On 1/25/21 7:47 PM, Pavel Tatashin wrote:
> 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 <j...@nvidia.com>
> Signed-off-by: Pavel Tatashin <pasha.tatas...@soleen.com>
> Reviewed-by: Jason Gunthorpe <j...@nvidia.com>
> ---
[...]
> /*
> * 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;
> }
>
With this, longterm gup will 'regress' for hugetlbfs e.g. from ~6k -> 32k usecs
when
pinning a 16G hugetlb file.
Splitting can only occur on THP right? If so, perhaps we could retain the @step
increment
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));
