[PATCH] f2fs: call unlock_new_inode() before d_instantiate()

[PATCH] f2fs: call unlock_new_inode() before d_instantiate()

[Eric Biggers]

From: Eric Biggers <ebiggers@google.com>

xfstest generic/429 sometimes hangs on f2fs, caused by a thread being
unable to take a directory's i_rwsem for write in vfs_rmdir().  In the
test, one thread repeatedly creates and removes a directory, and other
threads repeatedly look up a file in the directory.  The bug is that
f2fs_mkdir() calls d_instantiate() before unlock_new_inode(), resulting
in the directory inode being exposed to lookups before it has been fully
initialized.  And with CONFIG_DEBUG_LOCK_ALLOC, unlock_new_inode()
reinitializes ->i_rwsem, corrupting its state when it is already held.

Fix it by calling unlock_new_inode() before d_instantiate().  This
matches what other filesystems do.

Fixes: 57397d86c62d ("f2fs: add inode operations for special inodes")
Signed-off-by: Eric Biggers <ebiggers@google.com>
---
 fs/f2fs/namei.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/fs/f2fs/namei.c b/fs/f2fs/namei.c
index d5098efe577c..3a7ed962d2f7 100644
--- a/fs/f2fs/namei.c
+++ b/fs/f2fs/namei.c
@@ -294,8 +294,8 @@ static int f2fs_create(struct inode *dir, struct dentry *dentry, umode_t mode,
 
 	alloc_nid_done(sbi, ino);
 
-	d_instantiate(dentry, inode);
 	unlock_new_inode(inode);
+	d_instantiate(dentry, inode);
 
 	if (IS_DIRSYNC(dir))
 		f2fs_sync_fs(sbi->sb, 1);
@@ -597,8 +597,8 @@ static int f2fs_symlink(struct inode *dir, struct dentry *dentry,
 	err = page_symlink(inode, disk_link.name, disk_link.len);
 
 err_out:
-	d_instantiate(dentry, inode);
 	unlock_new_inode(inode);
+	d_instantiate(dentry, inode);
 
 	/*
 	 * Let's flush symlink data in order to avoid broken symlink as much as
@@ -661,8 +661,8 @@ static int f2fs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
 
 	alloc_nid_done(sbi, inode->i_ino);
 
-	d_instantiate(dentry, inode);
 	unlock_new_inode(inode);
+	d_instantiate(dentry, inode);
 
 	if (IS_DIRSYNC(dir))
 		f2fs_sync_fs(sbi->sb, 1);
@@ -713,8 +713,8 @@ static int f2fs_mknod(struct inode *dir, struct dentry *dentry,
 
 	alloc_nid_done(sbi, inode->i_ino);
 
-	d_instantiate(dentry, inode);
 	unlock_new_inode(inode);
+	d_instantiate(dentry, inode);
 
 	if (IS_DIRSYNC(dir))
 		f2fs_sync_fs(sbi->sb, 1);
-- 
2.17.0.484.g0c8726318c-goog

Re: [PATCH] f2fs: call unlock_new_inode() before d_instantiate()

[Chao Yu]

On 2018/4/19 6:48, Eric Biggers wrote:
> From: Eric Biggers <ebiggers@google.com>
> 
> xfstest generic/429 sometimes hangs on f2fs, caused by a thread being
> unable to take a directory's i_rwsem for write in vfs_rmdir().  In the
> test, one thread repeatedly creates and removes a directory, and other
> threads repeatedly look up a file in the directory.  The bug is that
> f2fs_mkdir() calls d_instantiate() before unlock_new_inode(), resulting
> in the directory inode being exposed to lookups before it has been fully
> initialized.  And with CONFIG_DEBUG_LOCK_ALLOC, unlock_new_inode()
> reinitializes ->i_rwsem, corrupting its state when it is already held.
> 
> Fix it by calling unlock_new_inode() before d_instantiate().  This
> matches what other filesystems do.
> 
> Fixes: 57397d86c62d ("f2fs: add inode operations for special inodes")
> Signed-off-by: Eric Biggers <ebiggers@google.com>

Reviewed-by: Chao Yu <yuchao0@huawei.com>

Thanks,