[ltt-dev] net isntrumentation: duplicated receive

[Benjamin Poirier]

On 24/10/10 09:38 PM, Mathieu Desnoyers wrote:
> * Benjamin Poirier (benjamin.poirier at polymtl.ca) wrote:
>> On 22/10/10 04:26 PM, Mathieu Desnoyers wrote:
>>> Hi Benjamin,
>>>
>>> Can you explain why you need two receive tracepoints in the recv path ?
>>
>> One tracepoint is in netif_rx() and is fired by drivers that use the
>> pre-NAPI receive path. The other tracepoint is in netif_receive_skb()
>> and is fired by drivers that use the NAPI receive path.
>>
>>>
>>> "The receive tracepoint is moved earlier in the netif_receive_skb() function.
>>> The tracepoint is also duplicated in the netif_rx() function. This enables
>>> more precise measurements for pre-NAPI drivers."
>>>
>>> @@ -2518,9 +2520,13 @@ int netif_rx(struct sk_buff *skb)
>>>           if (netpoll_rx(skb))
>>>                   return NET_RX_DROP;
>>>
>>> +       trace_net_dev_receive(skb);
>>> +
>>>           if (netdev_tstamp_prequeue)
>>>                   net_timestamp_check(skb);
>>>
>>> +       trace_net_dev_receive(skb);
>>>
>>> The side-effect is that two events will be generated for each packet, no ?
>>
>> No. Depending on the receive path, only one of the two tracepoints is
>> hit. netif_rx() is specific to pre-NAPI drivers. It eventually leads to
>> __netif_receive_skb() and therefore avoids the tracepoint in
>> netif_receive_skb(). NAPI drivers just don't call netif_rx() and
>> therefore avoid the tracepoint in there.
>>
>> It would be possible to have only one tracepoint in
>> __netif_receive_skb() but that would make the event be generated later
>> on the receive path (in all cases). One of the goals when tracing packet
>> reception is to get a timestamp as close as possible to the packet
>> arrival.
>>
>> Hope this clears things up,
>
> Not really. What I am saying here is that the instrumentation in "netif_rx()"
> really calls "trace_net_dev_receive(skb);" *twice*. Once before the call to
> "net_timestamp_check(skb)" and once after. I really think it is an
> implementation error (or maybe a patch merge error somewhere).

Yeah, that patch hunk looks wonky. The one in the before-last version 
(.232) of the kernel.org lttng repository is fine though.

>
> So all pre-NAPI receive paths which call into netif_rx will call the
> net_dev_receive event twice logically here, no ?
>
> Thanks,
>
> Mathieu
>