Re: [PATCH bpf-next v4 2/6] netfilter: bpf: Support BPF_F_NETFILTER_IP_DEFRAG in netfilter link

[Alexei Starovoitov]

On Wed, Jul 12, 2023 at 9:33 PM Daniel Xu <dxu@xxxxxxxxx> wrote:
>
> On Wed, Jul 12, 2023 at 06:26:13PM -0700, Alexei Starovoitov wrote:
> > On Wed, Jul 12, 2023 at 6:22 PM Daniel Xu <dxu@xxxxxxxxx> wrote:
> > >
> > > Hi Alexei,
> > >
> > > On Wed, Jul 12, 2023 at 05:43:49PM -0700, Alexei Starovoitov wrote:
> > > > On Wed, Jul 12, 2023 at 4:44 PM Daniel Xu <dxu@xxxxxxxxx> wrote:
> > > > > +#if IS_ENABLED(CONFIG_NF_DEFRAG_IPV6)
> > > > > +       case NFPROTO_IPV6:
> > > > > +               rcu_read_lock();
> > > > > +               v6_hook = rcu_dereference(nf_defrag_v6_hook);
> > > > > +               if (!v6_hook) {
> > > > > +                       rcu_read_unlock();
> > > > > +                       err = request_module("nf_defrag_ipv6");
> > > > > +                       if (err)
> > > > > +                               return err < 0 ? err : -EINVAL;
> > > > > +
> > > > > +                       rcu_read_lock();
> > > > > +                       v6_hook = rcu_dereference(nf_defrag_v6_hook);
> > > > > +                       if (!v6_hook) {
> > > > > +                               WARN_ONCE(1, "nf_defrag_ipv6_hooks bad registration");
> > > > > +                               err = -ENOENT;
> > > > > +                               goto out_v6;
> > > > > +                       }
> > > > > +               }
> > > > > +
> > > > > +               err = v6_hook->enable(link->net);
> > > >
> > > > I was about to apply, but luckily caught this issue in my local test:
> > > >
> > > > [   18.462448] BUG: sleeping function called from invalid context at
> > > > kernel/locking/mutex.c:283
> > > > [   18.463238] in_atomic(): 0, irqs_disabled(): 0, non_block: 0, pid:
> > > > 2042, name: test_progs
> > > > [   18.463927] preempt_count: 0, expected: 0
> > > > [   18.464249] RCU nest depth: 1, expected: 0
> > > > [   18.464631] CPU: 15 PID: 2042 Comm: test_progs Tainted: G
> > > > O       6.4.0-04319-g6f6ec4fa00dc #4896
> > > > [   18.465480] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996),
> > > > BIOS rel-1.12.0-59-gc9ba5276e321-prebuilt.qemu.org 04/01/2014
> > > > [   18.466531] Call Trace:
> > > > [   18.466767]  <TASK>
> > > > [   18.466975]  dump_stack_lvl+0x32/0x40
> > > > [   18.467325]  __might_resched+0x129/0x180
> > > > [   18.467691]  mutex_lock+0x1a/0x40
> > > > [   18.468057]  nf_defrag_ipv4_enable+0x16/0x70
> > > > [   18.468467]  bpf_nf_link_attach+0x141/0x300
> > > > [   18.468856]  __sys_bpf+0x133e/0x26d0
> > > >
> > > > You cannot call mutex under rcu_read_lock.
> > >
> > > Whoops, my bad. I think this patch should fix it:
> > >
> > > ```
> > > From 7e8927c44452db07ddd7cf0e30bb49215fc044ed Mon Sep 17 00:00:00 2001
> > > Message-ID: <7e8927c44452db07ddd7cf0e30bb49215fc044ed.1689211250.git.dxu@xxxxxxxxx>
> > > From: Daniel Xu <dxu@xxxxxxxxx>
> > > Date: Wed, 12 Jul 2023 19:17:35 -0600
> > > Subject: [PATCH] netfilter: bpf: Don't hold rcu_read_lock during
> > >  enable/disable
> > >
> > > ->enable()/->disable() takes a mutex which can sleep. You can't sleep
> > > during RCU read side critical section.
> > >
> > > Our refcnt on the module will protect us from ->enable()/->disable()
> > > from going away while we call it.
> > >
> > > Signed-off-by: Daniel Xu <dxu@xxxxxxxxx>
> > > ---
> > >  net/netfilter/nf_bpf_link.c | 10 ++++++++--
> > >  1 file changed, 8 insertions(+), 2 deletions(-)
> > >
> > > diff --git a/net/netfilter/nf_bpf_link.c b/net/netfilter/nf_bpf_link.c
> > > index 77ffbf26ba3d..79704cc596aa 100644
> > > --- a/net/netfilter/nf_bpf_link.c
> > > +++ b/net/netfilter/nf_bpf_link.c
> > > @@ -60,9 +60,12 @@ static int bpf_nf_enable_defrag(struct bpf_nf_link *link)
> > >                         goto out_v4;
> > >                 }
> > >
> > > +               rcu_read_unlock();
> > >                 err = v4_hook->enable(link->net);
> > >                 if (err)
> > >                         module_put(v4_hook->owner);
> > > +
> > > +               return err;
> > >  out_v4:
> > >                 rcu_read_unlock();
> > >                 return err;
> > > @@ -92,9 +95,12 @@ static int bpf_nf_enable_defrag(struct bpf_nf_link *link)
> > >                         goto out_v6;
> > >                 }
> > >
> > > +               rcu_read_unlock();
> > >                 err = v6_hook->enable(link->net);
> > >                 if (err)
> > >                         module_put(v6_hook->owner);
> > > +
> > > +               return err;
> > >  out_v6:
> > >                 rcu_read_unlock();
> > >                 return err;
> > > @@ -114,11 +120,11 @@ static void bpf_nf_disable_defrag(struct bpf_nf_link *link)
> > >         case NFPROTO_IPV4:
> > >                 rcu_read_lock();
> > >                 v4_hook = rcu_dereference(nf_defrag_v4_hook);
> > > +               rcu_read_unlock();
> > >                 if (v4_hook) {
> > >                         v4_hook->disable(link->net);
> > >                         module_put(v4_hook->owner);
> > >                 }
> > > -               rcu_read_unlock();
> > >
> > >                 break;
> > >  #endif
> > > @@ -126,11 +132,11 @@ static void bpf_nf_disable_defrag(struct bpf_nf_link *link)
> > >         case NFPROTO_IPV6:
> > >                 rcu_read_lock();
> > >                 v6_hook = rcu_dereference(nf_defrag_v6_hook);
> > > +               rcu_read_unlock();
> >
> > No. v6_hook is gone as soon as you unlock it.
>
> I think we're protected here by the try_module_get() on the enable path.
> And we only disable defrag if enabling succeeds. The module shouldn't
> be able to deregister its hooks until we call the module_put() later.
>
> I think READ_ONCE() would've been more appropriate but I wasn't sure if
> that was ok given nf_defrag_v(4|6)_hook is written to by
> rcu_assign_pointer() and I was assuming symmetry is necessary.

Why is rcu_assign_pointer() used?
If it's not RCU protected, what is the point of rcu_*() accessors
and rcu_read_lock() ?

In general, the pattern:
rcu_read_lock();
ptr = rcu_dereference(...);
rcu_read_unlock();
ptr->..
is a bug. 100%.