Kefeng Wang <wangkefeng.wang@xxxxxxxxxx> writes:
> On 2024/10/30 9:04, Huang, Ying wrote:
>> David Hildenbrand <david@xxxxxxxxxx> writes:
>>
>>> On 29.10.24 14:04, Kefeng Wang wrote:
>>>>>>>>>>>
>>>>>>>>>>> That should all be cleaned up ... process_huge_page() likely
>>>>>>>>>>> shouldn't
>>>>>>>>>>
>>>>>>>>>> Yes, let's fix the bug firstly,
>>>>>>>>>>
>>>>>>>>>>> be even consuming "nr_pages".
>>>>>>>>>>
>>>>>>>>>> No sure about this part, it uses nr_pages as the end and calculate
>>>>>>>>>> the
>>>>>>>>>> 'base'.
>>>>>>>>>
>>>>>>>>> It should be using folio_nr_pages().
>>>>>>>>
>>>>>>>> But process_huge_page() without an explicit folio argument, I'd like to
>>>>>>>> move the aligned address calculate into the folio_zero_user and
>>>>>>>> copy_user_large_folio(will rename it to folio_copy_user()) in the
>>>>>>>> following cleanup patches, or do it in the fix patches?
>>>>>>>
>>>>>>> First, why does folio_zero_user() call process_huge_page() for *a small
>>>>>>> folio*? Because we like or code to be extra complicated to understand?
>>>>>>> Or am I missing something important?
>>>>>>
>>>>>> The folio_zero_user() used for PMD-sized THP and HugeTLB before, and
>>>>>> after anon mTHP supported, it is used for order-2~order-PMD-order THP
>>>>>> and HugeTLB, so it won't process a small folio if I understand correctly.
>>>>>
>>>>> And unfortunately neither the documentation nor the function name
>>>>> expresses that :(
>>>>>
>>>>> I'm happy to review any patches that improve the situation here :)
>>>>>
>>>> Actually, could we drop the process_huge_page() totally, from my
>>>> testcase[1], process_huge_page() is not better than clear/copy page
>>>> from start to last, and sequential clearing/copying maybe more
>>>> beneficial to the hardware prefetching, and is there a way to let lkp
>>>> to test to check the performance, since the process_huge_page()
>>>> was submitted by Ying, what's your opinion?
>> I don't think that it's a good idea to revert the commit without
>> studying and root causing the issues. I can work together with you on
>> that. If we have solid and well explained data to prove
>> process_huge_page() isn't benefitial, we can revert the commit.
>
>
> Take 'fallocate 20G' as an example， before
>
> Performance counter stats for 'taskset -c 10 fallocate -l 20G
> /mnt/hugetlbfs/test':
IIUC, fallocate will zero pages, but will not touch them at all, right?
If so, no cache benefit from clearing referenced page last.
> 3,118.94 msec task-clock # 0.999 CPUs
> utilized
> 30 context-switches # 0.010 K/sec
> 1 cpu-migrations # 0.000 K/sec
> 136 page-faults # 0.044 K/sec
> 8,092,075,873 cycles #
> 2.594 GHz (92.82%)
> 1,624,587,663 instructions # 0.20 insn per
> cycle (92.83%)
> 395,341,850 branches # 126.755 M/sec
> (92.82%)
> 3,872,302 branch-misses # 0.98% of all
> branches (92.83%)
> 1,398,066,701 L1-dcache-loads # 448.251 M/sec
> (92.82%)
> 58,124,626 L1-dcache-load-misses # 4.16% of all
> L1-dcache accesses (92.82%)
> 1,032,527 LLC-loads # 0.331 M/sec
> (92.82%)
> 498,684 LLC-load-misses # 48.30% of all
> LL-cache accesses (92.84%)
> 473,689,004 L1-icache-loads # 151.875 M/sec
> (92.82%)
> 356,721 L1-icache-load-misses # 0.08% of all
> L1-icache accesses (92.85%)
> 1,947,644,987 dTLB-loads # 624.458 M/sec
> (92.95%)
> 10,185 dTLB-load-misses # 0.00% of all
> dTLB cache accesses (92.96%)
