On Mon, Apr 30, 2018 at 09:40:24PM +0200, Peter Zijlstra wrote: > On Mon, Apr 30, 2018 at 06:45:46PM +0200, Oleg Nesterov wrote: > > On 04/30, Peter Zijlstra wrote: > > > > > > --- a/kernel/signal.c > > > +++ b/kernel/signal.c > > > @@ -1968,7 +1968,7 @@ static void ptrace_stop(int exit_code, i > > > * atomic with respect to siglock and should be done after the arch > > > * hook as siglock is released and regrabbed across it. > > > */ > > > - set_current_state(TASK_TRACED); > > > + set_special_state(TASK_TRACED); > > > > Yes, but please note the comment above, we need a barrier after state = > > TASK_TRACED, > > that is why ptrace_stop() does set_current_state(), not > > __set_current_state(). > > OK, so I got properly lost in that stuff. > > The comment says it we need to set TASK_TRACED before clearing > JOBCTL_TRAPPING because of do_wait(), but I cannot seem to locate code > in do_wait() and below that cares about JOBCTL_TRAPPING. > > Could you elucidate my tired brain?
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. 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.
