On Tue, May 31, 2016 at 09:53:55AM -0700, Stephen Hemminger wrote:
> On Sun, 29 May 2016 22:16:46 +0530
> Jerin Jacob <jerin.jacob at caviumnetworks.com> wrote:
>
> > +
> > +static struct itimerspec alarm_time = {
> > + .it_interval = {
> > + .tv_sec = 0,
> > + .tv_nsec = NICVF_INTR_POLL_INTERVAL_MS * 1000000,
> > + },
> > + .it_value = {
> > + .tv_sec = 0,
> > + .tv_nsec = NICVF_INTR_POLL_INTERVAL_MS * 1000000,
> > + },
> > +};
> > +
> > +static void
> > +nicvf_interrupt(struct rte_intr_handle *hdl __rte_unused, void *arg)
> > +{
> > + struct nicvf *nic = (struct nicvf *)arg;
> > +
> > + nicvf_reg_poll_interrupts(nic);
> > +}
> > +
> > +static int
> > +nicvf_periodic_alarm_start(struct nicvf *nic)
> > +{
> > + int ret = -EBUSY;
> > +
> > + nic->intr_handle.type = RTE_INTR_HANDLE_ALARM;
> > + nic->intr_handle.fd = timerfd_create(CLOCK_MONOTONIC, TFD_NONBLOCK);
> > + if (nic->intr_handle.fd == -1)
> > + goto error;
> > + ret = rte_intr_callback_register(&nic->intr_handle,
> > + nicvf_interrupt, nic);
> > + ret |= timerfd_settime(nic->intr_handle.fd, 0, &alarm_time, NULL);
> > +error:
> > + return ret;
> > +}
> > +
> > +static int
> > +nicvf_periodic_alarm_stop(struct nicvf *nic)
> > +{
> > + int ret;
> > +
> > + ret = rte_intr_callback_unregister(&nic->intr_handle,
> > + nicvf_interrupt, nic);
> > + ret |= close(nic->intr_handle.fd);
> > + return ret;
> > +}
>
> It would be good to have real link status interrupts or just report that
> device does not support Link State interrupt and let application poll.
> Having another
> thing going on (timerfd callback) seems like it would add more complexity to
> the environment of an already complex thread structure with DPDK.
But we would still need some polling infrastructure for some 'async'
mbox events and error events from PF.So, I think I can change to
rte_eal_alarm* infrastructure like bond pmd driver. Will fix it in V3.
>
> Also, timerfd() doesn't exist on all OS's.
