On Wed, Mar 10, 2021 at 04:54:49PM +0000, David Howells wrote:

Add a function, unlock_page_private_2(), to unlock PG_private_2 analogous
to that of PG_lock.  Add a kerneldoc banner to that indicating the example
usage case.

This isn't a problem with this patch per se, but I'm concerned about
private2 and expected page refcounts.

static inline int is_page_cache_freeable(struct page *page)
{
        /*
         * A freeable page cache page is referenced only by the caller
         * that isolated the page, the page cache and optional buffer
         * heads at page->private.
         */
        int page_cache_pins = thp_nr_pages(page);
        return page_count(page) - page_has_private(page) == 1 + page_cache_pins;
}

static inline int page_has_private(struct page *page)
{
        return !!(page->flags & PAGE_FLAGS_PRIVATE);
}

#define PAGE_FLAGS_PRIVATE                              \
        (1UL << PG_private | 1UL << PG_private_2)

So ... a page with both flags cleared should have a refcount of N.
A page with one or both flags set should have a refcount of N+1.

How is a poor filesystem supposed to make that true?  Also btrfs has this
problem since it uses private_2 for its own purposes.