Re: [PATCH net,v2] team: fix possible deadlock in team_port_change_check
From: Jeongjun Park
Date: Fri Aug 02 2024 - 21:37:11 EST
> Jiri Pirko wrote:
>
> Fri, Aug 02, 2024 at 06:25:31PM CEST, aha310510@xxxxxxxxx wrote:
>> Eric Dumazet wrote:
>>>
>>>> On Fri, Aug 2, 2024 at 5:00 PM Jeongjun Park <aha310510@xxxxxxxxx> wrote:
>>>>>
>>
>> [..]
>>
>> @@ -2501,6 +2470,11 @@ int team_nl_options_get_doit(struct sk_buff *skb, struct genl_info *info)
>> int err;
>> LIST_HEAD(sel_opt_inst_list);
>>
>> + if (!rtnl_is_locked()) {
>
> This is completely wrong, other thread may hold the lock.
>
>
>> + rtnl_lock();
>
> NACK! I wrote it in the other thread. Don't take rtnl for get options
> command. It is used for repeated fetch of stats. It's read only. Should
> be converted to RCU.
>
I see. But, in the current, when called through the following path:
team_nl_send_event_options_get()->
team_nl_send_options_get()->
team_nl_fill_one_option_get()
, it was protected through rtnl. Does this mean that rcu should be
used instead of rtnl in this case as well?
> Why are you so obsessed by this hypothetical syzcaller bug? Are you
> hitting this in real? If not, please let it go. I will fix it myself
> when I find some spare cycles.
Sorry for the inconvenience, but I don't want to give up on this bug
so easily since it is a valid bug that we have started analyzing
anyway and the direction of how to fix it is clear. I hope you
understand and I will send you a patch that uses rcu instead
of rtnl soon.
Regards,
Jeongjun Park