Re: [PATCH V5] Allow compaction of unevictable pages

[Vlastimil Babka <vbabka@suse.cz>]

[CC += linux-api@]

Since this is a kernel-user-space API change, please CC linux-api@.
The kernel source file Documentation/SubmitChecklist notes that all
Linux kernel patches that change userspace interfaces should be CCed
to linux-api@vger.kernel.org, so that the various parties who are
interested in API changes are informed. For further information, see
https://www.kernel.org/doc/man-pages/linux-api-ml.html


On 03/13/2015 09:19 PM, Michal Hocko wrote:
> On Fri 13-03-15 15:09:15, Eric B Munson wrote:
>> On Fri, 13 Mar 2015, Rik van Riel wrote:
>>
>>> On 03/13/2015 01:26 PM, Eric B Munson wrote:
>>>
>>>> --- a/mm/compaction.c
>>>> +++ b/mm/compaction.c
>>>> @@ -1046,6 +1046,8 @@ typedef enum {
>>>>   	ISOLATE_SUCCESS,	/* Pages isolated, migrate */
>>>>   } isolate_migrate_t;
>>>>
>>>> +int sysctl_compact_unevictable;

A comment here would be useful I think, as well as explicit default 
value. Maybe also __read_mostly although I don't know how much that matters.

I also wonder if it might be confusing that "compact_memory" is a 
write-only trigger that doesn't even show under "sysctl -a", while 
"compact_unevictable" is a read/write setting. But I don't have a better 
suggestion right now.

>>>> +
>>>>   /*
>>>>    * Isolate all pages that can be migrated from the first suitable block,
>>>>    * starting at the block pointed to by the migrate scanner pfn within
>>>
>>> I suspect that the use cases where users absolutely do not want
>>> unevictable pages migrated are special cases, and it may make
>>> sense to enable sysctl_compact_unevictable by default.
>>
>> Given that sysctl_compact_unevictable=0 is the way the kernel behaves
>> now and the push back against always enabling compaction on unevictable
>> pages, I left the default to be the behavior as it is today.
>
> The question is _why_ we have this behavior now. Is it intentional?

It's there since 748446bb6 ("mm: compaction: memory compaction core"). 
Commit c53919adc0 ("mm: vmscan: remove lumpy reclaim") changes the 
comment in __isolate_lru_page() handling of unevictable pages to mention 
compaction explicitly. It could have been accidental in 748446bb6 
though, maybe it just reused __isolate_lru_page() for compaction - it 
seems that the skipping of unevictable was initially meant to optimize 
lumpy reclaim.

> e46a28790e59 (CMA: migrate mlocked pages) is a precedence in that

Well, CMA and realtime kernels are probably mutually exclusive enough.

> direction. Vlastimil has then changed that by edc2ca612496 (mm,
> compaction: move pageblock checks up from isolate_migratepages_range()).
> There is no mention about mlock pages so I guess it was more an
> unintentional side effect of the patch. At least that is my current
> understanding. I might be wrong here.

Although that commit did change unintentionally more details that I 
would have liked (unfortunately), I think you are wrong on this one. 
ISOLATE_UNEVICTABLE is still passed from isolate_migratepages_range() 
which is used by CMA, while the compaction variant 
isolate_migratepages() does not pass it. So it's kept CMA-specific as 
before.

> The thing about RT is that it is not usable with the upstream kernel
> without the RT patchset AFAIU. So the default should be reflect what is
> better for the standard kernel. RT loads have to tune the system anyway
> so it is not so surprising they would disable this option as well. We
> should help those guys and do not require them to touch the code but the
> knob is reasonable IMHO.
>
> Especially when your changelog suggests that having this enabled by
> default is beneficial for the standard kernel.

I agree, but if there's a danger of becoming too of a bikeshed topic, 
I'm fine with keeping the default same as current behavior and changing 
it later. Or maybe we should ask some -rt mailing list instead of just 
Peter and Thomas?