Re: CFQ: async queue blocks the whole system

[Tao Ma <tm@tao.ma>]

On 06/14/2011 09:30 PM, Vivek Goyal wrote:
> On Tue, Jun 14, 2011 at 03:03:24PM +0800, Tao Ma wrote:
>> Hi Vivek,
>> On 06/14/2011 05:41 AM, Vivek Goyal wrote:
>>> On Mon, Jun 13, 2011 at 06:08:40PM +0800, Tao Ma wrote:
>>>
>>> [..]
>>>>> You can also run iostat on disk and should be able to see that with
>>>>> the patch you are dispatching writes more often than before.
>>>> Sorry, the patch doesn't work.
>>>>
>>>> I used trace event to capture all the blktraces since it doesn't
>>>> interfere with the tests, hope it helps.
>>>
>>> Actually I was looking for CFQ traces. This seems to be generic block
>>> layer trace points. May be you can use "blktrace -d /dev/<device>"
>>> and then blkparse. It also gives the aggregate view which is helpful.
>>>
>>>>
>>>> Please downloaded it from http://blog.coly.li/tmp/blktrace.tar.bz2
>>>
>>> What concerns me is following.
>>>
>>> 5255.521353: block_rq_issue: 8,0 W 0 () 571137153 + 8 [attr_set]
>>> 5578.863871: block_rq_issue: 8,0 W 0 () 512950473 + 48 [kworker/0:1]
>>>
>>> IIUC, we dispatched second write more than 300 seconds after dispatching
>>> 1 write. What happened in between. We should have dispatched more writes.
>>>
>>> CFQ traces might give better idea in terms of whether wl_type for async
>>> queues was scheduled or not at all.
>> I tried several times today, but it looks like that if I enable
>> blktrace, the hung_task will not show up in the message. So do you think
>> the blktrace at that time is still useful? If yes, I can capture 1
>> minute for you. Thanks.
> 
> Capturing 1 min output will also be good. 
OK, I captured 2 mins blkparse log before the hung. You can downloaded
it from http://blog.coly.li/tmp/blkparse.tar.bz2

> 
> You can do one more thing. Mount block IO controller. It has the stats for
> sync and async dispatch (blkio.io_serviced or blkio.io_service_bytes). You
> can write a simple script to read and print these files every few seconds.
> That will also tell whether CFQ is dispatching async requests for the 
> said device regularly or not.
OK, I will try block IO controller tomorrow to see whether we can find
some useful info. Anyway, thanks for the diagnose.

Thanks
Tao