Re: [PATCH v1 0/4]mm: convert totalram_pages, totalhigh_pages and managed pages to atomic

[Vlastimil Babka <vbabka@suse.cz>]

On 11/7/18 8:02 AM, Konstantin Khlebnikov wrote:
> On 06.11.2018 11:43, Arun KS wrote:
>> On 2018-11-06 14:07, Konstantin Khlebnikov wrote:
>>> On 06.11.2018 11:30, Arun KS wrote:
>>>> On 2018-11-06 13:47, Konstantin Khlebnikov wrote:
>>>>> On 06.11.2018 8:38, Arun KS wrote:
>>>>>> Any comments?
>>>>>
>>>>> Looks good.
>>>>> Except unclear motivation behind this change.
>>>>> This should be in comment of one of patch.
>>>>
>>>> totalram_pages, zone->managed_pages and totalhigh_pages are sometimes modified outside managed_page_count_lock. Hence convert these 
>>>> variable to atomic to avoid readers potentially seeing a store tear.
>>>
>>> So, this is just theoretical issue or splat from sanitizer.
>>> After boot memory online\offline are strictly serialized by rw-semaphore.
>>
>> Few instances which can race with hot add. Please see below,
>> https://patchwork.kernel.org/patch/10627521/
> Could you point what exactly are you fixing with this set?
> 
> from v2:
> 
>  > totalram_pages, zone->managed_pages and totalhigh_pages updates
>  > are protected by managed_page_count_lock, but readers never care
>  > about it. Convert these variables to atomic to avoid readers
>  > potentially seeing a store tear.
> 
> This?
> 
> 
> Aligned unsigned long almost always stored at once.

The point is "almost always", so better not rely on it :) But the main
motivation was that managed_page_count_lock handling was complicating
Arun's "memory_hotplug: Free pages as higher order" patch and it seemed
a better idea to just remove and convert this to atomics, with
preventing potential store-to-read tearing as a bonus.

It would be nice to mention it in the changelogs though.

> To make it completely correct you could replace
> 
> a += b;
> 
> with
> 
> WRITE_ONCE(a, a + b);

Wouldn't be enough to get rid of the locks.