Re: commit 5fe212364 causes division by zero with large bauds

[Felipe Balbi]

On Wed, Sep 11, 2013 at 09:48:13PM +0300, Alexey Pelykh wrote:
> On Wed, Sep 11, 2013 at 9:38 PM, Felipe Balbi <balbi@xxxxxx> wrote:
> > Hi,
> >
> > On Wed, Sep 11, 2013 at 09:22:26AM +0300, Alexey Pelykh wrote:
> >> Hi Felipe,
> >>
> >> Thanks for finding this issue. Indeed, there is a bug on 3M+ baud
> >> rates. First patch is close to a complete fix, but still contains
> >> div-by-zero issue. Here is my version:
> >>
> >> diff --git a/drivers/tty/serial/omap-serial.c b/drivers/tty/serial/omap-serial.c
> >> index 816d1a2..808a880 100644
> >> --- a/drivers/tty/serial/omap-serial.c
> >> +++ b/drivers/tty/serial/omap-serial.c
> >> @@ -240,8 +240,8 @@ serial_omap_baud_is_mode16(struct uart_port *port,
> >> unsigned int baud)
> >>  {
> >>         unsigned int n13 = port->uartclk / (13 * baud);
> >>         unsigned int n16 = port->uartclk / (16 * baud);
> >> -       int baudAbsDiff13 = baud - (port->uartclk / (13 * n13));
> >> -       int baudAbsDiff16 = baud - (port->uartclk / (16 * n16));
> >> +       int baudAbsDiff13 = n13 ? (baud - (port->uartclk / (13 * n13))) : INT_MAX;
> >> +       int baudAbsDiff16 = n16 ? (baud - (port->uartclk / (16 * n16))) : INT_MAX;
> >
> > IOW:
> >
> > int baudAbsDiff13 = 0;
> >
> > if (n13)
> >         baudAbsDiff13 = (baud - (port->uartclk / (13 * n13)));
> 
> Not quite same code, INT_MAX instead of 0. With 0 a div-by-zero
> exception will still occur on 3686400.

why, there's no division after that point, right ? Besides,
serial_omap_baud_is_mode16() is supposed to return a boolean value.

Setting baudAbsDiff1[36] to 0 will cause no problems, you're only using
that value with a boolean expression, no divisions whatsoever. Division
is done after using serial_omap_baud_is_mode16() to initialize divisor
to 13 or 16 (which is a misnamer, since that's the oversampling
parameter).

> > which is exactly what my patch did. I fail to see where division by zero
> > would be coming from.
> >
> >>         if(baudAbsDiff13 < 0)
> >>                 baudAbsDiff13 = -baudAbsDiff13;
> >>         if(baudAbsDiff16 < 0)
> >>
> >>
> >> With 48MHz UART clock, it will give
> >> 300: divisor = 12307 (13), real rate 300 (0.000000%)
> >> 600: divisor = 6153 (13), real rate 600 (0.000000%)
> >> 1200: divisor = 3076 (13), real rate 1200 (0.000000%)
> >> 2400: divisor = 1538 (13), real rate 2400 (0.000000%)
> >
> > TRM has these all set with oversampling of 16. In fact only 460800,
> > 921600, 1843200 and 3686400 should be using oversampling of 13.
> >
> 
> That's true, but TRM anyways does not contain all possible baud rates
> (1M e.g.). IMO, as long as error rate is the same as in TRM,
> it makes no difference what combination of (mode, divisor) to use.
> 
> > --
> > balbi
> 
> A complex solution may be implemented: use LUT for baud rates that TRM
> defines explicitly, and use calculation if lookup failed.

why would you try calculating anything if there is nothing in the table
which can support it ? Whatever is in the lookup table, are the only
baud rates the SoC supports, right ?

cheers

-- 
balbi

Attachment: signature.asc
Description: Digital signature