Re: [PATCH v5 6/9] slub: Delay freezing of partial slabs

[Chengming Zhou <chengming.zhou@linux.dev>]

On 2023/11/22 21:19, Vlastimil Babka wrote:
> On 11/22/23 12:54, Chengming Zhou wrote:
>> On 2023/11/22 19:40, Vlastimil Babka wrote:
>>> On 11/22/23 12:35, Chengming Zhou wrote:
>>>> On 2023/11/22 17:37, Vlastimil Babka wrote:
>>>>> On 11/20/23 19:49, Mark Brown wrote:
>>>>>> On Thu, Nov 02, 2023 at 03:23:27AM +0000, chengming.zhou@linux.dev wrote:
>>>>>>> From: Chengming Zhou <zhouchengming@bytedance.com>
>>>>>>>
>>>>>>> Now we will freeze slabs when moving them out of node partial list to
>>>>>>> cpu partial list, this method needs two cmpxchg_double operations:
>>>>>>>
>>>>>>> 1. freeze slab (acquire_slab()) under the node list_lock
>>>>>>> 2. get_freelist() when pick used in ___slab_alloc()
>>>>>>
>>>>>> Recently -next has been failing to boot on a Raspberry Pi 3 with an arm
>>>>>> multi_v7_defconfig and a NFS rootfs, a bisect appears to point to this
>>>>>> patch (in -next as c8d312e039030edab25836a326bcaeb2a3d4db14) as having
>>>>>> introduced the issue.  I've included the full bisect log below.
>>>>>>
>>>>>> When we see problems we see RCU stalls while logging in, for example:
>>>>>
>>>>> Can you try this, please?
>>>>>
>>>>
>>>> Great! I manually disabled __CMPXCHG_DOUBLE to reproduce the problem,
>>>> and this patch can solve the machine hang problem.
>>>>
>>>> BTW, I also did the performance testcase on the machine with 128 CPUs.
>>>>
>>>> stress-ng --rawpkt 128 --rawpkt-ops 100000000
>>>>
>>>> base    patched
>>>> 2.22s   2.35s
>>>> 2.21s   3.14s
>>>> 2.19s   4.75s
>>>>
>>>> Found this atomic version performance numbers are not stable.
>>>
>>> That's weirdly too bad. Is that measured also with __CMPXCHG_DOUBLE
>>> disabled, or just the patch? The PG_workingset flag change should be
>>
>> The performance test is just the patch.
>>
>>> uncontended as we are doing it under list_lock, and with __CMPXCHG_DOUBLE
>>> there should be no interfering PG_locked interference.
>>>
>>
>> Yes, I don't know. Maybe it's related with my kernel config, making the
>> atomic operation much expensive? Will look again..
> 
> I doubt it can explain going from 2.19s to 4.75s, must have been some
> interference on the machine?
> 

Yes, it looks so. There are some background services on the 128 CPUs machine.
Although "stress-ng --rawpkt 128 --rawpkt-ops 100000000" has so much regression,
I tried other less contented testcases:

1. stress-ng --rawpkt 64 --rawpkt-ops 100000000
2. perf bench sched messaging -g 5 -t -l 100000

The performance numbers of this atomic version are pretty much the same.

So this atomic version should be good in most cases IMHO.

>> And I also tested the atomic-optional version like below, found the
>> performance numbers are much stable.
> 
> This gets rather ugly and fragile so I'd maybe rather go back to the
> __unused field approach :/
> 

Agree. If we don't want this atomic version, the __unused field approach
seems better.

Thanks!