> 474,622,896 iTLB-loads # 152.174 M/sec
> (92.95%)
> 94 iTLB-load-misses # 0.00% of all
> iTLB cache accesses (85.69%)
>
> 3.122844830 seconds time elapsed
>
> 0.000000000 seconds user
> 3.107259000 seconds sys
>
> and after（clear from start to end）
>
> Performance counter stats for 'taskset -c 10 fallocate -l 20G
> /mnt/hugetlbfs/test':
>
> 1,135.53 msec task-clock # 0.999 CPUs
> utilized
> 10 context-switches # 0.009 K/sec
> 1 cpu-migrations # 0.001 K/sec
> 137 page-faults # 0.121 K/sec
> 2,946,673,587 cycles #
> 2.595 GHz (92.67%)
> 1,620,704,205 instructions # 0.55 insn per
> cycle (92.61%)
> 394,595,772 branches # 347.499 M/sec
> (92.60%)
> 130,756 branch-misses # 0.03% of all
> branches (92.84%)
> 1,396,726,689 L1-dcache-loads # 1230.022 M/sec
> (92.96%)
> 338,344 L1-dcache-load-misses # 0.02% of all
> L1-dcache accesses (92.95%)
> 111,737 LLC-loads # 0.098 M/sec
> (92.96%)
> 67,486 LLC-load-misses # 60.40% of all
> LL-cache accesses (92.96%)
> 418,198,663 L1-icache-loads # 368.285 M/sec
> (92.96%)
> 173,764 L1-icache-load-misses # 0.04% of all
> L1-icache accesses (92.96%)
> 2,203,364,632 dTLB-loads # 1940.385 M/sec
> (92.96%)
> 17,195 dTLB-load-misses # 0.00% of all
> dTLB cache accesses (92.95%)
> 418,198,365 iTLB-loads # 368.285 M/sec
> (92.96%)
> 79 iTLB-load-misses # 0.00% of all
> iTLB cache accesses (85.34%)
>
> 1.137015760 seconds time elapsed
>
> 0.000000000 seconds user
> 1.131266000 seconds sys
>
> The IPC improved a lot，less LLC-loads and more L1-dcache-loads， but
> this depends on the implementation of the microarchitecture.
Anyway, we need to avoid (or reduce at least) the pure memory clearing
performance. Have you double checked whether process_huge_page() is
inlined? Perf-profile result can be used to check this too.
When you say from start to end, you mean to use clear_gigantic_page()
directly, or change process_huge_page() to clear page from start to end?
> 1) Will test some rand test to check the different of performance as
> David suggested.
>
> 2) Hope the LKP to run more tests since it is very useful(more test
> set and different machines)
I'm starting to use LKP to test.
--
Best Regards,
Huang, Ying
>
>>
>>> I questioned that just recently [1], and Ying assumed that it still
>>> applies [2].
>>>
>>> c79b57e462b5 ("mm: hugetlb: clear target
>>> sub-page last when clearing huge page”) documents the scenario where
>>> this matters -- anon-w-seq which you also run below.
>>>
>>> If there is no performance benefit anymore, we should rip that
>>> out. But likely we should check on multiple micro-architectures with
>>> multiple #CPU configs that are relevant. c79b57e462b5 used a Xeon E5
>>> v3 2699 with 72 processes on 2 NUMA nodes, maybe your test environment
>>> cannot replicate that?>>
>>>
>>> [1]
>>> https://lore.kernel.org/linux-mm/b8272cb4-aee8-45ad-8dff-353444b3fa74@xxxxxxxxxx/
>>> [2]
>>> https://lore.kernel.org/linux-mm/878quv9lhf.fsf@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx/
>>>
>>>> [1]https://lore.kernel.org/linux-mm/2524689c-08f5-446c-8cb9-924f9db0ee3a@xxxxxxxxxx/
>>>> case-anon-w-seq-mt (tried 2M PMD THP/ 64K mTHP)
>>>> case-anon-w-seq-hugetlb (2M PMD HugeTLB)
>>>
>>> But these are sequential, not random. I'd have thought access +
>>> zeroing would be sequentially either way. Did you run with random
>>> access as well>
>
> Will do.
>> > --
>> Best Regards,
>> Huang, Ying
>>
