Re: [RFC PATCH v1] tools/mm: Add thpmaps script to dump THP usage info

[David Hildenbrand <david@redhat.com>]

On 11.01.24 21:21, Barry Song wrote:
> On Fri, Jan 12, 2024 at 2:18 AM David Hildenbrand <david@redhat.com> wrote:
>>
>> On 11.01.24 13:25, Ryan Roberts wrote:
>>> On 10/01/2024 22:14, Barry Song wrote:
>>>> On Wed, Jan 10, 2024 at 7:59 PM Ryan Roberts <ryan.roberts@arm.com> wrote:
>>>>>
>>>>> On 10/01/2024 11:38, Barry Song wrote:
>>>>>> On Wed, Jan 10, 2024 at 7:21 PM Ryan Roberts <ryan.roberts@arm.com> wrote:
>>>>>>>
>>>>>>> On 10/01/2024 11:00, David Hildenbrand wrote:
>>>>>>>> On 10.01.24 11:55, Ryan Roberts wrote:
>>>>>>>>> On 10/01/2024 10:42, David Hildenbrand wrote:
>>>>>>>>>> On 10.01.24 11:38, Ryan Roberts wrote:
>>>>>>>>>>> On 10/01/2024 10:30, Barry Song wrote:
>>>>>>>>>>>> On Wed, Jan 10, 2024 at 6:23 PM Ryan Roberts <ryan.roberts@arm.com> wrote:
>>>>>>>>>>>>>
>>>>>>>>>>>>> On 10/01/2024 09:09, Barry Song wrote:
>>>>>>>>>>>>>> On Wed, Jan 10, 2024 at 4:58 PM Ryan Roberts <ryan.roberts@arm.com> wrote:
>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>> On 10/01/2024 08:02, Barry Song wrote:
>>>>>>>>>>>>>>>> On Wed, Jan 10, 2024 at 12:16 PM John Hubbard <jhubbard@nvidia.com> wrote:
>>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>> On 1/9/24 19:51, Barry Song wrote:
>>>>>>>>>>>>>>>>>> On Wed, Jan 10, 2024 at 11:35 AM John Hubbard <jhubbard@nvidia.com>
>>>>>>>>>>>>>>>>>> wrote:
>>>>>>>>>>>>>>>>> ...
>>>>>>>>>>>>>>>>>>> Hi Ryan,
>>>>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>>>> One thing that immediately came up during some recent testing of mTHP
>>>>>>>>>>>>>>>>>>> on arm64: the pid requirement is sometimes a little awkward. I'm
>>>>>>>>>>>>>>>>>>> running
>>>>>>>>>>>>>>>>>>> tests on a machine at a time for now, inside various containers and
>>>>>>>>>>>>>>>>>>> such, and it would be nice if there were an easy way to get some
>>>>>>>>>>>>>>>>>>> numbers
>>>>>>>>>>>>>>>>>>> for the mTHPs across the whole machine.
>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>> Just to confirm, you're expecting these "global" stats be truely global
>>>>>>>>>>>>>>> and not
>>>>>>>>>>>>>>> per-container? (asking because you exploicitly mentioned being in a
>>>>>>>>>>>>>>> container).
>>>>>>>>>>>>>>> If you want per-container, then you can probably just create the container
>>>>>>>>>>>>>>> in a
>>>>>>>>>>>>>>> cgroup?
>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>>>> I'm not sure if that changes anything about thpmaps here. Probably
>>>>>>>>>>>>>>>>>>> this is fine as-is. But I wanted to give some initial reactions from
>>>>>>>>>>>>>>>>>>> just some quick runs: the global state would be convenient.
>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>> Thanks for taking this for a spin! Appreciate the feedback.
>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>>> +1. but this seems to be impossible by scanning pagemap?
>>>>>>>>>>>>>>>>>> so may we add this statistics information in kernel just like
>>>>>>>>>>>>>>>>>> /proc/meminfo or a separate /proc/mthp_info?
>>>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>> Yes. From my perspective, it looks like the global stats are more useful
>>>>>>>>>>>>>>>>> initially, and the more detailed per-pid or per-cgroup stats are the
>>>>>>>>>>>>>>>>> next level of investigation. So feels odd to start with the more
>>>>>>>>>>>>>>>>> detailed stats.
>>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>> probably because this can be done without the modification of the kernel.
>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>> Yes indeed, as John said in an earlier thread, my previous attempts to add
>>>>>>>>>>>>>>> stats
>>>>>>>>>>>>>>> directly in the kernel got pushback; DavidH was concerned that we don't
>>>>>>>>>>>>>>> really
>>>>>>>>>>>>>>> know exectly how to account mTHPs yet
>>>>>>>>>>>>>>> (whole/partial/aligned/unaligned/per-size/etc) so didn't want to end up
>>>>>>>>>>>>>>> adding
>>>>>>>>>>>>>>> the wrong ABI and having to maintain it forever. There has also been some
>>>>>>>>>>>>>>> pushback regarding adding more values to multi-value files in sysfs, so
>>>>>>>>>>>>>>> David
>>>>>>>>>>>>>>> was suggesting coming up with a whole new scheme at some point (I know
