Root Cause Analysis
The KCSAN data-race occurs in fs/kernfs/inode.c due to kernfs_iop_permission() mutating VFS struct inode fields while holding only a shared reader lock (root->kernfs_iattr_rwsem).
Mechanism and Involved Code
-
Locking and VFS Entry Points:
During path resolution (link_path_walk() -> may_lookup() -> lookup_inode_permission_may_exec() -> inode_permission() -> do_inode_permission() in fs/namei.c), the VFS invokes inode->i_op->permission, which for kernfs files, directories, and symlinks is kernfs_iop_permission() (fs/kernfs/inode.c:273). VFS inode_permission() does not acquire inode->i_rwsem or any other per-inode exclusive lock, so multiple tasks performing concurrent lookups (such as concurrent newfstatat() syscalls) can enter kernfs_iop_permission() simultaneously for the same struct inode.
-
Shared Lock Protecting Inode Mutation:
In kernfs_iop_permission() (fs/kernfs/inode.c:280-290):
- If
MAY_NOT_BLOCK is set (RCU-walk mode), it returns -ECHILD, causing the VFS to fall back to ref-walk mode.
- In ref-walk mode, it acquires
down_read(&root->kernfs_iattr_rwsem) and calls kernfs_refresh_inode(kn, inode) before calling generic_permission(&nop_mnt_idmap, inode, mask).
The same pattern also exists in kernfs_iop_getattr() (fs/kernfs/inode.c:182-196), which acquires down_read(&root->kernfs_iattr_rwsem) and calls kernfs_refresh_inode(kn, inode).
-
Concurrent Writes in kernfs_refresh_inode():
kernfs_refresh_inode() (fs/kernfs/inode.c:165-180) copies persistent attributes from struct kernfs_node *kn onto the VFS struct inode:
inode->i_mode = kn->mode; (fs/kernfs/inode.c:169) β a plain 2-byte write to i_mode at offset 0 of struct inode, which is the exact access flagged in the KCSAN report (write to 0xffff88810051c000 of 2 bytes).
set_inode_attr(inode, attrs) (fs/kernfs/inode.c:155-163) β writes inode->i_uid and inode->i_gid (plain writes) as well as i_atime, i_mtime, and i_ctime (via inode_set_*time_to_ts(), which use WRITE_ONCE()).
set_nlink(inode, kn->dir.subdirs + 2) (fs/kernfs/inode.c:178-179) β writes inode->__i_nlink (via fs/inode.c:449).
-
Violated Invariant:
root->kernfs_iattr_rwsem is taken as a write lock (down_write) when modifying struct kernfs_node attributes (kernfs_setattr(), kernfs_iop_setattr(), kernfs_link_sibling(), kernfs_unlink_sibling(), __kernfs_remove()), so down_read protects kn's fields from changing concurrently. However, down_read allows multiple readers (kernfs_iop_permission() and kernfs_iop_getattr()) to execute concurrently, turning kernfs_refresh_inode() into concurrent unsynchronized writes to the shared struct inode fields (i_mode, i_uid, i_gid, __i_nlink), as well as racing with lockless RCU-walk readers of inode->i_mode/i_uid/i_gid in inode_permission() (sb_permission() and HAS_UNMAPPED_ID() at fs/namei.c:633-650).