On 05/05/17 18:44, Oleg Nesterov wrote: > On 05/05, Vegard Nossum wrote: >> >> If a kthread forks (e.g. usermodehelper since commit 1da5c46fa965) but >> fails in copy_process() between calling dup_task_struct() and setting >> p->set_child_tid, then the value of p->set_child_tid will be inherited >> from the parent and get prematurely freed by free_kthread_struct(). > > Aaah... thanks! > >> --- a/kernel/fork.c >> +++ b/kernel/fork.c >> @@ -518,6 +518,13 @@ static struct task_struct *dup_task_struct(struct task_struct *orig, int node) >> atomic_set(&tsk->stack_refcount, 1); >> #endif >> >> + /* >> + * Forking kthreads (e.g. usermodehelper) should not inherit this >> + * field since it's a pointer to a 'struct kthread' which is not >> + * reference counted. >> + */ >> + tsk->set_child_tid = NULL; >> + > > Can't we just move both > > p->set_child_tid = (clone_flags & CLONE_CHILD_SETTID) ? child_tidptr : NULL; > /* > * Clear TID on mm_release()? > */ > p->clear_child_tid = (clone_flags & CLONE_CHILD_CLEARTID) ? child_tidptr : NULL; > > lines here? clone_flags is not available in dup_task_struct(), but we could move those lines higher in copy_process(). The reason we didn't do it was that we thought it was a little fragile/unobvious that this has to happen before free_task() is called and that it was safer to clear it in dup_task_struct() (which also contains zeroing of other fields). The newly attached patch has been tested and seems to work, if you prefer it. Vegard