Re: [PATCH 03/11] drbd_transport_rdma: put kref for cm in dtr_path_established in error path

在 2024/7/1 星期一 上午 10:07, Dongsheng Yang 写道:
> 
> 
> 在 2024/6/28 星期五 下午 5:40, Philipp Reisner 写道:
>> Hello Dongsheng,
>>
>> Please add more information why you think this change fixes a bug.
>> Have you experienced a leak of cm structs?
>> We got a RDMA_CM_EVENT_ESTABLISHED event. Even if DRBD does not do
>> anything with this cm, we sill expect a RDMA_CM_EVENT_DISCONNECTED in
>> the future. Is a problem in the handling of the disconnect?
> 
> If dtr_path_established() go into this branch, it will not 
> schedule_work(&cm->establish_work);
> 
> That means path->cm->state = DSM_CONNECTED; will not be done in 
> dtr_path_established_work_fn(), so __dtr_disconnect_path() will not call 
> rdma_disconnect(). That means this reference will never be put.

let me consider this  example:
a) rdma_connect() called and RDMA_CM_EVENT_ESTABLISHED received.

b) network failure and dtr_path_established() go into error path.

c) establish_work will not be scheduled.

d) drbdadm down test will hang because cm ref is not put.
>>
>> best regards,
>>   Philipp
>>
>> On Mon, Jun 24, 2024 at 9:28 AM zhengbing.huang
>> <[email protected]> wrote:
>>>
>>> From: Dongsheng Yang <[email protected]>
>>>
>>> Signed-off-by: Dongsheng Yang <[email protected]>
>>> ---
>>>   drbd/drbd_transport_rdma.c | 1 +
>>>   1 file changed, 1 insertion(+)
>>>
>>> diff --git a/drbd/drbd_transport_rdma.c b/drbd/drbd_transport_rdma.c
>>> index cfbae0e78..eccd0c6ce 100644
>>> --- a/drbd/drbd_transport_rdma.c
>>> +++ b/drbd/drbd_transport_rdma.c
>>> @@ -922,6 +922,7 @@ static void dtr_path_established(struct dtr_cm *cm)
>>>                          atomic_set(&cs->active_state, PCS_INACTIVE);
>>>                          wake_up(&cs->wq);
>>>                  }
>>> +               kref_put(&cm->kref, dtr_destroy_cm);
>>>                  return;
>>>          }
>>>
>>> -- 
>>> 2.27.0
>>>