>>>>>>>>>>>>>>> /proc/meminfo isn't sysfs, but the equivalent files for NUMA nodes and
>>>>>>>>>>>>>>> cgroups
>>>>>>>>>>>>>>> do live in sysfs).
>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>> Anyway, this script was my attempt to 1) provide a short term solution
>>>>>>>>>>>>>>> to the
>>>>>>>>>>>>>>> "we need some stats" request and 2) provide a context in which to explore
>>>>>>>>>>>>>>> what
>>>>>>>>>>>>>>> the right stats are - this script can evolve without the ABI problem.
>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>> The detailed per-pid or per-cgroup is still quite useful to my case in
>>>>>>>>>>>>>>>> which
>>>>>>>>>>>>>>>> we set mTHP enabled/disabled and allowed sizes according to vma types,
>>>>>>>>>>>>>>>> eg. libc_malloc, java heaps etc.
>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>> Different vma types can have different anon_name. So I can use the
>>>>>>>>>>>>>>>> detailed
>>>>>>>>>>>>>>>> info to find out if specific VMAs have gotten mTHP properly and how many
>>>>>>>>>>>>>>>> they have gotten.
>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>> However, Ryan did clearly say, above, "In future we may wish to
>>>>>>>>>>>>>>>>> introduce stats directly into the kernel (e.g. smaps or similar)". And
>>>>>>>>>>>>>>>>> earlier he ran into some pushback on trying to set up /proc or /sys
>>>>>>>>>>>>>>>>> values because this is still such an early feature.
>>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>> I wonder if we could put the global stats in debugfs for now? That's
>>>>>>>>>>>>>>>>> specifically supposed to be a "we promise *not* to keep this ABI stable"
>>>>>>>>>>>>>>>>> location.
>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>> Now that I think about it, I wonder if we can add a --global mode to the
>>>>>>>>>>>>>>> script
>>>>>>>>>>>>>>> (or just infer global when neither --pid nor --cgroup are provided). I
>>>>>>>>>>>>>>> think I
>>>>>>>>>>>>>>> should be able to determine all the physical memory ranges from
>>>>>>>>>>>>>>> /proc/iomem,
>>>>>>>>>>>>>>> then grab all the info we need from /proc/kpageflags. We should then be
>>>>>>>>>>>>>>> able to
>>>>>>>>>>>>>>> process it all in much the same way as for --pid/--cgroup and provide the
>>>>>>>>>>>>>>> same
>>>>>>>>>>>>>>> stats, but it will apply globally. What do you think?
>>>>>>>>>>>>>
>>>>>>>>>>>>> Having now thought about this for a few mins (in the shower, if anyone wants
>>>>>>>>>>>>> the
>>>>>>>>>>>>> complete picture :) ), this won't quite work. This approach doesn't have the
>>>>>>>>>>>>> virtual mapping information so the best it can do is tell us "how many of
>>>>>>>>>>>>> each
>>>>>>>>>>>>> size of THP are allocated?" - it doesn't tell us anything about whether they
>>>>>>>>>>>>> are
>>>>>>>>>>>>> fully or partially mapped or what their alignment is (all necessary if we
>>>>>>>>>>>>> want
>>>>>>>>>>>>> to know if they are contpte-mapped). So I don't think this approach is
>>>>>>>>>>>>> going to
>>>>>>>>>>>>> be particularly useful.
>>>>>>>>>>>>>
>>>>>>>>>>>>> And this is also the big problem if we want to gather stats inside the
>>>>>>>>>>>>> kernel;
>>>>>>>>>>>>> if we want something equivalant to /proc/meminfo's
>>>>>>>>>>>>> AnonHugePages/ShmemPmdMapped/FilePmdMapped, we need to consider not just the
>>>>>>>>>>>>> allocation of the THP but also whether it is mapped. That's easy for
>>>>>>>>>>>>> PMD-mappings, because there is only one entry to consider - when you set it,
>>>>>>>>>>>>> you
>>>>>>>>>>>>> increment the number of PMD-mapped THPs, when you clear it, you decrement.
>>>>>>>>>>>>> But
>>>>>>>>>>>>> for PTE-mappings it's harder; you know the size when you are mapping so its
>>>>>>>>>>>>> easy
>>>>>>>>>>>>> to increment, but you can do a partial unmap, so you would need to scan the
>>>>>>>>>>>>> PTEs
>>>>>>>>>>>>> to figure out if we are unmapping the first page of a previously
>>>>>>>>>>>>> fully-PTE-mapped THP, which is expensive. We would need a cheap mechanism to
>>>>>>>>>>>>> determine "is this folio fully and contiguously mapped in at least one
>>>>>>>>>>>>> process?".
>>>>>>>>>>>>
>>>>>>>>>>>> as OPPO's approach I shared to you before is maintaining two mapcount
