Re: [PATCH v1] usb: dwc3: gadget: fix gadget workqueue use-after-free

[Thinh Nguyen]

On Sat, Feb 01, 2025, Alan Stern wrote:
> On Fri, Jan 31, 2025 at 11:44:17PM +0000, Thinh Nguyen wrote:
> > Cc Alan
> > 
> > On Fri, Jan 31, 2025, Roy Luo wrote:
> > > On Mon, Jan 27, 2025 at 5:44 PM Thinh Nguyen <Thinh.Nguyen@xxxxxxxxxxxx> wrote:
> > > >
> > > > On Wed, Jan 22, 2025, Roy Luo wrote:
> > > > > `dwc3_gadget_soft_disconnect` function, called as part of
> > > >
> > > > The dwc3_gadget_soft_disconnect() isn't directly part of
> > > > device_del(&gadget->dev). It should be part of disconnect.
> > > >
> > > > Can you provide the full sequence of events so I can have more context?
> > > > The handling of the flushing of gadget->work should not be part of the
> > > > dwc3.
> > > 
> > > 
> > > Yes, it's a part of disconnect, and disconnect is a part of gadget unbind.
> > > Let me try my best to explain. Here's the call stack for usb_del_gadget:
> > > -> usb_del_gadget
> > >     -> flush_work(&gadget->work)
> > >     -> device_del
> > >         -> bus_remove_device
> > >         -> device_release_driver
> > >         -> gadget_unbind_driver
> > >         -> usb_gadget_disconnect_locked
> > >         -> dwc3_gadget_pullup
> > >         -> dwc3_gadget_soft_disconnect
> > >         -> usb_gadget_set_state
> > >         -> schedule_work(&gadget->work)
> > > 
> > > Then when usb_put_gadget is called, gadget could be freed before
> > > gadget->work is executed.
> > > -> usb_put_gadget
> > > -> put_device
> > > -> kobject_put
> > > -> device_release
> > > -> dwc_gadget_release
> > > -> kfree(gadget)
> > > 
> > 
> > Thanks for the details!
> > 
> > The UDC core is initiating and handling the gadget->work, so the
> > flushing of the gadget->work should also be handled there.
> > 
> > Since the usb_gadget_disconnect_locked() may trigger a state change work
> > on unbind, the flushing of the gadget->work should to be moved to
> > gadget_unbind_driver() instead:
> > 
> > diff --git a/drivers/usb/gadget/udc/core.c b/drivers/usb/gadget/udc/core.c
> > index f8c1ef465e45..9e4abd6e40f8 100644
> > --- a/drivers/usb/gadget/udc/core.c
> > +++ b/drivers/usb/gadget/udc/core.c
> > @@ -1568,7 +1568,6 @@ void usb_del_gadget(struct usb_gadget *gadget)
> >  
> >         kobject_uevent(&udc->dev.kobj, KOBJ_REMOVE);
> >         sysfs_remove_link(&udc->dev.kobj, "gadget");
> > -       flush_work(&gadget->work);
> >         device_del(&gadget->dev);
> >         ida_free(&gadget_id_numbers, gadget->id_number);
> >         cancel_work_sync(&udc->vbus_work);
> > @@ -1694,6 +1693,8 @@ static void gadget_unbind_driver(struct device *dev)
> >                 synchronize_irq(gadget->irq);
> >         mutex_unlock(&udc->connect_lock);
> >  
> > +       flush_work(&gadget->work);
> > +
> >         udc->driver->unbind(gadget);
> >  
> >         mutex_lock(&udc->connect_lock);
> 
> What about instead moving the flush_work() call down just one line, 
> after the device_del(&gadget->dev) call rather than before it?
> 
> The work queue doesn't need to be flushed every time a driver unbinds 
> from the gadget, only when the gadget is about to be deallocated.
> 

That sounds good to me.

Thanks,
Thinh