Re: [PATCH v15 3/7] mm: Add function __putback_isolated_page

[David Hildenbrand <david@redhat.com>]

> Am 17.12.2019 um 19:25 schrieb Alexander Duyck <alexander.h.duyck@linux.intel.com>:
> 
> On Tue, 2019-12-17 at 18:24 +0100, David Hildenbrand wrote:
>>>>> Also there are some scenarios where __page_to_pfn is not that simple a
>>>>> call with us having to get the node ID so we can find the pgdat structure
>>>>> to perform the calculation. I'm not sure the compiler would be ble to
>>>>> figure out that the result is the same for both calls, so it is better to
>>>>> make it explicit.
>>>> 
>>>> Only in case of CONFIG_SPARSEMEM we have to go via the section - but I
>>>> doubt this is really worth optimizing here.
>>>> 
>>>> But yeah, I'm fine with this change, only "IMHO
>>>> get_pageblock_migratetype() would be nicer" :)
>>> 
>>> Aren't most distros running with CONFIG_SPARSEMEM enabled? If that is the
>>> case why not optimize for it?
>> 
>> Because I tend to dislike micro-optimizations without performance
>> numbers for code that is not on a hot path. But I mean in this case, as
>> you said, you need the pfn either way, so it's completely fine with.
>> 
>> I do wonder, however, if you should just pass in the migratetype from
>> the caller. That would be even faster ;)
> 
> The problem is page isolation. We can end up with a page being moved to an
> isolate pageblock while we aren't holding the zone lock, and as such we
> likely need to test it again anyway. So there isn't value in storing and
> reusing the value for cases like page reporting.
> 
> In addition, the act of isolating the page can cause the migratetype to
> change as __isolate_free_page will attempt to change the migratetype to
> movable if it is one of the standard percpu types and we are pulling at
> least half a pageblock out. So storing the value before we isolate it
> would be problematic as well.
> 
> Undoing page isolation is the exception to the issues pointed out above,
> but in that case we are overwriting the pageblock migratetype anyway so
> the cache lines involved should all be warm from having just set the
> value.

Nothing would speak against querying the migratetype in the caller and passing it on. After all you‘re holding the zone lock, so it can‘t change.
>