[PATCH v2] mm/hugetlb: avoid looping to the same hugepage if !pages and !vmas

[PATCH v2] mm/hugetlb: avoid looping to the same hugepage if !pages and !vmas

[Zhigang Lu]

From: Zhigang Lu <tonnylu@tencent.com>

When mmapping an existing hugetlbfs file with MAP_POPULATE, we find
it is very time consuming. For example, mmapping a 128GB file takes
about 50 milliseconds. Sampling with perfevent shows it spends 99%
time in the same_page loop in follow_hugetlb_page().

samples: 205  of event 'cycles', Event count (approx.): 136686374
-  99.04%  test_mmap_huget  [kernel.kallsyms]  [k] follow_hugetlb_page
        follow_hugetlb_page
        __get_user_pages
        __mlock_vma_pages_range
        __mm_populate
        vm_mmap_pgoff
        sys_mmap_pgoff
        sys_mmap
        system_call_fastpath
        __mmap64

follow_hugetlb_page() is called with pages=NULL and vmas=NULL, so for
each hugepage, we run into the same_page loop for pages_per_huge_page()
times, but doing nothing. With this change, it takes less then 1
millisecond to mmap a 128GB file in hugetlbfs.

Signed-off-by: Zhigang Lu <tonnylu@tencent.com>
Reviewed-by: Haozhong Zhang <hzhongzhang@tencent.com>
Reviewed-by: Zongming Zhang <knightzhang@tencent.com>
Acked-by: Matthew Wilcox <willy@infradead.org>
---
 mm/hugetlb.c | 11 +++++++++++
 1 file changed, 11 insertions(+)

diff --git a/mm/hugetlb.c b/mm/hugetlb.c
index 6d7296d..2df941a 100644
--- a/mm/hugetlb.c
+++ b/mm/hugetlb.c
@@ -4391,6 +4391,17 @@ long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
 				break;
 			}
 		}
+
+		if (!pages && !vmas && !pfn_offset &&
+		    (vaddr + huge_page_size(h) < vma->vm_end) &&
+		    (remainder >= pages_per_huge_page(h))) {
+			vaddr += huge_page_size(h);
+			remainder -= pages_per_huge_page(h);
+			i += pages_per_huge_page(h);
+			spin_unlock(ptl);
+			continue;
+		}
+
 same_page:
 		if (pages) {
 			pages[i] = mem_map_offset(page, pfn_offset);
-- 
1.8.3.1

Re: [PATCH v2] mm/hugetlb: avoid looping to the same hugepage if !pages and !vmas

[Mike Kravetz]

On 8/29/19 6:50 AM, Zhigang Lu wrote:
> From: Zhigang Lu <tonnylu@tencent.com>
> 
> When mmapping an existing hugetlbfs file with MAP_POPULATE, we find
> it is very time consuming. For example, mmapping a 128GB file takes
> about 50 milliseconds. Sampling with perfevent shows it spends 99%
> time in the same_page loop in follow_hugetlb_page().
> 
> samples: 205  of event 'cycles', Event count (approx.): 136686374
> -  99.04%  test_mmap_huget  [kernel.kallsyms]  [k] follow_hugetlb_page
>         follow_hugetlb_page
>         __get_user_pages
>         __mlock_vma_pages_range
>         __mm_populate
>         vm_mmap_pgoff
>         sys_mmap_pgoff
>         sys_mmap
>         system_call_fastpath
>         __mmap64
> 
> follow_hugetlb_page() is called with pages=NULL and vmas=NULL, so for
> each hugepage, we run into the same_page loop for pages_per_huge_page()
> times, but doing nothing. With this change, it takes less then 1
> millisecond to mmap a 128GB file in hugetlbfs.

Thanks for the analysis!

Just curious, do you have an application that does this (mmap(MAP_POPULATE)
for an existing hugetlbfs file), or was this part of some test suite or
debug code?

