Re: [PATCH] bpf, sockmap: defer sk_psock_free_link() using RCU

[Jakub Sitnicki <jakub@cloudflare.com>]

On Wed, May 22, 2024 at 07:33 PM +08, Hillf Danton wrote:
> On Wed, 22 May 2024 11:50:49 +0200 Jakub Sitnicki <jakub@cloudflare.com>
> On Wed, May 22, 2024 at 06:59 AM +08, Hillf Danton wrote:
>> > On Tue, 21 May 2024 08:38:52 -0700 Alexei Starovoitov <alexei.starovoitov@gmail.com>
>> >> On Sun, May 12, 2024 at 12:22=E2=80=AFAM Tetsuo Handa <penguin-kernel@i-love.sakura.ne.jp> wrote:
>> >> > --- a/net/core/sock_map.c
>> >> > +++ b/net/core/sock_map.c
>> >> > @@ -142,6 +142,7 @@ static void sock_map_del_link(struct sock *sk,
>> >> >         bool strp_stop =3D false, verdict_stop =3D false;
>> >> >         struct sk_psock_link *link, *tmp;
>> >> >
>> >> > +       rcu_read_lock();
>> >> >         spin_lock_bh(&psock->link_lock);
>> >> 
>> >> I think this is incorrect.
>> >> spin_lock_bh may sleep in RT and it won't be safe to do in rcu cs.
>> >
>> > Could you specify why it won't be safe in rcu cs if you are right?
>> > What does rcu look like in RT if not nothing?
>> 
>> RCU readers can't block, while spinlock RT doesn't disable preemption.
>> 
>> https://docs.kernel.org/RCU/rcu.html
>> https://docs.kernel.org/locking/locktypes.html#spinlock-t-and-preempt-rt
>> 
>> I've finally gotten around to testing proposed fix that just disallows
>> map_delete_elem on sockmap/sockhash from BPF tracing progs
>> completely. This should put an end to this saga of syzkaller reports.
>> 
>> https://lore.kernel.org/all/87jzjnxaqf.fsf@cloudflare.com/
>> 
> The locking info syzbot reported [2] suggests a known issue that like Alexei
> you hit the send button earlier than expected.
>
> 4 locks held by syz-executor361/5090:
>  #0: ffffffff8e334d20 (rcu_read_lock){....}-{1:2}, at: rcu_lock_acquire include/linux/rcupdate.h:329 [inline]
>  #0: ffffffff8e334d20 (rcu_read_lock){....}-{1:2}, at: rcu_read_lock include/linux/rcupdate.h:781 [inline]
>  #0: ffffffff8e334d20 (rcu_read_lock){....}-{1:2}, at: map_delete_elem+0x388/0x5e0 kernel/bpf/syscall.c:1695
>  #1: ffff88807b2af8f8 (&htab->buckets[i].lock){+...}-{2:2}, at: spin_lock_bh include/linux/spinlock.h:356 [inline]
>  #1: ffff88807b2af8f8 (&htab->buckets[i].lock){+...}-{2:2}, at: sock_hash_delete_elem+0x17c/0x400 net/core/sock_map.c:945
>  #2: ffff88801c2a4290 (&psock->link_lock){+...}-{2:2}, at: spin_lock_bh include/linux/spinlock.h:356 [inline]
>  #2: ffff88801c2a4290 (&psock->link_lock){+...}-{2:2}, at: sock_map_del_link net/core/sock_map.c:145 [inline]
>  #2: ffff88801c2a4290 (&psock->link_lock){+...}-{2:2}, at: sock_map_unref+0xcc/0x5e0 net/core/sock_map.c:180
>  #3: ffffffff8e334d20 (rcu_read_lock){....}-{1:2}, at: rcu_lock_acquire include/linux/rcupdate.h:329 [inline]
>  #3: ffffffff8e334d20 (rcu_read_lock){....}-{1:2}, at: rcu_read_lock include/linux/rcupdate.h:781 [inline]
>  #3: ffffffff8e334d20 (rcu_read_lock){....}-{1:2}, at: __bpf_trace_run kernel/trace/bpf_trace.c:2380 [inline]
>  #3: ffffffff8e334d20 (rcu_read_lock){....}-{1:2}, at: bpf_trace_run2+0x114/0x420 kernel/trace/bpf_trace.c:2420
>
> [2] https://lore.kernel.org/all/000000000000d0b87206170dd88f@google.com/
>
>
> If CONFIG_PREEMPT_RCU=y rcu_read_lock() does not disable
> preemption. This is even true for !RT kernels with CONFIG_PREEMPT=y
>
> [3] Subject: Re: [patch 30/63] locking/spinlock: Provide RT variant
> https://lore.kernel.org/all/874kc6rizr.ffs@tglx/

That locking issue is related to my earlier, as it turned out -
incomplete, fix:

https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=ff91059932401894e6c86341915615c5eb0eca48

We don't expect map_delete_elem to be called from map_update_elem for
sockmap/sockhash, but that is what syzkaller started doing by attaching
BPF tracing progs which call map_delete_elem.