Re: [PATCH v4 2/8] nvme-tcp: Update number of hardware queues before using them

From: Sagi Grimberg <sagi@grimberg.me>
To: Daniel Wagner <dwagner@suse.de>
Cc: linux-nvme@lists.infradead.org, linux-kernel@vger.kernel.org,
	James Smart <james.smart@broadcom.com>,
	Keith Busch <kbusch@kernel.org>, Ming Lei <ming.lei@redhat.com>,
	Hannes Reinecke <hare@suse.de>, Wen Xiong <wenxiong@us.ibm.com>
Subject: Re: [PATCH v4 2/8] nvme-tcp: Update number of hardware queues before using them
Date: Tue, 10 Aug 2021 18:00:37 -0700	[thread overview]
Message-ID: <01d7878c-e396-1d6b-c383-8739ca0b3d11@grimberg.me> (raw)
In-Reply-To: <20210809085250.xguvx5qiv2gxcoqk@carbon>

On 8/9/21 1:52 AM, Daniel Wagner wrote:
> Hi Sagi,
> 
> On Fri, Aug 06, 2021 at 12:57:17PM -0700, Sagi Grimberg wrote:
>>> -	ret = nvme_tcp_start_io_queues(ctrl);
>>> -	if (ret)
>>> -		goto out_cleanup_connect_q;
>>> -
>>> -	if (!new) {
>>> -		nvme_start_queues(ctrl);
>>> +	} else if (prior_q_cnt != ctrl->queue_count) {
>>
>> So if the queue count did not change we don't wait to make sure
>> the queue g_usage_counter ref made it to zero? What guarantees that it
>> did?
> 
> Hmm, good point. we should always call nvme_wait_freeze_timeout()
> for !new queues. Is this what you are implying?

I think we should always wait for the freeze to complete.

> 
> 
>>>    		if (!nvme_wait_freeze_timeout(ctrl, NVME_IO_TIMEOUT)) {
>>>    			/*
>>>    			 * If we timed out waiting for freeze we are likely to
>>> @@ -1828,6 +1822,10 @@ static int nvme_tcp_configure_io_queues(struct nvme_ctrl *ctrl, bool new)
>>>    		nvme_unfreeze(ctrl);
>>>    	}
>>> +	ret = nvme_tcp_start_io_queues(ctrl);
>>> +	if (ret)
>>> +		goto out_cleanup_connect_q;
>>> +
>>
>> Did you test this with both heavy I/O, reset loop and ifdown/ifup
>> loop?
> 
> Not sure if this classifies as heavy I/O (on 80 CPU machine)
> 
> fio --rw=readwrite --name=test --filename=/dev/nvme16n1 --size=50M \
>      --direct=1 --bs=4k --numjobs=40 --group_reporting --runtime=4h \
>      --time_based
> 
> and then I installed iptables rules to block the traffic on the
> controller side. With this test it is pretty easily to get
> the host hanging. Let me know what test you would like to see
> from me. I am glad to try to get them running.

Lets add iodepth=128

>> If we unquiesce and unfreeze before we start the queues the pending I/Os
>> may resume before the connect and not allow the connect to make forward
>> progress.
> 
> So the unfreeze should happen after the connect call? What about the
> newly created queues? Do they not suffer from the same problem? Isn't
> the NVME_TCP_Q_LIVE flag not enough?

Q_LIVE will protect against the transport itself from queueing, however
when multipath is not used, the transport will return BLK_STS_RESOURCE
which will immediately trigger re-submission, in an endless loop, and
that can prevent forward progress. It is also consistent with what 
nvme-pci does.