Re: [PATCH bpf-next v5 1/6] bpf/helpers: introduce sleepable bpf_timers

[Alexei Starovoitov]

On Wed, Apr 3, 2024 at 11:50 AM Alexei Starovoitov
<alexei.starovoitov@xxxxxxxxx> wrote:
>
> On Wed, Mar 27, 2024 at 10:02 AM Benjamin Tissoires
> <benjamin.tissoires@xxxxxxxxxx> wrote:
> > > >                 goto out;
> > > >         }
> > > > +       spin_lock(&t->sleepable_lock);
> > > >         drop_prog_refcnt(t);
> > > > +       spin_unlock(&t->sleepable_lock);
> > >
> > > this also looks odd.
> >
> > I basically need to protect "t->prog = NULL;" from happening while
> > bpf_timer_work_cb is setting up the bpf program to be run.
>
> Ok. I think I understand the race you're trying to fix.
> The bpf_timer_cancel_and_free() is doing
> cancel_work()
> and proceeds with
> kfree_rcu(t, rcu);
>
> That's the only race and these extra locks don't help.
>
> The t->prog = NULL is nothing to worry about.
> The bpf_timer_work_cb() might still see callback_fn == NULL
> "when it's being setup" and it's ok.
> These locks don't help that.
>
> I suggest to drop sleepable_lock everywhere.
> READ_ONCE of callback_fn in bpf_timer_work_cb() is enough.
> Add rcu_read_lock_trace() before calling bpf prog.
>
> The race to fix is above 'cancel_work + kfree_rcu'
> since kfree_rcu might free 'struct bpf_hrtimer *t'
> while the work is pending and work_queue internal
> logic might UAF struct work_struct work.
> By the time it may luckily enter bpf_timer_work_cb() it's too late.
> The argument 'struct work_struct *work' might already be freed.
>
> To fix this problem, how about the following:
> don't call kfree_rcu and instead queue the work to free it.
> After cancel_work(&t->work); the work_struct can be reused.
> So set it up to call "freeing callback" and do
> schedule_work(&t->work);
>
> There is a big assumption here that new work won't be
> executed before cancelled work completes.
> Need to check with wq experts.
>
> Another approach is to do something smart with
> cancel_work() return code.
> If it returns true set a flag inside bpf_hrtimer and
> make bpf_timer_work_cb() free(t) after bpf prog finishes.

Looking through wq code... I think I have to correct myself.
cancel_work and immediate free is probably fine from wq pov.
It has this comment:
        worker->current_func(work);
        /*
         * While we must be careful to not use "work" after this, the trace
         * point will only record its address.
         */
        trace_workqueue_execute_end(work, worker->current_func);

the bpf_timer_work_cb() might still be running bpf prog.
So it shouldn't touch 'struct bpf_hrtimer *t' after bpf prog returns,
since kfree_rcu(t, rcu); could have freed it by then.
There is also this code in net/rxrpc/rxperf.c
        cancel_work(&call->work);
        kfree(call);

So it looks like it's fine to drop sleepable_lock,
add rcu_read_lock_trace() and things should be ok.