On Thu, Mar 23, 2017 at 06:51:07PM +0000, Eads, Gage wrote:

Hi Jerin,

Hi Gage,

<snip>

quoted

 +force_inline uint16_t __hot
 +ssows_deq_timeout(void *port, struct rte_event *ev, uint64_t
 +timeout_ticks) {
 +	struct ssows *ws = port;
 +	uint64_t iter;
 +	uint16_t ret = 1;
 +
 +	ssows_swtag_wait(ws);
 +	if (ws->swtag_req) {
 +		ws->swtag_req = 0;
 +	} else {
 +		ret = ssows_get_work(ws, ev);
 +		for (iter = 1; iter < timeout_ticks && (ret == 0); iter++)
 +			ret = ssows_get_work(ws, ev);
 +	}
 +	return ret;
 +}

If I understand this correctly, each ssows_get_work() call will wait up to N ns, where N is the dequeue_timeout_ns value supplied to ssovf_mbox_getwork_tmo_set() in ssovf_configure().

So in ssows_deq_timeout, the wait time is (worst case) timeout_ticks * (N * (ns to tick conversion factor)) ticks, which depends on the user-supplied N at eventdev configuration time.

Perhaps in ssovf_configure, if the RTE_EVENT_DEV_CFG_PER_DEQUEUE_TIMEOUT flag is used, the getwork timeout should be set to 1 tick?

I agree. That will improve the timeout accuracy.I will change to
ssovf_mbox_getwork_tmo_set(conf->min_dequeue_timeout_ns) in
ssovf_configure if RTE_EVENT_DEV_CFG_PER_DEQUEUE_TIMEOUT flag is set.

Thanks

Thanks,
Gage