Re: [PATCH v2 1/2] mm: collect the number of anon large folios

[David Hildenbrand <david@redhat.com>]

On 22.08.24 10:44, Barry Song wrote:
> On Thu, Aug 22, 2024 at 12:52 PM Barry Song <21cnbao@gmail.com> wrote:
>>
>> On Thu, Aug 22, 2024 at 5:34 AM David Hildenbrand <david@redhat.com> wrote:
>>>
>>> On 12.08.24 00:49, Barry Song wrote:
>>>> From: Barry Song <v-songbaohua@oppo.com>
>>>>
>>>> Anon large folios come from three places:
>>>> 1. new allocated large folios in PF, they will call folio_add_new_anon_rmap()
>>>> for rmap;
>>>> 2. a large folio is split into multiple lower-order large folios;
>>>> 3. a large folio is migrated to a new large folio.
>>>>
>>>> In all above three counts, we increase nr_anon by 1;
>>>>
>>>> Anon large folios might go either because of be split or be put
>>>> to free, in these cases, we reduce the count by 1.
>>>>
>>>> Folios that have been added to the swap cache but have not yet received
>>>> an anon mapping won't be counted. This is consistent with the AnonPages
>>>> statistics in /proc/meminfo.
>>>
>>> Thinking out loud, I wonder if we want to have something like that for
>>> any anon folios (including small ones).
>>>
>>> Assume we longterm-pinned an anon folio and unmapped/zapped it. It would
>>> be quite interesting to see that these are actually anon pages still
>>> consuming memory. Same with memory leaks, when an anon folio doesn't get
>>> freed for some reason.
>>>
>>> The whole "AnonPages" counter thingy is just confusing, it only counts
>>> what's currently mapped ... so we'd want something different.
>>>
>>> But it's okay to start with large folios only, there we have a new
>>> interface without that legacy stuff :)
>>
>> We have two options to do this:
>> 1. add a new separate nr_anon_unmapped interface which
>> counts unmapped anon memory only
>> 2. let the nr_anon count both mapped and unmapped anon
>> folios.
>>
>> I would assume 1 is clearer as right now AnonPages have been
>> there for years. and counting all mapped and unmapped together,
>> we are still lacking an approach to find out anon memory leak
>> problem you mentioned.
>>
>> We are right now comparing nr_anon(including mapped folios only)
>> with AnonPages to get the distribution of different folio sizes in
>> performance profiling.
>>
>> unmapped_nr_anon should be normally always quite small. otherwise,
>> something must be wrong.
>>
>>>
>>>>
>>>> Signed-off-by: Barry Song <v-songbaohua@oppo.com>
>>>> ---
>>>>    Documentation/admin-guide/mm/transhuge.rst |  5 +++++
>>>>    include/linux/huge_mm.h                    | 15 +++++++++++++--
>>>>    mm/huge_memory.c                           | 13 ++++++++++---
>>>>    mm/migrate.c                               |  4 ++++
>>>>    mm/page_alloc.c                            |  5 ++++-
>>>>    mm/rmap.c                                  |  1 +
>>>>    6 files changed, 37 insertions(+), 6 deletions(-)
>>>>
>>>> diff --git a/Documentation/admin-guide/mm/transhuge.rst b/Documentation/admin-guide/mm/transhuge.rst
>>>> index 058485daf186..9fdfb46e4560 100644
>>>> --- a/Documentation/admin-guide/mm/transhuge.rst
>>>> +++ b/Documentation/admin-guide/mm/transhuge.rst
>>>> @@ -527,6 +527,11 @@ split_deferred
>>>>            it would free up some memory. Pages on split queue are going to
>>>>            be split under memory pressure, if splitting is possible.
>>>>
>>>> +nr_anon
>>>> +       the number of anon huge pages we have in the whole system.
>>>
>>> "transparent ..." otherwise people might confuse it with anon hugetlb
>>> "huge pages" ... :)
>>>
>>> I briefly tried coming up with a better name than "nr_anon" but failed.
>>>
>>>
>>
>> if we might have unmapped_anon counter later, maybe rename it to
>> nr_anon_mapped? and the new interface we will have in the future
>> might be nr_anon_unmapped?

We really shouldn't be using the mapped/unmapped terminology here ... we 
allocated pages and turned them into anonymous folios. At some point we 
free them. That's the lifecycle.

> 
> On second thought, this might be incorrect as well. Concepts like 'anon',
> 'shmem', and 'file' refer to states after mapping. If an 'anon' has been
> unmapped but is still pinned and not yet freed, it isn't technically an
> 'anon' anymore?

It's just not mapped, and cannot get mapped, anymore. In the memdesc 
world, we'd be freeing the "struct anon" or "struct folio" once the last 
refcount goes to 0, not once (e.g., temporarily during a failed 
migration?) unmapped.

The important part to me would be: this is memory that was allocated for 
anonymous memory, and it's still around for some reason and not getting 
freed. Usually, we would expect anon memory to get freed fairly quickly 
once unmapped. Except when there are long-term pinnings or other types 
of memory leaks.

You could happily continue using these anon pages via vmsplice() or 
similar, even thought he original page table mapping was torn down.

> 
> On the other hand, implementing nr_anon_unmapped could be extremely
> tricky. I have no idea how to implement it as we are losing those mapping
> flags.

folio_mapcount() can tell you efficiently whether a folio is mapped or 
not -- and that information will stay for all eternity as long as we 
have any mapcounts :) . It cannot tell "how many" pages of a large folio 
are mapped, but at least "is any page of this large folio mapped".

> 
> However, a page that is read-ahead but not yet mapped can still become
> an anon, which seems slightly less tricky to count though seems still
> difficult - except anon pages, shmem can be also swapped-backed?

Yes. I'm sure there would be ways to achieve it, but I am not sure if 
it's worth the churn. These pages can be reclaimed easily (I would 
assume? They are not even mapped and were never accessible via GUP), and 
can certainly not have any longterm pinnings or similar. There are more 
like "cached things that could become an anon folio".

-- 
Cheers,

David / dhildenb