On 1/26/21 6:08 PM, Mike Kravetz wrote:
> On 1/25/21 12:57 PM, Joao Martins wrote:
>> For a given hugepage backing a VA, there's a rather ineficient
>> loop which is solely responsible for storing subpages in the passed
>> pages/vmas array. For each subpage we check whether it's within
>> range or size of @pages and keep incrementing @pfn_offset and a couple
>> other variables per subpage iteration.
>>
>> Simplify this logic and minimize ops per iteration to just
>> store the output page/vma. Instead of incrementing number of @refs
>> iteratively, we do it through a precalculation of @refs and having
>> only a tight loop for storing pinned subpages/vmas.
>>
>> pinning consequently improves considerably, bringing us close to
>> {pin,get}_user_pages_fast:
>>
>> - 16G with 1G huge page size
>> gup_test -f /mnt/huge/file -m 16384 -r 10 -L -S -n 512 -w
>>
>> PIN_LONGTERM_BENCHMARK: ~11k us -> ~4400 us
>> PIN_FAST_BENCHMARK: ~3700 us
>>
>> Signed-off-by: Joao Martins <joao.m.martins@xxxxxxxxxx>
>> ---
>> mm/hugetlb.c | 49 ++++++++++++++++++++++++++++---------------------
>> 1 file changed, 28 insertions(+), 21 deletions(-)
>>
>> diff --git a/mm/hugetlb.c b/mm/hugetlb.c
>> index 016addc8e413..1f7a95bc7c87 100644
>> --- a/mm/hugetlb.c
>> +++ b/mm/hugetlb.c
>> @@ -4789,6 +4789,20 @@ int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
>> goto out;
>> }
>>
>> +static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma,
>> + int refs, struct page **pages,
>> + struct vm_area_struct **vmas)
>> +{
>> + int nr;
>> +
>> + for (nr = 0; nr < refs; nr++) {
>> + if (likely(pages))
>> + pages[nr] = page++;
>> + if (vmas)
>> + vmas[nr] = vma;
>> + }
>> +}
>> +
>> long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
>> struct page **pages, struct vm_area_struct **vmas,
>> unsigned long *position, unsigned long *nr_pages,
>> @@ -4918,28 +4932,16 @@ long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
>> continue;
>> }
>>
>> - refs = 0;
>> + refs = min3(pages_per_huge_page(h) - pfn_offset,
>> + (vma->vm_end - vaddr) >> PAGE_SHIFT, remainder);
>>
>> -same_page:
>> - if (pages)
>> - pages[i] = mem_map_offset(page, pfn_offset);
>> + if (pages || vmas)
>> + record_subpages_vmas(mem_map_offset(page, pfn_offset),
>
> The assumption made here is that mem_map is contiguous for the range of
> pages in the hugetlb page. I do not believe you can make this assumption
> for (gigantic) hugetlb pages which are > MAX_ORDER_NR_PAGES. For example,
>
That would mean get_user_pages_fast() and put_user_pages_fast() are broken for anything
handling PUDs or above? See record_subpages() in gup_huge_pud() or even gup_huge_pgd().
It's using the same page++.
This adjustment below probably is what you're trying to suggest.
Also, nth_page() is slightly more expensive and so the numbers above change from ~4.4k
usecs to ~7.8k usecs.
diff --git a/mm/hugetlb.c b/mm/hugetlb.c
index 1f7a95bc7c87..cf66f8c2f92a 100644
--- a/mm/hugetlb.c
+++ b/mm/hugetlb.c
@@ -4789,15 +4789,16 @@ int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
goto out;
}
-static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma,
+static void record_subpages_vmas(struct page *page, unsigned long pfn_offset,
+ struct vm_area_struct *vma,
int refs, struct page **pages,
struct vm_area_struct **vmas)
{
- int nr;
+ unsigned long nr;
for (nr = 0; nr < refs; nr++) {
if (likely(pages))
- pages[nr] = page++;
+ pages[nr] = mem_map_offset(page, pfn_offset + nr);
if (vmas)
vmas[nr] = vma;
}
@@ -4936,8 +4937,7 @@ long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct
*vma,
(vma->vm_end - vaddr) >> PAGE_SHIFT, remainder);
if (pages || vmas)
- record_subpages_vmas(mem_map_offset(page, pfn_offset),
- vma, refs,
+ record_subpages_vmas(page, pfn_offset, vma, refs,
likely(pages) ? pages + i : NULL,
vmas ? vmas + i : NULL);
