Re: [PATCH] hw/usb/hcd-dwc2: Enforce epnum to 0 for the control endpoint to avoid the assertion failure in usb_ep_get()

[Qiang Liu]

Hi Paul,

On Mon, Jul 5, 2021 at 6:27 AM Paul Zimmerman <pauldzim@gmail.com> wrote:
>
> On Sat, Jun 26, 2021 at 10:21 PM Qiang Liu <cyruscyliu@gmail.com> wrote:
> >
> > Hi folks,
> >
> > I found this bug by my dwc2 fuzzer.
> > It seems that
> > * https://bugs.launchpad.net/qemu/+bug/1907042
> > * https://bugs.launchpad.net/qemu/+bug/1525123
> > or
> > * https://gitlab.com/qemu-project/qemu/-/issues/119
> > * https://gitlab.com/qemu-project/qemu/-/issues/303
> > have reported similar issues.
> >
> > Would it be better to consider and fix them together?
> >
> > Best,
> > Qiang
> >
> > On Sun, Jun 27, 2021 at 11:28 AM Qiang Liu <cyruscyliu@gmail.com> wrote:
> > >
> > > When eptype is USB_ENDPOINT_XFER_CONTROL and pid is
> > > TSIZ_SC_MC_PID_SETUP, usb_ep_get() should return the control endpoint.
> > > In hw/usb/core.c, the assumed epnum of the control endpoint is 0. When
> > > epnum is not 0, usb_ep_get() will crash due to the check assert(pid ==
> > > USB_TOKEN_IN || pid == USB_TOKEN_OUT).
> > >
> > > The description
> > > http://www.capital-micro.com/PDF/CME-M7_Family_User_Guide_EN.pdf
> > > (18.5.3.4 (14), 18.5.3.4 (10)) a) mentions that the pid is maintained by
> > > the host, b) but doesn't mention that whether the epnum should be 0 for
> > > the control endpoint. However, usb_ep_get() assumes it is 0. To avoid
> > > potential assertion failure in usb_ep_get(), we could enforce epnum to 0
> > > and warn users.
> > >
> > > Fixes: 153ef1662c3 ("dwc-hsotg (dwc2) USB host controller emulation")
> > > Signed-off-by: Qiang Liu <cyruscyliu@gmail.com>
> > > ---
> > >  hw/usb/hcd-dwc2.c | 5 +++++
> > >  1 file changed, 5 insertions(+)
> > >
> > > diff --git a/hw/usb/hcd-dwc2.c b/hw/usb/hcd-dwc2.c
> > > index e1d96ac..65d9d46 100644
> > > --- a/hw/usb/hcd-dwc2.c
> > > +++ b/hw/usb/hcd-dwc2.c
> > > @@ -636,6 +636,11 @@ static void dwc2_enable_chan(DWC2State *s,  uint32_t 
> > > index)
> > >      }
> > >
> > >      if (eptype == USB_ENDPOINT_XFER_CONTROL && pid == 
> > > TSIZ_SC_MC_PID_SETUP) {
> > > +        if (epnum != 0) {
> > > +            qemu_log_mask(LOG_GUEST_ERROR,
> > > +                          "epnum should be 0 for the control 
> > > endpoint\n");
> > > +            epnum = 0;
> > > +        }
> > >          pid = USB_TOKEN_SETUP;
> > >      } else {
> > >          pid = epdir ? USB_TOKEN_IN : USB_TOKEN_OUT;
> > > --
> > > 2.7.4
> > >
>
> Hi Qiang,
>
> Sorry for the late reply, I've had a busy week.
> Yes, I think it would be best to fix this in the core since it affects more
> than one host.
My current fix is in the following.

diff --git a/hw/usb/core.c b/hw/usb/core.c
index 975f762..a29b378 100644
--- a/hw/usb/core.c
+++ b/hw/usb/core.c
@@ -735,10 +735,11 @@ void usb_ep_dump(USBDevice *dev)

 struct USBEndpoint *usb_ep_get(USBDevice *dev, int pid, int ep)
 {
     struct USBEndpoint *eps;

     assert(dev != NULL);
-    if (ep == 0) {
+    if (ep == 0 || pid == USB_TOKEN_SETUP) {
         return &dev->ep_ctl;
     }
     assert(pid == USB_TOKEN_IN || pid == USB_TOKEN_OUT);

>I'm not sure that forcing the Control endpoint to 0 is
> the best solution though, perhaps it would be better to print an error
> message and fail the operation? AFAIK there are no real-world devices
> that have Control endpoints other than 0, although I believe it is allowed
> by the USB spec.
I think, as a malicious guest, if the spec allows, it can set the
control endpoint to non-zero.

> Let's wait and see what Gerd thinks.
Yes!

Best,
Qiang