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

[Barry Song <21cnbao@gmail.com>]

On Wed, Jan 10, 2024 at 6:58 PM Ryan Roberts <ryan.roberts@arm.com> wrote:
>
> On 10/01/2024 10:54, David Hildenbrand wrote:
> > On 10.01.24 11:48, Barry Song wrote:
> >> On Wed, Jan 10, 2024 at 6:38 PM Ryan Roberts <ryan.roberts@arm.com> 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.
> >>
> >> My current implementation does use cont_pte but i don't think it is a must-have.
> >> we don't need a bit in PTE to know if we are partially unmapping a large folio
> >> at all.
> >>
> >> as long as we are unmapping a part of a large folio, we do know what we are
> >> doing. if a large folio is mapped entirely in a process, we get only
> >> entire_map +1,
> >> if we are unmapping a subpage of it, we get entire_map -1 and remained subpage's
> >> mapcount + 1. if we are only mapping a part of this large folio, we
> >> only increase
> >> its subpages' mapcount.
> >
> > That doesn't work as soon as you unmap a second subpage. Not to mention that
> > people ( :) ) are working on removing the subpage mapcounts.
>
> Yes, that was my point - Oppo's implementation relies on the bit in the PTE to
> tell the difference between unmapping the first subpage and unmapping the
> others. We don't have that luxury here.

right. The devil is in the details :-)

>
> >
> > I'm going propose that as a topic for LSF/MM soon, once I get to it.
> >
>