f2fs: decrease the lock granularity during write_begin

This patch reduces the lock granularity during write_begin.
When the system is under memory pressure, it would be better to reduce
the locking time for the data pages.
Signed-off-by: NJaegeuk Kim <jaegeuk.kim@samsung.com>

fs/f2fs/data.c

@@ -909,6 +909,10 @@ static int f2fs_write_begin(struct file *file, struct address_space *mapping,
 	page = grab_cache_page_write_begin(mapping, index, flags);
 	if (!page)
 		return -ENOMEM;
+
+	/* to avoid latency during memory pressure */
+	unlock_page(page);
+
 	*pagep = page;
 
 	if (f2fs_has_inline_data(inode) && (pos + len) <= MAX_INLINE_DATA)
@@ -920,10 +924,18 @@ static int f2fs_write_begin(struct file *file, struct address_space *mapping,
 	f2fs_unlock_op(sbi);
 
 	if (err) {
-		f2fs_put_page(page, 1);
+		f2fs_put_page(page, 0);
 		return err;
 	}
 inline_data:
+	lock_page(page);
+	if (unlikely(page->mapping != mapping)) {
+		f2fs_put_page(page, 1);
+		goto repeat;
+	}
+
+	f2fs_wait_on_page_writeback(page, DATA);
+
 	if ((len == PAGE_CACHE_SIZE) || PageUptodate(page))
 		return 0;