[lkml]   [2017]   [May]   [8]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: Large latency on blk_queue_enter
On 05/08/2017 09:49 AM, Javier González wrote:
>> On 8 May 2017, at 17.40, Jens Axboe <> wrote:
>> On 05/08/2017 09:38 AM, Javier González wrote:
>>>> On 8 May 2017, at 17.25, Jens Axboe <> wrote:
>>>> On 05/08/2017 09:22 AM, Javier González wrote:
>>>>> Javier
>>>>>> On 8 May 2017, at 17.14, Jens Axboe <> wrote:
>>>>>> On 05/08/2017 09:08 AM, Jens Axboe wrote:
>>>>>>> On 05/08/2017 09:02 AM, Javier González wrote:
>>>>>>>>> On 8 May 2017, at 16.52, Jens Axboe <> wrote:
>>>>>>>>> On 05/08/2017 08:46 AM, Javier González wrote:
>>>>>>>>>>> On 8 May 2017, at 16.23, Jens Axboe <> wrote:
>>>>>>>>>>> On 05/08/2017 08:20 AM, Javier González wrote:
>>>>>>>>>>>>> On 8 May 2017, at 16.13, Jens Axboe <> wrote:
>>>>>>>>>>>>> On 05/08/2017 07:44 AM, Javier González wrote:
>>>>>>>>>>>>>>> On 8 May 2017, at 14.27, Ming Lei <> wrote:
>>>>>>>>>>>>>>> On Mon, May 08, 2017 at 01:54:58PM +0200, Javier González wrote:
>>>>>>>>>>>>>>>> Hi,
>>>>>>>>>>>>>>>> I find an unusual added latency(~20-30ms) on blk_queue_enter when
>>>>>>>>>>>>>>>> allocating a request directly from the NVMe driver through
>>>>>>>>>>>>>>>> nvme_alloc_request. I could use some help confirming that this is a bug
>>>>>>>>>>>>>>>> and not an expected side effect due to something else.
>>>>>>>>>>>>>>>> I can reproduce this latency consistently on LightNVM when mixing I/O
>>>>>>>>>>>>>>>> from pblk and I/O sent through an ioctl using liblightnvm, but I don't
>>>>>>>>>>>>>>>> see anything on the LightNVM side that could impact the request
>>>>>>>>>>>>>>>> allocation.
>>>>>>>>>>>>>>>> When I have a 100% read workload sent from pblk, the max. latency is
>>>>>>>>>>>>>>>> constant throughout several runs at ~80us (which is normal for the media
>>>>>>>>>>>>>>>> we are using at bs=4k, qd=1). All pblk I/Os reach the nvme_nvm_submit_io
>>>>>>>>>>>>>>>> function on lightnvm.c., which uses nvme_alloc_request. When we send a
>>>>>>>>>>>>>>>> command from user space through an ioctl, then the max latency goes up
>>>>>>>>>>>>>>>> to ~20-30ms. This happens independently from the actual command
>>>>>>>>>>>>>>>> (IN/OUT). I tracked down the added latency down to the call
>>>>>>>>>>>>>>>> percpu_ref_tryget_live in blk_queue_enter. Seems that the queue
>>>>>>>>>>>>>>>> reference counter is not released as it should through blk_queue_exit in
>>>>>>>>>>>>>>>> blk_mq_alloc_request. For reference, all ioctl I/Os reach the
>>>>>>>>>>>>>>>> nvme_nvm_submit_user_cmd on lightnvm.c
>>>>>>>>>>>>>>>> Do you have any idea about why this might happen? I can dig more into
>>>>>>>>>>>>>>>> it, but first I wanted to make sure that I am not missing any obvious
>>>>>>>>>>>>>>>> assumption, which would explain the reference counter to be held for a
>>>>>>>>>>>>>>>> longer time.
>>>>>>>>>>>>>>> You need to check if the .q_usage_counter is working at atomic mode.
>>>>>>>>>>>>>>> This counter is initialized as atomic mode, and finally switchs to
>>>>>>>>>>>>>>> percpu mode via percpu_ref_switch_to_percpu() in blk_register_queue().
>>>>>>>>>>>>>> Thanks for commenting Ming.
>>>>>>>>>>>>>> The .q_usage_counter is not working on atomic mode. The queue is
>>>>>>>>>>>>>> initialized normally through blk_register_queue() and the counter is
>>>>>>>>>>>>>> switched to percpu mode, as you mentioned. As I understand it, this is
>>>>>>>>>>>>>> how it should be, right?
>>>>>>>>>>>>> That is how it should be, yes. You're not running with any heavy
>>>>>>>>>>>>> debugging options, like lockdep or anything like that?
>>>>>>>>>>>> No lockdep, KASAN, kmemleak or any of the other usual suspects.
>>>>>>>>>>>> What's interesting is that it only happens when one of the I/Os comes
>>>>>>>>>>>> from user space through the ioctl. If I have several pblk instances on
>>>>>>>>>>>> the same device (which would end up allocating a new request in
>>>>>>>>>>>> parallel, potentially on the same core), the latency spike does not
>>>>>>>>>>>> trigger.
>>>>>>>>>>>> I also tried to bind the read thread and the liblightnvm thread issuing
>>>>>>>>>>>> the ioctl to different cores, but it does not help...
>>>>>>>>>>> How do I reproduce this? Off the top of my head, and looking at the code,
>>>>>>>>>>> I have no idea what is going on here.
>>>>>>>>>> Using LightNVM and liblightnvm [1] you can reproduce it by:
