Re: [PATCH] lib/raid6: fix abnormally high latency

[yajun.deng@linux.dev]

December 16, 2021 2:39 PM, "Song Liu" <song@kernel.org> wrote:

> On Wed, Dec 15, 2021 at 6:15 PM <yajun.deng@linux.dev> wrote:
> 
>> December 16, 2021 12:52 AM, "Song Liu" <song@kernel.org> wrote:
>> 
>> On Tue, Dec 14, 2021 at 6:14 PM <yajun.deng@linux.dev> wrote:
>> 
>> December 15, 2021 1:27 AM, "Song Liu" <song@kernel.org> wrote:
>> 
>> On Mon, Dec 13, 2021 at 7:17 PM Yajun Deng <yajun.deng@linux.dev> wrote:
>> 
>> We found an abnormally high latency when executing modprobe raid6_pq, the
>> latency is greater than 1.2s when CONFIG_PREEMPT_VOLUNTARY=y, greater than
>> 67ms when CONFIG_PREEMPT=y, and greater than 16ms when CONFIG_PREEMPT_RT=y.
>> This is caused by disable the preemption, this time is too long and
>> unreasonable. We just need to disable migration. so used migrate_disable()/
>> migrate_enable() instead of preempt_disable()/preempt_enable(). This is
>> beneficial for CONFIG_PREEMPT=y or CONFIG_PREEMPT_RT=y, but no effect for
>> CONFIG_PREEMPT_VOLUNTARY=y.
>> 
>> Fixes: fe5cbc6e06c7 ("md/raid6 algorithms: delta syndrome functions")
>> Signed-off-by: Yajun Deng <yajun.deng@linux.dev>
>> 
>> We measure the speed of different RAID algorithms.If we don't disable
>> preempt, the result may be inaccurate, right? IIUC, we only disable preempt
>> for 16 jiffies. Why do we see 1.2 second delay?
>> 
>> Here are the command of my test：
>> Execute "sudo cyclictest -S -p 95 -d 0 -i 1000 -D 24h -m" in one terminal and "sudo modprobe
>> raid6_pq" in the other terminal.
>> 
>> Here are the results of my test：
>> CONFIG_PREEMPT_VOLUNTARY=y，CONFIG_HZ_250=y
>> T: 0 ( 3092) P:95 I:1000 C: 8514 Min: 1 Act: 2 Avg: 1 Max: 6
>> T: 1 ( 3093) P:95 I:1000 C: 8511 Min: 1 Act: 2 Avg: 1 Max: 14
>> 
>> I am not very familiar with the RT work, so please forgive me for some
>> rookie questions.
>> 
>> From the result, I think the CONFIG_PREEMPT_VOLUNTARY=y and the
>> CONFIG_PREEMPT=y cases failed to preempt during the preempt enabled period in
>> raid6_choose_gen(). Is this expected?
>> 
>> No, This is due to disable preemption causing ksoftirqd fail to schedule, we can use bcc tools see
>> that.
>> 
>> OTOH, the 16ms latency with CONFIG_PREEMPT_RT=y is more or less expected.
>> Is this acceptable? If not, is 1ms latency acceptable?
>> 
>> Here are the test results after adding patch：
>> CONFIG_PREEMPT=y or CONFIG_PREEMPT_RT=y
>> T: 0 ( 3167) P:95 I:1000 C: 13958 Min: 1 Act: 2 Avg: 1 Max: 5
>> T: 1 ( 3168) P:95 I:1000 C: 13956 Min: 1 Act: 2 Avg: 1 Max: 7
>> T: 2 ( 3169) P:95 I:1000 C: 13946 Min: 1 Act: 2 Avg: 1 Max: 12
