On Mon, Dec 21, 2020 at 03:12:57PM +0100, Jan Kara wrote: > But overall even with GUP woes fixed up, set_page_dirty() called by a PUP > user could still see already truncated page. So it has to deal with it. Thanks! That was really helpful. We have a number of currently-buggy filesystems which assume they can do inode = page->mapping->host without checking that page->mapping is not NULL. Anyway, since I'm changing the set_page_dirty signature for folios, this feels like the right time to pass in the page's mapping. __set_page_dirty() rechecks the mapping under the i_pages lock, so we won't do anything inappropriate if the page has been truncated. You can find the whole thing at https://git.infradead.org/users/willy/pagecache.git/shortlog/refs/heads/folio but the important bit is: - /* Set a page dirty. Return true if this dirtied it */ - int (*set_page_dirty)(struct page *page); + /* Set a folio dirty. Return true if this dirtied it */ + bool (*set_page_dirty)(struct address_space *, struct folio *); I'm kind of tempted to rename it to ->dirty_folio(), but I'm also fine with leaving it this way.
