[PATCH v2 7/8] nvme-rdma: fix timeout handler

[Sagi Grimberg]

>> +static void nvme_rdma_complete_timed_out(struct request *rq)
>> +{
>> +    struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
>> +    struct nvme_rdma_queue *queue = req->queue;
>> +    struct nvme_rdma_ctrl *ctrl = queue->ctrl;
>> +
>> +    /* fence other contexts that may complete the command */
>> +    mutex_lock(&ctrl->teardown_lock);
>> +    nvme_rdma_stop_queue(queue);
>> +    if (blk_mq_request_completed(rq))
>> +        goto out;
>> +    nvme_req(rq)->status = NVME_SC_HOST_ABORTED_CMD;
>> +    blk_mq_complete_request(rq);
>> +out:
>> +    mutex_unlock(&ctrl->teardown_lock);
>> +}
>> +
> 
> I believe there should be some comment explaining why it's ok to leave 
> the rdma queue stopped.
> I think it's ok as:
> resetting: the controller will be reset, so the queue will be deleted
> connecting: init io failures will teardown partially initialized 
> controller, so the queue will be deleted

I can add this comment.

> 
>>   static enum blk_eh_timer_return
>>   nvme_rdma_timeout(struct request *rq, bool reserved)
>>   {
>> @@ -1961,29 +1979,43 @@ nvme_rdma_timeout(struct request *rq, bool 
>> reserved)
>>       dev_warn(ctrl->ctrl.device, "I/O %d QID %d timeout\n",
>>            rq->tag, nvme_rdma_queue_idx(queue));
>> -    /*
>> -     * Restart the timer if a controller reset is already scheduled. Any
>> -     * timed out commands would be handled before entering the 
>> connecting
>> -     * state.
>> -     */
>> -    if (ctrl->ctrl.state == NVME_CTRL_RESETTING)
>> +    switch (ctrl->ctrl.state) {
>> +    case NVME_CTRL_RESETTING:
>> +        if (!nvme_rdma_queue_idx(queue)) {
>> +            /*
>> +             * if we are in teardown we must complete immediately
>> +             * because we may block the teardown sequence (e.g.
>> +             * nvme_disable_ctrl timed out).
>> +             */
>> +            nvme_rdma_complete_timed_out(rq);
>> +            return BLK_EH_DONE;
>> +        }
>> +        /*
>> +         * Restart the timer if a controller reset is already scheduled.
>> +         * Any timed out commands would be handled before entering the
>> +         * connecting state.
>> +         */
>>           return BLK_EH_RESET_TIMER;
> 
> If you're in RESETTING, why do you need to qualify ios only on the admin 
> queue. Can't all ios, regardless of queue, just be complete_timed_out() 
> ?  Isn't this just a race between the io timeout and the resetting 
> routine reaching the io ?

You are correct, given that we are serialized against the reset/error 
recovery we can just do the same for both. The request is going to
be cancelled anyways.

> 
> 
>> -
>> -    if (ctrl->ctrl.state != NVME_CTRL_LIVE) {
>> +    case NVME_CTRL_CONNECTING:
>> +        if (reserved || !nvme_rdma_queue_idx(queue)) {
>> +            /*
>> +             * if we are connecting we must complete immediately
>> +             * connect (reserved) or admin requests because we may
>> +             * block controller setup sequence.
>> +             */
>> +            nvme_rdma_complete_timed_out(rq);
>> +            return BLK_EH_DONE;
>> +        }
> 
> This is reasonable.  But I'm wondering why this too isn't just 
> completing any io that timed out.  For the non-controller create/init 
> ios - they'll either bounce back to the multipather or will requeue. 
> With the requeue, there's an opportunity for Viktor Gladko'vs "reject 
> I/O to offline device" to bounce it if it's been waiting a while.

You are right, I can do that to any state that is not LIVE.

Thanks for the review!