Test Then Set

A first attempt at a lock, using only ordinary loads and stores:

typedef struct { int flag; } lock_t;

void lock(lock_t *m) {
    while (m->flag == 1)      // TEST
        ;                     // spin
    m->flag = 1;              // SET
}
void unlock(lock_t *m) { m->flag = 0; }

Two threads on two cores, flag starting at 0:

step	Thread 1	Thread 2	flag
1	reads flag == 0, exits the loop		0
2	preempted	reads flag == 0, exits the loop	0
3		sets flag = 1, enters	1
4	sets flag = 1, enters		1

Which replacement lock() body restores mutual exclusion on a multiprocessor?