> Signed-off-by: Zhigang Lu <tonnylu@tencent.com>
> Reviewed-by: Haozhong Zhang <hzhongzhang@tencent.com>
> Reviewed-by: Zongming Zhang <knightzhang@tencent.com>
> Acked-by: Matthew Wilcox <willy@infradead.org>
> ---
>  mm/hugetlb.c | 11 +++++++++++
>  1 file changed, 11 insertions(+)
> 
> diff --git a/mm/hugetlb.c b/mm/hugetlb.c
> index 6d7296d..2df941a 100644
> --- a/mm/hugetlb.c
> +++ b/mm/hugetlb.c
> @@ -4391,6 +4391,17 @@ long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
>  				break;
>  			}
>  		}

It might be helpful to add a comment here to help readers of the code.
Something like:

		/*
		 * If subpage information not requested, update counters
		 * and skip the same_page loop below.
		 */
> +
> +		if (!pages && !vmas && !pfn_offset &&
> +		    (vaddr + huge_page_size(h) < vma->vm_end) &&
> +		    (remainder >= pages_per_huge_page(h))) {
> +			vaddr += huge_page_size(h);
> +			remainder -= pages_per_huge_page(h);
> +			i += pages_per_huge_page(h);
> +			spin_unlock(ptl);
> +			continue;
> +		}
> +
>  same_page:
>  		if (pages) {
>  			pages[i] = mem_map_offset(page, pfn_offset);
> 

With a comment added to the code,
Reviewed-by: Mike Kravetz <mike.kravetz@oracle.com>
-- 
Mike Kravetz

Re: [PATCH v2] mm/hugetlb: avoid looping to the same hugepage if !pages and !vmas

[陆志刚]

Mike Kravetz <mike.kravetz@oracle.com> 于2019年9月4日周三 上午5:26写道：
>
> On 8/29/19 6:50 AM, Zhigang Lu wrote:
> > From: Zhigang Lu <tonnylu@tencent.com>
> >
> > When mmapping an existing hugetlbfs file with MAP_POPULATE, we find
> > it is very time consuming. For example, mmapping a 128GB file takes
> > about 50 milliseconds. Sampling with perfevent shows it spends 99%
> > time in the same_page loop in follow_hugetlb_page().
> >
> > samples: 205  of event 'cycles', Event count (approx.): 136686374
> > -  99.04%  test_mmap_huget  [kernel.kallsyms]  [k] follow_hugetlb_page
> >         follow_hugetlb_page
> >         __get_user_pages
> >         __mlock_vma_pages_range
> >         __mm_populate
> >         vm_mmap_pgoff
> >         sys_mmap_pgoff
> >         sys_mmap
> >         system_call_fastpath
> >         __mmap64
> >
> > follow_hugetlb_page() is called with pages=NULL and vmas=NULL, so for
> > each hugepage, we run into the same_page loop for pages_per_huge_page()
> > times, but doing nothing. With this change, it takes less then 1
> > millisecond to mmap a 128GB file in hugetlbfs.
>
> Thanks for the analysis!
>
> Just curious, do you have an application that does this (mmap(MAP_POPULATE)
> for an existing hugetlbfs file), or was this part of some test suite or
> debug code?

Yes, DPDK and SPDK actually do this in vhost_user.c.
vhost_setup_mem_table() {
...
mmap_size = RTE_ALIGN_CEIL(mmap_size, alignment);

mmap_addr = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE,
MAP_SHARED | MAP_POPULATE, fd, 0);
...
}

>
> > Signed-off-by: Zhigang Lu <tonnylu@tencent.com>
> > Reviewed-by: Haozhong Zhang <hzhongzhang@tencent.com>
> > Reviewed-by: Zongming Zhang <knightzhang@tencent.com>
> > Acked-by: Matthew Wilcox <willy@infradead.org>
> > ---
> >  mm/hugetlb.c | 11 +++++++++++
> >  1 file changed, 11 insertions(+)
> >
> > diff --git a/mm/hugetlb.c b/mm/hugetlb.c
> > index 6d7296d..2df941a 100644
> > --- a/mm/hugetlb.c
> > +++ b/mm/hugetlb.c
> > @@ -4391,6 +4391,17 @@ long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
> >                               break;
> >                       }
> >               }
>
> It might be helpful to add a comment here to help readers of the code.
> Something like:
>

Thanks, I will add this comment and send a new version.

>                 /*
>                  * If subpage information not requested, update counters
>                  * and skip the same_page loop below.
>                  */
> > +
> > +             if (!pages && !vmas && !pfn_offset &&
> > +                 (vaddr + huge_page_size(h) < vma->vm_end) &&
> > +                 (remainder >= pages_per_huge_page(h))) {
> > +                     vaddr += huge_page_size(h);
> > +                     remainder -= pages_per_huge_page(h);
> > +                     i += pages_per_huge_page(h);
> > +                     spin_unlock(ptl);
> > +                     continue;
> > +             }
> > +
> >  same_page:
> >               if (pages) {
> >                       pages[i] = mem_map_offset(page, pfn_offset);
> >
>
> With a comment added to the code,
> Reviewed-by: Mike Kravetz <mike.kravetz@oracle.com>
> --
> Mike Kravetz