Re: [RFC v2 0/9] khugepaged: mTHP support

[Dev Jain <dev.jain@arm.com>]

On 15/02/25 6:22 am, Nico Pache wrote:
> On Thu, Feb 13, 2025 at 7:02 PM Dev Jain <dev.jain@arm.com> wrote:
>>
>>
>>
>> On 14/02/25 1:09 am, Nico Pache wrote:
>>> On Thu, Feb 13, 2025 at 1:26 AM Dev Jain <dev.jain@arm.com> wrote:
>>>>
>>>>
>>>>
>>>> On 12/02/25 10:19 pm, Nico Pache wrote:
>>>>> On Tue, Feb 11, 2025 at 5:50 AM Dev Jain <dev.jain@arm.com> wrote:
>>>>>>
>>>>>>
>>>>>>
>>>>>> On 11/02/25 6:00 am, Nico Pache wrote:
>>>>>>> The following series provides khugepaged and madvise collapse with the
>>>>>>> capability to collapse regions to mTHPs.
>>>>>>>
>>>>>>> To achieve this we generalize the khugepaged functions to no longer depend
>>>>>>> on PMD_ORDER. Then during the PMD scan, we keep track of chunks of pages
>>>>>>> (defined by MTHP_MIN_ORDER) that are utilized. This info is tracked
>>>>>>> using a bitmap. After the PMD scan is done, we do binary recursion on the
>>>>>>> bitmap to find the optimal mTHP sizes for the PMD range. The restriction
>>>>>>> on max_ptes_none is removed during the scan, to make sure we account for
>>>>>>> the whole PMD range. max_ptes_none will be scaled by the attempted collapse
>>>>>>> order to determine how full a THP must be to be eligible. If a mTHP collapse
>>>>>>> is attempted, but contains swapped out, or shared pages, we dont perform the
>>>>>>> collapse.
>>>>>>>
>>>>>>> With the default max_ptes_none=511, the code should keep its most of its
>>>>>>> original behavior. To exercise mTHP collapse we need to set max_ptes_none<=255.
>>>>>>> With max_ptes_none > HPAGE_PMD_NR/2 you will experience collapse "creep" and
>>>>>>> constantly promote mTHPs to the next available size.
>>>>>>>
>>>>>>> Patch 1:     Some refactoring to combine madvise_collapse and khugepaged
>>>>>>> Patch 2:     Refactor/rename hpage_collapse
>>>>>>> Patch 3-5:   Generalize khugepaged functions for arbitrary orders
>>>>>>> Patch 6-9:   The mTHP patches
>>>>>>>
>>>>>>> ---------
>>>>>>>      Testing
>>>>>>> ---------
>>>>>>> - Built for x86_64, aarch64, ppc64le, and s390x
>>>>>>> - selftests mm
>>>>>>> - I created a test script that I used to push khugepaged to its limits while
>>>>>>>        monitoring a number of stats and tracepoints. The code is available
>>>>>>>        here[1] (Run in legacy mode for these changes and set mthp sizes to inherit)
>>>>>>>        The summary from my testings was that there was no significant regression
>>>>>>>        noticed through this test. In some cases my changes had better collapse
>>>>>>>        latencies, and was able to scan more pages in the same amount of time/work,
>>>>>>>        but for the most part the results were consistant.
>>>>>>> - redis testing. I tested these changes along with my defer changes
>>>>>>>       (see followup post for more details).
>>>>>>> - some basic testing on 64k page size.
>>>>>>> - lots of general use. These changes have been running in my VM for some time.
>>>>>>>
>>>>>>> Changes since V1 [2]:
>>>>>>> - Minor bug fixes discovered during review and testing
>>>>>>> - removed dynamic allocations for bitmaps, and made them stack based
>>>>>>> - Adjusted bitmap offset from u8 to u16 to support 64k pagesize.
>>>>>>> - Updated trace events to include collapsing order info.
>>>>>>> - Scaled max_ptes_none by order rather than scaling to a 0-100 scale.
>>>>>>> - No longer require a chunk to be fully utilized before setting the bit. Use
>>>>>>>        the same max_ptes_none scaling principle to achieve this.
>>>>>>> - Skip mTHP collapse that requires swapin or shared handling. This helps prevent
>>>>>>>        some of the "creep" that was discovered in v1.
>>>>>>>
>>>>>>> [1] - https://gitlab.com/npache/khugepaged_mthp_test
>>>>>>> [2] - https://lore.kernel.org/lkml/20250108233128.14484-1-npache@redhat.com/
>>>>>>>
>>>>>>> Nico Pache (9):
>>>>>>>       introduce khugepaged_collapse_single_pmd to unify khugepaged and
>>>>>>>         madvise_collapse
>>>>>>>       khugepaged: rename hpage_collapse_* to khugepaged_*
>>>>>>>       khugepaged: generalize hugepage_vma_revalidate for mTHP support
>>>>>>>       khugepaged: generalize alloc_charge_folio for mTHP support
>>>>>>>       khugepaged: generalize __collapse_huge_page_* for mTHP support
>>>>>>>       khugepaged: introduce khugepaged_scan_bitmap for mTHP support
>>>>>>>       khugepaged: add mTHP support
>>>>>>>       khugepaged: improve tracepoints for mTHP orders
>>>>>>>       khugepaged: skip collapsing mTHP to smaller orders
>>>>>>>
>>>>>>>      include/linux/khugepaged.h         |   4 +
>>>>>>>      include/trace/events/huge_memory.h |  34 ++-
>>>>>>>      mm/khugepaged.c                    | 422 +++++++++++++++++++----------
>>>>>>>      3 files changed, 306 insertions(+), 154 deletions(-)
>>>>>>>
>>>>>>
>>>>>> Does this patchset suffer from the problem described here:
>>>>>> https://lore.kernel.org/all/8abd99d5-329f-4f8d-8680-c2d48d4963b6@arm.com/
>>>>> Hi Dev,
>>>>>
>>>>> Sorry I meant to get back to you about that.
>>>>>
>>>>> I understand your concern, but like I've mentioned before, the scan
>>>>> with the read lock was done so we dont have to do the more expensive
>>>>> locking, and could still gain insight into the state. You are right
>>>>> that this info could become stale if the state changes dramatically,
