NVMeoF: multipath stuck after bringing one ethernet port down

shahar.salzman shahar.salzman at kaminario.com
Thu May 25 05:27:02 PDT 2017


Returning the QUEUE_ERROR also solved the "nvme list" issue reported, 
but the paths where not coming back after I reconnected the ports.

I played with the keep alive thread so that it didn't exit if it got an 
error, and that was resolved as well.


On 05/25/2017 02:06 PM, shahar.salzman wrote:
> OK, so the indefinite retries are due to the 
> nvme/host/rdma.c:nvme_rdma_queue_rq returning BLK_MQ_RQ_QUEUE_BUSY 
> when the queue is not ready, wouldn't it be better to return 
> BLK_MQ_RQ_QUEUE_ERROR so that the application may handle it? Or 
> alternatively return an error after several retries (per Q or something).
>
> I tested a patch (on top of Sagi's) converting the return value to 
> QUEUE_ERROR, and both dd, and multipath -ll return instantly.
>
>
> On 05/23/2017 08:31 PM, shahar.salzman wrote:
>> I used the patch with my Mellanox OFED (mlnx-nvme-rdma), it patrially 
>> helps but the overall behaviour is still not good...
>>
>> If I do a dd (direct_IO) or multipath -ll during the path disconnect, 
>> I get indefinite retries with nearly zero time between retries, until 
>> the path is reconnected, and then dd is completed successfully, paths 
>> come back to life, returning OK status to the process which completes 
>> successfully.I would expect IO to fail after a certain timeout, and 
>> that the retries should be paced somewhat
>>
>> If I do an nvme list (blk_execute_rq), then the process stays in D 
>> state (I still see the retries), the block layer (at leasy for this 
>> device) seems to be stuck, and I have to power cycle the server to 
>> get it to work...
>>
>> Just a reminder, I am using Mellanox OFED 4.0, and kernel 4.9.6, the 
>> plan here is to upgrade to the latest stable kernel but not sure when 
>> we'll get to it.
>>
>>
>> On 05/18/2017 08:52 AM, shahar.salzman wrote:
>>> I have seen this too, but was internally investigating it as I am 
>>> running 4.9.6 and not upstream. I will be able to check this patch 
>>> on Sunday/Monday.
>>>
>>> On my setup the IOs never complete even after the path is 
>>> reconnected, and the process remains stuck in D state, and the path 
>>> unusable until I power cycle the server.
>>>
>>> Some more information, hope it helps:
>>>
>>> Dumping the stuck process stack (multipath -ll):
>>>
>>> [root at kblock01-knode02 ~]# cat /proc/9715/stack
>>> [<ffffffff94354727>] blk_execute_rq+0x97/0x110
>>> [<ffffffffc0a72f8a>] __nvme_submit_user_cmd+0xca/0x300 [nvme_core]
>>> [<ffffffffc0a73369>] nvme_submit_user_cmd+0x29/0x30 [nvme_core]
>>> [<ffffffffc0a734bd>] nvme_user_cmd+0x14d/0x180 [nvme_core]
>>> [<ffffffffc0a7376c>] nvme_ioctl+0x7c/0xa0 [nvme_core]
>>> [<ffffffff9435d548>] __blkdev_driver_ioctl+0x28/0x30
>>> [<ffffffff9435dce1>] blkdev_ioctl+0x131/0x8f0
>>> [<ffffffff9428993c>] block_ioctl+0x3c/0x40
>>> [<ffffffff94260ae8>] vfs_ioctl+0x18/0x30
>>> [<ffffffff94261201>] do_vfs_ioctl+0x161/0x600
>>> [<ffffffff94261732>] SyS_ioctl+0x92/0xa0
>>> [<ffffffff9479cc77>] entry_SYSCALL_64_fastpath+0x1a/0xa9
>>> [<ffffffffffffffff>] 0xffffffffffffffff
>>>
>>> Looking at the dissasembly:
>>>
>>> 0000000000000170 <blk_execute_rq>:
>>>
>>> ...
>>>
>>>  1fa:   e8 00 00 00 00          callq  1ff <blk_execute_rq+0x8f>
>>>         /* Prevent hang_check timer from firing at us during very 
>>> long I/O */
>>>         hang_check = sysctl_hung_task_timeout_secs;
>>>         if (hang_check)
>>>                 while (!wait_for_completion_io_timeout(&wait, 
>>> hang_check * (HZ/2)));
>>>         else
>>>                 wait_for_completion_io(&wait);
>>>  1ff:   4c 89 e7                mov    %r12,%rdi
>>>  202:   e8 00 00 00 00          callq  207 <blk_execute_rq+0x97>
>>>
>>>         if (rq->errors)
>>>  207:   83 bb 04 01 00 00 01    cmpl   $0x1,0x104(%rbx)
>>>
>>> I ran btrace, and it seems that the IO running before the path 
>>> disconnect is properly cleaned, and only the IO running after gets 
