On Tue, Aug 25, 2020 at 10:59:54PM +0900, Masami Hiramatsu wrote:
> On Tue, 25 Aug 2020 15:30:05 +0200
> pet...@infradead.org wrote:
>
> > On Tue, Aug 25, 2020 at 10:15:55PM +0900, Masami Hiramatsu wrote:
> >
> > > > damn... one last problem is dangling instances.. so close.
> > > > We can apparently unregister a kretprobe while there's still active
> > > > kretprobe_instance's out referencing it.
> > >
> > > Yeah, kretprobe already provided the per-instance data (as far as
> > > I know, only systemtap depends on it). We need to provide it for
> > > such users.
> > > But if we only have one lock, we can avoid checking NMI because
> > > we can check the recursion with trylock. It is needed only if the
> > > kretprobe uses per-instance data. Or we can just pass a dummy
> > > instance on the stack.
> >
> > I think it is true in general, you can unregister a rp while tasks are
> > preempted.
>
> Would you mean the kretprobe handler (or trampoline handler) will be
> preempted? All kprobes (including kretprobe) handler is running in
> non-preemptive state, so it shouldn't happen...
I was thinking about something like:
for_each_process_thread(p, t) {
if (!t->kretprobe_instances.first)
continue;
again:
if (try_invoke_on_locked_down_task(t, unhook_rp_inst, tp))
continue;
smp_function_call(...);
if (!done)
goto again;
}
So then for each task that has a kretprobe stack, we iterate the stack
and set ri->rp = NULL, remotely when the task isn't running, locally if
the task is running.
I just need to change the semantics of try_invoke_on_locked_down_task()
a bit -- they're a tad weird atm.
> > Anyway,. I think I have a solution, just need to talk to paulmck for a
> > bit.
>
> Ah, you mentioned that the removing the kfree() from the trampline
> handler? I think we can make an rcu callback which will kfree() the
> given instances. (If it works in NMI)
Yes, calling kfree() from the trampoline seems dodgy at best. When
!ri->rp rcu_free() is a good option.
