Re: [PATCH RFC] mm: Use pr_warn_once() for min_free_kbytes warning

[Weilin Tong <tongweilin@linux.alibaba.com>]

在 2025/8/28 18:09, Michal Hocko 写道:
> On Thu 28-08-25 17:48:54, Weilin Tong wrote:
>> 在 2025/8/28 17:40, Michal Hocko 写道:
>>> On Thu 28-08-25 17:23:40, Weilin Tong wrote:
>>>> 在 2025/8/28 14:45, Michal Hocko 写道:
>>>>
>>>>> On Thu 28-08-25 11:06:02, Weilin Tong wrote:
>>>>>> When min_free_kbytes is user-configured, increasing system memory via memory
>>>>>> hotplug may trigger multiple recalculations of min_free_kbytes. This results
>>>>>> in excessive warning messages flooding the kernel log if several memory blocks
>>>>>> are added in a short period.
>>>>>>
>>>>>> Sample dmesg output before optimization:
>>>>>> ...
>>>>>> [ 1303.897214] min_free_kbytes is not updated to 126529 because user defined value 1048576 is preferred
>>>>>> [ 1303.960498] min_free_kbytes is not updated to 126529 because user defined value 1048576 is preferred
>>>>>> [ 1303.970116] min_free_kbytes is not updated to 126529 because user defined value 1048576 is preferred
>>>>>> [ 1303.979709] min_free_kbytes is not updated to 126529 because user defined value 1048576 is preferred
>>>>>> [ 1303.989254] min_free_kbytes is not updated to 126529 because user defined value 1048576 is preferred
>>>>>> [ 1303.999122] min_free_kbytes is not updated to 126529 because user defined value 1048576 is preferred
>>>>>> [ 1304.008644] min_free_kbytes is not updated to 126529 because user defined value 1048576 is preferred
>>>>>> [ 1304.018537] min_free_kbytes is not updated to 126529 because user defined value 1048576 is preferred
>>>>>> [ 1304.028054] min_free_kbytes is not updated to 126529 because user defined value 1048576 is preferred
>>>>>> [ 1304.037615] min_free_kbytes is not updated to 126529 because user defined value 1048576 is preferred
>>>>>> ...
>>>>>>
>>>>>> Replace pr_warn() with pr_warn_once() to ensure only one warning is printed,
>>>>>> preventing large volumes of repeated log entries and improving log readability.
>>>>> pr_warn_once seems too aggressive as we could miss useful events. On the
>>>>> other hand I agree that repeating the same message for each memory block
>>>>> onlining is not really helpful. Would it make sense to only pr_warn when
>>>>> new_min_free_kbytes differs from the previous one we have warned for?
>>>> Thanks for your feedback!
>>>>
>>>> The dmesg output above comes from hotplugging a large amount of memory into
>>>> ZONE_MOVABLE, where new_min_free_kbytes does not actually change, resulting
>>>> in repeated warnings with identical messages.
>>> Yes, this is clear from the changelog
>>>
>>>> However, if memory is hotplugged into ZONE_NORMAL (such as pmem-type
>>>> memory), new_min_free_kbytes changes on each operation, so we still get a
>>>> large number of warnings—even though the value is different each time.
>>> We can check whether the value has changed considerably.
>>>
>>>> If the concern is missing useful warnings, pr_warn_ratelimited() would be an
>>>> acceptable alternative, as it can reduce log spam without completely
>>>> suppressing potentially important messages. However I still think that
>>>> printing the warning once is sufficient to alert the user about the
>>>> overridden configuration, especially since this is not a particularly
>>>> critical warning.
>>> The thing is that kernel log buffer can easily overflow and you can lose
>>> those messages over time, especially for system with a large uptime -
>>> which is far from uncommon.
>>>
>>> I am not entirely enthusiastic about rate limiting because that is time
>>> rather than even driven. Anyway, if you can make ratelimiting work for
>>> your usecase, then no objection from me but I would rather make the
>>> reporting more useful than hack around it.
>> I agree with your suggestion.
>>
>> With respect to your suggestion that “we can check whether the value has
>> changed considerably” I would like to seek your advice on how to define what
>> constitutes a significant change in this context. Do you have any
>> recommended criteria or thresholds for determining when a difference in
>> min_free_kbytes should trigger a warning?
> No really. Certainly increasing min_free_kbytes by 1% would be barely
> noticeable but 10% might show some difference. This will likely need to
> be tuned on real life usecases so start with something and we can tune
> that based on future usecases.
>
Understood, thank you for your suggestion.

I'm also looking forward to additional discussion and input from the 
community.