Re: My vision of usbmon

[Pete Zaitcev <zaitcev@redhat.com>]

On Mon, 20 Dec 2004 15:25:52 +0100, Oliver Neukum <oliver@neukum.org> wrote:

> Am Montag, 20. Dezember 2004 08:04 schrieb Pete Zaitcev:
> > +               memcpy(&mbus->shim_ops, ubus->op, sizeof(struct usb_operations));
> > +               mbus->shim_ops.submit_urb = mon_submit;
> > +               mbus->saved_op = ubus->op;
> > +               ubus->op = &mbus->shim_ops;
> > +               ubus->monitored = 1;
> 
> I think you need smp_wmb() here to make sure that an irq taken
> on another CPU sees the manipulations in the correct order.

Hmm, it seems you are right. I forgot about reordering issues. I relied on
op being atomic, but if it points at an uninitialized shim, this will end
badly. How about this?

                memcpy(&mbus->shim_ops, ubus->op, sizeof(struct usb_operations));
                mbus->shim_ops.submit_urb = mon_submit;
                mbus->saved_op = ubus->op;
                smp_mb();       /* ubus->op is not protected by spinlocks */
                ubus->op = &mbus->shim_ops;
                ubus->monitored = 1;

Generally, the type of coding which requires a use of memory barriers in drivers
is a bug or a latent bug, so I am sorry for the above. It was a sacrifice to
make usbmon invisible if it's not actively monitoring. Sorry about that.

-- Pete