>>>>> but the collapse_isolate function will verify it and not collapse.
>>>>
>>>> If the state changes dramatically, the _isolate function will verify it,
>>>> and fallback. And this fallback happens after following this costly
>>>> path: retrieve a large folio from the buddy allocator -> swapin pages
>>>> from the disk -> mmap_write_lock() -> anon_vma_lock_write() -> TLB flush
>>>> on all CPUs -> fallback in _isolate().
>>>> If you do fail in _isolate(), doesn't it make sense to get the updated
>>>> state for the next fallback order immediately, because we have prior
>>>> information that we failed because of PTE state? What your algorithm
>>>> will do is *still* follow the costly path described above, and again
>>>> fail in _isolate(), instead of failing in hpage_collapse_scan_pmd() like
>>>> mine would.
>>>
>>> You do raise a valid point here, I can optimize my solution by
>>> detecting certain collapse failure types and jump to the next scan.
>>> I'll add that to my solution, thanks!
>>>
>>> As for the disagreement around the bitmap, we'll leave that up to the
>>> community to decide since we have differing opinions/solutions.
>>>
>>>>
>>>> The verification of the PTE state by the _isolate() function is the "no
>>>> turning back" point of the algorithm. The verification by
>>>> hpage_collapse_scan_pmd() is the "let us see if proceeding is even worth
>>>> it, before we do costly operations" point of the algorithm.
>>>>
>>>>>    From my testing I found this to rarely happen.
>>>>
>>>> Unfortunately, I am not very familiar with performance testing/load
>>>> testing, I am fairly new to kernel programming, so I am getting there.
>>>> But it really depends on the type of test you are running, what actually
>>>> runs on memory-intensive systems, etc etc. In fact, on loaded systems I
>>>> would expect the PTE state to dramatically change. But still, no opinion
>>>> here.
>>>
>>> Yeah there are probably some cases where it happens more often.
>>> Probably in cases of short lived allocations, but khugepaged doesn't
>>> run that frequently so those won't be that big of an issue.
>>>
>>> Our performance team is currently testing my implementation so I
>>> should have more real workload test results soon. The redis testing
>>> had some gains and didn't show any signs of obvious regressions.
>>>
>>> As for the testing, check out
>>> https://gitlab.com/npache/khugepaged_mthp_test/-/blob/master/record-khuge-performance.sh?ref_type=heads
>>> this does the tracing for my testing script. It can help you get
>>> started. There are 3 different traces being applied there: the
>>> bpftrace for collapse latencies, the perf record for the flamegraph
>>> (not actually that useful, but may be useful to visualize any
>>> weird/long paths that you may not have noticed), and the trace-cmd
>>> which records the tracepoint of the scan and the collapse functions
>>> then processes the data using the awk script-- the output being the
>>> scan rate, the pages collapsed, and their result status (grouped by
>>> order).
>>>
>>> You can also look into https://github.com/gormanm/mmtests for
>>> testing/comparing kernels. I was running the
>>> config-memdb-redis-benchmark-medium workload.
>>
>> Thanks. I'll take a look.
>>
>>>
>>>>
>>>>>
>>>>> Also, khugepaged, my changes, and your changes are all a victim of
>>>>> this. Once we drop the read lock (to either allocate the folio, or
>>>>> right before acquiring the write_lock), the state can change. In your
>>>>> case, yes, you are gathering more up to date information, but is it
>>>>> really that important/worth it to retake locks and rescan for each
>>>>> instance if we are about to reverify with the write lock taken?
>>>>
>>>> You said "reverify": You are removing the verification, so this step
>>>> won't be reverification, it will be verification. We do not want to
>>>> verify *after* we have already done 95% of latency-heavy stuff, only to
>>>> know that we are going to fail.
>>>>
>>>> Algorithms in the kernel, in general, are of the following form: 1)
>>>> Verify if a condition is true, resulting in taking a control path -> 2)
>>>> do a lot of stuff -> "no turning back" step, wherein before committing
>>>> (by taking locks, say), reverify if this is the control path we should
>>>> be in. You are eliminating step 1).
>>>>
>>>> Therefore, I will have to say that I disagree with your approach.
>>>>
>>>> On top of this, in the subjective analysis in [1], point number 7 (along
>>>> with point number 1) remains. And, point number 4 remains.
>>>
>>> for 1) your worst case of 1024 is not the worst case. There are 8
>>> possible orders in your implementation, if all are enabled, that is
>>> 4096 iterations in the worst case.
>>
>> Yes, that is exactly what I wrote in 1). I am still not convinced that
>> the overhead you produce + 512 iterations is going to beat 4096
>> iterations. Anyways, that is hand-waving and we should test this.
>>
>>> This becomes WAY worse on 64k page size, ~45,000 iterations vs 4096 in my case.
>>
>> Sorry, I am missing something here; how does the number of iterations
>> change with page size? Am I not scanning the PTE table, which is
>> invariant to the page size?
> 
> I got the calculation wrong the first time and it's actually worst.
> Lets hope I got this right this time
> on ARM64 64k kernel:
> PMD size = 512M
> PTE= 64k
> PTEs per PMD = 8192

