Message-ID: <CAG48ez1arKO3uYzwng8fst-UHkcH6J7YzyHFN+vfXUT2=1HT+w@mail.gmail.com>
Date:   Tue, 23 Mar 2021 01:13:36 +0100
From:   Jann Horn <jannh@...gle.com>
To:     Mickaël Salaün <mic@...ikod.net>
Cc:     James Morris <jmorris@...ei.org>,
        "Serge E . Hallyn" <serge@...lyn.com>,
        Al Viro <viro@...iv.linux.org.uk>,
        Andrew Morton <akpm@...ux-foundation.org>,
        Andy Lutomirski <luto@...capital.net>,
        Anton Ivanov <anton.ivanov@...bridgegreys.com>,
        Arnd Bergmann <arnd@...db.de>,
        Casey Schaufler <casey@...aufler-ca.com>,
        David Howells <dhowells@...hat.com>,
        Jeff Dike <jdike@...toit.com>,
        Jonathan Corbet <corbet@....net>,
        Kees Cook <keescook@...omium.org>,
        Michael Kerrisk <mtk.manpages@...il.com>,
        Richard Weinberger <richard@....at>,
        Shuah Khan <shuah@...nel.org>,
        Vincent Dagonneau <vincent.dagonneau@....gouv.fr>,
        Kernel Hardening <kernel-hardening@...ts.openwall.com>,
        Linux API <linux-api@...r.kernel.org>,
        linux-arch <linux-arch@...r.kernel.org>,
        "open list:DOCUMENTATION" <linux-doc@...r.kernel.org>,
        linux-fsdevel <linux-fsdevel@...r.kernel.org>,
        kernel list <linux-kernel@...r.kernel.org>,
        "open list:KERNEL SELFTEST FRAMEWORK" 
        <linux-kselftest@...r.kernel.org>,
        linux-security-module <linux-security-module@...r.kernel.org>,
        "the arch/x86 maintainers" <x86@...nel.org>,
        Mickaël Salaün <mic@...ux.microsoft.com>
Subject: Re: [PATCH v30 07/12] landlock: Support filesystem access-control

 On Tue, Mar 16, 2021 at 9:43 PM Mickaël Salaün <mic@...ikod.net> 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.)


> +       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;
> +}

Powered by blists - more mailing lists