>>> stuck (in the btrace example bellow I run both multipath -ll, and a 
>>> dd):
>>>
>>> 259,2   15    27548     4.790727567  8527  I  RS 764837376 + 8 [fio]
>>> 259,2   15    27549     4.790728103  8527  D  RS 764837376 + 8 [fio]
>>> 259,2   15    27550     4.791244566  8527  A   R 738007008 + 8 <- 
>>> (253,0) 738007008
>>> 259,2   15    27551     4.791245136  8527  I  RS 738007008 + 8 [fio]
>>> 259,2   15    27552     4.791245803  8527  D  RS 738007008 + 8 [fio]
>>> <<<<< From this stage there are no more fio Queue requests, only 
>>> completions
>>> 259,2    8     9118     4.758116423  8062  C  RS 2294539440 + 8 [0]
>>> 259,2    8     9119     4.758559461  8062  C  RS 160307464 + 8 [0]
>>> 259,2    8     9120     4.759009990     0  C  RS 2330623512 + 8 [0]
>>> 259,2    8     9121     4.759457400     0  C  RS 2657710096 + 8 [0]
>>> 259,2    8     9122     4.759928113     0  C  RS 2988256176 + 8 [0]
>>> 259,2    8     9123     4.760388009     0  C  RS 789190936 + 8 [0]
>>> 259,2    8     9124     4.760835889     0  C  RS 2915035352 + 8 [0]
>>> 259,2    8     9125     4.761304282     0  C  RS 2458313624 + 8 [0]
>>> 259,2    8     9126     4.761779694     0  C  RS 2280411456 + 8 [0]
>>> 259,2    8     9127     4.762286741     0  C  RS 2082207376 + 8 [0]
>>> 259,2    8     9128     4.762783722     0  C  RS 2874601688 + 8 [0]
>>> ....
>>> 259,2    8     9173     4.785798485     0  C  RS 828737568 + 8 [0]
>>> 259,2    8     9174     4.786301391     0  C  RS 229168888 + 8 [0]
>>> 259,2    8     9175     4.786769105     0  C  RS 893604096 + 8 [0]
>>> 259,2    8     9176     4.787270594     0  C  RS 2308778728 + 8 [0]
>>> 259,2    8     9177     4.787797448     0  C  RS 2190029912 + 8 [0]
>>> 259,2    8     9178     4.788298956     0  C  RS 2652012944 + 8 [0]
>>> 259,2    8     9179     4.788803759     0  C  RS 2205242008 + 8 [0]
>>> 259,2    8     9180     4.789309423     0  C  RS 1948528416 + 8 [0]
>>> 259,2    8     9181     4.789831240     0  C  RS 2003421288 + 8 [0]
>>> 259,2    8     9182     4.790343528     0  C  RS 2464964728 + 8 [0]
>>> 259,2    8     9183     4.790854684     0  C  RS 764837376 + 8 [0]
>>> <<<<<< This is probably where the port had been disconnected
>>> 259,2   21        7     9.413088410  8574  Q   R 0 + 8 [multipathd]
>>> 259,2   21        8     9.413089387  8574  G   R 0 + 8 [multipathd]
>>> 259,2   21        9     9.413095030  8574  U   N [multipathd] 1
>>> 259,2   21       10     9.413095367  8574  I  RS 0 + 8 [multipathd]
>>> 259,2   21       11     9.413096534  8574  D  RS 0 + 8 [multipathd]
>>> 259,2   10        1    14.381330190  2016  C  RS 738007008 + 8 [7]
>>> 259,2   21       12    14.381394890     0  R  RS 0 + 8 [7]
>>> 259,2    8     9184    80.500537429  8657  Q   R 0 + 8 [dd]
>>> 259,2    8     9185    80.500540122  8657  G   R 0 + 8 [dd]
>>> 259,2    8     9186    80.500545289  8657  U   N [dd] 1
>>> 259,2    8     9187    80.500545792  8657  I  RS 0 + 8 [dd]
>>> <<<<<< At this stage, I reconnected the path:
>>> 259,2    4        1  8381.791090134 11127  Q   R 0 + 8 [dd]
>>> 259,2    4        2  8381.791093611 11127  G   R 0 + 8 [dd]
>>> 259,2    4        3  8381.791098288 11127  U   N [dd] 1
>>> 259,2    4        4  8381.791098791 11127  I  RS 0 + 8 [dd]
>>>
>>>
>>> As I wrote above, I will check the patch on Sunday/Monday.
>>>
>>> On 05/17/2017 08:28 PM, Sagi Grimberg wrote:
>>>> Hi Alex,
>>>>
>>>>> I am trying to test failure scenarios of NVMeoF + multipath. I bring
>>>>> one of ports down and expect to see failed paths using "multipath
>>>>> -ll". Instead I see that "multipath -ll" get stuck.
>>>>>
>>>>> reproduce:
>>>>> 1. Connected to NVMeoF device through 2 ports.
>>>>> 2. Bind them with multipath.
>>>>> 3. Bring one port down (ifconfig eth3 down)
>>>>> 4. Execute "multipath -ll" command and it will get stuck.
>>>>> From strace I see that multipath is stuck in io_destroy() during
>>>>> release of resources. As I understand io_destroy is stuck because of
>>>>> io_cancel() that failed. And io_cancel() failed because of port that
