Re: [dpdk-dev] [Linuxarm] Re: [PATCH V2 04/14] net/hns3: add Rx and Tx bytes stats

[Ferruh Yigit <ferruh.yigit@intel.com>]

On 3/3/2021 2:08 PM, oulijun wrote:
> 
> 
> 在 2021/3/3 21:28, Ferruh Yigit 写道:
>> On 3/2/2021 1:58 PM, Lijun Ou wrote:
>>> From: "Min Hu (Connor)" <humin29@huawei.com>
>>>
>>> In current HNS3 PMD, Rx/Tx bytes from packet stats are not
>>> implemented.
>>>
>>> This patch implemented Rx/Tx bytes using soft counters.
>>> Rx/Tx bytes stats will be enabled if the macro
>>> RTE_LIBRTE_HNS3_PMD_SOFT_COUNTERS is defined.
>>>
>>> Signed-off-by: Min Hu (Connor) <humin29@huawei.com>
>>> Signed-off-by: Lijun Ou <oulijun@huawei.com>
>>> ---
>>>   drivers/net/hns3/hns3_rxtx.c          | 24 ++++++++++++++++++++++++
>>>   drivers/net/hns3/hns3_rxtx_vec_neon.h | 15 +++++++++++++++
>>>   drivers/net/hns3/hns3_rxtx_vec_sve.c  | 11 +++++++++++
>>>   drivers/net/hns3/hns3_stats.c         | 22 ++++++++++++++++++----
>>>   4 files changed, 68 insertions(+), 4 deletions(-)
>>>
>>> diff --git a/drivers/net/hns3/hns3_rxtx.c b/drivers/net/hns3/hns3_rxtx.c
>>> index 5e79177..a8bd2cc 100644
>>> --- a/drivers/net/hns3/hns3_rxtx.c
>>> +++ b/drivers/net/hns3/hns3_rxtx.c
>>> @@ -2181,6 +2181,10 @@ hns3_recv_pkts(void *rx_queue, struct rte_mbuf 
>>> **rx_pkts, uint16_t nb_pkts)
>>>                              cksum_err);
>>>           hns3_rxd_to_vlan_tci(rxq, rxm, l234_info, &rxd);
>>> +#ifdef RTE_LIBRTE_HNS3_PMD_SOFT_COUNTERS
>>> +        /* Increment bytes counter  */
>>> +        rxq->basic_stats.bytes += rxm->pkt_len;
>>> +#endif
>>
>> copy/paste from previous version:
>>
>> Why statistics enabled only with macro?
>> It is not common to use macro to enable the stats, what do you think to remove 
>> it, to be consistent with rest of the PMDs?
> I'm sorry. I thought it was a success.
>   Firstly, the macro is used to control the statistics to ensure performance and 
> facilitate flexible usage. For example, the macro needs to be disabled when high 
> performance is required.
> secondly the byte statistics of other vendors are implemented by reading and 
> writing registers. Therefore, macros are not used.By the way, the MLX driver has 
> a precedent (code snippets can be intercepted here).

It is not convenient for a user re-compile the DPDK to be able to get the ethdev 
byte statistics, stats are not developer/debug information, end user may need 
them. And this recompilation may not be an option for the distributed software.

How much performance affect it has if you enable it always?

And just to double check, isn't there a way to get this information from HW 
without calculating it in the driver?

The compile time flags, specially after meson switch, hard to enable and a 
little hidden, it is very easy to have broken code there, that is why it better 
to prevent compile time flags as much as possible.