On Wed, Jun 3, 2015 at 7:54 AM, Bruce Richardson <bruce.richardson at intel.com > wrote:
> On Wed, Jun 03, 2015 at 02:31:47PM +0800, Selmon Yang wrote: > > Hi, > > > > I found that, in dpdk 2.0, rte_eal_alarm_set() is affected by > > discontinuous jumps in the system time because eal_alarm_callback() > > and rte_eal_alarm_set() use gettimeofday() to get the current time. > > > > Here is what I encountered. > > I set up a rte eal alarm as below, and I like it to be triggered every > second. > > #define USE_PER_S 1000 * 1000 > > void my_alarm_cb(void *arg) > > { > > /* send heartbeat signal out, etc. */ > > > > rte_eal_alarm_set(1 * US_PER_S, my_alarm_cb, NULL); > > return; > > } > > > > int main(void) > > { > > /* ..., do something */ > > rte_eal_alarm_set(1 * US_PER_S, my_alarm_cb, NULL); > > /* ... do something else */ > > } > > > > It works fine in most of time. > > However, if I change system time manually, it is possible that rte alarm > > function works out of my expectation. > > Suppose that current time is 11:00:00 AM, and eal_alarm_callback() > > is triggered because I executed > > rte_eal_alarm_set(1 * US_PER_S, my_alarm_cb, NULL) at 10:59:59 AM. > > eal_alarm_callback() gets the current time (11:00:00 AM) > > and calls my_alarm_cb() as below. > > while ((ap = LIST_FIRST(&alarm_list)) !=NULL && > > gettimeofday(&now, NULL) == 0 && > > (ap->time.tv_sec < now.tv_sec || > > (ap->time.tv_sec == now.tv_sec && > > ap->time.tv_usec <= > > now.tv_usec))){ > > ap->executing = 1; > > ap->executing_id = pthread_self(); > > rte_spinlock_unlock(&alarm_list_lk); > > > > ap->cb_fn(ap->cb_arg); > > > > rte_spinlock_lock(&alarm_list_lk); > > > > LIST_REMOVE(ap, next); > > rte_free(ap); > > } > > > > In my_alarm_cb(), rte_eal_alarm_set() is called again. > > rte_eall_alarm_set() gets the current time (11:00:00 AM), plus 1 second, > > and adds the new alarm entry to alarm_list. > > /* use current time to calculate absolute time of alarm */ > > gettimeofday(&now, NULL); > > > > new_alarm->cb_fn = cb_fn; > > new_alarm->cb_arg = cb_arg; > > new_alarm->time.tv_usec = (now.tv_usec + us) % US_PER_S; > > new_alarm->time.tv_sec = now.tv_sec + ((now.tv_usec + us) / > US_PER_S); > > > > rte_spinlock_lock(&alarm_list_lk); > > if (!handler_registered) { > > ret |= rte_intr_callback_register(&intr_handle, > > eal_alarm_callback, NULL); > > handler_registered = (ret == 0) ? 1 : 0; > > } > > > > if (LIST_EMPTY(&alarm_list)) > > LIST_INSERT_HEAD(&alarm_list, new_alarm, next); > > else { > > LIST_FOREACH(ap, &alarm_list, next) { > > if (ap->time.tv_sec > new_alarm->time.tv_sec || > > (ap->time.tv_sec == > > new_alarm->time.tv_sec && > > > > ap->time.tv_usec > new_alarm->time.tv_usec)){ > > LIST_INSERT_BEFORE(ap, new_alarm, next); > > break; > > } > > if (LIST_NEXT(ap, next) == NULL) { > > LIST_INSERT_AFTER(ap, new_alarm, next); > > break; > > } > > } > > } > > > > After the new alarm entry is added to alarm_list, if current time is > > set to 8:00:00 AM manually, the current time in eal_alarm_callback() > > will be updated to 8:00:00 AM, too. > > Then the new alarm entry will be triggered after 3 hours and 1 second. > > > > I think rte alarm should not be affected by discontinuous jumps in > > the system time. > > I tried to replace gettimeofday() with > clock_gettime(CLOCK_MONOTONIC_RAW, &now), > > and it looks work fine. > > What do you think about this modification? > > Will you consider to modify rte_alarm functions to be not affected > > by discontinuous jumps in the system time? > > I agree with you that the alarm functionality should not be affected by > jumps > in system time. If you have a patch that fixes this bug, it would be great > if > you could upstream it here. > > Thanks, > /Bruce > I haven't looked through the RTE alarm code, but one thing to consider is whether you want to use CLOCK_MONOTONIC_RAW or just CLOCK_MONOTONIC. The RAW version is ~10x slower than the CLOCK_MONOTONIC. CLOCK_MONOTONIC isn't completely protected from NTP frequency adjustments, but it won't have discontinuities. We've found the rte_eal_alarm calls to be surprisingly (ie., we hadn't bothered to look at the implementation yet) variable and intermittently slow, so the 10x difference on the call to clock_gettime() may be relevant. Jay