>>>>> was disabled in step 3.
>>>>
>>>> Hmm, it looks like we do take care of failing fast pending IO, but 
>>>> once
>>>> we schedule periodic reconnects the request queues are already stopped
>>>> and new incoming requests may block until we successfully reconnect.
>>>>
>>>> I don't have too much time for it at the moment, but here is an 
>>>> untested
>>>> patch for you to try out:
>>>>
>>>> -- 
>>>> [PATCH] nvme-rdma: restart queues after we at error recovery to fast
>>>>  fail incoming io
>>>>
>>>> When we encounter an transport/controller errors, error recovery
>>>> kicks in which performs:
>>>> 1. stops io/admin queues
>>>> 2. moves transport queues out of LIVE state
>>>> 3. fast fail pending io
>>>> 4. schedule periodic reconnects.
>>>>
>>>> But we also need to fast fail incoming IO taht enters after we
>>>> already scheduled. Given that our queue is not LIVE anymore, simply
>>>> restart the request queues to fail in .queue_rq
>>>>
>>>> Signed-off-by: Sagi Grimberg <sagi at grimberg.me>
>>>> ---
>>>>  drivers/nvme/host/rdma.c | 20 +++++++++++---------
>>>>  1 file changed, 11 insertions(+), 9 deletions(-)
>>>>
>>>> diff --git a/drivers/nvme/host/rdma.c b/drivers/nvme/host/rdma.c
>>>> index dd1c6deef82f..a0aa2bfb91ee 100644
>>>> --- a/drivers/nvme/host/rdma.c
>>>> +++ b/drivers/nvme/host/rdma.c
>>>> @@ -753,28 +753,26 @@ static void 
>>>> nvme_rdma_reconnect_ctrl_work(struct work_struct *work)
>>>>         if (ret)
>>>>                 goto requeue;
>>>>
>>>> - blk_mq_start_stopped_hw_queues(ctrl->ctrl.admin_q, true);
>>>> -
>>>>         ret = nvmf_connect_admin_queue(&ctrl->ctrl);
>>>>         if (ret)
>>>> -               goto stop_admin_q;
>>>> +               goto requeue;
>>>>
>>>>         set_bit(NVME_RDMA_Q_LIVE, &ctrl->queues[0].flags);
>>>>
>>>>         ret = nvme_enable_ctrl(&ctrl->ctrl, ctrl->cap);
>>>>         if (ret)
>>>> -               goto stop_admin_q;
>>>> +               goto requeue;
>>>>
>>>>         nvme_start_keep_alive(&ctrl->ctrl);
>>>>
>>>>         if (ctrl->queue_count > 1) {
>>>>                 ret = nvme_rdma_init_io_queues(ctrl);
>>>>                 if (ret)
>>>> -                       goto stop_admin_q;
>>>> +                       goto requeue;
>>>>
>>>>                 ret = nvme_rdma_connect_io_queues(ctrl);
>>>>                 if (ret)
>>>> -                       goto stop_admin_q;
>>>> +                       goto requeue;
>>>>         }
>>>>
>>>>         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
>>>> @@ -782,7 +780,6 @@ static void 
>>>> nvme_rdma_reconnect_ctrl_work(struct work_struct *work)
>>>>         ctrl->ctrl.opts->nr_reconnects = 0;
>>>>
>>>>         if (ctrl->queue_count > 1) {
>>>> -               nvme_start_queues(&ctrl->ctrl);
>>>>                 nvme_queue_scan(&ctrl->ctrl);
>>>>                 nvme_queue_async_events(&ctrl->ctrl);
>>>>         }
>>>> @@ -791,8 +788,6 @@ static void 
>>>> nvme_rdma_reconnect_ctrl_work(struct work_struct *work)
>>>>
>>>>         return;
>>>>
>>>> -stop_admin_q:
>>>> -       blk_mq_stop_hw_queues(ctrl->ctrl.admin_q);
>>>>  requeue:
>>>>         dev_info(ctrl->ctrl.device, "Failed reconnect attempt %d\n",
>>>> ctrl->ctrl.opts->nr_reconnects);
>>>> @@ -823,6 +818,13 @@ static void 
>>>> nvme_rdma_error_recovery_work(struct work_struct *work)
>>>>         blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
>>>>                                 nvme_cancel_request, &ctrl->ctrl);
>>>>
>>>> +       /*
>>>> +        * queues are not a live anymore, so restart the queues to 
>>>> fail fast
>>>> +        * new IO
>>>> +        */
>>>> + blk_mq_start_stopped_hw_queues(ctrl->ctrl.admin_q, true);
>>>> +       nvme_start_queues(&ctrl->ctrl);
>>>> +
>>>>         nvme_rdma_reconnect_or_remove(ctrl);
>>>>  }
>>>>
>>>> -- 
>>>>
>>>> _______________________________________________
>>>> Linux-nvme mailing list
>>>> Linux-nvme at lists.infradead.org
>>>> http://lists.infradead.org/mailman/listinfo/linux-nvme
>>>
>>
>




More information about the Linux-nvme mailing list