Re: Prerequisites for Large Anon Folios

[Ryan Roberts <ryan.roberts@arm.com>]

On 30/08/2023 17:20, David Hildenbrand wrote:
> On 30.08.23 12:44, Ryan Roberts wrote:
>> Hi All,
>>
> 
> Hi Ryan,
> 
> I'll be back from vacation next Wednesday.
> 
> Note that I asked David R. to have large anon folios as topic for the next
> bi-weekly mm meeting.

Ahh great! I don't have an invite to this meeting - is that something I can get
added to?

> 
> There, we should discuss things like
> * naming
> * accounting (/proc/meminfo)
> * required toggles (especially, to ways to disable it, as we want to
>   keep toggles minimal)
> 
> David R. raised that there are certainly workloads where the additional memory
> overhead is usually not acceptable. So it will be valuable to get input from
> others.
> 
>>
>> I want to get serious about getting large anon folios merged. To do that, there
>> are a number of outstanding prerequistes. I'm hoping the respective owners may
>> be able to provide an update on progress?
> 
> I shared some details in the last meeting when you were on vacation :)
> 
> High level update below.
> 
> [...]
> 
>>>
>>> - item:
>>>      shared vs exclusive mappings
>>>
>>>    priority:
>>>      prerequisite
>>>
>>>    description: >-
>>>      New mechanism to allow us to easily determine precisely whether a given
>>>      folio is mapped exclusively or shared between multiple processes. Required
>>>      for (from David H):
>>>
>>>      (1) Detecting shared folios, to not mess with them while they are shared.
>>>      MADV_PAGEOUT, user-triggered page migration, NUMA hinting, khugepaged ...
>>>      replace cases where folio_estimated_sharers() == 1 would currently be the
>>>      best we can do (and in some cases, page_mapcount() == 1).
>>>
>>>      (2) COW improvements for PTE-mapped large anon folios after fork(). Before
>>>      fork(), PageAnonExclusive would have been reliable, after fork() it's not.
>>>
>>>      For (1), "MADV_PAGEOUT" maps to the "madvise" item captured in this list. I
>>>      *think* "NUMA hinting" maps to "numa balancing" (but need confirmation!).
>>>      "user-triggered page migration" and "khugepaged" not yet captured (would
>>>      appreciate someone fleshing it out). I previously understood migration
>>> to be
>>>      working for large folios - is "user-triggered page migration" some specific
>>>      aspect that does not work?
>>>
>>>      For (2), this relates to Large Anon Folio enhancements which I plan to
>>>      tackle after we get the basic series merged.
>>>
>>>    links:
>>>      - 'email thread: Mapcount games: "exclusive mapped" vs. "mapped shared"'
>>>
>>>    location:
>>>      - shrink_folio_list()
>>>
>>>    assignee:
>>>      David Hildenbrand <david@redhat.com>
>>
>> Any comment on this David? I think the last comment I saw was that you were
>> planning to start an implementation a couple of weeks back? Did that get
>> anywhere?
> 
> The math should be solid at this point and I had a simple prototype running --
> including fairly clean COW reuse handling.
> 
> I started cleaning it all up before my vacation. I'll first need the total
> mapcount (which I sent), and might have to implement rmap patching during THP
> split (easy), but I first have to do more measurements.
> 
> Willies patches to free up space in the first tail page will be required. In
> addition, my patches to free up ->private in tail pages for THP_SWAP. Both
> things on their way upstream.
> 
> Based on that, I need a bit spinlock to protect the total mapcount+tracking
> data. There are things to measure (contention) and optimize (why even care about
> tracking shared vs. exclusive if it's pretty guaranteed to always be shared --
> for example, shared libraries).
> 
> So it looks reasonable at this point, but I'll have to look into possible
> contentions and optimizations once I have the basics implemented cleanly.
> 
> It's a shame we cannot get the subpage mapcount out of the way immediately, then
> it wouldn't be "additional tracking" but "different tracking" :)
> 
> Once back from vacation, I'm planning on prioritizing this. Shouldn't take ages
> to get it cleaned up. Measurements and optimizations might take a bit longer.

That's great - thanks for the update. I'm obviously happy to help with any
benchmarking/testing - just shout.


> 
> [...]
> 
> 
>>>
>>>    assignee:
>>>      Yin, Fengwei <fengwei.yin@intel.com>
>>
>> As I understand it: initial solution based on folio_estimated_sharers() has gone
>> into v6.5. Have a dependecy on David's precise shared vs exclusive work for an
> 
> shared vs. exclusive in place would replace folio_estimated_sharers() users and
> most sub-page mapcount users.
> 
>> improved solution. And I think you mentioned you are planning to do a change
>> that avoids splitting a large folio if it is entirely covered by the range?
> 
> [..]
>>>
>>> - item:
>>>      numa balancing
>>>
>>>    priority:
>>>      prerequisite
>>>
>>>    description: >-
>>>      Large, pte-mapped folios are ignored by numa-balancing code. Commit comment
>>>      (e81c480): "We're going to have THP mapped with PTEs. It will confuse
>>>      numabalancing. Let's skip them for now." Likely depends on "shared vs
>>>      exclusive mappings". >>
>>>    links: []
>>>
>>>    location:
>>>      - do_numa_page()
>>>
>>>    assignee:
>>>      <none>
>>>
>>
>> Vaguely sounded like David might be planning to tackle this as part of his work
>> on "shared vs exclusive mappings" ("NUMA hinting"??). David?
> 
> It should be easy to handle it based on that. Similarly, khugepaged IIRC.

OK that's good to hear. I missed it off the list, but I have a regression with
large anon folios currently in the khugepaged mm selftest, which I think should
be fixed by this.

Thanks,
Ryan


>