>>>>>>>>>>>> 1. entire map
>>>>>>>>>>>> 2. subpage's map
>>>>>>>>>>>> 3. if 1 and 2 both exist, it is DoubleMapped.
>>>>>>>>>>>>
>>>>>>>>>>>> This isn't a problem for us. and everytime if we do a partial unmap,
>>>>>>>>>>>> we have an explicit
>>>>>>>>>>>> cont_pte split which will decrease the entire map and increase the
>>>>>>>>>>>> subpage's mapcount.
>>>>>>>>>>>>
>>>>>>>>>>>> but its downside is that we expose this info to mm-core.
>>>>>>>>>>>
>>>>>>>>>>> OK, but I think we have a slightly more generic situation going on with the
>>>>>>>>>>> upstream; If I've understood correctly, you are using the PTE_CONT bit in the
>>>>>>>>>>> PTE to determne if its fully mapped? That works for your case where you only
>>>>>>>>>>> have 1 size of THP that you care about (contpte-size). But for the upstream, we
>>>>>>>>>>> have multi-size THP so we can't use the PTE_CONT bit to determine if its fully
>>>>>>>>>>> mapped because we can only use that bit if the THP is at least 64K and aligned,
>>>>>>>>>>> and only on arm64. We would need a SW bit for this purpose, and the mm would
>>>>>>>>>>> need to update that SW bit for every PTE one the full -> partial map
>>>>>>>>>>> transition.
>>>>>>>>>>
>>>>>>>>>> Oh no. Let's not make everything more complicated for the purpose of some stats.
>>>>>>>>>>
>>>>>>>>>
>>>>>>>>> Indeed, I was intending to argue *against* doing it this way. Fundamentally, if
>>>>>>>>> we want to know what's fully mapped and what's not, then I don't see any way
>>>>>>>>> other than by scanning the page tables and we might as well do that in user
>>>>>>>>> space with this script.
>>>>>>>>>
>>>>>>>>> Although, I expect you will shortly make a proposal that is simple to implement
>>>>>>>>> and prove me wrong ;-)
>>>>>>>>
>>>>>>>> Unlikely :) As you said, once you have multiple folio sizes, it stops really
>>>>>>>> making sense.
>>>>>>>>
>>>>>>>> Assume you have a 128 kiB pageache folio, and half of that is mapped. You can
>>>>>>>> set cont-pte bits on that half and all is fine. Or AMD can benefit from it's
>>>>>>>> optimizations without the cont-pte bit and everything is fine.
>>>>>>>
>>>>>>> Yes, but for debug and optimization, its useful to know when THPs are
>>>>>>> fully/partially mapped, when they are unaligned etc. Anyway, the script does
>>>>>>> that for us, and I think we are tending towards agreement that there are
>>>>>>> unlikely to be any cost benefits by moving it into the kernel.
>>>>>>
>>>>>> frequent partial unmap can defeat all purpose for us to use large folios.
>>>>>> just imagine a large folio can soon be splitted after it is formed. we lose
>>>>>> the performance gain and might get regression instead.
>>>>>
>>>>> nit: just because a THP gets partially unmapped in a process doesn't mean it
>>>>> gets split into order-0 pages. If the folio still has all its pages mapped at
>>>>> least once then no further action is taken. If the page being unmapped was the
>>>>> last mapping of that page, then the THP is put on the deferred split queue, so
>>>>> that it can be split in future if needed.
>>>>>>
>>>>>> and this can be very frequent, for example, one userspace heap management
>>>>>> is releasing memory page by page.
>>>>>>
>>>>>> In our real product deployment, we might not care about the second partial
>>>>>> unmapped,  we do care about the first partial unmapped as we can use this
>>>>>> to know if split has ever happened on this large folios. an partial unmapped
>>>>>> subpage can be unlikely re-mapped back.
>>>>>>
>>>>>> so i guess 1st unmap is probably enough, at least for my product. I mean we
>>>>>> care about if partial unmap has ever happened on a large folio more than how
>>>>>> they are exactly partially unmapped :-)
>>>>>
>>>>> I'm not sure what you are suggesting here? A global boolean that tells you if
>>>>> any folio in the system has ever been partially unmapped? That will almost
>>>>> certainly always be true, even for a very well tuned system.
>>>>
>>>> not a global boolean but a per-folio boolean. in case userspace maps a region
>>>> and has no userspace management, then we are fine as it is unlikely to have
>>>> partial unmap/map things; in case userspace maps a region, but manages it
