mm/gup: Handle page split race more efficiently

If we hit the page split race, the current code returns NULL which will
presumably trigger a retry under the mmap_lock.  This isn't necessary;
we can just retry the compound_head() lookup.  This is a very minor
optimisation of an unlikely path, but conceptually it matches (eg)
the page cache RCU-protected lookup.

Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: John Hubbard <jhubbard@nvidia.com>
Reviewed-by: Jason Gunthorpe <jgg@nvidia.com>
Reviewed-by: William Kucharski <william.kucharski@oracle.com>

diff --git a/mm/gup.c b/mm/gup.c
index ad120f4..d3e8266 100644 (file)
--- a/mm/gup.c
+++ b/mm/gup.c
@@ -68,7 +68,10 @@ static void put_page_refs(struct page *page, int refs)
  */
 static inline struct page *try_get_compound_head(struct page *page, int refs)
 {
-       struct page *head = compound_head(page);
+       struct page *head;
+
+retry:
+       head = compound_head(page);
 
        if (WARN_ON_ONCE(page_ref_count(head) < 0))
                return NULL;
@@ -86,7 +89,7 @@ static inline struct page *try_get_compound_head(struct page *page, int refs)
         */
        if (unlikely(compound_head(page) != head)) {
                put_page_refs(head, refs);
-               return NULL;
+               goto retry;
        }
 
        return head;