qla3xxx: Odd likely incorrect use of test_bit in qla3xxx.c

qla3xxx: Odd likely incorrect use of test_bit in qla3xxx.c

[Joe Perches]

In ql_reset_work (line 3620)

The ql_reset_work function uses an "or" of 2 enum values.

enum { QL_RESET_DONE = 1,	/* Reset finished. */
	QL_RESET_ACTIVE = 2,	/* Waiting for reset to finish. */
	QL_RESET_START = 3,	/* Please reset the chip. */
	QL_RESET_PER_SCSI = 4,	/* SCSI driver requests reset. */
	QL_TX_TIMEOUT = 5,	/* Timeout in progress. */
	QL_LINK_MASTER = 6,	/* This driver controls the link. */
	QL_ADAPTER_UP = 7,	/* Adapter has been brought up. */

The values (3 and 4) or'd (added) together are
QL_ADAPTER_UP

If that's really what's desired, it'd be better to use
QL_ADAPTER_UP directly, though I'd expect this should
test both bits independently instead.

---

static void ql_reset_work(struct work_struct *work)
{
	struct ql3_adapter *qdev =
		container_of(work, struct ql3_adapter, reset_work.work);
	struct net_device *ndev = qdev->ndev;
	u32 value;
	struct ql_tx_buf_cb *tx_cb;
	int max_wait_time, i;
	struct ql3xxx_port_registers __iomem *port_regs =
		qdev->mem_map_registers;
	unsigned long hw_flags;

	if (test_bit((QL_RESET_PER_SCSI | QL_RESET_START), &qdev->flags)) {
		clear_bit(QL_LINK_MASTER, &qdev->flags);

Re: qla3xxx: Odd likely incorrect use of test_bit in qla3xxx.c

[Dan Carpenter]

Good eye, Joe.

I wrote a Smatch check to find similar bugs.  They weren't any other
places which tried to do bitwise OR.  The bug that happens occasionally
is:

#define MY_FLAG BIT(1)

	if (test_bit(MY_FLAG, &map)) {
		...

It's not normally harmful if it's used consistently, but ath9k had
memory corruption because they do:

	set_bit(BIT(6), &some_unsigned_long);

Anyway, I'll send patches for the bugs I found and push the Smatch
check. Thanks!

regards,
dan carpenter

Re: qla3xxx: Odd likely incorrect use of test_bit in qla3xxx.c

[Joe Perches]

On Wed, 2015-05-13 at 16:19 +0300, Dan Carpenter wrote:
> I wrote a Smatch check to find similar bugs.

Nice.

It didn't look like something coccinelle could do and
I didn't know smatch syntax well enough (OK, I don't
know smatch syntax at all), to write something to find
these defects but I thought, if anyone, you could.

Thanks Dan for doing it.

> ath9k had memory corruption because they do:
> 	set_bit(BIT(6), &some_unsigned_long);
> Anyway, I'll send patches for the bugs I found and push the Smatch
> check.

Re: qla3xxx: Odd likely incorrect use of test_bit in qla3xxx.c

[Joe Perches]

On Wed, 2015-05-13 at 16:19 +0300, Dan Carpenter wrote:
> Good eye, Joe.
> 
> I wrote a Smatch check to find similar bugs.  Te rhey weren't any other
> places which tried to do bitwise OR.  The bug that happens occasionally
> is:
> 
> #define MY_FLAG BIT(1)
> 
> 	if (test_bit(MY_FLAG, &map)) {
> 		...
> 
> It's not normally harmful if it's used consistently, but ath9k had
> memory corruption because they do:
> 
> 	set_bit(BIT(6), &some_unsigned_long);
> 
> Anyway, I'll send patches for the bugs I found and push the Smatch
> check. Thanks!

In case the patches you found weren't of this type,
there are other possibly suspicious uses with & like:

drivers/usb/host/ehci-tegra.c:		set_bit((wIndex & 0xff) - 1, &ehci->suspended_ports);

Re: qla3xxx: Odd likely incorrect use of test_bit in qla3xxx.c

[Dan Carpenter]

On Wed, May 13, 2015 at 10:34:51AM -0700, Joe Perches wrote:
> On Wed, 2015-05-13 at 16:19 +0300, Dan Carpenter wrote:
> > Good eye, Joe.
> > 
> > I wrote a Smatch check to find similar bugs.  Te rhey weren't any other
> > places which tried to do bitwise OR.  The bug that happens occasionally
> > is:
> > 
> > #define MY_FLAG BIT(1)
> > 
> > 	if (test_bit(MY_FLAG, &map)) {
> > 		...
> > 
> > It's not normally harmful if it's used consistently, but ath9k had
> > memory corruption because they do:
> > 
> > 	set_bit(BIT(6), &some_unsigned_long);
> > 
> > Anyway, I'll send patches for the bugs I found and push the Smatch
> > check. Thanks!
> 
> In case the patches you found weren't of this type,
> there are other possibly suspicious uses with & like:
> 
> drivers/usb/host/ehci-tegra.c:		set_bit((wIndex & 0xff) - 1, &ehci->suspended_ports);
>

That's puzzling, yes, but my instinct without looking at the context is
that it's probably valid.  I didn't look at that that one, but I looked
at a bunch of similar cases.

My test only looked at test_bit().  I'll update it to look at set_bit()
and clear_bit() as well.  It only complains about "test_bit(x | y, " and
"test_bit(x < y, ".

Btw, Smatch already has a check for when people do:

	set_bit(FOO_BIT, &flags);

	...

	if (flags & FOO_BIT) {

It looks uses the macro names, so if you open code it then it won't
catch the inconsistency, but mostly it works.

regards,
dan carpenter

Re: qla3xxx: Odd likely incorrect use of test_bit in qla3xxx.c

[Joe Perches]

On Thu, 2015-05-14 at 11:01 +0300, Dan Carpenter wrote:
> On Wed, May 13, 2015 at 10:34:51AM -0700, Joe Perches wrote:
> > In case the patches you found weren't of this type,
> > there are other possibly suspicious uses with & like:
> > drivers/usb/host/ehci-tegra.c:		set_bit((wIndex & 0xff) - 1, &ehci->suspended_ports);
> That's puzzling, yes, but my instinct without looking at the context is
> that it's probably valid.

It may be valid, but at a minimum it's bad code.

Assuming wIndex is guaranteed non-zero, it likely
should be something like:

(wIndex % BITS_PER_LONG) - 1

or suspended_ports should be more than a single ulong

Anyway, thanks.