On Thu, Dec 2, 2021 at 12:18 AM Matthew Wilcox (Oracle)
<willy@xxxxxxxxxxxxx> wrote:
> Commit 17839856fd58 ("gup: document and work around "COW can break either
> way" issue") forces a COW break, even for read-only GUP. This interacts
> badly with CONFIG_READ_ONLY_THP_FOR_FS as it tries to write to a read-only
> PMD and follow_trans_huge_pmd() returns NULL which induces an endless
> loop as __get_user_pages() interprets that as page-not-present, tries
> to fault it in and retries the follow_page_mask().
>
> The issues fixed by 17839856fd58 don't apply to files. We know which way
> the COW breaks; the page cache keeps the original and any modifications
> are private to that process. There's no optimisation that allows a
> process to reuse a file-backed MAP_PRIVATE page. So we can skip the
> breaking of the COW for file-backed mappings.
>
> This problem only exists in v5.4.y; other stable kernels either predate
> CONFIG_READ_ONLY_THP_FOR_FS or they include commit a308c71bf1e6 ("mm/gup:
> Remove enfornced COW mechanism").
>
> Signed-off-by: Matthew Wilcox (Oracle) <willy@xxxxxxxxxxxxx>
> ---
> mm/gup.c | 3 ++-
> 1 file changed, 2 insertions(+), 1 deletion(-)
>
> diff --git a/mm/gup.c b/mm/gup.c
> index 3ef769529548..d55e02411010 100644
> --- a/mm/gup.c
> +++ b/mm/gup.c
> @@ -176,7 +176,8 @@ static inline bool can_follow_write_pte(pte_t pte, unsigned int flags)
> */
> static inline bool should_force_cow_break(struct vm_area_struct *vma, unsigned int flags)
> {
> - return is_cow_mapping(vma->vm_flags) && (flags & FOLL_GET);
> + return is_cow_mapping(vma->vm_flags) && vma_is_anonymous(vma) &&
> + (flags & FOLL_GET);
> }
To be fully correct, the check would have to check for PageAnon(), not
whether the mapping is anonymous, right? Since a private file mapping
can still contain anonymous pages from a prior CoW?
