Re: [PATCH 2/2] sched: Introduce set_special_state()

[Oleg Nesterov <oleg@redhat.com>]

On 05/01, Peter Zijlstra wrote:
>
> On Tue, May 01, 2018 at 03:59:24PM +0200, Oleg Nesterov wrote:
> > On 05/01, Peter Zijlstra wrote:
> > >
> > > The only code I found that seems to care is ptrace_attach(), where we
> > > wait for JOBCTL_TRAPPING to get cleared. That same function has a
> > > comment about hiding the STOPPED -> RUNNING -> TRACED transition. So I'm
> > > assuming it needs to observe TRACED if it observes !TRAPPING.
> >
> > Yes, exactly.
> >
> > > But I don't think there's enough barriers on that end to guarantee this.
> > > Any ->state load after wait_on_bit() is, afact, free to have happened
> > > before the ->jobctl load.
> >
> > do_wait() does set_current_state() before it checks ->state or anything else.
>
> But how are ptrace_attach() and do_wait() related?

Yes.

> I guess I'm missing
> something fairly fundamental here.

You are missing the fact that ptrace API is very old and ugly ;)

Just one example. If the debugger knows that the task is STOPPED then it has
all rights to do, say,

	ptrace(PTRACE_ATTACH, pid);
	BUG_ON(pid != waitpid(pid, WNOHANG));

Or even do another ptrace() request right after ptrace(PTRACE_ATTACH) returns,
without do_wait().

And unless my memory fools me, gdb even has some test-cases for this... Not sure,
but it certainly looks at tracee->state in /proc before it does PTRACE_ATTACH,
because if it was already STOPPED then gdb won't have any notification from the
tracee.

> Anyway, does the below look ok?

Yes, thanks.

Oleg.