Technical Mechanism and Root Cause Analysis
1. Context and Involved Components
- Source Files:
mm/shmem.c: shmem_mknod(), shmem_create(), shmem_link(), shmem_unlink(), shmem_rename2(), shmem_symlink()
fs/stat.c: generic_fillattr(), shmem_getattr()
include/linux/fs.h: i_size_read(), i_size_write()
- Structures and Fields:
struct inode: dir->i_size (offset 0x50, 8 bytes)
2. Root Cause Mechanism
In the Linux VFS layer, querying file or directory attributes via stat() / newfstatat() / statx() invokes vfs_getattr_nosec() and subsequent filesystem-specific getattr operations (in tmpfs, shmem_getattr()). To preserve high concurrency and avoid deadlocks, attribute inspection does not acquire inode->i_rwsem. Instead, generic_fillattr() relies on i_size_read(inode) to inspect inode->i_size concurrently with writers:
/* include/linux/fs.h */
static inline loff_t i_size_read(const struct inode *inode)
{
#if BITS_PER_LONG==32 && defined(CONFIG_SMP)
...
#else
/* Pairs with smp_store_release() in i_size_write() */
return smp_load_acquire(&inode->i_size);
#endif
}
To support lockless i_size_read() calls safely across architectures and prevent data races and tearing, any filesystem code updating inode->i_size must use i_size_write(inode, new_size). On 64-bit architectures, i_size_write() emits an smp_store_release(&inode->i_size, i_size) to pair with smp_load_acquire(), while on 32-bit SMP architectures it manages inode->i_size_seqcount to prevent readers from observing torn 64-bit values.
In mm/shmem.c, tmpfs maintains a fake directory size heuristic using BOGO_DIRENT_SIZE (20 bytes per directory entry) so that directory sizes reported by stat() reflect entry counts. However, shmem_mknod() and other tmpfs directory modification routines modify the parent directory's size via direct arithmetic operations without calling i_size_write():
/* mm/shmem.c:4092 in shmem_mknod() */
dir->i_size += BOGO_DIRENT_SIZE;
(Similar unsynchronized modifications exist in shmem_link(), shmem_unlink(), shmem_rename2(), and shmem_symlink().)
3. Bug Trigger and Failure Condition
When task 8131 (udevd) creates a file or node in a tmpfs directory via openat(..., O_CREAT) -> lookup_open() -> shmem_create() -> shmem_mknod(), it performs a plain read-modify-write on dir->i_size without release semantics or seqcount tracking. Simultaneously, task 4073 issues newfstatat() on the same directory, calling generic_fillattr() -> i_size_read(inode) which executes a marked smp_load_acquire(&inode->i_size).
Under the Linux Kernel Concurrency Sanitizer (KCSAN), concurrent access between a plain/unmarked write and an atomic/marked read constitutes an unsynchronized data race and triggers this KCSAN report. The appropriate fix is updating dir->i_size using i_size_write(dir, dir->i_size + BOGO_DIRENT_SIZE) (or - BOGO_DIRENT_SIZE).