Re: [PATCH 3/3] mm: Allow find_get_page to be used for large pages

["Kirill A. Shutemov" <kirill@shutemov.name>]

On Thu, Sep 05, 2019 at 11:23:48AM -0700, Matthew Wilcox wrote:
> From: "Matthew Wilcox (Oracle)" <willy@infradead.org>
> 
> Add FGP_PMD to indicate that we're trying to find-or-create a page that
> is at least PMD_ORDER in size.  The internal 'conflict' entry usage
> is modelled after that in DAX, but the implementations are different
> due to DAX using multi-order entries and the page cache using multiple
> order-0 entries.
> 
> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
> ---
>  include/linux/pagemap.h |  9 +++++
>  mm/filemap.c            | 82 +++++++++++++++++++++++++++++++++++++----
>  2 files changed, 84 insertions(+), 7 deletions(-)
> 
> diff --git a/include/linux/pagemap.h b/include/linux/pagemap.h
> index d2147215d415..72101811524c 100644
> --- a/include/linux/pagemap.h
> +++ b/include/linux/pagemap.h
> @@ -248,6 +248,15 @@ pgoff_t page_cache_prev_miss(struct address_space *mapping,
>  #define FGP_NOFS		0x00000010
>  #define FGP_NOWAIT		0x00000020
>  #define FGP_FOR_MMAP		0x00000040
> +/*
> + * If you add more flags, increment FGP_ORDER_SHIFT (no further than 25).

Maybe some BUILD_BUG_ON()s to ensure FGP_ORDER_SHIFT is sane?

> + * Do not insert flags above the FGP order bits.
> + */
> +#define FGP_ORDER_SHIFT		7
> +#define FGP_PMD			((PMD_SHIFT - PAGE_SHIFT) << FGP_ORDER_SHIFT)
> +#define FGP_PUD			((PUD_SHIFT - PAGE_SHIFT) << FGP_ORDER_SHIFT)
> +
> +#define fgp_order(fgp)		((fgp) >> FGP_ORDER_SHIFT)
>  
>  struct page *pagecache_get_page(struct address_space *mapping, pgoff_t offset,
>  		int fgp_flags, gfp_t cache_gfp_mask);
> diff --git a/mm/filemap.c b/mm/filemap.c
> index ae3c0a70a8e9..904dfabbea52 100644
> --- a/mm/filemap.c
> +++ b/mm/filemap.c
> @@ -1572,7 +1572,71 @@ struct page *find_get_entry(struct address_space *mapping, pgoff_t offset)
>  
>  	return page;
>  }
> -EXPORT_SYMBOL(find_get_entry);
> +
> +static bool pagecache_is_conflict(struct page *page)
> +{
> +	return page == XA_RETRY_ENTRY;
> +}
> +
> +/**
> + * __find_get_page - Find and get a page cache entry.
> + * @mapping: The address_space to search.
> + * @offset: The page cache index.
> + * @order: The minimum order of the entry to return.
> + *
> + * Looks up the page cache entries at @mapping between @offset and
> + * @offset + 2^@order.  If there is a page cache page, it is returned with

Off by one? :P

> + * an increased refcount unless it is smaller than @order.
> + *
> + * If the slot holds a shadow entry of a previously evicted page, or a
> + * swap entry from shmem/tmpfs, it is returned.
> + *
> + * Return: the found page, a value indicating a conflicting page or %NULL if
> + * there are no pages in this range.
> + */
> +static struct page *__find_get_page(struct address_space *mapping,
> +		unsigned long offset, unsigned int order)
> +{
> +	XA_STATE(xas, &mapping->i_pages, offset);
> +	struct page *page;
> +
> +	rcu_read_lock();
> +repeat:
> +	xas_reset(&xas);
> +	page = xas_find(&xas, offset | ((1UL << order) - 1));

Hm. '|' is confusing. What is expectation about offset?
Is round_down(offset, 1UL << order) expected to be equal offset?
If yes, please use '+' instead of '|'.

> +	if (xas_retry(&xas, page))
> +		goto repeat;
> +	/*
> +	 * A shadow entry of a recently evicted page, or a swap entry from
> +	 * shmem/tmpfs.  Skip it; keep looking for pages.
> +	 */
> +	if (xa_is_value(page))
> +		goto repeat;
> +	if (!page)
> +		goto out;
> +	if (compound_order(page) < order) {
> +		page = XA_RETRY_ENTRY;
> +		goto out;
> +	}

compound_order() is not stable if you don't have pin on the page.
Check it after page_cache_get_speculative().

> +
> +	if (!page_cache_get_speculative(page))
> +		goto repeat;
> +
> +	/*
> +	 * Has the page moved or been split?
> +	 * This is part of the lockless pagecache protocol. See
> +	 * include/linux/pagemap.h for details.
> +	 */
> +	if (unlikely(page != xas_reload(&xas))) {
> +		put_page(page);
> +		goto repeat;
> +	}
> +	page = find_subpage(page, offset);
> +out:
> +	rcu_read_unlock();
> +
> +	return page;
> +}
>  
>  /**
>   * find_lock_entry - locate, pin and lock a page cache entry
> @@ -1614,12 +1678,12 @@ EXPORT_SYMBOL(find_lock_entry);
>   * pagecache_get_page - find and get a page reference
>   * @mapping: the address_space to search
>   * @offset: the page index
> - * @fgp_flags: PCG flags
> + * @fgp_flags: FGP flags
>   * @gfp_mask: gfp mask to use for the page cache data page allocation
>   *
>   * Looks up the page cache slot at @mapping & @offset.
>   *
> - * PCG flags modify how the page is returned.
> + * FGP flags modify how the page is returned.
>   *
>   * @fgp_flags can be:
>   *
> @@ -1632,6 +1696,10 @@ EXPORT_SYMBOL(find_lock_entry);
>   * - FGP_FOR_MMAP: Similar to FGP_CREAT, only we want to allow the caller to do
>   *   its own locking dance if the page is already in cache, or unlock the page
>   *   before returning if we had to add the page to pagecache.
> + * - FGP_PMD: We're only interested in pages at PMD granularity.  If there
> + *   is no page here (and FGP_CREATE is set), we'll create one large enough.
> + *   If there is a smaller page in the cache that overlaps the PMD page, we
> + *   return %NULL and do not attempt to create a page.

Is it really the best inteface?

Maybe allow user to ask bitmask of allowed orders? For THP order-0 is fine
if order-9 has failed.

>   *
>   * If FGP_LOCK or FGP_CREAT are specified then the function may sleep even
>   * if the GFP flags specified for FGP_CREAT are atomic.
> @@ -1646,9 +1714,9 @@ struct page *pagecache_get_page(struct address_space *mapping, pgoff_t offset,
>  	struct page *page;
>  
>  repeat:
> -	page = find_get_entry(mapping, offset);
> -	if (xa_is_value(page))
> -		page = NULL;
> +	page = __find_get_page(mapping, offset, fgp_order(fgp_flags));
> +	if (pagecache_is_conflict(page))
> +		return NULL;
>  	if (!page)
>  		goto no_page;
>  
> @@ -1682,7 +1750,7 @@ struct page *pagecache_get_page(struct address_space *mapping, pgoff_t offset,
>  		if (fgp_flags & FGP_NOFS)
>  			gfp_mask &= ~__GFP_FS;
>  
> -		page = __page_cache_alloc(gfp_mask);
> +		page = __page_cache_alloc_order(gfp_mask, fgp_order(fgp_flags));
>  		if (!page)
>  			return NULL;
>  
> -- 
> 2.23.0.rc1
> 

-- 
 Kirill A. Shutemov