Re: [PATCH] ptrace: make PTRACE_DETACH work on non-stopped tracees.

[Oleg Nesterov <oleg@redhat.com>]

On 06/20, Denys Vlasenko wrote:
>
> On 06/19/2013 06:32 PM, Oleg Nesterov wrote:
> > On 06/19, Denys Vlasenko wrote:
> >>
> >> This is a user-visible behavior change.
> >> Do we really have to introduce a separate
> >> PTRACE_NOT_STUPID_DETACH? I hope not.
> >
> > Oh, I think yes.
> >
> >> @@ -1062,7 +1060,8 @@ SYSCALL_DEFINE4(ptrace, long, request, long, pid, unsigned long, addr,
> >>  	}
> >>
> >>  	ret = ptrace_check_attach(child, request == PTRACE_KILL ||
> >> -				  request == PTRACE_INTERRUPT);
> >> +				  request == PTRACE_INTERRUPT ||
> >> +				  request == PTRACE_DETACH);
> >
> > There doesn't look right.
> >
> > For example ptrace_disable(). See the comment set_task_blockstep().
>
> I see the comment. I think it implies that TF-induced debug
> interrupt may happen on the running task after it is detached,
> which will result in SIGTRAP being sent to it.

No. The comment means that set/clear of TIF_BLOCKSTEP is not safe unless
the tracee can't run. If we race with __switch_to() we can set the wrong
debugctlmsr.

> If so, do we have the same problem if tracer exits
> and implicit detach is performed?

No. If the tracer exits it doesn't do the "cleanups" like ptrace_disable().
That is why this potentially leaves the tracee in the inconsistent state.

Oleg.