[PATCH] watchdog: aspeed: fix integer overflow in set_timeout handler

Guenter Roeck linux at roeck-us.net
Fri Apr 16 11:47:50 AEST 2021


On Thu, Apr 15, 2021 at 06:04:45PM -0700, Tao Ren wrote:
> On Thu, Apr 15, 2021 at 05:50:32PM -0700, Guenter Roeck wrote:
> > On 4/15/21 5:12 PM, rentao.bupt at gmail.com wrote:
> > > From: Tao Ren <rentao.bupt at gmail.com>
> > > 
> > > Fix the time comparison (timeout vs. max_hw_heartbeat_ms) in set_timeout
> > > handler to avoid potential integer overflow when the supplied timeout is
> > > greater than aspeed's maximum allowed timeout (4294 seconds).
> > > 
> > > Fixes: efa859f7d786 ("watchdog: Add Aspeed watchdog driver")
> > > Reported-by: Amithash Prasad <amithash at fb.com>
> > > Signed-off-by: Tao Ren <rentao.bupt at gmail.com>
> > > ---
> > >  drivers/watchdog/aspeed_wdt.c | 5 ++---
> > >  1 file changed, 2 insertions(+), 3 deletions(-)
> > > 
> > > diff --git a/drivers/watchdog/aspeed_wdt.c b/drivers/watchdog/aspeed_wdt.c
> > > index 7e00960651fa..9f77272dc906 100644
> > > --- a/drivers/watchdog/aspeed_wdt.c
> > > +++ b/drivers/watchdog/aspeed_wdt.c
> > > @@ -145,9 +145,8 @@ static int aspeed_wdt_set_timeout(struct watchdog_device *wdd,
> > >  	struct aspeed_wdt *wdt = to_aspeed_wdt(wdd);
> > >  	u32 actual;
> > >  
> > > -	wdd->timeout = timeout;
> > > -
> > > -	actual = min(timeout, wdd->max_hw_heartbeat_ms * 1000);
> > > +	actual = min(timeout, wdd->max_hw_heartbeat_ms / 1000);
> > > +	wdd->timeout = actual;
> > >  
> > >  	writel(actual * WDT_RATE_1MHZ, wdt->base + WDT_RELOAD_VALUE);
> > >  	writel(WDT_RESTART_MAGIC, wdt->base + WDT_RESTART);
> > > 
> > 
> > If the provided timeout is larger than the supported hardware timeout,
> > the watchdog core will ping the hardware on behalf of userspace.
> > The above code would defeat that mechanism for no good reason.
> > 
> > NACK.
> > 
> > Guenter
> 
> Thanks Guenter for Joel for the quick review!
> 
> The integer overflow happens at (actual * WDT_RATE_1MHZ). For example,
> if a user tries to set timeout to 4295 seconds, then the hardware would
> be programmed to timeout after about 32 milliseconds. I would say this
> behavior is not expected?

The fix would be

-	actual = min(timeout, wdd->max_hw_heartbeat_ms * 1000);
+	actual = min(timeout, wdd->max_hw_heartbeat_ms / 1000);

without modifying
	wdd->timeout = timeout;

The real problem here is that "wdd->max_hw_heartbeat_ms * 1000"
is simply wrong. The overflow is a side effect of the wrong
calculation, and concentrating on it disguises the real problem.

Guenter


More information about the openbmc mailing list