check_signed.c false negative

check_signed.c false negative

[John Levon]

Given:

enum ib_qp_state {
        IB_QPS_RESET,
        IB_QPS_INIT,
        IB_QPS_RTR,
        IB_QPS_RTS,
        IB_QPS_SQD,
        IB_QPS_SQE,
        IB_QPS_ERR
};

int p(enum ib_qp_state cur_state, enum ib_qp_state next_state)
{
#if 0
       if (cur_state  < 0 || cur_state > IB_QPS_ERR ||
         next_state < 0 || next_state > IB_QPS_ERR) {
                return (5);
        }
#else
       if (next_state < 0 || next_state > IB_QPS_ERR) {
                return (5);
        }
#endif

        return (0);
}

current smatch is happy. But "#if 1" above instead, and:

jlevon@sent:~/src/smatch$ ./smatch a.c
./smatch: a.c:15 p() warn: unsigned 'cur_state' is never less than zero.

Why is "next_state" not reported equally?

thanks
john

Re: check_signed.c false negative

[Dan Carpenter]

On Mon, Nov 11, 2019 at 11:21:15PM +0000, John Levon wrote:
> 
> Given:
> 
> enum ib_qp_state {
>         IB_QPS_RESET,
>         IB_QPS_INIT,
>         IB_QPS_RTR,
>         IB_QPS_RTS,
>         IB_QPS_SQD,
>         IB_QPS_SQE,
>         IB_QPS_ERR
> };
> 
> int p(enum ib_qp_state cur_state, enum ib_qp_state next_state)
> {
> #if 0
>        if (cur_state  < 0 || cur_state > IB_QPS_ERR ||
>          next_state < 0 || next_state > IB_QPS_ERR) {
>                 return (5);
>         }
> #else
>        if (next_state < 0 || next_state > IB_QPS_ERR) {
>                 return (5);
>         }
> #endif
> 
>         return (0);
> }
> 
> current smatch is happy. But "#if 1" above instead, and:
> 
> jlevon@sent:~/src/smatch$ ./smatch a.c
> ./smatch: a.c:15 p() warn: unsigned 'cur_state' is never less than zero.
> 
> Why is "next_state" not reported equally?

My plan was that neither one would print a warning.  When people do

	if (foo < 0 || foo > max)

then normally it's not a real life bug...  Linus said at kernel summit
that he doesn't like when people send patches for those.

The problem is that when I wrote the code to silence the warning then I
hadn't created all the expr_get_parent_expr() function so I couldn't
silence it properly.

regards,
dan carpenter

Re: check_signed.c false negative

[John Levon]

On Wed, Nov 13, 2019 at 08:39:13PM +0300, Dan Carpenter wrote:

> On Mon, Nov 11, 2019 at 11:21:15PM +0000, John Levon wrote:
> > 
> > Given:
> > 
> > enum ib_qp_state {
> >         IB_QPS_RESET,
> >         IB_QPS_INIT,
> >         IB_QPS_RTR,
> >         IB_QPS_RTS,
> >         IB_QPS_SQD,
> >         IB_QPS_SQE,
> >         IB_QPS_ERR
> > };
> > 
> > int p(enum ib_qp_state cur_state, enum ib_qp_state next_state)
> > {
> > #if 0
> >        if (cur_state  < 0 || cur_state > IB_QPS_ERR ||
> >          next_state < 0 || next_state > IB_QPS_ERR) {
> >                 return (5);
> >         }
> > #else
> >        if (next_state < 0 || next_state > IB_QPS_ERR) {
> >                 return (5);
> >         }
> > #endif
> > 
> >         return (0);
> > }
> > 
> > current smatch is happy. But "#if 1" above instead, and:
> > 
> > jlevon@sent:~/src/smatch$ ./smatch a.c
> > ./smatch: a.c:15 p() warn: unsigned 'cur_state' is never less than zero.
> > 
> > Why is "next_state" not reported equally?
> 
> My plan was that neither one would print a warning.  When people do
> 
> 	if (foo < 0 || foo > max)
> 
> then normally it's not a real life bug...  Linus said at kernel summit
> that he doesn't like when people send patches for those.

Right, in fact the code this comes from is similar: it's a
belt-and-braces check that we didn't somehow get a bad state value.

There's definitely other cases where this might be useful, though I
wouldn't object to disabling it by default altogether.

thanks
john