On Sun, Nov 23, 2025 at 11:54 AM Qianchang Zhao pioooooooooip@gmail.com wrote:
ksmbd maintains delete-on-close and pending-delete state in ksmbd_inode->m_flags. In vfs_cache.c this field is accessed under inconsistent locking: some paths read and modify m_flags under ci->m_lock while others do so without taking the lock at all.
Examples:
- ksmbd_query_inode_status() and __ksmbd_inode_close() use ci->m_lock when checking or updating m_flags.
- ksmbd_inode_pending_delete(), ksmbd_set_inode_pending_delete(), ksmbd_clear_inode_pending_delete() and ksmbd_fd_set_delete_on_close() used to read and modify m_flags without ci->m_lock.
This creates a potential data race on m_flags when multiple threads open, close and delete the same file concurrently. In the worst case delete-on-close and pending-delete bits can be lost or observed in an inconsistent state, leading to confusing delete semantics (files that stay on disk after delete-on-close, or files that disappear while still in use).
Fix it by:
- Making ksmbd_query_inode_status() look at m_flags under ci->m_lock after dropping inode_hash_lock.
- Adding ci->m_lock protection to all helpers that read or modify m_flags (ksmbd_inode_pending_delete(), ksmbd_set_inode_pending_delete(), ksmbd_clear_inode_pending_delete(), ksmbd_fd_set_delete_on_close()).
- Keeping the existing ci->m_lock protection in __ksmbd_inode_close(), and moving the actual unlink/xattr removal outside the lock.
This unifies the locking around m_flags and removes the data race while preserving the existing delete-on-close behaviour.
Reported-by: Qianchang Zhao pioooooooooip@gmail.com Reported-by: Zhitong Liu liuzhitong1993@gmail.com Cc: stable@vger.kernel.org Signed-off-by: Qianchang Zhao pioooooooooip@gmail.com
I have directly updated your patch and applied it to #ksmbd-for-next-next. Please check the attached patch and let me know if you find any issues. Thanks.