ext4: fix data corruption for mmap writes

mpage_submit_page() can race with another process growing i_size and
writing data via mmap to the written-back page. As mpage_submit_page()
samples i_size too early, it may happen that ext4_bio_write_page()
zeroes out too large tail of the page and thus corrupts user data.

Fix the problem by sampling i_size only after the page has been
write-protected in page tables by clear_page_dirty_for_io() call.

Reported-by: Michael Zimmer <michael@swarm64.com>
CC: stable@vger.kernel.org
Fixes: cb20d51
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>

fs/ext4/inode.c

@@ -2124,15 +2124,29 @@ static int ext4_writepage(struct page *page,
 static int mpage_submit_page(struct mpage_da_data *mpd, struct page *page)
 {
 	int len;
-	loff_t size = i_size_read(mpd->inode);
+	loff_t size;
 	int err;
 
 	BUG_ON(page->index != mpd->first_page);
+	clear_page_dirty_for_io(page);
+	/*
+	 * We have to be very careful here!  Nothing protects writeback path
+	 * against i_size changes and the page can be writeably mapped into
+	 * page tables. So an application can be growing i_size and writing
+	 * data through mmap while writeback runs. clear_page_dirty_for_io()
+	 * write-protects our page in page tables and the page cannot get
+	 * written to again until we release page lock. So only after
+	 * clear_page_dirty_for_io() we are safe to sample i_size for
+	 * ext4_bio_write_page() to zero-out tail of the written page. We rely
+	 * on the barrier provided by TestClearPageDirty in
+	 * clear_page_dirty_for_io() to make sure i_size is really sampled only
+	 * after page tables are updated.
+	 */
+	size = i_size_read(mpd->inode);
 	if (page->index == size >> PAGE_SHIFT)
 		len = size & ~PAGE_MASK;
 	else
 		len = PAGE_SIZE;
-	clear_page_dirty_for_io(page);
 	err = ext4_bio_write_page(&mpd->io_submit, page, len, mpd->wbc, false);
 	if (!err)
 		mpd->wbc->nr_to_write--;