Oliver Endriss wrote:
> Johannes Stezenbach wrote:
> > I just fixed a bug in timeout handling:
> >
> > start = jiffies;
> > while (rdebi(av7110, DEBINOSWAP, COMMAND, 0, 2 )) {
> > - msleep(1);
> > if (time_after(jiffies, start + ARM_WAIT_FREE)) {
> > printk(KERN_ERR "dvb-ttpci: %s(): timeout waiting for COMMAND idle\n", __FUNCTION__);
> > return -ETIMEDOUT;
> > }
> > + msleep(1);
> > }
> >
> > (multiple occurences)
> >
> > When the machine is busy, msleep(1) can sleep unexpectedly long.
> > Additionally, it doesn't make sense to sleep and then exit with
> > a timeout without retesting the condition we're waiting for.
>
> The code above is better but does not solve the problem completely.
> I still had problems under high CPU load with a preemptible kernel.
>
> Imho it should be done this way:
>
> while (1) {
> err = time_after(jiffies, start + ARM_WAIT_FREE);
> if (rdebi(av7110, DEBINOSWAP, COMMAND, 0, 2) == 0)
> break;
> if (err) {
> printk(KERN_ERR "dvb-ttpci: %s(): timeout waiting for COMMAND idle\n", __FUNCTION__);
> return -ETIMEDOUT;
> }
> msleep(1);
> }
>
> This way rdebi can always be executed at the end of the timeout period.
> Now the scheduler may suspend the task everywhere in the loop for an
> extended period of time.
>
> If no one objects I will apply the patch to av7110_hw.c and
> saa7146_core.c/wait_for_debi_done().
Good catch! Pretty subtle, though...
Johannes
