Re: [PATCH bpf-next v2 02/11] net, sk_msg: Annotate lockless access to sk_prot on clone

[Jakub Sitnicki <jakub@cloudflare.com>]

On Mon, Jan 20, 2020 at 07:11 PM CET, Jakub Sitnicki wrote:
> On Mon, Jan 20, 2020 at 06:00 PM CET, John Fastabend wrote:
>> Jakub Sitnicki wrote:
>>> On Sun, Jan 12, 2020 at 12:14 AM CET, John Fastabend wrote:
>>> > Jakub Sitnicki wrote:
>>> >> sk_msg and ULP frameworks override protocol callbacks pointer in
>>> >> sk->sk_prot, while TCP accesses it locklessly when cloning the listening
>>> >> socket.
>>> >>
>>> >> Once we enable use of listening sockets with sockmap (and hence sk_msg),
>>> >> there can be shared access to sk->sk_prot if socket is getting cloned while
>>> >> being inserted/deleted to/from the sockmap from another CPU. Mark the
>>> >> shared access with READ_ONCE/WRITE_ONCE annotations.
>>> >>
>>> >> Signed-off-by: Jakub Sitnicki <jakub@cloudflare.com>
>>> >
>>> > In sockmap side I fixed this by wrapping the access in a lock_sock[0]. So
>>> > Do you think this is still needed with that in mind? The bpf_clone call
>>> > is using sk_prot_creater and also setting the newsk's proto field. Even
>>> > if the listening parent sock was being deleted in parallel would that be
>>> > a problem? We don't touch sk_prot_creator from the tear down path. I've
>>> > only scanned the 3..11 patches so maybe the answer is below. If that is
>>> > the case probably an improved commit message would be helpful.
>>>
>>> I think it is needed. Not because of tcp_bpf_clone or that we access
>>> listener's sk_prot_creator from there, if I'm grasping your question.
>>>
>>> Either way I'm glad this came up. Let's go though my reasoning and
>>> verify it. tcp stack accesses the listener sk_prot while cloning it:
>>>
>>> tcp_v4_rcv
>>>   sk = __inet_lookup_skb(...)
>>>   tcp_check_req(sk)
>>>     inet_csk(sk)->icsk_af_ops->syn_recv_sock
>>>       tcp_v4_syn_recv_sock
>>>         tcp_create_openreq_child
>>>           inet_csk_clone_lock
>>>             sk_clone_lock
>>>               READ_ONCE(sk->sk_prot)
>>>
>>> It grabs a reference to the listener, but doesn't grab the sk_lock.
>>>
>>> On another CPU we can be inserting/removing the listener socket from the
>>> sockmap and writing to its sk_prot. We have the update and the remove
>>> path:
>>>
>>> sock_map_ops->map_update_elem
>>>   sock_map_update_elem
>>>     sock_map_update_common
>>>       sock_map_link_no_progs
>>>         tcp_bpf_init
>>>           tcp_bpf_update_sk_prot
>>>             sk_psock_update_proto
>>>               WRITE_ONCE(sk->sk_prot, ops)
>>>
>>> sock_map_ops->map_delete_elem
>>>   sock_map_delete_elem
>>>     __sock_map_delete
>>>      sock_map_unref
>>>        sk_psock_put
>>>          sk_psock_drop
>>>            sk_psock_restore_proto
>>>              tcp_update_ulp
>>>                WRITE_ONCE(sk->sk_prot, proto)
>>>
>>> Following the guidelines from KTSAN project [0], sk_prot looks like a
>>> candidate for annotating it. At least on these 3 call paths.
>>>
>>> If that sounds correct, I can add it to the patch description.
>>>
>>> Thanks,
>>> -jkbs
>>>
>>> [0] https://github.com/google/ktsan/wiki/READ_ONCE-and-WRITE_ONCE
>>
>> Hi Jakub, can push this to bpf tree as well? There is another case
>> already in-kernel where this is needed. If the map is removed while
>> a recvmsg is in flight.
>>
>>  tcp_bpf_recvmsg()
>>   psock = sk_psock_get(sk)                         <- refcnt 2
>>   lock_sock(sk);
>>   ...
>>                                   sock_map_free()  <- refcnt 1
>>   release_sock(sk)
>>   sk_psock_put()                                   <- refcnt 0
>>
>> Then can you add this diff as well I got a bit too carried away
>> with that. If your busy I can do it as well if you want. Thanks!
>
> Hi John, I get the race between map_free and tcp_bpf_recvmsg, and how we
> end up dropping psock on a path where we don't hold the sock lock. What
> a rare case, since we don't destory maps that often usually.
>
> However, I'm not sure I follow where shared lockless access to
> sk->sk_prot is in this case?
>
> Perhaps between drop path:
>
> sk_psock_put
>   sk_psock_drop
>     sk_psock_restore_proto
>       WRITE_ONCE(sk->sk_prot, proto)
>
> ... and update path where we grab sk_callback_lock a little too late,
> that is after updating the proto?
>
> sock_map_update_common
>   sock_map_link
>     tcp_bpf_init
>       tcp_bpf_update_sk_prot
>         sk_psock_update_proto
>           WRITE_ONCE(sk->sk_prot, ops)
>
> I'm getting v3 ready to post, so happy to help you spin these bits.
> I'll need to do it with a fresh head tomorrow, though.
>
> If I don't see any patches from you hit the ML, I'll split out the
> chunks that annotate sk_prot access in sk_psock_{retore,update}_proto
> and post them together with the revert you suggested below.

I've sent out the partial revert you wanted:

https://lore.kernel.org/netdev/20200121123147.706666-1-jakub@cloudflare.com/T/#u

But otherwise didn't manage to convince myself that we need to annotate
access to sk_prot with READ/WRITE_ONCE in sk_psock_update/restore_proto.

Instead, I believe we might need to the extend critical section that
grabs sk_callback_lock in sock_map_link over tcp_bpf_init/reinit() to
serialize the writers.

Unless I'm missing the point here and you had some other race in mind?

-jkbs