Re: [PATCH bpf-next v3 01/10] net: introduce preferred busy-polling

["Björn Töpel" <bjorn.topel@intel.com>]

On 2020-11-24 01:04, Jakub Kicinski wrote:
> On Thu, 19 Nov 2020 09:30:15 +0100 Björn Töpel wrote:
>> +	/* The NAPI context has more processing work, but busy-polling
>> +	 * is preferred. Exit early.
>> +	 */
>> +	if (napi_prefer_busy_poll(n)) {
>> +		if (napi_complete_done(n, work)) {
>> +			/* If timeout is not set, we need to make sure
>> +			 * that the NAPI is re-scheduled.
>> +			 */
>> +			napi_schedule(n);
>> +		}
>> +		goto out_unlock;
>> +	}
> 
> Do we really need to go through napi_complete_done() here?
> 
> Isn't it sufficient to check:
> 
> 	if (napi_prefer_busy_poll(n) &&
> 	    hrtimer_active(&n->timer)) // not 100% sure this is the
> 	                               // right helper for the check
> 
> If timer is scheduled it will fire and worst case sirq will kick back
> in after timeout. napi_complete_done() should had been called by the
> driver already to schedule the timer. If the driver doesn't call
> napi_complete_done() we should not allow it to use busy_poll() anyway.
> 

No, it's not. For a heavy traffic load, the napi_complete_done() will
never be called by the driver. It'll just keep on spinning in the
ksoftirqd. This code is to force out of that loop, so we need to call
napi_complete_done() explicitly (which will set the timeout).

Without the explicit napi_complete_done(), the ksoftirqd will not stop,
and the busy-polling will never allow to enter.


Björn