Re: [PATCH] ethdev: document that stats reset APIs are not thread-safe

["Mattias Rönnblom" <hofors@lysator.liu.se>]

On 2024-04-29 09:57, Morten Brørup wrote:
>> From: Mattias Rönnblom [mailto:hofors@lysator.liu.se]
>> Sent: Sunday, 28 April 2024 17.49
>>
>> On 2024-04-26 14:20, Morten Brørup wrote:
>>>> From: Ferruh Yigit [mailto:ferruh.yigit@amd.com]
>>>> Sent: Thursday, 25 April 2024 18.53
>>>>
>>>> Making 'rte_eth_stats_reset()' and 'rte_eth_xstats_reset()' APIs thread
>>>> safe has performance impact on datapath.
>>>>
>>>> Instead document APIs as not thread safe and add condition for reliable
>>>> stats reset functionality, forwarding should be stopped.
>>>
>>
>> What does "forwarding" mean here? Sounds like something from testpmd.
>>
>> Isn't what needs to stop here other threads calling into the ethdev API
>> for this particular device?
> 
> Doesn't suffice.
> The NIC hardware counters increment if the NIC's DMA engine receives or transmits packets.
> E.g. in the RX direction, the NIC hardware counters increment according to what the NIC detects happening on the wire. If some packets are DMA'ed into memory and the associated RX descriptors are filled, the NIC's hardware counters progress. It doesn't matter if the software has looked at those RX descriptors or not.
> 
>>
>>> I'm not sure stopping forwarding suffices.
>>> NIC hardware counters will keep progressing unless RX and TX is stopped at
>> NIC level.
>>>
>>
>> Why would that be an issue?
> 
> My point is:
> Unless you stop everything, including the NIC hardware RX & TX DMA, the counters might not be zero when returning from the stats_reset() functions.
> 

The reset can still be MT safe, even though there is no way to retrieve 
the exact counter state prior to the reset and also is no guarantee that 
the counters are zero immediately after the function returns (whatever 
that means).

If you do want to be able to do that, the means to that end is to teach 
the reset function to return the prior counter state as a part of the 
reset call, not to force the user to have the system grind to a halt.

>>
>>> I don't have any suggestions for a better wording, though. :-(
> 
> I think it is fine providing a notice about MT-unsafety in the API documentation.
> But the only way to prevent it is useless in real applications, as it would cause packet loss. So it's not helpful describing how to do that.
> 
>>>
>>> Anyway, better with the patch than without...
>>> Acked-by: Morten Brørup <mb@smartsharesystems.com>
>>>