On 2/3/21 2:51 PM, Pavel Tatashin wrote:

On Wed, Feb 3, 2021 at 8:23 AM Joao Martins [off-list ref] wrote:

quoted

On 1/25/21 7:47 PM, Pavel Tatashin wrote:

quoted

When pages are isolated in check_and_migrate_movable_pages() we skip
compound number of pages at a time. However, as Jason noted, it is
not necessary correct that pages[i] corresponds to the pages that
we skipped. This is because it is possible that the addresses in
this range had split_huge_pmd()/split_huge_pud(), and these functions
do not update the compound page metadata.

The problem can be reproduced if something like this occurs:

1. User faulted huge pages.
2. split_huge_pmd() was called for some reason
3. User has unmapped some sub-pages in the range
4. User tries to longterm pin the addresses.

The resulting pages[i] might end-up having pages which are not compound
size page aligned.

Fixes: aa712399c1e8 ("mm/gup: speed up check_and_migrate_cma_pages() on huge page")
Reported-by: Jason Gunthorpe <jgg@nvidia.com>
Signed-off-by: Pavel Tatashin <pasha.tatashin@soleen.com>
Reviewed-by: Jason Gunthorpe <jgg@nvidia.com>
---

[...]

quoted

              /*
               * If we get a page from the CMA zone, since we are going to
               * be pinning these entries, we might as well move them out

@@ -1599,8 +1596,6 @@ static long check_and_migrate_cma_pages(struct mm_struct *mm,
                              }
                      }
              }
-
-             i += step;
      }

Hi Joao,

quoted

With this, longterm gup will 'regress' for hugetlbfs e.g. from ~6k -> ~32k usecs when
pinning a 16G hugetlb file.

Estimate or you actually measured?

It's what I had measured before sending. The ~ is because there's error variance.

quoted

quoted

Splitting can only occur on THP right? If so, perhaps we could retain the @step increment

Yes, I do not think we can split HugePage, only THP.

Right, that's my impression too.

quoted

for compound pages but when !is_transparent_hugepage(head) or just PageHuge(head) like:

+               if (!is_transparent_hugepage(head) && PageCompound(page))
+                       i += (compound_nr(head) - (pages[i] - head));

Or making specific to hugetlbfs:

+               if (PageHuge(head))
+                       i += (compound_nr(head) - (pages[i] - head));

Yes, this is reasonable optimization. I will submit a follow up patch
against linux-next.