[PATCH 90] ptrace_wake_up: add "bool force_wakeup" argument for implicit detach

[Oleg Nesterov]

On 10/15, Roland McGrath wrote:
>
> > 	ptrace_detach(sig) checks valid_signal(sig) to detect the
> > 	explicit detach and passes "bool voluntary" to ptrace_wake_up().
>
> ptrace_detach_task() does.  ptrace_detach() has already bailed out
> if !valid_signal(sig) was really pass in from userland.
>
> valid_signal(0) => true, so this is checking specifically for the
> exit_ptrace() call that passes -1 to ptrace_do_detach()->ptrace_detach_task().
>
> That is all rather nonobvious.  But once you do figure it all out, the
> "voluntary" variable name stops being confusing. :-)

Yes, agreed, this needs a comment. I tried, but can't invent a clean
and concise one... Perhaps,

	/*
	 * If we are called when the tracer does PTRACE_DETACH
	 * valid_signal(sig) == T. Otherwise the caller does the
	 * implicit detach during exit.
	 */

> It is strange to use valid_signal() at all inside ptrace_detach_task(),
> where the possible values for sig are -1 from exit_ptrace(), 0, or a valid
> signal number.

OK, agreed. We can check sig >= 0 instead

> Passing 0 to send_sig_info looks like it will get through and then confuse
> the lower layers.

No, send_sig_info() is called under "if (sig)", but I guess I misunderstood
you.

> If sig==0, remind me why we're not just doing action=UTRACE_DETACH anyway?

Hmm. Again, I am not sure I understand you correctly.
But, if sig == 0, then we only do utrace_control(DETACH), because
sig == 0 means PTRACE_DETACH wants to cancel the signal.

The code does

	action = UTRACE_DETACH;

	if (sig) {
		... change if needed ...
	}

	utrace_control(action);

> But remember that ptrace-tests has only the problem
> cases we've noticed before, not necessarily everything that matters.

Yes, yes, I know.

> We
> should feel comfortable when running the gdb testsuite has no regressions
> vs the vanilla kernel, for a variety of past and present gdb versions.
> (That suite tends to have many failures, but all we need to do is match the
> vanilla kernel's behavior in which ones pass or fail.)

You or Jan have alredy mentioned this testsuite... is it gdb-6.8/sim/testsuite/
which I see in gdb-6.8.tar.gz or something else?

Oleg.