out of bounds access on array error_text[] because of -ETIMEDOUT return from __send_command()

[Markus Mayer]

On Thu, 20 Aug 2020 at 10:23, Markus Mayer <mmayer at broadcom.com> wrote:
>
> On Wed, 19 Aug 2020 at 11:34, Florian Fainelli <f.fainelli at gmail.com> wrote:
> >
> > On 8/18/20 5:21 AM, Colin Ian King wrote:
> > > Hi,
> > >
> > > static analysis with coverity has found a buffer overflow issue with the
> > > brcmstb driver, I believe it may have been introduced with the following
> > > commit:
> > >
> > > commit a7c25759d8d84b64c437a78f05df7314b02934e5
> > > Author: Markus Mayer <mmayer at broadcom.com>
> > > Date:   Tue Apr 2 16:01:00 2019 -0700
> > >
> > >     memory: brcmstb: dpfe: wait for DCPU to be ready
> > >
> > > The static analysis is as follows for the source file
> > > /drivers/memory/brcmstb_dpfe.c :
> > >
> > > 684 static ssize_t generic_show(unsigned int command, u32 response[],
> > > 685                            struct brcmstb_dpfe_priv *priv, char *buf)
> > > 686 {
> > > 687        int ret;
> > > 688
> > >    1. Condition !priv, taking false branch.
> > >
> > > 689        if (!priv)
> > > 690                return sprintf(buf, "ERROR: driver private data not
> > > set\n");
> > > 691
> > >    2. return_constant: Function call __send_command(priv, command,
> > > response) may return -110.
> > >    3. assignment: Assigning: ret = __send_command(priv, command,
> > > response). The value of ret is now -110.
> > >
> > > 692        ret = __send_command(priv, command, response);
> > >    4. Condition ret < 0, taking true branch.
> > >
> > > 693        if (ret < 0)
> > >
> > > Out-of-bounds read (OVERRUN)
> > >    5. overrun-local: Overrunning array error_text of 6 8-byte elements
> > > at element index 110 (byte offset 887) using index -ret (which evaluates
> > > to 110).
> > > 694                return sprintf(buf, "ERROR: %s\n", error_text[-ret]);
> > > 695
> > > 696        return 0;
> > > 697 }
> > >
> > >
> > > Function __send_command() can return -ETIMEDOUT and this causes an
> > > out-of-bounds access on error_text[].
> >
> > Markus, what do you think of this:
> >
> > diff --git a/drivers/memory/brcmstb_dpfe.c b/drivers/memory/brcmstb_dpfe.c
> > index 60e8633b1175..b41c6251ddc3 100644
> > --- a/drivers/memory/brcmstb_dpfe.c
> > +++ b/drivers/memory/brcmstb_dpfe.c
> > @@ -445,7 +445,7 @@ static int __send_command(struct brcmstb_dpfe_priv
> > *priv, unsigned int cmd,
> >         }
> >         if (resp != 0) {
> >                 mutex_unlock(&priv->lock);
> > -               return -ETIMEDOUT;
> > +               return -ffs(DCPU_RET_ERR_TIMEDOUT);
> >         }
> >
> >         /* Compute checksum over the message */
> >
> > That way we only return DCPU-style error code from __send_command and we
> > de-reference error_text accordingly? Or we could just introduce a proper
> > lookup with a function instead of a direct array de-reference.
>
> Let me do some experiments. What you are proposing should work and is
> in line with the current code. A lookup function might be cleaner,
> though.

I submitted a patch for review. See https://lkml.org/lkml/2020/8/20/2291.

Regards,
-Markus