[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Date: Fri, 7 Jul 2023 08:22:43 +0200
From: Oliver Hartkopp <socketcan@...tkopp.net>
To: "Ziyang Xuan (William)" <william.xuanziyang@...wei.com>,
mkl@...gutronix.de, davem@...emloft.net, edumazet@...gle.com,
kuba@...nel.org, pabeni@...hat.com, linux-can@...r.kernel.org,
netdev@...r.kernel.org, penguin-kernel@...ove.SAKURA.ne.jp
Subject: Re: [PATCH net] can: raw: fix receiver memory leak
On 07.07.23 03:59, Ziyang Xuan (William) wrote:
>> On 06.07.23 14:48, Ziyang Xuan (William) wrote:
>>
>> (..)
>>
>>>>> }
>>>>> out:
>>>>> release_sock(sk);
>>>>> + rtnl_unlock();
>>>>
>>>> Would it also fix the issue when just adding the rtnl_locks to raw_bind() and raw_release() as suggested by you?
>>>
>>> This patch just add rtnl_lock in raw_bind() and raw_release(). raw_setsockopt() has rtnl_lock before this. raw_notify()
>>> is under rtnl_lock. My patch has been tested and solved the issue before send. I don't know if it answered your doubts.
>>
>> My question was whether adding rtnl_locks to raw_bind() and raw_release() would be enough to fix the issue.
>>
>> Without introducing the additional ro->dev element!?
>
> Understand. Just add rtnl_lock to raw_bind() and raw_release() can not fix the issue. I tested.
>
> We should understand that unregister a net device is divided into two stages generally.
> Fistly, call unregister_netdevice_many() to remove net_dev from device list and add
> net_dev to net_todo_list. Secondly, free net_dev in netdev_run_todo().
>
> In my issue. Firstly, unregister_netdevice_many() removed can_dev from device
> list and added can_dev to net_todo_list. Then got NULL by dev_get_by_index()
> and receivers in dev_rcv_lists would not be freed in raw_release().
> After raw_release(), ro->bound would be set 0. When NETDEV_UNREGISTER event
> arrived raw_notify(), receivers in dev_rcv_lists would not be freed too
> because ro->bound was already 0. Thus receivers in dev_rcv_lists would be leaked.
Thanks for the clarification and the testing!
I really assumed rtnl_lock would do this job and also protect the entire
sequence starting with unregister_netdevice_many() !?!
Looking forward to the V2 patch then.
Many thanks,
Oliver
>
> cpu0 cpu1
> unregister_netdevice_many(can_dev)
> unlist_netdevice(can_dev) // dev_get_by_index() return NULL after this
> net_set_todo(can_dev)
> raw_release(can_socket)
> dev = dev_get_by_index(, ro->ifindex); // dev == NULL
> if (dev) { // receivers in dev_rcv_lists not free because dev is NULL
> raw_disable_allfilters(, dev, );
> dev_put(dev);
> }
> ...
> ro->bound = 0;
> ...
>
> netdev_wait_allrefs_any()
> call_netdevice_notifiers(NETDEV_UNREGISTER, )
> raw_notify(, NETDEV_UNREGISTER, )
> if (ro->bound) // invalid because ro->bound has been set 0
> raw_disable_allfilters(, dev, ); // receivers in dev_rcv_lists will never be freed
>
>
> Thanks,
> William Xuan
>
>>
>> Best regards,
>> Oliver
>> .
Powered by blists - more mailing lists