[PATCH] nvmet-rdma: Release connections synchronously

[Sagi Grimberg]

>>>> @@ -1582,11 +1566,6 @@ static int nvmet_rdma_queue_connect(struct 
>>>> rdma_cm_id *cm_id,
>>>>           goto put_device;
>>>>       }
>>>> -    if (queue->host_qid == 0) {
>>>> -        /* Let inflight controller teardown complete */
>>>> -        flush_workqueue(nvmet_wq);
>>>> -    }
>>>> -
>>>>       ret = nvmet_rdma_cm_accept(cm_id, queue, &event->param.conn);
>>>>       if (ret) {
>>>>           /*
>>>
>>> You could have simply removed this hunk alone to make lockdep quiet on
>>> this, without the need to rework the async queue removal.
>>>
>>> The flush here was added to prevent a reset/connect/disconnect storm
>>> causing the target to run out of resources (which we have seen reports
>>> about in the distant past). What prevents it now?
>>>
>>> And you both reworked the teardown, and still removed the flush, I don't
>>> get why both are needed.
>>
>> Hi Sagi,
>>
>> My understanding is that the above flush_workqueue() call waits for 
>> prior release_work to complete. If the release_work instance is 
>> removed, I don't think that the above flush_workqueue() call is still 
>> necessary.
> 
> I'm wandering if making delete_ctrl synchronous may be a problem
> in some cases, nvmet_fatal_error can be triggered from the
> rdma completion path (rdma core workqueue context). Wondering if there
> may be some inter-dependency there...

For example, nvmet_rdma_comp_error -> nvmet_fatal_error -> delete_ctrl
-> __nvmet_rdma_queue_disconnect -> nvmet_rdma_free_queue ->
nvmet_rdma_destroy_queue_ib -> ib_drain_qp

ib_drain_qp cannot complete because the completion handler is
synchronously deleting the queue, hence cannot reap the drain
completion...