On Thu, 24 Apr 2008, Wolfram Sang wrote:
> On Mon, Apr 21, 2008 at 10:20:17AM -0700, Trent Piepho wrote:
> > > > Still, if you want to wait at least 25 ms, on a HZ=1000 system you
> > > > might wait only 3 ms.
> > > I'm sorry, I fail to see this. If there are more than three retries,
> > > then there is still the time_before-condition which keeps the loop
> > > running until the timeout is reached, no?
> > Except for the timing problem I pointed out before. The timeout is
> > checked before the write takes place. So if after the 3rd attempt the
> > msleep(), or kernel preemption, etc., delays for 22 ms or more, the
> > next write will never happen.
> Got it now, I misunderstood you before. We do wait 25ms in total, it is
> just that the last write-try happened at 3ms. This is indeed bad. The
> following code should handle it better. (Skipping retries for now)
I guess I should have said "only give the chip 3ms to respond" instead of
"only wait 3ms", since we do wait the 25ms.
> timeout = jiffies + msecs_to_jiffies(write_timeout);
> do {
> keep_trying = time_before(jiffies, timeout);
>
> transfer();
>
> if (success)
> return count;
>
> msleep(1);
> } while (keep_trying);
>
> return -ETIMEDOUT;
That works. Or a very minor optimization not to evaluate the time_before()
unless the write failes:
timeout = jiffies + msecs_to_jiffies(write_timeout);
do {
unsigned long wtime = jiffies;
transfer();
if (success)
return count;
msleep(1);
} while (time_before(wtime, jiffies));
return -ETIMEDOUT;
_______________________________________________
i2c mailing list
[email protected]
http://lists.lm-sensors.org/mailman/listinfo/i2c
