On Fri, 12 Sep 2014, Petr Mladek wrote:
> > I don't think you can eliminate the lock quite so easily. This patch
> > introduces some nasty races.
> >
> > > @@ -577,18 +571,20 @@ static int hub_port_status(struct usb_hub *hub, int
> > > port1,
> > >
> > > static void kick_khubd(struct usb_hub *hub)
> > > {
> > > - unsigned long flags;
> > > -
> > > - spin_lock_irqsave(&hub_event_lock, flags);
> > > - if (!hub->disconnected && list_empty(&hub->event_list)) {
> > > - list_add_tail(&hub->event_list, &hub_event_list);
> > > -
> > > - /* Suppress autosuspend until khubd runs */
> > > + if (!hub->disconnected && !work_pending(&hub->events)) {
> >
> > Here you test hub->disconnected, with no lock for protection.
>
> This should not be that big problem. It will schedule hub_event() but
> it will do basically nothing. This is why I thought that the lock was
> not needed.
What do you mean "basically nothing"? hub_event will be scheduled, via
a work_struct that is embedded in the usb_hub structure. But that
structure will be deallocated by hub_disconnect, so you will create a
"use after free" bug.
> > (Also, note that work_pending is not synchronized with anything. What
> > happens if two threads call this routine at the same time?)
>
> You are right! This is a real problem because it might call
> usb_autopm_put_interface_no_suspend() twice but it might schedule
> hub_event() and call usb_autopm_put_interface() only once.
>
> Well, it might be possible to check the return value of
> queue_work and do something like:
>
> if (!hub->disconnected && !work_pending(&hub->events)) {
> usb_autopm_get_interface_no_resume(
> to_usb_interface(hub->intfdev));
> if (!queue_work(hub_wq, &hub->events))
> usb_autopm_put_interface_no_suspend(intf);
> }
>
> But there is still problem that we need to call
> "INIT_WORK(&hub->events, hub_event)" somewhere and do it only once
> before calling kick_hub_wq(). I wonder if it might be safe to do
> so in hub_activate().
If I thought this was the right way to go, I would suggest initializing
hub->events in hub_probe, where the structure is created.
> Hmm, I am not longer that optimistic about it. After all, it might
> be better to put the lock back. Would you prefer it, please?
Here's what I think. If you want to make khubd into a work queue
thread, you can. But it should be invoked only once, and the routine
it runs should be hub_thread, not hub_events. Overall I don't see any
advantage in making this change.
> > And here you set hub->disconnected with no lock for protection. So
> > what happens if one thread calls kick_khubd at the same time as another
> > thread calls hub_disconnect?
>
> This should not be that big problem as explained above. Note that
> hub->disconnected was tested in hub_events() without the lock
> even before this patch. Hence I thought that the new code was as racy
> as before.
But you ignored what the comment says about "don't let it be added
again".
Alan Stern
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