*facepalm* my bad, thanks. I got thrown off thinking HPAGE_PMD_NR won't 
depend on page size, but #pte entries = PAGE_SIZE / sizeof(pte) = 
PAGE_SIZE / 8. So it does depend. You are correct, the PTEs per PMD is 1 
<< 13.

> log2(8192) = 13 - 2 = 11 number of (m)THP sizes including PMD (the
> first and second order are skipped)
> 
> Assuming I understand your algorithm correctly, in the worst case you
> are scanning the whole PMD for each order.
> 
> So you scan 8192 PTEs 11 times. 8192 * 11 = 90112.

Yup. Now it seems that the bitmap overhead may just be worth it; for the 
worst case the bitmap will give us an 11x saving...for the average case, 
it will give us 2x, but still, 8192 is a large number. I'll think of 
ways to test this out.

Btw, I was made aware that an LWN article just got posted on our work!
https://lwn.net/Articles/1009039/

> 
> Please let me know if I'm missing something here.
>>
>>>>
>>>> [1]
>>>> https://lore.kernel.org/all/23023f48-95c6-4a24-ac8b-aba4b1a441b4@arm.com/
>>>>
>>>>>
>>>>> So in my eyes, this is not a "problem"
>>>>
>>>> Looks like the kernel scheduled us for a high-priority debate, I hope
>>>> there's no deadlock :)
>>>>
>>>>>
>>>>> Cheers,
>>>>> -- Nico
>>>>>
>>>>>
>>>>>>
>>>>>
>>>>
>>>
>>
>