>>>> by itself, such as heap things, we might result in lots of partial map/unmap,
>>>> which can lead to 3 problems:
>>>> 1. potential memory footprint increase, for example, while userspace releases
>>>> some pages in a folio, we might still keep it as frequent splitting folio into
>>>> basepages and releasing the unmapped subpage might be too expensive.
>>>> 2. if cont-pte is involved, frequent dropping cont-pte/tlb shootdown
>>>> might happen.
>>>> 3. other maintenance overhead such as splitting large folios etc.
>>>>
>>>> We'd like to know how serious partial map things are happening. so either
>>>> we will disable mTHP in this kind of VMAs, or optimize userspace to do
>>>> some alignment according to the size of large folios.
>>>>
>>>> in android phones, we detect lots of apps, and also found some apps might
>>>> do things like
>>>> 1. mprotect on some pages within a large folio
>>>> 2. mlock on some pages within a large folio
>>>> 3. madv_free on some pages within a large folio
>>>> 4. madv_pageout on some pages within a large folio.
>>>>
>>>> it would be good if we have a per-folio boolean to know how serious userspace
>>>> is breaking the large folios. for example, if more than 50% folios in a vma has
>>>> this problem, we can find it out and take some action.
>>>
>>> The high level value of these stats seems clear - I agree we need to be able to
>>> get these insights. I think the issues are more around the implementation
>>> though. I'm struggling to understand exactly how we could implement a lot of
>>> these things cheaply (either in the kernel or in user space).
>>>
>>> Let me try to work though what I think you are suggesting:
>>>
>>>    - every THP is initially fully mapped
>>
>> Not for pagecache folios.
>>
>>>    - when an operation causes a partial unmap, mark the folio as having at least
>>>      one partial mapping
>>>    - on transition from "no partial mappings" to "at least one partial mapping"
>>>      increment a "anon-partial-<size>kB" (one for each supported folio size)
>>>      counter by the folio size
>>>    - on transition from "at least one partial mapping" to "fully unampped
>>>      everywhere" decrement the counter by the folio size
>>>
>>> I think the issue with this is that a folio that is fully mapped in a process
>>> that gets forked, then is partially unmapped in 1 process, will be accounted as
>>> partially mapped even after the process that partially unmapped it exits, even
>>> though that folio is now fully mapped in all processes that map it. Is that a
>>> problem, perhaps not? I'm not sure.
>>
>> What I can offer with my total mapcount I am working on (+ entire/pmd
>> mapcount, but let's put that aside):
>>
>> 1) total_mapcount not multiples of folio_nr_page -> at least one process
>> currently maps the folio partially
>>
>> 2) total_mapcount is less than folio_nr_page -> surely partially mapped
>>
>> I think for most of anon memory (note that most folios are always
>> exclusive in our system, not cow-shared) 2) would already be sufficient.
> 
> if we can improve Ryan's "mm: Batch-copy PTE ranges during fork()" to
> add nr_pages in copy_pte_range for rmap.
> copy_pte_range()
> {
>             folio_try_dup_anon_rmap_ptes(...nr_pages....)
> }
> and at the same time, in zap_pte_range(), we remove the whole anon_rmap
> if the zapped-range covers the whole folio.
> 
> Replace the for-loop
> for (i = 0; i < nr; i++, page++) {
>          add_rmap(1);
> }
> for (i = 0; i < nr; i++, page++) {
>          remove_rmap(1);
> }
> by always using add_rmap(nr_pages) and remove_rmap(nr_pages) if we
> are doing the entire mapping/unmapping

That's precisely what I have already running as protoypes :) And I 
promised Ryan to get to this soon, clean it up and sent it out.

.
> 
> then we might be able to TestAndSetPartialMapped flag for this folio anywhile
> 1. someone is adding rmap with a number not equal nr_pages
> 2. someone is removing rmap with a number not equal nr_pages
> That means we are doing partial mapping or unmapping.
> and we increment partialmap_count by 1, let debugfs or somewhere present
> this count.

Yes. The only "ugly" corner case if you have a split VMA. We're not 
batching rmap exceeding that.

-- 
Cheers,

David / dhildenb