Re: [PATCH v2 5/8] mm/gup: Accelerate thp gup even for "pages != NULL"

[David Hildenbrand <david@redhat.com>]

On 20.06.23 01:10, Peter Xu wrote:
> The acceleration of THP was done with ctx.page_mask, however it'll be
> ignored if **pages is non-NULL.
> 
> The old optimization was introduced in 2013 in 240aadeedc4a ("mm:
> accelerate mm_populate() treatment of THP pages").  It didn't explain why
> we can't optimize the **pages non-NULL case.  It's possible that at that
> time the major goal was for mm_populate() which should be enough back then.

In the past we had these sub-page refcounts for THP. My best guess (and 
I didn't check if that was still the case in 2013) would be that it was 
simpler regarding refcount handling to to do it one-subpage at a time.

But I might be just wrong.

> 
> Optimize thp for all cases, by properly looping over each subpage, doing
> cache flushes, and boost refcounts / pincounts where needed in one go.
> 
> This can be verified using gup_test below:
> 
>    # chrt -f 1 ./gup_test -m 512 -t -L -n 1024 -r 10
> 
> Before:    13992.50 ( +-8.75%)
> After:       378.50 (+-69.62%)
> 
> Signed-off-by: Peter Xu <peterx@redhat.com>
> ---
>   mm/gup.c | 51 ++++++++++++++++++++++++++++++++++++++++++++-------
>   1 file changed, 44 insertions(+), 7 deletions(-)
> 
> diff --git a/mm/gup.c b/mm/gup.c
> index 4a00d609033e..b50272012e49 100644
> --- a/mm/gup.c
> +++ b/mm/gup.c
> @@ -1199,16 +1199,53 @@ static long __get_user_pages(struct mm_struct *mm,
>   			goto out;
>   		}
>   next_page:
> -		if (pages) {
> -			pages[i] = page;
> -			flush_anon_page(vma, page, start);
> -			flush_dcache_page(page);
> -			ctx.page_mask = 0;
> -		}
> -
>   		page_increm = 1 + (~(start >> PAGE_SHIFT) & ctx.page_mask);
>   		if (page_increm > nr_pages)
>   			page_increm = nr_pages;
> +
> +		if (pages) {
> +			struct page *subpage;
> +			unsigned int j;
> +
> +			/*
> +			 * This must be a large folio (and doesn't need to
> +			 * be the whole folio; it can be part of it), do
> +			 * the refcount work for all the subpages too.
> +			 *
> +			 * NOTE: here the page may not be the head page
> +			 * e.g. when start addr is not thp-size aligned.
> +			 * try_grab_folio() should have taken care of tail
> +			 * pages.
> +			 */
> +			if (page_increm > 1) {
> +				struct folio *folio;
> +
> +				/*
> +				 * Since we already hold refcount on the
> +				 * large folio, this should never fail.
> +				 */
> +				folio = try_grab_folio(page, page_increm - 1,
> +						       foll_flags);
> +				if (WARN_ON_ONCE(!folio)) {
> +					/*
> +					 * Release the 1st page ref if the
> +					 * folio is problematic, fail hard.
> +					 */
> +					gup_put_folio(page_folio(page), 1,
> +						      foll_flags);
> +					ret = -EFAULT;
> +					goto out;
> +				}
> +			}
> +
> +			for (j = 0; j < page_increm; j++) {
> +				subpage = nth_page(page, j);
> +				pages[i+j] = subpage;

Doe checkpatch like pages[i+j]? I'd have used spaces around the +.

> +				flush_anon_page(vma, subpage, start + j * PAGE_SIZE);
> +				flush_dcache_page(subpage);
> +			}
> +		}
> +
>   		i += page_increm;
>   		start += page_increm * PAGE_SIZE;
>   		nr_pages -= page_increm;


So, we did the first try_grab_folio() while our page was PMD-mapped 
udner the PT lock and we had sufficient permissions (e.g., mapped 
writable, no unsharing required). With FOLL_PIN, we incremented the 
pincount.


I was wondering if something could have happened ever since we unlocked 
the PT table lock and possibly PTE-mapped the THP. ... but as it's 
already pinned, it cannot get shared during fork() [will stay exclusive].

So we can just take additional pins on that folio.


LGTM, although I do like the GUP-fast way of recording+ref'ing it at a 
central place (see gup_huge_pmd() with record_subpages() and friends), 
not after the effects.

-- 
Cheers,

David / dhildenb