>> T: 3 ( 3170) P:95 I:1000 C: 13951 Min: 1 Act: 2 Avg: 1 Max: 5
>> T: 4 ( 3171) P:95 I:1000 C: 13949 Min: 1 Act: 2 Avg: 1 Max: 3
>> T: 5 ( 3172) P:95 I:1000 C: 13947 Min: 1 Act: 2 Avg: 1 Max: 16
>> T: 6 ( 3173) P:95 I:1000 C: 13945 Min: 1 Act: 2 Avg: 2 Max: 7
>> T: 7 ( 3174) P:95 I:1000 C: 13942 Min: 1 Act: 2 Avg: 1 Max: 3
>> T: 8 ( 3175) P:95 I:1000 C: 13940 Min: 1 Act: 2 Avg: 1 Max: 3
>> T: 9 ( 3176) P:95 I:1000 C: 13938 Min: 1 Act: 1 Avg: 1 Max: 3
>> T:10 ( 3177) P:95 I:1000 C: 13936 Min: 1 Act: 2 Avg: 1 Max: 6
>> T:11 ( 3178) P:95 I:1000 C: 13933 Min: 1 Act: 2 Avg: 1 Max: 3
>> T:12 ( 3179) P:95 I:1000 C: 13931 Min: 1 Act: 2 Avg: 1 Max: 4
>> T:13 ( 3180) P:95 I:1000 C: 13929 Min: 1 Act: 2 Avg: 1 Max: 7
>> T:14 ( 3181) P:95 I:1000 C: 13927 Min: 1 Act: 2 Avg: 1 Max: 6
>> T:15 ( 3182) P:95 I:1000 C: 13925 Min: 1 Act: 2 Avg: 1 Max: 4
>> T:16 ( 3183) P:95 I:1000 C: 13923 Min: 1 Act: 2 Avg: 1 Max: 5
>> T:17 ( 3184) P:95 I:1000 C: 13921 Min: 1 Act: 2 Avg: 1 Max: 5
>> T:18 ( 3185) P:95 I:1000 C: 13919 Min: 1 Act: 2 Avg: 1 Max: 4
>> T:19 ( 3186) P:95 I:1000 C: 13916 Min: 1 Act: 2 Avg: 1 Max: 4
>> T:20 ( 3187) P:95 I:1000 C: 13914 Min: 1 Act: 2 Avg: 1 Max: 4
>> T:21 ( 3188) P:95 I:1000 C: 13912 Min: 1 Act: 2 Avg: 1 Max: 10
>> T:22 ( 3189) P:95 I:1000 C: 13910 Min: 1 Act: 2 Avg: 1 Max: 5
>> T:23 ( 3190) P:95 I:1000 C: 13908 Min: 1 Act: 2 Avg: 1 Max: 5
>> T:24 ( 3191) P:95 I:1000 C: 13906 Min: 1 Act: 2 Avg: 1 Max: 18
>> T:25 ( 3192) P:95 I:1000 C: 13904 Min: 1 Act: 2 Avg: 1 Max: 7
>> T:26 ( 3193) P:95 I:1000 C: 13902 Min: 1 Act: 2 Avg: 1 Max: 5
>> T:27 ( 3194) P:95 I:1000 C: 13900 Min: 1 Act: 1 Avg: 1 Max: 11
>> T:28 ( 3195) P:95 I:1000 C: 13898 Min: 1 Act: 2 Avg: 1 Max: 3
>> T:29 ( 3196) P:95 I:1000 C: 13896 Min: 1 Act: 2 Avg: 1 Max: 5
>> T:30 ( 3197) P:95 I:1000 C: 13894 Min: 1 Act: 2 Avg: 1 Max: 4
>> T:31 ( 3198) P:95 I:1000 C: 13892 Min: 1 Act: 2 Avg: 1 Max: 3
>> 
>> we can see the latency will not greater than 100us，so 1ms latency is also too long for
>> CONFIG_PREEMPT=y or CONFIG_PREEMPT_RT=y.
>> use migrate_disable()/migrate_enable() instead of preempt_disable()/preempt_enable() will not
>> affect the speed of different RAID algorithms and the latency can be reduced to a reasonable range.
> 
> I think allowing preempt may still affect the speed comparison. But
> such discrepancy
> should be acceptable. I will apply this to md-next.

Thank you, would you mind backport this patch to all LTS kernel? As you can see, the latency is not a little higher, but too much higher, it seems unreasonable， I think this may be a bug. We are using the 5.10 kernel and it would be very convenient for us if that was done.

> Thanks,
> Song