On 17/03/2023 08:23, Yin, Fengwei wrote:
[...]
>>>>> FYI, I'm seeing a perf regression of about 1% when compiling the kernel on
>>>>> Ampere Altra (arm64) with this whole series on top of v6.3-rc1 (In a VM using
>>>>> ext4 filesystem). Looks like instruction aborts are taking much longer and a
>>>>> selection of syscalls are a bit slower. Still hunting down the root cause. Will
>>>>> report once I have conclusive diagnosis.
>>>>
>>>> I'm sorry - I'm struggling to find the exact cause. But its spending over 2x the
>>>> amount of time in the instruction abort handling code once patches 32-36 are
>>>> included. Everything in the flame graph is just taking longer. Perhaps we are
>>>> getting more instruction aborts somehow? I have the flamegraphs if anyone wants
>>>> them - just shout and I'll email them separately.
>>> Thanks a lot to Ryan for sharing the flamegraphs to me. I found the __do_fault()
>>> is called with patch 32-36 while no __do_fault() just with first 31 patches. I
>>> suspect the folio_add_file_rmap_range() missed some PTEs population. Please give
>>> me few days to find the root cause and fix. Sorry for this.
>>
>> You're welcome. Give me a shout once you have a re-spin and I'll rerun the tests.
> Could you please help to try following changes? Thanks in advance.
>
> diff --git a/mm/filemap.c b/mm/filemap.c
> index 40be33b5ee46..137011320c68 100644
> --- a/mm/filemap.c
> +++ b/mm/filemap.c
> @@ -3504,15 +3504,16 @@ static vm_fault_t filemap_map_folio_range(struct vm_fault *vmf,
> if (!pte_none(vmf->pte[count]))
> goto skip;
>
> - if (vmf->address == addr)
> - ret = VM_FAULT_NOPAGE;
> -
> count++;
> continue;
> skip:
> if (count) {
> set_pte_range(vmf, folio, page, count, addr);
> folio_ref_add(folio, count);
> + if ((vmf->address < (addr + count * PAGE_SIZE)) &&
> + (vmf->address >= addr))
> + ret = VM_FAULT_NOPAGE;
> +
> }
>
> count++;
> @@ -3525,6 +3526,9 @@ static vm_fault_t filemap_map_folio_range(struct vm_fault *vmf,
> if (count) {
> set_pte_range(vmf, folio, page, count, addr);
> folio_ref_add(folio, count);
> + if ((vmf->address < (addr + count * PAGE_SIZE)) &&
> + (vmf->address >= addr))
> + ret = VM_FAULT_NOPAGE;
> }
>
> vmf->pte = old_ptep;
>
I'm afraid this hasn't fixed it, and I still see __do_fault(). I'll send the
flame graph over separately.
Given I'm running on ext4, I wasn't expecting to see any large page cache
folios? So I don't think we would have expected this patch to help anyway? (or
perhaps there are still THP folios? But I think they will get PMD mapped?).
>
> Regards
> Yin, Fengwei
>
>>
>>>
>>>
>>> Regards
>>> Yin, Fengwei
>>>
>>>>
>>>>>
>>>>> Thanks,
>>>>> Ryan
>>>>>
>>>>>
>>>>>> + }
>>>>>> +
>>>>>> + if (first)
>>>>>> + nr++;
>>>>>> + } while (page++, --nr_pages > 0);
>>>>>> } else if (folio_test_pmd_mappable(folio)) {
>>>>>> /* That test is redundant: it's for safety or to optimize out */
>>>>>>
>>>>>> @@ -1354,6 +1362,30 @@ void page_add_file_rmap(struct page *page, struct vm_area_struct *vma,
>>>>>> mlock_vma_folio(folio, vma, compound);
>>>>>> }
>>>>>>
>>>>>> +/**
>>>>>> + * page_add_file_rmap - add pte mapping to a file page
>>>>>> + * @page: the page to add the mapping to
>>>>>> + * @vma: the vm area in which the mapping is added
>>>>>> + * @compound: charge the page as compound or small page
>>>>>> + *
>>>>>> + * The caller needs to hold the pte lock.
>>>>>> + */
>>>>>> +void page_add_file_rmap(struct page *page, struct vm_area_struct *vma,
>>>>>> + bool compound)
>>>>>> +{
>>>>>> + struct folio *folio = page_folio(page);
>>>>>> + unsigned int nr_pages;
>>>>>> +
>>>>>> + VM_WARN_ON_ONCE_PAGE(compound && !PageTransHuge(page), page);
>>>>>> +
>>>>>> + if (likely(!compound))
>>>>>> + nr_pages = 1;
>>>>>> + else
>>>>>> + nr_pages = folio_nr_pages(folio);
>>>>>> +
>>>>>> + folio_add_file_rmap_range(folio, page, nr_pages, vma, compound);
>>>>>> +}
>>>>>> +
>>>>>> /**
>>>>>> * page_remove_rmap - take down pte mapping from a page
>>>>>> * @page: page to remove mapping from
>>>>>
>>>>
>>
>>
