Re: KMSAN: uninit-value in cdc_ncm_set_dgram_size

[Alexander Potapenko <glider@google.com>]

+ Greg K-H
On Tue, Nov 5, 2019 at 1:25 PM Bjørn Mork <bjorn@mork.no> wrote:
>
> Oliver Neukum <oneukum@suse.com> writes:
> > Am Montag, den 04.11.2019, 22:22 +0100 schrieb Bjørn Mork:
> >> This looks like a false positive to me. max_datagram_size is two bytes
> >> declared as
> >>
> >>         __le16 max_datagram_size;
> >>
> >> and the code leading up to the access on drivers/net/usb/cdc_ncm.c:587
> >> is:
> >>
> >>         /* read current mtu value from device */
> >>         err = usbnet_read_cmd(dev, USB_CDC_GET_MAX_DATAGRAM_SIZE,
> >>                               USB_TYPE_CLASS | USB_DIR_IN | USB_RECIP_INTERFACE,
> >>                               0, iface_no, &max_datagram_size, 2);
> >
> > At this point err can be 1.
> >
> >>         if (err < 0) {
> >>                 dev_dbg(&dev->intf->dev, "GET_MAX_DATAGRAM_SIZE failed\n");
> >>                 goto out;
> >>         }
> >>
> >>         if (le16_to_cpu(max_datagram_size) == ctx->max_datagram_size)
> >>
> >>
> >>
> >> AFAICS, there is no way max_datagram_size can be uninitialized here.
> >> usbnet_read_cmd() either read 2 bytes into it or returned an error,
> >
> > No. usbnet_read_cmd() will return the number of bytes transfered up
> > to the number requested or an error.
>
> Ah, OK. So that could be fixed with e.g.
>
>   if (err < 2)
>        goto out;
It'd better be (err < sizeof(max_datagram_size)), and probably in the
call to usbnet_read_cmd() as well.
>
> Or would it be better to add a strict length checking variant of this
> API?  There are probably lots of similar cases where we expect a
> multibyte value and a short read is (or should be) considered an error.
> I can't imagine any situation where we want a 2, 4, 6 or 8 byte value
> and expect a flexible length returned.
This is really a widespread problem on syzbot: a lot of USB devices
use similar code calling usb_control_msg() to read from the device and
not checking that the buffer is fully initialized.

Greg, do you know how often usb_control_msg() is expected to read less
than |size| bytes? Is it viable to make it return an error if this
happens?
Almost nobody is using this function correctly (i.e. checking that it
has read the whole buffer before accessing it).

> >> causing the access to be skipped.  Or am I missing something?
> >
> > Yes. You can get half the MTU. We have a similar class of bugs
> > with MAC addresses.
>
> Right.  And probably all 16 or 32 bit integer reads...
>
> Looking at the NCM spec, I see that the wording is annoyingly flexible
> wrt length - both ways.  E.g for GetNetAddress:
>
>   To get the entire network address, the host should set wLength to at
>   least 6. The function shall never return more than 6 bytes in response
>   to this command.
>
> Maybe the correct fix is simply to let usbnet_read_cmd() initialize the
> full buffer regardless of what the device returns?  I.e.
>
> diff --git a/drivers/net/usb/usbnet.c b/drivers/net/usb/usbnet.c
> index dde05e2fdc3e..df3efafca450 100644
> --- a/drivers/net/usb/usbnet.c
> +++ b/drivers/net/usb/usbnet.c
> @@ -1982,7 +1982,7 @@ static int __usbnet_read_cmd(struct usbnet *dev, u8 cmd, u8 reqtype,
>                    cmd, reqtype, value, index, size);
>
>         if (size) {
> -               buf = kmalloc(size, GFP_KERNEL);
> +               buf = kzalloc(size, GFP_KERNEL);
>                 if (!buf)
>                         goto out;
>         }
> @@ -1992,7 +1992,7 @@ static int __usbnet_read_cmd(struct usbnet *dev, u8 cmd, u8 reqtype,
>                               USB_CTRL_GET_TIMEOUT);
>         if (err > 0 && err <= size) {
>          if (data)
> -            memcpy(data, buf, err);
> +            memcpy(data, buf, size);
>          else
>              netdev_dbg(dev->net,
>                  "Huh? Data requested but thrown away.\n");
>
>
>
>
> What do you think?
>
> Personally, I don't think it makes sense for a device to return a 1-byte
> mtu or 3-byte mac address. But the spec allows it and this would at
> least make it safe.
>
> We have a couple of similar bugs elsewhere in the same driver, BTW..
>
>
> Bjørn



-- 
Alexander Potapenko
Software Engineer

Google Germany GmbH
Erika-Mann-Straße, 33
80636 München

Geschäftsführer: Paul Manicle, Halimah DeLaine Prado
Registergericht und -nummer: Hamburg, HRB 86891
Sitz der Gesellschaft: Hamburg