On Tue, Sep 10, 2019 at 03:27:02PM +0200, Christian Brauner wrote: > On Tue, Sep 10, 2019 at 03:10:48PM +0200, Christian Brauner wrote: > > On Tue, Sep 10, 2019 at 03:09:35PM +0200, Christian Brauner wrote: > > > On Tue, Sep 10, 2019 at 02:44:41PM +0200, Oleg Nesterov wrote: > > > > On 09/10, Eugene Syromiatnikov wrote: > > > > > > > > > > --- a/kernel/fork.c > > > > > +++ b/kernel/fork.c > > > > > @@ -2562,6 +2562,9 @@ noinline static int > > > > > copy_clone_args_from_user(struct kernel_clone_args *kargs, > > > > > if (copy_from_user(&args, uargs, size)) > > > > > return -EFAULT; > > > > > > > > > > + if (unlikely(((unsigned int)args.exit_signal) != > > > > > args.exit_signal)) > > > > > + return -EINVAL; > > > > > > > > Hmm. Unless I am totally confused you found a serious bug... > > > > > > > > Without CLONE_THREAD/CLONE_PARENT copy_process() blindly does > > > > > > > > p->exit_signal = args->exit_signal; > > > > > > > > the valid_signal(sig) check in do_notify_parent() mostly saves us, but > > > > we > > > > must not allow child->exit_signal < 0, if nothing else this breaks > > > > thread_group_leader(). > > > > > > > > And afaics this patch doesn't fix this? I think we need the > > > > valid_signal() > > > > check... > > > > > > Thanks for sending this patch so quickly after our conversation > > > yesterday, Eugene! > > > We definitely want valid_signal() to verify the signal is ok. > > So we could do your check in copy_clone_args_from_user(), and then we do > another valid_signal() check in clone3_args_valid()? We could do the > latter in copy_clone_args_from_user() too but it's nicer to do it along > the other checks in clone3_args_valid().
There's also a discrepancy between CSIGNAL (0xff) and _NSIG, used in valid_signal (which is between 32 and 128, depending on architecture), it seems it doesn't break thread_group_leader, but definitely allows passing some invalid signal numbers via legacy clone-like syscalls—I'm not sure if that's important. > Christian
