Re: [PATCH 5/7] blk-mq: remove REQ_ATOM_COMPLETE usages from blk-mq

From: "jianchao.wang" <jianchao.w.wang@oracle.com>
To: Tejun Heo <tj@kernel.org>
Cc: jbacik@fb.com, jack@suse.cz, axboe@kernel.dk, clm@fb.com,
	kernel-team@fb.com, linux-kernel@vger.kernel.org,
	linux-btrfs@vger.kernel.org, peterz@infradead.org,
	Bart.VanAssche@wdc.com
Subject: Re: [PATCH 5/7] blk-mq: remove REQ_ATOM_COMPLETE usages from blk-mq
Date: Fri, 22 Dec 2017 12:02:20 +0800	[thread overview]
Message-ID: <4e5aa629-f341-d077-961b-c778ceb31154@oracle.com> (raw)
In-Reply-To: <20171221135051.GE1084507@devbig577.frc2.facebook.com>

Sorry for my non-detailed description. 

On 12/21/2017 09:50 PM, Tejun Heo wrote:
> Hello,
> 
> On Thu, Dec 21, 2017 at 11:56:49AM +0800, jianchao.wang wrote:
>> It's worrying that even though the blk_mark_rq_complete() here is intended to synchronize with
>> timeout path, but it indeed give the blk_mq_complete_request() the capability to exclude with 
There could be scenario where the driver itself stop a request itself with blk_mq_complete_request() or
some other interface that will invoke it, races with the normal completion path where a same request comes.
For example:
a reset could be triggered through sysfs on nvme-rdma
Then the driver will cancel all the reqs, including in-flight ones.
nvme_rdma_reset_ctrl_work()
    nvme_rdma_shutdown_ctrl()
    >>>>
        if (ctrl->ctrl.queue_count > 1) {
            nvme_stop_queues(&ctrl->ctrl); //quiesce the queue
            blk_mq_tagset_busy_iter(&ctrl->tag_set,
                        nvme_cancel_request, &ctrl->ctrl); //invoke blk_mq_complete_request()
            nvme_rdma_destroy_io_queues(ctrl, shutdown);
        }
    >>>>

These operations could race with the normal completion path of in-flight ones.
It should drain all the in-flight ones first here. But there maybe some other
places similar with this.

>> itself. Maybe this capability should be reserved.
> 
> Can you explain further how that'd help?  The problem there is that if
> you have two competing completions, where one isn't a timeout, there's
> nothing synchronizing the reuse of the request.  IOW, the losing on
> can easily complete the next recycle instance.  The atomic bitops
> might feel like more protection but it's just feels.

In above case, the request may simultaneously enter requeue and end path.

Thanks
Jianchao