f2fs: fix the inconsistent state of data pages

In get_lock_data_page, if there is a data race between get_dnode_of_data for
node and grab_cache_page for data, f2fs is able to face with the following
BUG_ON(dn.data_blkaddr == NEW_ADDR).

kernel BUG at /home/zeus/f2fs_test/src/fs/f2fs/data.c:251!
 [<ffffffffa044966c>] get_lock_data_page+0x1ec/0x210 [f2fs]
Call Trace:
 [<ffffffffa043b089>] f2fs_readdir+0x89/0x210 [f2fs]
 [<ffffffff811a0920>] ? fillonedir+0x100/0x100
 [<ffffffff811a0920>] ? fillonedir+0x100/0x100
 [<ffffffff811a07f8>] vfs_readdir+0xb8/0xe0
 [<ffffffff811a0b4f>] sys_getdents+0x8f/0x110
 [<ffffffff816d7999>] system_call_fastpath+0x16/0x1b

This bug is able to be occurred when the block address of the data block is
changed after f2fs_put_dnode().
In order to avoid that, this patch fixes the lock order of node and data
blocks in which the node block lock is covered by the data block lock.

Signed-off-by: Jaegeuk Kim <jaegeuk.kim@samsung.com>

diff --git a/fs/f2fs/data.c b/fs/f2fs/data.c
index 91ff93b..05fb5c6 100644 (file)
--- a/fs/f2fs/data.c
+++ b/fs/f2fs/data.c
@@ -233,18 +233,23 @@ struct page *get_lock_data_page(struct inode *inode, pgoff_t index)
        struct page *page;
        int err;
 
+repeat:
+       page = grab_cache_page(mapping, index);
+       if (!page)
+               return ERR_PTR(-ENOMEM);
+
        set_new_dnode(&dn, inode, NULL, NULL, 0);
        err = get_dnode_of_data(&dn, index, LOOKUP_NODE);
-       if (err)
+       if (err) {
+               f2fs_put_page(page, 1);
                return ERR_PTR(err);
+       }
        f2fs_put_dnode(&dn);
 
-       if (dn.data_blkaddr == NULL_ADDR)
+       if (dn.data_blkaddr == NULL_ADDR) {
+               f2fs_put_page(page, 1);
                return ERR_PTR(-ENOENT);
-repeat:
-       page = grab_cache_page(mapping, index);
-       if (!page)
-               return ERR_PTR(-ENOMEM);
+       }
 
        if (PageUptodate(page))
                return page;