lkml.org 
[lkml]   [2020]   [Sep]   [21]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [v4] mm: khugepaged: avoid overriding min_free_kbytes set by user
From
Date


On 9/17/2020 10:56 PM, Michal Hocko wrote:
> On Thu 17-09-20 11:16:55, Vijay Balakrishna wrote:
>>
>>
>> On 9/17/2020 10:52 AM, Michal Hocko wrote:
>>> On Thu 17-09-20 10:27:16, Vijay Balakrishna wrote:
>>>>
>>>>
>>>> On 9/17/2020 2:28 AM, Michal Hocko wrote:
>>>>> On Wed 16-09-20 23:39:39, Vijay Balakrishna wrote:
>>>>>> set_recommended_min_free_kbytes need to honor min_free_kbytes set by the
>>>>>> user. Post start-of-day THP enable or memory hotplug operations can
>>>>>> lose user specified min_free_kbytes, in particular when it is higher than
>>>>>> calculated recommended value.
>>>>>
>>>>> I was about to recommend a more detailed explanation when I have
>>>>> realized that this patch is not really needed after all. Unless I am
>>>>> missing something.
>>>>>
>>>>> init_per_zone_wmark_min ignores the newly calculated min_free_kbytes if
>>>>> it is lower than user_min_free_kbytes. So calculated min_free_kbytes >=
>>>>> user_min_free_kbytes.
>>>>>
>>>>> Except for value clamping when the value is reduced and this likely
>>>>> needs fixing. But set_recommended_min_free_kbytes should be fine.
>>>>>
>>>>
>>>> IIUC, after start-of-day if a user performs
>>>> - THP disable
>>>> - modifies min_free_bytes
>>>> - THP enable
>>>> above sequence currently wouldn't result in calling init_per_zone_wmark_min.
>>>
>>> I will not, but why do you think this matters? All we should care about
>>> is that auto-tuning shouldn't reduce user provided value [1] and that
>>> the memory hotplug should be consistent with the boot time heuristic.
>>> init_per_zone_wmark_min should make sure that the user value is not
>>> reduced and thp heuristic makes sure it will not reduce this value.
>>> So the property should be transitive with the existing code (modulo the
>>> problem I have highlighted).
>>>
>>> [1] one could argue that it shouldn't even increase the value strictly
>>> speaking because an admin might have a very good reason to decrease the
>>> value but this has never been the semantic and changing it now might be
>>> problematic
>>>
>>
>> I made an attempt to address Kirill A. Shutemov's comment.
>
> This is for Kirill to comment on but my take would be that memory
> hotplug really has to alter the user defined min_free_kbytes because it
> is manipulating the amount of memory. There are usecases which are
> adding a lot of memory.
>
> We are trying to not decrease the value which is arguably a weird semantic
> but this is what've been doing for years. We would need to hear a
> specific usecase where this matters (e.g. memory hotremove heavy
> workalod with manually tuned min_free_kbytes) that misbehaves.

In our use case memory hotremove done normally during shutdown and we
aren't manually tuning min_free_kbytes.

>
>> And incrased
>> min_free_kbytes to see the issue in my testing and attempted a fix. I'm ok
>> leaving as it is. Do not want introduce any changes that may cause
>> regression.
>
> I would recommend reposting the patch which adds heuristic for THP (if
> THP is enabled) into the hotplug path, arguing with the consistency and
> surprising results when adding memory decreases the value.

I hope my reposted patch
([v3 1/2] mm: khugepaged: recalculate min_free_kbytes after memory
hotplug as expected by khugepaged)
change log is ok:

When memory is hotplug added or removed the min_free_kbytes must be
recalculated based on what is expected by khugepaged. Currently
after hotplug, min_free_kbytes will be set to a lower default and higher
default set when THP enabled is lost. This change restores
min_free_kbytes as expected for THP consumers.


> Your initial
> problem is in sizing as mentioned in other email thread and you should
> be investigating more but this inconsistency might really come as a
> surprise.
>
> All that if Kirill is reconsidering his initial position of course.
>

Kirill, can you comment or share your opinion?

Thanks,
Vijay

\
 
 \ /
  Last update: 2020-09-21 21:08    [W:0.151 / U:0.144 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site