Re: USB fuzzing with syzbot

[Andrey Konovalov <andreyknvl@google.com>]

On Fri, Apr 19, 2019 at 4:39 PM Alan Stern <stern@rowland.harvard.edu> wrote:
>
> On Fri, 19 Apr 2019, Greg Kroah-Hartman wrote:
>
> > > 3. Regarding that GadgetFS-like interface.
> > >
> > > Initially I was using GadgetFS (together with the Dummy HCD/UDC
> > > module) to perform emulation of USB devices for fuzzing, but later
> > > switched to a custom written interface. This interface is essentially
> > > implemented in the following patch [3]. An example that emulates a USB
> > > keyboard through this interface can be found here [4]. And the
> > > syzkaller parts responsible for USB fuzzing are here [5], [6]. The
> > > incentive to implement a different interface was to provide a somewhat
> > > raw and direct access to the USB Gadget layer for the userspace, where
> > > every USB request is passed to the userspace to get a response.
> > >
> > > The main differences between this interface (referred to as usb-fuzzer
> > > for now) and GadgetFS are:
> > >
> > > 1) GadgetFS does some sanity checks on the provided USB descriptors,
> > > which is something we don't want for fuzzing. We want the descriptors
> > > to be as corrupted as they can.
> > >
> > > 2) GadgetFS handles some of the USB requests internally based on the
> > > provided device descriptor, which is also something we don't want. For
> > > example we may want to be able to provide differently corrupted
> > > responses to the same request.
> > >
> > > 3) usb-fuzzer has ioctl-based interface instead of a filesystem-based
> > > one. I wouldn't say it's that big of a deal, but it makes it somewhat
> > > easier to incorporate into a fuzzer.
> > >
> > > 4) Somewhat related to the previous point: usb-fuzzer uses predictable
> > > endpoint names across different UDCs.
> > >
> > > Right now each UDC driver defines endpoint names via EP_INFO() as it
> > > pleases. And GadgetFS uses those names to create file entries for each
> > > of the endpoints. As a result, endpoint names for different UDCs will
> > > be different and it requires more work to write a portable userspace
> > > gadget. The usb-fuzzer interface auto selects and assigns an endpoint
> > > based on the required features like the transfer type.
> > >
> > > 5) GadgetFS binds to the first available UDC, usb-fuzzer provides a
> > > way to select a UDC to bind to.
> > >
> > > Since the fuzzing happens in multiple processes each of which has its
> > > own Dummy UDC assigned, we want to have control over which UDC we bind
> > > to. This part is a bit confusing, but what I found out is that a UDC
> > > is selected based on two different identifying names. I call the first
> > > one "udc_device_name" and the second one "udc_driver_name".
> > > "udc_device_name" has to be assigned to usb_gadget_driver->udc_name
> > > when usb_gadget_probe_driver() is called, and "udc_driver_name" is
> > > what we have to compare usb_gadget->name with inside of the
> > > usb_gadget_driver->bind() callback. For example, Dummy UDC has
> > > "dummy_udc" as its "udc_driver_name" and "dummy_udc.N" as its
> > > "udc_device_name". At the same time the dwc2 driver that is used on
> > > Raspberry Pi Zero, has "20980000.usb" as both "udc_driver_name" and
> > > "udc_device_name".
> > >
> > > Overall, the usb-fuzzer interface implementation has a similar
> > > structure to that of GadgetFS, but looks way simpler (although that
> > > might be because I've missed to implement some functionality :).
> > >
> > > We'd like to get this upstreamed, but I'm not sure whether this should
> > > be a separate interface (which we can rebrand as a raw usb gadget
> > > interface or something like that) or we should try to make it a
> > > special mode of GadgetFS. I like the former approach more, as GadgetFS
> > > looks quite complicated from my point of view and making fundamental
> > > changes to it doesn't seem like an easy task. This is where we'd like
> > > to get your input.
> >
> > I'll leave the gadgetfs questions to others that know that api better
> > than I do :)
>
> Assuming the usb-fuzzer implementation gets merged, I think it
> definitely should be kept separate from gadgetfs.  The differences in
> functionality are too great; it doesn't make sense to combine them in
> a single driver.
>
> Also, while it often is reasonable for people to build gadgetfs for
> production use, usb-fuzzer is very much more of a developer/debugger
> type of thing.

OK, I'll clean up the patches and send them as an RFC. Thanks!