On Thu, 19 Jan 2023 17:37:11 -0800 Suren Baghdasaryan <surenb@xxxxxxxxxx> wrote:
> On Thu, Jan 19, 2023 at 5:31 PM Hillf Danton <hdanton@xxxxxxxx> wrote:
> > On Thu, 19 Jan 2023 13:01:42 -0800 Suren Baghdasaryan <surenb@xxxxxxxxxx> wrote:
> > >
> > > Hi Folks,
> > > I spent some more time digging into the details and this is what's
> > > happening. When we call rmdir to delete the cgroup with the pressure
> > > file being epoll'ed, roughly the following call chain happens in the
> > > context of the shell process:
> > >
> > > do_rmdir
> > > cgroup_rmdir
> > > kernfs_drain_open_files
> > > cgroup_file_release
> > > cgroup_pressure_release
> > > psi_trigger_destroy
> > >
> > > Later on in the context of our reproducer, the last fput() is called
> > > causing wait queue removal:
> > >
> > > fput
> > > ep_eventpoll_release
> > > ep_free
> > > ep_remove_wait_queue
> > > remove_wait_queue
> > >
> > > By this time psi_trigger_destroy() already destroyed the trigger's
> > > waitqueue head and we hit UAF.
> > > I think the conceptual problem here (or maybe that's by design?) is
> > > that cgroup_file_release() is not really tied to the file's real
> > > lifetime (when the last fput() is issued). Otherwise fput() would call
> > > eventpoll_release() before f_op->release() and the order would be fine
> > > (we would remove the wait queue first in eventpoll_release() and then
> > > f_op->release() would cause trigger's destruction).
> >
> > eventpoll_release
> > eventpoll_release_file
> > ep_remove
> > ep_unregister_pollwait
> > ep_remove_wait_queue
> >
>
> Yes but fput() calls eventpoll_release() *before* f_op->release(), so
> waitqueue_head would be removed before trigger destruction.
Then check if file is polled before destroying trigger.
+++ b/kernel/sched/psi.c
@@ -1529,6 +1529,7 @@ static int psi_fop_release(struct inode
{
struct seq_file *seq = file->private_data;
+ eventpoll_release_file(file);
psi_trigger_destroy(seq->private);
return single_release(inode, file);
}
