Re: [PATCH net] r8169: fix NAPI handling under high load

[Eric Dumazet <eric.dumazet@gmail.com>]

On 10/16/2018 03:17 PM, Stephen Hemminger wrote:
> On Tue, 16 Oct 2018 22:37:31 +0200
> Heiner Kallweit <hkallweit1@gmail.com> wrote:
> 
>> rtl_rx() and rtl_tx() are called only if the respective bits are set
>> in the interrupt status register. Under high load NAPI may not be
>> able to process all data (work_done == budget) and it will schedule
>> subsequent calls to the poll callback.
>> rtl_ack_events() however resets the bits in the interrupt status
>> register, therefore subsequent calls to rtl8169_poll() won't call
>> rtl_rx() and rtl_tx() - chip interrupts are still disabled.
>>
>> Fix this by calling rtl_rx() and rtl_tx() independent of the bits
>> set in the interrupt status register. Both functions will detect
>> if there's nothing to do for them.
>>
>> This issue has been there more or less forever (at least it exists in
>> 3.16 already), so I can't provide a "Fixes" tag. 
>>
>> Signed-off-by: Heiner Kallweit <hkallweit1@gmail.com>
> 
> Another issue is this:
> 
> 	if (work_done < budget) {
> 		napi_complete_done(napi, work_done);
> 
> 		rtl_irq_enable(tp, enable_mask);
> 		mmiowb();
> 	}
> 
> 	return work_done;
> }
> 
> 
> The code needs to check return value of napi_complete_done.
> 
> 	if (work_done < budget &&
> 	    napi_complete_done(napi, work_done) {
> 		rtl_irq_enable(tp, enable_mask);
> 		mmiowb();
> 	}
> 
> 	return work_done;
> }
> 
> Try that, it might fix the problem and your logic would
> be unnecessary 
> 

Well, I do not believe this is related.

Testing napi_complete_done() is not mandatory, it is only an optimization [1]
and only for busy polling users.

In short, by default, this should not matter, since busy-polling is not enabled by default.


[1] busy polling users are spinning anyway, so it is not even clear if this
   is really an optimization, unless maybe the cost of irq enabling is really really high...