After doing more drop_caches stress test on
our products, I found the mistake introduced by
a very recent cleanup [1].

The current rule is that "erofs_get_meta_page"
should be returned with page locked (although
it's mostly unnecessary for read-only fs after
pages are PG_Uptodate), but a fix should be
done for this.

[1] https://lore.kernel.org/r/[email protected]
Fixes: 618f40ea026b ("erofs: use read_cache_page_gfp for erofs_get_meta_page")
Signed-off-by: Gao Xiang <[email protected]>
---
 fs/erofs/data.c | 12 +++++++++---
 1 file changed, 9 insertions(+), 3 deletions(-)

diff --git a/fs/erofs/data.c b/fs/erofs/data.c
index 8a9fcbd0e8ac..e0207dba31cb 100644
--- a/fs/erofs/data.c
+++ b/fs/erofs/data.c
@@ -34,11 +34,17 @@ static void erofs_readendio(struct bio *bio)
 
 struct page *erofs_get_meta_page(struct super_block *sb, erofs_blk_t blkaddr)
 {
-       struct inode *const bd_inode = sb->s_bdev->bd_inode;
-       struct address_space *const mapping = bd_inode->i_mapping;
+       struct address_space *const mapping = sb->s_bdev->bd_inode->i_mapping;
+       struct page *page;
 
-       return read_cache_page_gfp(mapping, blkaddr,
+       page = read_cache_page_gfp(mapping, blkaddr,
                                   mapping_gfp_constraint(mapping, ~__GFP_FS));
+       if (IS_ERR(page))
+               return PTR_ERR(page);
+
+       /* should already be PageUptodate */
+       lock_page(page);
+       return page;
 }
 
 static int erofs_map_blocks_flatmode(struct inode *inode,
-- 
2.17.1

Reply via email to