Received-SPF: pass (google.com: domain of linux-kernel+bounces-214327-linux.lists.archive=gmail.com@vger.kernel.org designates 2604:1380:45d1:ec00::1 as permitted sender) client-ip=2604:1380:45d1:ec00::1;
Message-ID: <522c0b17-c515-475d-8224-637ca0eaf6a2@I-love.SAKURA.ne.jp>
Date: Fri, 14 Jun 2024 13:00:30 +0900
Precedence: bulk
MIME-Version: 1.0
User-Agent: Mozilla Thunderbird
Subject: Re: [net/sched] Question: Locks for clearing ERR_PTR() value from
 idrinfo->action_idr ?
To: Pedro Tammela <pctammela@mojatatu.com>,
        Jamal Hadi Salim <jhs@mojatatu.com>,
        Cong Wang <xiyou.wangcong@gmail.com>, Jiri Pirko <jiri@resnulli.us>
Cc: "David S. Miller" <davem@davemloft.net>,
        Eric Dumazet <edumazet@google.com>, Jakub Kicinski <kuba@kernel.org>,
        Paolo Abeni <pabeni@redhat.com>,
        Network Development
 <netdev@vger.kernel.org>,
        LKML <linux-kernel@vger.kernel.org>
References: <8d61200a-a739-4200-a8a3-5386a834d44f@I-love.SAKURA.ne.jp>
 <de8e2709-8d7f-4e51-a4a4-35bad72ba136@mojatatu.com>
Content-Language: en-US
From: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
In-Reply-To: <de8e2709-8d7f-4e51-a4a4-35bad72ba136@mojatatu.com>
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 7bit

On 2024/06/14 11:47, Pedro Tammela wrote:
> On 13/06/2024 21:58, Tetsuo Handa wrote:
>>
>> Is there a possibility that tcf_idr_check_alloc() is called without holding
>> rtnl_mutex?
> 
> There is, but not in the code path of this reproducer.
> 
>> If yes, adding a sleep before "goto again;" would help. But if no,
>> is this a sign that some path forgot to call tcf_idr_{cleanup,insert_many}() ?
> 
> The reproducer is sending a new action message with 2 actions.
> Actions are committed to the idr after processing in order to make them visible
> together and after any errors are caught.
> 
> The bug happens when the actions in the message refer to the same index. Since
> the first processing succeeds, adding -EBUSY to the index, the second processing,
> which references the same index, will loop forever.
> 
> After the change to rely on RCU for this check, instead of the idr lock, the hangs
> became more noticeable to syzbot since now it's hanging a system-wide lock.

Thank you for explanation. Then, what type of sleep do we want?

schedule_timeout_uninteruptible(1)
(based on an assumption that conflict will be solved shortly) ?

wait_event()/wake_up_all() using one global waitqueue
(based on an assumption that conflict is rare) ?

wait_event()/wake_up_all() using per struct tcf_idrinfo waitqueue
(based on an assumption that conflict might not be rare) ?