Re: [PATCH v4 2/6] can: etas_es58x: add devlink support

[Vincent MAILHOL <mailhol.vincent@wanadoo.fr>]

On Tue. 27 Nov. 2022 at 01:51, Andrew Lunn <andrew@lunn.ch> wrote:
> > @@ -2196,11 +2198,12 @@ static struct es58x_device *es58x_init_es58x_dev(struct usb_interface *intf,
> >               ops = &es581_4_ops;
> >       }
> >
> > -     es58x_dev = devm_kzalloc(dev, es58x_sizeof_es58x_device(param),
> > -                              GFP_KERNEL);
> > -     if (!es58x_dev)
> > +     devlink = devlink_alloc(&es58x_dl_ops, es58x_sizeof_es58x_device(param),
> > +                             dev);
> > +     if (!devlink)
> >               return ERR_PTR(-ENOMEM);
> >
> > +     es58x_dev = devlink_priv(devlink);
>
> That is 'interesting'.

Another interesting thing I found is:
https://elixir.bootlin.com/linux/v6.1-rc6/source/drivers/net/ethernet/intel/ice/ice_devlink.c#L866

Because devlink does not have an equivalent to devm_kzalloc(), that
driver uses devm_add_action_or_reset() instead. But any other drivers
will call devlink_free() in their disconnect function. So here, I just
followed the trend.

> Makes me wonder about lifetimes of different
> objects. Previously your es58x_dev structure would disappear when the
> driver is released, or an explicit call to devm_kfree(). Now it
> disappears when devlink_free() is called.

Even before that, this driver used to release es58x_dev in its
disconnect() function. I changed it to use devm_kzalloc() last year
after discovering its existence.
  https://git.kernel.org/torvalds/linux/c/6bde4c7fd845

>Any danger of use after free here?

devlink_alloc() allocates one continuous block for both the devlink
and the device priv (struct es58x_dev here):
  https://elixir.bootlin.com/linux/v6.1-rc6/source/net/core/devlink.c#L9629

So calling devlink_free() also releases struct es58x_dev.

> USB devices always make me wonder about life times rules since they
> are probably the mode dynamic sort of device the kernel has the
> handle, them just abruptly disappearing.
>
> >       es58x_dev->param = param;
> >       es58x_dev->ops = ops;
> >       es58x_dev->dev = dev;
> > @@ -2247,6 +2250,8 @@ static int es58x_probe(struct usb_interface *intf,
> >       if (ret)
> >               return ret;
> >
> > +     devlink_register(priv_to_devlink(es58x_dev));
> > +
> >       for (ch_idx = 0; ch_idx < es58x_dev->num_can_ch; ch_idx++) {
> >               ret = es58x_init_netdev(es58x_dev, ch_idx);
> >               if (ret) {
> > @@ -2272,8 +2277,10 @@ static void es58x_disconnect(struct usb_interface *intf)
> >       dev_info(&intf->dev, "Disconnecting %s %s\n",
> >                es58x_dev->udev->manufacturer, es58x_dev->udev->product);
> >
> > +     devlink_unregister(priv_to_devlink(es58x_dev));
> >       es58x_free_netdevs(es58x_dev);
> >       es58x_free_urbs(es58x_dev);
> > +     devlink_free(priv_to_devlink(es58x_dev));
> >       usb_set_intfdata(intf, NULL);
>
> Should devlink_free() be after usb_set_inftdata()?

A look at
  $ git grep -W "usb_set_intfdata(.*NULL)"

shows that the two patterns (freeing before or after
usb_set_intfdata()) coexist.

You are raising an important question here. usb_set_intfdata() does
not have documentation that freeing before it is risky. And the
documentation of usb_driver::disconnect says that:
  "@disconnect: Called when the interface is no longer accessible,
   usually because its device has been (or is being) disconnected
   or the driver module is being unloaded."
  Ref: https://elixir.bootlin.com/linux/v6.1-rc6/source/include/linux/usb.h#L1130

So the interface no longer being accessible makes me assume that the
order does not matter. If it indeed matters, then this is a foot gun
and there is some clean-up work waiting for us on many drivers.

@Greg, any thoughts on whether or not the order of usb_set_intfdata()
and resource freeing matters or not?


Yours sincerely,
Vincent Mailhol