On 03/15, Roland McGrath wrote:
>
>
> > > Does that need a barrier pair here and in
> >
> > No, set_notify_resume()->test_and_set_tsk_thread_flag() implies mb(),
>
> Ah, ok.
>
> > > tracehook_notify_resume()?
> >
> > Ah. I think you are right, and I think it needs the barrier even without
> > this change. Say, UTRACE_REPORT does:
> >
> > utrace->report = 1;
> > set_notify_resume();
> >
> > Without mb() there is no guarantee that utrace_resume() will notice and
> > clear ->report.
>
> Wait, what? You just said that set_notify_resume() already implies an mb().
Yes, but the other side lacks a barrier. UTRACE_REPORT does
utrace->report = 1;
wmb(); // actually mb, but wmb is enough
set _TIF_NOTIFY_RESUME;
do_notify_resume()->utrace_resume()->start_report() path does
if (_TIF_NOTIFY_RESUME)
// !!! we need rmb in between !!!
if (utrace->report)
...
and it can miss ->report.
> But we don't want a solution that requires changing
> arch code.
Yes, agreed.
Oleg.
