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

[David Hildenbrand <david@redhat.com>]

On 10.01.24 16:19, Zi Yan wrote:
> On 10 Jan 2024, at 7:12, David Hildenbrand wrote:
> 
>> On 10.01.24 13:05, 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.
>>>
>>> yes. That is exactly what the kernel is doing, but this is not so
>>> important for us
>>> to resolve performance issues.
>>>
>>>>>
>>>>> 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.
>>>>
>>>>>
>>>>>>
>>>>>>>
>>>>>>> We want simple stats that tell us which folio sizes are actually allocated. For
>>>>>>> everything else, just scan the process to figure out what exactly is going on.
>>>>>>>
>>>>>>
>>>>>> Certainly that's much easier to do. But is it valuable? It might be if we also
>>>>>> keep stats for the number of failures to allocate the various sizes - then we
>>>>>> can see what percentage of high order allocation attempts are successful, which
>>>>>> is probably useful.
>>>
>>> My point is that we split large folios into two simple categories,
>>> 1. large folios which have never been partially unmapped
>>> 2. large folios which have ever been partially unmapped.
>>>
>>
>> With the rmap batching stuff I am working on, you get the complete thing unmapped in most cases (as long as they are in one VMA) -- for example during munmap()/exit()/etc.
> 
> IIUC, there are two cases:
> 
> 1. munmap() a range within a VMA, the rmap batching can void temporary partially unmapped folios, since it does the range operations as a whole.
> 
> 2. Barry has a case that userspace, e.g., the heap management, releases
> memory page by page, which rmap batching cannot help, unless either userspace
> batches memory releases or kernel delays and aggregates these memory releasing
> syscalls.

Exactly. And for 2. you immediately know that someone is partially 
unmapping a large folio. At least temporarily. Compared to doing a 
MADV_DONTNEED that covers a whole large folio (e.g., THP).

-- 
Cheers,

David / dhildenb