>>>>>>>>>> 1. Instantiate a pblk instance on the first channel (luns 0 - 7):
>>>>>>>>>> sudo nvme lnvm create -d nvme0n1 -n test0 -t pblk -b 0 -e 7 -f
>>>>>>>>>> 2. Write 5GB to the test0 block device with a normal fio script
>>>>>>>>>> 3. Read 5GB to verify that latencies are good (max. ~80-90us at bs=4k, qd=1)
>>>>>>>>>> 4. Re-run 3. and in parallel send a command through liblightnvm to a
>>>>>>>>>> different channel. A simple command is an erase (erase block 900 on
>>>>>>>>>> channel 2, lun 0):
>>>>>>>>>> sudo nvm_vblk line_erase /dev/nvme0n1 2 2 0 0 900
>>>>>>>>>> After 4. you should see a ~25-30ms latency on the read workload.
>>>>>>>>>> I tried to reproduce the ioctl in a more generic way to reach
>>>>>>>>>> __nvme_submit_user_cmd(), but SPDK steals the whole device. Also, qemu
>>>>>>>>>> is not reliable for this kind of performance testing.
>>>>>>>>>> If you have a suggestion on how I can mix an ioctl with normal block I/O
>>>>>>>>>> read on a standard NVMe device, I'm happy to try it and see if I can
>>>>>>>>>> reproduce the issue.
>>>>>>>>> Just to rule out this being any hardware related delays in processing
>>>>>>>>> IO:
>>>>>>>>> 1) Does it reproduce with a simpler command, anything close to a no-op
>>>>>>>>> that you can test?
>>>>>>>> Yes. I tried with a 4KB read and with a fake command I drop right after
>>>>>>>> allocation.
>>>>>>>>> 2) What did you use to time the stall being blk_queue_enter()?
>>>>>>>> I have some debug code measuring time with ktime_get() in different
>>>>>>>> places in the stack, and among other places, around blk_queue_enter(). I
>>>>>>>> use them then to measure max latency and expose it through sysfs. I can
>>>>>>>> see that the latency peak is recorded in the probe before
>>>>>>>> blk_queue_enter() and not in the one after.
>>>>>>>> I also did an experiment, where the normal I/O path allocates the
>>>>>>>> request with BLK_MQ_REQ_NOWAIT. When running the experiment above, the
>>>>>>>> read test fails since we reach:
>>>>>>>> if (nowait)
>>>>>>>> return -EBUSY;
>>>>>>>> in blk_queue_enter.
>>>>>>> OK, that's starting to make more sense, that indicates that there is indeed
>>>>>>> something wrong with the refs. Does the below help?
>>>>>> No, that can't be right, it does look balanced to begin with.
>>>>>> blk_mq_alloc_request() always grabs a queue ref, and always drops it. If
>>>>>> we return with a request succesfully allocated, then we have an extra
>>>>>> ref on it, which is dropped when it is later freed.
>>>>> I agree, it seems more like a reference is put too late. I looked into
>>>>> into the places where the reference is put, but it all seems normal. In
>>>>> any case, I run it (just to see), and it did not help.
>>>>>> Something smells fishy, I'll dig a bit.
>>>>> Thanks! I continue looking into it myself; let me know if I can help
>>>>> with something more specific.
>>>> What exact kernel are you running? And does the device have a scheduler
>>>> attached, or is it set to "none"?
>>> I can reproduce the issue on 4.11-rc7. I will rebase on top of your
>>> for-4.12/block, but I cannot see any patches that might be related. If
>>> it changes I'll ping you.
>> I don't suspect it will do anything for you. I just ask to know what
>> base you are on.
>>> I mentioned the problem to Christoph last week and disabling the
>>> schedulers was the first thing he recommended. I measured time around
>>> blk_mq_sched_get_request and for this particular test the choose of
>>> scheduler (including BFQ and kyber) does not seem to have an effect.
>> kyber vs none would be the interesting test. Some of the paths are a
>> little different depending if there's a scheduler attached or not, so
>> it's good to know that we're seeing this in both cases.
> I just tested on your for-4.12/block with none and kyber and the latency
> spike appears in both cases.

OK good. I looked at your reproduction case. Looks like we ultimately
end up submitting IO through nvme_nvm_submit_user_cmd() when you do the
nvm_vblk line_erase, which is basically the same code as
NVME_IOCTL_SUBMIT_IO as far as request alloc, setup, issue, free goes.
So does it reproduce for you as well on a normal nvme device, if you run
a nvme read /dev/nvme0 [...] while running the same read fio job?

Jens Axboe

 \ /
  Last update: 2017-05-08 18:07    [W:0.093 / U:2.800 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site