Re: [PATCH 2/3] ftrace: use struct pid

[Dave Hansen <dave-23VcF4HTsmIX0ybBhKVfKdBPR1lH4CV8@public.gmane.org>]

On Thu, 2008-12-04 at 05:40 -0800, Eric W. Biederman wrote:
> Dave Hansen <dave-23VcF4HTsmIX0ybBhKVfKdBPR1lH4CV8@public.gmane.org> writes:
> > On Thu, 2008-12-04 at 04:56 -0800, Dave Hansen wrote:
> >> On Thu, 2008-12-04 at 04:42 -0800, Eric W. Biederman wrote:
> >> > 
> >> > > +static void clear_ftrace_pid_task(struct pid **pid)
> >> > > +{
> >> > > +     struct task_struct *p;
> >> > > +
> >> >         rcu_read_lock();
> >> > 
> >> > > +     do_each_pid_task(*pid, PIDTYPE_PID, p) {
> >> > > +             clear_tsk_trace_trace(p);
> >> > > +     } while_each_pid_task(*pid, PIDTYPE_PID, p);
> >> >         rcu_read_unlock()
> >> > 
> >> > > +     put_pid(*pid);
> >> > > +
> >> > > +     *pid = NULL;
> >> > > +}
> >> 
> >> Could we get away with sticking the rcu_read_{un}lock() inside
> those
> >> macros?  Those are going to get used in pretty high level code and
> we're
> >> allowed to nest rcu_read_lock().  No danger of deadlocks or lock
> >> inversions.
> >
> > Why don't any of the other users of do_each_pid_task() use
> > rcu_read_lock()?  They all seem to be under
> read_lock(&tasklist_lock)
> > (except one is under a write lock of the same).
> 
> We probably should.  Historically read_lock(&tasklist_lock) implies
> rcu_read_lock().  And the tasklist lock is what we hold when it is
> safe.

So, Dipankar tells me that you really do need rcu_read_lock/unlock() for
the guarantee here; the tasklist lock is not sufficient.  The realtime
kernel will preempt even those sections covered by spinlocks.

-- Dave