Root Cause Analysis
1. Mechanism and Involved Entities
- Files and Functions:
fs/kernfs/inode.c: kernfs_iop_permission(), kernfs_refresh_inode(), set_inode_attr()
fs/namei.c: inode_permission(), do_inode_permission()
include/linux/device_cgroup.h: devcgroup_inode_permission()
- Structs and Fields:
struct inode: i_mode (offset 0, 16-bit umode_t), i_uid, i_gid, __i_nlink, timestamps
struct kernfs_node: mode, iattr
struct kernfs_root: kernfs_iattr_rwsem
During path resolution (e.g., link_path_walk(), may_lookup()), VFS calls inode_permission() to verify access rights. In inode_permission():
retval = do_inode_permission(idmap, inode, mask);
if (unlikely(retval))
return retval;
retval = devcgroup_inode_permission(inode, mask);
if (unlikely(retval))
return retval;
For inodes backed by kernfs (such as sysfs and cgroupfs), do_inode_permission() invokes the filesystem permission callback kernfs_iop_permission():
int kernfs_iop_permission(const struct mnt_idmap *idmap,
struct inode *inode, int mask)
{
struct kernfs_node *kn;
struct kernfs_root *root;
int ret;
if (mask & MAY_NOT_BLOCK)
return -ECHILD;
kn = inode->i_private;
root = kernfs_root(kn);
down_read(&root->kernfs_iattr_rwsem);
kernfs_refresh_inode(kn, inode);
ret = generic_permission(&nop_mnt_idmap, inode, mask);
up_read(&root->kernfs_iattr_rwsem);
return ret;
}
In kernfs_refresh_inode() (fs/kernfs/inode.c:165):
static void kernfs_refresh_inode(struct kernfs_node *kn, struct inode *inode)
{
struct kernfs_iattrs *attrs;
inode->i_mode = kn->mode;
...
2. Locking Invariant Violation and Data Race Trigger
-
Shared Read Lock Mutating Inode Fields:
kernfs_iop_permission() acquires down_read(&root->kernfs_iattr_rwsem) before calling kernfs_refresh_inode(). kernfs_iattr_rwsem is a read-write semaphore per kernfs root, and down_read allows multiple readers to enter concurrently. Because multiple tasks concurrently performing path lookups on the same kernfs inode both hold only the read lock, both tasks can simultaneously enter kernfs_refresh_inode() and execute plain, unsynchronized writes to inode->i_mode = kn->mode; (as well as to inode->i_uid, inode->i_gid, inode->__i_nlink, etc.).
-
Unsynchronized Reads in Common VFS Path:
kernfs_iop_permission() releases root->kernfs_iattr_rwsem via up_read() before returning to inode_permission(). Immediately after do_inode_permission() returns, inode_permission() invokes devcgroup_inode_permission(inode, mask). In devcgroup_inode_permission():
if (likely(!S_ISBLK(inode->i_mode) && !S_ISCHR(inode->i_mode)))
return 0;
This reads inode->i_mode without holding kernfs_iattr_rwsem (or any inode-level lock). When Task 1 is executing devcgroup_inode_permission() on CPU 0 reading inode->i_mode, Task 2 on CPU 1 concurrently executing kernfs_iop_permission() -> kernfs_refresh_inode() writes to inode->i_mode, resulting in the reported read-write data race on inode->i_mode.