security/keys/core.rst

and since key_user_put() takes key_user_lock with irqs enabled,
key_put()->key_user_put() doesn't look correct...

Oleg.

On 03/18, Oleg Nesterov wrote:
>
> On 03/18, David Howells wrote:
> >
> > --- a/security/keys/key.c
> > +++ b/security/keys/key.c
> > @@ -645,21 +645,30 @@ EXPORT_SYMBOL(key_reject_and_link);
> >   */
> >  void key_put(struct key *key)
> >  {
> > +	int quota_flag;
> > +	unsigned short len;
> > +	struct key_user *user;
> > +
> >  	if (key) {
> >  		key_check(key);
> >
> > +		quota_flag = test_bit(KEY_FLAG_IN_QUOTA, &key->flags);
> > +		len = key->quotalen;
> > +		user = key->user;
> > +		refcount_inc(&user->usage);
> >  		if (refcount_dec_and_test(&key->usage)) {
> >  			unsigned long flags;
> >
> >  			/* deal with the user's key tracking and quota */
> > -			if (test_bit(KEY_FLAG_IN_QUOTA, &key->flags)) {
> > -				spin_lock_irqsave(&key->user->lock, flags);
> > -				key->user->qnkeys--;
> > -				key->user->qnbytes -= key->quotalen;
> > -				spin_unlock_irqrestore(&key->user->lock, flags);
> > +			if (quota_flag) {
> > +				spin_lock_irqsave(&user->lock, flags);
> > +				user->qnkeys--;
> > +				user->qnbytes -= len;
> > +				spin_unlock_irqrestore(&user->lock, flags);
> >  			}
> >  			schedule_work(&key_gc_work);
> >  		}
> > +		key_user_put(user);
>
> Do we really need the unconditional refcount_inc / key_user_put ?
>
> 	void key_put(struct key *key)
> 	{
> 		if (key) {
> 			struct key_user *user = NULL;
> 			unsigned short len;
>
> 			key_check(key);
>
> 			if (test_bit(KEY_FLAG_IN_QUOTA, &key->flags)) {
> 				len = key->quotalen;
> 				user = key->user;
> 				refcount_inc(&user->usage);
> 			}
>
> 			if (refcount_dec_and_test(&key->usage)) {
> 				unsigned long flags;
>
> 				/* deal with the user's key tracking and quota */
> 				if (user) {
> 					spin_lock_irqsave(&user->lock, flags);
> 					user->qnkeys--;
> 					user->qnbytes -= len;
> 					spin_unlock_irqrestore(&user->lock, flags);
> 				}
> 				schedule_work(&key_gc_work);
> 			}
>
> 			if (user)
> 				key_user_put(user);
> 		}
> 	}
>
> looks a bit more clear/simple to me...
>
> Oleg.

Oleg Nesterov <oleg@redhat.com> wrote:

> and since key_user_put() takes key_user_lock with irqs enabled,
> key_put()->key_user_put() doesn't look correct...

Meh.  Yeah.  I think it's time to do it the other way (i.e. putting keys to be
destroyed onto an explicit cleanup queue).

David